Industrial equipment fault diagnosis method and system based on multilevel federated learning

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an industrial equipment fault diagnosis method and system based on multilevel federated learning, and the method comprises the steps: collecting operation time sequence data of industrial equipment, training a fault diagnosis model deployed at a current equipment end, and obtaining trained local model parameters; based on a dynamic weight distribution strategy, performing weighted aggregation on the received local model parameters of the plurality of equipment ends to generate region-level model parameters; integrating all region-level model parameters, and injecting differential privacy noise to generate a global fault diagnosis model; carrying out collaborative optimization compression on the global fault diagnosis model, and distributing the compressed model to each equipment end to update the fault diagnosis model currently deployed by the equipment end; and based on the contribution data of each equipment end in the current round and historical federated learning, executing a contribution evaluation and incentive distribution process, and updating the dynamic weight distribution strategy with an evaluation result. And the quick response requirement of an industrial scene is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to an industrial equipment fault diagnosis method and system based on multi-federated learning. BACKGROUND

[0002] Industrial equipment fault diagnosis is a key technology to ensure the continuity, safety and reliability of modern industrial production. With the deep integration of Industrial Internet of Things (IIoT) and artificial intelligence (AI) technologies, data-driven intelligent diagnosis methods have become the mainstream. However, the centralized data analysis paradigm widely adopted by the current industry is facing severe challenges. This paradigm requires all the massive equipment operation data distributed in different geographical areas and different plants to be transmitted to the cloud or data center for unified modeling, and fault prediction is achieved through unified modeling. Such methods rely on a single subject to control all data, and there are the following problems: First, industrial production data contains core process parameters and operating status, which are highly sensitive assets for enterprises. There are strict "data silos" between different enterprises and even different plants under the same group, which refuse to directly share raw data. This results in the centralized model being unable to train with global data, having weak generalization ability, and being difficult to cope with diverse fault patterns.

[0003] Second, the brands, models, working conditions and loads of industrial field devices are very different, and the data generated by them presents typical Non-IID characteristics. Traditional global models based on the assumption of independent and identically distributed data have poor adaptability and significantly reduced diagnostic accuracy when faced with such complex cross-domain data distribution.

[0004] To solve the above problems, federated learning (Federated Learning) is introduced into the industrial field as a distributed machine learning paradigm that does not move the model. It allows the global model to be trained collaboratively by exchanging model parameters without the local data leaving the device, which theoretically solves the data privacy and island problems. However, there are still many defects in the direct application of the standard federated learning framework in complex industrial scenarios: first, its flat architecture has low communication efficiency when faced with thousands of heterogeneous devices, and is prone to network congestion; second, it lacks effective processing mechanisms for Non-IID data distribution, and the model generalization ability is insufficient; third, it does not design a reasonable contribution evaluation and incentive mechanism, and there is no economic motivation for enterprises to participate in collaboration; fourth, the privacy protection means is single, and there is still a risk of leaking raw data through reverse analysis of model parameters.

[0005] Therefore, there is an urgent need for a new generation of distributed intelligent diagnosis solution that can balance data privacy and cross-domain generalization. SUMMARY

[0006] To solve the above problems, the application provides an industrial equipment fault diagnosis method and system based on multi-level federated learning.

[0007] In a first aspect, the technical scheme of the application provides an industrial equipment fault diagnosis method based on multi-level federated learning, comprising the following steps: S1, a three-level federated learning network composed of a device end, a regional edge server and a central cloud platform is constructed; S2, at the device end, runtime sequence data of the industrial equipment is collected, the collected sequence data is used to train the fault diagnosis model currently deployed at the device end, and local model parameters after training are obtained; S3, the local model parameters are encrypted and desensitized, and uploaded to the corresponding regional edge server; S4, at the regional edge server, based on a dynamic weight distribution strategy, the local model parameters received from multiple device ends are weighted and aggregated to generate regional level model parameters; S5, the regional level model parameters of each region are uploaded to the central cloud platform; S6, at the central cloud platform, all regional level model parameters are integrated, and differential privacy noise is injected to generate a global fault diagnosis model; S7, the global fault diagnosis model is collaboratively optimized and compressed, and the compressed model is distributed to each device end to update the fault diagnosis model currently deployed at the device end for real-time fault diagnosis; S8, at the central cloud platform, based on the contribution data of each device end in the current round and historical federated learning, a contribution evaluation and incentive distribution process is performed, and the evaluation result is used to update the dynamic weight distribution strategy.

[0008] As a further limitation of the technical scheme of the application, the step S2 comprises: S21, collecting original runtime sequence data of the industrial equipment, and performing denoising, resampling and normalization preprocessing on the original runtime sequence data; S22, extracting time domain, frequency domain and time-frequency domain features for fault diagnosis from the preprocessed time sequence data to form a feature vector; S23, inputting the feature vector into the currently deployed fault diagnosis model for training to obtain local model parameters after training.

