A cloud-edge collaborative health state diagnosis method and system for multiple shearer

Abstract

The application discloses a cloud-edge collaborative health state diagnosis method and system for multiple coal mining machines, comprising an edge device and a cloud server, the edge device collects multi-modal data of the coal mining machine and runs an edge model, the edge model is constructed based on a Transformer, and fuses a deep residual shrinkage network, a multi-scale linear attention mechanism and an LSTM decoding layer. The cloud server receives and aggregates edge model parameters of each edge device through a federated learning mechanism, thereby updating a global model, and distributes the optimized global model parameters to each edge device. The application introduces an edge device model upload relevance judgment mechanism and a cloud server model download weight distribution mechanism, effectively reducing communication overhead and improving model training efficiency. While protecting the privacy of edge data of each coal mining machine, the application realizes edge data sharing and cloud model sharing among multiple coal mining machines, and has good universality.

Description

TECHNICAL FIELD

[0001] The present application relates to a cloud-edge collaborative health state diagnosis method and system for multiple coal mining machines, and belongs to the technical field of intelligent coal mining machines. BACKGROUND

[0002] With the popularity of Internet of Things devices, the amount of data is growing explosively. Transmitting all data to the cloud for processing not only has high costs, but also has privacy and security problems. Edge computing allows data processing near the source of data generation, and federated learning allows multiple devices to collaboratively train models without sharing raw data. The combination of these two technologies can better utilize data on edge devices while protecting user privacy. However, existing federated learning methods have limitations. These methods are mainly designed for single-modal data and are difficult to effectively handle multi-modal data. In practical applications, edge devices can collect multiple types of data, such as temperature, pressure, vibration, etc. These different types of data are collectively referred to as multi-modal data. Effectively fusing and processing multi-modal data is crucial to improving model performance.

[0003] In many practical application scenarios, model training needs to use data from multiple data owners. Each data owner has a part of the training sample data. These data owners expect to jointly use each other's data to train the model uniformly. At the same time, they do not want to provide their own data to other data owners to prevent their data from being leaked.

[0004] The computing resources of edge devices are usually limited, so lightweight and efficient models need to be designed to adapt to these environments. At the same time, the model needs to have the ability to capture long-distance dependencies in order to effectively process time series data. In the federated learning framework, different edge devices may produce model parameters of different qualities, and there is a lack of effective mechanism to judge the correlation of model parameters in existing technologies, so the error of model prediction is large. SUMMARY

[0005] The present application aims to solve the problems of data sharing difficulty and low model training efficiency in the prior art, and provides a cloud-edge collaborative health state diagnosis method and system for multiple coal mining machines. This method and system can realize collaborative training of models on multiple devices without sharing the data of all edge devices, effectively protect the privacy of each edge device, and effectively fuse and process multi-modal data.

[0006] This invention provides a cloud-edge collaborative health status diagnosis method for multiple coal mining machines, with the edge device as the execution subject, including the following steps: collecting multimodal data of the coal mining machine's operating status; extracting features from the multimodal data to obtain single-modal data; receiving a global model sent by a cloud server; training the global model using the single-modal data to obtain an edge model; sending edge model parameters that meet preset conditions to the cloud server; receiving updated global model parameters sent by the cloud server; updating the edge model based on the updated global model parameters; and using the edge model to predict and obtain coal mining machine lifespan data.

[0007] Prior to this, the edge model is built on an improved Transformer block, which includes multiple pooling layers, convolutional layers, several deep residual shrinking network layers, several Itransformer Block layers, linear layers, normalization layers, long short-term linear memory layers, and fully connected layers connected in sequence. Each Itransformer Block layer includes a linear attention layer, a first Norm&Add layer, a feedforward layer, and a second Norm&Add layer connected in sequence.

[0008] Prior to this, the multi-scale linear attention mechanism includes multiple linear transformations at different scales; the deep residual shrinking network includes multiple residual blocks, each of which includes at least one shrinking operation.

[0009] Firstly, feature extraction is performed on the multimodal data to obtain single-modal data, including: firstly, segmenting the multimodal data; then performing linear mapping on the segmented data; and finally embedding modal information to obtain single-modal data.

