A computer system based on a deep learning model

Abstract

The application relates to the field of deep learning and relates to a computer system based on a deep learning model, wherein a communication module obtains to-be-corrected equipment data and edge model parameters of a plurality of to-be-corrected equipment in real time from an edge end, and the edge equipment model is used for performing preliminary correction on the to-be-corrected equipment data; a clustering module performs clustering processing on the to-be-corrected equipment data, and obtains an association factor corresponding to a data cluster; a training module trains a basic model by using the edge model parameters and sample equipment data, and obtains a central correction model when the basic model meets a preset deep learning convergence condition; the convergence condition is that the basic model learns an associated operation mode of the to-be-corrected equipment through the association factor; and a correction module inputs the preliminary correction parameters into the central correction model, and obtains secondary correction data. The application extracts the internal association features among the to-be-corrected equipment data through the association factor, so that the model has the ability to consider the data association, and effectively reduces the equipment error.

Description

Technical Field

[0001] This invention relates to the field of deep learning, and more particularly to a computer system based on a deep learning model. Background Technology

[0002] Currently, computer systems often face the challenge of processing multi-source heterogeneous data. "Multi-source" refers to data originating from different data sources, which are various access devices. "Heterogeneous" often indicates differences in the data structure, storage format, and access interface of the data output by the access devices to the computer system (collectively referred to as device data). When a computer system communicates with multiple access devices, it is precisely because of these differences in data type, data structure, storage format, and / or access interface that the computer system needs to correct the device data of the access devices. However, existing correction methods often struggle to balance efficiency and accuracy when dealing with complex multi-source data, limiting the system's application in dynamic environments.

[0003] Traditional calibration methods typically optimize data streams from a single device, lacking analysis of the correlation between data from multiple access devices. In reality, when there is correlation between data from multiple devices, the calibration results obtained by traditional calibration methods will have significant errors. Summary of the Invention

[0004] This invention provides a computer system based on a deep learning model to solve the technical problem of how to reduce the error in correcting device data.

[0005] To address the aforementioned technical problems, this invention provides a computer system based on a deep learning model, wherein the computer system is deployed with a basic model;

[0006] The computer system includes a communication module, a clustering module, a training module, and a correction module; wherein,

[0007] The communication module is used to acquire, in real time, data of multiple devices to be calibrated and edge model parameters of the edge device model deployed on the edge terminal based on communication with at least one edge terminal; the edge device model is used to perform preliminary calibration processing on the data of the devices to be calibrated to obtain preliminary calibration parameters.

[0008] The clustering module is used to perform clustering processing on the data of the device to be corrected, and obtain the correlation factor corresponding to each data cluster based on the clustering results.

[0009] The training module is used to acquire sample device data; train the base model using the edge model parameters and the sample device data; and obtain a center correction model when the base model meets the preset deep learning convergence conditions; wherein, the deep learning convergence conditions include the base model learning the associated operating mode of the device to be corrected through the association factor.

[0010] The correction module is used to input the preliminary correction parameters into the central correction model and obtain secondary correction data based on the output of the central correction model.

[0011] As a preferred embodiment, the edge device model is used to perform preliminary correction processing on the data of the device to be corrected to obtain preliminary correction parameters, including:

[0012] The edge device model uses a mutual information calculation method to filter out associated device data from a preset historical device database that have a mutual information value greater than a preset association threshold with the data of the device to be corrected.

[0013] A knowledge graph is constructed by using the associated device data as nodes and the mutual information value as edge weights; the historical drift rate of the nodes is labeled, and the associated feature carrier is obtained based on the labeled knowledge graph; wherein, the historical drift rate is calculated based on the average change of parameters of the associated devices over multiple operating cycles;

[0014] Joint feature extraction is performed on the data of the device to be corrected and the associated feature carrier to obtain fused features; wavelet threshold denoising is then applied to the fused features.