[0009] Through standardized preprocessing and feature engineering of the original time sequence data, the original, noisy and dimensionless industrial data are converted into high-quality and standardized feature vectors. This step improves the convergence speed and stability of subsequent model training, and improves the discrimination of fault features from the data source.

[0010] As a further limitation of the technical solution of the present invention, the fault diagnosis model in S2 is a dual-stream feature decoupling network; The dual-stream feature decoupling network extracts device-domain-independent general fault features through a shared encoder and device-specific features through a private encoder. The shared encoder and a domain classifier are trained adversarially through a gradient inversion layer to force the general fault features to be untraceable to a specific device. The general fault features and device-specific features are dynamically weighted and fused by an attention fusion module to form the final feature representation for fault diagnosis.

[0011] By employing a dual-stream feature decoupling network and optimizing it through adversarial training, the general fault features in equipment data can be actively separated from equipment-specific features. This solves the model generalization problem caused by the non-independent and identically distributed (Non-IID) nature of industrial equipment data, enabling the trained fault diagnosis model to maintain excellent diagnostic accuracy when facing equipment of unknown models or operating conditions, thus enhancing the system's cross-domain adaptability.

[0012] As a further limitation of the technical solution of the present invention, in S4, weighted aggregation based on a dynamic weight allocation strategy specifically includes: The regional edge server calculates the number of... Aggregate weight of individual devices :

[0013]

[0014]

[0015]

[0016] In the formula, For the first Data quality factor for each device For the first The signal-to-noise ratio of data from each device This represents the average signal-to-noise ratio of data from all devices. For the first Health rating of each device For the first The normalized value of the historical contribution of each device is the historical contribution factor. In each round of federal learning, the first The quantitative value of the contribution of each device, where N is the total number of devices participating in federated learning; Regional model parameters It is generated using the following weighted average formula:

[0017] wherein, is the regional level model parameter, is the dynamic weight of the i-th device, is the local model parameter of the i-th device, is the total number of devices participating in the regional aggregation. M

[0018] By introducing a dynamic weight distribution strategy determined by data quality, device health and historical contribution, the contribution of participants is finely and fairly evaluated. This strategy can effectively encourage participants to provide high-quality data, thereby guiding the federal learning alliance to evolve in a healthier and more efficient direction, and ultimately aggregating more reliable regional level model parameters to accelerate the convergence of the global model.

[0019] As a further limitation of the technical solution of the application, the injection of differential privacy noise in S6 is specifically: adding random noise satisfying ( )-differential privacy definition to the integrated model parameters.

[0020] Applying noise satisfying the differential privacy definition at the global model level provides strict and quantifiable privacy protection guarantees for the model. Even if an attacker obtains the final global model, they cannot infer any sensitive information about a single training sample, solving the concern of enterprises about privacy leakage after participating in joint modeling.

[0021] As a further limitation of the technical solution of the application, the process of injecting differential privacy noise is an adaptive process, including: calculating the average data quality of the current round of federal learning ; the average data quality is the average value of the data quality factors of all devices participating in aggregation; dynamically adjusting the privacy budget according to the average data quality : setting a baseline privacy budget , when is higher than the preset threshold, using a smaller reinforced privacy budget value than ; when is lower than the preset threshold, using a larger weakened privacy budget value than ; based on the dynamically adjusted value, injecting a corresponding amount of noise into the integrated model parameters.

[0022] ​​​By dynamically adjusting the privacy budget according to the average data quality, an intelligent balance between privacy protection strength and model utility is achieved. When the overall data quality is high, privacy protection is enhanced while the model performance loss is controllable; when the data quality is generally low, privacy is appropriately relaxed to ensure the basic performance of the model. This adaptive mechanism overcomes the rigid defects of fixed privacy budget, enabling the system to maintain the best state in different data environments.

[0023] As a further limitation of the technical solution of the application, the cooperative optimization compression in S7 specifically includes the following steps: S71, in the center cloud platform, the global fault diagnosis model is used as a teacher model; S72, for the hardware resource constraints of the target edge device, define the search space of model compression, the search space includes quantization bit width and channel pruning rate; S73, through an automatic search algorithm, find the optimal compression strategy in the search space that meets the preset performance target; S74, according to the optimal compression strategy, distill a lightweight student model from the teacher model; S75, distribute the student model to the corresponding device end to update the model currently deployed on the device end.

[0024] Through the cooperative optimization compression of automatic search and knowledge distillation, a high-precision lightweight diagnostic model that matches different edge devices of different computing power levels can be automatically generated. Once trained, it can be deployed anywhere, greatly reducing the deployment threshold and adaptation cost of the model on resource-constrained devices, while ensuring the high performance of the compressed model.

