Intelligent equipment health monitoring method, equipment, storage medium and program product

Abstract

The embodiment of the invention provides an intelligent equipment health monitoring method, equipment, a storage medium and a program product. The method comprises the following steps: performing knowledge distillation through a first teacher model and a second teacher model to generate a student model; deploying parameters of the student model and the global model to an edge node, and performing health monitoring on the equipment by the edge node; obtaining an optimized student model based on the received global parameter state information; performing edge node fault diagnosis on the received edge node data based on a fault diagnosis model; and maintaining the edge node according to the edge node fault data. According to the embodiment of the invention, the student model is constructed by utilizing knowledge distillation, the student model is optimized through federated training, health monitoring of equipment in the edge node is realized, fault diagnosis is carried out on the edge node data, the problem of limited edge node resources is effectively solved, the fault diagnosis capability is improved, the problem of edge node failure coverage is solved, and the fault diagnosis efficiency is improved. And the cloud edge collaboration efficiency is improved.

Description

Technical Field

[0001] This application belongs to the field of industrial Internet of Things (IoT) technology, and in particular relates to an intelligent device health monitoring method, device, storage medium, and program product. Background Technology

[0002] With the development of the Industrial Internet of Things (IIOT), the level of intelligence in manufacturing plants is constantly increasing. This has led to concerns about equipment stability, making equipment health monitoring a hot research topic. Timely detection of equipment failures and ensuring the continuous and reliable operation of smart factories are of great significance.

[0003] Industrial manufacturing equipment can experience various types of malfunctions with varying degrees of severity during use, threatening enterprise production and even operational safety. The sheer variety of equipment and the large number of instruments make single monitoring methods ineffective and difficult to deploy; furthermore, processing massive amounts of monitoring data presents a significant challenge.

[0004] Current technologies utilize edge cloud infrastructure, where sensors on devices perceive data, the edge performs real-time analysis, and complex tasks are uploaded to the cloud for processing. However, these technologies suffer from limitations such as the limited computing power of edge devices, which prevents the execution of complex AI models, and insufficient training data for high-precision models due to factory data. Traditional modules also suffer from drawbacks, including high latency in fault detection due to periodic data uploads, inability to handle complex operating data caused by device heterogeneity, and inability to address coverage issues arising from edge device failures. Summary of the Invention

[0005] This application provides an intelligent device health monitoring method, device, storage medium, and program product, which can effectively solve the problem of limited edge node resources, improve fault diagnosis capabilities, address edge node failure coverage issues, and improve cloud-edge collaboration efficiency.

[0006] In a first aspect, embodiments of this application provide an intelligent device health monitoring method applied to a cloud server, comprising: generating a student model by performing knowledge distillation through a pre-constructed first teacher model and a second teacher model; distributing the student model and global model parameters to a regional coordination server, so that the regional coordination server deploys the student model and global model parameters to edge nodes, enabling the edge nodes to perform device health monitoring; receiving global parameter status information sent by the regional coordination server, performing parameter aggregation processing based on the global parameter status information, updating the first teacher model and the second teacher model, and obtaining an optimized student model, wherein the global parameter status information is forwarded by the edge nodes through the regional coordination server after edge federated learning training of the student model using the global model parameters; receiving edge node data sent by the regional coordination server, performing edge node fault diagnosis on the edge node data based on a pre-constructed fault diagnosis model, obtaining edge node fault data; and sending the edge node fault data to the regional coordination server, so that the regional coordination server maintains the edge nodes according to the edge node fault data.

[0007] In one feasible implementation, the step of generating a student model through knowledge distillation using a pre-constructed first teacher model and a second teacher model includes: extracting features from time-series data using the first teacher model to obtain corresponding first teacher features; extracting features from image data using the second teacher model to obtain corresponding second teacher features; and extracting features from multimodal input data using a student training model to obtain corresponding student training features; projecting features onto the first teacher features and the second teacher features respectively to obtain corresponding first teacher projection features and second teacher projection features; performing weighted fusion processing on the first teacher projection features and the second teacher projection features using pre-set modality fusion weights to obtain teacher fusion features; determining a mean squared error loss value, a relation matrix similarity, and a dynamic temperature parameter based on the teacher fusion features and the student training features; determining a loss function value based on the mean squared error loss value, the relation matrix similarity, and the dynamic temperature parameter; and training the student training model using the loss function value to generate the student model.

[0008] In one feasible implementation, the step of performing parameter aggregation processing based on the global parameter state information to update the first teacher model and the second teacher model, and to obtain the optimized student model, includes: assigning corresponding edge node aggregation weight ratios to the global parameter state information using pre-set device importance scores for each edge node; dynamically aggregating parameters of the global parameter state information according to the edge node aggregation weight ratios, updating the first teacher model and the second teacher model, and obtaining the optimized student model.

[0009] In one feasible implementation, the step of dynamically aggregating global parameter state information based on the aggregated weight ratio of edge nodes to obtain an optimized student model includes: determining the importance score of equipment according to the key equipment and basic general equipment corresponding to the equipment groups in each edge node; normalizing the importance score of each edge node, and determining the original weight ratio of the edge node based on the product of the normalized importance score and the local training data of each edge node; determining the original weight ratio of a single edge node based on the original weight ratio of the edge nodes and the number of edge nodes; normalizing the original weight ratio of a single edge node, and dynamically aggregating parameters based on the normalized original weight ratio of the single edge node and the local model parameters of the edge node to determine global model parameters; and obtaining the optimized student model using the global model parameters.

[0010] In one feasible implementation, the device importance score is determined in the following manner: ,in, The score indicates the importance of the equipment. This refers to the basic equipment in the corresponding device group of the edge node. This represents the key equipment in the corresponding device group within the edge node; the normalized equipment importance score is determined in the following way: ,in, This represents the normalized score indicating the importance of the equipment. This represents the lower bound of normalization. This represents the upper limit of normalization. This represents the minimum value of the device importance score corresponding to all the aforementioned edge nodes. This represents the maximum value of the device importance score corresponding to all the edge nodes; the original weight percentage of the edge nodes is determined in the following way: ,in, This represents the original weight percentage of the edge nodes. This represents the amount of local data at the i-th edge node; the original weight percentage of a single edge node is determined in the following way: ,in, This represents the original weight percentage of a single edge node; the normalized original weight percentage of a single edge node is determined in the following way: ,in, This represents the original weight percentage of a single edge node after normalization; the global model parameters are determined in the following way: ,in, Represents global model parameters. This represents the local model parameters.

[0011] In one feasible implementation, the method further includes: receiving device health monitoring results sent by a regional coordination server, the device health monitoring results being obtained by the edge node performing health monitoring on the device based on the student model; and, if the device health monitoring results are of an uncertain type, diagnosing the device health of the uncertain type using the first teacher model and the second teacher model.

[0012] In one feasible implementation, diagnosing the health of the uncertain type of device using the first teacher model and the second teacher model includes: performing diagnostic processing on the uncertain type of device using the first teacher model and the second teacher model respectively, and determining whether the diagnostic results of the first teacher model and the second teacher model are consistent; if the diagnostic results of the first teacher model and the second teacher model are consistent, taking the diagnostic result of the first teacher model as the health diagnosis result of the device; if the diagnostic results of the first teacher model and the second teacher model are inconsistent, performing weighted processing based on the accuracy rates of the historical diagnostic results corresponding to the first teacher model and the second teacher model, and taking the weighted historical diagnostic result as the health diagnosis result of the device.

[0013] In one feasible implementation, the first teacher model and the second teacher model are determined by: training a temporal convolutional network model using a time-series data training dataset to construct the first teacher model; and training an object detection model using an image data training dataset to construct the second teacher model.

[0014] Secondly, embodiments of this application provide an intelligent device health monitoring method applied to a regional coordination server, comprising: receiving student models and global model parameters sent by a cloud server, and deploying the global model parameters and the student model to edge nodes; receiving edge node data sent by the edge nodes, sending the edge node data to the cloud server, and receiving edge node fault data returned by the cloud server for the edge node data, and maintaining the edge nodes according to the edge node fault data, wherein the edge node fault data is obtained by the cloud server performing edge node fault diagnosis on the edge node data based on a pre-built fault diagnosis model; utilizing the pre-built global model parameters and global model parameters, the method further comprises: receiving student models and global model parameters sent by a cloud server, and deploying the global model parameters and the student model to edge nodes; receiving edge node data sent by the edge nodes, sending the edge node data to the cloud server, and receiving edge node fault data returned by the cloud server for the edge node data; and maintaining the edge nodes according to the edge node fault data, wherein the edge node fault data is obtained by the cloud server performing edge node fault diagnosis on the edge node data based on a pre-built fault diagnosis model; and ... and deploying the global model parameters and the student model to edge nodes according to edge nodes according to edge nodes according to edge nodes according to edge nodes according to edge nodes according to edge nodes according to edge nodes according to edge nodes according to edge nodes according to edge nodes according to edge nodes according to edge nodes according to edge nodes according to edge nodes according to edge nodes according The local optimization objective involves performing edge node adjustment analysis on edge node state data to obtain global adjustment parameters for edge nodes, and sending these global adjustment parameters to the corresponding edge nodes so that the edge nodes can adjust their parameters using these parameters. It also receives global parameter state information from multiple edge nodes, which is obtained by the edge nodes using the global model parameters to perform edge federated learning training on the student model. Finally, it sends the global parameter state information to the cloud server, enabling the cloud server to perform parameter aggregation processing based on the global parameter state information, update the first teacher model and the second teacher model, and obtain the optimized student model.

[0015] In one feasible implementation, the step of performing edge node global adjustment analysis on edge node state data using a pre-constructed global optimization objective to obtain edge node global adjustment parameters includes: calculating global indices of edge nodes based on node state data; determining the theoretical optimal solution of the edge node using the global optimization objective; and determining the edge node global adjustment parameters by comparing the global indices and the theoretical optimal solution.

[0016] In one feasible implementation, the global optimization objective is obtained by setting a global optimization objective and a basic sub-objective; wherein, the global optimization objective is... G g E represents the global node coverage performance. i Let β represent the energy consumption of node i, where i∈A, A represents the set of edge nodes in the active state, and β represents the influence factor. ,in, Let A1 and A2 represent the coverage area of ​​the i-th active node (cover1 and cover2 represent general nodes and key nodes, respectively, with key nodes having a greater weight in the coverage performance index), and let R represent the coverage area of ​​the entire smart factory where the intelligent device is located. The basic sub-objective is... , This represents the global node coverage rate. Indicates the basic coverage target; , Let R represent the coverage area of ​​the i-th active node, R represent the coverage area of ​​the entire smart factory where the smart device is located, and x represent the number of nodes; based on the global optimization objective and the basic sub-objective, the global optimization objective corresponding to the edge node is constructed according to the pre-set global objective constraints.

[0017] In one feasible implementation, the maintenance of edge nodes based on the edge node fault data includes: determining global parameters and LA learning automata objects corresponding to global nodes, wherein the global nodes include device nodes, edge nodes, and cloud servers; updating the state of edge nodes in the global parameters based on the edge node fault data and a preset edge node state machine; and determining a neighbor node selection strategy based on the state of edge nodes in the global parameters, the LA learning automata objects, and a neighbor comprehensive evaluation function, wherein the neighbor node selection strategy is used to select neighbor nodes to be activated, and the neighbor nodes to be activated are used to back up the content of the faulty edge nodes.

[0018] In one feasible implementation, before receiving the global parameter status information sent by multiple edge nodes, the method further includes: obtaining a list of state attribute values ​​of the edge nodes; determining a set of candidate edge nodes to participate in federated training based on the list of state attribute values; selecting edge nodes to participate in federated training from the set of candidate edge nodes based on preset bandwidth constraints; and sending a command to the edge nodes participating in federated training to execute federated training, so that the edge node receiving the command uses the global model parameters to perform edge federated learning training on the student model to obtain the global parameter status information.

