Event camera and spiking neural network based mechanical fault diagnosis method under varying working conditions

Abstract

The application discloses a variable working condition mechanical fault diagnosis method based on an event camera and a pulse neural network, and has the advantages that the event camera is used to asynchronously collect vibration signals of a rotating machine, so that original event stream data of the rotating machine are obtained; pulse data representation is constructed through the original event stream data; a pulse event data set is constructed through the pulse data representation; a convolution type pulse neural network intelligent diagnosis model is constructed, and a training set of the pulse event data set is sent into the intelligent diagnosis model for training, so that deep pulse feature representation is obtained; the features in the time dimension are compressed, each loss is calculated, and the parameters of the model are optimized and updated until the model converges; the application can extract domain invariant features from different working conditions without target domain machine fault data, saves the calculation amount of the deep intelligent diagnosis model, and improves the applicability of the deep intelligent diagnosis model in actual engineering scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of rotating machinery fault diagnosis, and particularly relates to a variable working condition mechanical fault diagnosis method based on an event camera and a pulse neural network. BACKGROUND

[0002] Rotating machinery has been widely used in the field of industrial manufacturing and is one of the key parts indispensable to various mechanical equipment. Once the rotating machinery fails, it will affect the overall performance of the mechanical equipment, and even cause serious safety accidents. Therefore, it is particularly important to develop an efficient rotating machinery fault diagnosis method, and the intelligent fault diagnosis method for rotating machinery has developed rapidly in recent years.

[0003] At present, the most commonly used signal type in the field of fault diagnosis is the vibration acceleration data collected by a contact type acceleration sensor ([1] Bian Wenbin, Deng Aidong, Liu Dongchuan, et al. Wind turbine rolling bearing fault diagnosis method based on improved deep residual shrinkage network [J]. Journal of Mechanical Engineering, 2023, 59(12): 202-214; [2] Meng Zong, Guan Yang, Pan Zuozhou, et al. Rolling bearing fault diagnosis based on secondary data enhancement and deep convolution [J]. Journal of Mechanical Engineering, 2021, 57(23): 106-115), however, the deployment of the contact type acceleration sensor in industrial applications has great limitations, and the existing non-contact sensors are usually high in cost or low in efficiency.

[0004] Meanwhile, the current mainstream methods based on traditional deep neural networks ([3] J. Lee, M. Kim, JUKo, JHJung and KHSun, “Asymmetric inter-intra domain alignments (AIIDA) method for intelligent fault diagnosis of rotating machinery,” Reliability Engineering & System Safety, vol. 218, part. B, pp. 108-186, Feb. 2022; [4] WTMao, GSWang, LLKou, and XHLiang, “Deep domain-adversarial anomaly detection with one-class transfer learning,” IEEE / CAA J. Autom. Sinica, vol. 10, no. 2, pp. 524-546, Feb. 2023) have high computational requirements, which makes it difficult for intelligent diagnostic models to be deployed in real time on edge computing hardware, and the performance of the models is difficult to be further improved. This makes the current mainstream intelligent fault diagnosis methods less applicable to practical engineering problems.

[0005] In addition, since there are often slight differences between the available laboratory data and the data collected in actual engineering scenarios, the distribution of training data and test data often becomes inconsistent. Especially in actual engineering scenarios, fault data is difficult to obtain, so conventional transfer learning methods ([5] Ye Nan, Chang Peize, Zhang Luyu et al. Research on multi-condition bearing fault diagnosis based on improved semi-supervised deep belief network [J]. Journal of Mechanical Engineering, 2021, 57(15): 80-9) are difficult to apply. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the present invention aims to provide a fault diagnosis method for rotating machinery based on an event camera and a pulse neural network. First, an event camera is used to collect vibration signals of rotating machinery. Then, a pulse convolutional neural network is used to extract features from the event data. Furthermore, cross-domain diagnosis of rotating machinery is achieved in the absence of target domain fault data, thereby improving the applicability of deep intelligent diagnosis methods in engineering practice.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A fault diagnosis method for machinery under varying operating conditions based on event cameras and spiking neural networks includes the following steps:

[0009] Step 1: Use an event camera to asynchronously acquire vibration signals from the rotating machinery to obtain raw event stream data of the rotating machinery;

[0010] Step 2: Process the raw event stream data obtained in Step 1 to construct a pulse data representation;

[0011] Step 3: Divide the pulse data representation constructed in Step 2 to construct a pulse event dataset;

[0012] Step 4: Construct an initial spiking neural network intelligent diagnostic model in convolutional form. First, use two convolutional layers to extract shallow and deep features from the data, and then use two max pooling layers to compress the shallow and deep features. After each max pooling layer, use a LIF spiking layer as the activation function. Finally, the processed high-dimensional features are flattened and output through two linear layers and one LIF spiking layer.