[0010] Prioritizes meeting preset conditions, including: the model correlation between the edge model and the global model is higher than a first threshold; the model update trend correlation between the edge model and the global model is higher than a second threshold; wherein, the formula for calculating the model correlation D between the edge model and the global model is: ,in, Indicates the dimension of the edge model parameters. In the edge model, the first The model parameter values ​​of each edge model parameter Represents the first in the global model The model parameter values ​​of each global model parameter; calculate the correlation R between the update trends of the edge models and the global model: ,in, It is a value-finding function. Indicates the dimension of the edge model parameters; Represents the first edge in the edge model The change values ​​of the parameters of each edge model Represents the first in the global model If the model parameter changes, and If the symbols are the same, then and The value is 1, otherwise and The value is 0.

[0011] First, sort the edge models according to their contribution and determine their weights, including: calculating the contribution score (Score(i)) of the edge model parameters. ,in, It is the correlation between the update trends of the i-th edge model and the global model. , and It is the weighting coefficient, D i This represents the relevance of the i-th model, where ε is the minimum value. It is the logarithm of the total number of multimodal data points on the coal mining machine's operating status collected by the i-th edge device; calculate the weight of the i-th edge device. : ,in, is the contribution of the i-th edge device, e=2.718 is a mathematical constant, k is the total number of edge devices, and t is the control factor.

[0012] Prior to this, a cloud-edge collaborative health status diagnosis method for multi-coal mining machines, with a cloud server as the execution entity, includes: receiving edge model parameters that meet preset conditions sent by edge devices; aggregating the edge model parameters and updating the global model in the cloud server; and sending the updated global model parameters to the edge devices.

[0013] Prior to this, a cloud-edge collaborative health status diagnosis method for multiple coal mining machines includes the following steps: collecting multimodal data of the coal mining machine's operating status using edge devices, and extracting features from the multimodal data to obtain single-modal data; receiving a global model sent by a cloud server using the edge devices; training the received global model using the edge devices based on the single-modal data to obtain an edge model; aggregating the model parameters of the edge model using the cloud server and updating the global model in the cloud server; sending the updated global model parameters to the edge devices using the cloud server; updating the edge model using the edge devices based on the updated global model parameters; and predicting the coal mining machine's lifespan data using the edge model.

[0014] Prior to this, a cloud-edge collaborative health status diagnosis system for multiple coal mining machines includes the following modules: a feature extraction module, used to collect multimodal data of the coal mining machine's operating status using edge devices, and extract features from the multimodal data to obtain single-modal data; a model parameter acquisition module, used to train a received global model using edge devices based on the single-modal data to obtain an edge model; aggregate the model parameters of the edge model using a cloud server to update the global model; send the updated global model parameters to the edge devices using the cloud server; update the edge model using the edge devices based on the updated global model parameters; and a model prediction module, used to predict coal mining machine lifespan data using the edge model. Prior to this, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory. When the processor executes the program, it can implement any step of the method described in the present invention. Furthermore, the present invention provides a computer-readable storage medium storing a computer program implementing the method described in the present invention.

[0015] The beneficial effects achieved by this invention are as follows: Based on an edge computing architecture, this invention distributes computing tasks across edge devices and a cloud server. Edge devices are responsible for receiving the global model sent by the cloud server and collecting multimodal data from the environment for preliminary feature extraction. The edge devices then use local data to train and optimize the global model, generating preliminary model parameters. The cloud server aggregates edge model parameters from multiple edge devices and updates the global model through aggregation operations.

[0016] The cloud server sends the updated global model parameters back to the edge device for updating and optimizing the edge model, enabling model iteration. This invention can effectively process multimodal data and improve the accuracy of fault diagnosis. This invention employs federated learning technology to achieve model training while protecting data privacy. This invention reduces communication costs and computing power consumption through collaborative work between the edge device and the cloud server.

