Electromechanical composite transmission system health state identification method based on dynamic kernel width

Abstract

The invention discloses an electromechanical composite transmission system health state identification method based on dynamic kernel width, which belongs to the technical field of intelligent operation and maintenance of mechanical equipment and comprises the following steps: acquiring an electromechanical composite transmission system data set; building a health state identification network framework; parameters of each layer of the identification network framework are initialized, convolution kernel weight parameters of a sparse convolution layer are divided into a plurality of groups along an output channel, and each group corresponds to a search area; performing forward propagation and calculating loss; performing dynamic kernel wide layer convolution calculation optimization and sparse function training optimization; repeating forward propagation, loss calculation, sparse updating and gradient updating until the iteration task is finished, and obtaining a corresponding electromechanical composite transmission system health state identification model; and health state identification of the electromechanical composite transmission system is realized based on the model. According to the method, convolution calculation of a dynamic kernel width layer is guided by using a sparse function, the training complexity is reduced, and the balance of a large receptive field and low calculation complexity is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of mechanical equipment, and more specifically to a method for identifying the health status of an electromechanical composite transmission system based on dynamic kernel width. Background Art

[0002] The electromechanical transmission system is a key subsystem of the vehicle, playing an important role in transmitting traction to drive the vehicle forward. Due to the complex and changeable operating environment of the vehicle, various faults will inevitably occur in the electromechanical transmission system, posing safety risks to the normal service of the vehicle. Therefore, it is necessary to adopt an effective health status identification method to detect faulty components in the early stage of transmission system abnormality and repair them in time to ensure safe and reliable operation of the vehicle. In recent years, health status identification methods based on artificial intelligence have gradually been applied to the field of intelligent operation and maintenance. They use deep learning network models to analyze the signals collected by sensors, get rid of the dependence on expert experience, and achieve high-accuracy health status identification, which is highly respected.

[0003] The existing smart device health status identification system framework usually interacts with each sensor node through hardware such as gateways and data acquisition boards, and then transmits the data to high-computing power devices for health status identification. This method has disadvantages such as long transmission time and poor real-time identification. Edge computing devices can be deployed near sensor nodes to carry out data processing in a timely manner, thereby effectively improving the above problems and realizing real-time status identification. In the research on edge computing device networks, convolutional neural networks have become the preferred network model due to their efficient spatial weight sharing mechanism and relatively few trainable parameters. However, the electromechanical hybrid transmission system is a typical device with variable working conditions. The change in working speed causes the vibration signal to show different characteristics on the time scale. When performing health status identification, it is necessary to use multi-scale convolutional neural networks to fully extract features of different scales, but the multi-core wide channel causes the network model parameters and computational complexity to increase sharply, making it difficult to deploy on edge devices with insufficient computing and storage capabilities. Therefore, researchers often achieve model lightweighting by improving network structure, pruning, knowledge distillation, etc., but the existing lightweight multi-scale convolutional neural networks still have the following problems: manual analysis of data features and tedious debugging of convolution kernel width are required to effectively extract fault features; sparse training of network parameters layer by layer ignores the characteristics of local areas, and global optimization requires consideration of all weights of the entire convolution kernel, resulting in huge computational complexity. In summary, there is currently a lack of an effective health status identification method that adaptively selects convolution kernels to achieve good identification accuracy on the basis of low parameter count and computational complexity. Summary of the invention

[0004] In view of this, the present invention provides a method for identifying the health status of an electromechanical hybrid transmission system based on dynamic kernel width, which uses sparse functions to guide the convolution calculation of the dynamic kernel width layer, solves the problem that multi-scale convolutional neural networks require manual parameter adjustment, reduces the complexity of training, and achieves a balance between a large receptive field and low computational complexity.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] The present invention provides a method for identifying the health status of an electromechanical hybrid transmission system based on dynamic kernel width, comprising the following steps:

[0007] S1. Acquire a mechatronic transmission system data set; the data set includes multi-source sensor signal data and corresponding labels;

[0008] S2. Build a health status recognition network framework; the framework includes multiple dynamic kernel width layers, maximum pooling layers and fully connected layers, wherein the dynamic kernel width layer includes a sparse convolution layer and a 1x1 convolution layer;

[0009] S3, initializing the parameters of each layer of the recognition network framework, and dividing the convolution kernel weight parameters of the sparse convolution layer into multiple groups along the output channel, each group corresponding to a search area;

