A federal adaptive fuzzy rough neural network industrial equipment fault diagnosis method

Abstract

This invention belongs to the field of industrial equipment fault diagnosis technology, specifically disclosing a federated adaptive fuzzy coarse neural network method for industrial equipment fault diagnosis. The method includes: a server initializing globally shared parameters and class condition parameters and distributing them to each client; each client using the received parameters as initial parameters for its local adaptive fuzzy coarse neural network for local training, calculating the fault class boundary reliability corresponding to each fault category; the client uploading the updated parameters, sample count, and fault class boundary reliability to the server; the server weighting and aggregating the globally shared parameters according to the sample count, and independently weighting and aggregating the class condition parameters for each category according to the fault class boundary reliability, iteratively updating until convergence, obtaining the final model, and distributing it to each client for local fault classification. This invention can characterize the uncertainty of fault decision boundaries, improve the diagnostic accuracy and robustness in industrial non-IID scenarios, while protecting data privacy.

Description

Technical Field

[0001] This invention belongs to the field of industrial equipment fault diagnosis technology, specifically relating to a federated adaptive fuzzy coarse neural network method for industrial equipment fault diagnosis. Background Technology

[0002] With the deepening development of intelligent manufacturing, modern factories have deployed a large number of production equipment, such as CNC machine tools, motors, and bearings. These devices are typically equipped with multiple types of sensors for real-time monitoring of their operating status. Intelligent fault diagnosis based on multi-sensor data is of great significance for ensuring production safety. However, equipment fault data involves trade secrets, and direct centralized collection poses a risk of privacy leakage. Furthermore, the workload and operating conditions of different devices vary significantly, resulting in monitoring data exhibiting non-independent identically distributed (non-IID) characteristics. Centralized models struggle to adapt to the differences in operating conditions across various devices, exhibiting poor generalization capabilities. Therefore, how to utilize multi-source data for collaborative fault diagnosis while protecting data privacy has become a critical issue that urgently needs to be addressed in the field of industrial intelligent operation and maintenance.

[0003] Federated learning, which trains models collaboratively between a server and multiple clients, allows devices to upload only model parameters without uploading raw data, and is gradually being introduced into the field of industrial fault diagnosis. Existing typical methods, such as FedAvg, use parameter-weighted averaging to obtain a global model; to alleviate data heterogeneity issues, FedProx introduces a proximal regularization term to limit local model bias; FedFNN and other federated fuzzy neural network methods handle non-IID and uncertainty issues through rule-based collaboration and evolutionary mechanisms. Furthermore, rough set theory and fuzzy rough set theory, through lower and upper approximations, characterize the determination and possible attribution relationships of sample pairs, and have been used for classification and uncertainty modeling in centralized environments.

[0004] However, existing technologies still have the following shortcomings: Existing federated learning methods struggle to characterize inconsistencies such as inter-class overlap, blurred boundaries, and similar samples corresponding to different labels, and their expressive power for blurred boundaries is limited. Server-side aggregation fails to distinguish the reliability differences in how different devices model the boundaries of different fault categories. When fault categories are missing, samples are imbalanced, or there is insufficient reference from other classes, it easily introduces unreliable information, reducing the accuracy of the global model. In summary, existing federated fault diagnosis methods are insufficient to effectively address the complex problems of inter-class overlap, blurred boundaries, and class imbalance in industrial non-IID scenarios. Summary of the Invention

[0005] The purpose of this invention is to address the problem that existing federated fault diagnosis methods are unable to effectively handle inter-class overlap, boundary ambiguity, and class imbalance in industrial non-IID scenarios. A federated adaptive fuzzy coarse neural network method for fault diagnosis of industrial equipment is proposed.

[0006] The technical solution of this invention is as follows: This invention provides a federated adaptive fuzzy coarse neural network method for fault diagnosis of industrial equipment, applied to a federated learning system consisting of a central diagnostic server and multiple production equipment clients, including: Step 1: The central diagnostic server initializes the model parameters of the global fault diagnosis model and sends the initialized model parameters to each production equipment client. The model parameters of the global fault diagnosis model include: globally shared parameters and class condition parameters. Step 2: Each production equipment client trains and updates the parameters of the local adaptive fuzzy coarse neural network based on the received model parameters and local multi-sensor monitoring data, and calculates the reliability of the fault class boundary corresponding to each fault category locally; Step 3: Each production equipment client uploads the updated global shared parameters, updated class condition parameters, total number of samples, number of samples corresponding to each fault category, and the reliability of the fault category boundary to the central diagnostic server. Step 4: The central diagnostic server performs weighted aggregation of the received global shared parameters according to the total number of samples from each production equipment client, and independently aggregates the received class condition parameters according to the fault category based on the reliability of the fault class boundary, to obtain the updated global fault diagnosis model. Step 5: Send the updated global fault diagnosis model parameters to each production equipment client. Repeat steps 2 to 5 until the preset stopping condition is met to obtain the final global fault diagnosis model. Send the final global fault diagnosis model parameters to each production equipment client to achieve fault classification on each production equipment client.

