Unmanned aerial vehicle fault diagnosis method and system fusing random convolution data enhancement and deep learning

Abstract

The invention discloses an unmanned aerial vehicle fault diagnosis method and system fusing random convolution data enhancement and deep learning, and effectively alleviates the problem of poor model generalization and robustness caused by the class imbalance of an unmanned aerial vehicle flight data set. According to the method, a random convolution data enhancement technology is provided, convolution calculation is performed on an original fault sample, diversified enhanced time sequence samples which maintain original fault features but introduce diversified noise and deformation into local details are generated, and weighted fusion is performed on the enhanced time sequence samples and the original samples, so that a fault sample library is effectively expanded, and the balance of a data set is improved; in order to efficiently capture the time sequence characteristics of the unmanned aerial vehicle fault signals, a fault diagnosis model based on a time convolution network is adopted, and a decision basis is provided for the accuracy of the fault diagnosis model. According to the method, the problem caused by data set category imbalance is effectively solved, the fault diagnosis performance can be improved, and a new thought and method are provided for the unmanned aerial vehicle fault detection method based on data driving.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle technology, and in particular to a method and system for unmanned aerial vehicle fault diagnosis integrating random convolution data enhancement and deep learning. Background Art

[0002] In recent years, drone technology has developed rapidly and is widely used in agriculture, logistics, security, surveying and mapping and other fields. However, drones will inevitably encounter various faults during actual operation, such as mechanical failures, sensor failures, communication failures, etc. These failures may affect the efficiency of mission execution, or even cause the drone to crash, resulting in property damage or even casualties. Therefore, how to improve the accuracy and efficiency of drone fault detection and diagnosis has become a key issue that needs to be urgently addressed in the development of drone technology. Traditional drone fault detection and diagnosis methods mainly rely on expert experience, rule bases, and fault tree analysis. These methods have limitations such as relying on expert experience, difficulty in covering all fault types, difficulty in dealing with complex and changeable fault scenarios, and low diagnostic efficiency. With the rapid development of machine learning technology, data-driven drone fault detection and diagnosis methods have gradually become a research hotspot.

[0003] However, such methods rely on a large amount of high-quality fault data for model training, while the acquisition of UAV fault samples faces challenges such as the scarcity of fault sample types, high cost of fault sample acquisition, and difficulty in fault sample labeling. In order to solve the problem of scarce UAV fault samples, data enhancement technology came into being. By performing a series of transformation operations on the original data, new data samples are generated, thereby expanding the scale of the data set and improving the generalization ability and robustness of the model. In the field of UAV fault diagnosis, data enhancement technology can expand the number of fault samples, enrich the types of fault samples, and enhance the robustness of the model. The current mainstream data enhancement methods can be divided into three categories: geometric transformation methods mainly generate new samples by changing the spatial position or shape of the data, such as flipping, translating, rotating, scaling, shearing and other operations, but these methods can usually only change the data in a small range and it is difficult to simulate complex failure modes; noise addition methods, such as Gaussian noise and salt and pepper noise, can simulate some environmental conditions or sensor errors, but their noise types are limited and it is difficult to cover all possible failure modes. Considering that the original flight data of the drone itself has a certain amount of noise, under excessive noise intensity, it may cause data distortion; generative model methods such as adversarial neural networks (GANs) require a large amount of training data for their own training, and the process is relatively complicated, which is prone to mode collapse or poor quality of generated samples. Summary of the invention

[0004] Purpose of the invention: In order to solve the problem of scarcity of fault data sets when training fault diagnosis models mentioned in the above background technology, the present invention provides a UAV fault diagnosis method and system that integrates random convolution data enhancement and deep learning, and proposes random convolution data enhancement technology to effectively expand the fault sample library and improve the balance of the data set.