[0010] S4, forward propagation and loss calculation, including: taking the data set as the input of the recognition network framework, and calculating the loss according to the health status recognition result and the true label;

[0011] S5, dynamic kernel width layer convolution calculation optimization, including: guiding the activation of convolution kernel weights of sparse convolution layers through sparse functions, and dynamically adjusting the receptive field size;

[0012] S6, sparse function training optimization, including: in the forward propagation process, updating the sparse function by group search, and updating the number of weights according to the cosine annealing function decay in each Δt iteration;

[0013] S7, repeating steps S4 to S6 until the iterative task is completed, and obtaining a corresponding mechatronic hybrid transmission system health status identification model; and realizing health status identification of the mechatronic hybrid transmission system based on the model.

[0014] It can be seen from the above technical solution that compared with the prior art, the present invention has the following advantages:

[0015] 1) There is no need to manually adjust the width of the convolution kernel. The convolution is guided by the sparse function, which realizes the adaptive extraction of multi-scale features and effectively reduces the number of parameters and computational complexity.

[0016] 2) The training method of sparse functions is improved. By grouping the convolution kernel weights for search, the calculation scope of each pruning and growing is reduced, and the training cost is reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0018] Figure 1 A flow chart of the method for identifying the health status of an electromechanical hybrid transmission system based on dynamic kernel width provided by the present invention.

[0019] Figure 2 Schematic diagram of the health status identification network framework based on dynamic kernel width provided by the present invention.

[0020] Figure 3 This is a schematic diagram of the sparsification principle of the dynamic kernel width layer provided by the present invention.

[0021] Figure 4 This is a schematic diagram for comparing the performance of the method provided by the present invention with the existing method. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] Reference Figure 1 As shown, the embodiment of the present invention discloses a method for identifying the health status of an electromechanical hybrid transmission system based on dynamic kernel width, comprising the following steps:

[0024] S1. Obtain a mechatronic transmission system data set; the data set includes multi-source sensor signal data and corresponding labels. The sensor signal data includes vibration, current, speed and other signals of components such as motors, gear boxes, and axle boxes. The label indicates the health status of the transmission system, such as normal, faulty, etc.

[0025] Normal status: The transmission system operates normally without any faults.

[0026] Fault status: The transmission system has a fault and further identification of the fault type and severity is required.

[0027] Fault types: including the following:

[0028] Gearbox failure: such as gear wear, broken teeth, etc.

[0029] Axle box bearing failure: such as inner ring failure, outer ring failure, etc.

[0030] Motor failure: For example, motor bearing failure, short circuit, etc.

[0031] Other faults: such as loose connections, oil leakage, etc.

[0032] During the training process, the model needs to learn to associate the input multi-source sensor signal data with the corresponding health status labels, so as to learn to recognize different health states.

[0033] During the evaluation process, the model needs to predict the input multi-source sensor signal data into the corresponding health status label and compare it with the true label to evaluate the accuracy of the model.

[0034] S2. Build a health status recognition network framework; the framework includes multiple dynamic kernel width layers, maximum pooling layers and fully connected layers, wherein the dynamic kernel width layer includes a sparse convolution layer and a 1x1 convolution layer. The sparse convolution layer is used to extract multi-scale features, and the 1x1 convolution layer is used for dimensionality reduction and feature fusion.

[0035] S3. Initialize the parameters of each layer of the recognition network framework, and divide the convolution kernel weight parameters of the sparse convolution layer into multiple groups along the output channel, each group corresponds to a search area; the size of the search area can be pre-set or determined through experiments.

[0036] S4, forward propagation and loss calculation, including: using the data set as the input of the recognition network framework, and calculating the loss according to the health state recognition result and the true label. For example, using the training data set as the input of the network framework, forward propagation is performed, and the loss, such as cross entropy loss, is calculated according to the health state recognition result and the true label.

[0037] S5, dynamic kernel width layer convolution calculation optimization, including: guiding the activation of convolution kernel weights of sparse convolution layers through sparse functions, and dynamically adjusting the receptive field size;

[0038] In this step, the convolution kernel weight activation of the sparse convolution layer is guided by the sparse function, and only part of the weight is used for calculation. The receptive field size is dynamically adjusted to adapt to feature extraction at different scales.

