Equipment fault diagnosis method and system based on cloud edge cooperation

Abstract

The invention discloses an equipment fault diagnosis method and system based on cloud edge cooperation, and relates to the field of complex equipment fault diagnosis. According to the method, global model centralized training is realized at a cloud end, lightweight model optimization is realized at an edge end, and a periodic closed-loop system of cloud end retraining-knowledge issuing-edge updating-data returning is constructed, so that continuous optimization and system-level intelligent collaboration of the model are realized.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis for complex equipment, and specifically to a method and system for fault diagnosis of equipment based on cloud-edge collaboration. Background Technology

[0002] The statements in this section are provided only as background information in connection with this disclosure and may not constitute prior art.

[0003] With the continuous development of industrial automation, aerospace equipment, and intelligent manufacturing technologies, modern complex systems are gradually evolving towards higher integration, higher reliability, and multi-device collaborative control. These systems typically contain multiple heterogeneous sub-devices and distributed control units, often operating for extended periods in complex environments such as high temperature, high pressure, and electromagnetic interference. Their operating conditions are highly variable and susceptible to external disturbances. Under these extreme conditions, equipment is prone to problems such as material aging, structural fatigue, and sensor malfunction, leading to system performance degradation or even functional failure. Failure to identify and address these issues promptly can trigger a chain reaction, causing serious safety and economic losses.

[0004] Therefore, achieving accurate and reliable equipment fault diagnosis is one of the core technical challenges in the field of intelligent operation and maintenance in industry and aerospace. Existing equipment fault diagnosis methods are mainly divided into two categories: The first category is traditional fault diagnosis methods based on signal processing and pattern recognition. These methods typically extract time-domain, frequency-domain, or time-frequency-domain features from vibration, current, and temperature signals, combining this with feature statistical analysis and shallow learning algorithms such as principal component analysis, support vector machines, and Gaussian mixture models to establish normal and fault modes of the equipment. When the monitored signal differs significantly from the normal mode, the corresponding fault type can be determined. This type of method has advantages such as simple implementation and strong interpretability, but it relies too heavily on manual feature design and struggles to handle complex nonlinear dynamic systems and feature variations under multiple operating conditions.

[0005] The second category comprises intelligent fault diagnosis methods based on deep learning. These methods utilize models such as convolutional neural networks, recurrent neural networks, autoencoders, time-frequency convolutional networks, or Transformers to automatically extract multi-level feature representations from the original signal and identify the fault type through a classifier. Compared to traditional methods, deep learning models can effectively learn the implicit structure and temporal relationships of complex signals, improving diagnostic accuracy. However, their performance depends on sufficient labeled samples and a stable data distribution. When the number of samples is limited or operating conditions change, the model's generalization performance and cross-device adaptability significantly decrease.

[0006] In real-world industrial and aerospace systems, equipment is often distributed across multiple heterogeneous nodes, such as production lines, propulsion systems, and satellite attitude control units, resulting in significantly different data acquisition environments. Due to mission constraints and data privacy protection, data collected by different nodes typically cannot be directly shared, creating data silos. To address the challenges of expanding industrial system scale, widespread equipment distribution, and increased data heterogeneity, recent patent research has explored cloud-edge collaborative intelligent operation and maintenance architectures. This architecture balances system computational efficiency with model performance by deploying high-performance computing and centralized models in the cloud and implementing state awareness and local inference at the edge.

[0007] Current cloud-edge collaborative fault diagnosis systems typically consist of three layers: a device perception layer responsible for multi-source signal acquisition and preliminary feature extraction; an edge computing layer performing local feature analysis and rapid judgment; and a cloud computing center executing model training, knowledge fusion, and global optimization. This architecture enables distributed monitoring and fault diagnosis of device operating status, but it still faces several challenges in practical applications, as follows: (1) Data security and privacy issues: In distributed industrial and aerospace systems, the data collected by each device node is limited by privacy and security requirements, and the raw data of different nodes are difficult to share directly.

[0008] (2) Insufficient knowledge transfer mechanism: Due to the limited computing resources and storage space of edge devices, diagnostic models deployed at the edge are usually lightweight models, while large-scale models with high complexity are deployed in the cloud. Due to the different structures of cloud and edge models, the model parameters in the cloud cannot be directly distributed or reused at the edge. Existing parameter compression or model pruning methods often fail to effectively transfer the discriminative knowledge in the cloud model to the edge model, resulting in low diagnostic performance of the edge under complex working conditions.