[0019] In one feasible implementation, the state attribute value list includes the remaining power evaluation index, edge node criticality evaluation index, and entropy evaluation index of the edge node; determining the candidate edge node set for participating in federated training based on the state attribute value list includes: performing priority evaluation processing on a pre-set index set using the remaining power evaluation index, the edge node criticality evaluation index, and the entropy evaluation index to obtain an evaluation set; determining whether the edge nodes in the evaluation set meet the pre-set initial screening conditions according to the pre-set initial screening conditions of the edge nodes; if the edge node does not meet the initial screening conditions, the edge node is removed from the evaluation set. The edge nodes are stored in a blacklist; if the edge nodes meet the initial screening conditions, they are stored in a whitelist; the edge nodes in the whitelist are sorted, and it is determined whether there are any edge nodes that meet the preset total bandwidth requirements among the top target number of edge nodes; if there are no edge nodes that meet the total bandwidth requirements among the top target number of edge nodes, the edge nodes are sparsified; if there are edge nodes that meet the total bandwidth requirements among the top target number of edge nodes, the edge nodes that meet the total bandwidth requirements are added to the candidate edge node set for participating in federated training.

[0020] In one feasible implementation, the edge node criticality evaluation index is determined by the following formula: Among them, b i2 Let m represent the criticality evaluation index of the i-th edge node, k represent the number of critical devices covered by the edge node, k represent the types of devices included in the edge node, and N represent the total number of devices covered by the edge node; the entropy evaluation index is determined in the following way: Among them, b i3 Let p(x) represent the data distribution entropy value of the i-th edge node, where the i-th edge node contains m types of devices, and p(x) represent the proportion of data volume of the x-th type of device; the total bandwidth requirement is determined in the following way: in, Total bandwidth 'b' represents the parameter size of the edge node, and 'b' represents the parameter precision.

[0021] In one feasible implementation, before receiving the edge node data sent by the edge node, the method further includes: determining a device distance matrix based on the device nodes in the device layer; setting an optimization objective relation for the device nodes in the device distance matrix according to a pre-set grouping rule, wherein the optimization objective relation is obtained using a priority factor, an objective relation for assigning adjacent devices to the same group, an objective relation for each device group including multiple device types, and an objective relation for assigning key devices to different groups; determining the priority constraints of the device nodes in the device layer; and solving a multi-objective nonlinear programming problem using a genetic algorithm or a particle swarm optimization algorithm based on the optimization objective relation and the priority constraints to obtain a device grouping result, wherein the device grouping result is used to determine the correspondence between edge nodes and device groups.

[0022] Thirdly, embodiments of this application provide an intelligent device health monitoring method applied to edge nodes, comprising: receiving global model parameters and a student model from a regional coordination server; deploying the student model and the global model parameters on the edge node; the student model being generated by a cloud server through knowledge distillation using a pre-built first teacher model and a second teacher model; training the student model locally based on the global model parameters and an edge federated learning algorithm to obtain local training model parameters; sending the local training model parameters to the regional coordination server, causing the regional coordination server to send the local training model parameters of each edge node as global parameter status information to the cloud server for parameter aggregation processing, thereby updating the first teacher model and the second teacher model, and obtaining an optimized student model; collecting edge node data and sending it to the regional coordination server, causing the regional coordination server to send the edge node data to the cloud server, causing the cloud server to perform edge node fault diagnosis on the edge node data based on a pre-built fault diagnosis model to obtain edge node fault data; and performing health monitoring on the device based on the student model to obtain the device health monitoring result.

[0023] In one feasible implementation, the collection of edge node data includes: performing self-checks on edge nodes to obtain edge node status indicators; collecting neighbor information to establish a neighbor node table for each edge node; using a pre-built node state machine to process the state changes of the edge node status indicators; using a node selection strategy to select corresponding target neighbor nodes for the edge nodes after the state changes; and collecting edge node data of neighbor nodes based on the target neighbor nodes.

[0024] Fourthly, embodiments of this application provide an intelligent device health monitoring device, the device comprising: a processor, and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the intelligent device health monitoring method as described in any one of the first, second, or third aspects.

[0025] Fifthly, embodiments of this application provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the intelligent device health monitoring method as described in any one of the first, second, or third aspects.

[0026] Sixthly, embodiments of this application provide a computer program product, including a computer program, which, when executed, implements the intelligent device health monitoring method as described in any one of the first, second, or third aspects.

[0027] This application discloses an intelligent device health monitoring method, device, storage medium, and program product. It generates a student model through knowledge distillation using a pre-built first teacher model and a second teacher model; deploys the student model and global model parameters to edge nodes, enabling edge nodes to monitor device health; aggregates parameters based on received global parameter status information to update the first and second teacher models, resulting in an optimized student model; performs edge node fault diagnosis on received edge node data based on a pre-built fault diagnosis model, obtaining edge node fault data; and maintains the edge nodes based on the edge node fault data. By utilizing knowledge distillation to construct the student model and then optimizing it through federated training, it achieves health monitoring of devices in edge nodes and fault diagnosis of edge node data, effectively solving the problem of limited edge node resources, improving fault diagnosis capabilities, addressing edge node failure coverage issues, and improving cloud-edge collaboration efficiency. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a schematic diagram of a device health monitoring structure provided in an embodiment of this application; Figure 2 This is a flowchart illustrating an intelligent device health monitoring method provided in an embodiment of this application; Figure 3This is a flowchart illustrating a student model generation process provided in an embodiment of this application; Figure 4 This is a flowchart illustrating a teacher model projection process provided in an embodiment of this application; Figure 5 This is a flowchart illustrating a student model optimization process provided in an embodiment of this application; Figure 6 This is a flowchart illustrating a dynamic parameter aggregation process provided in an embodiment of this application; Figure 7 This is a flowchart illustrating another intelligent device health monitoring method provided in an embodiment of this application; Figure 8 This is a flowchart illustrating yet another intelligent device health monitoring method provided in this application embodiment; Figure 9 This is a flowchart illustrating a teacher model construction process provided in an embodiment of this application; Figure 10 This is a flowchart illustrating the operation of a cloud server according to an embodiment of this application; Figure 11 This is a flowchart illustrating another intelligent device health monitoring method provided in an embodiment of this application; Figure 12 This is a flowchart illustrating a method for determining global adjustment parameters of an edge node, as provided in an embodiment of this application. Figure 13 This is a flowchart illustrating a method for determining global adjustment parameters of an edge node, as provided in an embodiment of this application. Figure 14 This is a schematic diagram of an edge node maintenance process provided in an embodiment of this application; Figure 15 This is a flowchart illustrating a method for determining a node state machine, as provided in an embodiment of this application. Figure 16 This is a flowchart illustrating a node maintenance process provided in an embodiment of this application; Figure 17 This is a flowchart illustrating a federated training process provided in an embodiment of this application; Figure 18 This is a flowchart illustrating another federated training process provided in an embodiment of this application; Figure 19 This is a schematic diagram of a process for determining a set of candidate edge nodes provided in an embodiment of this application; Figure 20 This is a flowchart illustrating a device grouping process provided in an embodiment of this application; Figure 21 This is a flowchart illustrating another intelligent device health monitoring method provided in an embodiment of this application; Figure 22 This is a schematic diagram of a blacklist dynamic update process provided in an embodiment of this application; Figure 23 This is a schematic diagram of a process for determining candidate edge nodes provided in an embodiment of this application; Figure 24 This is a flowchart illustrating a process for collecting edge nodes provided in an embodiment of this application; Figure 25 This is a schematic diagram of the structure of an intelligent device health monitoring device provided in an embodiment of this application. Detailed Implementation

[0030] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0031] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations 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..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0032] It should be noted that the acquisition, storage, use, and processing of data in this application embodiment all comply with the relevant provisions of national laws and regulations.

[0033] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0034] The Temporal Convolutional Network (TCN) model is a deep learning architecture specifically designed for temporal data. Its core utilizes causal convolution, dilated convolution, and residual connections to address the vanishing gradient problem of traditional recurrent neural networks (such as LSTM) and supports efficient parallel computation.

[0035] The YOLOv8-Heatmap Fusion model is a module in the YOLOv8 object detection framework used for heatmap generation and fusion. Its core function is to visualize the model's attention to different regions in an image in the form of a heatmap through multi-scale feature fusion and attention mechanisms, thereby intuitively showing the confidence distribution of the detection results.

[0036] The MobileViT model is a lightweight visual Transformer model that combines the local feature extraction capabilities of CNNs with the global modeling advantages of ViT to achieve efficient image processing. Its core innovation is the MobileViT Block: it first encodes local information using convolution, then learns the global representation through a Transformer, and finally fuses the two to output the final result.

[0037] Currently, the approach used is to build an edge cloud, where devices perceive data through sensor nodes, and the edge performs real-time analysis of the data. Complex tasks are then uploaded to the cloud for processing. However, this solution has the following problems: ① Edge devices have limited computing power and cannot directly run complex AI models; ② Training high-precision models requires a large amount of rich data on factory equipment operation, but some of this data may involve security and confidentiality issues; ③ Traditional modules rely on periodically uploading data to the cloud for analysis, resulting in high latency (seconds or even minutes) in fault detection and an inability to handle complex operating data caused by heterogeneous equipment; ④ It cannot address the coverage issues caused by edge device failures.

[0038] To address the aforementioned technical problems, embodiments of this application provide an intelligent device health monitoring method, device, storage medium, and program product. A student model is generated by knowledge distillation using a pre-built first teacher model and a second teacher model; the student model is trained using federated learning, updating the first and second teacher models, and the resulting optimized student model, to monitor the health of devices in edge nodes; edge node fault diagnosis is performed based on a pre-built fault diagnosis model to obtain edge node fault data; and edge nodes are maintained based on the edge node fault data. This effectively solves the problem of limited edge node resources, improves fault diagnosis capabilities, addresses edge node failure coverage issues, and improves cloud-edge collaboration efficiency. The intelligent device health monitoring method provided in this application embodiment is described below.

[0039] Figure 1 This is a schematic diagram of a device health monitoring system provided in an embodiment of this application. Figure 1 As shown in the diagram, the device health monitoring structure includes a device layer S1, a data mapping layer S2, and an edge node layer S3 (i.e., the edge nodes in subsequent embodiments). The device layer S1 is responsible for the data collection, aggregation, and uploading of smart devices, forming the foundation for cloud-edge collaboration. The data mapping layer S2 is responsible for extracting and stitching features from heterogeneous sensor data (used to collect image and time-series data from devices) to ensure consistency in the input data features of each edge node. The edge node layer S3 has a built-in lightweight small model responsible for rapid diagnosis and fault classification of device data. It also serves as a trainer for local fault data in the federated learning architecture and is managed by the regional coordination layer S4, which periodically uploads local log information from the edge nodes. A regional coordination layer S4 (i.e., the regional coordination server described in subsequent embodiments) is located between the edge node layer S3 and the cloud service layer S5. The regional coordination layer S4 is responsible for cloud-edge collaboration and scheduling. By collecting the reported information from each edge node, it calculates global optimization indicators, dynamically adjusts reward and penalty parameters, and ensures that all nodes are in optimal working condition. It also has a self-protection function for failed edge nodes. The cloud service layer S5 (i.e., the cloud server described in the following embodiments) is an important component of the federated learning architecture, and is responsible for distributing policy parameters, secondary diagnosis of uncertain faults, predictive maintenance, and providing a remote operating condition visualization interface.

[0040] This application applies to smart factories, smart cities, smart transportation, and other areas with a large number of smart devices.

[0041] Figure 2 This diagram illustrates a flowchart of an intelligent device health monitoring method provided in an embodiment of this application. It is applied to a cloud server. For example... Figure 1 As shown, the intelligent device health monitoring method specifically includes the following steps: S201~S205.

