Description:
1 Title of the Invention
AI based Aircraft Airspeed Prediction.
2 Field of the Invention
The invention presented herewith in general relates to aviation industry and is Artificial Intelligence based method and apparatus for prediction of airspeed of an aircraft in real time and usage of the said predicted airspeed to provide redundancy to a critical air data measurement in case of an emergency and for fault detection and fault isolation of air data system
3 Background of the Invention
Airspeed is one of the basic parameters in flying an aircraft. The most common and widely used arrangement for measuring airspeed is Pitot - static tube mounted typically near the nose of an aircraft. This sensor measures the total and static pressure and the difference, which is dynamic pressure, is then used by Air Data Computers (ADC) on-board aircraft to compute the airspeed.
The airspeed value is not only used by pilots while flying for reference, but also, in modern aircrafts, by a host of systems such as Automatic Flight Control System (AFCS), Health and Usage monitoring System (HUMS), Full Authority Digital Engine Control (FADEC) system, Flight Data Recorders (FDR) etc. Airspeed is therefore a very critical parameter and failure of the Airspeed measurement and Indicator system, be it due to Pitot - static tube blockage because of icing, flying debris, impact damage etc or ADC failure, can prove to be even catastrophic.
The most practical solution to handle such failures is to have redundancy.
3.1 Prior Art
A most often used approach is to have physical hardware redundancy. This includes dual, triple or even more hardware redundancy both in terms of plurality of sensor which in this case is typically pitot – static tube and also plurality of sensor data acquisition and processing units such as ADC.
However, hardware redundancy comes with its own challenges. Each redundant system adds to the weight, increases power consumption requirement, constraints already cramped space and increases maintenance and logistic activities in an aircraft, thus affecting the performance and mission parameters. Another challenge in dual redundant systems is to identify which of the systems is accurate/ reliable especially in an event of discrepancy in the values as measured by the redundant systems.
Another approach is to have analytical redundancy, in place of hardware redundancy, in form of Virtual Sensor, which is in principle, a computational model that estimates the state or property of an entity with information from extra / alternate measurements.
Some of the proposed solutions for a virtual sensor are to use Kalman Filters. Application of Kalman Filters and its improvements in terms of Unscented Kalman Filter (UKF), Extended Kalman Filter (EKF), and Adaptive Extended Kalman Filter etc. However, Kalman Filter based approaches require healthy sensor signals to start the estimation and also are observed to sometimes diverge. The Kalman Filter based approaches are mostly not independent of the pitot-static or other speed measurement systems.
Other proposed solutions have used various input parameters acquired from different sensors onboard aircraft to compute airspeed. Some of these solutions have used neural networks to establish relationship between the input parameters and airspeed. However, many of these solutions call for sensors and instrumentation such as forward rake angle of rotor, velocity vector of rotor blade, airflow etc. that are not available on an average flying aircraft, even by modern standards.
Some solutions require weight and Center of Gravity (CG) parameters of aircraft for which there do not exist both accurate and widely available methods of estimation.
Some other solutions have utilized few air data parameters like altitude, temperature, air density etc., which may not be available in case of failure of air data measurement system. Also, it may be noted that more the number of parameters are used for estimation, the more the probability for error in estimation of airspeed, due to inaccuracy or non-availability of these parameters.
The objective of the present invention is to address the afore-mentioned problems by a low cost, small footprint, easily maintainable virtual sensor using a continuously learning Artificial Intelligence network based method and apparatus to predict airspeed using only non-pneumatic parameters.
Another objective of this invention is to use parameters which are available ordinarily in any aircraft and not require any additional sensors and instrumentation.
Yet another objective of this invention is to reduce the number of input parameters which are deemed to be necessary and sufficient for the purpose of airspeed prediction with reasonable accuracy and reliability to be fit for aviation industry application.
4 Brief Summary of the Invention
In an aspect described herein, AI based speed prediction for aircraft comprises of CNN architecture as its core. The CNN architecture is trained using flight test data. In this present invention, the pre-trained CNN model is used to predict the speed with using data only from non-pneumatic sensors. The input to the pre-trained model comprises only of control inputs and gyro sensors outputs namely pitch attitude, roll attitude, pitch rate, roll rate and yaw rate. The uniqueness of this invention is in its very small input sample space. As the input parameters are very less, this present invention is suitable for implementation on a dedicated airborne hardware or can be easily integrated with an existing airborne line replace unit.