[0025] As a further limitation of the technical solution of the application, the loss function L used in S74 for distilling the student model is:

[0026] In the formula, is the cross-entropy loss between the student model output and the true label; is the KL divergence loss between the student model output and the teacher model output; is the mean square error loss between the intermediate layer feature map of the student model and the teacher model; and alpha and gamma are hyperparameters for balancing the weights of each loss term.

[0027] The multi-task loss function simultaneously constrains the final output, output distribution and internal attention mechanism of the student model, achieving deep imitation of the teacher model from inside to outside. The lightweight student model can maximize the retention of the reasoning ability and knowledge density of the teacher model at a very high compression ratio, thereby achieving the diagnostic accuracy of approximating the cloud-side large model on the edge.

[0028] As a further limitation of the technical solution of the application, S8 is specifically implemented through a blockchain smart contract, comprising: S81, uploading the contribution data of each device end in the current round of federated learning to the blockchain network; S82, automatically calculating the contribution degree of each device end in the current round based on the Shapley value algorithm through the smart contract on the blockchain; S83, the smart contract issues digital assets or service vouchers as incentives to the party to which the corresponding device end belongs according to the calculated contribution degree in the current round; S84, the smart contract updates the historical contribution factor of the device according to the contribution degree in the current round, and uses the updated factor in the dynamic weight distribution strategy of step S4 in the next round of federated learning for aggregate weight calculation.

[0029] The contribution evaluation and incentive allocation are implemented through a blockchain smart contract, building a decentralized, transparent, trustworthy, and automatically executed closed-loop system that fairly quantifies and rewards the contributions of each participant, solves the intrinsic motivation problem of their participation in collaboration, establishes a solid trust foundation through unalterable records on the chain, and guarantees the long-term vitality and stability of the federated learning alliance.

[0030] In a second aspect, the technical solution of the application also provides an industrial equipment fault diagnosis system based on multi-level federated learning for implementing the method of the first aspect, comprising: A device end module is deployed in an industrial site and is configured to: Collect raw runtime sequence data of industrial equipment; perform denoising, resampling, and normalization preprocessing on the raw data; extract time domain, frequency domain, and time-frequency domain features from the preprocessed time sequence data to form a feature vector; train a locally deployed fault diagnosis model using the feature vector to generate local model parameters; upload the local model parameters after encryption and desensitization processing; and perform real-time fault diagnosis using the updated model; A regional edge server module is deployed at the plant or workshop level and is configured to: Receive and decrypt the local model parameters uploaded by each device end module in the jurisdiction; based on a dynamic weight distribution strategy, weight and aggregate the received local model parameters to generate regional-level model parameters; upload the regional-level model parameters to the central cloud platform; A central cloud platform module is configured to: receive and integrate regional level model parameters from each regional edge server module; inject differential privacy noise into the integrated model parameters to generate a new generation of global fault diagnosis model; perform collaborative optimization compression on the global fault diagnosis model; distribute the compressed model to each device end module to update its local model; based on the contribution data of each device end module in the current round and historical federated learning, perform contribution evaluation and incentive allocation process, and use the evaluation results to update the dynamic weight allocation strategy.

[0031] The responsibilities and cooperative relationship of the device end, edge side and cloud end are clarified, ensuring that the entire multi-level federated learning based fault diagnosis method can be efficiently and stably operated in actual industrial environment.

[0032] As a further limitation of the technical solutions of the application, a lightweight fault diagnosis model is integrated in the device end module, which is a double-flow feature decoupling network. A dynamic weight calculation unit is integrated in the regional edge server module for executing the dynamic weight allocation strategy. A meta-learner, a differential privacy injection unit, a model collaborative compression unit and a blockchain smart contract interaction interface are integrated in the center cloud platform module.

[0033] As a further limitation of the technical solutions of the application, the device end module is deployed on a programmable logic controller (PLC) or an industrial gateway; the regional edge server module is deployed on a workshop level industrial server; and the center cloud platform module is deployed on a cloud server cluster.

[0034] As can be seen from the above technical solutions, the present application has the following advantages: it realizes privacy-safe collaborative training of "data not moving and model moving", breaks down the data barriers between enterprises; reduces network communication load through hierarchical architecture, improves system scalability; and finally realizes millisecond-level real-time fault diagnosis by deploying lightweight models on the device end, meeting the rapid response needs of industrial high-risk scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the present application, the drawings needed in the description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0036] Figure 1 The flowchart of the method provided by the embodiment of the present application is shown.

[0037] Figure 2 The block diagram of the system provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0038] In order to make the application purposes, features and advantages of the present application more obvious and easy to understand, the technical solutions protected by the present application will be described clearly and completely below by using specific examples and drawings. Obviously, the examples described below are only some of the examples of the present application, but not all. Based on the examples in the present application, all other examples obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0039] Unless otherwise defined, all technical and scientific terms used in the present application have the same meanings as commonly understood by those skilled in the art to which the present application belongs. The terms used in the specification of the present application are only for the purpose of describing the specific examples, and are not intended to limit the present application.