[0017] To address the problems of excessive communication overhead, low model training efficiency, and limited system scalability in existing federated learning methods, this invention provides a cloud-edge collaborative health status diagnosis method for multi-mining machines. This method has the following characteristics: After the edge device completes edge model training, it evaluates the impact of edge model updates on global model updates and only sends edge model parameters that meet preset conditions to the cloud server, thereby reducing communication overhead. After receiving the global model parameters from the cloud server, the edge device updates its edge model based on these parameters. The system design ensures that the locally trained edge model parameters are only uploaded to the cloud server and participate in the global model update when the updated edge model parameters meet predetermined conditions.

[0018] This method utilizes federated learning to achieve collaborative model training without sharing raw data. By training the model locally on edge devices, the amount of data transmitted to the cloud server is reduced, effectively lowering communication overhead. Simultaneously, the use of Transformer networks for feature extraction and sequence data modeling improves model training efficiency. Optimizing the model structure and training algorithm enables efficient model training even on resource-constrained edge devices, thus addressing the issue of system scalability limitations. After each edge device completes its edge model training, the following steps are performed: calculating the model correlation between the locally trained edge model and the received global model. When the model correlation reaches a predetermined threshold, the model update trend correlation between the edge model and the global model is further calculated to assess whether the edge model's update direction aligns with the global model's update trend. Only when both the model correlation and the model update trend correlation meet the corresponding threshold conditions will the edge device upload the model parameters of its locally trained edge model to the cloud server for global model updates. This dual-screening mechanism ensures that only local parameters that significantly contribute to global model optimization and have a consistent update direction participate in the global model's iterative process. This invention, by reducing communication overhead and improving model training efficiency, can expand the scale of federated learning systems, accommodating more member devices to participate in model training, thereby better meeting the needs of various application scenarios. This invention can be applied to various scenarios requiring collaborative model training among multiple data owners. Attached Figure Description

[0019] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a block diagram illustrating the principle of the federated learning model.

[0021] Figure 2 This is a flowchart of the federated learning method.

[0022] Figure 3 This is a flowchart of the data conversion process at the edge layer.

[0023] Figure 4 This is a diagram of the improved Transformer algorithm architecture.

[0024] Figure 5 This is a flowchart of data transmission for a coal mining machine. Detailed Implementation

[0025] See Figure 1This application discloses a cloud-edge collaborative health status diagnosis method and system for multi-coal mining machines. The core working principle of the system is based on the edge computing architecture, which distributes computing tasks on edge devices and cloud servers.

[0026] The edge device receives the global model from the cloud server, then collects multimodal data from the field environment and performs preliminary feature extraction on this data. Next, based on the received global model, the edge device trains and optimizes the global model using local multimodal data to generate preliminary model parameters.

[0027] The cloud server is responsible for aggregating edge model parameters from multiple edge devices and updating the global model through aggregation. Subsequently, the cloud server sends the updated global model parameters back to the edge devices for updating and optimizing the edge models, which include Model 1, Model 2, and Model 3, thus enabling model iteration. Figure 1 As shown.

[0028] Multimodal data includes the following types: pressure data segments collected by pressure sensors from the coal mining machine, vibration data segments collected by vibration sensors from the coal mining machine, and temperature data segments collected by temperature sensors from the coal mining machine. Single-modal data can be any of the above data segments.

[0029] Furthermore, an innovative data processing method is employed for feature extraction from edge devices. This method aims to address potential modality loss and input data misalignment issues in multimodal federated learning. The specific processing steps are as follows.

[0030] 1. To divide the raw data of different modalities into smaller, uniformly processed segments, the multimodal data is first segmented: Where X represents a complete sample, x1, x2, ..., xn. n This represents the n data segments obtained after dividing sample X.

[0031] 2. Perform linear mapping, modal information embedding, and serialization operations on the segmented data using the following formula: Where Y is the sequence resulting from the linear mapping, It is for a specific mode Optimized personalized linear mapping matrix. .in For local mapping matrix, It is a mapping matrix applied to aggregation. It is a hyperparameter between 0 and 1. Through Transformer-based multimodal collaborative training, this invention realizes a federated learning framework for different modal combinations. Each participant can adjust the model according to its own modal composition and task requirements, thereby obtaining a personalized model.