[0007] Preferably, the globally shared parameter is a globally shared feature weight, which is used to represent the contribution of different sensor feature dimensions to the fault diagnosis task, and satisfies that the weights of each sensor feature dimension are non-negative and their sum is 1; The class condition parameter is a class condition variance parameter, which is used to adaptively control the scale of fuzzy similarity relationships. Each fault category corresponds to a class condition variance parameter in each sensor feature dimension.

[0008] Preferably, step 2 includes: Step 2.1: Each production equipment client uses the received model parameters as the initial parameters for its local adaptive fuzzy coarse neural network; Step 2.2: Construct a corresponding fault category sample set and a heterogeneous sample set for each fault category based on the local multi-sensor monitoring data, and calculate a fuzzy rough approximation result to represent the degree of separation between the sample and the heterogeneous sample set based on the constructed fuzzy similarity relationship; Step 2.3: Calculate the degree to which each sample belongs to each fault category based on the approximate results under fuzzy coarse conditions; Step 2.4: Based on the degree of belonging and the preset local training loss function, calculate the local classification loss, and update the global shared parameters and class condition parameters of the local adaptive fuzzy coarse neural network by minimizing the local classification loss to obtain the updated global shared parameters and the updated class condition parameters; Step 2.5: Calculate the fault class boundary reliability corresponding to each fault category in the local area based on the degree of attribution.

[0009] Preferably, the fuzzy similarity relationship is defined as: ; In the formula, For the first The first production equipment client collected the first Multiple sensor signal samples, For the first The first production equipment client collected the first Multiple sensor signal samples, Indicates sample In the The values ​​of each sensor feature dimension Indicates sample In the The values ​​of each sensor feature dimension Indicates the first The fault category is in the first Class conditional variance parameters on each sensor feature dimension A positive smoothing coefficient. Indicates sample and samples In the The fault category and the first Similarity across sensor feature dimensions.

[0010] Preferably, the approximate result under fuzzy coarsening is calculated according to the following formula: ; In the formula, For the first The first production equipment client collected the first Multiple sensor signal samples, For the first The first production equipment client collected the first Multiple sensor signal samples, Indicates the first Among the production equipment clients, those not belonging to the first A set of outlier samples for each fault category. Indicates sample and samples In the The fault category and the first Similarity across sensor feature dimensions Indicates sample In the The fault category and the first Approximate results under fuzzy and rough conditions in each sensor feature dimension.

[0011] Preferably, step 2.3 includes: The fuzzy and coarse approximation results under each sensor feature dimension are weighted and fused using the globally shared parameters to obtain the degree of each sample's association with each fault category. The fault category with the highest degree of association is taken as the predicted category of the sample; where, the sample For the The degree of attribution for each fault category is calculated using the following formula: ; in, Indicates the first d Globally shared feature weights for each sensor feature dimension. This represents the total number of sensor feature dimensions. Indicates sample In the The fault category and the first Approximate results under fuzzy and rough conditions in each sensor feature dimension.

[0012] Preferably, the local training loss function is expressed as: ; In the formula, For the first The local training loss function for each production equipment client. For the first Total number of samples for each production equipment client Number of fault categories For indicator functions, For the sample Corresponding fault category labels, For the first Each fault category Indicates sample Belongs to the The predicted probability of each fault category. , Indicates sample For the The degree to which each fault category belongs.

[0013] Preferably, step 2.5 includes: For the current production equipment client, the first The fault category will belong to the first The average degree of belonging to all samples of a fault category is used as the original fault class boundary reliability of that fault category. A heterogeneous reference sufficiency factor is constructed to correct the original fault class boundary reliability, thereby obtaining the fault class boundary reliability corresponding to each local fault category.

[0014] Preferably, the heterogeneous reference sufficiency factor is expressed as: ; in, Indicates the first Among the production equipment clients, those not belonging to the first The number of outlier samples for each fault category The smoothing coefficient is positive.