[0013] Step 5: Input the training set of the pulse event dataset obtained in Step 3 into the initial spiking neural network intelligent diagnostic model constructed in Step 4 for training, so as to obtain deep pulse feature representations of the source domain and the target domain.

[0014] Step 6: Compress the temporal features of the deep impulse features in the source and target domains obtained in Step 5.

[0015] Step 7: Using the compressed source and target domain deep pulse features obtained in Step 6, calculate each loss, optimize and update the parameters of the intelligent diagnostic model;

[0016] Step 8: Repeat steps 5-7, iterating and optimizing the intelligent diagnostic model until training is complete, and obtain the final spiking neural network intelligent diagnostic model;

[0017] Step 9: Input the test set containing fault data into the final spiking neural network intelligent diagnostic model to obtain the label of the corresponding category in the target domain.

[0018] Step 1 specifically includes:

[0019] Vibration signals from rotating machinery are collected over a continuous period of time. These signals are stored as events in chronological order of occurrence. A series of consecutive events constitutes an event stream, which is represented as follows: Where e i Let N represent the i-th event, and N represent the n-th event. EThis represents the total number of events that occurred within this continuous time period; a single event is a four-dimensional array containing the event's time information, spatial information, and polarity. A single event is represented as:

[0020] e i =(t i x i y i p i )

[0021] Where t i Let x represent the time when the i-th event occurs. i and y i p represents the x-coordinate and y-coordinate position of the i-th event, respectively. i Indicate the polarity of the i-th event, when p i When p = 1, it indicates that the event is a positive event, meaning that an increase in brightness has occurred at that location, and the magnitude of the brightness change exceeds the threshold. i When the value is 0, it indicates that the event is negative, meaning that a decrease in brightness has occurred at that location, and the magnitude of the brightness change exceeds the threshold.

[0022] Step 2 specifically refers to:

[0023] For the raw event stream data acquired using the event camera in step 1, a rectangular region needs to be set as the boundary, denoted as (x... bl x bu y bl y bu ), where x bu and x bl These represent the upper and lower boundary values ​​of the rectangular region on the x-axis, and y-axis, respectively. bu and y bl These represent the upper and lower boundary values ​​of the rectangular region on the y-axis, respectively. Events outside the boundaries are discarded, and events inside the boundaries are subjected to spatial and temporal standardization.

[0024] x i =x i -x bl

[0025] y i =y i -y bl

[0026] t i =t i -t0+1

[0027] Where t0 is the time of the first event occurring within the boundary, the event stream after filtering and standardization is represented as follows: Where NROI The total number of events that occur within the boundary;

[0028] Based on the time width t0 of a single sample and the number of samples N s Confirm the total duration T0 of the required event stream:

[0029] T0 = ​​t0 * N s

[0030] Construct pulse data representations, representing each sample as a three-dimensional tensor (v) containing temporal and spatial information. t v x v y ), where v t The time dimension of the sample, its value is related to the time width t0 of a single sample; v x With v y This represents the spatial dimension of the sample, and its value is related to the specifications of the selected rectangular region.

[0031] Step 3 specifically includes:

[0032] The pulse data representation constructed in step 2 is first divided into the source domain. and target domain Where N sou N represents the total number of samples in the source domain. tar d represents the total number of samples in the target domain. j For the j-th impulse event sample, l j For sample d j The corresponding labels; the data in the source domain contains the labels corresponding to the categories, while the data in the target domain is not processed; then the training set is constructed as follows:

[0033]

[0034] Where N train The target set contains the total number of samples used for training. The training set includes all data from the source domain and some data from the target domain that does not contain fault states. The test set consists of the remaining data from the target domain that includes all healthy states.

[0035]

[0036] Where N test This represents the total number of samples used for testing in the target domain.