[0032] After linear mapping and modal embedding, the subsequent model learning and training process can treat the above results as data of the same modality. For example... Figure 2 As shown, the system can use federated learning for single-modality processing without worrying about the model being unable to perceive the differences between data from different modalities. This process effectively reduces the complexity of data processing and provides a standardized input format for subsequent model training.

[0033] Improved Transformer blocks such as Figure 4 As shown, its structure includes an input layer, a pooling layer, a convolutional layer, three DRSN layers (Deep Residual Shrinking Network layers), four Itransformer Block layers (Improved Transformer layers), a first linear layer, a first normalization layer, a long short-term linear memory layer, a second normalization layer, a second linear layer, a fully connected layer, and an output layer, all connected in sequence. Each Itransformer Block layer includes a Linear Attention layer, a first Norm & Add layer, a Feed forward layer, and a second Norm & Add layer, all connected in sequence.

[0034] The improved Transformer block's encoding layer replaces the traditional multi-head self-attention mechanism with a multi-scale linear attention mechanism. It also introduces a Deep Residual Shrinking Network (DRSN), where each residual block contains a shrinking operation and an attention mechanism to enhance local feature extraction capabilities and improve the model's accuracy in identifying key features. Specifically, the local feature extraction layer consists of three deep residual shrinking network blocks, each containing a convolutional layer, residual connections, and an attention mechanism to extract local features from the signal and remove redundant information. The global feature extraction layer consists of four ImTransformerBlocks (improved Transformer blocks). Each ImTransformer block first uses a multi-scale linear attention mechanism to extract dependencies between features, then uses residual connections and normalization to ensure training stability, followed by nonlinear transformation, residual connections, and normalization through a feedforward neural network. This invention stacks these structures multiple times to collectively create the model's powerful feature extraction capabilities. The improved Transformer algorithm's decoding layer uses an LSTM network instead of traditional linear layers, better handling time-series data and significantly improving the model's prediction accuracy.

[0035] Furthermore, such as Figure 3 As shown, the edge device incorporates a model relevance judgment mechanism, a model update trend relevance judgment mechanism, and a model download weight allocation mechanism. This mechanism determines the model relevance between the locally trained edge model and the received global model. Only when the model relevance meets predetermined conditions will the edge device provide the locally trained model parameters to the cloud server. The cloud server updates the global model based on the model parameters received from each edge device. The specific process is as follows: The edge device receives the global model from the cloud server, which is the updated global model from the previous training round. The edge device trains the global model using locally collected multimodal data to obtain the edge model. After training, the edge device calculates the model relevance between the locally trained edge model and the received global model. ,in, Indicates the dimension of model parameters. Represents the first edge in the edge model The model parameter values ​​of each edge model parameter Represents the first in the global model The model parameter values ​​of each global model parameter are used. Based on a preset first threshold, the edge device determines the level of model relevance. If the model relevance meets the condition, the model update trend relevance is further calculated. ,in, This is a function used to determine the sign of a parameter's change. Indicates the dimension of the edge model parameters. Represents the first edge in the edge model The change values ​​of the parameters of each edge model Represents the first in the global model If the global model parameter changes, and If the symbols are the same, then and The value is 1, otherwise and The value is 0. According to the preset second threshold, the edge device will only send the locally trained model parameters to the cloud server when both the model relevance and the model update trend relevance meet the preset conditions. If either condition is not met, the model parameters will not be sent.

[0036] Model download weight allocation mechanism: The cloud server dynamically allocates weights to edge devices based on their contribution. The formula for calculating the contribution is as follows: ,in, It is the correlation between the update trends of the i-th edge model and the global model, where, This is the correlation of the i-th model update trend; a larger value is better; D i This is the relevance of the i-th model, where ε is the minimum value; avoid dividing by zero. It is the logarithm of the total number of multimodal data collected by the i-th edge device, to avoid weight imbalance caused by excessive data volume. , and It is the weighting coefficient.

