DESC:FORM 2
The Patent Act 1970
(39 of 1970)
&
The Patent Rules, 2003

COMPLETE SPECIFICATION
(See Section 10 and Rule 13)

TITLE OF THE INVENTION

A SYSTEM AND METHOD FOR ESTIMATING A STATE OF CHARGE (SOC) OF A BATTERY

KPIT Technologies Limited
Plot -17, Rajiv Gandhi Infotech Park,
MIDC-SEZ, Phase-III, Maan, Hinjawadi,
Taluka-Mulshi, Pune 411057, Maharashtra, India
an Indian Company

The following specification particularly describes the invention and the manner in which it is to be performed:

TECHNICAL FIELD
The present invention relates to a battery and more particularly related to a system and method for accurately estimating a state of charge (SOC) of the battery.
BACKGROUND OF INVENTION
The automotive industry is currently experiencing a paradigm shift from conventional, diesel and gasoline-propelled vehicles into second-generation hybrid and electric vehicles. The electric vehicles are mainly equipped with a Rechargeable battery, which is the most critical component in the electric vehicle.
It is very important for a user to determine remaining percentage of charge available in the battery. State of charge (SOC) estimation is one of the most important functions of battery management systems (BMSs), which is defined as the percentage of the remaining charge inside the battery to its maximum capacity. The SOC gives an indication to the user that when the battery needs to be recharged. The battery SOC estimation is very crucial for many battery management functions, for example, but not limited to, charge/discharge control, remaining useful time/ driving range predictions, and battery power capability estimations. As an essential performance indicator, the SOC reflects a residual capacity of the battery. To ensure a safe operation of systems, it is very essential to estimate the SOC of the battery accurately. However, an inaccurate SOC estimation of the battery lead to the user dissatisfaction, mission failures, inefficient performance of the vehicles, premature battery failures, etc. While a system or component exhibits degradation during its life cycle, there are various methods to predict its future performance and assess the time frame until it no longer performs its desired functionality.
The conventional techniques that are currently being used in the vehicles for the estimation of the SoC of Li-ion batteries are highly dependent on accuracy of physical modeling of the Li-ion batteries, i.e., accuracy in measuring current (I), and an initial value of the SoC when we start using the battery. Based on these dependencies, existing solutions not only suffer in terms of accuracy in predicting the SOC of batteries, but also in terms of the speed of prediction.
There is a trade-off between the speed and accuracy of the SoC estimation. Most of the battery management systems calculate SoC using the Coulomb Counting based approach which has its own limitations. The limitations mainly include the requirement of the initial state, dependency only on the current (I) value, and error accumulation over time. Thus, the limitations enable the Coulomb Counting based approach less accurate in practical scenarios. The other conventional method used for the estimation of the SoC of the battery is the Kalman Filters which has higher accuracy but involve a higher-order matrix inversion while updating the state value which is computationally expensive. Hence, they are mathematically more complex which makes the estimation speed very slow. The Kalman filter based SoC estimation also requires recalibration of initial values and kalman constants. In case of some wrong initial state, it takes significantly larger amount of time to settle to the true state of SoC.
Further, the current machine learning methods are lacking generalization ability due to their data-driven nature and, they are sensitive to the amount and quality of training data. Further, in the case of similar values for two different regions of flat SOC, existing methods may get confused and predict the wrong SOC value. All these limitations hinder the model’s effectiveness in estimating the SOC of the battery.
Thus, there is a need for the battery SOC estimation system and method which utilizes data driven techniques and overcomes the above limitations. There is a need for a system and method for accurately estimating the SOC of the battery. There is a need for a system for the SOC estimation that is way faster than the bulky and time-consuming solutions like the Kalman filter. Further, there is also a need for a system for the battery SOC estimation that does not require knowledge of the battery model and the initial states of various parameters. There is also a need for the battery SOC estimation system that eliminates the need for recalibration. Further, there is need for a system for the SOC estimation that is simpler, more accurate and easily employable on the vehicle/on the fly. There is need for a system for the battery SOC estimation that, irrespective of similar values of two different regions of flat SOC predicts the SOC in the correct region. There is also a need for a real time battery SOC estimation which is quick and accurate.
Background description includes information that may be useful in understanding the present invention. It is not an admission that any of the information provided herein is prior art or relevant to the presently claimed invention, or that any publication specifically or implicitly referenced is prior art.
OBJECT OF INVENTION
The principal object of the embodiments herein is to provide a system and method for accurately estimating a state of charge (SOC) of a battery.
Yet another object of the embodiments herein is to provide a data-driven module for accurately estimating the SOC.
Yet another object of the embodiments herein is to provide a system and method for a battery SOC estimation that eliminates a need for recalibration of a various battery parameters.
Yet another object of the embodiments herein is to provide a system and method for a battery SOC estimation that does not require knowledge of the battery model and initial states of various parameters.
Yet another object of the of the embodiments herein is to provide a system and method for a battery SOC estimation that is efficient and quick and not time consuming and bulky like the Kalman filter.
Yet another object of the embodiments herein is to provide a system and method for a battery SOC estimation that irrespective of similar values of two different regions of flat SOC predicts the SOC in a correct region.
Yet another object of the embodiments herein is to provide a system and method for a battery SOC estimation that is simpler, more accurate and easily employable on the vehicle.
Yet another object of the embodiments herein is to provide a system and method for a battery SOC estimation for a hybrid/electric vehicle battery.
Still another object of the embodiments herein is to provide a system and method for a battery SOC estimation that can be implemented over edge device and can predict the SOC value in real time/on the fly.
SUMMARY
Accordingly, the embodiments herein provide a method for estimating a state of charge (SOC) of a battery. The method comprising extracting raw features from the battery at different time intervals, wherein the raw features comprise a current, a voltage and a temperature of the battery. Further, the method comprises receiving the extracted raw features at different time intervals as an input. Further, the method comprises generating by the pre-processor module one or more processed outputs corresponding to the extracted raw features at different time intervals to derive one or more feature vectors, wherein each of a processed output of the one or more processed outputs comprises a current entropy (IE), a voltage entropy (VE), and a difference in open circuit voltage (DOCV), the voltage, the current and the temperature. Further, the method comprises transmitting by the pre-processor module the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals, to a deep learning module. Further, the method includes estimating by the deep learning module the SOC of the battery using the derived one or more feature vectors from the one or more processed outputs corresponding to the different time intervals, wherein the deep learning module uses an attention-based Long short-term memory (LSTM) network for estimating the SOC of the battery.