[0005] Technical solution: To achieve the above-mentioned invention object, the present invention adopts the following technical solution:

[0006] A UAV fault diagnosis method integrating random convolution data enhancement and deep learning includes the following steps:

[0007] (1) Collect flight data of the UAV under normal flight conditions and fault conditions;

[0008] (2) Taking the time when the fault occurs as the boundary point, the data on both sides of the boundary point are divided into sliding windows to obtain window data of equal length under different flight conditions. Each window corresponds to a sample data.

[0009] (3) Randomly generate a set of convolution kernels with random parameters that conform to the normal distribution, and dynamically adjust the convolution width, height, zero padding, and step size so that the convolution input and output have the same size;

[0010] (4) Perform weighted fusion on the sample data and its random convolution results to maximize the retention of the fault characteristics of the original sample while introducing local noise and deformation according to the weights;

[0011] (5) The original sample data, random convolution enhanced sample data and label training are used to develop a fault diagnosis model based on temporal convolutional network to realize UAV fault diagnosis.

[0012] Furthermore, step (1) includes:

[0013] Various sensors installed on the drone collect data during flight, including attitude, position, speed and altitude;

[0014] A fault signal is artificially introduced at the ground station, and the flight data of the UAV under artificially simulated fault conditions is collected, covering all flight parameter trajectories in the time interval from take-off to landing.

[0015] Furthermore, in step (2), for a complete flight data X∈R N×T Sliding window segmentation is performed, where N is the number of flight parameters, T is the time length, and the time when the fault occurs is t f The data on both sides of the boundary point are divided using sliding windows of length M, with a step size of S. The i-th window corresponds to the i-th sample data, and finally the window data of equal length under different flight states are obtained.i ∈R N×M ;

[0016] During the sliding process of the window, for the pre-fault state, the sliding window starts from t=0 until

[0017] t=t f -1 ends, the window index range is W i =X[:,i·S:i·S+M], where the value range of i is

[0018] For the post-fault state, the sliding window starts from t = t f Starting from t = T-1, the window index range is W i =X[:,t f +i·S:T f +i·S+M], where i ranges from

[0019] The label y of the i-th window sample i Depends on the last time point t in its window end With the boundary point t f The relationship between them is expressed as:

[0020]

[0021] 0 indicates a normal flight, and Tag indicates the type of injected fault in the current flight.

[0022] Furthermore, in step (3), for the convolution kernel with width w and height h respectively, its parameters are randomly selected from the normal distribution, expressed as:

[0023]

[0024] Where Keernel[i][j] represents the weight of the convolution kernel at position (i, j). and σ 2 Represents the mean and variance of the normal distribution; in the fault sample Window i ∈R N×M The zero padding Pad and stride during convolution are constrained as follows:

[0025]

[0026] Among them, Pad w and Pad h Indicates the number of zero padding layers in the width and height dimensions of the input sample.

[0027] Furthermore, step (3) includes:

[0028] Generate a k×k random convolution kernel Kernel. Each element of the convolution kernel is randomly selected from a normal distribution with a mean of 0 and a standard deviation of 1, expressed as:

[0029]

[0030] The calculation process of random convolution is as follows: for the input sample x with width W and height H, the border is zero-filled, that is, padding is done in the width W dimension w Layer zero, padding in height H dimension h Layer zero, get the padded sample data x p , and then use the generated random convolution kernel Kernel to the padded data x p Perform convolution calculation with a step size of s to obtain the output y. The calculation process can be expressed as:

[0031]

[0032] The value range of a and b is:

[0033]

[0034] In order to ensure that the width W' and height H' after convolution are consistent with the width W and height H of the input sample, that is, W' = W, H' = H, the following conditions need to be met:

[0035]

[0036] Furthermore, in step (4), for the input sample x and the result of random convolution x rc ,The fusion method is expressed as:

[0037] x new =α*x+(1-α)*x rc

[0038] Where α is the fusion weight, α∈{0.1, 0.3, 0.5, 0.7, 0.9}.

