An unmanned aerial vehicle motor fault diagnosis method based on a hollow convolution residual

Abstract

This invention relates to a method for diagnosing UAV motor faults based on dilated convolution residuals. The method comprises a data preprocessing module that converts one-dimensional vibration signals collected during UAV motor operation into two-dimensional grayscale images; a feature extraction module that extracts features from the input grayscale images to generate multi-channel feature maps; and a fault classification module that compresses the feature dimensions of the multi-channel feature maps, performs dimensionality mapping, and outputs the probabilities of various fault types, which constitute the fault diagnosis results. This invention improves fault identification accuracy by expanding the receptive field through dilated convolution and focusing on key features using a CBAM attention module, effectively capturing early, weak fault signals in the motor. The grayscale image conversion method in the data preprocessing stage does not require manual preset feature parameters, and the model can adaptively learn motor fault features under different operating conditions (such as different speeds and loads), with a generalization error of less than 5%.

Description

Technical Field

[0001] This invention relates to the field of UAV fault diagnosis technology, specifically a UAV motor fault diagnosis method based on vacuous convolution residuals. This method enables real-time monitoring of the UAV motor's operating status and accurate fault identification, and is particularly suitable for diagnosing early and minor faults in UAV motors under various operating conditions, providing technical support for the safe and stable operation of UAVs. Background Technology

[0002] As a type of aircraft capable of remote control or autonomous flight, drones are widely used in various fields such as aerial surveying, power line inspection, agricultural plant protection, and emergency rescue. The motor, as the core power component of a drone, directly determines its flight performance, endurance, and operational safety. A motor malfunction can lead to anything from operational interruption and mission failure to drone crashes, equipment damage, and even personal injury or death. Therefore, timely and accurate diagnosis of drone motor malfunctions is of significant practical importance.

[0003] Currently, UAV motor fault diagnosis mainly relies on traditional signal analysis methods and conventional machine learning algorithms. Traditional signal analysis methods use vibration and current signals as time-frequency analysis objects, extracting features in the time domain (e.g., peak value, root mean square, kurtosis), frequency domain (e.g., power spectrum peak value, frequency centroid), and time-frequency domain (e.g., wavelet transform coefficients) to achieve fault identification. Commonly used methods include short-time Fourier transform and Wigner-Ville distribution. However, these methods have significant limitations: on the one hand, feature extraction is highly dependent on human experience, requiring professionals to preset feature parameters based on the motor fault type, resulting in poor adaptability to motor diagnosis under multiple fault types and complex operating conditions; on the other hand, the time-frequency resolution of short-time Fourier transform is limited by the window function, and the Wigner-Ville distribution is prone to cross-term interference, leading to distortion in the extraction of early weak fault features, making it difficult to meet the needs of accurate diagnosis.

[0004] With the development of machine learning technology, algorithms such as Support Vector Machines (SVMs), Random Forests, and Backpropagation (BP) Neural Networks have been applied to UAV motor fault diagnosis, reducing manual intervention to some extent. However, these algorithms still have shortcomings: SVMs have weak generalization ability when processing high-dimensional feature data, Random Forests have limited fitting ability for nonlinear complex fault patterns, and traditional BP Neural Networks are prone to getting trapped in local optima and lack the ability to mine deep features. In recent years, deep learning technology has provided a new direction for motor fault diagnosis, among which Convolutional Neural Networks (CNNs) have received widespread attention due to their advantages in automatic feature extraction. However, traditional CNNs face many problems in application: First, increasing network depth can easily lead to gradient vanishing or gradient exploding, making model training convergence difficult; second, the receptive field of ordinary convolutional kernels is limited, making it difficult to capture global correlation information in motor fault signals, especially the weak feature capture ability of early faults; third, the computational load during model training is huge, requiring high computing power from UAV onboard hardware, making it difficult to meet the needs of real-time diagnosis.