This invention also encompasses a unique technique for batch incremental training of the pre-trained model. This technique ensures that the model weights and biases get tuned to the aircraft on which this invention is installed. A fault detection and isolation technique is also part of this invention. This aspect of the invention helps fault detection and isolation in case of multi, dual or single ADU configuration system using the predicted airspeed output from the CNN.
5 Detail Description of the Drawings
The foregoing objects and other advantages of the present invention will be more fully understood by reference to the following description taken in conjunction with the accompanying drawings wherein like reference numerals refer to like or corresponding elements throughout and wherein:
Figure 1: is a symbolic representation of one embodiment of the present invention installed on an aircraft;
Figure 2: is a flow chart of a method for practicing the present invention in case of dual ADU sensor configuration;
Figure 3: is a flow chart of a method for practicing the present invention in case of single ADU sensor configuration;
Figure 4: is a representation of CNN architecture with feedback used as part of present invention;
Figure 5: is a flow chart of a method for practicing online batch-wise incremental tuning of the neural network;
6 Detail Description of the Invention
A detailed description of the present invention is provided herein for purposes of explanation and not limitation, specific details are provided herewith to provide a better understanding of the present disclosure. It will be apparent to one skilled in the art that the present disclosure may be practiced in other embodiments that depart from these specific details.
An embodiment of the present invention is disclosed in Figure 1. Aircraft (10) control inputs (12) from pilot control sticks are measured using displacement sensors such as potentiometers, LVDTs etc. and acquired (14) and processed to convert to their engineering unit values in terms of percentage of full displacement. In another embodiment, in addition to pilot control stick, the aircraft (10) control surfaces can be controlled by Automatic Flight Control System (AFCS), if installed in the aircraft (10). In such embodiment, the degree of control surface actuation as commanded by AFCS is converted to its equivalent pilot control stick position displacement and added to the physical pilot control stick position displacement to get final effective aircraft (10) control input (12) parameter value. In yet another embodiment, if there exist any system on the aircraft (10) which can actuate the aircraft (10) control surfaces, the equivalent effect of actuation of the aircraft (10) control surfaces by said system in terms of pilot control stick position displacement is computed and added to physical control stick position displacement to get the final control inputs (12). In an embodiment of the present invention these control inputs (12) typically are pilot collective stick position, pilot pedal position, and pilot cyclic stick position in both lateral and longitudinal directions, which are control input parameters in a helicopter.
In addition to control input (12) parameters, gyroscopic sensors are used to measure gyro input (16) parameters such as the attitude of the aircraft (10) in particular pitch attitude and roll attitude parameters, attitude rate of the aircraft (10), namely pitch rate, roll rate and yaw rate. These measurements are acquired (14) and processed to convert to their respective engineering unit values.
The control input (12) parameters and gyroscopic sensor parameters (16) as disclosed herein form necessary and sufficient non-pneumatic input parameters referred to as ‘input flight parameters’ (12 and 16) hereinafter for prediction of airspeed. These parameters do not require any additional complex instrumentation and are usually available in almost all aircrafts.
The input parameters (12 and 16) are stored into a buffer (18) at a pre-determined uniform sampling rate of ‘F’ Hz. In case of parameters whose acquisition unit samples the parameter at a lower rate than the pre-determined sampling rate, the parameter is up-sampled before storing into buffer (18) and in case of parameters whose acquisition unit samples the parameter at a higher rate than the pre-determined sampling rate then the parameter is down-sampled before storing into buffer (18). In other embodiments, the uniform sampling rate of the input parameter (12 and 16) can range from 1 Hz to any higher limit as deemed fit for the purpose and practical from computation point of view.
In the present invention, the said buffer (18) is a dedicated unit present in the prediction device. In another embodiment, the buffer (18) is a shared memory used in a mission computer installed on-board aircraft (10). In yet another embodiment, the said buffer (18) is shared memory used in Automatic Flight Control Computer (AFCC) installed on-board aircraft (10). Other embodiments are possible wherein the said buffer (18) is shared memory used by a computing device available on-board aircraft (10) which has access to the said input parameters (12 and 16).
The said buffer (18) is used to store the input flight parameters (12 and 16) at pre-determined sampling rate. A total of latest ‘X’ samples, per input flight parameter, is stored in the buffer (18) at any point of time, by discarding the oldest sample set as soon as the newest sample set arrives. In other embodiments, different number of samples may be stored to be considered for the prediction task.