[0039] Furthermore, the loss function of the fault diagnosis model training is designed as follows: Assume that the total number of training set samples is Num, and for one of the fault status categories is y i The input sample x i , the result predicted by the model is The result of the sample after data enhancement by the jth random convolution kernel is predicted by the model: The total task loss is given by the task loss L task and consistency loss L cons Composition, where the task loss L taskUsed to measure the difference between the model prediction results and the true label, consistency loss L cons It is used to measure the consistency between the prediction results of different enhanced samples. The consistency loss is calculated by KL divergence and is expressed as in Represents the fault sample x i The mean of the probability distribution predicted by the model after C times of data augmentation, λ represents the impact of consistency loss.

[0040] A UAV fault diagnosis system integrating random convolution data enhancement and deep learning includes the following modules:

[0041] The UAV flight data collection module is used to collect the flight data of the UAV under normal flight conditions and fault conditions;

[0042] The sliding window segmentation module is used to perform sliding window segmentation on the data on both sides of the boundary point, taking the time when the fault occurs as the boundary point, to obtain window data of equal length under different flight states, and each window corresponds to a sample data;

[0043] The random convolution kernel generation module is used to randomly generate a set of convolution kernels with random parameters that conform to the normal distribution, and dynamically adjust the convolution width, height, zero padding and step size so that the convolution input and output have the same size;

[0044] The weighted fusion module is used to perform weighted fusion on the sample data and its random convolution results to maximize the preservation of the fault characteristics of the original sample while introducing local noise and deformation according to the weights;

[0045] The fault diagnosis model training module is used to use the original sample data, random convolution enhanced sample data and labels to train the fault diagnosis model based on the temporal convolution network to realize UAV fault diagnosis.

[0046] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of the drone fault diagnosis method integrating random convolution data enhancement and deep learning.

[0047] Beneficial effects: The present invention uses a data enhancement method based on random convolution, and performs weighted fusion based on the result of random convolution and the input sample to maximize the maintenance of the fault characteristics of the original sample, while introducing local noise and deformation according to the weight, solving the problem of scarce UAV fault samples, enriching the UAV fault sample library, and training an end-to-end fault diagnosis model based on a temporal convolutional network, improving the accuracy and efficiency of the results of the UAV fault diagnosis model, providing new ideas and methods for data-driven UAV fault detection methods, promoting the development of UAV fault diagnosis technology, improving the safety and reliability of UAVs, and promoting the application and promotion of UAV technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is an overall flow chart of an embodiment of the present invention.

[0049] Figure 2 Detailed flowchart of an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to more clearly understand the working principle and application of the present invention and further explain the present invention, the implementation steps of the present invention will be described in detail below in conjunction with specific embodiments.

[0051] like Figure 1 and Figure 2 As shown, an embodiment of the present invention discloses a method for diagnosing a drone fault by integrating random convolution data enhancement and deep learning, and the main steps are as follows:

[0052] Step 1: Collect the flight data of the drone under normal flight conditions and fault conditions. Mainly includes:

[0053] Various sensors installed on the drone collect data during flight, including attitude, position, speed and altitude;

[0054] A fault signal is artificially introduced at the ground station, and the flight data of the UAV under artificially simulated fault conditions is collected, covering all flight parameter trajectories in the time interval from take-off to landing.

[0055] Specifically, first, a data acquisition platform is built, and ROS (Robot Operating System) is used as the data acquisition platform to build a communication link between the UAV flight control system and the ground station. Based on a variety of sensors installed on the UAV, such as IMU (inertial measurement unit), GPS, barometer, magnetometer, etc., it is used to collect data such as attitude, position, speed, altitude, etc. during the flight. All sensor data is transmitted to the ground station in real time through the ROS message mechanism. When simulating faulty flight, a fault injection module is developed on the ground station to simulate various faults that may occur in the UAV during flight, such as motor failure, rudder failure, elevator failure, aileron failure, etc. The fault injection module artificially introduces fault signals by modifying sensor data or sending control instructions.

[0056] Use the ROS rosbag tool to record all sensor data and fault injection information collected during the flight. rosbag can store data in binary format for easy playback and analysis. Through the above steps, a UAV flight data set containing normal flight data and fault data can be constructed to provide a data basis for subsequent fault diagnosis model training.