[0005] To address the aforementioned technical challenges, there is an urgent need for a method for diagnosing UAV motor faults that can expand the receptive field, avoid gradient problems, reduce computational costs, and improve fault identification accuracy. This invention combines dilated convolution, residual learning, and channel-space attention (CBAM) to construct a diagnostic model based on CBAM dilated convolution residuals. This effectively overcomes the shortcomings of traditional methods and provides an efficient and accurate technical solution for UAV motor fault diagnosis. Summary of the Invention

[0006] The purpose of this invention is to provide a method for diagnosing motor faults in unmanned aerial vehicles (UAVs) based on dilated convolution residuals.

[0007] The technical solution adopted by this invention to achieve the above objectives is: a method for diagnosing UAV motor faults based on dilated convolution residuals, comprising the following steps:

[0008] The data preprocessing module converts the one-dimensional vibration signal collected during the operation of the drone motor into a two-dimensional grayscale image;

[0009] The feature extraction module extracts features from the input grayscale image and generates a multi-channel feature map.

[0010] The fault classification module compresses the feature dimensions of the multi-channel feature map and, after dimension mapping, outputs the probability of various faults, which is the fault diagnosis result.

[0011] The data preprocessing module is configured to perform the following steps:

[0012] 1) Signal segmentation: The one-dimensional vibration signal is divided into equal parts according to a fixed window to ensure that each signal segment can completely contain the fault characteristic cycle, thus obtaining several signal segments;

[0013] 2) Feature extraction: Extract time-domain features from each signal segment to form a feature matrix;

[0014] 3) Normalization and image conversion: The eigenvalues ​​in the feature matrix are mapped to the [0,255] interval through linear normalization, which are used as pixel values ​​of the two-dimensional grayscale image to generate the grayscale image, thus completing the dimensional conversion from one-dimensional signal to two-dimensional image.

[0015] The feature extraction module is configured to perform the following steps:

[0016] Initial convolutional layer: A 3×3 convolutional kernel with a stride of 1 is used to perform preliminary feature extraction on the input grayscale image and output a 64-channel feature map.

[0017] Dilated Convolutional Residual Blocks: Multiple residual blocks are set, each containing two dilated convolutional layers with dilation rates of 1 and 2, respectively, to expand the receptive field of the convolutional kernels and capture global correlation features in motor fault signals. Simultaneously, the residual blocks introduce skip connections, directly superimposing the input features onto the output of the dilated convolutional layers, as expressed in the formula: ,in x Input for residual blocks, W For convolution kernel parameters, F(x,W) This is the result of dilated convolution processing. y Output for residual blocks;

[0018] CBAM Attention Module: Through the channel attention submodule and the spatial attention submodule, the weights of the channel and spatial dimensions of the feature map are adjusted respectively to highlight key fault features.

[0019] The CBAM attention module is configured to perform the following steps:

[0020] Channel attention submodule: Performs global average pooling and global max pooling on the feature map to obtain two 1×1×C vectors, where C is the number of channels in the feature map; after fusing through two fully connected layers, a channel weight vector is generated by the Sigmoid activation function to weight each channel of the feature map to highlight fault-related channel features;

[0021] Spatial attention submodule: Performs global average pooling and global max pooling on the channel-weighted feature map to obtain two H×W×1 matrices, where H and W are the height and width of the feature map, respectively. After concatenation, the matrices are compressed to one channel through a 3×3 convolutional layer. The spatial weight matrix is ​​generated by the Sigmoid activation function to weight the spatial position of the feature map in order to focus on the fault feature region.

[0022] The fault classification module is configured as follows:

[0023] Global average pooling layer: converts the multi-channel feature map output by the feature extraction module into a 1×1×C feature vector;

[0024] Fully connected layer: Two fully connected layers are set up, in which the number of neurons in the second layer is consistent with the number of motor fault categories, so as to realize the dimensionality mapping of the feature vector;

[0025] Softmax layer: Converts the output of the fully connected layer into a probability distribution of various faults, and the category with the highest probability is the fault diagnosis result.

[0026] The fault diagnosis results include: normal state, rotor imbalance fault, bearing wear fault, and stator winding short circuit fault.

[0027] A fault diagnosis system for UAV motors based on dilated convolution residuals, comprising:

[0028] The data preprocessing module is used to convert the one-dimensional vibration signal collected during the operation of the UAV motor into a two-dimensional grayscale image;

[0029] The feature extraction module is used to extract features from the input grayscale image and generate a multi-channel feature map;

[0030] The fault classification module is used to compress the feature dimensions of the multi-channel feature map and output the probability of various faults after dimension mapping, which is the fault diagnosis result.