[0037] Contribution score for all edge devices Scaling is performed, and the final weight assigned to the i-th edge device is calculated: ,in: is the contribution of the i-th edge device, e=2.718 is a mathematical constant, k is the total number of edge devices, and t is a control factor. When t approaches 0, the weight distribution will be extremely biased towards the device with the highest score; when t is large, the weight distribution will tend to be average.

[0038] The cloud server is responsible for maintaining the global model to be trained and coordinating the entire federated learning process. The specific steps are as follows: The cloud server sends the initial global model to each edge device. The cloud server receives model parameters from the edge devices that meet preset conditions. The cloud server updates the global model based on the received model parameters from each edge device and sends the updated global model as the global model for the next round of training to each edge device, thus completing one federated learning cycle. The above process is repeated for the next round of model training and updating.

[0039] By assessing model relevance and model update trend relevance, federated learning systems can enable multiple data owners to collaboratively train a global model and selectively update model parameters, while protecting data privacy. This mechanism reduces unnecessary data transmission and communication overhead, while ensuring the high quality and representativeness of uploaded edge model parameters, thereby improving the efficiency of global model construction.

[0040] Furthermore, to improve the performance of the Transformer model in fault diagnosis when processing coal mining machine segment data, especially in terms of local feature extraction capability and computational efficiency, the edge device adopts a multi-scale linear attention mechanism: First, the input matrix X is transformed linearly to generate three new matrices: query matrix Q, key matrix K, and value matrix V, as shown in the following formula: , , Where Wq, Wk, and Wv are trainable weight matrices. N is the length of the input sequence, and d is the channel dimension; next, the feature representation function is used. A new feature representation is obtained by performing a nonlinear transformation on Q and K: ,in, It can be various nonlinear functions. Finally, through calculation... and The similarity between the elements is used to obtain attention weights, which are then applied to the value matrix V to obtain the final output. In this approach, linear attention avoids calculating the similarity between all query matrices Q and key matrices K, instead using learnable convolutions to directly compute attention weights, significantly reducing computational cost. Furthermore, convolution operations with kernels of different scales are applied to the query matrix Q and key matrix K, enabling the model to simultaneously focus on feature information at different scales.

[0041] Furthermore, to effectively capture key local fault features, especially when there is a large amount of redundant information in the data fragments, a residual block structure of a deep residual shrinkage network is introduced into the local feature extraction layer of the Transformer used by the edge device. This can extract local features more effectively and improve the model's noise resistance.

[0042] Furthermore, the edge model combines the advantages of parallel computation of Transformer and serial computation of LSTM, such as... Figure 5 As shown, the Transformer coding region model excels at parallel data processing, efficiently capturing global information within the data; while the LSTM model excels at serial processing of time series data, effectively capturing temporal information. This combined serial and parallel processing approach enables more effective prediction of coal mining machine lifespan data. By capturing global trends in coal mining machine data through parallel processing and capturing temporal variations through serial processing, the model's predictive ability for coal mining machine data with different lifespans is enhanced, making it suitable for various application scenarios and improving its robustness and generalization ability.

[0043] This invention enhances the Transformer network's ability to extract local fault features and improve its noise resistance by introducing DRSN. At the same time, it replaces the traditional multi-head self-attention mechanism with a multi-scale linear attention mechanism, which not only reduces computational complexity but also enhances the model's ability to capture feature information at different scales through multi-scale convolution and adaptive weighting, ultimately improving the model's performance in coal mining machine signal fault diagnosis tasks.

[0044] In this embodiment of the application, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described above.

[0045] In this application embodiment, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0046] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0047] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention described herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not invented herein. The specification and embodiments are to be considered exemplary only.

[0048] The above specific embodiments further illustrate the purpose, technical solution and beneficial effects of this application. It should be understood that the above are only specific embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of this application should be included within the scope of protection of this application.