[0042] S201. Knowledge distillation is performed using a pre-built first teacher model and a second teacher model to generate a student model.

[0043] The first teacher model mentioned here can be understood as a model built using the TCN model, used for processing time-series data. The second teacher model mentioned here can be understood as a model built using the YOLOv8 model, used for processing image data.

[0044] In the device health detection structure consisting of cloud servers, regional coordination servers, and edge nodes, a first teacher model and a second teacher model are pre-built in the cloud server. By performing knowledge distillation on the first teacher model and the second teacher model, they are merged into a lightweight small model, which serves as the student model.

[0045] By leveraging the powerful computing resources of cloud servers and publicly available small-scale factory equipment datasets, a high-precision large model is trained, and a small model that can be deployed on edge nodes is extracted using distillation techniques.

[0046] S202. Send the student model and global model parameters to the regional coordination server so that the regional coordination server can deploy the student model and global model parameters to the edge nodes, enabling the edge nodes to perform health monitoring on the devices.

[0047] The obtained student model is sent to the regional coordination server via the cloud server. The regional service coordinator then sends the received student model and its own generated global model parameters to the edge nodes. The student model and its own generated global model parameters are deployed on the edge nodes, making it easier for the edge nodes to use the student model to perform health diagnoses on the devices.

[0048] S203. Receive global parameter status information sent by the regional coordination server, perform parameter aggregation processing based on the global parameter status information, update the first teacher model and the second teacher model, and obtain the optimized student model. The global parameter status information is forwarded by the regional coordination server after the edge node uses the global model parameters to train the student model through edge federated learning.

[0049] After the student model is deployed on the edge nodes, edge federated learning training is performed on the student model. Global single-state information is generated by using edge node data obtained from the regional coordination server and sent to the cloud server. The cloud server performs parameter aggregation processing on the global parameter state information, updates and optimizes the first teacher model and the second teacher model, and obtains the optimized student model. This realizes the optimization of the student model using the edge federated learning algorithm, while feeding back to the cloud server to optimize and adjust the first teacher model and the second teacher model, providing a more accurate model for the next step of cloud-edge collaborative control.

[0050] S204. Receive edge node data sent by the regional coordination server, perform edge node fault diagnosis on the edge node data based on the pre-built fault diagnosis model, and obtain edge node fault data.

[0051] S205. Send the edge node fault data to the regional coordination server so that the regional coordination server can maintain the edge nodes based on the edge node fault data.

[0052] After obtaining the optimized student model, the edge nodes not only perform health checks on the devices using the optimized student model, but also diagnose whether their own edge nodes have experienced faults. First, edge node data is generated by the edge nodes and sent to the cloud server through the regional coordination server. The cloud server then performs edge node fault diagnosis on the edge node data to determine whether the edge nodes have experienced faults, anomalies, or node failures, obtaining edge node fault data. In the event of a problem with an edge node, the cloud server sends the edge node fault data to the regional coordination server, which then maintains the edge nodes corresponding to the edge node fault data. This resolves edge node failures or fault diagnosis issues, effectively addresses the problem of limited edge node resources, improves fault diagnosis capabilities, addresses edge node failure coverage issues, and enhances cloud-edge collaboration efficiency.

[0053] This application discloses an intelligent device health monitoring method, device, storage medium, and program product. It generates a student model through knowledge distillation using a pre-built first teacher model and a second teacher model; deploys the student model and global model parameters to edge nodes, enabling edge nodes to monitor device health; aggregates parameters based on received global parameter status information to update the first and second teacher models, resulting in an optimized student model; performs edge node fault diagnosis on received edge node data based on a pre-built fault diagnosis model, obtaining edge node fault data; and maintains the edge nodes based on the edge node fault data. This effectively solves the problem of limited edge node resources, improves fault diagnosis capabilities, addresses edge node failure coverage issues, and enhances cloud-edge collaboration efficiency.

[0054] Figure 3 This is a flowchart illustrating a student model generation process provided in an embodiment of this application. It focuses on the student model generation process in step S201. Figure 3 The description is based on the above embodiments. Figure 3 The provided diagram illustrates the specific steps in the student model generation process, including S301 to S306.

[0055] S301. The first teacher model is used to extract features from the time series data to obtain the corresponding first teacher features. The second teacher model is used to extract features from the image data to obtain the corresponding second teacher features. The student training model is used to extract features from the multimodal input data to obtain the corresponding student training features.

[0056] First, let's represent the time series data X t The input is fed into the TCN model corresponding to the first teacher model to perform TCN feature extraction, thus obtaining the first teacher feature Y. t .

[0057] Among them, time series data X t for First Teacher Characteristic Y t for X t Y represents the input of time series data. t This represents the output of the 4th layer of the TCN model (the 4th layer was chosen to strike a balance between preserving temporal details and capturing long-range dependencies). B, T, and D represent the batch size, time step (i.e., the number of sequences sampled), and feature dimension, respectively.

[0058] Secondly, the image data X will be represented i The input is fed into the YOLOv8 model corresponding to the second teacher model to perform YOLOv8 feature extraction, thus obtaining the second teacher feature Y. i .

[0059] Among them, image data X i for Second teacher characteristic Y i for X i This represents the input image data, with a default feature of 3*640*640 (in actual use, images of other sizes can be changed to 640*640 through compression, padding, etc.), Y i This is a multi-scale output feature map (containing deep and shallow features) of the Neck layer of the YOLOv8 model.

[0060] Finally, it will represent multimodal X. m The input is fed into the MobileViT model corresponding to the student training model for MobileViT feature extraction to obtain the student training features Y. m .

[0061] Among them, multimodal X m for Student training characteristics Y m for X m For multimodal input, the first 3 channels are multi-scale feature maps Y. i Channel 4 is the timing signal Y. t The pseudo-images obtained through time-frequency transformation are finally stitched together to form a 4-channel input; Y m These are the fused features output by the MobileViT model Stage3.

[0062] Based on the features obtained from the three models, the output features of the three models are standardized according to Equation 1: Equation 1 Among them, Y kFor the standardized output, u, The mean and standard deviation of each modal feature are given. , To learn parameters (normalization may destroy relevant features). Used for scaling different modalities or different feature channels. Used to translate different features , It is usually initialized to 1 and 0, and learns automatically through backpropagation.

[0063] S302. Project the features of the first teacher and the second teacher respectively to obtain the corresponding projected features of the first teacher and the second teacher.

[0064] A bi-branch learnable projection head is used to utilize the features of the first teacher and the second teacher. By projecting the features of the first teacher and the second teacher respectively, feature projection processing of the time series data branch and the image data branch is realized, resulting in the projection features of the first teacher and the second teacher. This facilitates subsequent special processing of the time series data branch and the image data branch before projecting them onto the same feature dimension.

[0065] In one possible example scenario, Figure 4 This is a flowchart illustrating a teacher model projection process provided in an embodiment of this application. It focuses on the teacher model projection process in step S302. According to... Figure 4 The provided diagram shows that, firstly, in step S41, the TCN output features under the first teacher model are processed; secondly, in step S42, the TCN output features are input into the 1D convolution module for the first layer of convolution processing; thirdly, in step S43, they are input into the temporal attention module for attention training; and finally, in step S44, the obtained temporal features are processed using the projection matrix. Feature projection is performed to obtain the first teacher's projected features. Similarly, the YOLOv8 output features of the second teacher model in step S45 are processed by inputting them into the spatial pyramid pooling module in step S46 for pooling. Then, attention training is performed through the multi-channel attention module in step S47. Finally, the image features are input into the projection matrix in step S48. and The second teacher projection feature is obtained by performing feature projection.

[0066] By using TCN model features Features of the YOLOv8 model After projection, all models are mapped to a 512-dimensional common space. The projection features of the TCN model corresponding to the first teacher model are: The YOLOv8 projection features corresponding to the second teacher model are: .

[0067] S303. The first teacher projection feature and the second teacher projection feature are weighted and fused using pre-set modal fusion weights to obtain the teacher fused features.

[0068] Modality fusion weights are generated using a gating network. First, the features projected from the TCN model corresponding to the first teacher model and the YOLOv8 model corresponding to the second teacher model are unified in spatial dimension.

[0069] Features of the YOLOv8 model In and Downsample to 20*20 and then stitch together the channels: .

[0070] TCN model features After expansion, the matching spatial dimensions are obtained. Then, the modalities are spliced ​​and fused to obtain the features. .

[0071] Then After tiling, the data is fed into an MLP perceptron (responsible for capturing complex cross-modal relationships and mapping high-dimensional features to interpretable weights). After dimensionality compression and activation by a fully connected layer within the perceptron using a non-linear activation function (such as ReLU), the data is finally passed through a sigmoid function. Generate weight parameters Weight parameters We obtain the following from Equation 2: Equation 2 Furthermore, modal fusion weights are used to perform weighted fusion of the two types of features. The result after weighted fusion is expressed by Equation 3: Equation 3 Among them, F proj The result is a weighted fusion of temporal and image features. yes The range of the results after expanding the spatial dimension matching is (∈B*T*512*32*32). It is achieved through 1*1 convolution The range of the result after three-channel feature compression is (∈B*512*32*32), W i and W t (Range ∈ 512*D) represent the weight matrices for time series and image, respectively, where D represents the target dimension, which is adjusted according to task requirements (different tasks may have different dimensional requirements) to make it more flexible. Generated through an MLP gated network, the model can automatically adjust the contribution of time series and image modalities according to different samples, making it more flexible and efficient.

[0072] S304. Determine the mean squared error loss value, relation matrix similarity, and dynamic temperature parameters based on teacher fusion characteristics and student training characteristics.

[0073] S305. Determine the loss function value based on the mean square error loss value, the similarity of the relation matrix, and the dynamic temperature parameter.

[0074] S306. Use the loss function value to train the student training model and generate the student model.

[0075] Then obtain the teacher integration feature F proj Based on this, plus student training characteristics Y m The mean squared error loss (MSE loss, which is also the square of the Euclidean distance) is calculated using Equation 3: Equation 3 Where L is the MSE loss value and B is the batch size.

[0076] Furthermore, the similarity of the relationship matrices between the feature maps of teachers and students is calculated.

[0077] First, construct the relationship matrix, and then calculate the cosine similarity between each position in the teacher model and the other positions in both the teacher and student models. This is represented by Equations 4 and 5: Equation 4 Equation 5 Among them, G T (i, j) and G S (i, j) represent the relationship matrices of the teacher model and the student model, respectively, and c_s represents the cosine similarity function.

[0078] Normalize the vector at each position in the relation matrix into a probability to obtain the normalized probabilities of the teacher model and the student model (softmax represents the probability normalization formula).

[0079] Among them, the teacher model normalized probability Equation 6 represents: Equation 6 Normalized probability of student model Equation 7 represents: Equation 7 Furthermore, by calculating the mean of the KL divergence: Equation 8 Equation 9 Among them, Ksum The KL divergence, representing the total number of positions, is primarily used to measure the degree of difference between two probability distributions.

[0080] Furthermore, dynamic temperature parameters are designed to automatically adjust the soft target distribution based on feature differences, resulting in Equation 10 representing the dynamic temperature parameters, and Equation 11 determining the learnable parameters: Equation 10 Equation 11 Where β represents the learnable parameter, u represents the decay coefficient, which controls the rate at which β decreases, and x represents the training epoch; β has a larger value in the initial stage to learn similarity, and decreases in the convergence stage to focus on model details; MSE represents the mean squared error. This is a very small constant (e.g., 1e-6) used to prevent the temperature coefficient from being zero. Traditional schemes often use a fixed temperature coefficient, resulting in the same soft target distribution for all samples, making it difficult to adapt to all samples. This application implements an adaptive teaching strategy in the knowledge distillation process through a feature difference feedback mechanism and learnable parameters, thereby improving model compression and transfer efficiency.