In addition to the input flight parameters (12 and 16), the predicted airspeed values are also stored in the buffer (18). The number of samples of predicted airspeed stored in the buffer (18) is the same as what is used for input flight parameters (12 and 16).
On initialization of the processor (20) at the start of a flight, the said processor (20) starts a Neural Network (22) initialization routine. The initialization routine is presented in Figure 5 and described herein later. The processor (20) fetches the final selected weights and biases from memory (24) and programs the neural network (22). After programming, the processor (20) fetches the samples of input flight parameter (12 and 16) and previous predicted airspeed values from the buffer (18), normalizes the inputs and feeds it to the input of neural network (22).
The artificial neural network (22) presented in this invention is a 1D Convolutional Neural Network (CNN). The architecture of this network is as shown in Figure 4. The input to the CNN network (300) is the input flight parameters (12 and 16). The said network consists of first a CNN layer (302) with ‘I’ number of kernel filters. The size of the kernel filter in the present invention is ‘T’. This layer is followed by a Max pooling layer (304). Another layer of CNN (306) with ‘J’ number of kernel filters follows. Here too, the size of the kernel filter is ‘T’. This layer is again followed by a Max pooling layer (308). The output of this layer is fed to a dense layer (310) with ‘K’ neurons. The output of dense layer is fed to a single neuron which gives the predicted airspeed output (312). The activation function used in this last neuron is linear activation. In all the previous layers, the activation function used in Rectified Linear (ReLU).
As mentioned earlier, the predicted airspeed (312) is stored into the buffer (18) just the way the input flight parameters (12 and 16) are stored and fetched along with the input flight parameters (12 and 16) to form the final input to the neural network (22). This means that the predicted airspeed output (312) is fed back to the input of neural network (22) for next prediction thus forming a closed loop AI based prediction network (22).
Other embodiments of this invention are possible and not limited by changing the number of CNN and pooling layers, changing the number of kernel filters used in each CNN layer, changing the size of kernel filter, changing pooling layer from Max pooling to Min or Average pooling etc., changing the number of denes layers, inclusion or deletion of bias values, changing the activation function of each layer such as and not limited to tanh, sigmoid, leaky ReLU etc.
Further embodiments are also possible by using other neural network architectures such as and not limited to feed forward network, Jordan network, Elman network, 2D CNN, Recurrent Neural Networks (RNN) like LSTM, GRU etc., Encoder-Decoder architecture, Transformers, Auto-encoders etc.
The training process of neural network (22) used in the present invention is as disclosed herein. The training utilizes back propagation method. The optimization algorithm used is ‘ADAM’. L2 regularization, Dropout, Early stopping, adaptive learning rate reduction techniques are used for regularization and combat over-fitting. Root Mean Squared Error (RMSE) is used as loss function. RMSE, Mean Absolute Error (MAE), R2 score and Correlation coefficient have been used as performance metrics. In other embodiments, different training processes by changing optimization technique (even nature inspired techniques like Genetic algorithm, Particle Swarm Optimization, Ant Bee Colony etc.), regularization techniques, loss function, performance metrics in any permutation and combination can be used.
Detailed steps of neural network training process are as disclosed herein:
a) Gathering data
Gather flight dataset consisting of historical flight data of the aircraft (10) of interest. The dataset should be diverse to cover almost the entire flight envelope of the aircraft (10). This includes different altitudes, airspeed, gross weight, CG location, temperature, weather conditions, terrain etc.
b) Data Pre-processing
Select only the input flight parameters (12 and 16) and airspeed. Keep the sampling rate of all parameters to the pre-determined rate. Normalize all the parameter values using min-max normalization. Other normalization techniques may also be used. Prepare data to make it suitable for acceptance by CNN. This preparation is demonstrated in the following example. Consider a flight with ‘N’ lines of data. Let there be ‘P’ parameters. This dataset can be seen as a matrix with dimension ‘N x P’. Let the number of samples of data required for prediction of airspeed by the neural network (22) be ‘M’. Now split the dataset to form a tensor of shape ‘N x M x P’. This essentially means that there are ‘N’ data-points of dimension ‘M x P’. This can be achieved by selecting first ‘M’ lines of data with ‘P’ parameters as one data-point. The output corresponding to this data-point is obtained by shifting the column of airspeed up by one sample and then choosing the ‘Mth’ airspeed value as output. Next data-point is formed by starting from the second row of the dataset and selecting ‘M’ lines of data. Similarly, the corresponding output value will be ‘(M+1)th’ value in the shifted airspeed column. Vertically stack the data-points to form the aforementioned tensor. Repeat these steps for all the flights in the dataset. Let the final input dataset be of shape ‘Y x M x P’ after inclusion of all flights, then the corresponding output dataset shape is ‘Y x 1’. Randomize the order of this input – output dataset pair to get better training results.