[0057] Step 2: Preprocess the flight data and introduce a sliding window-based segmentation method. The moment when the fault occurs is used as the boundary point, and the data on both sides of the boundary point are segmented by sliding windows to obtain window data of equal length under different flight states. Each window corresponds to a sample data.

[0058] Specifically, for a complete flight data X∈R N×T Sliding window segmentation is performed, where N is the number of flight parameters, T is the time length, and the time when the fault occurs is t f The data on both sides of the boundary point are divided using sliding windows of length M, with a step size of S. The i-th window corresponds to the i-th sample data, and finally the window data of equal length under different flight states are obtained. i ∈R N×M ;

[0059] During the sliding process of the window, for the state before the fault, the sliding window starts from t = 0 and ends at t = t f -1 ends, the window index range is W i =X[:,i·S:i·S+M], where the value range of i is

[0060] For the post-fault state, the sliding window starts from t = t f Starting from t = T-1, the window index range is Wi i =X[:,tf +i·S:t f +i·S+M], where i ranges from

[0061] The label y of the i-th window sample i Depends on the last time point t in its window end With the boundary point t f The relationship between them is expressed as:

[0062]

[0063] 0 indicates a normal flight, and Tag indicates the type of injected fault in the current flight.

[0064] In this way, a series of window data of equal length can be obtained, each window corresponds to a sample data, which provides a basis for subsequent analysis and model training.

[0065] Step 3: Introduce random convolution data enhancement and perform random convolution operations on the original fault samples. Randomly generate a set of convolution kernels with random parameters that conform to the normal distribution, and dynamically adjust the convolution width, height, zero padding, and step size so that the size of the convolution input and output are the same.

[0066] Specifically, for a convolution kernel with width w and height h respectively, its parameters are randomly taken from a normal distribution, expressed as:

[0067]

[0068] Where Kernel[i][j] represents the weight of the convolution kernel at position (i, j). and σ 2 Represents the mean and variance of the normal distribution; in the fault sample Window i ∈R N×M The zero padding Pad and stride during convolution are constrained as follows:

[0069]

[0070] Among them, Pad w and Pad h Indicates the number of zero padding layers in the width and height dimensions of the input sample.

[0071] In this embodiment, a k×k random convolution kernel Kernel is first generated, and each element of the convolution kernel is randomly selected from a normal distribution with a mean of 0 and a standard deviation of 1, which is expressed as:

[0072]

[0073] Then, zero padding is performed on the boundary of the input data to obtain the padded data, and the generated random convolution kernel is used to perform convolution calculation on the padded data to obtain the output data. Specifically, the calculation process of random convolution is as follows:

[0074] For the input sample x with width W and height H, zero padding is performed on the border, that is, padding is performed on the width W dimension. w Layer zero, fill with Pad in the height H dimension h Layer zero, get the padded sample data x p , and then use the generated random convolution kernel Kernel to the padded data x p Perform convolution calculation with a step size of s to obtain the output y. The calculation process can be expressed as:

[0075]

[0076] The value range of a and b is:

[0077]

[0078] In order to ensure that the width W' and height H' after convolution are consistent with the width W and height G of the input sample, that is, W' = W, G' = G, the following conditions need to be met:

[0079]

[0080] In this way, diverse enhanced time series samples can be generated that retain the original fault features but introduce diverse noise and deformation in local details.

[0081] Step 4: Introduce the weighted fusion method to perform weighted fusion on the sample data and its random convolution results to maximize the retention of the fault characteristics of the original sample, while introducing local noise and deformation according to the weight.

[0082] In this step, the random convolution result x rc The input sample x is weighted and fused to maximize the fault characteristics of the original sample, while introducing local noise and deformation according to the weight. The final new sample is x new The value range of weight α is 0.1 to 0.9, which ensures that the fused data retains the original features while introducing appropriate amount of noise and deformation.