[0015] Preliminary correction parameters are obtained by processing the fused features after wavelet threshold denoising using a preset deviation-aware correction model.

[0016] As a preferred embodiment, the fusion features include the deviation values ​​of the data of the device to be corrected, the deviation values ​​of the associated device data, and the real-time operating temperature of the device to be corrected; the edge device model processes the fusion features after wavelet threshold denoising through a preset deviation-aware correction model to obtain preliminary correction parameters, including:

[0017] The edge device model inputs the fused features into a preset deviation-aware correction model; the preset deviation-aware correction model includes a deviation attributor and a correction predictor; the deviation attributor uses a decision tree, and the correction predictor uses a long short-term memory network;

[0018] The bias attributor is used to identify the fused features and output the dominant source of the bias.

[0019] The correction predictor is used to determine the corresponding deviation-dominant attention weight based on the deviation-dominant source, and to predict the correction value corresponding to the fused feature using the deviation-dominant attention weight in combination with a preset attention coefficient to obtain preliminary correction parameters; the preset attention coefficient includes aging deviation coefficient, operating condition interference coefficient and coupling deviation coefficient.

[0020] As a preferred embodiment, the edge device model performs wavelet threshold denoising on the fused features, including:

[0021] The edge device model uses Sym8 wavelet basis functions to denoise the fused features. During the denoising process, the fluctuation frequency of the data of the device to be corrected is calculated in real time, and the wavelet threshold is dynamically adjusted according to the fluctuation frequency.

[0022] As a preferred embodiment, the clustering module performs clustering processing on the data of the device to be corrected, and obtains the correlation factor corresponding to each data cluster based on the clustering results, including:

[0023] The clustering module performs a difference operation on the data of the device to be calibrated and the preliminary calibration parameters to obtain parameter deviation characteristics;

[0024] Obtain the equipment operating status label, and construct a multimodal feature matrix based on the equipment operating status label, the type of the deviation main source, the parameter deviation characteristics, and the equipment data to be corrected;

[0025] The multimodal feature matrix is ​​reconstructed using a preset adversarial autoencoder to obtain enhanced feature vectors;

[0026] Density peak clustering and fuzzy C-means clustering are performed sequentially on the enhanced feature vectors to obtain multidimensional correlation factors for each data cluster; the multidimensional correlation factors include intra-cluster correlation factors, inter-cluster correlation factors and equipment operation status correlation factors.

[0027] As a preferred embodiment, the intra-cluster correlation factor is calculated by constructing a causal graph of intra-cluster parameters through Bayesian network structure learning.

[0028] As a preferred embodiment, the inter-cluster correlation factor is obtained through the inter-cluster correlation degree matrix; the inter-cluster correlation degree matrix is ​​generated based on the nonlinear correlation degree combined with the connection weight determined based on the physical connection relationship of the devices; the nonlinear correlation degree is calculated by using the maximum information coefficient to determine the correlation degree between the enhanced feature vectors of each data cluster.

[0029] As a preferred embodiment, the time-series correlation factor is obtained by slicing the data cluster according to the device operating status and using the Granger causality test to calculate the lag effect coefficient between adjacent slices.

[0030] As a preferred embodiment, the training module trains the base model using the edge model parameters and the sample device data. When the base model satisfies the preset deep learning convergence condition, a center correction model is obtained, including:

[0031] Based on the edge model parameters, extract the association factor summary, and determine the weight of each edge end in the basic model process according to the matching degree between the association factor summary and the association factor corresponding to the data cluster.

[0032] Based on the weights corresponding to each edge end, the basic model is trained using the edge model parameters and the sample device data. When the basic model learns the associated operating mode of the device to be corrected through the association factor, the central correction model is obtained.

[0033] As a preferred embodiment, the edge terminal communicates with at least one device to be calibrated, and the edge terminal is an edge gateway.