In an embodiment, wherein generating each of the processed output comprises enabling by a first feature extraction module (FE1) of the pre-processor module to receive the current from the battery module to generate the current entropy (IE), enabling by a second feature extraction module (FE2) of the pre-processor module to receive the current and the voltage from the battery module to generate the difference in open circuit voltage (Docv), enabling by a third feature extraction module (FE3) of the pre-processor module to receive the voltage from the battery module to generate the voltage entropy (VE) and enabling by at least one filter to receive the voltage and the temperature from the battery module to generate a noise free voltage and a noise free temperature.
In an embodiment, wherein the deep learning module is adapted to differentiate two different constant SOC regions using the current entropy (IE), the voltage entropy (VE) and difference in open circuit voltage (Docv) to predict the SOC in a correct region.
In an embodiment, wherein the method is adapted to work over an edge device to predict the SOC value in real time.
Accordingly, the embodiments herein provide a system for estimating a state of charge (SOC) of a battery. The system comprising a battery module configured to extract raw features from the battery at different time intervals, wherein the raw features include a current, a voltage and a temperature of the battery. Further, the system comprising a pre-processor module configured to receive the extracted raw features at different time intervals as an input. Further, the pre-processor module configured to generate one or more processed outputs corresponding to the extracted raw features at different time intervals to derive one or more feature vectors, wherein each of a processed output of the one or more processed outputs comprises a current entropy (IE), a voltage entropy (VE), and a difference in open circuit voltage (DOCV), the voltage, the current and the temperature. Further, the pre-processor module configured to transmit the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals, to a deep learning module. Further, the system comprising a deep learning module configured to estimate the SOC of the battery using the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals, wherein the deep learning module uses an attention-based Long short-term memory (LSTM) network for estimating the SOC of the battery.
In an embodiment, wherein the pre-processor module configured to generate each of the processed output by: enabling a first feature extraction module (FE1) of the pre-processor module to receive the current from the battery module to generate the current entropy (IE); enabling a second feature extraction module (FE2) of the pre-processor module to receive the current and the voltage from the battery module to generate the difference in open circuit voltage (Docv); enabling a third feature extraction module (FE3) of the pre-processor module to receive the voltage from the battery module to generate the voltage entropy (VE); and enabling at least one filter of the pre-processor module to receive the voltage and the temperature from the battery module to generate a noise free voltage and a noise free temperature.
Accordingly, the embodiments herein provide a system for estimating a state of charge (SOC) of a battery of a hybrid/electric vehicle. The system comprising a Cell Management Unit (CMU) configured to receive a current, a voltage, and a temperature signals from one or more battery packs of the vehicle, wherein the current, the voltage and the temperature signals are received from respective sensors associated with the one or more battery packs of the electric vehicle. Further, the CMU configured to transmit the current, the voltage, and the temperature signals received from the one or more battery packs of the electric vehicle to a Battery Management System (BMS) present in an Electric Controlling Unit (ECU) of the electric vehicle, over an asynchronous communication protocol. Further, the BMS configured to convert the current, the voltage, and the temperature signals into a physical current, a cell voltage, and a module temperature to generate a feature vector. Further, the BMS configured to transmit the generated feature vector to a deep learning module to predict the SOC of the battery using the generated feature vector. Further, the BMS configured to communicate the predicted SOC of the battery to a user of the electric vehicle through an audio-visual display unit of the electric vehicle.
These and other aspects of the embodiments herein will be better appreciated and understood when considered in conjunction with the following description and the accompanying drawings. It should be understood, however, that the following descriptions, while indicating preferred embodiments and numerous specific details thereof, are given by way of illustration and not of limitation. Many changes and modifications may be made within the scope of the embodiments herein without departing from the spirit thereof, and the embodiments herein include all such modifications.
BRIEF DESCRIPTION OF THE DRAWINGS
Embodiments herein are illustrated in the accompanying drawings, throughout which like reference letters indicate corresponding parts in the various figures. The example embodiments herein will be better understood from the following description with reference to the drawings, in which:
FIG. 1 is a system block diagram illustrating a method for estimating a State of Charge (SOC) of a battery, according to an embodiment as disclosed herein;
FIG. 2 illustrates a resistive circuit representing a battery module, according to an embodiment as disclosed herein;
FIG. 3 illustrates a pre-processing module for generating one or more processed outputs corresponding to the extracted raw features at different time intervals to derive one or more feature vectors, according to an embodiment as disclosed herein;
FIG. 4 illustrates a restructuring block diagram for merging and transmitting the one or more feature vectors derived from one or more processed outputs corresponding to the different time intervals to the deep learning module, according to an embodiment as disclosed herein;
FIG. 5 is a table illustrating different values of various battery features at different time instances, according to an embodiment as disclosed herein;
FIG. 6 illustrates a deep learning module architecture for battery SOC estimation, according to an embodiment as disclosed herein;
FIG. 7 illustrates one node of the LSTM layer, according to an embodiment as disclosed herein;
FIG.8 is flow diagram illustrating a method for estimating the SOC of the battery, according to the embodiments as disclosed herein;
FIG.9 is an example illustration of a system for estimating a state of charge (SOC) of a battery of a hybrid/electric vehicle, according to an embodiment as disclosed herein; and
FIG.10 is flow diagram illustrating a method for estimating the SOC of the battery of a hybrid or electric vehicle, according to the embodiments as disclosed herein.