[0009] In summary, while existing cloud-edge collaborative fault diagnosis technologies have made some progress in achieving distributed device monitoring, they still have shortcomings in system collaboration and knowledge transfer mechanisms. Therefore, there is an urgent need for a cloud-edge collaborative fault diagnosis method that can effectively utilize data and efficiently transfer knowledge between high-performance cloud models and lightweight edge models while ensuring data security, thereby improving the overall intelligence level and operational reliability of the system. Summary of the Invention

[0010] The purpose of this invention is to provide a cloud-edge collaborative equipment fault diagnosis method and system to address the problems existing in current cloud-edge collaborative fault diagnosis systems for industrial and aerospace equipment, such as difficulty in sharing original data, limited model knowledge transfer methods, and unstable edge diagnosis performance.

[0011] Specifically, the present invention mainly solves the following technical problems: (1) The system model is not updated in a timely manner: the existing cloud-edge collaborative system mostly operates by static model distribution or manual triggering of updates. The models in the cloud and the edge are not synchronized for a long time, and there is a lack of periodic retraining, knowledge feedback and model version management mechanism, which leads to the degradation of diagnostic performance over time.

[0012] (2) Difficulties in multi-node data fusion and unified modeling: In scenarios with multiple devices and multiple operating conditions, the data types, frequencies and formats uploaded by each node are quite different, making it difficult to achieve efficient aggregation and unified modeling in the cloud. The model update cycle is long and the fusion efficiency is low.

[0013] (3) Problem of a single knowledge transfer mechanism: Existing systems mostly rely on direct parameter synchronization or model file replacement, which makes it difficult to achieve dynamic migration and collaborative optimization of knowledge at the system level, resulting in edge models being unable to effectively inherit the discrimination capabilities of the cloud.

[0014] Therefore, this invention proposes a centralized modeling and dynamic knowledge update system architecture based on a cloud-edge collaborative mechanism. It realizes centralized training of the global model in the cloud, lightweight model optimization at the edge, and constructs a closed-loop system of periodic "cloud retraining - knowledge distribution - edge update - data feedback" to achieve continuous model optimization and system-level intelligent collaboration.

[0015] The technical solution of the present invention is as follows: A device fault diagnosis method based on cloud-edge collaboration includes: Step S1: The edge node collects the device's operating data and uploads it to the cloud; Step S2: The cloud integrates the operational data uploaded by each node for centralized model training to obtain a cloud model with global discrimination capabilities; the cloud model is used to generate knowledge information about the device's operating status. Step S3: The cloud sends the knowledge information to the edge nodes, and the edge nodes optimize and train the edge model based on the knowledge according to the local data, so as to realize the transfer of cloud model knowledge to the edge model; Step S4: The optimized edge model performs fault diagnosis on the equipment site and sends the new operating data back to the cloud to form a cloud-edge collaborative closed loop of model retraining, knowledge distribution and edge update.

[0016] This invention also proposes a cloud-edge collaborative device fault diagnosis system to implement the above-mentioned device fault diagnosis method, comprising: The cloud-based model training module is used to receive runtime data uploaded from multiple edge nodes, fuse the runtime data to construct a joint dataset, and perform centralized model training to obtain a cloud-based model with global discriminative capabilities; the cloud-based model is used to generate knowledge information about the device's operating status. The edge model training module is used to optimize and train the edge model at the edge node based on locally collected data and under the guidance of knowledge information generated by the cloud model, so as to realize the transfer of knowledge from the cloud model to the edge model. The cloud-edge data interaction and knowledge update module is used to upload runtime data, update cloud model training, and distribute cloud model knowledge between edge nodes and the cloud, so as to support a cloud-edge collaborative closed loop of cloud retraining, knowledge distribution, and edge updates.

[0017] Furthermore, the device operation signals collected by each edge node are aggregated in the cloud to form a global training set.

[0018] Furthermore, the training of the cloud-based model is as follows: After normalizing, denoising, and performing time-series segmentation preprocessing on the running signals uploaded by each edge node, a deep learning model is used to extract multi-layer time-frequency features and perform centralized training to establish a cloud model with global discrimination capabilities.

[0019] Furthermore, during the cloud-based model training phase, the cloud uses cross-entropy loss as the optimization objective:

[0020] in, For the first The true class label of each sample The knowledge information output by the model; the model parameters are optimized using this loss function. The trained cloud model is obtained. .