[0037] The deep pulse features of the source and target domains in step 5 are represented as follows:

[0038]

[0039]

[0040] Where u is the membrane potential of the LIF neuron, o is the impulse output of the LIF neuron, m and n are the m-th and n-th input synapses, respectively, and t s This represents the time step, and q represents the network layer number. The delay effect of the membrane potential, f(·) is the step function, u th This is the potential threshold, which directly affects the pulse output.

[0041] In step 6, compression is represented as follows:

[0042]

[0043] Where v represents the eigenvalue, T represents the total time step, and s t This represents the peak value of the spiking neural network output at time t.

[0044] Step 7 specifically includes:

[0045] Using the compressed source and target domain deep impulse features obtained in step 6, the classification loss L is calculated respectively. cls Domain migration loss L MMD Inter-class separation loss L inter and intraclass dense loss L intra This optimizes and updates the model's parameters.

[0046]

[0047]

[0048]

[0049]

[0050] Where cls is the number of categories, y c p represents the true label of the c-th class of data. c For the predicted value of the c-th type of data, Let E represent the reproducing kernel Hilbert space with characteristic kernel k. S With E T Both represent the expectation, φ(s) s ) and φ(s t ) represent the source domain x respectively s With the target domain x t The distribution of w (q) (s (c) ) represents the features of the c-th class of data after the q-th layer of the intelligent diagnostic model; the overall optimization objective is expressed as:

[0051]

[0052] Where α and β are the weight hyperparameters for inter-class separation loss and intra-class density loss, respectively.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] This invention proposes a fault diagnosis method for machinery under varying operating conditions based on an event camera and a spiking neural network. The method employs an event camera to non-contactly acquire vibration signals from a visual perspective, and uses a convolutional spiking neural network intelligent diagnostic model to process the event stream, extract features, and diagnose faults. This intelligent diagnostic model can also extract domain-invariant features from different operating conditions without requiring mechanical fault data in the target domain. This invention further reduces the computational burden of deep intelligent diagnostic models, improves their applicability in practical engineering scenarios, and provides an efficient method for deploying deep intelligent diagnostic models in real-world engineering environments. Attached Figure Description

[0055] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0056] Figure 2 This is a schematic diagram of cross-domain diagnosis and optimization of a spiking neural network according to an embodiment of the present invention. Detailed Implementation

[0057] The present invention will now be described in detail with reference to the embodiments and accompanying drawings.

[0058] Reference Figure 1 A method for diagnosing mechanical faults under varying operating conditions based on event cameras and spiking neural networks includes the following steps:

[0059] Step 1: Use an event camera to asynchronously acquire vibration signals from the rotating machinery to obtain raw event stream data of the rotating machinery;

[0060] This embodiment acquires vibration signals from rotating machinery over a continuous period of time. The vibration signals are stored as events in chronological order of occurrence. A series of consecutive events constitutes an event stream, which can be represented as follows: Where e i Let N represent the i-th event, and N represent the n-th event. E This represents the total number of events that occurred within this continuous time period; a single event is a four-dimensional array containing the event's time information, spatial information, and polarity. A single event can be represented as:

[0061] e i =(t i x i y i p i )

[0062] Where t i Let x represent the time when the i-th event occurs. i and y i p represents the x-coordinate and y-coordinate position of the i-th event, respectively. i Indicate the polarity of the i-th event, when p i When p = 1, it indicates that the event is a positive event, meaning that an increase in brightness has occurred at that location, and the magnitude of the brightness change exceeds the threshold. i When = 0, it indicates that the event is a negative event, that is, an event of decreased brightness has occurred at this location, and the magnitude of the brightness change exceeds the threshold;

[0063] Step 2: Process the raw event stream data obtained in Step 1 to construct a pulse data representation;

[0064] This embodiment focuses on the raw event stream data acquired using the event camera in step 1. First, a rectangular region needs to be set as the boundary to filter events within the key region. This rectangular region can be represented as (x... bl x bu y bl y bu ), where x bu and x bl These represent the upper and lower boundary values ​​of the rectangular region on the x-axis, and y-axis, respectively. bu and y bl These represent the upper and lower boundary values ​​of the rectangular region on the y-axis, respectively. After selecting the size of the rectangular region, all events outside the boundaries are discarded, and spatial and temporal standardization is performed on events inside the boundaries.