DETAILED DESCRIPTION OF THE INVENTION
The embodiments herein and the various features and advantageous details thereof are explained more fully with reference to the non-limiting embodiments that are illustrated in the accompanying drawings and detailed in the following description. Descriptions of well-known components and processing techniques are omitted, so as not to unnecessarily obscure the embodiments herein. The examples used herein are intended merely to facilitate an understanding of ways in which the embodiments herein can be practiced and to further enable those of skill in the art to practice the embodiments herein. Accordingly, the examples should not be construed as limiting the scope of the embodiments herein.
The embodiments herein provide a method for estimating a state of charge (SOC) of a battery. The method comprising extracting raw features from the battery at different time intervals, wherein the raw features include a current, a voltage and a temperature of the battery. Further, the method comprises receiving the extracted raw features at different time intervals as an input. Further, the method comprises generating one or more processed outputs corresponding to the extracted raw features at different time intervals to derive one or more feature vectors, wherein each of a processed output of the one or more processed outputs comprises a current entropy (IE), a voltage entropy (VE), and a difference in open circuit voltage (DOCV), the voltage, the current and the temperature. Further, the method comprises transmitting the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals, to a deep learning module. Further, the method comprises estimating the SOC of the battery using the derived one or more feature vectors from the one or more processed outputs corresponding to the different time intervals, wherein the deep learning module uses an attention-based Long short-term memory (LSTM) network for estimating the SOC of the battery.
The embodiments herein provide a system for estimating a state of charge (SOC) of a battery. The system comprising a battery module configured to extract raw features from the battery at different time intervals, wherein the raw features include a current, a voltage and a temperature of the battery. Further, the system comprising a pre-processor module configured to receive the extracted raw features at different time intervals as an input. Further, the pre-processor module configured to generate one or more processed outputs corresponding to the extracted raw features at different time intervals to derive one or more feature vectors, wherein each of a processed output of the one or more processed outputs comprises a current entropy (IE), a voltage entropy (VE), and a difference in open circuit voltage (DOCV), the voltage, the current and the temperature. Further, the pre-processor module configured to transmit the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals, to a deep learning module. Further, the system comprising a deep learning module configured to estimate the SOC of the battery using the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals, wherein the deep learning module uses an attention-based Long short-term memory (LSTM) network for estimating the SOC of the battery.
The embodiments herein provide a system for estimating a state of charge (SOC) of a battery of a hybrid/electric vehicle. The system comprising a Cell Management Unit (CMU) configured to receive a Current, a Voltage, and a Temperature signals from one or more battery packs of the vehicle, wherein the current, the voltage and the temperature signals are received from respective sensors associated with the one or more battery packs of the electric vehicle. Further, the CMU configured to transmit the current, the voltage, and the temperature signals received from the one or more battery packs of the electric vehicle to a Battery Management System (BMS) present in an Electric Controlling Unit (ECU) of the electric vehicle, over an asynchronous communication protocol. Further, the BMS configured to convert the current, the voltage, and the temperature signals into a physical current, a cell voltage, and a module temperature to generate a feature vector. Further, the BMS configured to transmit the generated feature vector to a deep learning module to predict the SOC of the battery using the generated feature vector. Further, the BMS configured to communicate the predicted SOC of the battery to a user of the electric vehicle through an audio-visual display unit of the electric vehicle.
Referring now to the drawings, and more particularly to FIGS. 1 through 10, where similar reference characters denote corresponding features consistently throughout the figures, there are shown preferred embodiments.
FIG. 1 is a system 100 block diagram illustrating a method for estimating a State of Charge (SOC) of a battery, according to an embodiment as disclosed herein.