[0040] As shown in Figure 1 The embodiment of the present application provides an industrial equipment fault diagnosis method based on multi-federated learning, which comprises the following steps: S1, a three-level federated learning network composed of a device end, a regional edge server and a central cloud platform is constructed; The specific implementation is as follows: The device end is composed of sensors (such as vibration sensors, temperature sensors) and embedded devices (such as programmable logic controllers (PLCs), industrial gateways, or edge computing modules) deployed in industrial sites with computing capabilities. The device is pre-installed with a lightweight fault diagnosis model of the present application and a local training program, and is configured with a network interface for communication with the upper layer.

[0041] The regional edge server is composed of computing nodes (such as industrial servers, high-performance industrial computers) deployed at the plant workshop or plant level. The server is deployed with a model aggregation module and a dynamic weight calculation unit, and is responsible for communication with all device ends in the jurisdiction through a factory local area network or an industrial bus.

[0042] The central cloud platform is composed of a cloud server cluster located in a data center, and is deployed with a global model management module, a meta-learner, a differential privacy engine and a blockchain smart contract interaction interface. The central cloud platform establishes a secure communication connection with each regional edge server through a wide area network (such as the Internet / VPN dedicated line).

[0043] By configuring the above hardware entities and deploying corresponding software modules thereon, the three-level federated learning network architecture is established, so that data, model parameters and instructions can flow between the device end, the regional edge server and the central cloud platform in a secure and orderly manner according to the preset rules.

[0044] S2, collecting runtime sequence data of the industrial equipment at the device end, training a fault diagnosis model currently deployed at the device end using the collected sequence data to obtain local model parameters after training; S3, encrypting and desensitizing the local model parameters and uploading them to a corresponding regional edge server; S4, at the regional edge server, weighting and aggregating the local model parameters received from multiple device ends based on a dynamic weight distribution strategy to generate regional-level model parameters; S5, uploading the regional-level model parameters of each region to the central cloud platform; S6, at the central cloud platform, integrating all regional-level model parameters and injecting differential privacy noise to generate a global fault diagnosis model; S7, collaboratively optimizing and compressing the global fault diagnosis model, and distributing the compressed model to each device end to update the fault diagnosis model currently deployed at the device end for real-time fault diagnosis; S8, at the central cloud platform, performing contribution evaluation and incentive distribution processes based on the contribution data of each device end in the current round and historical federated learning, and using the evaluation results to update the dynamic weight distribution strategy.

[0045] In some embodiments, step S2 includes: S21, collecting original runtime sequence data of the industrial equipment and performing denoising, resampling, and normalization preprocessing on the original data; S22, extracting time domain, frequency domain, and time-frequency domain features for fault diagnosis from the preprocessed sequence data to form a feature vector; S23, inputting the feature vector into the currently deployed fault diagnosis model for training to obtain local model parameters after training. The specific training process of this step is as follows: 1. Forward propagation Input: The feature vector obtained from step S22 is simultaneously input into two parts of the dual-flow feature decoupling network: a shared encoder (Φs) and a private encoder (Φp).

[0046] Feature extraction: The shared encoder processes the input and outputs a general fault feature (Zcommon) that is independent of the device domain. This feature aims to capture common fault patterns (such as typical vibration features of bearing wear) for all devices.

[0047] The private encoder processes the same input and outputs a device-specific feature (Zspecific). This feature captures individualized patterns specific to the current device (such as unique signals caused by installation gaps or minor wear).

[0048] Feature fusion and classification: Zcommon and Zspecific are fed into the attention fusion module. This module learns a set of dynamic weights, according to the formula The fused feature representation (Zfinal) for final diagnosis is generated.

[0049] Zfinal is fed into the fault classifier (usually a fully connected layer + Softmax), which outputs a predicted probability distribution over different fault types, thus computing the fault classification loss (Ltask), e.g. cross-entropy loss.

[0050] 2. Adversarial training (for feature decoupling) Zcommon is also fed into a domain classifier.

[0051] The goal of the domain classifier is to accurately tell from which device these features come from.

[0052] During backpropagation, GRL negates the gradient from the domain classifier before passing it back to the shared encoder. The shared encoder is thus incentivized to generate features that "confuse" the domain classifier, thus forcing the common features to gradually decouple from the device identity.

[0053] 3. Loss computation and backpropagation The total loss for local training is composed of multiple task losses weighted together:

[0054] Ltask (fault classification loss): ensures the core diagnostic function of the model is accurate.

[0055] Ldomain (domain adversarial loss): drives the decoupling of common features from device identity. The domain classifier tries to minimize it, while the shared encoder (via GRL) tries to maximize it, a minimax game.

[0056] Lrecon (feature reconstruction loss): ensures private features do not learn common information. For example, one can require that the original input can be reconstructed from the private features, thus ensuring they carry all information beyond the common features.