[0065] x i =x i -x bl

[0066] y i =y i -y bl

[0067] t i =t i -t0+1

[0068] Where t0 is the time of the first event occurring within the boundary, the event stream after filtering and standardization can be represented as: Where N ROI The total number of events that occur within the boundary;

[0069] Obtain the event stream after filtering and standardization. Then, it is necessary to determine the time width t0 of a single sample and the number of samples N. sConfirm the total duration T0 of the required event stream:

[0070] T0 = ​​t0 * N s

[0071] After obtaining the event stream required to generate each sample, it is necessary to construct an impulse data representation, representing each sample as a three-dimensional tensor (v) containing temporal and spatial information. t v x v y ), where v t This is the time dimension of the sample, and this value is related to the time width t0 of a single sample, v x With v y This represents the spatial dimension of the sample, and its value is related to the dimensions of the selected rectangular region.

[0072] Step 3: Divide the pulse data representation constructed in Step 2 to construct a pulse event dataset;

[0073] In this embodiment, the pulse data representation constructed in step 2 is first divided into the source domain. and target domain Where N sou N represents the total number of samples in the source domain. tar d represents the total number of samples in the target domain. j For the j-th impulse event sample, l j For sample d j The corresponding labels; the data in the source domain contains the labels corresponding to the categories, while the data in the target domain is not processed; then the training set is constructed as follows:

[0074]

[0075] Where N train The target set contains the total number of samples used for training. The training set includes all data from the source domain and some data from the target domain that does not contain fault states. The test set consists of the remaining data from the target domain that includes all healthy states.

[0076]

[0077] Where N test This represents the total number of samples used for testing in the target domain.

[0078] Step 4: Construct an initial spiking neural network intelligent diagnostic model in convolutional form. First, two convolutional layers are used to extract shallow and deep features from the data, respectively. Then, two max pooling layers are used to compress these shallow and deep features. After each max pooling layer, a LIF spiking layer is used as the activation function. Finally, the processed high-dimensional features are flattened and output through two linear layers and one LIF spiking layer.

[0079] Step 5: Input the training set of the pulse event dataset obtained in Step 3 into the initial spiking neural network intelligent diagnostic model constructed in Step 4 for training, so as to obtain deep pulse feature representations of the source domain and the target domain.

[0080] The deep pulse characteristics of the source and target domains are represented as follows:

[0081]

[0082]

[0083] Where u is the membrane potential of the LIF neuron, o is the impulse output of the LIF neuron, m and n are the m-th and n-th input synapses, respectively, and t s This represents the time step, and q represents the network layer number. The delay effect of the membrane potential, f(·) is the step function, u th This is the potential threshold, which directly affects the pulse output;

[0084] Step 6: As Figure 2 As shown, the deep pulse feature representations of the source and target domains obtained in step 5 are compressed in the time dimension, and the compressed representation is as follows:

[0085]

[0086] Where v represents the eigenvalue, T represents the total time step, and s t This represents the peak value of the spiking neural network output at time t.

[0087] Step 7: Using the compressed deep impulse features of the source and target domains obtained in Step 6, calculate the classification loss L respectively. cls Domain migration loss L MMD Inter-class separation loss L inter and intraclass dense loss L intra This optimizes and updates the model's parameters.

[0088]

[0089]

[0090]

[0091]

[0092] Where cls is the number of categories, y c p represents the true label of the c-th class of data. c For the predicted value of the c-th type of data, Let E represent the reproducing kernel Hilbert space with characteristic kernel k. S With E T Both represent the expectation, φ(x) s ) and φ(x t ) represent the source domain x respectively s With the target domain x t The distribution of w (q) (s (c) () represents the features of the c-th class of data after the q-th layer of the intelligent diagnostic model; the overall optimization objective can be expressed as:

[0093]

[0094] Where α and β are the weight hyperparameters of the inter-class separation loss and the intra-class dense loss, respectively;

[0095] Step 8: Repeat steps 5-7, iterating and optimizing the intelligent diagnostic model until training is complete, and obtain the final spiking neural network intelligent diagnostic model;

[0096] Step 9: Input the test set containing fault data into the final spiking neural network intelligent diagnostic model to obtain the label of the corresponding category in the target domain.