[0081] Based on the above parameters, the loss function value is determined using the mean squared error loss value, the similarity of the relation matrix, and the dynamic temperature parameter. The loss function is expressed using Equation 12: Equation 12 Where L represents hard loss, L1 represents soft loss, and T 2 This is used to compensate for the decrease in gradient magnitude of KL divergence at high temperatures. The trained MobileViT student model is converted to TensorRT format, thus obtaining the student model, and deployed on the edge nodes.

[0082] This application embodiment extracts features from the time-series data and image data of the absolute device by pre-constructing a first teacher model and a second teacher model in a cloud server. The obtained features are then projected and fused to obtain fused features. The fused features are then normalized and multiple parameter data are calculated to obtain a loss function. The student model is then trained by knowledge distillation using the loss function to obtain a student model. The resulting lightweight student model solves the problems of lightweighting and initial capability of edge models.

[0083] Figure 5 This is a flowchart illustrating a student model optimization process provided in an embodiment of this application. It focuses on the student model optimization process in step S203. Figure 5 The description is based on the above embodiments. Figure 5 The provided diagram illustrates the specific steps of the student model optimization process, including S501~S502.

[0084] S501. Using the pre-set device importance score of each edge node, assign the corresponding edge node aggregation weight ratio to the global parameter status information.

[0085] S502. Based on the weight ratio of edge nodes, dynamically aggregate the global parameter state information, update the first teacher model and the second teacher model, and obtain the optimized student model.

[0086] Based on the obtained student model, in order to further optimize the student model, edge nodes use global model parameters to perform edge federated learning training on local training data. After obtaining the local model training parameters, the edge nodes send the local model training parameters to the regional coordination server, which then collects the local model training parameters of each edge node, obtains global parameter status information, and sends the global parameter status information to the cloud server. The cloud server uses a pre-set device importance score for each edge node to assign corresponding aggregation weights to the global parameter status information. By using different aggregation weights, the global parameter status information is dynamically aggregated to obtain new global model parameters. Finally, the obtained global model parameters are fed back to the student model to optimize the student model, and the first teacher model and the second teacher model are also updated.

[0087] This application embodiment, based on the construction of a student model, further feeds back the corresponding global parameter status information to the cloud server through a regional coordinator, and then uses the device importance score of each edge node to assign a corresponding aggregate weight ratio to the global parameter status information. The aggregate weight ratio is used to dynamically aggregate the global parameter status information to optimize the student model and further improve the monitoring accuracy of the lightweight student model.

[0088] Figure 6 This is a flowchart illustrating a dynamic parameter aggregation process provided in an embodiment of this application. It describes the dynamic parameter aggregation process in step S502. Figure 6 This description is based on the previous embodiment. Figure 6 The provided diagram illustrates the specific steps of the parameter dynamic aggregation process, including S601~S605.

[0089] S601. Determine the equipment importance score based on the key equipment and basic general equipment corresponding to the equipment groups in each edge node.

[0090] S602. Normalize the device importance scores in each edge node, and determine the original weight ratio of the edge node based on the product of the normalized device importance scores and the local training data of each edge node.

[0091] The traditional FedAvg algorithm calculates global model parameters by weighting the data volume of clients (edge ​​nodes) according to their proportion. In this application, the importance of the devices in the area where the edge nodes are located is taken into account, and the "node priority contribution" is superimposed on the "data volume contribution". That is, some edge nodes may not have a large amount of data, but they have a large number of devices and many key devices, and their scenario importance is high.

[0092] Assuming an edge node manages a device group consisting of x basic devices and y critical devices, the importance score of the device group for the i-th edge node is defined by Equation 13: Equation 13 This represents the importance score of the device group for the i-th edge node. This represents the i-th basic device. This represents the i-th critical device.

[0093] The ratio of basic equipment to critical equipment can be defined as 1 part for basic equipment and 2 parts for critical equipment.

[0094] To prevent individual edge nodes from having excessively high weights due to large differences in the number of devices, thus monopolizing the global model parameters, the importance scores of all edge nodes need to be normalized and constrained before calculating the aggregated global model parameters.

[0095] First, the device importance scores of the edge nodes are normalized to obtain the normalized device importance scores of the edge nodes, which are represented by Equation 14: Equation 14 in, This represents the device importance score of the normalized edge nodes (which is ultimately used to calculate the weights). This represents the lower bound of normalization (it can be set to 0.1 to avoid the score being too low and the weight being 0). This indicates the upper limit of normalization (can be set to 1 to avoid excessively high scores). This represents the minimum value of the device importance score corresponding to all edge nodes. This represents the maximum value of the device importance score corresponding to all edge nodes. If ,make .

[0096] Based on the normalized device importance score of the edge nodes, the original weight ratio of the edge nodes is further calculated.

[0097] The original weight percentage of edge nodes is determined by Equation 15: Equation 15 in, This represents the original weight percentage of the edge nodes. This represents the amount of local data at the i-th edge node.

[0098] S603. Determine the original weight percentage of a single edge node based on the original weight percentage of the edge node and the number of edge nodes.

[0099] S604. Normalize the original weight ratio of a single edge node, and dynamically aggregate the parameters based on the normalized original weight ratio of the single edge node and the local model parameters of the edge node to determine the global model parameters.

[0100] S605. Obtain the optimized student model using global model parameters.

[0101] Based on the original weight percentages of the edge nodes, the original weight percentage of each individual edge node is calculated. Specifically, the weight percentage of each individual edge node is determined using Equation 16: Equation 16 in, This represents the original weight percentage of a single edge node, and n represents the number of aggregation devices (edge ​​nodes) participating in this round.

[0102] After obtaining the original weight percentage of a single edge node, in order to further reflect the importance of edge nodes, the original weight percentage of a single edge node is further restricted according to the remaining constraints: like Then the constraints are set as follows: ; like Then the constraints are set as follows: ; like and Then the constraints are set as follows: constant.

[0103] The original weight percentages of individual edge nodes after constraint limitation are normalized again to obtain the normalized original weight percentages of individual edge nodes, which are expressed using Equation 17: Equation 17 in, This represents the original weight percentage of a single edge node after normalization.

[0104] Based on the normalized original weight proportions of individual edge nodes and the local model parameters of edge nodes, the parameters are dynamically aggregated to determine the global model parameters, which are then determined using Equation 18: Equation 18 in, Represents global model parameters. This represents the local model parameters uploaded by the i-th edge node after training.

[0105] By dynamically aggregating the original weight ratio of a single edge node with the local model parameters, the student model is updated, enabling further optimization of the student model. The first teacher model and the second teacher model are also updated, thus achieving the effect of feeding back from the edge nodes to the cloud server.

[0106] This application embodiment calculates the device importance score of edge nodes, further obtains the original weight ratio of individual edge nodes, and performs dynamic aggregation processing on the student model using the original weight ratio of individual edge nodes to obtain global model parameters, further optimizing the student model. While updating the teacher model, the student model is optimized, thereby further improving the accuracy of the student model in diagnosing the health of devices in edge nodes.

[0107] Figure 7 This is a flowchart illustrating another intelligent device health monitoring method provided in an embodiment of this application. Figure 7 The description is based on the above embodiments. Figure 7 The provided diagram illustrates the specific steps of the intelligent device health monitoring method, including: S701~S702.

[0108] S701: Receives device health monitoring results sent by the regional coordination server. The device health monitoring results are obtained by the edge node based on the student model to monitor the health of the device.

[0109] S702. When the equipment health monitoring result is uncertain, the uncertain equipment health is diagnosed by using the first teacher model and the second teacher model.

[0110] After obtaining the optimized student model, the edge node performs device health monitoring on the device through the student model and sends the device health monitoring results to the regional coordination server. If the device monitoring results contain uncertain types of device faults, the device health monitoring results are sent to the cloud server through the regional coordination server. The first teacher model and the second teacher model in the cloud server then perform a second diagnosis of the device health to determine the device health diagnosis result.

[0111] This application embodiment uses a student model at the edge node to monitor the health of the device. When the device health monitoring result is uncertain about the type of device failure, the device health is re-diagnosed using a first teacher model and a second teacher model in the cloud server to achieve device failure diagnosis and improve the accuracy of device health diagnosis.

[0112] Figure 8 This is a flowchart illustrating another intelligent device health monitoring method provided in this application embodiment. Figure 8 This refers to the process of the intelligent device health monitoring method in step S702. Figure 8 This description is based on the previous embodiment. Figure 8 The provided diagram illustrates the specific steps of the intelligent device health monitoring method, including: S801~S803.

[0113] S801. Diagnose the uncertain type of equipment using the first teacher model and the second teacher model respectively, and determine whether the diagnostic results of the first teacher model and the diagnostic results of the second teacher model are consistent.

[0114] S802. If the diagnostic results of the first teacher model are consistent with the diagnostic results of the second teacher model, the diagnostic results of the first teacher model shall be taken as the health diagnostic results of the equipment.

[0115] S803. If the diagnostic results of the first teacher model and the diagnostic results of the second teacher model are inconsistent, the historical diagnostic results corresponding to the first teacher model and the second teacher model are weighted and processed, and the weighted historical diagnostic results are used as the health diagnostic results of the equipment.

[0116] When an edge node detects an uncertain fault in the health monitoring results of a device, it reports the abnormal data to the cloud server through the regional coordinator. Two teacher models in the cloud server diagnose the abnormal data respectively. If the diagnosis results are consistent, the diagnosis result is output immediately. If the diagnosis results are inconsistent, the corresponding teacher model performs feature-weighted voting based on the historical diagnosis accuracy rate of the corresponding category as the confidence level, and then outputs the diagnosis result.

[0117] Optionally, the two teacher models on the cloud server also have the ability to grow. They continuously feed back to the teacher models by periodically collecting historical diagnostic data from edge nodes, thereby improving the diagnostic accuracy of the teacher models. This process is executed periodically and a fixed period can be set according to the operation of the cloud server.

[0118] This application embodiment uses a student model deployed on an edge node to perform health monitoring on the device. When the device health monitoring results indicate an uncertain fault, a teacher model is used again to diagnose the fault. This secondary diagnosis further improves the accuracy of device diagnosis and enhances the processing precision of both the teacher and student models.

[0119] Figure 9 This is a flowchart illustrating a teacher model construction process provided in an embodiment of this application. Figure 9 This refers to the process of dynamically aggregating parameters in step S502. Figure 9 yes Figure 2 The description is based on the provided embodiments. Figure 9 The provided diagram illustrates the specific steps of the parameter dynamic aggregation process, including S901~S902.

[0120] S901. Use time-series data to train the temporal convolutional network model and construct the first teacher model.

[0121] S902. Use image data to train the target detection model and construct a second teacher model.

[0122] Considering that time-series data typically includes vibrations, currents, etc., a first teacher model is obtained by training a temporal convolutional network model using the TCN model. Meanwhile, to address the challenge of processing heterogeneous machine data, particularly image data which often includes surface defects and thermal imaging, a second teacher model is obtained by training the image data using the YOLOv8 model.

[0123] In one possible example scenario, Figure 10 This is a flowchart illustrating the operation of a cloud server according to an embodiment of this application. Figure 10The provided diagram illustrates the following steps: In step S111, when the student model in the edge node performs health monitoring on the device, if the device health diagnosis result is normal, step S112 periodically sends the local data on whether the edge node has failed to the regional coordination server. The regional coordination server then transmits this data to the cloud server. In step S113, the cloud server further diagnoses the edge node. When a fault is diagnosed, in step S114, the cloud server sends a high-risk warning to the regional coordination server. Finally, in step S115, the regional coordination server decides whether to activate a self-protection mechanism to adjust the currently faulty edge node, achieving the purpose of cloud-edge coordinated scheduling. Alternatively, according to step S116, if the device health monitoring result is uncertain about the type of device fault, the first teacher model in step S117 and the second teacher model in step S118 need to be used again for device fault diagnosis. A comprehensive decision is made in step S119, and in step S120, the device diagnosis result of the teacher model yields the final diagnostic report.