[0015] Preferably, the received class condition parameters are independently aggregated according to the fault category based on the reliability of the fault class boundary, including: For the For each fault category, the central diagnostic server selects production equipment clients with samples of that fault category to participate in the aggregation of class condition parameters for that fault category. The aggregation weight is calculated according to the following formula: ; in, Indicates the first In the first round of training A collection of production equipment clients for each fault category sample. Indicates the first Among the production equipment clients, the one belonging to the first The number of samples for each fault category For the first The first production equipment client Fault class boundary reliability for each fault category Indicates the first The first production equipment client Aggregate weights of class condition parameters for each fault category; Based on the aggregation weight, the first The class condition parameters of each sensor feature dimension for each fault category are weighted and aggregated: ; in, Indicates the first After the first round of federated communication, the central diagnostic server aggregates the results. The fault category is in the first Class conditional variance parameter on each sensor feature dimension; Indicates the first In the first round of federal communications, The first production equipment client local training update The fault category is in the first Class conditional variance parameters on each sensor feature dimension This represents the total number of sensor feature dimensions.

[0016] The beneficial effects of this invention are: 1. The federated adaptive fuzzy coarse neural network method for industrial equipment fault diagnosis of the present invention introduces the degree of sample attribution to each fault category during the training process of each production equipment client. This allows for explicit characterization of whether a sensor signal sample is located in a definite region of a certain fault category or in a fuzzy region near the category boundary. For transitional regions between normal states and early faults, and feature overlap regions between different fault types, which are common in industrial fault diagnosis, the method can effectively distinguish between the classification certainty and uncertainty of samples. This improves the model's fault diagnosis accuracy and robustness in scenarios with overlapping fault classes, fuzzy boundaries, and noise interference, and reduces the rate of false and false diagnoses.

[0017] 2. The federated adaptive fuzzy coarse neural network method for industrial equipment fault diagnosis of the present invention, in the parameter aggregation stage of the central diagnostic server, independently aggregates the received class condition parameters according to the fault category. This ensures that for a given fault category, only production equipment clients whose local datasets actually contain samples of that category participate in the aggregation and updating of the class condition parameters; clients that do not contain samples of that category do not participate in the aggregation of those parameters. This avoids the situation in federated learning non-IID scenarios where a device, having never experienced a certain type of fault, uploads parameters unrelated to that fault category, which could unreasonably affect the global class condition parameters. This improves the global fault diagnosis model's ability to identify various faults, especially local and rare faults. Attached Figure Description

[0018] Figure 1 The diagram shows a flowchart of a federated adaptive fuzzy coarse neural network method for fault diagnosis of industrial equipment. Figure 2 The diagram shown is a flowchart of a federated adaptive fuzzy coarse neural network method for fault diagnosis of industrial equipment. Figure 3 The diagram shown is a schematic diagram of the approximate calculation under local fuzzy roughness for a certain production equipment; Figure 4 The figure shown is a graph of experimental results under different Dirichlet distribution parameters; Figure 5 The figure shown represents the experimental results for different numbers of clients. Detailed Implementation

[0019] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the drawings are merely exemplary and are intended to illustrate the principles and spirit of the invention, and are not intended to limit the scope of the invention.

[0020] This invention provides a federated adaptive fuzzy coarse neural network method for fault diagnosis of industrial equipment, applied to a federated learning system consisting of a central diagnostic server and multiple production equipment clients. The core of this method lies in constructing an adaptive fuzzy coarse neural network on each production equipment client. It constructs fuzzy similarity relationships between sensor signal samples using learnable class condition parameters, uses fuzzy coarseness to approximate and explicitly characterize the degree of definite attribution of sensor signal samples relative to the decision boundaries of each fault category, and fuses the approximate results of multiple sensor features through learnable feature weights to model the local uncertain fault classification relationships.

[0021] Please see Figure 1 and Figure 2 , Figure 1 The diagram shows a flowchart of a federated adaptive fuzzy coarse neural network method for fault diagnosis of industrial equipment. Figure 2 The diagram shows a process flowchart of a federated adaptive fuzzy coarse neural network method for fault diagnosis of industrial equipment.

[0022] like Figure 1 As shown in this embodiment, the federated adaptive fuzzy coarse neural network method for industrial equipment fault diagnosis includes the following steps: Step 1: The central diagnostic server initializes the model parameters of the global fault diagnosis model and distributes the initialized model parameters to the clients of each production equipment. The model parameters of the global fault diagnosis model include: globally shared parameters and class condition parameters.

[0023] In this embodiment, the globally shared parameter is the globally shared feature weight. The class conditional parameter is the class conditional variance parameter. The model parameters of the global fault diagnosis model can be expressed as: .

[0024] Among them, globally shared feature weights This is used to represent the contribution of different sensor feature dimensions to the fault diagnosis task, satisfying... and , For sensor feature dimensions. Class conditional variance parameter. The scale used for adaptive control of fuzzy similarity relationships, where each fault category corresponds to a class conditional variance parameter on each sensor feature dimension, is... The fault category is in the first The class conditional variance parameter on each sensor feature dimension is denoted as . .