[0034] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0035] The clustering module of the computer system proposed in this invention clusters the data of the devices to be calibrated to obtain the correlation factors of each data cluster. When training the basic model, it not only integrates the parameters of the marginal model with the sample device data, but also uses "the basic model learns the correlation operation mode of the devices to be calibrated through the correlation factors" as the deep learning convergence condition, and finally generates the central calibration model. The intrinsic correlation features between the data of the devices to be calibrated are extracted through the correlation factors, and the convergence condition forces the basic model to learn the correlation operation mode, so that the central calibration model has the ability to optimize the calibration results based on the data correlation. This solves the defect of traditional calibration methods that ignore the correlation of multi-device data, and reduces the calibration error caused by not considering the data correlation from the root.

[0036] Furthermore, the communication module synchronously acquires the data of the device to be calibrated and the parameters of the edge device model through real-time communication with the edge end. The edge device model has already completed the preliminary calibration of the data of the device to be calibrated. On the one hand, this application moves the data calibration work that does not need to consider the correlation between devices to the edge end. On the other hand, it puts the secondary calibration of the device based on the correlation operation mode into the computer system and completes it by the central calibration model. In this way, the relatively simple calibration tasks with relatively low requirements for computing resources can be completed by the edge end, reducing the requirements for the computing power and performance of the edge end. At the same time, the relatively complex secondary calibration with relatively high requirements for computing resources can be completed by the relatively high-performance computer system, which can optimize the overall allocation of computing resources and performance and improve the efficiency of the overall calibration process.

[0037] Furthermore, the edge-end preliminary correction of this application focuses on efficiency and quickly completes the preliminary correction processing of the data of the device to be corrected; the central correction model, having learned the associated operating mode of the device, can focus on the accuracy level and optimize the preliminary correction results through associated features. The correction method of this application can perform targeted preliminary correction on the data of the device to be corrected at the edge end, while the common correction work is completed by the central correction model, thus achieving a balance between efficiency and accuracy in complex multi-source heterogeneous data correction scenarios.

[0038] Finally, the training module introduces edge device model parameters during the basic model training process, so that the model training not only relies on sample device data, but also integrates the model parameters of actual operation at the edge, avoiding the disconnect between the central model and the actual device operation scenario caused by relying solely on sample data for training; combined with the device association operation mode learned by the association factor, it further ensures that the secondary correction data output by the central correction model is highly matched with the actual operating state of the device to be corrected, which can significantly improve the reliability of the correction results. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of an embodiment of a computer system based on a deep learning model provided in this application.

[0040] Figure 2 This is a flowchart illustrating a preferred embodiment of the computer system based on a deep learning model provided in this application.

[0041] Figure 3 This is a flowchart illustrating a preferred embodiment of the computer system based on a deep learning model provided in this application.

[0042] Figure 4 This is a flowchart illustrating a preferred embodiment three of the computer system based on a deep learning model provided in this application.

[0043] Figure 5 This is a flowchart illustrating an embodiment of the device data correction method based on a deep learning model provided in this invention application. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1:

[0046] Please refer to Figure 1, Figure 1 This is a schematic diagram of the structure of an embodiment of a computer system 100 based on a deep learning model provided in this invention.

[0047] In this embodiment, the computer system 100 can be applied to computer devices, including but not limited to smartphones, laptops, tablets, desktop computers, physical servers, and cloud servers.

[0048] In this embodiment, the computer system 100 is deployed with a basic model, and the computer system 100 communicates in real time with at least one edge terminal. Each edge terminal is deployed with an edge device model and communicates with at least one device to be calibrated. The edge terminal may be an edge gateway.

[0049] When these devices to be calibrated are connected to a system (such as a power system) (including both the moment of connection and the ongoing connection), the system may receive device data with diverse and heterogeneous characteristics due to the large number of connected devices. To integrate and analyze this diverse data, it is necessary to perform calibration processing on this massive amount of device data, which differs in data type, data structure, storage format, and access interface (primarily correcting for differences caused by the operating environment and device type of different connected devices). This calibration process serves two purposes: firstly, to facilitate integration, processing, and analysis in subsequent steps; and secondly, to ensure the accuracy of the acquired device data while maintaining stable system operation, reducing the deviation between the acquired device data and the actual output data of the devices.