[0039] S6, sparse function training optimization, including: in the forward propagation process, updating the sparse function by group search, and updating the number of weights according to the cosine annealing function decay in each Δt iteration;

[0040] In this step, during the forward propagation process, the sparse function is updated by group search; each group independently optimizes the search area to reduce the calculation range of each pruning and growth. The number of weights is updated according to the cosine annealing function decay to improve the search efficiency.

[0041] S7, repeating steps S4 to S6 until the iterative task is completed, and obtaining a corresponding mechatronic hybrid transmission system health status identification model; and realizing health status identification of the mechatronic hybrid transmission system based on the model.

[0042] This method uses a dynamic kernel width layer to adaptively capture multi-scale features from small receptive fields to large receptive fields, improving recognition accuracy. It also uses a convolution kernel weight grouping search training method to improve training efficiency through finer-grained optimization and reduce the computational complexity of each weight update. This method achieves the unity of lightweight model and high accuracy through dynamic kernel width layer and sparse function training optimization, and is suitable for the health status identification of electromechanical hybrid transmission systems with variable working conditions.

[0043] The following is a detailed description of each of the above steps:

[0044] The method for identifying the health status of a mechatronic transmission system based on dynamic kernel width includes the following steps:

[0045] 1. Obtain the mechatronic transmission system data set. Obtain the mechatronic transmission system fault data set , where any data can be expressed as , Represents the multi-source sensor signal data of the transmission system, l is the signal sampling length, c is the number of signal channels, is the corresponding label. And complete operations such as data standardization and data set division, such as dividing into training set, test set and validation set.

[0046] 2. Build a health status identification network framework. Figure 2 As shown, the health status identification network framework It is composed of multiple dynamic kernel width layers, maximum pooling layers and fully connected layers. x is the input data of network F; are the learnable network parameters of network F. At the same time, set the hyperparameters, including loss function, batch size, sparsity rate S, total number of training iterations T, and learning rate , cosine annealing function , sparse update frequency Δt, etc.

[0047] 3. Initialize the parameters of each layer of the identification network. Figure 3The dynamic kernel width layer sparsification principle diagram, the sparse function is an important function used to record the weight state of the convolution kernel, where the activated weights participate in the calculation. It should be noted in the initialization of each layer that the dynamic kernel width layer includes sparse convolution and 1×1 convolution, where the sparse convolution kernel weight parameters are divided into different groups to facilitate the subsequent training and learning of the sparse function. Without loss of generality, the convolution kernel weight of the lth layer Divided into N groups along the output channel, namely . is the number of output channels of layer l, is the number of output channels of the l-1th layer, that is, the number of input channels of the lth layer. Each group has a corresponding search area , for example, the exploration area in the i-th group of the l-th layer Defined as the front of each core in the group Position, only the search area participates in the update learning. When the dynamic kernel width layer is initialized, the weight of each group i in each layer l , according to the sparse rate S and the search area Randomly activate some weights and use a sparse function Record the activation of weights.

[0048] 4. Forward propagation and calculation of loss. The multi-source sensor signal data of the input data set is forward propagated, passing through the dynamic kernel width layer, the maximum pooling layer, and the fully connected layer in sequence, and the loss is calculated based on the health status recognition results and the true label. The specific calculation method is as follows:

[0049]

[0050] in, Represents the cross entropy loss function, and batch represents the training batch size. There are two special optimization parts in the process of updating parameters in the forward propagation and back propagation of the network model, including dynamic kernel width optimization and sparse function training process optimization.

[0051] 5 Dynamic kernel width layer specific convolution calculation optimization process. The dynamic kernel width layer is a sparse convolution part with a large and sparse convolution kernel, which can capture the adaptive data receptive field through learning, thereby dynamically selecting the most suitable kernel width. Specifically, given the input feature map of layer l, and convolution kernel weights , where h is the height of the feature map after the convolution operation, w is the width of the feature map after the convolution operation; k is the convolution kernel size, is the number of output channels of layer l, is the number of output channels of the l-1th layer, that is, the number of input channels of the lth layer. The sparse convolution operation formula with a step size of 1 and 0 padding in the proposed dynamic kernel width layer is as follows:

[0052]

[0053] in, represents the output feature representation of the j-th output channel, represents a set of integers, is a sparse function, express The activation weight is The core in the channel, represents the element-wise product, and Represents the convolution operator. is a sparse function learned during training that satisfies , where the sparseness rate , represents the L0 norm, ,when The convolution kernel is sparse. It is then input into the convolution layer with a kernel width of 1 and then activated by ReLU. The convolution output of the jth input channel is :

[0054]

[0055] O j represents the output feature map of the 1x1 convolution in the dynamic kernel width layer, and its dimension is the same as the input feature map; w1 represents the weight corresponding to the convolution, and b represents the bias term;

[0056] Therefore, k can be dynamically adjusted in the dynamic kernel width layer to obtain different receptive fields for efficient feature extraction and significantly reduce the computational cost.