[0025] Optionally, the central diagnostic server randomly initializes the globally shared feature weights. For a uniform distribution, the class conditional variance parameter If the value is positive, such as 1.0, it ensures that the weights of the globally shared features are non-negative and the variance is positive, and then it is distributed to all clients of the production equipment.

[0026] Step 2: Each production equipment client trains and updates the parameters of the local adaptive fuzzy coarse neural network based on the received model parameters and local multi-sensor monitoring data, and calculates the reliability of the fault class boundary corresponding to each fault category.

[0027] In this embodiment, step 2 includes: Step 2.1: Each production equipment client uses the received model parameters as the initial parameters for the local adaptive fuzzy coarse neural network.

[0028] Step 2.2: Construct a corresponding fault category sample set and a heterogeneous sample set for each fault category based on local multi-sensor monitoring data, and calculate a fuzzy rough approximation result to represent the degree of separation between the sample and the heterogeneous sample set based on the constructed fuzzy similarity relationship.

[0029] In this embodiment, it is assumed that the system has a total of Each production equipment client, such as CNC machine tools, motor units, and pump and valve equipment, is equipped with multiple sensors including vibration sensors, temperature sensors, and current sensors. The number of fault categories is... The sensor feature dimension is , No. The local multi-sensor monitoring dataset of each production equipment client is represented as follows: ; in, , indicating the first The first production equipment client collected the first Multiple sensor signal samples, , indicating sample Corresponding fault category labels, Indicates the first Total number of samples for each production equipment client.

[0030] For each fault category, a sample set and a set of outlier samples are constructed on the client side of each production equipment. The first production equipment client The sample set for each fault category is represented as: ; The corresponding set of outlier samples is represented as follows: ; in, Including normal samples and excluding the first Samples of other fault categories besides the one fault category.

[0031] In this embodiment, fuzzy similarity relationship is defined as: ; In the formula, For the first The first production equipment client collected the first Multiple sensor signal samples, For the first The first production equipment client collected the first Multiple sensor signal samples, Indicates sample In the The values ​​of each sensor feature dimension Indicates sample In the The values ​​of each sensor feature dimension Indicates the first The fault category is in the first Class conditional variance parameters on each sensor feature dimension A positive smoothing coefficient. Indicates sample and samples In the The fault category and the first Similarity across sensor feature dimensions.

[0032] The value range is [0,1]. The larger the value, the more similar the two samples are under the fault category and the feature dimension.

[0033] In this embodiment, each production equipment client uses fuzzy rough approximation to calculate the degree to which a sample is far from the heterogeneous set. In its first The fault category and the first The approximate result under fuzzy coarseness in each sensor feature dimension is calculated according to the following formula: ; In the formula, For the first The first production equipment client collected the first Multiple sensor signal samples, For the first The first production equipment client collected the first Multiple sensor signal samples, Indicates the first Among the production equipment clients, those not belonging to the first A set of outlier samples for each fault category. Indicates sample and samples In the The fault category and the first Similarity across sensor feature dimensions Indicates sample In the The fault category and the first Approximate results under fuzzy and rough conditions in each sensor feature dimension.

[0034] Please see Figure 3 , Figure 3 The figure shown is a schematic diagram of the approximate calculation under local fuzzy roughness for a certain production equipment. Used to characterize samples The degree of separation between the most similar outlier sample and the most similar outlier sample, when The larger the value, the better the sample size. The further away from the first The more certain the boundary region of a fault category is, the more certain it is to belong to that fault category; when The smaller the size, the better the sample size. More similar to outlier samples, they may be in the transitional region between normal and faulty conditions or the overlapping region between different fault types.

[0035] Step 2.3: Calculate the degree of each sample’s association with each fault category based on the approximate results under fuzzy coarse conditions.

[0036] Specifically, step 2.3 includes: using globally shared parameters to perform weighted fusion of the fuzzy and coarse approximation results under the feature dimensions of each sensor to obtain the degree of each sample’s belonging to each fault category, and taking the fault category with the highest degree of belonging as the predicted category of the sample.

[0037] In this embodiment, the sample For the The degree of attribution for each fault category is calculated using the following formula: ; in, Indicates the first dGlobally shared feature weights for each sensor feature dimension. This represents the total number of sensor feature dimensions. Indicates sample In the The fault category and the first Approximate results under fuzzy and rough conditions in each sensor feature dimension.

[0038] After calculating the degree of attribution for all fault categories, the predicted category of the sample is represented as: .

[0039] Step 2.4: Based on the degree of belonging and the preset local training loss function, calculate the local classification loss. Update the global shared parameters and class condition parameters of the local adaptive fuzzy coarse neural network by minimizing the local classification loss, and obtain the updated global shared parameters and updated class condition parameters.