[0097] Based on experimental data from rolling bearings, the effectiveness of the method in this embodiment was verified through experiments:

[0098] The test bench used in this embodiment is driven by a motor, which drives the shaft through a coupling. The bearing on the shaft is model ER-16K. The event camera is placed directly in front of the bearing. The event camera used is Prophesee version 3.1. The specific parameters of the camera are shown in Table 1. Four states of the rolling bearing were collected, including healthy, outer ring fault, inner ring fault, and rolling element fault. The source domain data was collected at a rotation speed of 40Hz, and the target domain data was collected at a rotation speed of 30Hz. When generating samples, the number of time steps was 10, the duration of a single time step was 1ms, and the RoI size was 30×30. In this embodiment, a total of 1000 samples were generated for each class, the training set size was 800, and the test set size was 200. The constructed event representation dataset is shown in Table 2.

[0099] Table 1 Event camera parameters used in the embodiments

[0100]

[0101] Table 2 shows the constructed event representation dataset.

[0102]

[0103] After dividing the dataset, both the source domain data and the target domain data (excluding fault states) are input into the intelligent diagnostic model. The basic training settings are as follows: mini-batch size is 128, number of iterations is 200, learning rate is 0.0001, and the Adam optimizer is used. After inputting the source domain data and target domain data into the intelligent diagnostic model, the domain transfer loss, classification loss, and two distance metric losses are calculated respectively. The weights of the two distance metric losses are both 1. After calculating the total loss, the parameters are optimized and updated. The process is iterated until the final intelligent diagnostic model is obtained. The target domain test set, which is unlabeled and contains other fault data, is input into the intelligent diagnostic model to obtain the classification results of various rolling bearing health states.

[0104] The experiment was conducted three times on the test set to minimize the influence of randomness in deep learning. Three other different methods were selected for comparison with the method of this invention: Method 1 uses a mainstream convolutional neural network as the basic feature extraction network and also uses the domain transfer method; Method 2 uses a mainstream convolutional neural network as the basic feature extraction network and does not use the domain adaptation method; Method 3 uses the spiking neural network constructed in this invention as the basic feature extraction network and also does not use the domain adaptation method.

[0105] As shown in Table 3, based on the results of the three experiments, it can be concluded that the method of the present invention achieved diagnostic accuracies of 98.13%, 98.28%, and 97.97% on the test set. This verifies the feasibility of using an event camera and a spiking neural network for cross-domain fault diagnosis in the present invention.

[0106] Table 3 Comparison of diagnostic results from different methods

[0107]

[0108] Furthermore, the experimental results show that: First, the method of this invention is highly competitive compared to traditional CNN-based methods. Using this method, the average accuracy can reach 98.12%, slightly higher than the 95.99% of CNN-based methods, and the convergence speed during training is faster. In addition, this method improves the model generalization performance at different rotation speeds; compared to methods without domain transfer, the average accuracy of this method is improved by approximately 13%, and it has good application prospects. Due to the advantages of spiking neural networks, this method is more energy-efficient than traditional CNN-based methods.

Claims

1. A method for diagnosing mechanical faults under varying operating conditions based on an event camera and a spiking neural network, characterized in that, Includes the following steps: Step 1: Use an event camera to asynchronously acquire vibration signals from the rotating machinery to obtain raw event stream data of the rotating machinery; Step 2: Process the raw event stream data obtained in Step 1. For the raw event stream data collected by the event camera in Step 1, set a rectangular area as the boundary, and perform spatial and temporal standardization on the events within the boundary to construct a pulse data representation. Step 3: Divide the pulse data representation constructed in Step 2 to construct a pulse event dataset; Step 3 specifically includes: The pulse data representation constructed in step 2 is first divided into the source domain. and target domain Where N sou N represents the total number of samples in the source domain. tar d represents the total number of samples in the target domain. j For the j-th impulse event sample, l j For sample d j The corresponding labels; the data in the source domain contains the labels corresponding to the categories, while the data in the target domain is not processed; then the training set is constructed as follows: Where N train The target set contains the total number of samples used for training. The training set includes all data from the source domain and some data from the target domain that does not contain fault states. The test set consists of the remaining data from the target domain that includes all healthy states. Where N test This represents the total number of samples used for testing in the target domain. Step 4: Construct an initial spiking neural network intelligent diagnostic model in convolutional form. First, use two convolutional layers to extract shallow and deep features from the data, and then use two max pooling layers to compress the shallow and deep features. After each max pooling layer, use a LIF spiking layer as the activation function. Finally, the processed high-dimensional features are flattened and output through two linear layers and one LIF spiking layer. Step 5: Input the training set of the pulse event dataset obtained in Step 3 into the initial spiking neural network intelligent diagnostic model constructed in Step 4 for training, so as to obtain deep pulse feature representations of the source domain and the target domain. Step 6: Compress the temporal features of the deep impulse features in the source and target domains obtained in Step 5. Step 7: Using the compressed deep impulse features of the source and target domains obtained in Step 6, calculate the classification loss L respectively. cls Domain migration loss L MMD Inter-class separation loss L inter and intraclass dense loss L intra Optimize and update the parameters of the intelligent diagnostic model; Step 8: Repeat steps 5-7, iterating and optimizing the intelligent diagnostic model until training is complete, and obtain the final spiking neural network intelligent diagnostic model; Step 9: Input the test set containing fault data into the final spiking neural network intelligent diagnostic model to obtain the label of the corresponding category in the target domain.