[0031] The feature extraction module includes:

[0032] Initial convolutional layer: Uses a 3×3 kernel with a stride of 1 to perform preliminary feature extraction on the input grayscale image and output a 64-channel feature map;

[0033] Dilated Convolutional Residual Blocks: Multiple residual blocks are set, each containing two dilated convolutional layers with dilation rates of 1 and 2, respectively. These are used to expand the receptive field of the convolutional kernels and capture globally correlated features in motor fault signals. Simultaneously, the residual blocks introduce skip connections, directly superimposing the input features onto the output of the dilated convolutional layers. The formula is expressed as follows: ,in x Input for residual blocks, W For convolution kernel parameters, F(x,W) This is the result of dilated convolution processing. y Output for residual blocks;

[0034] CBAM Attention Module: Used to adjust the weights of the channel and spatial dimensions of the feature map through the channel attention submodule and the spatial attention submodule, respectively, to highlight key fault features.

[0035] A UAV motor fault diagnosis device based on dilated convolution residuals includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement the UAV motor fault diagnosis method based on dilated convolution residuals when the computer program is executed.

[0036] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned method for diagnosing unmanned aerial vehicle (UAV) motor faults based on dilated convolution residuals.

[0037] The present invention has the following beneficial effects and advantages:

[0038] 1. Improve fault identification accuracy: By expanding the receptive field through dilated convolution and focusing on key features by combining the CBAM attention module, the early weak fault signals of the motor can be effectively captured. Experimental verification shows that the identification accuracy of 4 typical faults can reach more than 98.5%, which is 10-15 percentage points higher than that of traditional CNN.

[0039] 2. Avoid gradient problems: The skip connection design of residual blocks solves the gradient vanishing and gradient exploding problems in deep network training, enabling the model to converge stably even with a depth of more than 18 layers.

[0040] 3. Reduced computational cost: Dilated convolution expands the receptive field without increasing the number of convolution kernel parameters, reducing computational cost by 30%-40% compared to traditional CNNs. It can be adapted to UAV-borne embedded hardware (such as NVIDIA Jetson Nano) to meet real-time diagnostic needs (single sample diagnosis time ≤ 0.1s).

[0041] 4. Enhanced generalization ability: The grayscale image conversion method in the data preprocessing stage does not require manual preset feature parameters. The model can adaptively learn motor fault characteristics under different operating conditions (such as different speeds and loads), with a generalization error of less than 5%.

[0042] 5. High practicality: The diagnostic process is highly automated, requiring no professional intervention. It can be integrated into the UAV flight control system to achieve real-time fault monitoring and alarm, reducing UAV operation and maintenance costs and safety risks. Attached Figure Description

[0043] Figure 1 The residual block structure of this invention;

[0044] Figure 2 CBAM structure diagram of this invention;

[0045] Figure 3 Overall structural diagram of the model of this invention;

[0046] Figure 4Example diagram of the signal to two-dimensional grayscale image of the present invention;

[0047] Figure 5 A comparison diagram of ordinary convolution and dilated convolution in this invention. Detailed Implementation

[0048] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0049] The technical solution adopted in this invention is as follows:

[0050] (1) Construction of UAV motor fault diagnosis model based on CBAM dilated convolution residual

[0051] The diagnostic model constructed in this invention mainly consists of three parts: a data preprocessing module, a feature extraction module, and a fault classification module. The specific structure is as follows:

[0052] (1-1) Data Preprocessing Module: This module converts the one-dimensional vibration signals collected during the operation of the UAV motor into two-dimensional grayscale images, which are then used as model input. The specific steps are as follows:

[0053] 1) Signal segmentation: A fixed window is used, and the window size can be adjusted according to the motor speed (e.g., 256×256). The one-dimensional vibration signal is averaged to ensure that each signal segment contains the complete fault characteristic cycle.