[0083] In this embodiment, for the input sample x and the random convolution result x rc ,The fusion method is expressed as:

[0084] x new =α*x+(1-α)*x rc

[0085] Where α is the fusion weight, α∈{0.1, 0.3, 0.5, 0.7, 0.9}.

[0086] Step 5: Use the original sample data, random convolution enhanced sample data and labels to train the fault diagnosis model based on the temporal convolution network to realize UAV fault diagnosis.

[0087] In this embodiment, a fault diagnosis model based on a temporal convolutional network (TCN) is introduced to extract features in terms of timing, flight parameter dimensions, fault characteristics, etc. from the original samples and enhanced samples, and based on this, the mapping relationship between learning features and flight status labels is learned. Next, the definition of the relevant model of this embodiment and the structure of each part are introduced in detail, and the key steps and strategies in the training process are explained.

[0088] Temporal Convolutional Network (TCN) is a neural network model specifically designed to process time series data. TCN captures long-distance dependencies in time series by using causal convolution and dilated convolution. The following is a detailed definition of the TCN model:

[0089] 1) Input layer: For an input sample x of shape (w,h), w represents the width of the input sample, which represents the length in the time dimension; h represents the height of the sample, that is, the sensor parameter dimension.

[0090] 2) Input residual connection layer: In order to avoid the problem of gradient vanishing during back propagation as the number of network layers increases, a residual connection edge is established between the input layer and the dilated convolution layer so that the gradient can be directly transmitted through this connection.

[0091] 3) Causal convolution layer: When processing time series data, in order to prevent the model from using future information to predict the output of the current time step, it is necessary to ensure that the convolution operation does not leak future information. Taking a one-dimensional causal convolution kernel as an example, its length is L, and the convolution step is 1. By padding zeros on the left side of the convolution kernel to ensure that the output of the current time step only depends on the current and past time steps, the calculation process can be expressed as:

[0092]

[0093] where x' t is the value of the output sequence at time step t, Weight i is the weight of the convolution kernel at position i, and x' is the output of the causal convolution layer.

[0094] 4) Dilated convolution layer: Dilated convolution is used to increase the receptive field and capture long-distance dependencies. The dilation rate starts from 1 and increases layer by layer. Taking a one-dimensional dilated convolution kernel as an example, its length is L and the convolution step is 1. Dilated convolution is achieved by inserting zeros between the convolution kernels, thereby expanding the receptive field without increasing the computational complexity. The calculation process is as follows:

[0095]

[0096] where x' t-d*i It is the value of the input sequence at time step td*i, d represents the dilation rate, that is, the interval between the convolution kernels, and the output x'' after residual addition to the input sample x is used as the output of the dilated convolution layer.

[0097] 5) Multi-head self-attention mechanism: The multi-head self-attention mechanism is used for the dependencies between different positions in the time series. It captures richer feature representations by computing multiple attention heads in parallel, with each head focusing on a different subspace of the input sequence. Specifically, the multi-head self-attention mechanism maps the input sequence to multiple low-dimensional spaces, then calculates the attention weights in each low-dimensional space, and finally linearly combines the outputs of multiple attention heads to obtain the final output representation.

[0098] Based on the output x'' of the dilated convolutional layer, the single attention mechanism is calculated as follows:

[0099] Use linear transformation to map the input x'' to query, key, and value matrices respectively. The calculation formula is:

[0100] Q=W Q *x″

[0101] K=W K *x″

[0102] V=W V *x″

[0103] Where W Q , W K , W V is a learnable weight matrix with the same size and height denoted by d.

[0104] Next, calculate the attention matrix A, the calculation formula is as follows:

[0105]

[0106] O=A*V

[0107] The value matrix V is weighted and summed using the attention weight matrix A to obtain the output O of this attention mechanism calculation. The multi-head attention mechanism repeats the above calculations multiple times to capture multiple long-distance dependencies between sequences, then concatenates all the outputs and obtains the final output through a linear transformation:

[0108]

[0109] where λ i is a learnable weight matrix, indicating that the i-th attention head outputs the result x in this layer tmp The weight size.

