Dynamic vision-enabled non-contact equipment migration fault diagnosis method

Abstract

A non-contact equipment migration fault diagnosis method empowered by dynamic vision is proposed. First, dynamic visual data of equipment vibration is acquired non-contactly using an event camera. Then, the dynamic visual data is segmented to construct an initial intelligent diagnostic dataset. Next, a cross-domain diffusion generation model for dynamic visual data is established to supplement missing fault data in the target domain. Then, an intelligent diagnostic model based on neuromorphic computing technology for dynamic visual data feature extraction and intelligent identification of equipment fault modes is constructed. The training set from the final intelligent diagnostic dataset is input into the intelligent diagnostic model for training to obtain deep feature representations. Then, various losses are calculated, and the parameters of the intelligent diagnostic model are iteratively updated and optimized until training is complete, obtaining the final intelligent diagnostic model. Finally, the test set is input into the final intelligent diagnostic model to obtain the labels corresponding to the categories of the test set data. This invention achieves non-contact intelligent migration fault diagnosis of mechanical equipment.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment fault diagnosis technology, specifically relating to a dynamic vision-enabled non-contact equipment migration fault diagnosis method. Background Technology

[0002] Mechanical equipment is crucial in fields such as intelligent manufacturing, transportation, and aerospace, and timely and accurate fault diagnosis is key to ensuring its safe and reliable operation. Although good fault diagnosis results have been achieved in experimental scenarios, how to achieve efficient and accurate fault diagnosis in engineering scenarios remains an urgent problem to be solved.

[0003] To obtain health status information of mechanical equipment, contact measurement methods are usually used to collect vibration signals. However, these methods have strict requirements on factors such as sensor installation space and deployment environment. Therefore, contact vibration measurement fault diagnosis methods are often difficult to apply in many engineering scenarios.

[0004] Existing non-contact vibration measurement methods often suffer from drawbacks such as high cost, difficult deployment, and poor robustness. Event cameras, a high-speed asynchronous dynamic vision sensor inspired by biological vision, possess excellent characteristics such as extremely high temporal resolution, high dynamic range, low data transmission bandwidth, and low power consumption, showing broad application prospects in the field of non-contact mechanical vibration monitoring and fault diagnosis. However, research on dynamic vision-based mechanical vibration measurement and diagnosis is currently very scarce.

[0005] And it is usually based on the premise that experimental data and measured data follow the same distribution ([1] LI Xiang, YUShupeng, LEI Yaguo, et al. Intelligent machinery fault diagnosis with event-based camera[J]. IEEE Transactions on Industrial Informatics, 2023: 1-10; [2] ZHAO Mingzhe, SHEN Xiaojun, JIANG Fei. Research on mechanical vibration measurement method based on event camera[C] / / 2023 3rd International Conference on Energy Engineering and Power Systems. Dali: IEEE, 2023, 528-532.), however, in engineering scenarios, the actual operating state of equipment often differs greatly from experimental data; and fault data is usually difficult to obtain during actual operation of equipment, so existing transfer learning methods are also difficult to apply ([3] SCHWENDEMANN S, SIKORA Z A. Bearing fault diagnosis with intermediate domainbased layered maximum mean discrepancy: A new transfer learning approach[J]. Engineering Applications of Artificial Intelligence, 2021, 105: 10445; [4] Jia Sixiang, Sun Dingyi, Mao Gang, et al. Fault diagnosis method for rotor system under cross-condition based on anti-entropy [J]. Journal of Mechanical Engineering, 2023, 59(15): 110-120. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, the present invention aims to provide a dynamic vision-enabled non-contact equipment migration fault diagnosis method, which solves the cross-domain diagnosis problem in scenarios where it is difficult to obtain fault data, and realizes non-contact intelligent migration fault diagnosis of mechanical equipment.

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

[0008] A non-contact equipment migration fault diagnosis method empowered by dynamic vision is proposed. First, dynamic visual data of equipment vibration is collected using an event camera. Second, a cross-domain diffusion generation model for dynamic visual data is established to realize the intelligent generation of dynamic visual data of unknown equipment fault states under measured data. Finally, an intelligent diagnosis model for dynamic visual data feature extraction and equipment fault modes based on brain-like computing technology is established.

[0009] A dynamic vision-enabled non-contact equipment migration fault diagnosis method includes the following steps:

[0010] Step 1: Use an event camera to non-contactly acquire dynamic visual data of equipment vibration to characterize the health status of the equipment.