[0050] In this embodiment, the computer system 100 includes a communication module 101, a clustering module 102, a training module 103, and a calibration module 104. The communication module 101 is used to acquire, in real-time, data of multiple devices to be calibrated and edge model parameters of an edge device model deployed at the edge, based on communication with at least one edge terminal. The edge device model is used to perform preliminary calibration processing on the data of devices to be calibrated to obtain preliminary calibration parameters. The clustering module 102 is used to perform clustering processing on the data of devices to be calibrated and obtain correlation factors corresponding to each data cluster based on the clustering results. The training module 103 is used to acquire sample device data and train the base model using the edge model parameters and the sample device data. When the base model meets a preset deep learning convergence condition, a central calibration model is obtained. The deep learning convergence condition includes the base model learning the correlation operation mode of the devices to be calibrated through the correlation factors. The calibration module 104 is used to input the preliminary calibration parameters into the central calibration model and obtain secondary calibration data based on the output of the central calibration model.

[0051] In this embodiment, when the edge device acquires the data of the device to be calibrated, it can quickly complete the preliminary calibration of the device data by focusing on efficiency through preliminary calibration. This preliminary calibration does not need to consider the correlation between different devices, but only considers the calibration of the data quality of a single device. For example, simple normalization, outlier removal, and missing value imputation can be performed. On the other hand, the central calibration model, having learned the associated operating patterns of the devices, can focus on the accuracy level and optimize the preliminary calibration results through device association features. The calibration method of this invention allows the edge device to perform separate and targeted preliminary calibration for each device data, while the calibration work with common characteristics of each device to be calibrated is completed by the central calibration model, achieving a balance between efficiency and accuracy in complex multi-source heterogeneous data calibration scenarios. This approach avoids placing excessive demands on the computing power and performance of the edge device, reducing the cost of the edge device, while the high-computing-power calibration task considering the associated operating patterns of the devices is completed by the central calibration model, optimizing the overall allocation of computing resources and performance.

[0052] In this embodiment, both the sample device data and the data to be calibrated are device data of the access device. The difference between the two is that the former is sample device data (used for training the basic model), while the latter is data of the device to be calibrated. Both contain device data regarding data type, data structure, storage format, and access interface, respectively. This device data (sample device data or data to be calibrated) can reflect the differences between different access devices during the transmission process between the system and the access device (e.g., different access devices use different communication protocols or require different data storage methods). It can be understood that the sample device data in this application serves as a training set for the basic model. When it is estimated or envisioned that multiple devices to be calibrated may include various device types such as A, B, and C, the sample device data can also be pre-configured to include device data of various device types such as A, B, and C. Therefore, when using it in conjunction with edge model parameters to train the basic model, it can ensure that the basic model can learn the associated operating mode, and at the same time, it can make the performance of the trained central calibration model match as closely as possible to the device types that the devices to be calibrated may contain.

[0053] like Figure 2 As shown, in a preferred embodiment, the edge device model is used to perform preliminary correction processing on the data of the device to be corrected, and the preliminary correction parameters are obtained by steps S201 to S204 (the executing entity can be the edge device model or the edge terminal); the details of each step are as follows:

[0054] Step S201: Using the mutual information calculation method, filter out associated device data from the preset historical device database that have a mutual information value greater than a preset association threshold with the data of the device to be corrected;

[0055] Step S202: Construct a knowledge graph by using the associated device data as nodes and the mutual information value as edge weights; label the historical drift rate of the nodes, and obtain the associated feature carrier based on the labeled knowledge graph; wherein, the historical drift rate is calculated based on the average change of parameters of the associated devices over multiple operating cycles;

[0056] Step S203: Perform joint feature extraction on the data of the device to be corrected and the associated feature carrier to obtain fused features; perform wavelet threshold denoising on the fused features;

[0057] Step S204: The fused features after wavelet threshold denoising are processed by a preset deviation perception correction model to obtain preliminary correction parameters.