[0021] Furthermore, edge models are deployed on each device node to enable real-time fault diagnosis and adaptive model optimization locally.

[0022] Furthermore, the edge model structure is a lightweight neural network; the edge model training process is conducted under the guidance of a cloud model, achieving knowledge transfer and capability inheritance through knowledge distillation; the overall loss function of the edge model is as follows:

[0023] in, For local cross-entropy loss, For knowledge distillation loss, This is the weighting balance coefficient. This is the distillation temperature coefficient.

[0024] Furthermore, the local cross-entropy loss includes:

[0025] in, For the sample The true label, The knowledge information output by the edge model; The knowledge distillation loss includes:

[0026] in, and Cloud models and edge models are categorized separately. The output logits distribution on the [database name].

[0027] Furthermore, during cloud-based retraining, the training objective function is cross-entropy loss: .

[0028] Furthermore, during the edge update process, the edge model uses a joint loss function as the optimization objective, which is as follows:

[0029] Edge models optimize parameters and absorb knowledge by minimizing the joint loss function, enabling their output to gradually approach that of cloud models in terms of knowledge information.

[0030] Compared with existing cloud-edge collaborative fault diagnosis methods based on parameter synchronization or model distribution, this invention, by introducing a knowledge distillation mechanism and a periodic cloud-edge collaborative update strategy, has the following beneficial effects: 1. Improved fault diagnosis performance of edge model: The cloud model is trained with data uploaded by each node to obtain a cloud model with global discrimination capability; the edge model is distilled with the probability distribution output by the cloud to align the decision boundary and confidence, so as to maintain fault identification accuracy close to that of the cloud under lightweight conditions.

[0031] 2. Continuous Optimization Capability of Cloud-Edge Collaborative System: Through periodic cloud-edge data interaction and knowledge update mechanisms, the system can continuously optimize model parameters and discrimination capabilities based on the latest operational data. This enables the system to continuously evolve, thus maintaining the effectiveness of the diagnostic model over the long term. Compared with existing static parameter synchronization methods, this invention can achieve adaptive model updates when equipment status or operating conditions change, improving the maintainability and long-term stability of the system.

[0032] 3. A Balanced Approach Between Modeling and Edge Inference: At the system level, a balance is achieved between cloud-based modeling and lightweight edge inference. The cloud handles efficient training on large-scale data, while the edge only needs to deploy a distilled and optimized lightweight model to complete fault diagnosis, reducing the computational load and resource consumption at the edge. Knowledge distribution ensures the accuracy of the edge model's discrimination, achieving a balance between performance and resources. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in the embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0034] Figure 1 This is a diagram showing the overall structure of a cloud-edge collaborative device fault diagnosis system. Figure 2 This is a flowchart of a device fault diagnosis method based on cloud-edge collaboration. Detailed Implementation

[0035] It should be noted that 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 thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0036] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0037] Example 1 Please see Figure 1 The present invention also proposes a cloud-edge collaborative equipment fault diagnosis system to implement the above-mentioned equipment fault diagnosis method, including: The cloud-based model training module is used to receive runtime data uploaded from multiple edge nodes, fuse the runtime data to construct a joint dataset, and perform centralized model training to obtain a cloud-based model with global discriminative capabilities. The cloud-based model is used to generate knowledge information such as the category probability distribution of device operating status (this embodiment uses the category probability distribution as an example for explanation). That is, high-performance computing nodes are deployed in the cloud to centrally fuse runtime data uploaded from multiple device nodes. The edge model training module is used to optimize and train the edge model at the edge node based on locally collected data and guided by the category probability distribution generated by the cloud model, so as to realize the transfer of knowledge from the cloud model to the edge model.

[0038] In this embodiment, specifically, to maintain consistency and continuous optimization between cloud and edge models, this embodiment also includes a cloud-edge data interaction and knowledge update module, which is used to realize the uploading of running data, cloud model training updates, and cloud model knowledge distribution between edge nodes and the cloud, so as to support a cloud-edge collaborative closed loop of cloud retraining, knowledge distribution, and edge updates.

[0039] In this embodiment, specifically, the device operation signals collected by each edge node... After being aggregated in the cloud, a global training set is formed:

[0040] in, This indicates the number of device nodes participating in the training.

[0041] In this embodiment, the training of the cloud model is specifically as follows: After normalizing, denoising, and performing time-series segmentation preprocessing on the running signals uploaded by each edge node, a deep learning model is used to extract multi-layer time-frequency features and perform centralized training to establish a cloud model with global discrimination capabilities.