[0011] Step 2: Divide the dynamic visual data collected in Step 1 to construct an initial intelligent diagnostic dataset;

[0012] Step 3: Establish a cross-domain diffusion generation model for dynamic visual data to supplement the missing fault data in the target domain, thereby obtaining the final intelligent diagnostic dataset after sample generation and supplementation.

[0013] Step 4: Construct an intelligent diagnostic model for dynamic visual data feature extraction and equipment fault mode intelligent recognition based on neuromorphic computing technology; the temporal dimension of the dynamic visual data will be merged first; then the data will be processed through four pulsed convolutional modules to extract deep features. Between the first three adjacent pulsed convolutional modules, two summing pooling layers will be used to compress the data features. After that, these deep pulsed features will be flattened to one dimension and output through a linear pulsed layer; finally, the temporal dimension of the output features will be decompressed and classified by a classifier.

[0014] Step 5: Input the training set of the final intelligent diagnostic dataset obtained in Step 3 into the intelligent diagnostic model constructed in Step 4, and extract the deep feature representations of the source domain and target domain data in the final intelligent diagnostic dataset;

[0015] Step 6: Using the deep feature representations of the source and target domain data obtained in Step 5, calculate the cross-domain diagnostic loss and the classification diagnostic loss respectively, and optimize and update the model parameters;

[0016] Step 7: Repeat steps 5-6, iterating and updating the parameters of the intelligent diagnostic model until training is complete and the final intelligent diagnostic model is obtained.

[0017] Step 8: Input the test set from the intelligent diagnostic dataset into the trained final intelligent diagnostic model to obtain the label of the category corresponding to the test set data.

[0018] Step 1 specifically includes:

[0019] Over a period of time, the event camera continuously captures images of the equipment in the target area. The resulting dynamic visual data is output as an event stream, and the events are stored in chronological order of occurrence. The event stream is represented as follows: ,in Indicates the first An event that occurred. This represents the total number of events that occurred; a single event is represented as a quadruple:

[0020]

[0021] In the formula, where Indicates the time when the event occurred. and These represent the horizontal and vertical coordinates of the location where the event occurred in the camera coordinate system. Indicates the polarity of the event, with two possible values: -1 and 1; when When a positive event occurs, it indicates that a positive event has occurred, meaning the brightness of that pixel increases; conversely, when a negative event occurs... At that time, a negative polarity event (negative event) occurred in the pixel, that is, the brightness of the point decreased.

[0022] Step 2 specifically includes:

[0023] For the collected dynamic visual data The dynamic visual data collected in the source domain is defined as... The dynamic visual data collected in the target domain is defined as In the target domain, a subset of health status data is used for training the intelligent diagnostic model, defined as... The remaining data is used for testing the intelligent diagnostic model, defined as follows: Therefore, the training set of the initial intelligent diagnostic dataset With test set Represented as:

[0024]

[0025] In the formula, and These represent the total number of events in the source domain and the target domain, respectively. and These represent the total number of events used for model training and testing in the target domain, respectively.

[0026] Step 3 specifically includes:

[0027] The process is divided into a training process and a sampling process: first, source domain data is used. Train a diffusion generation model from healthy to faulty states for source domain data. Collect overall health status data The forward process of the input diffusion generation model is gradually noise-added. Let be the total number of health state events in the source domain; let the noise sequence be . ,from Choose any value This indicates the superposition of the original samples. Sub-Gaussian noise; the forward diffusion process is represented as:

[0028]

[0029] In the formula, This represents the input source domain health status data sample. Indicates superposition of them The noisy sample following the secondary noise. This represents a sample of health status data from the input source domain. The superposition of the two is derived Noisy samples after secondary noise The conditional probability distribution, It is the identity matrix. These are parameters related to noise; they are used to obtain noisy samples. Then, samples are generated step by step through a reverse process:

[0030]

[0031] In the formula, These are the parameters to be optimized in the diffusion generation model. Indicates the presence of noisy samples Denoising can be used to advance the noise-added samples in the previous step. The conditional probability distribution, This represents the mean of the noise distribution in the reverse process. The distribution is Gaussian; to generate source domain fault state samples across domains from source domain health state samples, source domain fault data is required. Supervise the reverse generation process. The number of fault data events in the source domain, therefore the optimization function Defined as:

[0032] In the formula, Indicates the expectation. This represents the mean of the actual noise distribution during the forward noise addition process of the source domain fault data. The mean noise distribution predicted by the diffusion generation model when the input is source domain health status data. as well as These are source domain fault data samples and samples with added noise. Subsequent samples;

[0033] During the sampling process, target domain health data are collected from the training set. As input to the diffusion generation model, corresponding fault state pseudo-data is generated. , The number of pseudo-data events in the target domain generated by the diffusion generation model; the pseudo-data generated by the diffusion generation model is used to supplement the initial intelligent diagnostic dataset, resulting in the final training set of the intelligent diagnostic dataset. That is, expanded to:

[0034] .