[0057] The device end utilizes the locally stored data to minimize this total loss function Ltotal via stochastic gradient descent (SGD) or its variants (e.g. Adam). Through multiple iterations, the model parameters (including the weights and biases of the shared encoder, private encoder, attention module, and classifiers) are continuously updated.

[0058] 4. Output After training is complete, these updated model parameters are the trained local model parameters output in step S2. These parameters encapsulate the knowledge learned by the device in this round of training and will then be encrypted and uploaded.

[0059] The fault diagnosis model in S2 is a dual-stream feature decoupling network; The dual-stream feature decoupling network extracts device-domain-independent general fault features through a shared encoder and device-specific features through a private encoder. The shared encoder and a domain classifier are trained adversarially through a gradient inversion layer to force the general fault features to be untraceable to a specific device. The general fault features and device-specific features are dynamically weighted and fused by an attention fusion module to form the final feature representation for fault diagnosis.

[0060] In some embodiments, the encryption and desensitization processing of local model parameters described in S3 includes: The local model parameters are then subjected to privacy-enhancing desensitization processing, the desensitized parameters are then encrypted, and the encrypted data is then uploaded to the corresponding regional edge server. Among them, the privacy-enhancing desensitization process involves applying differential privacy noise and encrypting using a homomorphic encryption method.

[0061] In some embodiments, S4 involves weighted aggregation based on a dynamic weight allocation strategy, specifically including: The regional edge server calculates the first step according to the following formula. Aggregate weight of individual devices :

[0062]

[0063]

[0064]

[0065] In the formula, For the first Data quality factor for each device For the first The signal-to-noise ratio of a device's data indicates the signal-to-noise ratio of the preprocessed signal; a higher value indicates better data quality. This represents the average signal-to-noise ratio of data from all devices. For the first The health score of each device is based on the normalized value of the device's continuous fault-free operation time, with a value range of [0, 1]. For the first The normalized value of the historical contribution degree of the device, i.e., the historical contribution factor, is calculated based on the Shapley value algorithm. is the quantified value of the contribution degree of the device in each round of federated learning, and N is the total number of devices participating in federated learning. is the quantified value of the contribution degree of the device in each round of federated learning, and N is the total number of devices participating in federated learning. is the quantified value of the contribution degree of the device in each round of federated learning, and N is the total number of devices participating in federated learning. is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps:

[0066] is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps:

[0067] is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps:

[0068] is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: M is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps:

[0069] is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps:

[0070] is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: is the cumulative contribution degree of the device in multiple rounds of federated learning, which can be calculated by the following steps: Set a baseline privacy budget ,when When the value exceeds a preset threshold, use a ratio Small enhanced privacy budget value; when When the value is below a preset threshold, use a ratio Larger privacy budget values; Based on the dynamically adjusted This value injects a corresponding amount of noise into the integrated model parameters. The specific implementation is as follows: First, the sensitivity of the regional model parameters is calibrated by setting the clipping norm of the model parameters to determine the global sensitivity Δ. Subsequently, based on the dynamically adjusted privacy budget Calculate the distribution parameters of the noise to be added, based on the global sensitivity Δ. Finally, noise is sampled from a Laplace or Gaussian distribution with a mean of zero and scaled by the aforementioned distribution parameters, and this noise is added to the integrated model parameters.

[0071] In federated learning, model parameters (especially gradients) are typically pruned either locally on the client or on the server. For example, all client-side update gradients may be pruned to a fixed L2 norm C.

[0072] For aggregation operations that calculate the average, the global sensitivity Δ is defined as Δ = 2C / Nn, where C is the clipping norm and Nn is the number of clients.

[0073] In some embodiments, the collaborative optimization compression in S7 specifically includes the following steps: S71. On the central cloud platform, the global fault diagnosis model is used as the teacher model; S72. Based on the hardware resource constraints of the target edge device, define a search space for model compression, wherein the search space includes quantization bit width and channel pruning rate; S73. Using an automated search algorithm, find the optimal compression strategy that meets the preset performance target in the search space; S74. Based on the optimal compression strategy, a lightweight student model is generated from the teacher model by distillation. S75. Distribute the student model to the corresponding device to update the model currently deployed on the device.

[0074] In some embodiments, the student model is generated by distillation in S74, and the loss function L used for training is:

[0075] In the formula, Output the cross-entropy loss between the student model and the true label; KL divergence loss between student model output and teacher model output; mean square error loss between intermediate layer feature maps of student model and teacher model; and a and g are hyperparameters balancing the weights of each loss term.

[0076] In some embodiments, S8 is implemented specifically through a blockchain smart contract, including: S81, upload the contribution data of each device end in this round of federated learning to the blockchain network; this step specifically includes: Data preparation: the central cloud platform collates and summarizes the contribution data of each device end in this round of learning. These data are necessary for subsequent calculation of Shapley value, and usually include: Hash value of local model parameters (as a unique proof of participation and contribution carrier of this device).