[0054] 2) Feature extraction: Extract time-domain features (such as peak value, root mean square, kurtosis, and skewness) for each signal segment, extracting a total of 8-12 feature parameters to form a feature matrix;

[0055] 3) Normalization and image transformation: The eigenvalues ​​in the feature matrix are mapped to the [0,255] interval through linear normalization, which are used as the pixel values ​​of the two-dimensional grayscale image to generate a grayscale image with a size of 256×256. This completes the dimensional transformation from one-dimensional signal to two-dimensional image, which is convenient for the convolutional layer to extract spatial correlation features.

[0056] (1-2) Feature extraction module: This module is the core of the model and consists of "initial convolutional layer + dilated convolutional residual block + CBAM attention module":

[0057] 1) Initial convolutional layer: Uses 3×3 convolutional kernels (64 kernels) with a stride of 1 to perform preliminary feature extraction on the input grayscale image and output a 64-channel feature map. According to the convolution operation rules, an N-channel image requires an N-channel convolutional kernel for convolution operation, and the result will only output a single-channel feature map. Therefore, if you want the output feature map to have 64 channels, you need 64 convolutional kernels.

[0058] 2) Dentulated convolution residual block: The residual block structure is as follows Figure 1As shown. Each residual block contains two dilated convolutional layers with dilation rates of 1 and 2, respectively. By expanding the receptive field of the convolutional kernel (the receptive field can reach 7×7), it captures the global correlation features in the motor fault signal. Simultaneously, the residual block introduces skip connections, directly superimposing the input features onto the output of the dilated convolutional layer, effectively avoiding the gradient vanishing problem. The formula is expressed as: ,in x For a hollow convolution residual block, W For convolution kernel parameters, F(x,W) This is the result of dilated convolution processing. y Outputting residual blocks; A comparison of ordinary convolution and dilated convolution is shown below. Figure 5 As shown.

[0059] 3) CBAM Attention Module: Embedded after the dilated convolution residual block, it is divided into a channel attention submodule (CA) and a spatial attention submodule (SA):

[0060] i) Channel attention submodule: such as Figure 2 As shown, global average pooling and global max pooling are performed on the feature map to compress the spatial dimension of the input feature map, generating two different spatial context descriptions to obtain two 1×1×C vectors (C is the number of channels in the feature map). These vectors are then fused through two fully connected layers (the number of neurons in the hidden layer is C / 8). After fusion, a channel weight vector is generated by the Sigmoid activation function to weight each channel of the feature map, highlighting the fault-related channel features. In the first fully connected layer, the features are reduced in dimensionality and mapped to the hidden layer (the purpose is to reduce the number of parameters and extract the nonlinear relationship between channels). The second fully connected layer then maps the features back to the initial dimension.

[0061] ii) Spatial attention submodule: such as Figure 2 As shown, the channel-weighted feature maps are subjected to global average pooling and global max pooling respectively, resulting in two H×W×1 matrices (H and W are the height and width of the feature maps). These matrices are concatenated and compressed into one channel using a 3×3 convolutional layer (stride 1). A spatial weight matrix is ​​then generated using the Sigmoid activation function to weight the spatial location of the feature map, focusing on fault feature regions. Simultaneously, global pooling and max pooling are performed, and the max-pooled and average-pooled features are concatenated along the channel dimension to obtain a feature map with contextual information at different scales. This feature map is then processed by a convolutional layer to generate spatial attention weights.

[0062] The feature extraction module introduces an organic combination of dilated convolution and attention mechanism. Dilated convolution expands the receptive field to capture more features while reducing the amount of computation. CBAM reweights each channel of the feature map, allowing the model's vision to focus on fault information.

[0063] (1-3) Fault classification module: Consists of a global average pooling layer, a fully connected layer, and a Softmax layer.

[0064] 1) Global average pooling layer: Converts the multi-channel feature map output by the feature extraction module into a 1×1×C feature vector, reducing the number of parameters and avoiding overfitting;

[0065] 2) Fully connected layers: Two fully connected layers are set up. The first layer has 256 neurons, and the second layer has the same number of neurons as the number of motor fault categories (e.g., 4 fault categories correspond to 4 neurons) to realize the dimensionality mapping of feature vectors;

[0066] 3) Softmax layer: Converts the output of the fully connected layer into a probability distribution of various faults, and the category with the highest probability is the fault diagnosis result.