Claims

1. A cloud-edge collaborative health status diagnosis method for multi-mining machines, characterized in that, Using edge devices as the execution entity, the process includes the following steps: Collect multimodal data on the operating status of the coal mining machine; Feature extraction is performed on multimodal data to obtain single-modal data; Receive the global model sent by the cloud server; The edge model is obtained by training the global model using single-modal data; Send the edge model parameters that meet the preset conditions to the cloud server; Receive updated global model parameters sent by the cloud server; Update the edge model based on the updated global model parameters; The lifespan data of the coal mining machine is obtained by using edge-end models for prediction. The edge model is built on an improved Transformer block, which includes a pooling layer, a convolutional layer, several deep residual shrinking network layers, several Itransformer Block layers, a linear layer, a normalization layer, a long short-term linear memory layer, and a fully connected layer, connected in sequence. Each Itransformer Block layer includes a linear attention layer, a first Norm & Add layer, a feedforward layer, and a second Norm & Add layer, connected in sequence. Multiscale linear attention mechanisms involve multiple linear transformations at different scales; The deep residual shrinking network consists of multiple residual blocks, each of which includes at least one shrinking operation; Meets preset conditions, including: The model correlation between the edge model and the global model is higher than the first threshold; The correlation between the model update trends of the edge model and the global model is higher than the second threshold; The formula for calculating the model correlation D between the edge model and the global model is as follows: ， in, Indicates the dimension of the edge model parameters. In the edge model, the first The model parameter values ​​of each edge model parameter Represents the first in the global model The model parameter values ​​of each global model parameter; Calculate the correlation R between the update trends of the edge model and the global model: ， in, It is a value-finding function. Indicates the dimension of the edge model parameters; Represents the first edge in the edge model The change values ​​of the parameters of each edge model Represents the first in the global model If the model parameter changes, and If the symbols are the same, then and The value is 1, otherwise and The value is 0.

2. The cloud-edge collaborative health status diagnosis method for multi-mining machines according to claim 1, characterized in that, Feature extraction from multimodal data to obtain single-modal data includes: first, segmenting the multimodal data; then, performing linear mapping on the segmented data; and finally, embedding modal information to obtain single-modal data.

3. The cloud-edge collaborative health status diagnosis method for multi-mining machines according to claim 1, characterized in that, The edge models are sorted according to their contribution, and their weights are determined, including: Calculate the contribution score (i) of the edge model parameters: ， in, It is the correlation between the update trends of the i-th edge model and the global model. , and It is the weighting coefficient, D i This represents the relevance of the i-th model, where ε is the minimum value. It is the logarithm of the total number of multimodal data on the operating status of the coal mining machine collected by the i-th edge device; Calculate the weight of the i-th edge device : ， in, is the contribution of the i-th edge device, e=2.718 is a mathematical constant, k is the total number of edge devices, and t is the control factor.

4. A cloud-edge collaborative health status diagnosis method for multi-mining machines, characterized in that, The execution entity is a cloud server, including: Receive edge model parameters that meet preset conditions sent by the edge device; Aggregate the parameters of the edge model and update the global model in the cloud server; Send the updated global model parameters to the edge device; The edge model is built on an improved Transformer block, which includes a pooling layer, a convolutional layer, several deep residual shrinking network layers, several Itransformer Block layers, a linear layer, a normalization layer, a long short-term linear memory layer, and a fully connected layer, connected in sequence. Each Itransformer Block layer includes a linear attention layer, a first Norm & Add layer, a feedforward layer, and a second Norm & Add layer, connected in sequence. Multiscale linear attention mechanisms involve multiple linear transformations at different scales; The deep residual shrinking network consists of multiple residual blocks, each of which includes at least one shrinking operation; Meets preset conditions, including: The model correlation between the edge model and the global model is higher than the first threshold; The correlation between the model update trends of the edge model and the global model is higher than the second threshold; The formula for calculating the model correlation D between the edge model and the global model is as follows: ， in, Indicates the dimension of the edge model parameters. In the edge model, the first The model parameter values ​​of each edge model parameter Represents the first in the global model The model parameter values ​​of each global model parameter; Calculate the correlation R between the update trends of the edge model and the global model: ， in, It is a value-finding function. Indicates the dimension of the edge model parameters; Represents the first edge in the edge model The change values ​​of the parameters of each edge model Represents the first in the global model If the model parameter changes, and If the symbols are the same, then and The value is 1, otherwise and The value is 0.