[0040] In this embodiment, the local training loss function is expressed as: ; In the formula, For the first The local training loss function for each production equipment client. For the first Total number of samples for each production equipment client Number of fault categories For indicator functions, For the sample Corresponding fault category labels, For the first Each fault category Indicates sample Belongs to the The predicted probability of each fault category. , Indicates sample For the The degree to which each fault category belongs.

[0041] Step 2.5: Calculate the fault class boundary reliability corresponding to each fault category in the local system based on the degree of attribution.

[0042] In this embodiment, step 2.5 includes: for the current production equipment client... The fault category will belong to the first The average degree of belonging to all samples of a fault category is used as the original fault class boundary reliability of that fault category; a heterogeneous reference sufficiency factor is constructed to correct the original fault class boundary reliability, thus obtaining the fault class boundary reliability corresponding to each local fault category.

[0043] Specifically, for the first The first production equipment client Each fault category has a sample size represented as follows: , No. The production equipment client for the first The primitive class boundary reliability definition for each fault category is: .

[0044] This value is used to reflect the first The production equipment client for the first The credibility of boundary modeling for each fault category. The larger the value, the further away the fault category samples in the client of the production equipment are from the set of outliers, and the clearer the fault boundary.

[0045] To avoid overestimating reliability due to insufficient out-of-class reference samples, an out-of-class reference sufficiency factor is further constructed, expressed as: ; in, , indicating the first Among the production equipment clients, those not belonging to the first The number of outlier samples for each fault category A positive smoothing coefficient. The more abundant the outlier the reference samples, the better. The closer to 1, the less likely there are out-of-class reference samples. Decrease.

[0046] The reliability of the original fault class boundary is corrected using the heterogeneous reference sufficiency factor. The production equipment client for the first The fault class boundary reliability, after correction for each fault category, is defined as follows: .

[0047] Step 3: Each production equipment client uploads the updated global shared parameters, updated class condition parameters, total number of samples, number of samples corresponding to each fault category, and fault class boundary reliability to the central diagnostic server.

[0048] Step 4: The central diagnostic server performs weighted aggregation of the received global shared parameters according to the total number of samples from each production equipment client, and independently aggregates the received class condition parameters according to the fault category and the reliability of the fault class boundary to obtain the updated global fault diagnosis model.

[0049] In this embodiment, for those participating in the first Production equipment client collection for round training The central diagnostic server weights and aggregates the received globally shared parameters according to the total number of samples from each production equipment client. The specific calculation process is as follows: ; In the formula, Indicates the first After rounds of federated communication, the central diagnostic server aggregates the globally shared feature weights. Indicates the first In the first round of federal communications, The globally shared feature weights are updated locally on each production equipment client. Indicates the first Total number of local samples for each production equipment client.

[0050] In this embodiment, the received class condition parameters are independently aggregated according to the fault category and based on the reliability of the fault class boundary, including: For the For each fault category, the central diagnostic server selects production equipment clients with samples of that fault category to participate in the aggregation of class condition parameters for that fault category. The aggregation weight is calculated according to the following formula: ; in, Indicates the first In the first round of training A collection of production equipment clients for each fault category sample. Indicates the first Among the production equipment clients, the one belonging to the first The number of samples for each fault category For the first The first production equipment client Fault class boundary reliability for each fault category Indicates the first The first production equipment client Aggregate weights of class condition parameters for each fault category.

[0051] Based on the aggregation weight, the first The class condition parameters of each sensor feature dimension for each fault category are weighted and aggregated: ; in, Indicates the first After the first round of federated communication, the central diagnostic server aggregates the results. The fault category is in the first Class conditional variance parameter on each sensor feature dimension; Indicates the first In the first round of federal communications, The first production equipment client local training update The fault category is in the first Class conditional variance parameters on each sensor feature dimension This represents the total number of sensor feature dimensions.

[0052] In this embodiment, model parameters are divided into globally shared parameters and class-conditional parameters. Globally shared parameters are collaboratively aggregated based on the total number of samples from each production equipment client. Class-conditional parameters are aggregated separately according to fault categories, ensuring that only devices with samples of the corresponding fault category participate in updating the parameters for that fault category, thus avoiding the introduction of invalid information from devices with missing fault categories. Class-conditional parameters are weighted and aggregated based on the reliability of fault category boundaries. Fault category-level reliability is constructed using the local fuzzy coarse approximation results from each production equipment client, and a heterogeneous reference sufficiency factor is introduced to correct the reliability. This allows the central diagnostic server to allocate aggregation weights based on the credibility of boundary modeling for each production equipment client in different fault categories, thereby enabling devices with more sufficient and reliable fault boundary estimations to contribute more to the aggregation of corresponding fault category parameters.