[0058] In this preferred embodiment, the preset association threshold can be 0.6. Step S201 filters out associated device data from the preset historical device database. This increases the sample size used to construct the knowledge graph while filtering out device data with low association, ensuring the quality of the constructed knowledge graph.

[0059] In this preferred embodiment, the preset historical device database can be local data stored on the edge. Therefore, it should be noted that the "separate" and "targeted" meanings of "the edge performs separate and targeted preliminary corrections on the device data" mentioned above are "separate" and "targeted" relative to the device to be corrected. The preliminary correction can use the existing device data stored locally on the corresponding edge, without considering the device data to be corrected on other edge devices.

[0060] Compared with conventional edge-end correction operations in the prior art, this preferred embodiment uses mutual information calculation to filter associated devices and knowledge graph to construct parameter association relationships. This can transform isolated single device data to be corrected into a multi-device association feature set of data to be corrected and associated devices, enabling the preliminary correction to actively adapt to the coupling effects between devices within the local data range of a single edge end, thus solving the correction deviation problem caused by traditional solutions.

[0061] Furthermore, the edge device model performs wavelet threshold denoising on the fused features, including: the edge device model uses the Sym8 wavelet basis function to denoise the fused features; during the denoising process, the fluctuation frequency of the data to be corrected is calculated in real time, and the wavelet threshold is dynamically adjusted according to the fluctuation frequency. For example, when the fluctuation frequency of the data to be corrected meets the definition of "high frequency," the wavelet threshold can be set to 0.05~0.1; when it meets the definition of "low frequency," the wavelet threshold can be set to 0.01~0.03. In this way, data with different fluctuation frequencies can be targeted for denoising through dynamic wavelet thresholding.

[0062] In addition, before inputting the preliminary correction parameters into the center correction model, the fused features that have undergone wavelet threshold denoising can be dimensionality reduced, for example, by mapping them from a high-dimensional space to a low-dimensional feature space of 10 to 15 dimensions. This preserves the parameter correlation information and reduces the computational load of the center correction model.

[0063] In some other preferred embodiments, the fusion features include the deviation value of the data of the device to be calibrated, the deviation value of the associated device data, and the real-time operating temperature of the device to be calibrated.

[0064] like Figure 3 As shown, the edge device model processes the fused features after wavelet threshold denoising using a preset deviation-aware correction model to obtain preliminary correction parameters, including steps S301 to S303; each step is detailed below:

[0065] Step S301: Input the fused features into a preset deviation perception correction model; the preset deviation perception correction model includes a deviation attributor and a correction predictor; the deviation attributor uses a decision tree, and the correction predictor uses a long short-term memory network.

[0066] Step S302, the bias attributor is used to identify the fused features and output the dominant source of the bias;

[0067] In step S303, the correction predictor is used to determine the corresponding deviation-dominant attention weight based on the deviation-dominant source, and use the deviation-dominant attention weight in combination with a preset attention coefficient to predict the correction value corresponding to the fusion feature to obtain preliminary correction parameters; the preset attention coefficient includes aging deviation coefficient, operating condition interference coefficient and coupling deviation coefficient.

[0068] In this embodiment, the preset deviation perception correction model can be pre-trained before step S301. During training, historical correction data from the past 20 running cycles can be used (each sample contains fusion features from 20 cycles and corresponding preliminary correction values).

[0069] For example, the bias attributor can be a lightweight decision tree with a depth of 5 to 8 layers and a leaf node sample number greater than or equal to 5.