The embodiments herein disclose the system 100 and method for estimating the SoC of the battery. The method comprises extracting by a battery module 102, raw features 104 from the battery at different time intervals. The raw features 104 include a current I, a voltage V, and a temperature T of the battery. In an embodiment, as illustrated in FIG. 2, a simple resistive circuit 200 is used as the battery module 102, wherein R is a hyper-parameter tuned based on the extracted raw features 104. Further, the method comprises receiving by a pre-processor module 106, the extracted raw features 104 at different time intervals as an input to generate one or more processed outputs/Input features 108 corresponding to the extracted raw features to derive one or more feature vectors, wherein each of a processed output of the one or more processed outputs comprises a current entropy (IE), a voltage entropy (VE), and a difference in open circuit voltage (DOCV), the voltage, the current and the temperature. Further, the method comprises transmitting by the pre-processor module 106, the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals to a deep learning module 110. Further, the method comprises estimating by the deep learning module 110, the SOC of the battery using the derived one or more feature vectors from the one or more processed outputs corresponding to the different time intervals. The deep learning module 110 uses an attention-based Long short-term memory (LSTM) network for estimating the SOC of the battery.
FIG. 1 shows exemplary units of the system 100, but it is to be understood that other embodiments are not limited thereon. In other embodiments, the system 100 may include less or more number of units. Further, the labels or names of the units are used only for illustrative purpose and does not limit the scope of the embodiments herein. One or more units can be combined to perform same or substantially similar function in the system 100.
FIG. 2 illustrates the resistive circuit representing the battery module 102, according to an embodiment as disclosed herein. In an embodiment, the difference in open circuit voltage (DOCV) is calculated using the simple resistive circuit 200 as the battery module 102 in which the R is treated as the hyper-parameter and tuned based on the on the extracted raw features 104.
OCVt = Vt - It * R, R is hyperparameter
?OCVt = OCVt - OCVt-1
FIG. 3 illustrates a pre-processing module 106 for generating one or more processed outputs 108 corresponding to the extracted raw features 104 at different time intervals to derive one or more feature vectors, according to an embodiment as disclosed herein.
The pre-processor module 106 comprises one or more Feature Engineering (FE) blocks i.e., FE1, FE2 and FE3, one or more filters and a restructuring block 302. The raw features 104 extracted from the battery module 102, i.e., current I, and voltage V are given as input to the Feature Engineering (FE) blocks.
In an embodiment, the current I and the voltage V are given as inputs to the FE-1 and FE-2. The FE-1 calculates the current entropy IE based on the current I. The FE-2 calculates a difference in an open circuit voltage Docv based on the current I and voltage V. Further, the voltage V is given as input to the FE-3 which estimates the voltage entropy VE based on the voltage V. The Voltage V and the temperature T are given as input to a restructuring block 302 after passing through the one or more filters. In an embodiment, the one or more filters used can be any filter known in the art, including, but not limited to, a Savgol filter. Finally, the current entropy IE, the voltage entropy VE, the open circuit voltage Docv, the current I, the voltage V, and the temperature T through the restructuring block 302 to generate the one or more processed outputs to supply as input features to the deep learning module.
Current Entropy (IE): A ramp variable which measures the duration of time for which value of current I keep changing beyond the threshold.
I(t) – I(t-1) ? [-a, + a], a is a Hyperparameter
Voltage Entropy (VE): A ramp variable which keeps track of the Voltage V by monitoring the period of time for which voltage variation was negligible.
V(t) – V(t-1) ? [-ß,+ß],ß is a Hyperparameter
FIG. 4 illustrates a restructuring block 302 diagram for merging and transmitting the one or more feature vectors derived from the one or more processed outputs 108 corresponding to the different time intervals to the deep learning module 110, according to an embodiment as disclosed herein. The restructuring block 302 generates sets of individual inputs/feature vectors corresponding to the different time intervals.
One feature vector = Xt = [It, Vt, Tt, DOCVt, IEt, VEt]
According to the embodiments, the restructuring block 302 merges n feature vectors together corresponding to the different time intervals and transmits to the deep learning module 110, instead of feeding the individual feature vector at once. This helps in improving the accuracy of the SoC prediction. The value of n is a hyperparameter and can vary based on the dataset. As illustrated in FIG. 4, each orange box represents one input unit to a LSTM model.
Xt = {x(t-time_steps), x(t-time_steps+1), x(t-time_steps+2), .........., x(t-1), x(t)}
Xt+1 = {x(t-time_steps+1), x(t-time_steps+2), x(t-time_steps+3), ......., x(t), x(t+1)}
Xt+2 = {x(t-time_steps+2), x(t-time_steps+3), x(t-time_steps+4), ......, x(t+1), x(t+2)}