[0053] Step 5: Send the updated global fault diagnosis model parameters to each production equipment client. Repeat steps 2 to 5 until the preset stopping condition is met to obtain the final global fault diagnosis model. Send the final global fault diagnosis model parameters to each production equipment client to achieve fault classification on each production equipment client.

[0054] In this embodiment, the preset stopping condition can be reaching a preset number of training rounds. After training is complete, the model parameters of the final global fault diagnosis model are used. and Fault classification is performed on the real-time multi-sensor monitoring signals of the equipment to be diagnosed. For each sample to be diagnosed, the degree of its belonging to each fault category is calculated according to the fuzzy similarity relation, fuzzy coarse approximation, and feature weight fusion process, and the category with the highest degree of belonging is output as the fault diagnosis result. The fault diagnosis results can be output to the equipment operation and maintenance management system for fault early warning, maintenance scheduling or equipment status assessment.

[0055] The federated adaptive fuzzy coarse neural network method for industrial equipment fault diagnosis of this invention introduces an adaptive fuzzy coarse neural network fault modeling mechanism into each production equipment client. By approximating the separation degree of sensor signal samples relative to the heterogeneous class set under fuzzy similarity relations and fuzzy coarsening, it can explicitly characterize the uncertainty of the fault decision boundary at the level of the relationship between sensor signal samples and fault categories. For common situations in industrial fault diagnosis such as overlapping fault classes, blurred boundaries, noise interference, and similar signal samples corresponding to different fault labels, this invention can distinguish whether a sample is in a fault-determined region or a boundary-uncertain region. Therefore, it can improve the recognition stability of the fault diagnosis model under complex industrial data distributions and reduce the rate of false and false diagnoses.

[0056] Furthermore, the effectiveness of the federated adaptive fuzzy coarse neural network industrial equipment fault diagnosis method of the present invention is illustrated through simulation experiments.

[0057] 1. Experimental Preparation To verify the effectiveness of the method of this invention in industrial equipment fault diagnosis scenarios, this experiment selected three publicly available industrial datasets: SteelPlates Faults (SPF), SECOM, and Sensorless Drive Diagnosis (SDD). The SPF dataset is used for steel plate surface defect type identification and includes 7 types of steel plate defects; the SECOM dataset originates from semiconductor manufacturing processes and belongs to industrial process anomaly detection datasets; the SDD dataset is used for motor drive state diagnosis and includes 11 types of motor drive states.

[0058] To simulate the data distribution differences among different production equipment in real-world industrial scenarios, each dataset is divided into multiple clients, each representing either a production equipment or an edge diagnostic node. The experiment uses a Dirichlet distribution to construct the non-IID data partition, where the Dirichlet parameter α controls the degree of data heterogeneity among the production equipment clients. A smaller α indicates a greater difference in class distribution among different production equipment clients, and a stronger degree of non-IID; a larger α indicates that the data distribution of the production equipment clients is closer to the overall data distribution. The experiment uses Accuracy and Macro-F1 as evaluation metrics. Accuracy measures the overall classification accuracy, while Macro-F1 measures the average classification performance of each class under imbalanced conditions.

[0059] 2. Comparative Experiment In the comparative experiments, the FedAFRNN method proposed in this invention was compared with FedAvg, FedProx, and FedFNN. FedAvg is a basic federated averaging method, FedProx is a federated optimization method for heterogeneous data scenarios, FedFNN is a federated fuzzy neural network method, and FedAFRNN is the federated adaptive fuzzy coarse neural network method proposed in this invention. With Dirichlet parameter α=0.5, the classification accuracy results of different methods on the SPF, SECOM, and SDD datasets are shown in Table 1.

[0060] Table 1

[0061] As shown in Table 1, the method of this invention achieved the highest classification accuracy on all three industrial datasets. These results demonstrate that by introducing a fuzzy, coarse approximation mechanism on the client side and a class boundary reliability-weighted aggregation mechanism on the server side, this invention can more effectively address the problems of imbalanced class distribution, fuzzy fault boundaries, and differences in local model reliability on the client side in industrial non-IID data, thereby improving the classification performance of the global fault diagnosis model.

[0062] 3. Ablation test To further verify the effectiveness of the key technical features of this invention, an ablation experiment was designed. Table 2 shows the four different methods used in the ablation experiment: Table 2

[0063] The classification results of each ablation experiment on the SPF and SDD datasets are shown in Tables 3 and 4.

[0064] Table 3

[0065] Table 4

[0066] As shown in Tables 3 and 4, with the gradual introduction of category-aware aggregation, fuzzy coarse boundary reliability, and class boundary reliability weighted aggregation mechanisms, the model's accuracy and Macro-F1 scores on the SPF and SDD datasets have both improved. This result indicates that the performance improvement of this invention does not stem from a single module, but rather from the synergistic effect of client-side fuzzy coarse boundary modeling and server-side reliability weighted aggregation.