[0110] 6) Multi-head attention residual connection: Add residual connection after the multi-head attention layer to speed up training and improve the stability of the model. The residual connection makes it easier for the model to learn the identity mapping by adding the input directly to the output, thereby alleviating the gradient vanishing problem. The calculation formula is as follows:

[0111] x'''=x tmp +x''

[0112] 7) Activation function: All activation functions in the model are ReLU activation functions. The ReLU function is simple and computationally efficient, and can effectively avoid the vanishing gradient problem.

[0113] 8) Output layer: The output layer uses a fully connected layer to map the convolution result to the category space. The fully connected layer maps the high-dimensional feature x''' to the low-dimensional category space through linear transformation to obtain the output result In order to achieve the classification task, the calculation process is as follows:

[0114]

[0115] Where W fc is the weight matrix of the learnable fully connected layer.

[0116] The total number of training set samples is Num, and one of the fault status categories is y i The input sample x i , the result predicted by the above model is The result of the sample after data enhancement by the jth random convolution kernel is predicted by the model: In order to ensure that the model introduces local noise and deformation while maintaining the original sample fault characteristics, this embodiment designs the following loss function:

[0117] 1) Mission loss L task : It is used to measure the difference between the model prediction result and the true label. The specific task loss used is the mean square error loss, which is expressed as:

[0118]

[0119] 2) Consistency loss L cons : It is used to measure the consistency between the prediction results of different enhanced samples. The consistency loss is calculated by KL divergence and is expressed as:

[0120]

[0121] in Represents the fault sample x i After C times of data enhancement, the mean of the probability distribution predicted by the model is obtained. λ is the regularization term, which indicates the influence of consistency loss. The total task loss is the task loss L task and consistency loss L cons Composition, expressed as:

[0122] L=L task +L cons

[0123] During the training process, the following strategies are used for model optimization:

[0124] 1) Use batch training method to randomly select a batch of samples from the training set for training each time to improve training efficiency and model stability.

[0125] 2) Use Adam learning rate adjustment strategy and adaptive learning rate method to improve the convergence speed and stability of the model.

[0126] 3) Use regularization methods, such as L2 regularization, to prevent model overfitting and improve the generalization ability of the model.

[0127] In this embodiment, when training the drone fault diagnosis model based on random convolution data enhancement and deep learning, the relevant parameters are selected as follows:

[0128] In the data preprocessing stage, the sliding window size determines the number of time steps for each sample. The window size M is set to 128 and the sliding window step size S is set to 16 to ensure that the fault characteristics can be captured while avoiding excessive computational burden and reducing the overlap between adjacent windows. A smaller step size can increase the number of samples.

[0129] In the random convolution operation, in order to keep the input and output sizes consistent, it is necessary to fill the boundaries of the input data with zeros, and the number of filling layers is 2. The random convolution parameters include the convolution kernel size K of 3, and each element of the convolution kernel is randomly selected from the distribution, where the mean μ of the distribution is 0 and the standard deviation σ is 1. The void rate d determines the interval between the convolution kernels, which increases layer by layer to expand the receptive field, and the layer-by-layer size is defined as d=1,2,4,8.

[0130] In the model training phase, the batch size (B) is 32, the learning rate (lr) is 0.0001, and the optimizer (Adam) is used. In addition, the L2 regularization parameter (λ) is introduced to prevent the model from overfitting.

[0131] The present invention adopts a method based on random convolution kernel transformation data enhancement and combines it with a temporal convolutional network for model training. It introduces local noise and deformation while maintaining the original sample fault characteristics, thereby improving the generalization ability and robustness of the model. Through the above-mentioned training method, the present invention can improve the performance of fault diagnosis when the fault data set is scarce, providing new ideas and methods for data-driven UAV fault detection methods.