[0042] In this embodiment, specifically during the cloud-based model training phase, the cloud uses cross-entropy loss as the optimization objective:

[0043] in, For the first The true class label of each sample The model outputs a probability prediction for the fault category; the model parameters are then optimized using this loss function. The trained cloud model is obtained. This is used for subsequent knowledge distillation and edge model guidance.

[0044] In this embodiment, specifically, the edge model is deployed on each device node to achieve real-time fault diagnosis and adaptive model optimization locally; the edge model structure is a lightweight neural network, denoted as... ,in This represents the model parameters; its training process takes place in the cloud. Under the guidance of [the relevant authorities], knowledge transfer and ability inheritance are achieved through knowledge distillation.

[0045] In this embodiment, it should be noted that during each training cycle, the edge nodes receive the knowledge distribution data sent by the cloud model. And combined with the locally collected running sample dataset Joint optimization is performed. The overall loss function consists of two parts: Local cross-entropy loss:

[0046] in, For the sample The true label, This is the predicted probability output by the marginal model; this item is used to ensure that your marginal model can correctly identify the actual fault category of the local sample. Knowledge distillation loss:

[0047] in, and Cloud models and edge models are categorized separately. The output logits distribution on, This is the distillation temperature coefficient, used to adjust the smoothness of the cloud model's output. This term is used to constrain the output distribution of the edge model to approximate the soft label distribution of the cloud model, thereby learning the discriminative knowledge formed by the cloud model at the task level.

[0048] Therefore, the overall loss function of the edge model is as follows:

[0049] in, For local cross-entropy loss, For knowledge distillation loss, This is a weighting balancing coefficient used to adjust the importance of supervised learning and distillation learning. This is the distillation temperature coefficient; The edge model is optimized under this joint loss function so that, under limited computing resources, it can maintain the discriminative ability of local data and inherit the global knowledge features of the cloud model.

[0050] In this embodiment, it should be noted that during system operation, edge nodes periodically upload the collected device operation data to the cloud; assuming the data originates from the first... The uploaded dataset of each edge node is The global dataset used for training the cloud model can then be represented as:

[0051] After receiving the latest data uploaded by each edge node, the cloud merges the samples from all device nodes and re-executes the complete model training to optimize the model parameters. The updated cloud model is obtained. .

[0052] In this embodiment, specifically during cloud-based retraining, the training objective function is cross-entropy loss:

[0053] The cloud-based model, once trained, possesses global discriminative capabilities, and its output class probability distribution... As a new knowledge representation, it is distributed to each edge node.

[0054] After receiving model knowledge from the cloud, the edge nodes then use it to build upon their local datasets. Distillation updates are performed. The edge model uses a joint loss function as the optimization objective, which is as follows:

[0055] Edge models optimize parameters and absorb knowledge by minimizing the joint loss function, allowing their output to gradually approximate cloud models in terms of probability distribution.

[0056] In this embodiment, specifically, the cloud-edge collaborative device fault diagnosis system proposed in this embodiment retrains the cloud model after receiving the running data uploaded by each edge node during use, so as to update the global diagnostic knowledge. Edge nodes perform edge model distillation optimization based on the knowledge distribution distributed from the cloud, inheriting and strengthening the model's discriminative ability; The new operational data is then collected in the cloud, and the cloud model undergoes the next round of training and knowledge distribution. This cyclical process ensures that cloud and edge models remain consistent at the knowledge level, enabling dynamic optimization of diagnostic capabilities and long-term stability of system performance.

[0057] Please see Figure 2 In this embodiment, it should also be noted that, through the above system design, this embodiment also proposes a device fault diagnosis method based on cloud-edge collaboration. Step S1: The edge node collects the device's operating data and uploads it to the cloud; Step S2: The cloud integrates the operational data uploaded by each node for centralized model training to obtain a cloud model with global discrimination capability; the cloud model is used to generate the category probability distribution of the device's operating status. Step S3: The cloud distributes the category probability distribution as cloud model knowledge to the edge nodes. The edge nodes optimize and train the edge model based on the knowledge and local data to realize the transfer of cloud model knowledge to the edge model. Step S4: The optimized edge model performs fault diagnosis on the equipment site and sends the new operating data back to the cloud to form a cloud-edge collaborative closed loop of model retraining, knowledge distribution and edge update.

[0058] Example 2 Example 2 is a further explanation of Example 1. The same components will not be described again here. The details are as follows: Cloud and edge models can employ deep neural networks with different structural forms to adapt to the signal characteristics of different types of devices.