[0067] (2) Model training and fault diagnosis process

[0068] (2-1) Dataset construction: Vibration signals were collected under four typical conditions of UAV motor: normal state, rotor imbalance, bearing wear, and stator winding short circuit. 1000-2000 samples were collected for each condition and divided into training set, validation set and test set in a ratio of 7:2:1.

[0069] (2-2) Model training: The cross-entropy loss function is used as the loss calculation index. The Adam optimization algorithm (the initial learning rate is set to 0.001 and decays to 1 / 10 of the original value every 50 rounds) is used to optimize the model parameters. The training rounds are set to 200 rounds. Training is stopped when the validation set loss does not decrease for 10 consecutive rounds, and the optimal model is saved.

[0070] (2-3) Real-time fault diagnosis: The motor vibration signal collected in real time by the UAV on-board sensor is input into the preprocessing module, and after generating a grayscale image, it is input into the trained model. The model outputs the probability of a fault category. When the probability of a certain fault category is greater than the preset threshold (preferably 0.9), it is determined that the motor has a corresponding fault and an alarm signal is triggered.

[0071] like Figure 3 The diagram shown is the overall structure of the UAV motor fault diagnosis model based on CBAM dilated convolution residuals according to the present invention. The present invention provides a UAV motor fault diagnosis method based on dilated convolution residuals, characterized by the following steps:

[0072] Step (1): Acquisition and preprocessing of UAV motor vibration signals, outputting two-dimensional grayscale image samples. Motor vibration signals are acquired through onboard sensors and preprocessed into two-dimensional grayscale images recognizable by the model. The specific operations are as follows:

[0073] (1-1) Vibration signal acquisition: An accelerometer is installed on the top of the UAV motor housing to collect vibration signals in real time during motor operation. The acquisition range covers four typical motor states: normal state (S0), rotor imbalance fault (S1), bearing wear fault (S2), and stator winding short circuit fault (S3). Sufficient samples are collected for each state to meet the model training requirements.

[0074] (1-2) Signal segmentation: The one-dimensional vibration signal is divided into segments according to a fixed window to ensure that each segment can completely contain the fault characteristic cycle, thus obtaining several signal segments.

[0075] (1-3) Time-domain feature extraction: Extract time-domain features (such as peak value, root mean square, kurtosis, etc.) for each signal segment to form a feature matrix. The number of features should be sufficient to fully reflect the motor's operating status.

[0076] (1-4) Grayscale image conversion: The feature matrix elements are mapped to the [0,255] interval through linear normalization, serving as the pixel values ​​of the two-dimensional grayscale image. Zero-padding is applied to feature matrices with insufficient size, ultimately generating a grayscale image of fixed size, such as... Figure 4 As shown, the generated grayscale images are divided into training, validation, and test sets according to a standard ratio for subsequent model training and validation.

[0077] Step (2): CBAM dilated convolutional residual model construction and training, outputting the optimal diagnostic model. The diagnostic model is constructed and trained based on the preprocessed samples. The specific process is as follows:

[0078] (2-1) Model Structure Construction

[0079] The model consists of an initial convolutional layer, dilated convolutional residual blocks, a CBAM attention module, a global average pooling layer, a fully connected layer, and a softmax layer, as follows: Figure 3 As shown:

[0080] 1) Initial convolutional layer: A conventionally sized convolutional kernel is used to perform preliminary feature extraction on the input grayscale image, and a multi-channel feature map is output.

[0081] 2) Dilated Convolutional Residual Blocks: Multiple residual blocks are set up, each containing a dilated convolutional layer (as shown in Figure 5, dilated convolution can expand the receptive field and improve the global feature capture capability), and gradient problems are avoided during model training through residual connections;

[0082] 3) Attention module: such as Figure 2 As shown, it includes channel attention and spatial attention sub-modules, which adjust the weights of the channel and spatial dimensions of the feature map respectively to highlight key fault features;

[0083] 4) Classification layer: The feature dimension is compressed by a global average pooling layer, mapped by a fully connected layer, and the probability of each type of fault is output by a Softmax layer.