[0077] Dynamic weight of this device in regional aggregation.

[0078] Data quality factor, health score, and other original indicators of this device for calculating weight.

[0079] Performance improvement indicators of this round of global model (e.g. accuracy improvement on validation set, loss function reduction value, etc.).

[0080] Data on-chain: the central cloud platform uploads the structured summary or hash value of the above data to the blockchain network through a transaction. Here, a "on-chain-off-chain" hybrid storage mode is usually used to reduce costs: only key indexes and hash values are stored on-chain to ensure non-tamperability, while complete, larger-volume contribution data are stored in the central cloud or distributed storage (such as IPFS) off-chain.

[0081] S82, automatically calculate the contribution of each device end in this round based on the Shapley value algorithm through the smart contract on the blockchain; specifically including: Contract triggering: after the data on-chain transaction is confirmed, the pre-deployed smart contract on the blockchain will be triggered.

[0082] Execute Shapley value calculation: The efficient approximate calculation algorithm of Shapley value coded in the smart contract is a random algorithm based on Monte Carlo sampling, or a simplified algorithm based on device weight ordering.

[0083] The execution process of the contract is: a. Construct device subsets: the algorithm simulates different combinations (subsets) of device participation.

[0084] b. Evaluate Model Utility: For each subset, the contract needs to evaluate the performance of the model it can train. Since the model cannot be retrained on-chain, a proxy metric will be used, such as the performance of the weighted aggregated model parameters of all devices in the subset on a public benchmark dataset. This benchmark dataset and the evaluation logic need to be predefined in the contract.

[0085] c. Calculate marginal contribution: Calculate the increase in model utility (i.e., marginal contribution) when a device is added to a subset.

[0086] d. Average marginal contribution: Iterate through all possible subset combinations and average all marginal contributions of the device to obtain the device's Shapley value in this round, which is its fair contribution metric.

[0087] S83. The smart contract, based on the calculated contribution level in this round, issues digital assets or service certificates as incentives to the owner of the corresponding device; specifically including: The smart contract maintains a registry that binds the device ID to the blockchain wallet address of its parent company.

[0088] As an alternative or supplement, the contract can issue a non-fungible token (NFT) to the enterprise address as a service credential. This NFT represents the right to claim specific services (such as additional cloud storage, a deep diagnostic report, computing power packages, etc.), which the enterprise can redeem and use in DApps within the alliance ecosystem.

[0089] S84. The smart contract updates the historical contribution factor of the device based on the contribution of this round, and uses the updated factor in the dynamic weight allocation strategy of step S4 in the next round of federated learning to calculate the aggregate weight.

[0090] The smart contract maintains a state variable—the historical contribution factor—for each device on the blockchain.

[0091] The contract uses a smooth iterative formula to update the factor. After the update, when the regional edge server executes S4 (dynamic weight calculation) in the next round of federated learning, it can query the latest historical contribution factor value of each device by calling the blockchain's read-only function. This updated historical contribution factor retrieved from the chain is substituted into the dynamic weight calculation formula, directly affecting the device's weight in the next round of aggregation. Devices with higher contributions will have greater influence and higher returns in the future, thus forming a self-reinforcing positive cycle.

[0092] like Figure 2 As shown, this embodiment of the invention also provides an industrial equipment fault diagnosis system based on multi-level federated learning for implementing the method described in the above embodiments, comprising: A device end module deployed in an industrial site, configured to: collect raw runtime data of industrial equipment; pre-process the raw data by denoising, resampling and normalization; extract time-domain, frequency-domain and time-frequency domain features from the pre-processed time series data to form a feature vector; train a locally deployed fault diagnosis model using the feature vector to generate local model parameters; upload the local model parameters after encryption and desensitization processing; perform real-time fault diagnosis using the updated model; A regional edge server module deployed at the plant or workshop level, configured to: receive and decrypt the local model parameters uploaded by each device end module in the jurisdiction; based on a dynamic weight allocation strategy, aggregate the received local model parameters by weighting to generate regional-level model parameters; upload the regional-level model parameters to the central cloud platform; A central cloud platform module, configured to: receive and integrate regional-level model parameters from each regional edge server module; inject differential privacy noise into the integrated model parameters to generate a new generation of global fault diagnosis model; perform collaborative optimization and compression of the global fault diagnosis model; distribute the compressed model to each device end module to update its local model; based on the contribution data of each device end module in the current and historical federated learning, perform contribution evaluation and incentive allocation process, and use the evaluation results to update the dynamic weight allocation strategy.

[0093] The responsibilities and collaboration relationship of the device end, edge side and cloud end are clarified, ensuring that the entire multi-level federated learning-based fault diagnosis method can run efficiently and stably in the actual industrial environment.