[0124] This application embodiment considers the heterogeneous data of time-series data and image data, and uses different processing models to construct different teacher models. The purpose is to achieve the goal of the first teacher model processing time-series data and the second teacher model processing image data, which facilitates the processing of time-series data and image data in equipment health monitoring, and more intuitively and comprehensively solves the technical effect of equipment health monitoring.

[0125] Figure 11 This is a flowchart illustrating another intelligent device health monitoring method provided in an embodiment of this application. It is applied to a regional coordination server. According to... Figure 11 The provided diagram illustrates the specific steps of the intelligent device health monitoring method, including: S1101~S1105.

[0126] S1101: Receives student model and global model parameters sent by the cloud server, and deploys the global model parameters and student model to the edge node.

[0127] The first teacher model and the second teacher model are fused in the cloud server, and then the student model is obtained through knowledge distillation. The global model parameters contained in the first teacher model and the second teacher model are obtained. The global model parameters and the student model are deployed to the edge node to provide a reference for the next step of edge node device health monitoring.

[0128] S1102. Receive edge node data sent by edge nodes, send the edge node data to the cloud server, and receive edge node fault data returned by the cloud server for the edge node data. Maintain the edge nodes according to the edge node fault data. The edge node fault data is obtained by the cloud server based on a pre-built fault diagnosis model to diagnose edge node faults in the edge node data.

[0129] The edge nodes perform self-checks and obtain edge node data, which is then periodically sent to the regional coordination server. The regional coordination server then sends the edge node data to the cloud server, which diagnoses the edge nodes. If the cloud server detects faults or failures in the nodes, it sends the monitoring results to the regional coordination server, which then performs maintenance and adjustments on the edge nodes.

[0130] S1103. Using a pre-constructed global optimization objective, perform edge node adjustment analysis on the edge node state data to obtain the edge node global adjustment parameters, and send the edge node global adjustment parameters to the corresponding edge nodes so that the edge nodes can adjust their parameters using the edge node global adjustment parameters.

[0131] A global optimization target is constructed in the regional coordination server to characterize the global node coverage performance. Edge node state data corresponding to different states of edge nodes is sent to the regional coordination server. The global optimization target analyzes the edge node state data to determine if the edge node parameters need adjustment. When global parameter adjustment is required, the regional coordination server adjusts the edge node state data to obtain global adjustment parameters for the edge nodes. This allows edge nodes to adjust their parameters using these global adjustment parameters, achieving the goal of maintaining effective edge nodes globally.

[0132] S1104. Receive global parameter status information sent by multiple edge nodes. The global parameter status information is obtained by the edge nodes after performing edge federated learning training on the student model using global model parameters.

[0133] S1105. Send the global parameter status information to the cloud server, so that the cloud server can perform parameter aggregation processing based on the global parameter status information, update the first teacher model and the second teacher model, and obtain the optimized student model.

[0134] Based on the student model deployed on edge nodes, edge federated learning is used to train the student model, resulting in a lightweight student model. Edge nodes use global model parameters to train their local student models using edge federated learning, obtaining local training model parameters. Then, the edge nodes send these local training model parameters to the regional coordination server, which collects the local training model parameters from each edge node, obtains global parameter status information, and aggregates this information through a cloud server. This achieves edge federated learning training, enabling feedback from edge nodes to the cloud server, further updating the first and second teacher models.

[0135] This application embodiment transmits data to the edge node monitoring device from the perspective of the regional coordination server, transmits the student model generated by the cloud server to the edge node for deployment, and simultaneously receives local data and data trained through federation from the edge node, realizing the connection and data transmission between the edge node and the cloud server. In addition, when the cloud server detects that the edge node is faulty or has failed, the regional coordination server maintains the edge node, improves data processing efficiency, and solves the problem of edge node failure.

[0136] Figure 12 This is a flowchart illustrating a method for determining global adjustment parameters of edge nodes, as provided in an embodiment of this application. Figure 12 This refers to the process of determining the global adjustment parameters of the edge nodes in step S1102. Figure 12 This description is based on the previous embodiment. Figure 12 The provided diagram illustrates the specific steps for determining the global adjustment parameters of edge nodes, including: S1201~S1203.

[0137] S1201. Calculate the global metrics of the edge nodes based on the node status data.

[0138] S1202. Determine the theoretical optimal solution for edge nodes using the global optimization objective.

[0139] S1203. By comparing the global index with the theoretical optimal solution, the global adjustment parameters of the edge nodes are determined.

[0140] The global optimization objective is obtained in the following way: First, set the global optimization goal and basic sub-goals. The global optimization goal is... G g E represents the global node coverage performance. i Let represent the energy consumption of node i, where i∈A, A represents the set of edge nodes in the active state, and β represents the influence factor. The global node coverage performance is represented by Equation 19: Equation 19 in, Let A1 and A2 represent the set of general nodes and the set of key nodes that are in the active state, respectively, and R represent the coverage of the entire smart factory where the smart device is located.

[0141] The basic sub-objective is , This represents the global node coverage rate. This represents the basic coverage target. The global node coverage is represented by Equation 20: Equation 20 in, Let represent the coverage area of ​​the i-th active node, R represent the coverage area of ​​the entire smart factory where the smart devices are located, and x represent the number of nodes. Based on the global optimization objective and the basic sub-objectives, and according to the pre-set global objective constraints, the global optimization objective corresponding to the edge nodes is constructed.

[0142] Furthermore, a constraint is set: when the i-th edge node selects a neighbor node j, then neighbor node j is removed from the list of remaining nodes and neighbors. action i Indicates the current edge node, a j Represents a neighboring node, b j Represents the remaining nodes, neighnors m This represents the list of neighbors. Meanwhile, the node probability vector Pi indicates that any value has a bounded range: , where P min P represents the minimum probability vector of a node. max This represents the maximum probability vector of a node. The critical node coverage sets a minimum coverage requirement, satisfying the following relationship: ,in, This indicates the target covered by the critical node.

[0143] Finally, the residual energy constraint at the edge nodes yields the corresponding expression: Meanwhile, the sum of the node probability vectors is constrained to be 1: After obtaining the constraints, the global optimization objective is mathematically restricted to obtain a global optimization objective that satisfies the constraints.

[0144] Furthermore, the system first receives status information processing reports from each node and calculates current global metrics, such as coverage, coverage performance, and energy consumption. Then, based on pre-set global optimization objectives and constraints, it calculates the theoretical optimal solution. Finally, it compares the current metrics with the theoretical values ​​to determine the global adjustment parameters for the edge nodes. , , Dynamic adjustment , , Finally, the adjusted parameters are sent to the corresponding edge nodes.

[0145] This application embodiment calculates the global index of the edge node by setting a global optimization objective and constraints, uses the global optimization objective to determine the theoretical optimal solution of the edge node, and determines the global adjustment parameters of the edge node by comparing the optimal solution with the global index, and adjusts the parameters of the edge node.

[0146] In one possible example scenario, Figure 13 This is a schematic flowchart illustrating a method for determining global adjustment parameters of edge nodes, as provided in an embodiment of this application. Figure 13 The provided diagram shows that, firstly, step S131 determines the ideal state vector and step S132 determines the actual state vector.

[0147] The state vector is defined according to equation 21-25: Equation 21 Equation 22 Equation 23 Equation 24 Equation 25 in, The values ​​represent the global parameters that need to be adjusted. a1 and a2 represent the node state vector table and global node index of the theoretical solution, respectively. c1 and c2 represent the current actual node state vector table and global node index. Each vector in a1 and c1 ∈ {-1, 0, 1} corresponds to the fault, sleep, and active states of the node. a2 and c2 represent the global node index (E represents node energy consumption).

[0148] Further, step S133 calculates the similarity between a1, c1 and a2, c2, resulting in the relationships in equations 26-28: Equation 26 Equation 27 Equation 28 The state vector table contains discrete values, and similarity is calculated using Hamming distance h, where h represents the number of unequal corresponding elements in a1 and c1. The similarity is denoted by s1. The global node index vector s2 contains continuous values, and similarity is tested using the vector angle, where s2 represents the similarity of the vector angle. Finally, a linearly weighted comprehensive similarity s is calculated. This is used to balance the importance of discrete and continuous indices.

[0149] Further, in step S134, for the global parameter table... Each parameter in Independently designed PID adaptive regulator: Assuming the overall similarity of the target is The parameters in the k-th iteration The corresponding current overall similarity is The deviation is then expressed by equation 29: Equation 29 Step S135, for parameters Design a PID controller, whose fine-tuning amount is expressed by Equation 30: Equation 30 in, This indicates the fine-tuning amount for the corresponding parameter. This represents the difference between the target similarity and the similarity in the current k-th iteration; , , These represent the parameters of the proportional, integral, and derivative terms of the PID adaptive regulator, respectively. This represents the cumulative sum of deviations from iterations 0 to k.

[0150] In actual parameter tuning, different constant terms are set according to the importance of different parameters, such as coverage performance indicators. The PID constant term should be greater than the energy index. The PID constant term is adjusted to make its response speed faster, for example... Meanwhile, when the similarity deviation is large, the parameter of the proportional term should be increased and the parameter of the integral term should be decreased. Conversely, when the similarity deviation is small, the parameter of the proportional term should be decreased and the parameter of the integral term should be increased to achieve adaptive gain scheduling.

[0151] To ensure the stability of parameter adjustments, constraints must be set, and the adjusted parameters must be limited to a reasonable range. Within, according to step S136, we obtain equation 31: Equation 31 in This indicates the parameter value before adjustment. This indicates the adjusted parameter value. This indicates the amount of parameter fine-tuning.

[0152] The magnitude of a single parameter adjustment must be limited to the threshold. Within this range, the corresponding limiting conditions are: .

[0153] When similarity deviation Less than the noise threshold If this occurs, it is determined to be noise interference, and no parameter adjustment is made. That is, if... Then fine-tuning amount Additionally, the global parameter table middle The weight parameters sum to 1.

[0154] Figure 14 This is a schematic diagram of an edge node maintenance process provided in an embodiment of this application. It describes the edge node maintenance process in step S1102. Figure 14 Is Figure 11 The description is based on the corresponding embodiments. Figure 14 The provided diagram illustrates the specific steps for edge node maintenance, including: S1401~S1403.

[0155] S1401. Determine the global parameters and LA learning automata objects corresponding to the global nodes. Global nodes include device nodes, edge nodes, and cloud servers.

[0156] S1402. Update the state of the edge nodes in the global parameters based on the edge node fault data and the preset edge node state machine.

[0157] S1403. Determine the neighbor node selection strategy based on the edge node status in the global parameters, the LA learning automaton object, and the neighbor comprehensive evaluation function. The neighbor node selection strategy is used to select neighbor nodes to be activated. The neighbor nodes to be activated are used to back up the content of the edge nodes that have failed.

[0158] First, determine the global parameters, as shown in Equation 32: Equation 32 Wherein, id represents the unique identifier of the node, position represents the node's location in the smart factory, usually in (x,y) plane coordinates; state represents the node's state, namely active state, sleep state, and fault state; type represents the node type, divided into critical edge nodes and general edge nodes; energy represents the node's remaining energy, expressed as a percentage; cover represents the coverage area (the range that each node can cover for surrounding nodes cover={node1, node2...}); la represents the built-in learning automaton object; and neighbors represents the list of neighboring nodes.