[0084] (2-2) Model Training

[0085] The model is built using mainstream deep learning frameworks, with appropriate loss functions and optimizers selected, and reasonable training epochs and batch sizes set. During training, the validation set loss is used as a reference; training stops when the validation set loss fails to decrease for several consecutive epochs, and the optimal model at that point is saved.

[0086] Step (3): Real-time fault diagnosis of UAV motors, outputting fault judgment results.

[0087] The trained model is deployed to the drone's onboard hardware to achieve real-time motor fault diagnosis. The specific steps are as follows:

[0088] (3-1) Airborne initialization

[0089] After the drone is powered on, the fault diagnosis module is activated to initialize the sensors, preprocessing program and optimal model. The sensors collect motor vibration signals at the set frequency.

[0090] (3-2) Real-time signal processing and inference

[0091] 1) Preprocessing: Convert the real-time acquired vibration signal into a fixed-size grayscale image according to the process of steps (1-2)-(1-4);

[0092] 2) Model inference: Input the grayscale image into the optimal model, and the model quickly outputs the probability of various faults to ensure that the total diagnostic time meets the real-time monitoring requirements of UAVs.

[0093] (3-3) Fault diagnosis and response

[0094] Set fault determination thresholds:

[0095] 1) If the probability of a certain type of fault exceeds the threshold, the motor is determined to have a corresponding fault and a fault code is generated;

[0096] 2) Develop response strategies for different fault types. Minor faults trigger alarm prompts, while serious faults automatically activate emergency procedures (such as controlling the drone to return to home) to ensure the flight safety of the drone.

[0097] Step (4): Model performance verification to ensure diagnostic accuracy and generalization ability.

[0098] The model's performance was validated using a test set and variable operating condition samples to ensure it meets the requirements of practical applications.

[0099] 1) Accuracy verification: Use test set samples to evaluate the model's accuracy in identifying various types of faults, ensuring that the average accuracy reaches a high level;

[0100] 2) Generalization verification: Adjust the motor operating conditions (such as changing the speed and load), collect new samples to test the model, and verify the model's adaptability under different operating conditions;

[0101] 3) Real-time verification: Test the diagnostic speed of the model on the airborne hardware to ensure that it can meet the frequency requirements of real-time monitoring of UAVs.

[0102] This invention, through the organic combination of the above steps, forms a complete fault diagnosis solution for UAV motors. By leveraging the synergistic effect of dilated convolution, residual learning, and CBAM attention mechanism, it effectively solves the problems of low accuracy and poor real-time performance of traditional diagnostic methods. It has the advantages of clear principles, low cost, and strong adaptability, and can be widely applied to the health monitoring of various UAV motors.

[0103] Those skilled in the art will understand that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for unmanned aerial vehicle motor fault diagnosis based on a cavity convolution residual, characterized in that, Includes the following steps: The data preprocessing module converts the one-dimensional vibration signal collected during the operation of the drone motor into a two-dimensional grayscale image; The feature extraction module extracts features from the input grayscale image and generates a multi-channel feature map. The fault classification module compresses the feature dimensions of the multi-channel feature map and, after dimension mapping, outputs the probability of various faults, which is the fault diagnosis result. 2.The UAV motor fault diagnosis method based on a hollow convolution residual error according to claim 1, characterized in that, The data preprocessing module is configured to perform the following steps: 1) Signal segmentation: The one-dimensional vibration signal is divided into equal parts according to a fixed window to ensure that each signal segment can completely contain the fault characteristic cycle, thus obtaining several signal segments; 2) Feature extraction: Extract time-domain features from each signal segment to form a feature matrix; 3) Normalization and image conversion: The eigenvalues ​​in the feature matrix are mapped to the [0,255] interval through linear normalization, which are used as pixel values ​​of the two-dimensional grayscale image to generate the grayscale image, thus completing the dimensional conversion from one-dimensional signal to two-dimensional image. 3.The UAV motor fault diagnosis method based on a hollow convolution residual error according to claim 1, characterized in that, The feature extraction module is configured to perform the following steps: Initial convolutional layer: A 3×3 convolutional kernel with a stride of 1 is used to perform preliminary feature extraction on the input grayscale image and output a 64-channel feature map. Dilated convolutional residual blocks: Multiple residual blocks are set, each containing two dilated convolutional layers with dilation rates of 1 and 2, respectively, to capture global correlation features in motor fault signals by expanding the receptive field of the convolutional kernel; At the same time, the residual block introduces a skip connection, which directly superimposes the input features on the output end of the dilated convolution layer, and the formula is expressed as: Wherein x is the input of the residual block, W is the convolution kernel parameter, F(x,W) is the dilated convolution processing result, y is the output of the residual block; CBAM Attention Module: Through the channel attention submodule and the spatial attention submodule, the weights of the channel and spatial dimensions of the feature map are adjusted respectively to highlight key fault features.