[0035] Step 6 specifically includes:

[0036] Calculate cross-domain diagnostic loss and classification diagnosis loss :

[0037]

[0038]

[0039] In the formula, For the output pulse of the last layer, For source domain sample labels, It is the number of time channels. Indicates having a characteristic kernel The regenerating nucleus Hilbert space, and Let represent the probability distributions of the deep features in the source and target domains, respectively; therefore, the overall optimization objective is expressed as:

[0040]

[0041] In the formula, For the total loss function, The weights are used for cross-domain diagnostic loss; the spatiotemporal backpropagation parameter optimization method based on the spiking neural network intelligent diagnostic model is as follows:

[0042]

[0043]

[0044]

[0045]

[0046] In the formula, Indicates the first Layer, number The first LIF neuron Each network weight, For time steps, and They represent in Time of the first The first layer Membrane potential and output of a single neuron Indicates in Time of the first The first layer One output, Indicates in Time of the first The first layer One output, Indicates in Time of the first The first layer A membrane potential, It is the number of time channels. This is the learning rate.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] This invention proposes a dynamic vision-enabled non-contact equipment migration fault diagnosis method. It employs a non-contact event camera to acquire dynamic visual data of equipment vibration, overcoming the difficulty of applying current contact vibration measurements in many engineering scenarios, and providing an innovative solution for mechanical vibration measurement and fault diagnosis. Furthermore, this invention proposes a diffusion generation model for dynamic visual data, enabling the generation of pseudo-data on unknown fault states in the target domain, laying a solid data foundation for migration fault diagnosis. Finally, this invention proposes a brain-like computing-based spiking neural network method, achieving intelligent feature extraction and fault mode recognition of dynamic visual data, significantly improving data processing efficiency and migration fault diagnosis effectiveness compared to traditional methods. Attached Figure Description

[0049] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram illustrating the generation of pseudo-data for unknown fault states in an embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram of the experimental platform structure according to an embodiment of the present invention. Detailed Implementation

[0052] The present invention will now be described in detail with reference to the embodiments and accompanying drawings.

[0053] Reference Figure 1 A dynamic vision-enabled non-contact equipment migration fault diagnosis method includes the following steps:

[0054] Step 1: Use an event camera to non-contactly acquire dynamic visual data of equipment vibration to characterize the health status of the mechanical equipment.

[0055] In this embodiment, an event camera continuously captures images of the equipment in the target area over a period of time. The obtained dynamic visual data is output as an event stream, and the events in the event stream are stored in chronological order of occurrence. The event stream can be represented as: ,in Indicates the first An event that occurred. This represents the total number of events that occurred; a single event can be represented as a quadruple:

[0056]

[0057] In the formula, where Indicates the time when the event occurred. and These represent the horizontal and vertical coordinates of the location where the event occurred in the camera coordinate system. Indicates the polarity of the event, generally with two possible values: -1 and 1; when When a positive event occurs, it indicates that a positive event has occurred, meaning the brightness of that pixel increases; conversely, when a negative event occurs... At that time, a negative polarity event (negative event) occurred in that pixel, meaning that the brightness of that point decreased;

[0058] Step 2: Divide the dynamic visual data collected in Step 1 to construct an initial intelligent diagnostic dataset;

[0059] This embodiment focuses on the acquired dynamic visual data. The dynamic visual data collected in the source domain is defined as... The dynamic visual data collected in the target domain is defined as In the target domain, a subset of health status data is used for training the intelligent diagnostic model, defined as... The remaining data is used for testing the intelligent diagnostic model, defined as follows: Therefore, the training set of the initial intelligent diagnostic dataset With test set It can be represented as:

[0060]

[0061] In the formula, and These represent the total number of events in the source domain and the target domain, respectively. and These represent the total number of events used for model training and testing in the target domain, respectively.

[0062] Step 3: As Figure 2 As shown, a cross-domain diffusion generation model for dynamic visual data is established to supplement the missing fault data in the target domain, thereby obtaining the final intelligent diagnostic dataset after sample generation and supplementation.