Xt+N = {x(t-time_steps+N), x(t-time_steps+N+1), x (t-time_steps+N+2) ...., x(t+N), x(t+N)}
In the above equation, time step (w) which can be tuned as per the dataset; and t represents present time.
FIG. 5 is a table illustrating different values of various battery features at different time instances according to an embodiment as disclosed herein. The pre-processor module 106 receive the raw features 104 such as the current I, the voltage V, and the temperature T from the battery module 102. Further, the pre-processor module 106 uses the raw features and generates processed outputs/input features. Finally, the pre-processor module 106 transmits the processed outputs/input features 108 as an input to the deep learning module 110. The deep learning module 110 uses an attention-based LSTM network for estimating the SOC of the battery.
FIG. 6 illustrates the deep learning module 110 architecture for battery SOC estimation, according to an embodiment as disclosed herein.
According to the embodiments, the pre-processor module 106 transmits the one or more feature vectors derived from the one or more processed outputs 108 corresponding to the different time intervals to the deep learning module 110 to estimate the SoC of the battery, wherein each of a processed output of the one or more processed outputs comprises the current entropy (IE), the voltage entropy (VE), and the difference in open circuit voltage (DOCV), the voltage V, the current I and the temperature T. The Current entropy and voltage entropy measure the amount of fluctuations in the current and the voltage feature. These two features help an attention-based LSTM model to learn flat SoC values easily. For two different constant regions, values of raw features i.e., current I, voltage V and temperature T remain the same. Hence, the LSTM model gets confused about which region to predict. However, the features such as the current entropy (IE) the voltage entropy (VE) and difference in open circuit voltage (Docv) enables the deep learning module 110 to differentiate the two different constant SOC regions and enables prediction of the SOC in a correct region.
The problem of estimating the State of Charge of the battery can be represented as,
SoCt = attention_lstm ([Xt-w+1, Xt-w+2, Xt-w+3, ......., Xt])
where Xt is the feature vector which is formed from one or more features from the list of current, voltage, temperature, IE, VE and Docv; t is a historical vector at sampling time t; the sliding window (SW) length is set as w ? [1, m] and m is the length of Xt. This m is a hyperparameter which can be tuned based on different datasets. SoCt is the estimated value of State of Charge at the sampling time t.
As illustrated in FIG. 6, the deep learning module 110 architecture contains one or more layers. The one or more layers further comprises but not limited to an attention layer, one or more LSTM layers, a batch normalization layer, and a dense layer.
In an embodiment, the attention layer is adapted to assign weightage tags to the one or more feature vectors received from the pre-processor module 106 and transmit the the one or more feature vectors along with weightage tags to one or more LSTM layers. The one or more LSTM layers are adapted to use the weighed tags associated with the one or more feature vectors to read an input sequence of the one or more feature vectors to accumulate one or more necessary features in a cell state i.e., memory cell. The one or more LSTM layers arranged in a stack enables in extracting a deeper meaning in the one or more feature vectors. The one or more LSTM layers are a type of a recurrent neural network and best deep-learning architectures for a time series prediction. In an embodiment, the one or more LSTM layers reads the feature vector sequence from backward and forward direction simultaneously using an overlapping sequence formed by the sliding window, wherein the length of the sliding window is treated as an hyperparameter and tuned based on a performance of the deep learning module 110 on different datasets. The batch normalization layer is adapted to optimize the pace at which the relationships and underlying patterns in one or more features vectors are being learnt. The dense layer observes all inputs and different dimensions of the inputs arrived at the dense layer to map the inputs to a corresponding output.
FIG. 7 illustrates one node of the LSTM layer according to an embodiment as disclosed herein.
According to the embodiments, the system 100 for the battery SOC estimation uses, for example, 200 LSTM nodes in the first layer and 50 LSTM nodes in the second layer. It is to be noted that these number of layers are exemplary only, and maybe modified as per the application requirements.
According to an embodiment of the present invention, the deep-learning module uses an attention mechanism and is referred as attentive neural network. Referring to FIG. 4, one orange block represents the input sequence for the attention layer. The attention layer helps in two things:
The attention layer allows the one or more LSTM layers to capture whole traffic dynamics (like differentiation, cross-correlation) in each feature vector input sequence. The attention layer informs later layers (i.e., LSTM layer) to accumulate one or more necessary feature vectors needed to attend over while reading an input sequence and accumulating to a representation in the cell state.
In an embodiment, when the future vector sequence [ I, V, T, DOCV, IE, VE] are passed through the attention mechanism, the attention layer sends the feature vector sequence to the LSTM layer with weightage tags assigned to every feature so that the consecutive layers will use those tags to read the input sequence and accumulate in cell-states i.e., memory cell. For example, if attention layer informs that for a particular input sequence, temperature feature is less important, then the LSTM layer does not save T values in the memory cell state. In an embodiment, stacking of two layers of the LSTM helps in extracting the deeper meanings hidden in a data sample. The accuracy of the deep learning module increases as the network becomes deeper. But as a consequence, it increases prediction time, and the deep learning module will become too bulky to use. Hence, the number of layers is optimized so as to provide a greater accuracy, while not increasing the prediction time or making the deep learning bulky.
In an embodiment, the deep learning module 110 architecture contains the batch normalization layer. Introduction of batch normalization layers ensures a faster learning process. As the data size increases, learning gets slower. Hence, to increase the efficiency of learning process, Batch Normalization is introduced. The batch normalization layer is adapted to optimize the pace at which the relationships and underlying patterns in one or more features vectors are being learnt.
In an embodiment, the deep learning module 110 architecture contains the dense layer is a fully connected layer. This layer observes all the inputs and different dimensions of the inputs that come to this layer and finally maps all those inputs to the corresponding output. Suppose at the dense layer, 100*100 matrix enters as an input, but since there is only 1 SOC output, this dense layer maps 100*100 matrix to a single SOC output.
According to an embodiment of the present invention, a reference SOC is taken from a dataset which is pre-obtained from a perfectly tuned Kalman filter. The reference SOC is more accurate and can be assumed to be perfectly correct as the system uses a perfectly tuned Kalman filter, which can be validated with the voltage error produced by the Kalman filter. The deep learning module does not use the reference SOC directly based on coulomb counting. However, the deep learning module trains and learns from the reference SOC, which is already validated and thus makes the hypothesis stronger and improves the accuracy.