[0094] It should be noted that the device end module integrates a lightweight fault diagnosis model, which is a dual-flow feature decoupling network; The regional edge server module integrates a dynamic weight calculation unit for executing the dynamic weight allocation strategy; The central cloud platform module integrates a meta-learner, a differential privacy injection unit, a model collaborative compression unit, and a blockchain smart contract interaction interface.

[0095] The dual-flow feature decoupling network consists of: (1) Shared encoder Φs: a multi-layer one-dimensional convolutional neural network (1D-CNN) for extracting shared, device-independent general fault features (such as common bearing wear and rotor imbalance vibration patterns) from the raw time series data of all devices; (2) Private encoder Φp: a lightweight gated recurrent unit (GRU) network to capture unique features specific to a certain device or type (e.g., minor variations due to installation gap, load difference); (3) Domain classifier D: connected after the shared encoder and trained adversarially via a gradient reversal layer (GRL). Its goal is to distinguish the domain of the features as much as possible, while the shared encoder aims to generate features that are indistinguishable to the domain classifier, thus forcing the general features to "de-identify"; (4) Attention fusion module A: receives the general and unique features and automatically learns a set of weight vectors to dynamically fuse them according to the formula , where the weight α is adaptively calculated from the content of the features themselves; Training process: the total loss function is: ; where Lrecon is the reconstruction loss term to ensure that the private features do not contain general information. Joint optimization is achieved through multi-task learning.

[0096] The device-side module is deployed on a programmable logic controller (PLC) or an industrial gateway; the regional edge server module is deployed on a workshop-level industrial server; and the central cloud platform module is deployed on a cloud server cluster.

[0097] The self-evolution diagnosis engine and the collaborative compression of the present application include the following specific implementation processes: Diagnosis self-evolution process: (1) New fault detection: while outputting the fault probability, the device-side model also calculates an uncertainty score. When the score exceeds a threshold and the confidence of the highest probability class is low, it is determined to be a potential new fault; (2) Trigger incremental learning: once a new fault is detected, the sample (and a subsequent time sequence) is labeled as "to be confirmed" and encrypted and uploaded to the regional edge server; (3) Artificial confirmation and memory bank update: the edge server pushes the alarm to the operation and maintenance personnel, and after confirmation, a new labeled sample is formed and stored in a fixed-size circular replay memory; (4) Prevention of forgetting: the regional model periodically samples old data from the replay memory and fine-tunes it with new data, effectively alleviating catastrophic forgetting. Experiments show that after introducing 10 new fault modes, the model's recognition accuracy for the original faults only decreases by <2%; Collaborative compression technology: (1) Search space definition: for edge devices, the search space includes: quantization bit width (8 / 16 bits), convolution layer channel pruning rate, whether to use depth separable convolution, etc. (2) Automatic compression: using reinforcement learning strategy of neural architecture search (NAS) to search the optimal compression configuration for different hardware configurations automatically with accuracy loss and model size as reward function; (3) Distillation reinforcement: using the global model optimized by element learning as a teacher model to guide the student model training.

[0098] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-level federated learning-based industrial equipment fault diagnosis method, characterized in that, The method comprises the following steps: S1, a three-level federated learning network composed of a device end, a regional edge server and a central cloud platform is constructed; S2, at the device end, runtime sequence data of the industrial equipment is collected, the collected sequence data is used to train the fault diagnosis model currently deployed at the device end, and local model parameters after training are obtained; S3, the local model parameters are encrypted and desensitized, and are uploaded to the corresponding regional edge server; S4, at the regional edge server, the local model parameters of multiple device ends received are weighted and aggregated based on a dynamic weight distribution strategy, and regional-level model parameters are generated; S5, the regional-level model parameters of each region are uploaded to the central cloud platform; S6, at the central cloud platform, all regional-level model parameters are integrated, and differential privacy noise is injected to generate a global fault diagnosis model; S7, the global fault diagnosis model is collaboratively optimized and compressed, and the compressed model is distributed to each device end to update the fault diagnosis model currently deployed at the device end for real-time fault diagnosis; S8, at the central cloud platform, based on the contribution data of each device end in the current round and historical federated learning, a contribution evaluation and incentive distribution process is performed, and the evaluation result is used to update the dynamic weight distribution strategy.

2. The industrial equipment fault diagnosis method based on multi-federated learning according to claim 1, characterized in that, The step S2 comprises: S21, original runtime sequence data of the industrial equipment is collected, and the original runtime sequence data is preprocessed by denoising, resampling and normalization; S22, time domain, frequency domain and time-frequency domain features for fault diagnosis are extracted from the preprocessed sequence data to form a feature vector; S23, the feature vector is input into the currently deployed fault diagnosis model for training to obtain local model parameters after training.

3. The industrial equipment fault diagnosis method based on multi-federated learning according to claim 2, characterized in that, The fault diagnosis model in S2 is a dual-flow feature decoupling network; The dual-flow feature decoupling network extracts general fault features irrelevant to the device domain through a shared encoder, and extracts device-specific features through a private encoder; the shared encoder and a domain classifier are trained through a gradient reversal layer to force the general fault features to be unable to be traced back to specific devices; the general fault features and the device-specific features are dynamically weighted and fused by an attention fusion module to form the final feature representation for fault diagnosis.