[0067] To verify the robustness of the method of the present invention under different non-IID levels, experiments were conducted on the SDD dataset with different Dirichlet parameters α.

[0068] Please see Figure 4 , Figure 4 The figure shown is a graph of experimental results under different Dirichlet distribution parameters. From... Figure 4 It can be seen that when the data is highly heterogeneous, traditional FedAvg and FedProx are easily affected by local class bias of the client. However, the method of this invention can reduce the impact of low-reliability client updates on the global model through class-aware aggregation and class boundary reliability weighted aggregation, thus maintaining good classification performance under different α settings.

[0069] To verify the applicability of the method of the present invention under different numbers of production equipment, experiments were conducted on the SDD dataset with different numbers of clients. The number of clients was set to 5, 6, 7, 8, 9, and 10, respectively.

[0070] Please see Figure 5 , Figure 5 The image shows the experimental results for different numbers of clients. Figure 5 As can be seen, with the increase in the number of clients, the traditional federated averaging method is easily affected by local client bias. However, the method of this invention can maintain relatively stable fault diagnosis performance when the number of clients changes by only allowing clients with samples of the corresponding fault category to participate in the aggregation of the category parameters and allocating aggregation weights according to the reliability of the class boundary.

[0071] The results of the comprehensive comparative experiments and ablation experiments show that the federated adaptive fuzzy coarse neural network method for industrial equipment fault diagnosis proposed in this invention outperforms existing federated learning methods such as FedAvg, FedProx, and FedFNN on industrial datasets such as SPF, SECOM, and SDD. This invention, through a combined technical solution of fuzzy coarse fault boundary modeling at the production equipment end, fault class condition parameter partitioning, fault category perception parameter aggregation, and fault class boundary reliability weighted aggregation, can improve the fault diagnosis accuracy, robustness, and interpretability of the federated fault diagnosis model in industrial non-IID data, fault class imbalance, fault class overlap, boundary ambiguity, and noise interference scenarios. It also reduces the adverse impact of low-reliability device client updates on global class condition parameters, while ensuring that the original industrial sensor data does not leave the domain, thus helping to meet industrial data security requirements.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0073] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0074] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A federated adaptive fuzzy coarse neural network method for fault diagnosis of industrial equipment, applied to a federated learning system consisting of a central diagnostic server and multiple production equipment clients, characterized in that, include: Step 1: The central diagnostic server initializes the model parameters of the global fault diagnosis model and sends the initialized model parameters to each production equipment client. The model parameters of the global fault diagnosis model include: globally shared parameters and class condition parameters. Step 2: Each production equipment client trains and updates the parameters of the local adaptive fuzzy coarse neural network based on the received model parameters and local multi-sensor monitoring data, and calculates the reliability of the fault class boundary corresponding to each fault category locally; Step 3: Each production equipment client uploads the updated global shared parameters, updated class condition parameters, total number of samples, number of samples corresponding to each fault category, and the reliability of the fault category boundary to the central diagnostic server. Step 4: The central diagnostic server performs weighted aggregation of the received global shared parameters according to the total number of samples from each production equipment client, and independently aggregates the received class condition parameters according to the fault category based on the reliability of the fault class boundary, to obtain the updated global fault diagnosis model. Step 5: Send the updated global fault diagnosis model parameters to each production equipment client. Repeat steps 2 to 5 until the preset stopping condition is met to obtain the final global fault diagnosis model. Send the final global fault diagnosis model parameters to each production equipment client to achieve fault classification on each production equipment client.

2. The industrial equipment fault diagnosis method using a federated adaptive fuzzy coarse neural network according to claim 1, characterized in that, The global shared parameter is the global shared feature weight, which is used to represent the contribution of different sensor feature dimensions to the fault diagnosis task, and satisfies that the weights of each sensor feature dimension are non-negative and the sum is 1. The class condition parameter is a class condition variance parameter, which is used to adaptively control the scale of fuzzy similarity relationships. Each fault category corresponds to a class condition variance parameter in each sensor feature dimension.

3. The federal adaptive fuzzy rough neural network industrial equipment fault diagnostic method of claim 1, wherein, Step 2 includes: Step 2.1: Each production equipment client uses the received model parameters as the initial parameters for its local adaptive fuzzy coarse neural network; Step 2.2: Construct a corresponding fault category sample set and a heterogeneous sample set for each fault category based on the local multi-sensor monitoring data, and calculate a fuzzy rough approximation result to represent the degree of separation between the sample and the heterogeneous sample set based on the constructed fuzzy similarity relationship; Step 2.3: Calculate the degree to which each sample belongs to each fault category based on the approximate results under fuzzy coarse conditions; Step 2.4: Based on the degree of belonging and the preset local training loss function, calculate the local classification loss, and update the global shared parameters and class condition parameters of the local adaptive fuzzy coarse neural network by minimizing the local classification loss to obtain the updated global shared parameters and the updated class condition parameters; Step 2.5: Calculate the fault class boundary reliability corresponding to each fault category in the local area based on the degree of attribution.