[0063] Step 3 of this embodiment includes two main processes: a training process and a sampling process. First, source domain data is used. Train a diffusion generation model from healthy to faulty states for source domain data. Collect overall health status data The forward process of the input diffusion generation model is gradually noise-added. Let be the total number of health state events in the source domain; let the noise sequence be . ,from Choose any value This indicates the superposition of the original samples. Sub-Gaussian noise; the forward diffusion process can be represented as...

[0064]

[0065] In the formula, This represents the input source domain health status data sample. Indicates superposition of them The noisy sample following the secondary noise. This represents a sample of health status data from the input source domain. The superposition of the two is derived Noisy samples after secondary noise The conditional probability distribution, It is the identity matrix. These are parameters related to noise; they are used to obtain noisy samples. Then, samples can be generated step by step through the reverse process:

[0066]

[0067] In the formula, These are the parameters to be optimized in the diffusion generation model. Indicates the presence of noisy samples Denoising can be used to advance the noise-added samples in the previous step. The conditional probability distribution, This represents the mean of the noise distribution in the reverse process. The distribution is Gaussian; to generate source domain fault state samples across domains from source domain health state samples, source domain fault data is required. Supervise the reverse generation process. The number of fault data events in the source domain, therefore the optimization function Defined as:

[0068] In the formula, Indicates the expectation. This represents the mean of the actual noise distribution during the forward noise addition process of the source domain fault data. The mean noise distribution predicted by the diffusion generation model when the input is source domain health status data. as well as These are source domain fault data samples and samples with added noise. Subsequent samples;

[0069] During the sampling process, target domain health data are collected from the training set. As input to the diffusion generation model, corresponding target domain fault pseudo-data can be generated. , The number of pseudo-data events in the target domain generated by the diffusion generation model; the pseudo-data generated by the diffusion generation model is used to supplement the initial intelligent diagnostic dataset, resulting in the final training set of the intelligent diagnostic dataset. That is, expanded to:

[0070] ;

[0071] Step 4: Construct an intelligent diagnostic model for dynamic visual data feature extraction and equipment fault mode intelligent recognition based on neuromorphic computing technology; specifically, the temporal dimension of the dynamic visual data will be merged first; then, the data will be processed through four pulsed convolutional modules to extract deep features. Between the first three adjacent pulsed convolutional modules, two summing pooling layers are used to compress the data features. After that, these deep pulsed features will be flattened to one dimension and output through a linear pulsed layer; finally, the temporal dimension of the output features will be decompressed and classified by a classifier.

[0072] Step 5: Input the training set of the final intelligent diagnostic dataset obtained in Step 3 into the intelligent diagnostic model constructed in Step 4, and extract the deep feature representations of the source domain and target domain data in the final intelligent diagnostic dataset;

[0073] Step 6: Using the deep feature representations of the source and target domain data obtained in Step 5, calculate the cross-domain diagnostic loss. and classification diagnosis loss :

[0074]

[0075]

[0076] In the formula, For the output pulse of the last layer, For source domain sample labels, It is the number of time channels. Indicates having a characteristic kernel The regenerating nucleus Hilbert space, and Let represent the probability distributions of the deep features in the source and target domains, respectively; therefore, the overall optimization objective can be expressed as:

[0077]

[0078] In the formula, For the total loss function, The weights are used for cross-domain diagnostic loss; the spatiotemporal backpropagation parameter optimization method based on the spiking neural network intelligent diagnostic model is as follows:

[0079]

[0080]

[0081]

[0082]

[0083] In the formula, Indicates the first Layer, number The first LIF neuron Each network weight, For time steps, and They represent in Time of the first The first layer Membrane potential and output of a single neuron Indicates in Time of the first The first layer One output, Indicates in Time of the first The first layer One output, Indicates in Time of the first The first layer A membrane potential, It is the number of time channels. The learning rate;

[0084] Step 7: Repeat steps 5-6, iterating and updating the parameters of the intelligent diagnostic model until training is complete and the final intelligent diagnostic model is obtained.

[0085] Step 8: Input the test set from the intelligent diagnostic dataset into the trained final intelligent diagnostic model to obtain the label of the category corresponding to the test set data.