4. The industrial equipment fault diagnosis method based on multi-federated learning according to claim 3, characterized in that, In S4, the weighted aggregation based on the dynamic weight distribution strategy specifically comprises: The regional edge server calculates the aggregate weight of the nth device according to the following formula :​ In the formula, is the data quality factor of the th device, is the data signal-to-noise ratio of the th device, is the average value of the data signal-to-noise ratios of all devices, is the health score of the th device, is the normalized value of the historical contribution degree of the th device, that is, a historical contribution factor, is the quantitative value of the contribution degree of the th device in each round of federated learning, and N is the total number of devices participating in federated learning. Region-level model parameters Generated by the following weighted average equation: wherein, is a zone-level model parameter, is a dynamic weight for the th device, is a local model parameter for the th device, M is the total number of devices participating in the zone aggregation.

5. The industrial equipment fault diagnosis method based on multi-federated learning according to claim 4, characterized in that, The injection differential privacy noise in S6 is specifically: adding random noise satisfying the differential privacy definition to the integrated model parameters. ) 6. The industrial equipment fault diagnosis method based on multi-federated learning according to claim 5, characterized in that, The process of injecting differential privacy noise is an adaptive process, comprising: computing an average data quality for the round of federated learning ; the average data quality is an average of data quality factors for all participating aggregated devices ; According to the average data quality Dynamic adjustment of privacy budget : Set a baseline privacy budget ,when When the value exceeds a preset threshold, use a ratio Small enhanced privacy budget value; when When the value is below a preset threshold, use a ratio Larger privacy budget values; based on the dynamically adjusted values, injects a corresponding amount of noise into the integrated model parameters.

7. The industrial equipment fault diagnosis method based on multi-federated learning according to claim 6, characterized in that, The collaborative optimization and compression in S7 specifically comprises the following steps: S71, at the central cloud platform, the global fault diagnosis model is taken as a teacher model; S72, a search space for model compression is defined according to the hardware resource constraints of the target edge device, and the search space includes quantization bit width and channel pruning rate; S73, an optimal compression strategy meeting a preset performance target is found in the search space through an automatic search algorithm; S74, a lightweight student model is distilled from the teacher model according to the optimal compression strategy; S75, the student model is distributed to the corresponding device end to update the model currently deployed at the device end.

8. The industrial equipment fault diagnosis method based on multi-federated learning according to claim 7, characterized in that, The student model is generated by distillation in S74, and the loss function L used in training is: wherein is the cross-entropy loss between the student model output and the true label; is the KL divergence loss between the student model output and the teacher model output; is the mean square error loss between the intermediate layer feature maps of the student model and the teacher model; and α and γ are hyperparameters balancing the weights of the loss terms.

9. The industrial equipment fault diagnosis method based on multi-federated learning according to claim 8, characterized in that, S8 is specifically implemented through a blockchain smart contract, including: S81, upload the contribution data of each device end in this round of federated learning to the blockchain network; S82, automatically calculate the contribution degree of each device end in this round based on the Shapley value algorithm through the smart contract on the blockchain; S83, the smart contract issues digital assets or service vouchers as incentives to the corresponding device end according to the calculated contribution degree in this round; S84, the smart contract updates the historical contribution factor of the device according to the contribution degree in this round, and uses the updated factor in the dynamic weight distribution strategy of step S4 in the next round of federated learning for aggregate weight calculation.

10. A multi-federated learning based industrial equipment fault diagnosis system for implementing the method of any one of claims 1 to 9, characterized in that, Including: The device end module is deployed in the industrial field and is configured to: Collect the original runtime sequence data of the industrial equipment; denoise, resample and normalize the original data; extract time domain, frequency domain and time-frequency domain features from the preprocessed time series data to form a feature vector; Use the feature vector to train the locally deployed fault diagnosis model to generate local model parameters; upload the encrypted and desensitized local model parameters; use the updated model for real-time fault diagnosis; The regional edge server module is deployed at the plant or workshop level and is configured to: Receive and decrypt the local model parameters uploaded by each device end module in the jurisdiction; based on the dynamic weight distribution strategy, weight the received local model parameters to generate regional-level model parameters; upload the regional-level model parameters to the central cloud platform; The central cloud platform module is configured to: Receive and integrate regional-level model parameters from each regional edge server module; inject differential privacy noise into the integrated model parameters to generate a new generation of global fault diagnosis model; collaboratively optimize and compress the global fault diagnosis model; distribute the compressed model to each device end module to update its local model; based on the contribution data of each device end module in this round and historical federated learning, execute the contribution evaluation and incentive distribution process, and use the evaluation results to update the dynamic weight distribution strategy.

Images