[0132] In summary, the present invention takes into account that random convolution kernel transformation can effectively capture the local and global characteristics of input data, and in time series data, the random convolution transformation still conforms to similar time series statistical characteristics before and after. By introducing random convolution data enhancement technology, random convolution operations are performed on the original fault samples to generate a variety of enhanced time series samples that retain the original fault characteristics but introduce a variety of noise and deformation in local details, and weighted fusion with the original samples, thereby effectively expanding the fault sample library and improving the balance of the data set; in addition, in order to efficiently capture the time series characteristics of drone fault signals, a time convolution network model is adopted, aiming to accurately identify complex fault modes, providing new ideas and methods for data-driven drone fault detection methods, and ensuring the reliability and accuracy of the diagnosis results. The present invention has broad application prospects in fault diagnosis tasks, especially in improving the generalization ability and robustness of the model.

[0133] Based on the same inventive concept, an embodiment of the present invention discloses a drone fault diagnosis system integrating random convolution data enhancement and deep learning, including the following modules:

[0134] The UAV flight data collection module is used to collect the flight data of the UAV under normal flight conditions and fault conditions;

[0135] The sliding window segmentation module is used to perform sliding window segmentation on the data on both sides of the boundary point, taking the time when the fault occurs as the boundary point, to obtain window data of equal length under different flight states, and each window corresponds to a sample data;

[0136] The random convolution kernel generation module is used to randomly generate a set of convolution kernels with random parameters that conform to the normal distribution, and dynamically adjust the convolution width, height, zero padding and step size so that the convolution input and output have the same size;

[0137] The weighted fusion module is used to perform weighted fusion on the sample data and its random convolution results to maximize the preservation of the fault characteristics of the original sample while introducing local noise and deformation according to the weights;

[0138] The fault diagnosis model training module is used to use the original sample data, random convolution enhanced sample data and labels to train the fault diagnosis model based on the temporal convolution network to realize UAV fault diagnosis.

[0139] An embodiment of the present invention also discloses a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the drone fault diagnosis method that integrates random convolution data enhancement and deep learning.

[0140] The above are preferred embodiments of the present invention. Those skilled in the art should know that the present invention is not limited to the above embodiments. Without departing from the spirit and scope of the present invention, the present invention will also have various variations and improvements, and such variations and improvements all fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A UAV fault diagnosis method integrating random convolution data enhancement and deep learning, characterized in that: The steps include: (1) Collect flight data of the UAV under normal flight conditions and fault conditions; (2) Taking the time when the fault occurs as the boundary point, the data on both sides of the boundary point are divided into sliding windows to obtain window data of equal length under different flight conditions. Each window corresponds to a sample data. (3) Randomly generate a set of convolution kernels with random parameters that conform to the normal distribution, and dynamically adjust the convolution width, height, zero padding, and step size so that the convolution input and output have the same size; (4) Perform weighted fusion on the sample data and its random convolution results to maximize the retention of the fault characteristics of the original sample while introducing local noise and deformation according to the weights; (5) The original sample data, random convolution enhanced sample data and label training are used to develop a fault diagnosis model based on temporal convolutional network to realize UAV fault diagnosis.

2. The UAV fault diagnosis method integrating random convolution data enhancement and deep learning according to claim 1 is characterized in that: Step (1) includes: Various sensors installed on the drone collect data during flight, including attitude, position, speed and altitude; A fault signal is artificially introduced at the ground station, and the flight data of the UAV under artificially simulated fault conditions is collected, covering all flight parameter trajectories in the time interval from take-off to landing.

3. The UAV fault diagnosis method integrating random convolution data enhancement and deep learning according to claim 1 is characterized in that: In step (2), a complete flight data X∈R N×T Sliding window segmentation is performed, where N is the number of flight parameters, T is the time length, and the time when the fault occurs is t f The data on both sides of the boundary point are divided using sliding windows of length M, with a step size of S. The i-th window corresponds to the i-th sample data, and finally the window data of equal length under different flight states are obtained. i ∈R N×M ; During the sliding process of the window, for the state before the fault, the sliding window starts from t = 0 and ends at t = t f -1 ends, the window index range is W i =X[:,i·S:i·S+M], where the value range of i is For the post-fault state, the sliding window starts from t = t f Starting from t = T-1, the window index range is W i =X[:,t f +i·S:t f +i·S+M], where i ranges from 4. The UAV fault diagnosis method integrating random convolution data enhancement and deep learning according to claim 3 is characterized in that: The label y of the i-th window sample i Depends on the last time point t in its window end With the boundary point t f The relationship between them is expressed as: 0 indicates a normal flight, and Tag indicates the type of injected fault in the current flight.