[0159] The learning automaton object, la, is determined by equation 33: Equation 33 in, This represents a set of actions, where each action involves selecting a corresponding neighboring node. This represents the probability of selecting a corresponding action. The probability update rule is determined by the feedback after random environment interaction and is decided by the regional coordination layer (i.e., the regional coordination server). When an action is selected, if global coverage performance improves and energy consumption decreases, a reward is given, increasing the probability of selecting that action and decreasing the probability of selecting other actions; conversely, a penalty is given, decreasing the probability of selecting that action and increasing the probability of selecting other actions. The specific reward and penalty rules are determined according to Equations 34 and 35: Equation 34 Equation 35 in, This represents the reward / penalty function, which adjusts the probability table based on the feedback from the global objective and sub-objectives after selecting the corresponding neighbor nodes. The reward / penalty function is determined by equations 36, 37, and 38. Equation 36 Equation 37 Equation 38 in, This represents the global reward and punishment factor, which is dynamically adjusted by the regional coordinator based on the global objective status. , These represent global coverage and coverage performance update status, respectively. This indicates the energy saving situation. w i This represents the weight parameter, since coverage is the basic sub-objective (set as...). ),when Once the basic coverage target is achieved, the weight parameters are dynamically allocated and adjusted by the regional coordinator. L represents the delay penalty (L0). a L represents the actual measurement delay. t (This represents the maximum allowable delay threshold); where L is determined by Equation 44.

[0160] The node state machine is determined based on the node state change rules. Specific node states include active, faulty, and sleep states. The edge node states in the global parameters are updated based on edge node fault data and the preset edge node state machine states.

[0161] Neighbor node selection is primarily based on the probability vector table within the la-learning automaton object and the states of neighbor nodes. First, the overall quality of each neighbor node is calculated, and the node with the highest quality is selected. Considering the global optimization objective and sub-objectives, given consistent probability vectors, nodes in an active state, with high remaining energy and low load are preferentially selected; their overall quality is determined by Equation 39. Equation 39 Among them, s i Indicates the node state ( ), e i and z i These represent the remaining energy and load rate of neighboring nodes, respectively. This leads to the neighbor node selection strategy.

[0162] In one possible example scenario, Figure 15 This is a flowchart illustrating a method for determining a node state machine, as provided in an embodiment of this application. Figure 15 The provided diagram illustrates that the edge node state machine has three states: active, sleep, and fault. When an edge node is forced to sleep by the regional coordination server, its state changes from active to sleep. When a neighbor request occurs or the regional coordination server wakes it up, the edge node's state changes from sleep to active. If a fault occurs after the edge node's self-test, its state changes from sleep to fault. After manual re-testing and reset, the edge node's state changes from fault to active. When a node self-test reports a fault, the edge node's state changes from active to fault. The edge node's state changes according to the different states in its state machine.

[0163] In one possible example scenario, Figure 16 This is a flowchart illustrating a node maintenance process provided in an embodiment of this application. According to... Figure 16The provided diagram illustrates that when a node fails or becomes ineffective, the regional coordination server activates neighboring nodes and backs up critical data based on a neighbor selection strategy. First, in step S161, an edge node failure is detected through edge node self-checking or by the regional coordination server. Next, in step S162, if the edge node's self-check result is positive, it sends the edge node failure information to the regional coordination server. Then, in step S163, the regional coordination server selects the optimal backup node based on the neighbor node selection strategy. Finally, in step S164, it sends an activation request using either the failed node or the regional coordination server. In step S165, the neighboring node confirms activation and uploads the information to the regional coordination server. In step S166, the regional coordination server sends the failure information to the cloud server. Finally, in step S167, the cloud server sends a maintenance notification, completing the edge node maintenance.

[0164] Figure 17 This is a flowchart illustrating a federated training process provided in an embodiment of this application. It describes the federated training process for step S1104. Figure 17 Is Figure 11 The description is based on the corresponding embodiments. Figure 17 The provided diagram illustrates the specific steps of the federated training process, including S1701 to S1704.

[0165] S1701. Obtain the list of state attribute values ​​of edge nodes.

[0166] S1702. Based on the list of state attribute values, determine the set of candidate edge nodes to participate in federated training.

[0167] S1703. Based on preset bandwidth constraints, select edge nodes from the candidate edge node set to participate in federated training.

[0168] S1704. Send a command to the edge nodes participating in federated training to execute federated training, so that the edge nodes receiving the command can use the global model parameters to perform edge federated learning training on the student model to obtain global parameter state information.

[0169] Assume the set of edge nodes is , For the i-th edge node, the participation ratio in each round is... .

[0170] Before each round, the state attribute values ​​of each edge node are collected, resulting in a list of state attribute values. Based on this list, a set of candidate edge nodes for participating in federated training is determined. The state attribute value list includes real-time bandwidth, node score, and number of participations. The set of real-time bandwidths of all nodes is represented by Equation 40: Equation 40 The set representing the historical participation counts of each node is expressed by Equation 41: Equation 41 Edge nodes participating in federated training are selected from the candidate edge node set. Each edge node is scored based on different metrics, namely, remaining battery power, criticality, and local data distribution entropy. Nodes with high criticality, high battery power, and low entropy are prioritized for training. The metric set for each node is defined by Equation 42: Equation 42 If the total bandwidth exceeds the maximum bandwidth limit q, then TOP-K parameter sparsity processing needs to be applied to some clients to reduce the total bandwidth value. Edge nodes receiving commands use global model parameters to perform edge federated learning training on the student model to obtain global parameter state information.

[0171] In one possible example scenario, Figure 18 This is a flowchart illustrating another federated training process provided in an embodiment of this application. According to... Figure 18 The provided diagram illustrates the following steps: Step S181: The cloud server and edge nodes are initialized, and then the cloud server distributes a basic small model to the edge nodes. Step S182: The cloud server distributes global mode parameters to each edge node. Step S183: Local training is performed on the nodes according to a pre-set node selection strategy. Step S184: Differential encryption is performed using weight parameters, and the training nodes are compressed before uploading based on actual bandwidth conditions. Step S185: The cloud server aggregates and updates the model based on the node scoring and model weights, and checks whether the aggregated model meets the accuracy requirements. If the aggregated model does not meet the requirements, the current loop ends.

[0172] Figure 19 This is a schematic flowchart illustrating a process for determining a set of candidate edge nodes according to an embodiment of this application. It describes the process for determining the set of candidate edge nodes in step S1702. Figure 19 This description is based on the previous embodiment. Figure 19 The provided diagram illustrates the specific steps for determining the candidate edge node set, including: S1901~S1907.

[0173] S1901. The status attribute value list contains the remaining power evaluation index, edge node criticality evaluation index, and entropy value evaluation index of the edge node. The remaining power evaluation index, edge node criticality evaluation index, and entropy value evaluation index are used to perform priority evaluation processing on the pre-set index set to obtain the evaluation set.

[0174] In the list of status attribute values This represents the remaining power of the i-th node, obtained through the module API. Let m represent the criticality of the i-th node, expressed by Equation 43. Let m represent the number of critical devices covered by the node, k represent the types of devices included in the node, and N represent the total number of devices covered by the node.

[0175] Equation 43 Let represent the local data distribution entropy value of the i-th node, expressed by Equation 44.

[0176] Equation 44 The i-th node contains m types of devices, p(x) represents the proportion of data in the x-th type of device, and the entropy value reflects the distribution of data. A low entropy value usually represents more unique data.

[0177] Furthermore, regarding the set of indicators Priority evaluation: Treat each node's index vector as a point in space, and calculate the best and worst points. Evaluate the priority of each node by comparing its distance from these points. The resulting evaluation set is represented by Equation 45. Equation 45 S1902. Based on the pre-set initial screening conditions for edge nodes, determine whether the edge nodes in the evaluation set meet the initial screening conditions.

[0178] S1903. If an edge node does not meet the initial screening criteria, the edge node will be added to the blacklist.

[0179] S1904. If the edge nodes meet the initial screening conditions, store the edge nodes in the whitelist.

[0180] First, set the initial screening criteria for edge nodes, and add abnormal nodes (such as sudden changes in local data distribution entropy or real-time bandwidth) to a blacklist for isolation and observation. The mutation rate is then represented by Equation 46. Equation 46 in, Indicates the current value. Represents the historical average. Let the dictionary... For blacklisted devices, id represents the edge node number, and num represents the number of anomalies. Let... , These are the entropy value and the bandwidth blacklist fluctuation threshold, respectively, and the entropy value fluctuation situation in a certain round. Or bandwidth fluctuations If num++, then num++; if 3 consecutive rounds If all values ​​are normal, then `num` is decremented. When `num` ≥ 3, the edge node is added to the blacklist; when `num` < 3, the edge node is removed from the blacklist. Simultaneously, to prevent prolonged anomalies in the early stages from rendering blacklisted devices unusable later, a cleanup of the blacklist is performed every N training iterations (e.g., 20), reducing all `num` values ​​by half. .

[0181] S1905. Sort the edge nodes in the whitelist and determine whether there are any edge nodes among the top target number of edge nodes that meet the preset total bandwidth requirements.

[0182] The total bandwidth requirement is determined using Equation 47: Equation 47 in, Total bandwidth 'b' represents the parameter size of the edge node, and 'b' represents the parameter precision.

[0183] S1906. If none of the target number of edge nodes in the sorted order meet the total bandwidth requirement, the edge nodes will be sparsified.

[0184] S1907. If there is an edge node that meets the total bandwidth requirement among the top target number of edge nodes, add the edge node that meets the total bandwidth requirement to the candidate edge node set for participating in federated training.

[0185] When the total bandwidth At that time, suitable edge nodes are found for sparsification processing, selecting those nodes with high bandwidth, high participation frequency, and low score. This represents the selection priority. In each round, all nodes are sorted in descending order of selection priority, and the resulting set is... The sorted set can be represented by equation 48: Equation 48 First select The first n / c nodes are used as training nodes for this round, and their calculations are performed. Value, if Then, the nodes with lower priority need to be sparsified.

[0186] The calculation process of the sparsified parameters: Parameter definition: m is the total number of parameters for a single edge node (all edge nodes have the same number of parameters), b is the precision of the parameters (in this scheme, b=4, which means 4 bytes), and the size of a single edge node model is mb.

[0187] (1) Select TOP-K sparsification, for example, select 30%. After sparsification, the total number of parameters becomes parameter values ​​+ parameter indices (since the student model has many parameters, a single parameter index is stored using 4 bytes). That is, the model size after sparsification is: .

[0188] (2) Perform 16-bit quantization on the parameter values, and the total number of parameters becomes: .

[0189] (3) Huffman compression is performed on the parameter index, which ranges from [0, m]. Since the index values ​​of important parameters have strong clustering, the difference between adjacent parameters is usually very small, and the distribution is highly biased towards small values. Therefore, differential coding can be used for storage. The average code length is calculated using Equation 49. Furthermore, since the actual compression effect is related to the difference distribution, the compression range can be determined according to the historical actual weight parameters of each node in engineering applications, as expressed in Equation 50. Equation 49 Equation 50 in, and Let represent the probability of occurrence of the range of differences of type i and the length of the Huffman code, respectively. Indicates the average code length; This indicates that the final compression ratio range is ∈ [z1, z2].

[0190] (4) Take the upper limit of compression z2, expressed by equation 50, and then... After the collection Sparsification is performed on (rounded up) edge nodes. The final sparsification result is obtained. Equation 51 represents the final sparsification result: Equation 51 To give some low-scoring nodes a chance to participate in collaborative training, during the priority ranking process in the middle of training, a certain probability is provided for randomly selecting nodes. This indicates that the selection is performed according to the priority sorting list. Random selection is indicated by the condition for selecting a node body, as expressed in Equation 52: Equation 52 If, among the top-ranked edge nodes after sparsification, there is an edge node that meets the total bandwidth requirement, then that edge node will be added to the candidate edge node set for federated training.