The system 100 and method of the present invention has several benefits and advantages. The system 100 and method of the present invention greatly improves the accuracy of the SOC estimation. The system 100 and method of the present invention does not require knowledge of battery model and initial states of various parameters. Further, the system 100 and method of the present invention eliminates the need for recalibration of various battery parameters. The method of the present invention is way faster than the existing bulky solutions like Kalman filter. The system 100 and method of the present invention irrespective of similar values of two different regions of flat SOC predicts the SOC in the correct region. The system and method of the present invention is simpler, more accurate and easily employable on the vehicle. The system 100 and method of the present invention can be implemented over edge device and can predict the SOC value in real time/on the fly.
FIG.8 is flow diagram 800 illustrating a method for estimating the SOC of the battery, according to the embodiments as disclosed herein.
At step 802, the method comprises extracting the raw features from the battery at different time intervals. The method allows the battery module 102 to extract the raw features from the battery at different time intervals. The raw features include a current, a voltage and a temperature of the battery.
At step 804, the method comprises receiving the extracted raw features at different time intervals as an input. The method allows the pre-processor module 106 to receive the extracted raw features at different time intervals as an input.
At step 806, the method comprises generating one or more processed outputs corresponding to the extracted raw features at different time intervals to derive one or more feature vectors. The method allows the pre-processor module 106 to generate one or more processed outputs corresponding to the extracted raw features at different time intervals to derive one or more feature vectors.
At step 808, the method comprises transmitting the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals, to a deep learning module 110. The method allows the pre-processor module 106 to transmit the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals, to the deep learning module 110.
At step 810, the method comprises estimating the SOC of the battery using the derived one or more feature vectors from the one or more processed outputs corresponding to the different time intervals. The method allows the the deep learning module 110 to estimate the SOC of the battery using the derived one or more feature vectors from the one or more processed outputs corresponding to the different time intervals.
The various actions, acts, blocks, steps, or the like in the method and the flow diagram 800 may be performed in the order presented, in a different order or simultaneously. Further, in some embodiments, some of the actions, acts, blocks, steps, or the like may be omitted, added, modified, skipped, or the like without departing from the scope of the invention.
FIG.9 is an example illustration of a system 900 for estimating a state of charge (SOC) of a battery of a hybrid/electric vehicle, according to an embodiment as disclosed herein.
The system 900 comprising a Cell Management Unit (CMU) 906 configured to receive a current, a voltage, and a temperature signals from one or more battery packs 902 of the vehicle, wherein the current, the voltage and the temperature signals are received from respective sensors 904 associated with the one or more battery packs 902 of the electric vehicle. Further, the CMU 906 is configured to transmit the current, the voltage, and the temperature signals received from the one or more battery packs 902 of the electric vehicle to a Battery Management System (BMS) 910 present in an Electric Controlling Unit (ECU) 908 of the electric vehicle, over an asynchronous communication protocol, like UART. Further, the BMS 910 of the system 900 configured to convert the current, the Voltage, and the temperature signals into a physical current, a cell voltage, and a module temperature to generate a feature vector. Further, the BMS 910 is configured to transmit the generated feature vector to a deep learning module 912 to predict the SOC of the battery using the generated feature vector. Further, the BMS 910 is configured to communicate the predicted SOC of the battery to a user of the electric vehicle through an audio-visual display unit 914 of the electric vehicle over a Controller Area Network (CAN). The estimated SoC value enables the user of the electric vehicle to know about a distance that he can drive based on the charge available, and/or the SOC value may be shared to the appropriate vehicle controls for further use, as required.
According to the embodiments, the system 900 accurately estimates the state of charge of the Li-ion batteries used in hybrid/electric vehicles. The system 900 and method can also be used for estimating the state of charge of various types of batteries used in various applications, including, but not limited to, hybrid vehicle, electric vehicle, inverter, industrial applications, consumer gadgets, consumer appliances, etc.
FIG. 9 shows exemplary units of the system 900, but it is to be understood that other embodiments are not limited thereon. In other embodiments, the system 900 may include less or more number of units. Further, the labels or names of the units are used only for illustrative purpose and does not limit the scope of the embodiments herein. One or more units can be combined to perform same or substantially similar function in the system 900.
FIG.10 is flow diagram 1000 illustrating a method for estimating the SOC of the battery of a hybrid or electric vehicle, according to the embodiments as disclosed herein.
At step 1002, the method comprises receiving the current, voltage, and temperature signals from one or more battery packs of the vehicle. The method allows the CMU 906 to receive the current, voltage, and temperature signals from one or more battery packs of the vehicle.
At step 1004, the method comprises transmitting the current, the voltage, and the temperature signals received from the one or more battery packs of the electric vehicle to the BMS 910 present in an Electric Controlling Unit (ECU) 908 of the electric vehicle, over an asynchronous communication protocol. The method allows the CMU 906 to transmit the current, the voltage, and the temperature signals received from the one or more battery packs of the electric vehicle to the BMS 910 present in the ECU 908 of the electric vehicle, over an asynchronous communication protocol.
At step 1006, the method comprises converting the current, the voltage, and the temperature signals into a physical current, a cell voltage, and a module temperature to generate a feature vector. The method allows the BMS 910 to convert the current, the voltage, and the temperature signals into a physical current, a cell voltage, and a module temperature to generate a feature vector.
At step 1008, the method comprises transmitting the generated feature vector to the deep learning module. The method allows the BMS 910 to transmit the generated feature vector to the deep learning module 912.
At step 1010, the method comprises enabling the deep learning module 912 to predict the SoC of the battery using the generated feature vector. The method allows the BMS 910 to enable the deep learning module 912 to predict the SoC of the battery using the generated feature vector.
At step 1012, the method comprises communicating the predicted SOC of the battery to a user of the electric vehicle through an audio-visual display unit 914 of the electric vehicle. The method allows the BMS 910 to communicate the predicted SOC of the battery to a user of the electric vehicle through the audio-visual display unit 914 of the electric vehicle.
. The various actions, acts, blocks, steps, or the like in the method and the flow diagram 1000 may be performed in the order presented, in a different order or simultaneously. Further, in some embodiments, some of the actions, acts, blocks, steps, or the like may be omitted, added, modified, skipped, or the like without departing from the scope of the invention.
In an embodiment, when the deep learning module 912 is trained once, it can predict the correct values of the SoC as soon as an input data is provided to the deep learning module 912.The deep learning module 912 can be easily employable on vehicles and embedded platforms.
In an embodiment, a training result and a test result is illustrated below:
Training Result: wherein MAE = Mean Absolute Error, RMSE = Root Mean Square Error.