c) Dataset splitting
Divide the dataset into training, validation and testing sets. The training set is used to train the neural network (22), while the validation set helps assess its performance on unseen data during the training process.
d) Initialization
Initialize the weights and biases of the neural network (22) randomly or using a specific initialization method to start the learning process.
e) Forward Propagation
Feed the training data forward through the network (22) to obtain predictions. Compute the loss, which represents the difference between the predicted airspeed and the actual airspeed.
f) Backward Propagation (Back-propagation)
Utilize optimization algorithms (e.g.,ADAM) to minimize the loss by adjusting the weights and biases backward through the network (22). This involves computing gradients with respect to the loss and updating the parameters accordingly.
g) Epochs and Batches
Repeat the forward and backward propagation steps for multiple epochs, with each epoch representing one pass through the entire training dataset. To improve efficiency, use mini-batches of data rather than the entire dataset in each epoch.
h) Validation
Periodically assess the performance of the neural network (22) on the validation set to monitor for over-fitting or under-fitting. Reduce the learning rate when the loss on validation data starts plateauing. This is known as ‘Adaptive learning rate reduction’. Stop the training once the validation data loss starts plateauing even after learning rate reduction. This is known as ‘Early stopping’.
i) Hyper-parameter Tuning
Experiment with various hyper-parameters (e.g., learning rate, batch size) to find the optimal configuration that results in the best airspeed predictions.
Another critical feature of this invention is a novel Online Batch-wise Incremental Training (26) module. Figure 5 represents the various steps involved in the neural network (22) initialization and online training process. Essentially two models of the neural network are used in online training and real time prediction process. The architectures of both the models are same as described in Figure 4. The trainable parameters of the prediction network are updated at the start of a flight based on the performance of weights and biases updated at the end of previous flight. The details of the neural network (22) initialization and online training as used in this invention are as presented herein. The pre-trained model weights, biases and performance metrics are stored in memory (24) at location ‘1’. These weights are loaded from memory (step 152) on power ON (step 150). The online batch-wise incrementally trained weights and biases obtained from training in previous flight are stored in memory (24) at location ‘2’ and read from the memory (24) during initialization (step 154). The test set containing flight data is stored in memory (24). This test dataset consists of aircraft flights which encompass the flight envelope as well as a few recent flights. This blend ensures a model with good fit for real flight data. A model with neural network architecture as shown in Figure 4 model is initialized with the weights and biases from location ‘2’ and tested with the test set data (step 156). If the performance metrics on test data set are better than the model loaded with weights and biases from location ‘1’ (step 158), then the weights and biases are updated in the location ‘1’ (step 160). If the performance metrics are not improved, then the weights and biases are disregarded (step 170). The updated model performance metrics are also updated in memory (24) at location ‘1’ (step 162).
Initialize the prediction model as well as the replica with the weights and biases available at location ‘1’ (step 164) in the memory (24). The training is started only after the flight condition is detected (step 166). Once a batch of ‘B’ samples of input flight parameters (12 and 16) along with actual airspeed from the ADU and predicted airspeed is received, the training of the replica neural network is initiated. Adaptive learning rate reduction technique is used during training. The training process for a batch is stopped when the training data loss starts plateauing or when the training is completed for a pre-determined number of epochs. After training on one batch of data, the model the trains on the next available latest batch of ‘B’ samples. At the end of flight, when ground condition is detected, the weights and biases of this online batch-wise incrementally trained model is stored in memory (24) at location ‘2’ (step 168). The online training process as disclosed herein is a novel process and designed considering the start-up duration requirement of airborne devices as well as the in-flight computation load on the processor bearing in mind the fact that the training and convergence process is non-deterministic in time.
In the present invention, the use of predicted airspeed (312) is envisaged but not limited to fault detection and fault isolation (28) of ADUs (30). Two embodiments of this method are presented in detail herein.