[0191] This application embodiment obtains an evaluation set by prioritizing the indicator set, and then filters out high-priority edge nodes by setting filtering conditions. At the same time, it gives edge nodes with low evaluation scores an opportunity to be accessed, thereby achieving the technical effect of allowing as many edge nodes as possible to participate in federated training, while allowing high-scoring nodes to participate in more rounds.

[0192] Figure 20 This is a schematic flowchart of a device grouping process provided in an embodiment of this application. It describes the device grouping process prior to receiving edge node data sent by the edge node in step S1102. Figure 20 Is Figure 11 The description is based on the corresponding embodiments. Figure 20 The provided diagram shows the specific steps of the equipment grouping process, including: S2001~S2004.

[0193] S2001. Determine the device distance matrix based on the device nodes in the device layer.

[0194] S2002. Set the optimization target relationship for the device nodes in the device distance matrix according to the pre-set grouping rules. The optimization target relationship is obtained by using priority factors, the target relationship of assigning adjacent devices to the same group, the target relationship of each device group including multiple device types, and the target relationship of assigning key devices to different groups.

[0195] Assuming each equipment group can accommodate a maximum of S m The number of devices is determined by the processing power of the edge module; the device set is... , This represents the i-th device, whose attributes are defined as follows: , Construct the device distance matrix A, represented by Equation 53: Equation 53 in, Indicates the type of the i-th device (e.g., motor, camera, etc.); Indicates whether it is a critical device (1 indicates yes, 0 indicates no); A represents the device distance matrix. (i=j), (i≠j), (If devices i and j are adjacent), if devices are not adjacent, round down based on the ratio of their distances to the adjacent devices.

[0196] Furthermore, adjacent devices should be assigned to the same group as much as possible, with the objective being: Where E is the set of adjacent device pairs, determined by the device distance matrix A, i.e. .

[0197] The goal is to include as many device types as possible in each device group. ; Assign critical equipment to different groups as much as possible, with the following objective: The final goal is to organize the information as follows: ,in, , , This is a priority factor; different objectives have different priorities.

[0198] This represents a collection of key equipment. This indicates whether device i is assigned to group j; , indicating whether group j is a critical group (at least one critical device exists); , indicating whether group j has a device of type t.

[0199] S2003. Determine the priority constraints of device nodes in the device layer.

[0200] S2004. Based on the optimization objective relation and priority constraints, use a genetic algorithm or particle swarm optimization algorithm to solve the multi-objective nonlinear programming problem and obtain the equipment grouping results. The equipment grouping results are used to determine the correspondence between edge nodes and equipment groups.

[0201] The priority constraint for device nodes in the device layer is a priority limit, prioritizing the allocation of critical devices to different groups, and ensuring that each device group includes as many device types as possible. And the number of devices in each group is limited: The constraints for the key equipment group are: The equipment type coverage constraint is: Equipment can only be assigned to one equipment group, while critical equipment can be assigned to multiple equipment groups, as expressed by equations 54 and 55: Equation 54 Equation 55 Based on the above-mentioned constraints and the final objective, it can be transformed into a multi-objective nonlinear programming problem. Genetic algorithms, particle swarm optimization algorithms, etc. can be used to solve the multi-objective nonlinear programming problem to obtain the grouped equipment and achieve the purpose of grouping the equipment.

[0202] Figure 21 This is a flowchart illustrating another intelligent device health monitoring method provided in this application embodiment. It is applied to edge nodes. According to... Figure 21The provided diagram illustrates the specific steps of the intelligent device health monitoring method, including: S2101~S2104.

[0203] S2101: Receive global model parameters and student models from the regional coordination server, deploy student models and global model parameters on edge nodes, and generate student models by knowledge distillation by the cloud server through pre-built first teacher models and second teacher models.

[0204] S2102. Based on the global model parameters and the edge federated learning algorithm, the student model is trained locally to obtain the local training model parameters. The local training model parameters are sent to the regional coordination server, which then sends the local training model parameters of each edge node as global parameter status information to the cloud server for parameter aggregation processing, so as to update the first teacher model and the second teacher model, and obtain the optimized student model.

[0205] Heterogeneous data is processed by the first and second teacher models in the cloud server, and the student model is obtained by knowledge distillation and fusion. The student model is then trained on local data of edge nodes using edge federated learning. This approach has low bandwidth requirements for the smart factory, high training efficiency, and good model quality, thus enabling the optimization of the student model.

[0206] S2103. Collect edge node data and send it to the regional coordination server, so that the regional coordination server sends the edge node data to the cloud server, so that the cloud server performs edge node fault diagnosis on the edge node data based on the pre-built fault diagnosis model to obtain edge node fault data.

[0207] S2104. Perform health monitoring on the equipment based on the student model to obtain the equipment health monitoring results.

[0208] This approach achieves device health monitoring from the perspective of edge nodes. First, a student model is obtained by merging the first and second teacher models in the cloud server. This student model is then deployed to the edge nodes, where federated learning is used to train and optimize it. The student model is then used for device health diagnosis. Simultaneously, the edge nodes perform self-checks and send their status information to the cloud server. When the cloud server detects a fault or failure in an edge node, it initiates maintenance and adjustments through a regional coordination server. Furthermore, by monitoring device health through the edge nodes, device health monitoring results are obtained, enabling the detection of device faults. This significantly improves cloud-edge collaboration efficiency and provides the capability for rapid dynamic adjustment of global indicators.

[0209] In one possible example scenario, Figure 22 This is a schematic diagram illustrating a process for dynamically updating a blacklist, as provided in an embodiment of this application.Figure 22 The provided diagram illustrates the process for edge nodes. Step S221: First, initialize the blacklist dictionary H, then set the number of loops; determine if the modulo elimination condition is met; Step S222: If the condition is met, clear the blacklist content; Step S223: If the condition is not met, Step S224: The edge node meets the blacklist entry condition, and then add the edge node to the blacklist; Step S225: Update the dictionary. By setting the blacklist and dynamically selecting nodes, training efficiency can be improved and the bandwidth requirements of the factory can be reduced.

[0210] In one possible example scenario, Figure 23 This is a schematic diagram illustrating a process for determining candidate edge nodes provided in an embodiment of this application. According to... Figure 23 The provided diagram shows the following steps: First, in step S231, obtain the list of state attribute values ​​for each edge node and determine whether the initial screening conditions are met. Second, in step S232, if the initial screening conditions are met, add the edge node to the whitelist, prioritize the edge nodes in the whitelist, select the top target number of edge nodes, and determine whether the total bandwidth requirement is met. Third, in step S233, if the initial screening conditions are not met, move the edge node to the blacklist. Fourth, in step S234, if the total bandwidth requirement is not met, initiate edge node sparsity processing. Fifth, in step S235, if the total bandwidth requirement is met, perform federated training and update the list of state attribute values ​​for each node.

[0211] Figure 24 This is a flowchart illustrating a process for collecting edge nodes according to an embodiment of this application. It focuses on the process of collecting edge nodes in step S2103. Figure 24 Is Figure 21 The description is based on the corresponding embodiments. Figure 24 The provided diagram illustrates the specific steps involved in collecting edge nodes, including: S2401~S2405.

[0212] S2401. Perform a self-check on the edge nodes to obtain the edge node status indicators.

[0213] S2402. Collect neighbor information and establish a neighbor node table for each edge node.

[0214] S2403. Use a pre-built node state machine to process the state changes of edge node state indicators.

[0215] S2404. Use a node selection strategy to select the corresponding target neighbor node for the edge node after the state change processing.

[0216] S2405. Collect edge node data of neighboring nodes based on the target neighboring nodes.

[0217] Based on the constructed student model and edge nodes, firstly, all edge nodes perform self-checks to confirm that their hardware is functioning correctly; then, they broadcast "HI" messages, collect neighbor information, and build a neighbor node table for each edge node; finally, the LA probability vector table is initialized, where... Indicates the node type (e.g., critical nodes are set to 1.5, and general nodes are set to 1). The remaining energy at a node is represented by Equation 56: Equation 56 Then, collect the status indicators of all edge nodes, such as load rate and power consumption; then, perform state changes according to the node state machine; after the state change, select neighboring nodes according to the edge node selection strategy and execute the edge node conversion and coverage task; next, update the probability table according to random environment feedback; finally, collect the edge node data of neighboring nodes according to the target neighboring nodes and report the processing report to the regional coordination server.

[0218] This application embodiment performs self-checks on edge nodes to obtain edge node status indicators, uses a node state machine to process the state changes of edge node status indicators, selects corresponding target neighbor nodes, and then collects edge node data from neighbor nodes to achieve the purpose of collecting edge nodes. At the same time, by designing a node state machine and a neighbor node selection strategy, each edge node can autonomously adjust its state based on global indicators.

[0219] Figure 25 This is a schematic diagram of the structure of an intelligent device health monitoring device provided in an embodiment of this application. The intelligent device health monitoring device may include a processor 2501 and a memory 2502 storing computer program instructions.

[0220] Specifically, the processor 2501 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0221] Memory 2502 may include mass storage for data or instructions. For example, and not limitingly, memory 2502 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. In one instance, memory 2502 may include removable or non-removable (or fixed) media, or memory 2502 may be non-volatile solid-state memory. Memory 2502 may be internal or external to the integrated gateway disaster recovery device.

[0222] In one instance, memory 2502 may be read-only memory (ROM). In one instance, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0223] Memory 2502 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Therefore, generally, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to the first aspect of this disclosure.

[0224] The processor 2501 reads and executes computer program instructions stored in memory 2502 to achieve... Figure 25 The intelligent device health monitoring method in the illustrated embodiment.

[0225] In one example, the intelligent device health monitoring device may also include a communication interface 2503 and a bus 2504. For example, Figure 25 As shown, the processor 2501, memory 2502, and communication interface 2503 are connected through bus 2504 and complete communication with each other.

[0226] The communication interface 2503 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0227] Bus 2504 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 2504 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0228] This intelligent device health monitoring device can execute the online data traffic billing method described in this application embodiment based on currently blocked spam SMS messages and SMS messages reported by users, thereby achieving a combination of... Figures 1-24 The method described is an intelligent device health monitoring method.

[0229] Furthermore, in conjunction with the intelligent device health monitoring method in the above embodiments, this application embodiment can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the intelligent device health monitoring methods in the above embodiments.

[0230] This application also provides a computer program product, including a computer program, which, when executed, implements any of the intelligent device health monitoring methods described in the above embodiments.

[0231] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0232] The functional blocks shown in the above-described block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, read-only memory (ROM), flash memory, erasable read-only memory (EROM), floppy disks, compact disc read-only memory (CD-ROM), optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0233] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0234] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0235] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method of intelligent device health monitoring, the method comprising: Be applied to cloud server, comprising: Knowledge distillation is carried out through the first teacher model and the second teacher model, and a student model is generated; The student model and global model parameters are sent to the regional coordination server, so that the regional coordination server deploys the student model and global model parameters to the edge node, and the edge node monitors the health of the device; Global parameter state information sent by the regional coordination server is received, parameter aggregation processing is carried out based on the global parameter state information, the first teacher model and the second teacher model are updated, and the optimized student model is obtained, and the global parameter state information is forwarded by the regional coordination server after the edge node trains the student model by using the global model parameters; Edge node data sent by the regional coordination server is received, edge node fault diagnosis is carried out on the edge node data based on the pre-constructed fault diagnosis model, and edge node fault data is obtained; The edge node fault data is sent to the regional coordination server, so that the regional coordination server maintains the edge node according to the edge node fault data.