[0070] The convergence condition for training the bias perception correction model can be that the average absolute error between the predicted correction value and the actual preliminary correction value is ≤0.02.

[0071] This preferred embodiment employs a dual-module deviation perception architecture consisting of a deviation attributor and a correction predictor. Compared to the existing technology's "one-size-fits-all" correction strategy that uses a single neural network to directly output correction values, this preferred embodiment uses a decision tree to accurately locate the dominant source of deviation and uses the dominant source of deviation to determine the corresponding deviation-dominant attention weight of the correction predictor. This allows the model to achieve differentiated optimization of the correction accuracy for different deviation sources (such as assigning higher weights to aging deviations with greater impact), realizing precise processing of attribution first and then correction, and breaking through the technical bottleneck of indiscriminate correction in traditional models.

[0072] In a preferred embodiment, such as Figure 4 As shown, the clustering module 102 performs clustering processing on the data of the device to be corrected, and obtains the correlation factor corresponding to each data cluster based on the clustering results, including steps S401 to S404; each step is detailed below:

[0073] Step S401: Perform differential operation on the data of the device to be calibrated and the preliminary calibration parameters to obtain parameter deviation characteristics;

[0074] Step S402: Obtain the equipment operating status label, and construct a multimodal feature matrix based on the equipment operating status label, the type of the deviation main source, the parameter deviation characteristics, and the equipment data to be corrected;

[0075] Step S403: The multimodal feature matrix is ​​reconstructed using a preset adversarial autoencoder to obtain an enhanced feature vector;

[0076] Step S404: Density peak clustering and fuzzy C-means clustering are performed sequentially on the enhanced feature vector to obtain the multidimensional correlation factor for each data cluster; the multidimensional correlation factor includes intra-cluster correlation factor, inter-cluster correlation factor and equipment operation status correlation factor.

[0077] In this embodiment, the parameter deviation characteristic can be expressed as:

[0078] D i =|P i -C i |;

[0079] Among them, D i Let P be the parameter deviation characteristic of the i-th device to be calibrated.i For the data of the i-th device to be calibrated, C i These are the initial calibration parameters for the i-th device to be calibrated.

[0080] The type of the dominant source of the deviation can be represented by a label of the dominant source of the deviation, which can be obtained based on the output of the deviation attributor.

[0081] During the adversarial training of the generator and discriminator, the preset adversarial autoencoder is trained in 50 to 80 rounds, with a discriminator accuracy threshold of 0.85. Furthermore, its output dimension is compressed to one-third of the original feature vector, retaining key features while eliminating redundant information.

[0082] For example, density peak clustering can optimize the traditional local density calculation method into spatiotemporal weighted density (time weight decays according to the parameter acquisition interval, and spatial weight is assigned according to the physical distance of the device), and automatically identify 10 to 20 initial cluster centers.

[0083] Fuzzy C-means clustering can perform fine clustering based on the initial cluster centers mentioned above, combined with an attention mechanism. It assigns attention weights corresponding to different cluster centers to each feature vector. Feature vectors with weights greater than 0.6 are assigned to the corresponding clusters, and the rest are assigned to the clusters by calculating the membership matrix using fuzzy C-means (with the membership threshold set to 0.7). Finally, it forms data clusters with clear boundaries.

[0084] This preferred embodiment constructs multimodal features and enhances them through adversarial autoencoder reconstruction, enabling the clustering features to simultaneously contain numerical differences, process attributes, and root cause information of deviations, thus solving the problem of insufficient intra-cluster homogeneity caused by the single feature dimension in traditional clustering.

[0085] In a further preferred embodiment, the intra-cluster correlation factor is calculated by learning a causal graph of intra-cluster parameters through a Bayesian network structure.

[0086] Furthermore, the intra-cluster association factor can be constructed by using Bayesian network structure learning (K2 algorithm) to build a causal graph of intra-cluster parameters, calculate the causal strength value between each node in the causal graph (represented by conditional probability increment), and take the causal relationship with the highest strength value to form a 5×5 matrix as the intra-cluster association factor.