[0059] Cloud-edge communication can be achieved through secure encryption protocols, ensuring data security and network stability during knowledge transmission.

[0060] The system can combine operation logs and diagnostic results to create a cloud-edge model performance visualization and remote monitoring interface for device status monitoring and model version management.

[0061] Both this invention and existing cloud-edge collaborative diagnostic systems adopt a layered architecture, including a cloud computing layer and an edge computing layer, and have the function of data acquisition and uploading of multiple device nodes.

[0062] This invention differs from traditional cloud-edge collaborative systems that rely solely on parameter synchronization. Instead, it proposes a knowledge transfer-based "edge-device-edge" collaborative update mechanism. By establishing a closed-loop process at the system level—involving cloud-based knowledge generation, edge knowledge reception, and edge-side model updates—it achieves automated distribution and periodic updates of model knowledge, thereby continuously improving the diagnostic performance of edge models without altering the original model structure.

[0063] This invention differs from the static deployment method of one-time model distribution by constructing a periodic cloud-edge collaborative update system. After receiving the running data uploaded by each node, the cloud automatically triggers a unified retraining task, and the updated model knowledge is distributed to each edge node for synchronous optimization via the task scheduling module.

[0064] It should be noted that this embodiment proposes a cloud-based centralized modeling mechanism: This invention integrates operational data from multiple edge nodes in the cloud to construct a centrally trained cloud model, enabling unified modeling and global knowledge extraction of data from multiple devices and operating conditions. This results in a diagnostic model with strong discriminative capabilities, providing a reliable knowledge source for knowledge transfer at the edge.

[0065] It should be noted that this embodiment proposes an edge-end model optimization method: Edge nodes receive probability distribution information output by the cloud model and train the edge model on local data using a joint loss function, achieving knowledge transfer from the cloud to the edge. This maintains the lightweight nature of the edge model while giving it discriminative and diagnostic performance similar to that of the cloud model.

[0066] It should be noted that this embodiment proposes a cloud-edge collaborative dynamic update mechanism: This invention constructs a cyclical system of "cloud retraining - knowledge distribution - edge distillation - data backhaul". The cloud retrains the model based on the new data uploaded from the edge, while the edge nodes update their local models based on the knowledge from the cloud model, achieving dynamic consistency and continuous optimization of the cloud-edge models at the knowledge level, thereby ensuring the long-term stability and diagnostic capabilities of the system under various operating conditions.

[0067] Example 3 Example 3 is a further explanation of Example 1.

[0068] (1) Overall system structure This embodiment provides a device fault diagnosis system based on cloud-edge collaborative knowledge distillation. The system mainly consists of a cloud-based model training module, an edge model training module, and a cloud-edge data interaction and knowledge update module. The cloud-based model training module receives device operation data from various edge nodes, merges them to form a joint dataset, and performs centralized modeling. The edge model training module optimizes the model using lightweight distillation on local data based on the knowledge distribution distributed from the cloud. The cloud-edge data interaction and knowledge update module enables periodic information exchange and knowledge synchronization between the cloud and edge models.

[0069] The cloud and edge nodes interact with each other through a secure communication network. The cloud deploys high-performance computing servers and centralized storage systems, while the edge nodes deploy lightweight computing modules and sensing and acquisition units.

[0070] (2) Cloud-based model training module In the initial stage of system operation, each edge node collects device operating signals, such as vibration signals, and uploads the pre-processed samples to the cloud to form a training dataset. .in, For the first Signal samples from each device The corresponding fault category labels are used. Multi-node data is fused in the cloud and centralized model training is performed. Deep neural networks are used to extract multi-layer features and build a global model. Its optimization objective is the cross-entropy loss function. This module provides the system with accurate global discriminative knowledge and is the basic structure for realizing cloud-edge knowledge transfer.

[0071] (3) Edge model training module Edge nodes receive model output distributions from the cloud. and with locally collected datasets Distillation training is performed to form an edge model. The optimization objectives of edge models include cross-entropy loss and knowledge distillation loss. By minimizing the joint loss function, edge models inherit the discriminative capabilities of cloud models under computationally limited conditions, achieving efficient knowledge transfer.