[0086] Based on experimental data from rolling bearings, the effectiveness of the method in this embodiment was verified through experiments:

[0087] Reference Figure 3 In this embodiment, the experimental platform is driven by motor 1 and connected to shaft 3 via coupling 2, thereby driving water pump 4. Shaft 3 is connected to base 6 via bearing 5, model ER-16K. Event camera 7 is placed directly opposite bearing 5, and the specific parameters of event camera 7 are shown in Table 1. The experiment considered four different health states of bearing 5, including normal state, outer ring failure, inner ring failure, and rolling element failure. Source domain data was collected at a rotational speed of 40Hz, and target domain data was collected at rotational speeds of 20Hz, 25Hz, and 30Hz, respectively. The total dataset size is 4000 samples, and the dataset settings are shown in Table 2.

[0088] Table 1 Event camera parameters used in the embodiments

[0089]

[0090] Table 2. Setup of the diagnostic dataset

[0091]

[0092] To fully verify the effectiveness of the method of this invention, five other mainstream methods were compared. CNN: As the basic comparison method, it uses the currently mainstream Convolutional Neural Networks (CNN) as the feature extraction and diagnostic network, without using transfer learning. TCNN: It uses a CNN as the feature extraction and diagnostic network and a transfer learning method to measure distribution differences; however, it does not use a diffusion generation model to supplement the fault samples in the target domain, i.e., the target domain only contains 400 samples of healthy state data. TDCNN: It uses a CNN as the feature extraction and diagnostic network, uses a transfer learning method to measure distribution differences, and uses a diffusion generation model to generate fault state data in the target domain, which contains 400 samples of each of four different health states. SNN: It uses the spiking neural network proposed in this invention as the feature extraction and diagnostic network, but does not use transfer learning. TSNN: It uses a spiking neural network as the feature extraction and diagnostic network, uses a transfer learning method to optimize distribution differences, but does not use a diffusion generation model to supplement the generated fault state samples in the target domain. To eliminate the randomness of model training, the problem of fault diagnosis during the migration of the target domain at three rotational speeds was considered. Each experiment was conducted five times, and the average and standard deviation of the test accuracy were calculated.

[0093] Table 3 Comparison of Migration Fault Diagnosis Experiment Results

[0094]

[0095] In three different transfer fault diagnosis experiments, the method of this invention achieved a test accuracy of over 92% and a standard deviation of less than 2%, verifying the effectiveness of the method. This invention proposes a cross-domain dynamic visual data generation method. The generated pseudo-fault state data effectively integrates the features of source domain fault state data and target domain health state data, thus better measuring the distribution differences between domains and improving the accuracy of transfer fault diagnosis. Compared to methods without pseudo-data generation, the test accuracy of this invention can be improved by more than 15%. This invention's method can learn richer temporal features, which is crucial for rotating machinery fault diagnosis. Even with only partial access to the target domain's health state data, the transfer fault diagnosis accuracy can be improved by nearly 10%, while CNN-based methods only improve by about 3%. Furthermore, after further supplementing the target domain's fault state data, the average transfer fault diagnosis accuracy of this invention is improved by about 3.5% compared to CNN-based methods, with a smaller standard deviation and more stable results.

Claims

1. A dynamic vision-enabled non-contact equipment migration fault diagnosis method, characterized in that, Includes the following steps: Step 1: Use an event camera to non-contactly acquire dynamic visual data of equipment vibration to characterize the health status of the equipment. Step 2: Divide the dynamic visual data collected in Step 1 to construct an initial intelligent diagnostic dataset; Step 3: Establish a cross-domain diffusion generation model for dynamic visual data to supplement the missing fault data in the target domain, thereby obtaining the final intelligent diagnostic dataset after sample generation and supplementation. Step 4: Construct an intelligent diagnostic model for dynamic visual data feature extraction and equipment fault mode intelligent recognition based on neuromorphic computing technology; the temporal dimension of the dynamic visual data will be merged first; then the data will be processed through four pulsed convolutional modules to extract deep features. Between the first three adjacent pulsed convolutional modules, two summing pooling layers will be used to compress the data features. After that, these deep pulsed features will be flattened to one dimension and output through a linear pulsed layer; finally, the temporal dimension of the output features will be decompressed and classified by a classifier. Step 5: Input the training set of the final intelligent diagnostic dataset obtained in Step 3 into the intelligent diagnostic model constructed in Step 4, and extract the deep feature representations of the source domain and target domain data in the final intelligent diagnostic dataset; Step 6: Using the deep feature representations of the source and target domain data obtained in Step 5, calculate the cross-domain diagnostic loss and the classification diagnostic loss respectively, and optimize and update the model parameters; Step 7: Repeat steps 5-6, iterating and updating the parameters of the intelligent diagnostic model until training is complete and the final intelligent diagnostic model is obtained. Step 8: Input the test set from the intelligent diagnostic dataset into the trained final intelligent diagnostic model to obtain the label of the category corresponding to the test set data.