4. The federated self-adaptive fuzzy rough neural network industrial equipment fault diagnostic method of claim 3, wherein, The fuzzy similarity relationship is defined as follows: ； In the formula, For the first The first production equipment client collected the first Multiple sensor signal samples, For the first The first production equipment client collected the first Multiple sensor signal samples, Indicates sample In the The values ​​of each sensor feature dimension Indicates sample In the The values ​​of each sensor feature dimension Indicates the first The fault category is in the first Class conditional variance parameters on each sensor feature dimension A positive smoothing coefficient. Indicates sample and samples In the The fault category and the first Similarity across sensor feature dimensions.

5. The industrial equipment fault diagnosis method using a federated adaptive fuzzy coarse neural network according to claim 3, characterized in that, The approximate result under fuzzy coarsening is calculated according to the following formula: ； In the formula, For the first The first production equipment client collected the first Multiple sensor signal samples, For the first The first production equipment client collected the first Multiple sensor signal samples, Indicates the first Among the production equipment clients, those not belonging to the first A set of outlier samples for each fault category. Indicates sample and samples In the The fault category and the first Similarity across sensor feature dimensions Indicates sample In the The fault category and the first Approximate results under fuzzy and rough conditions in each sensor feature dimension.

6. The federal adaptive fuzzy rough neural network industrial equipment fault diagnostic method of claim 3, wherein, Step 2.3 includes: The fuzzy rough lower approximation results of each sensor feature dimension are weighted and fused by using the global shared parameter to obtain the attribution degree of each sample to each fault category, and the fault category with the maximum attribution degree is taken as the predicted category of the sample; wherein the attribution degree of the sample to the first fault category is calculated according to the following formula: ； in, Indicates the first d Globally shared feature weights for each sensor feature dimension. This represents the total number of sensor feature dimensions. Indicates sample In the The fault category and the first Approximate results under fuzzy and rough conditions in each sensor feature dimension.

7. The industrial equipment fault diagnosis method using a federated adaptive fuzzy coarse neural network according to claim 3, characterized in that, The local training loss function is expressed as: ； In the formula, For the first The local training loss function for each production equipment client. For the first Total number of samples for each production equipment client Number of fault categories For indicator functions, For the sample Corresponding fault category labels, For the first Each fault category Indicates sample Belongs to the The predicted probability of each fault category. , Indicates sample For the first The degree to which each fault category belongs.

8. The federal adaptive fuzzy rough neural network industrial equipment fault diagnostic method of claim 3, wherein, Step 2.5 includes: For the first fault category of the current production equipment client, the average of the attribution degrees of all samples belonging to the first fault category is taken as the original fault category boundary reliability of the fault category. A heterogeneous reference sufficiency factor is constructed to correct the original fault class boundary reliability, thereby obtaining the fault class boundary reliability corresponding to each local fault category.

9. The federated self-adaptive fuzzy rough neural network industrial equipment fault diagnostic method of claim 8, wherein, The heterogeneous reference sufficiency factor is expressed as: ； wherein, represents the number of heterogeneous samples in the th production equipment client that do not belong to the th failure category, is a positive smoothing coefficient.

10. The federated self-adaptive fuzzy rough neural network industrial equipment fault diagnostic method of claim 1, wherein, The received class condition parameters are independently aggregated according to the fault category and based on the reliability of the fault class boundary, including: For the For each fault category, the central diagnostic server selects production equipment clients with samples of that fault category to participate in the aggregation of class condition parameters for that fault category. The aggregation weight is calculated according to the following formula: ； in, Indicates the first In the first round of training A collection of production equipment clients for each fault category sample. Indicates the first Among the production equipment clients, the one belonging to the first The number of samples for each fault category For the first The first production equipment client Fault class boundary reliability for each fault category Indicates the first The first production equipment client Aggregate weights of class condition parameters for each fault category; Based on the aggregation weight, the first The class condition parameters of each sensor feature dimension for each fault category are weighted and aggregated: ； in, Indicates the first After the first round of federated communication, the central diagnostic server aggregates the results. The fault category is in the first Class conditional variance parameter on each sensor feature dimension; Indicates the first In the first round of federal communications, The first production equipment client local training update The fault category is in the first Class conditional variance parameters on each sensor feature dimension This represents the total number of sensor feature dimensions.

Images