[0072] (4) Cloud-edge data interaction and knowledge update model This module is responsible for the periodic information exchange between the cloud and edge models. Edge nodes periodically upload newly collected runtime data, which is then fused in the cloud to retrain the model. The updated model parameters are obtained. Knowledge distribution of model output The data is distributed to the edge for the next round of model distillation. Edge nodes update their models accordingly, achieving knowledge synchronization between the cloud model and the lightweight edge model. The entire process forms a periodic closed loop of "cloud retraining - knowledge distribution - edge distillation - data feedback," enabling the system to continuously learn.

[0073] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

[0074] This background section is provided to generally present the context of the invention. The work of the currently named inventors, the work to the extent described in this background section, and aspects of this section that did not constitute prior art at the time of application are neither expressly nor impliedly acknowledged as prior art to the invention.

Claims

1. A cloud-edge collaboration based device fault diagnosis method, characterized in that, The method comprises the following steps: Step S1: The edge node collects the operation data of the equipment and uploads the operation data to the cloud; Step S2: The cloud fuses the operation data uploaded by the nodes to perform centralized model training, and obtains a cloud model with global discrimination capability; the cloud model is used to generate knowledge information of the operation state of the equipment; Step S3: The cloud sends the knowledge information to the edge node, and the edge node optimizes and trains the edge model under the guidance of the knowledge to realize the migration of the knowledge of the cloud model to the edge model; Step S4: The optimized edge model performs fault diagnosis on the equipment in the field, and returns new operation data to the cloud to form a cloud-edge collaborative closed loop of model retraining, knowledge sending and edge updating.

2. A cloud-edge collaboration based device fault diagnosis system, characterized in that, The method for realizing the device fault diagnosis method of claim 1 comprises: a cloud model training module, configured to receive operation data uploaded by a plurality of edge nodes, fuse the operation data to construct a joint data set, and perform centralized model training to obtain a cloud model with global discrimination capability; the cloud model is used to generate knowledge information of the operation state of the equipment; an edge model training module, configured to optimize and train the edge model under the guidance of the knowledge information generated by the cloud model at the edge node according to the locally collected data, to realize the migration of the knowledge of the cloud model to the edge model; a cloud-edge data interaction and knowledge updating module, configured to realize the uploading of operation data between the edge node and the cloud, the training and updating of the cloud model, and the sending of the knowledge of the cloud model, to support the cloud-edge collaborative closed loop of cloud retraining, knowledge sending and edge updating. 3.The cloud-edge collaborative based device fault diagnosis system according to claim 2, wherein, The device operation signals collected by each edge node form a global training set after being converged by the cloud.

4. The cloud-edge collaboration based device fault diagnosis system according to claim 3, wherein, The cloud model is trained as follows: After the operation signals uploaded by each edge node are normalized, denoised and time-series segmented, a deep learning model is used to extract multi-layer time-frequency features and perform centralized training to establish a cloud model with global discrimination capability.

5. The cloud-edge collaboration based device fault diagnosis system according to claim 4, wherein, During the cloud model training stage, the cloud takes cross-entropy loss as the optimization target: in, For the first The true class label of each sample The knowledge information output by the model; the model parameters are optimized using this loss function. The trained cloud model is obtained. .

6. The cloud-edge collaboration based device fault diagnosis system according to claim 5, wherein, The edge model is deployed on each device node and is used to realize real-time fault diagnosis and model self-adaptive optimization locally.

7. The cloud-edge collaboration based device fault diagnosis system according to claim 6, wherein, The edge model structure is a lightweight neural network; the edge model training process is performed under the guidance of the cloud model, and knowledge migration and capability inheritance are realized through knowledge distillation; the overall loss function of the edge model is as follows: wherein, is a local cross-entropy loss, is a knowledge distillation loss, is a weight balancing coefficient, is a distillation temperature coefficient. 8.The cloud-edge collaboration based device fault diagnosis system of claim 7, wherein, The local cross-entropy loss includes: wherein, is the true label of the sample is the knowledge information output by the edge model, is the knowledge information output by the edge model, The knowledge distillation loss includes: wherein, with are the output logits distributions of the cloud model and the edge model on the class respectively.

9. The cloud-edge collaboration based device fault diagnosis system according to claim 8, wherein, During the cloud retraining process, the training target function is the cross-entropy loss: 。 10. The cloud-edge collaboration based device fault diagnosis system according to claim 9, wherein, During the edge updating process, the edge model takes the joint loss function as the optimization target, and the joint loss function is as follows: The edge model minimizes the joint loss function to realize parameter optimization and knowledge absorption, so that the output result gradually approaches the cloud model at the knowledge information level.

Images