5. The UAV fault diagnosis method integrating random convolution data enhancement and deep learning according to claim 1 is characterized in that: In step (3), for the convolution kernel with width w and height h respectively, its parameters are randomly selected from the normal distribution, expressed as: Where Kernel[i][j] represents the weight of the convolution kernel at position (i, j). and σ 2 Represents the mean and variance of the normal distribution; in the fault sample Window i ∈R N×M The zero padding Pad and stride during convolution are constrained as follows: Among them, Pad w and Pad h Indicates the number of zero padding layers in the width and height dimensions of the input sample.

6. The UAV fault diagnosis method integrating random convolution data enhancement and deep learning according to claim 5 is characterized in that: Step (3) includes: Generate a k×k random convolution kernel Kernel. Each element of the convolution kernel is randomly selected from a normal distribution with a mean of 0 and a standard deviation of 1, expressed as: The calculation process of random convolution is as follows: for the input sample x with width W and height H, the border is zero-filled, that is, padding is done in the width W dimension w Layer zero, padding in height H dimension h Layer zero, get the padded sample data x p , and then use the generated random convolution kernel Kernel to the padded data x p Perform convolution calculation with a step size of s to obtain the output y. The calculation process can be expressed as: The value range of a and b is: In order to ensure that the width W' and height H' after convolution are consistent with the width W and height H of the input sample, that is, W' = W, H' = H, the following conditions need to be met:

7. The UAV fault diagnosis method integrating random convolution data enhancement and deep learning according to claim 1 is characterized in that: In step (4), for the input sample x and the result of random convolution x rc ,The fusion method is expressed as: x new =α*x+(1-α)*x rc Where α is the fusion weight, α∈{0.1, 0.3, 0.5, 0.7, 0.9}.

8. The UAV fault diagnosis method integrating random convolution data enhancement and deep learning according to claim 1 is characterized in that: The loss function of the fault diagnosis model training is designed as follows: Let the total number of training set samples be Num, and for one of the fault status categories be y i The input sample x i , the result predicted by the model is The result of the sample after data enhancement by the jth random convolution kernel is predicted by the model: The total task loss is given by the task loss L task and consistency loss L cons Composition, where the task loss L task Used to measure the difference between the model prediction results and the true label, consistency loss L cons It is used to measure the consistency between the prediction results of different enhanced samples. The consistency loss is calculated by KL divergence and is expressed as in Represents the fault sample x i The mean of the probability distribution predicted by the model after C times of data augmentation, λ represents the impact of consistency loss.

9. A UAV fault diagnosis system integrating random convolution data enhancement and deep learning, characterized in that: Includes the following modules: The UAV flight data collection module is used to collect the flight data of the UAV under normal flight conditions and fault conditions; The sliding window segmentation module is used to perform sliding window segmentation on the data on both sides of the boundary point, taking the time when the fault occurs as the boundary point, to obtain window data of equal length under different flight states, and each window corresponds to a sample data; The random convolution kernel generation module is used to randomly generate a set of convolution kernels with random parameters that conform to the normal distribution, and dynamically adjust the convolution width, height, zero padding and step size so that the convolution input and output have the same size; The weighted fusion module is used to perform weighted fusion on the sample data and its random convolution results to maximize the preservation of the fault characteristics of the original sample while introducing local noise and deformation according to the weights; The fault diagnosis model training module is used to use the original sample data, random convolution enhanced sample data and labels to train the fault diagnosis model based on the temporal convolution network to realize UAV fault diagnosis.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images