2. The method of claim 1, wherein, The knowledge distillation is carried out through the first teacher model and the second teacher model, and a student model is generated, comprising: Feature extraction is carried out on time series data through the first teacher model to obtain corresponding first teacher features, feature extraction is carried out on image data through the second teacher model to obtain corresponding second teacher features, and feature extraction is carried out on multi-modal input data through a student training model to obtain corresponding student training features; The first teacher features and the second teacher are projected respectively to obtain corresponding first teacher projection features and second teacher projection features; The first teacher projection features and the second teacher projection features are weighted and fused by using pre-set modal fusion weights to obtain teacher fusion features; The mean square error loss value, the relationship matrix similarity and the dynamic temperature parameter are determined based on the teacher fusion features and the student training features; The loss function value is determined according to the mean square error loss value, the relationship matrix similarity and the dynamic temperature parameter; The student model is generated by using the loss function value to train the student training model.

3. The method of claim 1, wherein, The parameter aggregation processing is carried out based on the global parameter state information, the first teacher model and the second teacher model are updated, and the optimized student model is obtained, comprising: The global parameter state information is allocated with corresponding edge node aggregation weight proportion by using pre-set device importance score of each edge node; The global parameter state information is dynamically aggregated according to the edge node aggregation weight proportion, the first teacher model and the second teacher model are updated, and the optimized student model is obtained.

4. The method of claim 3, wherein, The global parameter state information is dynamically aggregated according to the edge node aggregation weight proportion, and the optimized student model is obtained, comprising: The device importance score is determined according to the key device and the basic device corresponding to the device group in each edge node; normalizing the device importance scores of the edge nodes, and determining the original weight proportion of each edge node according to the product of the normalized device importance scores and the local training data of the edge node; determining the original weight proportion of each edge node based on the original weight proportion of the edge node and the number of edge nodes; normalizing the original weight proportion of each edge node, and performing parameter dynamic aggregation on the normalized original weight proportion of each edge node and the local model parameters of the edge node to determine the global model parameters; obtaining an optimized student model by using the global model parameters.

5. The method of claim 4, wherein, The device importance scores are determined by the following method: wherein, represents a device importance score, represents a base device in a corresponding device group in the edge node, represents a key device in a corresponding device group in the edge node; The normalized device importance scores are determined by the following method: wherein, denotes a normalized device importance score, denotes a lower normalization limit, denotes an upper normalization limit, denotes a minimum of the device importance scores of all the edge nodes, denotes a maximum of the device importance scores of all the edge nodes. The original weight proportion of the edge node is determined by the following method: wherein, represents the proportion of the original weight of the edge node, represents the local data volume of the i-th edge node; The original weight proportion of each edge node is determined by the following method: wherein, represents the original weight proportion of a single said edge node; The normalized original weight proportion of each edge node is determined by the following method: wherein, represents the proportion of the original weight of the single edge node after normalization; The global model parameters are determined by the following method: wherein, denotes global model parameters, denotes local model parameters.

6. The method of claim 1, wherein, The method further comprises: receiving a device health monitoring result sent by the regional coordination server, the device health monitoring result being obtained by the edge node based on the student model; in the case that the device health monitoring result is of an uncertain type, diagnosing the device health of the uncertain type by the first teacher model and the second teacher model.

7. The method of claim 6, wherein, The diagnosis of the device health of the uncertain type by the first teacher model and the second teacher model comprises: diagnosing the device of the uncertain type by the first teacher model and the second teacher model respectively, and determining whether the diagnosis results of the first teacher model and the second teacher model are consistent; in the case that the diagnosis results of the first teacher model and the second teacher model are consistent, taking the diagnosis result of the first teacher model as the health diagnosis result of the device; in the case that the diagnosis results of the first teacher model and the second teacher model are inconsistent, performing weighted processing on the historical diagnosis results corresponding to the first teacher model and the second teacher model according to the historical diagnosis result accuracy, and taking the weighted historical diagnosis result as the health diagnosis result of the device.

8. The method of claim 1, wherein, The first teacher model and the second teacher model are determined by the following method: training a time convolution network model by using a time series data training dataset to construct the first teacher model; training a target detection model by using an image data training dataset to construct the second teacher model.

9. An intelligent device health monitoring method, characterized by, applied to a regional coordination server, comprising: receiving a student model and global model parameters sent by a cloud server, and deploying the global model parameters and the student model to edge nodes; receive edge node data sent by the edge node, send the edge node data to the cloud server, and receive edge node fault data returned by the cloud server for the edge node data, and maintain the edge node according to the edge node fault data, wherein the edge node fault data is obtained by the cloud server based on a pre-constructed fault diagnosis model for edge node fault diagnosis on the edge node data; perform edge node adjustment analysis on the edge node state data using a pre-constructed global optimization objective to obtain edge node global adjustment parameters, and send the edge node global adjustment parameters to the corresponding edge node to enable the edge node to perform parameter adjustment using the edge node global adjustment parameters; receive global parameter state information sent by a plurality of edge nodes, the global parameter state information being obtained by the edge nodes after performing edge federated learning training on the student model using the global model parameters; send the global parameter state information to the cloud server to enable the cloud server to perform parameter aggregation processing based on the global parameter state information, update the first teacher model and the second teacher model, and obtain an optimized student model.

10. The method of claim 9, wherein, The edge node global adjustment parameters obtained by performing edge node global adjustment analysis on the edge node state data using a pre-constructed global optimization objective include: calculating a global index of the edge node according to node state data; determining a theoretical optimal solution of the edge node using a global optimization objective; comparing the global index and the theoretical optimal solution to determine edge node global adjustment parameters.

11. The method of claim 10, wherein, The global optimization objective is obtained by: setting a global optimization objective and a basic sub-objective; Wherein, the global optimization target is , G g represents the global node coverage performance, E i represents the node i energy consumption, i∈A, A represents the edge node set in the active state, and β represents an influence factor. wherein, cover1 and cover2 represent the coverage of the ith activated node (cover1 and cover2 represent the general node and the key node respectively, and the key node has a larger weight in the coverage performance index), A1 and A2 represent the set of general nodes and the set of key nodes in the activated state, and R represents the coverage of the smart factory where the entire smart device is located; The base sub-target is , represents the global node coverage, represents the base coverage target; , represents the coverage range of the ith activated state node, R represents the coverage range of the smart factory where the entire smart device is located, and x represents the number of nodes; constructing the global optimization objective corresponding to the edge node based on the global optimization objective and the basic sub-objective according to pre-set global objective constraint conditions.

12. The method of claim 9, wherein, The maintenance of the edge node according to the edge node fault data includes: determining global parameters and LA learning automaton objects corresponding to global nodes, the global nodes including device nodes, edge nodes, and a cloud server; updating the state of the edge node in the global parameters according to the edge node fault data and a pre-set edge node state machine; determining a neighbor node selection strategy according to the state of the edge node in the global parameters, the LA learning automaton object, and a neighbor comprehensive evaluation function, the neighbor node selection strategy being used to select a neighbor node to be activated, the neighbor node to be activated being used to make content backup for the edge node that has failed.

13. The method of claim 9, wherein, Before receiving the global parameter state information sent by the plurality of edge nodes, the method further includes: obtaining a list of state attribute values of the edge node; determining a candidate edge node set participating in federated training according to the list of state attribute values; selecting edge nodes participating in federated training from the candidate edge node set based on a pre-set bandwidth constraint condition; sending a command of performing federated training to the edge node participating in the federated training, so that the edge node receiving the command performs edge federated learning training on the student model by using the global model parameters to obtain the global parameter state information.

14. The method of claim 13, wherein, The state attribute value list includes a remaining power evaluation index, an edge node criticality evaluation index, and an entropy value evaluation index of the edge node. The method further includes: performing priority evaluation processing on a pre-set index set by using the remaining power evaluation index, the edge node criticality evaluation index, and the entropy value evaluation index to obtain an evaluation set; determining whether the edge nodes in the evaluation set meet the pre-set preliminary screening condition of the edge node; in a case where the edge node does not meet the preliminary screening condition, storing the edge node in a blacklist; in a case where the edge node meets the preliminary screening condition, storing the edge node in a whitelist; performing sorting processing on the edge nodes in the whitelist, and determining whether there is an edge node meeting a pre-set total bandwidth requirement in a target number of edge nodes in the front of the sorting; in a case where there is no edge node meeting the total bandwidth requirement in the target number of edge nodes in the front of the sorting, performing sparsification processing on the edge nodes; in a case where there is an edge node meeting the total bandwidth requirement in the target number of edge nodes in the front of the sorting, adding the edge node meeting the total bandwidth requirement to a candidate edge node set participating in federated training.

15. The method of claim 14, wherein, The edge node criticality evaluation index is determined by the following formula: wherein b i2 represents the criticality evaluation index of the i-th edge node, m represents the number of critical devices covered by the edge node, k represents the device types contained in the edge node, and N represents the total amount of devices covered by the edge node; The entropy value evaluation index is determined by the following method: wherein b i3 represents the data distribution entropy value of the i-th edge node, the i-th edge node contains m types of devices, and p(x) represents the data volume proportion of the x-th type of device. The total bandwidth requirement is determined by the following method: wherein, is the total bandwidth, denotes the parameter size of the edge node, b denotes the parameter precision.

16. The method of claim 9, wherein, Before receiving the edge node data sent by the edge node, the method further includes: determining a device distance matrix according to device nodes in a device layer; setting an optimization target relationship formula for the device nodes in the device distance matrix according to a pre-set grouping rule, the optimization target relationship formula being obtained by using a priority factor, a same group target formula for adjacent device allocation, a device type target formula for each device group including multiple device types, and a different group target formula for key device allocation; determining a priority constraint condition of the device nodes in the device layer; solving a multi-objective nonlinear programming problem by using a genetic algorithm or a particle swarm optimization algorithm based on the optimization target relationship formula and the priority constraint condition to obtain a device grouping result, the device grouping result being used to determine a corresponding relationship between the edge node and the device grouping.

17. An intelligent device health monitoring method, characterized by, The method is applied to an edge node and includes: receiving a global model parameter and a student model issued by a regional coordination server, and deploying the student model and the global model parameter on the edge node, the student model being generated by a cloud server through knowledge distillation of a pre-constructed first teacher model and a second teacher model. According to the global model parameters and the edge federated learning algorithm, the student model is locally trained to obtain local training model parameters, and the local training model parameters are sent to the regional coordination server, so that the regional coordination server sends the local training model parameters of each edge node to the cloud server as the global parameter state information for parameter aggregation processing, to update the first teacher model and the second teacher model, and obtain an optimized student model; Collect edge node data and send it to the regional coordination server, so that the regional coordination server sends the edge node data to the cloud server, so that the cloud server performs edge node fault diagnosis on the edge node data based on a pre-constructed fault diagnosis model to obtain edge node fault data; According to the student model, the health of the device is monitored to obtain the device health monitoring result of the device.

18. The method of claim 17, wherein, The edge node data is collected and sent to the regional coordination server, so that the regional coordination server sends the edge node data to the cloud server, so that the cloud server performs edge node fault diagnosis on the edge node data based on a pre-constructed fault diagnosis model to obtain edge node fault data; The edge node is self-checked to obtain edge node state indicators; Collect neighbor information to establish a neighbor node table for each edge node; The edge node state indicators are processed for state changes using a pre-constructed node state machine; The edge node is selected to select the corresponding target neighbor node after state change processing using a node selection strategy; According to the target neighbor node, the edge node data of the neighbor node is collected.

19. An intelligent device health monitoring device, characterized by The device includes a processor and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the intelligent device health monitoring method of any one of claims 1-8 or 9-16 or 17-18.

20. A computer-readable storage medium, characterized in that, The computer storage medium stores computer program instructions, and the computer program instructions are executed by the processor to implement the intelligent device health monitoring method of any one of claims 1-8 or 9-16 or 17-18.

21. A computer program product, characterised in that, The computer program is executed by the processor to implement the intelligent device health monitoring method of any one of claims 1-8 or 9-16 or 17-18.

Images