2. The method according to claim 1, characterized in that, Step 1 specifically includes: Over a period of time, the event camera continuously captures images of the equipment in the target area, and the obtained dynamic visual data is output as an event stream. The events are stored in chronological order of occurrence. Event stream representation is as follows: ,in Indicates the first An event that occurred. This represents the total number of events that occurred; a single event is represented as a quadruple: In the formula, where Indicates the time when the event occurred. and These represent the horizontal and vertical coordinates of the location where the event occurred in the camera coordinate system. Indicates the polarity of the event, with two possible values: -1 and 1; when When a pixel brightness increases, it indicates a positive event; conversely, when a negative event occurs... At that time, a negative event occurred in the pixel, that is, the pixel brightness decreased.

3. The method according to claim 2, characterized in that, Step 2 specifically includes: For the collected dynamic visual data The dynamic visual data collected in the source domain is defined as... The dynamic visual data collected in the target domain is defined as In the target domain, a subset of health status data is used for training the intelligent diagnostic model, defined as... The remaining data is used for testing the intelligent diagnostic model, defined as follows: ; Therefore, the training set of the initial intelligent diagnostic dataset With test set Represented as: In the formula, and These represent the total number of events in the source domain and the target domain, respectively. and These represent the total number of events used for model training and testing in the target domain, respectively.

4. The method according to claim 3, characterized in that, Step 3 specifically includes: The process is divided into a training process and a sampling process: first, source domain data is used. Train a diffusion generation model from healthy to faulty states for source domain data. Collect overall health status data The forward process of the input diffusion generation model is gradually noise-added. Let be the total number of health state events in the source domain; let the noise sequence be . ,from Choose any value This indicates the superposition of the original samples. Sub-Gaussian noise; the forward diffusion process is represented as: In the formula, This represents the input source domain health status data sample. Indicates superposition of them The noisy sample following the secondary noise. This represents a sample of health status data from the input source domain. The superposition of the two is derived Noisy samples after secondary noise The conditional probability distribution, It is the identity matrix. These are parameters related to noise; Obtaining noisy samples Then, samples are generated step by step through a reverse process: In the formula, These are the parameters to be optimized in the diffusion generation model. Indicates the presence of noisy samples Denoising can be used to advance the noise-added samples in the previous step. The conditional probability distribution, This represents the mean of the noise distribution in the reverse process. The distribution is Gaussian; to generate source domain fault state samples across domains from source domain health state samples, source domain fault data is required. Supervise the reverse generation process. The number of fault data events in the source domain, therefore the optimization function Defined as: In the formula, Indicates the expectation. This represents the mean of the actual noise distribution during the forward noise addition process of the source domain fault data. The mean noise distribution predicted by the diffusion generation model when the input is source domain health status data. as well as These are source domain fault data samples and samples with added noise. Subsequent samples; During the sampling process, target domain health data are collected from the training set. As input to the diffusion generation model, corresponding fault state pseudo-data is generated. , The number of target domain fault pseudo-data events generated for the diffusion generation model; The initial intelligent diagnostic dataset was supplemented with pseudo-data generated using a diffusion generation model, resulting in the final training set of the intelligent diagnostic dataset. That is, expanded to: 。 5. The method according to claim 4, characterized in that, Step 6 specifically includes: Calculate cross-domain diagnostic loss and classification diagnosis loss : In the formula, For the output pulse of the last layer, For source domain sample labels, It is the number of time channels. Indicates having a characteristic kernel The regenerating nucleus Hilbert space, and Let represent the probability distributions of the deep features in the source and target domains, respectively; therefore, the overall optimization objective is expressed as: In the formula, For the total loss function, The weights are used for cross-domain diagnostic loss; the spatiotemporal backpropagation parameter optimization method based on the spiking neural network intelligent diagnostic model is as follows: In the formula, Indicates the first Layer, number The first LIF neuron Each network weight, For time steps, and They represent in Time of the first The first layer Membrane potential and output of a single neuron Indicates in Time of the first The first layer One output, Indicates in Time of the first The first layer One output, Indicates in Time of the first The first layer A membrane potential, It is the number of time channels. This is the learning rate.

Images