[0057] 6 Sparse function training optimization process. The dynamic kernel width layer dynamically adjusts k to obtain different receptive fields through sparse function training optimization. In order to improve the search efficiency of the optimal activation weight of the sparse function and better cover the receptive fields of various scales, it is necessary to use a group search method to update the sparse function during the forward propagation process. As described in the network initialization section, the convolution kernel weight parameters of the dynamic kernel width layer are divided into different groups, and the weight states are selected in different groups ( ) are performed separately. Each group has a corresponding search area In the training process of sparse functions, the search area of ​​the weight is given first and the sparsity rate S, which is then updated every Δt iterations. The number of weights is updated over time according to the function The decay follows cosine annealing:

[0058]

[0059] in, is the ratio of the updated activation weights, i.e., the learning rate, t is the current training iteration number, and T is the total training iteration number. Without loss of generality, during the tth iteration, the number of updated activation weights in the i-th group of the l-th layer is ,in is the number of weights that can be searched in the region. The specific training process of the sparse function is as follows:

[0060] 6.1 Choose the right time to perform sparse updates:

[0061] If the current training ( ) is sparsely updated, where t is the current training iteration number, is the aforementioned sparse update frequency. Sparse update refers to adjusting the search area of ​​each group in each dynamic kernel width layer. The specific adjustment method is as follows:

[0062] 1) Calculate the number of updates.

[0063] Calculate the number of weights for sparse updates , the specific calculation method is as follows:

[0064]

[0065] in, is the number of weights that can be searched in the region.

[0066] 2) Weight pruning and growing.

[0067] According to ArgTOPK Find the num weight indices with the smallest absolute values, remove these weights from the activation state, and record them as ;

[0068] ArgTOPK is a function that selects the K largest values ​​from a set of values ​​and returns the indices corresponding to these values.

[0069] According to RandomK , randomly activate num weights that are not activated in the current exploration area, activate these weights, and record them as ; RandomK is a function that randomly selects K values ​​from a set of unactivated values ​​and returns the indices corresponding to these values.

[0070] 3) Update the sparse function and convolution kernel weights.

[0071] according to , the pruning and growing operations are recorded into the sparse function, where Refers to the sparse function of any dynamic kernel width layer;

[0072] according to , the convolution kernel weights activated by the sparse function are substantially updated.

[0073] 6.2 Choose the right time to update the gradient:

[0074] If the current training ( ) performs gradient update, calculates gradient according to the loss of forward propagation, and updates parameters through back propagation.

[0075]

[0076] in, refers to the network model parameters, is the learning rate, It refers to the loss function Find the gradient.

[0077] 7. Carry out the next iteration ( ), repeat the forward propagation, loss calculation, sparse update and gradient update process until the iterative task is completed ( ); obtain the corresponding mechatronic transmission system health status identification model; based on the model, realize the health status identification of the mechatronic transmission system.

[0078] For example:

[0079] Taking the fault diagnosis of the electromechanical transmission system of a train as an example, the effectiveness of the method of the present invention is verified by collecting simulation data through the test bench. The test bench used in this embodiment is a proportionally reduced model. Different speeds of the motor are used to simulate different operating speeds during driving, and hydraulic load equipment is used to apply loads to simulate different load-bearing conditions. The experimental collection includes the three-axis acceleration of the motor drive end, the motor non-drive end, the gearbox input end, the gear output end, the left and right axle boxes, the three-phase current and speed of the motor, and the sound intensity of the left and right axle boxes. The signal of 24 channels, with a sampling frequency of 64kHz.

[0080] The experiment considered 17 health states of the electromechanical transmission system, including normal state, 4 pinion support bearing faults, 4 pinion faults (crack, wear, broken teeth and missing teeth), 4 axle box bearing faults (inner ring fault, outer ring fault, rolling element fault and cage fault), and 4 motor faults (motor bearing fault, short circuit, rotor bending and rotor bar breakage). Each health state includes 2040 sample data with a length of 3200.