Testing Result:

The deep learning model and the traditional models are compared below based on certain parameters:
Constant current value: An offset of magnitude -150 Amps is introduced in the current value for 15 minutes (900 samples). In the below figures, left figure shows the profile of current which is received from Battery Management System (blue) and the current profile after introduction of constant current in between for few minutes. Right side image shows the comparison of prediction of the SoC using a new current value. The MAE of the deep learning method is = 5.39, the MAE of the Coulomb Counting method is = 7.60

Periodic noise insertion: A periodic noise is added to the current value with a period of 20 Seconds and of the magnitude -150 Amps. Below left figure shows the current profile before noise addition (blue) and after noise addition (orange). Below right figure shows the comparison of SoC estimation with a new current. The MAE of the deep learning method is = 5.30, the MAE of the Coulomb Counting method is = 10.17

Wrong initial state because of self-discharge of battery: In some scenarios, wherein the vehicle is charged to 100\% and then the vehicle is left unused for some time. Later, when the vehicle is started, the Coulomb Counting (CC) method can start estimating the SOC value by assuming initial state of 100\%, but the initial state of 100\%, may got discharged by 1\% or 2\%. However, the present deep learning method wherein the current value is not altered, instead initial state in the coulomb counting was 100\%. The below figure shows the results of the CC estimation and the predicted SOC values. The MAE of the deep learning method is = 1.4, the MAE of the Coulomb Counting method is = 4.0.

Offset noise in current value: An offset of 5\% of the maximum current is added to value of current for the entire dataset. Figure below shows the profile of original current and noisy current. Below right figure shows the comparison of the SoC predicted via Kalman filter and the deep-learning method. The MAE of the deep learning method is = 1.86, the MAE of the Kalman Filter = 7.30.

The foregoing description of the specific embodiments will so fully reveal the general nature of the embodiments herein that others can, by applying current knowledge, readily modify and/or adapt for various applications such specific embodiments without departing from the generic concept, and, therefore, such adaptations and modifications should and are intended to be comprehended within the meaning and range of equivalents of the disclosed embodiments. It is to be understood that the phraseology or terminology employed herein is for the purpose of description and not of limitation. Therefore, while the embodiments herein have been described in terms of preferred embodiments, those skilled in the art will recognize that the embodiments herein can be practiced with modification within the spirit and scope of the embodiments as described herein.
,CLAIMS:We claim:
1. A method for estimating a state of charge (SOC) of a battery, the method comprising:
extracting, by a battery module, raw features from the battery at different time intervals, wherein the raw features include a current, a voltage and a temperature of the battery;
receiving, by a pre-processor module, the extracted raw features at different time intervals as an input;
generating, by the pre-processor module, one or more processed outputs corresponding to the extracted raw features at different time intervals to derive one or more feature vectors, wherein each of a processed output of the one or more processed outputs comprises a current entropy (IE), a voltage entropy (VE), and a difference in open circuit voltage (DOCV), the voltage, the current and the temperature;
transmitting, by the pre-processor module, the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals, to a deep learning module; and
estimating, by the deep learning module, the SOC of the battery using the derived one or more feature vectors from the one or more processed outputs corresponding to the different time intervals, wherein the deep learning module uses an attention-based Long short-term memory (LSTM) network for estimating the SOC of the battery.