Figure 2 demonstrates the various steps involved in the identifying the faulty ADU in case of an embodiment with dual ADU configuration. Both ADU airspeed data is continuously acquired by the processor (20) (step 200). The difference between both the ADUs is continuously monitored by the processor (20) (step 202). In case, the difference between the ADUs exceeds a predefined threshold (Th1) for a predefined time (T1) (step 202), then (step 204) the predicted speed is used to find the malfunctioning ADU (step 206). The algorithm checks the closeness of the predicted speed with the both ADU speeds. The faulty ADU is identified when the closeness test fails (exceed a threshold Th2) for more than predefined time (T2 in secs) (step 208). The faulty ADU is isolated (step 210). In case of a remote possibility of both ADUs failing in the closeness test, then the digital predicted airspeed is disregarded. In such a scenario, the normal operating procedure in case of discrepancy between the ADU is followed. This embodiment can be easily adapted for aircrafts with more than two ADU systems (30). In such embodiments, a faulty system is diagnosed and isolated by following a majority voting logic wherein, in case of discrepancy between the values of these ADUs, identified by difference is airspeed measurement greater than a pre-defined threshold value (Th1) for a pre-defined time (T1), an ADU is flagged as faulty if its airspeed value differs from the majority. In this method of detection and isolation, when the number of healthy ADUs comes down to two, the method as described in Figure 2 kicks in.
Figure 3 represents the various steps involved in the identifying fault in ADU in case of an embodiment with single ADU configuration. The ADU airspeed data is continuously acquired by the processor (20) (step 250). In case, the difference between the ADU speed and the predicted airspeed exceeds a predefined threshold (Th1) for a predefined time (T1) (step 252), then (step 254) the pilot is indicated of a possible ADU failure (step 256).
The predicted airspeed values are not limited in their usage by these embodiments. The predicted airspeed can be fed to other consuming systems (32) as well such a Pilot Display Unit, Flight Data Recorder etc.
, C , Claims:We Claim,
1. A method and apparatus for AI based airspeed prediction for aircrafts comprising:
a collection of sensors in plurality for measuring aircraft control input parameters (12);
a gyro based sensor for measuring aircraft attitude and attitude rates (16);
a memory buffer device (18) for storing pre-determined number of samples of the said control input parameters (12), said attitude parameters, said attitude rate parameters (16) and predicted airspeed parameter (312);
a non-volatile memory (24) for storing test flight data, pre-trained weights and biases and online batch-wise incrementally updated weights and biases;
an AI based network (22) with closed loop output feedback for predicting airspeed;
a novel online batch-wise incremental training method (26) for training of said neural network (22);
a fault detection and isolation system (28) for using the predicted airspeed value for detecting and isolating faulty Air Data Unit (ADU) (30); and
a collection of consuming systems (32) in plurality for utilizing predicted airspeed;
wherein said sensors measure control input parameters (12), attitude and attitude rate parameters (16) and store said parameters (12,16) in said memory buffer device (18) which are fed as inputs to said closed loop AI based network (22), initialized with pre-trained weights and biases fetched from said non-volatile memory (24) and trained using said online batch-wise incremental training method (26), to predict airspeed which is further used by said fault detection and isolation system (28) to identify faulty ADU (30) and by said consuming systems (32).
2. The method and apparatus as claimed in claim 1, wherein control input parameters (12) comprise of pilot control inputs which actuate the aircraft (10) control surfaces and control input from any system that has authority to actuate the aircraft control surfaces.
3. The method and apparatus as claimed in claim 1, wherein the said attitude and attitude rate parameters (16) comprise of pitch attitude, roll attitude, pitch rate, roll rate and yaw rate parameters.
4. The method and apparatus as claimed in claim 2 and claim 3 wherein the said parameters together constitute the necessary and sufficient parameters for prediction of airspeed.
5. The method and apparatus as claimed in claim 1, wherein the said AI based network (22) is a pre-trained 1-D Convolutional Neural Network (CNN).
6. The method and apparatus as claimed in claim 1, wherein the said online batch-wise incremental training module trains a replica model of the said AI based network (22) in near real time and the said AI based network (22) is updated in next flight after evaluation of performance of incrementally trained model on test data.
7. The method and apparatus as claimed in claim 1, wherein the said fault detection and isolation system (28) is used for fault detection and isolation in aircrafts with Multi ADU system or Single ADU system configurations.