4. The UAV motor fault diagnosis method based on a hollow convolution residual according to claim 3, characterized in that, The CBAM attention module is configured to perform the following steps: Channel attention submodule: Performs global average pooling and global max pooling on the feature map to obtain two 1×1×C vectors, where C is the number of channels in the feature map; after fusing through two fully connected layers, a channel weight vector is generated by the Sigmoid activation function to weight each channel of the feature map to highlight fault-related channel features; Spatial attention submodule: Performs global average pooling and global max pooling on the channel-weighted feature map to obtain two H×W×1 matrices, where H and W are the height and width of the feature map, respectively. After concatenation, the matrices are compressed to one channel through a 3×3 convolutional layer. The spatial weight matrix is ​​generated by the Sigmoid activation function to weight the spatial position of the feature map in order to focus on the fault feature region.

5. The method for diagnosing UAV motor faults based on dilated convolution residuals according to claim 1, characterized in that, The fault classification module is configured as follows: Global average pooling layer: converts the multi-channel feature map output by the feature extraction module into a 1×1×C feature vector; Fully connected layer: Two fully connected layers are set up, in which the number of neurons in the second layer is consistent with the number of motor fault categories, so as to realize the dimensionality mapping of the feature vector; Softmax layer: Converts the output of the fully connected layer into a probability distribution of various faults, and the category with the highest probability is the fault diagnosis result.

6. The method for diagnosing UAV motor faults based on dilated convolution residuals according to claim 1, characterized in that, The fault diagnosis results include: normal state, rotor imbalance fault, bearing wear fault, and stator winding short circuit fault.

7. A fault diagnosis system for UAV motors based on dilated convolution residuals, characterized in that, include: The data preprocessing module is used to convert the one-dimensional vibration signal collected during the operation of the UAV motor into a two-dimensional grayscale image; The feature extraction module is used to extract features from the input grayscale image and generate a multi-channel feature map; The fault classification module is used to compress the feature dimensions of the multi-channel feature map and output the probability of various faults after dimension mapping, which is the fault diagnosis result.

8. The UAV motor fault diagnosis system based on dilated convolution residuals according to claim 7, characterized in that, The feature extraction module includes: Initial convolutional layer: Uses a 3×3 kernel with a stride of 1 to perform preliminary feature extraction on the input grayscale image and output a 64-channel feature map; Dilated Convolutional Residual Blocks: Multiple residual blocks are set, each containing two dilated convolutional layers with dilation rates of 1 and 2, respectively. These are used to expand the receptive field of the convolutional kernels and capture globally correlated features in motor fault signals. Simultaneously, the residual blocks introduce skip connections, directly superimposing the input features onto the output of the dilated convolutional layers. The formula is expressed as follows: ,in x Input for residual blocks, W For convolution kernel parameters, F(x,W) This is the result of dilated convolution processing. y Output for residual blocks; CBAM Attention Module: Used to adjust the weights of the channel and spatial dimensions of the feature map through the channel attention submodule and the spatial attention submodule, respectively, to highlight key fault features.

9. A UAV motor fault diagnosis device based on dilated convolution residuals, characterized in that, It includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement, when executing the computer program, a method for diagnosing UAV motor faults based on dilated convolution residuals as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements a method for diagnosing UAV motor faults based on dilated convolution residuals as described in any one of claims 1-6.

Images