[0081] Reference Figure 4As shown, the experiment uses the method proposed in the present invention to identify the health status of the motor, gearbox and axle box, and compares three advanced multi-scale convolution methods: Res2Net-GAF, AkNets and MPNet. All four methods share the same model structure and parameter settings, the only difference is the multi-scale module. The experimental results summarize the performance evaluation results of the four multi-scale convolution methods on the experimental dataset, including the classification accuracy, number of parameters and average FLOPs of 17 health status category samples in 10 diagnostic experiments. The basic structure of the network is shown in Table 1, and the training related parameters are shown in Table 2. The experimental results are summarized in Tables 3 and Figure 4 The experimental results show that the health recognition effect of the proposed method is significantly better than that of the three comparison methods. The proposed dynamic kernel width layer has excellent performance, which significantly reduces the number of parameters and FLOPs while maintaining higher accuracy.

[0082] Table 1. Summary of basic network structure

[0083]

[0084] Table 2. Summary of training related parameters

[0085]

[0086] Table 3. Summary of model size and accuracy

[0087]

[0088] The method for identifying the health status of an electromechanical composite transmission system based on dynamic kernel width provided by the present invention solves the problem of lightweighting the network for identifying the health status of an electromechanical composite transmission system based on multi-scale convolution, does not require tedious manual adjustment of hyperparameters, and significantly reduces the complexity of training, ultimately achieving the unity of lightweight models and high accuracy. Specifically, a dynamic kernel width layer is constructed, and the activation of convolution kernel weights is guided by sparse functions, and only a portion of the weights are used for forward and backward propagation, which greatly reduces the amount of calculation and storage space occupied; during sparse function training, searches are performed in groups, and each group independently optimizes the search area, which reduces the calculation range of each pruning and growing, and reduces the training cost; weights are dynamically adjusted through pruning and growing operations to adapt to multi-scale feature learning, effectively improving the network's ability to identify the health status.

[0089] The method of the present invention can adaptively capture features within different scale ranges and realize local to global feature learning. It is suitable for the health status identification of mechatronic transmission systems with variable working conditions. Even under high sparsity rates, it can still maintain good feature extraction capabilities. It is of great significance to the promotion and application of health status identification networks of mechatronic transmission systems based on deep learning in edge computing scenarios.

[0090] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0091] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying the health status of a mechatronic transmission system based on dynamic kernel width, characterized in that: The following steps are involved: S1. Acquire a mechatronic transmission system data set; the data set includes multi-source sensor signal data and corresponding labels; S2. Build a health status recognition network framework; the framework includes multiple dynamic kernel width layers, maximum pooling layers and fully connected layers, wherein the dynamic kernel width layer includes a sparse convolution layer and a 1x1 convolution layer; S3, initializing the parameters of each layer of the recognition network framework, and dividing the convolution kernel weight parameters of the sparse convolution layer into multiple groups along the output channel, each group corresponding to a search area; S4, forward propagation and loss calculation, including: taking the data set as the input of the recognition network framework, and calculating the loss according to the health status recognition result and the true label; S5, dynamic kernel width layer convolution calculation optimization, including: guiding the activation of convolution kernel weights of sparse convolution layers through sparse functions, and dynamically adjusting the receptive field size; S6, sparse function training optimization, including: in the forward propagation process, updating the sparse function by group search, and updating the number of weights according to the cosine annealing function decay in each Δt iteration; S7, repeating steps S4 to S6 until the iterative task is completed, and obtaining a corresponding mechatronic hybrid transmission system health status identification model; and realizing health status identification of the mechatronic hybrid transmission system based on the model.

2. The method for identifying the health status of a mechatronic transmission system based on dynamic kernel width according to claim 1, characterized in that: The step S1 comprises: Obtaining a mechatronic transmission system fault dataset ; represents the multi-source sensor signal data of the electromechanical hybrid transmission system, l is the signal sampling length, c is the number of signal channels, is a corresponding label, including: normal state, fault state; the fault state includes: the fault type of a specific component; After data standardization, it is divided into training set, test set and validation set.