5. A cloud-edge collaborative health status diagnosis method for multi-mining machines, characterized in that, Includes the following steps: Multimodal data of the coal mining machine's operating status are collected using edge devices, and features are extracted from the multimodal data to obtain single-modal data; Utilize edge devices to receive the global model sent by the cloud server; Based on single-modal data, the received global model is trained using edge devices to obtain the edge model; The model parameters of the edge model are aggregated using the cloud server, and the global model in the cloud server is updated. The updated global model parameters are sent to edge devices using a cloud server. Based on the updated global model parameters, update the edge model using edge devices; The lifespan data of the coal mining machine is obtained by using edge-end models for prediction. The edge model is built on an improved Transformer block, which includes a pooling layer, a convolutional layer, several deep residual shrinking network layers, several Itransformer Block layers, a linear layer, a normalization layer, a long short-term linear memory layer, and a fully connected layer, connected in sequence. Each Itransformer Block layer includes a linear attention layer, a first Norm & Add layer, a feedforward layer, and a second Norm & Add layer, connected in sequence. Multiscale linear attention mechanisms involve multiple linear transformations at different scales; The deep residual shrinking network consists of multiple residual blocks, each of which includes at least one shrinking operation; Meets preset conditions, including: The model correlation between the edge model and the global model is higher than the first threshold; The correlation between the model update trends of the edge model and the global model is higher than the second threshold; The formula for calculating the model correlation D between the edge model and the global model is as follows: ， in, Indicates the dimension of the edge model parameters. In the edge model, the first The model parameter values ​​of each edge model parameter Represents the first in the global model The model parameter values ​​of each global model parameter; Calculate the correlation R between the update trends of the edge model and the global model: ， in, It is a value-finding function. Indicates the dimension of the edge model parameters; Represents the first edge in the edge model The change values ​​of the parameters of each edge model Represents the first in the global model If the model parameter changes, and If the symbols are the same, then and The value is 1, otherwise and The value is 0.

6. A cloud-edge collaborative health status diagnosis system for multi-mining machines, characterized in that, Includes the following modules: The feature extraction module is used to collect multimodal data on the operating status of the coal mining machine using edge devices, and to extract features from the multimodal data to obtain single-modal data. The model parameter acquisition module is used to train the received global model using edge devices based on single-modal data to obtain the edge model; aggregate the model parameters of the edge model using a cloud server to update the global model; send the updated global model parameters to the edge devices using a cloud server; and update the edge model using the edge devices based on the updated global model parameters. The model prediction module is used to obtain coal mining machine life data by using edge model prediction. The edge model is built on an improved Transformer block, which includes a pooling layer, a convolutional layer, several deep residual shrinking network layers, several Itransformer Block layers, a linear layer, a normalization layer, a long short-term linear memory layer, and a fully connected layer, connected in sequence. Each Itransformer Block layer includes a linear attention layer, a first Norm & Add layer, a feedforward layer, and a second Norm & Add layer, connected in sequence. Multiscale linear attention mechanisms involve multiple linear transformations at different scales; The deep residual shrinking network consists of multiple residual blocks, each of which includes at least one shrinking operation; Meets preset conditions, including: The model correlation between the edge model and the global model is higher than the first threshold; The correlation between the model update trends of the edge model and the global model is higher than the second threshold; The formula for calculating the model correlation D between the edge model and the global model is as follows: ， in, Indicates the dimension of the edge model parameters. In the edge model, the first The model parameter values ​​of each edge model parameter Represents the first in the global model The model parameter values ​​of each global model parameter; Calculate the correlation R between the update trends of the edge model and the global model: ， in, It is a value-finding function. Indicates the dimension of the edge model parameters; Represents the first edge in the edge model The change values ​​of the parameters of each edge model Represents the first in the global model If the model parameter changes, and If the symbols are the same, then and The value is 1, otherwise and The value is 0.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 5.

Images