2. The method according to claim 1, characterized in that, Step 1 specifically includes: Vibration signals from rotating machinery are collected over a continuous period of time. These signals are stored as events in chronological order of occurrence. A series of consecutive events constitutes an event stream, which is represented as follows: Where e i Let N represent the i-th event, and N represent the n-th event. E This represents the total number of events that occurred within this continuous time period; a single event is a four-dimensional array containing the event's time information, spatial information, and polarity. A single event is represented as: e i =(t i ,x i ,y i ,p i ) Where t i Let x represent the time when the i-th event occurs. i and y i p represents the x-coordinate and y-coordinate position of the i-th event, respectively. i Indicate the polarity of the i-th event, when p i When p = 1, it indicates that the event is a positive event, meaning that an increase in brightness has occurred at that location, and the magnitude of the brightness change exceeds the threshold. i When the value is 0, it indicates that the event is negative, meaning that a decrease in brightness has occurred at that location, and the magnitude of the brightness change exceeds the threshold.

3. The method according to claim 2, characterized in that, Step 2 specifically refers to: The rectangular region is represented as (x bl ,x bu ,y bl ,y bu ), where x bu and x bl These represent the upper and lower boundary values ​​of the rectangular region on the x-axis, and y-axis, respectively. bu and y bl These represent the upper and lower boundary values ​​of the rectangular region on the y-axis, respectively, discarding all events outside the boundaries: x i =x i -x bl and i / and i -and bl t i =t i -t0+1 Where t0 is the time of the first event occurring within the boundary, the event stream after filtering and standardization is represented as follows: Where N ROI The total number of events that occur within the boundary; Based on the time width t0 of a single sample and the number of samples N s Confirm the total duration T0 of the required event stream: T0=t0*N s Construct a pulse data representation, representing each sample as a three-dimensional tensor (v) containing temporal and spatial information. t ,v x ,v y ), where v t The time dimension of the sample, its value is related to the time width t0 of a single sample; v x With v y This represents the spatial dimension of the sample, and its value is related to the specifications of the selected rectangular region.

4. The method according to claim 1, characterized in that, The deep pulse features of the source and target domains in step 5 are represented as follows: Where u is the membrane potential of the LIF neuron, o is the impulse output of the LIF neuron, m and n are the m-th and n-th input synapses, respectively, and t s This represents the time step, and q represents the network layer number. The delay effect of the membrane potential, f(·) is the step function, u th This is the potential threshold, which directly affects the pulse output.

5. The method according to claim 1, characterized in that, In step 6, compression is represented as follows: Where v represents the eigenvalue, T represents the total time step, and s t This represents the peak value of the spiking neural network output at time t.

6. The method according to claim 1, characterized in that, Step 7 specifically includes: Using the compressed source and target domain deep impulse features obtained in step 6, the classification loss L is calculated respectively. cls Domain migration loss L MMD Inter-class separation loss L inter and intraclass dense loss L intra This optimizes and updates the model's parameters. Where cls is the number of categories, y c p represents the true label of the c-th class of data. c For the predicted value of the c-th type of data, Let E represent the reproducing kernel Hilbert space with characteristic kernel K. S With E T Both represent the expectation, φ(s) s ) and φ(s t ) represent the source domain x respectively s With the target domain x t The distribution of w (q) (s (c) ) represents the features of the c-th class of data after the q-th layer of the intelligent diagnostic model; the overall optimization objective is expressed as: Where α and β are the weight hyperparameters for inter-class separation loss and intra-class density loss, respectively.

Images