2. The method as claimed in claim 1, wherein generating each of the processed output comprises:
enabling, by a first feature extraction module (FE1) of the pre-processor module, to receive the current from the battery module to generate the current entropy (IE);
enabling, by a second feature extraction module (FE2) of the pre-processor module, to receive the current and the voltage from the battery module to generate the difference in open circuit voltage (Docv);
enabling, by a third feature extraction module (FE3) of the pre-processor module, to receive the voltage from the battery module to generate the voltage entropy (VE); and
enabling, by at least one filter, to receive the voltage and the temperature from the battery module to generate a noise free voltage and a noise free temperature.

3. The method as claimed in claim 1, wherein transmitting, by the pre-processor module, the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals, to the deep learning module comprising:
merging, by a restructuring process module, the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals; and
transmitting, by the restructuring process module, the merged one or more feature vectors to the deep learning module to estimate the SOC.

4. The method as claimed in claim 1, wherein the deep learning module is adapted to differentiate two different constant SOC regions using the current entropy (IE), the voltage entropy (VE) and difference in open circuit voltage (Docv) to predict the SOC in a correct region.

5. The method as claimed in claim 1, wherein the deep learning module comprises one or more layers,
wherein the one or more layers include an attention layer, one or more Long short-term memory (LSTM) layers, a batch normalization layer, and a dense layer,
wherein the attention layer is adapted to assign weightage tags to the one or more feature vectors received from the pre-processor module and transmit the the one or more feature vectors along with weightage tags to one or more LSTM layers;
wherein the one or more LSTM layers are adapted to use the weighed tags associated with the one or more feature vectors to read an input sequence of the one or more feature vectors to accumulate one or more necessary features in a cell state;
wherein the one or more LSTM layers arranged in a stack enables in extracting a deeper meaning in the one or more feature vectors;
Wherein the one or more LSTM layers reads the feature vector sequence from backward and forward direction simultaneously using an overlapping sequence formed by a sliding window, wherein the length of the sliding window is treated as an hyperparameter and tuned based on a performance of the deep learning module on different datasets.
wherein the batch normalization layer is adapted to optimize the pace at which the relationships and underlying patterns in one or more features vectors are being learnt; and
wherein the dense layer observes all inputs and different dimensions of the inputs arrived at the dense layer to map the inputs to a corresponding output.

6. The method as claimed in claim 1, wherein the one or more LSTM layers for estimating the SOC of the battery, are optimized to provide a balance between accuracy in estimating the SOC and prediction time.

7. The method as claimed in claim 1, wherein the method is adapted to work over an edge device to predict the SOC value in real time.

8. A system for estimating a state of charge (SOC) of a battery, the system comprising:
a battery module configured to:
extract raw features from the battery at different time intervals, wherein the raw features include a current, a voltage and a temperature of the battery;
a pre-processor module configured to:
receive the extracted raw features at different time intervals as an input;
generate one or more processed outputs corresponding to the extracted raw features at different time intervals to derive one or more feature vectors, wherein each of a processed output of the one or more processed outputs comprises a current entropy (IE), a voltage entropy (VE), and a difference in open circuit voltage (DOCV), the voltage, the current and the temperature;
transmit the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals, to a deep learning module; and
the deep learning module configured to:
estimate the SOC of the battery using the one or more feature vectors derived from the one or more processed outputs corresponding to the different time intervals, wherein the deep learning module uses an attention-based Long short-term memory (LSTM) network for estimating the SOC of the battery.

9. The system as claimed in claim 8, wherein the pre-processor module configured to generate each of the processed output by:
enabling a first feature extraction module (FE1) of the pre-processor module to receive the current from the battery module to generate the current entropy (IE);
enabling a second feature extraction module (FE2) of the pre-processor module to receive the current and the voltage from the battery module to generate the difference in open circuit voltage (Docv);
enabling a third feature extraction module (FE3) of the pre-processor module to receive the voltage from the battery module to generate the voltage entropy (VE); and
enabling at least one filter of the pre-processor module to receive the voltage and the temperature from the battery module to generate a noise free voltage and a noise free temperature.
10. A system for estimating a state of charge (SOC) of a battery of a hybrid/electric vehicle, the system comprising:
a Cell Management Unit (CMU) configured to:
receive a current, a voltage, and a temperature signals from one or more battery packs of the vehicle, wherein the current, the voltage and the temperature signals are received from respective sensors associated with the one or more battery packs of the electric vehicle;
transmit the current, the voltage, and the temperature signals received from the one or more battery packs of the electric vehicle to a Battery Management System (BMS) present in an Electric Controlling Unit (ECU) of the electric vehicle, over an asynchronous communication protocol;
the BMS configured to:
convert the current, the voltage, and the temperature signals into a physical current, a cell voltage, and a module temperature to generate a feature vector;
transmit the generated feature vector to a deep learning module;
enable the deep learning model to predict the SOC of the battery using the generated feature vector; and
communicate the predicted SOC of the battery to a user of the electric vehicle through an audio-visual display unit of the electric vehicle.