3. The method for identifying the health status of a mechatronic transmission system based on dynamic kernel width according to claim 1, characterized in that: The step S2 comprises: Building a health status identification network framework ; Where x is the input data of network F; is a learnable network parameter of the network F; the health status recognition network framework is composed of a plurality of dynamic kernel width layers, a maximum pooling layer and a fully connected layer stacked together, wherein the dynamic kernel width layer includes a sparse convolution layer and a 1x1 convolution layer; Set hyperparameters, including loss function, batch size, sparsity rate S, total number of training iterations T, and learning rate , cosine annealing function , sparse update frequency Δt.

4. The method for identifying the health status of a mechatronic transmission system based on dynamic kernel width according to claim 1, characterized in that: The step S3 comprises: Determine the number and configuration of dynamic kernel width layers, max pooling layers, and fully connected layers; In the dynamic kernel width layer, the convolution kernel weight parameters are divided into different groups along the output channel; a corresponding search area is set for each group; According to the sparsity rate S and the search area, a portion of the weights in each group are randomly activated, and the activation of the weights is recorded using the sparsity function Z.

5. The method for identifying the health status of a mechatronic transmission system based on dynamic kernel width according to claim 1, characterized in that: The step S4 comprises: The multi-source sensor signal data of the input data set is forward propagated through the dynamic kernel width layer, the maximum pooling layer, and the fully connected layer in sequence. The loss is calculated based on the health status recognition results and the true label. The specific calculation method is as follows: ； in, Represents the cross entropy loss function, batch represents the training batch size; there are two optimization parts in the forward propagation and back propagation parameter update process of the network model, including dynamic kernel width optimization and sparse function training process optimization.

6. The method for identifying the health status of a mechatronic transmission system based on dynamic kernel width according to claim 1, characterized in that: The step S5 comprises: Given the input feature map of layer l and convolution kernel weights ; where k is the convolution kernel size, is the number of output channels of the lth layer, is the number of output channels of the l-1th layer, that is, the number of input channels of the lth layer; h is the height of the feature map after the convolution operation, and w is the width of the feature map after the convolution operation; The formula for the sparse convolution operation with a step size of 1 and 0 padding in the proposed dynamic kernel width layer is as follows: ； in, represents the output feature representation of the j-th output channel, represents a set of integers, is a sparse function, express The activation weight is The core in the channel, represents the element-wise product, and represents the convolution operator; is a sparse function learned during training that satisfies , where the sparseness rate , represents the L0 norm, ,when When the convolution kernel is sparse; The sparse function is then input into a convolutional layer with a kernel width of 1 and then activated by ReLU. The convolution output of the jth input channel is : ； O j represents the output feature map of the 1x1 convolution in the dynamic kernel width layer, and its dimension is the same as the input feature map; w1 represents the weight corresponding to the convolution, and b represents the bias term; In the dynamic kernel width layer, k is dynamically adjusted to obtain different receptive fields for efficient feature extraction.

7. The method for identifying the health status of a mechatronic transmission system based on dynamic kernel width according to claim 1, characterized in that: The step S6 comprises: (1) Choose an opportunity to perform sparse updates: Check whether the current training iteration number t satisfies the multiple of the sparse update frequency Δt, that is, t mod Δt = 0; If the conditions are met, a sparse update is performed, otherwise the sparse update step is skipped and the gradient update is performed directly; According to the current training iteration number t, the total training iteration number T, the sparsity rate S and the cosine annealing function Calculate the number of weight updates num for each layer and each group; For each group of weights, according to the absolute value of the weights, select num weights for pruning, that is, change them from activated state to inactive state, and record these weight indices as ; For each group of weights, randomly select num unactivated weights to grow, that is, change them from inactive to active, and record these weight indices as ; Record the pruning and growing operations into the sparse function Z, ensuring that the sparse function reflects the activation state of the current weights; Substantially update the convolution kernel weights activated by the sparse function, that is, update the weight values ​​according to the gradient back propagation; (2) Choose the right time to update the gradient: Check whether the current training iteration number t satisfies the multiple of the sparse update frequency Δt, that is, t mod Δt ≠ 0; If the conditions are met, the gradient is updated, that is, the gradient is calculated according to the loss of forward propagation, and the network parameters are updated by backpropagation.

8. The method for identifying the health status of a mechatronic transmission system based on dynamic kernel width according to claim 7, characterized in that: The step S7 comprises: Repeat the forward propagation, loss calculation, sparse update and gradient update process until the total number of training iterations T is reached, completing the model training; A corresponding mechatronic transmission system health status identification model is obtained; and the health status identification of the mechatronic transmission system is realized based on the model.

Images