[0087] Furthermore, the inter-cluster correlation factor is obtained through the inter-cluster correlation degree matrix; the inter-cluster correlation degree matrix is ​​generated based on the nonlinear correlation degree combined with the connection weight (1.2 times the weight of the physical connection device) determined based on the physical connection relationship of the equipment (such as transmission connection, electrical signal connection); the nonlinear correlation degree is calculated by the maximum information coefficient to determine the correlation between the enhanced feature vectors of each data cluster.

[0088] Furthermore, the time-series correlation factor is obtained by slicing the data cluster according to the device operating state (one slice for each operating state) and using the Granger causality test to calculate the lag effect coefficient between adjacent slices (the lag order is 2 to 3).

[0089] This preferred embodiment comprehensively considers intra-cluster correlation factors, inter-cluster correlation factors, and temporal correlation factors. It can avoid the influence of unstable clusters on the correlation factor calculation results in the clustering results, effectively improve the quality of the correlation factor calculation results, and ensure that the correlation factors effectively support the subsequent deep learning models.

[0090] Preferably, the training module 103 trains the base model using the edge model parameters and the sample device data. When the base model meets the preset deep learning convergence conditions, a center correction model is obtained. This includes: the training module 103 extracts a correlation factor summary based on the edge model parameters; determines the weights corresponding to each edge end during the base model process based on the matching degree between the correlation factor summary and the correlation factors corresponding to the data clusters; and then trains the base model using the edge model parameters and the sample device data based on the weights corresponding to each edge end. When it is determined that the base model has learned the correlation operation mode of the device to be corrected through the correlation factors, a center correction model is obtained.

[0091] In this preferred embodiment, the base model can adopt the Transformer architecture, and the matching degree can be calculated using cosine similarity. Edges with a matching degree greater than 0.8 are assigned basic weights (accounting for 60%), while edges with a matching degree of 0.5 to 0.8 are assigned adaptive weights (linearly scaled according to the matching degree, ranging from 30% to 50%).

[0092] This preferred solution innovatively introduces the concept of federated learning when calibrating the device to be calibrated. The association factor acts as the "anchor" for federated aggregation. By using the matching degree between the association factor summary and the association factor corresponding to the data cluster, the participation weight of the edge end in the basic model training process is dynamically allocated. This allows the aggregation process to take into account both the data scale and the consistency of the association pattern, and can avoid the problem of the central calibration model shift caused by blind aggregation.

[0093] Accordingly, this application also provides a device data correction method based on a deep learning model. Figure 5 The illustration shows a flowchart of an embodiment of the device data correction method based on a deep learning model provided in this application.

[0094] The device data correction method is applied to a computer system, which has a basic model deployed on it.

[0095] The device data correction method includes steps S501 to S504; each step is described in detail below:

[0096] Step S501: Based on communication with at least one edge terminal, acquire in real time the data of multiple devices to be calibrated and the edge model parameters of the edge device model deployed on the edge terminal; the edge device model is used to perform preliminary calibration processing on the data of the devices to be calibrated to obtain preliminary calibration parameters;

[0097] Step S502: Perform clustering processing on the data of the device to be corrected, and obtain the correlation factor corresponding to each data cluster based on the clustering results.

[0098] Step S503: Obtain sample device data; train the basic model using the edge model parameters and the sample device data; when the basic model meets the preset deep learning convergence conditions, obtain the center correction model; wherein, the deep learning convergence conditions include the basic model learning the associated operating mode of the device to be corrected through the association factor;

[0099] Step S504: Input the preliminary correction parameters into the central correction model, and obtain secondary correction data based on the output of the central correction model.

[0100] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0101] The clustering module of the computer system proposed in this invention clusters the data of the devices to be calibrated to obtain the correlation factors of each data cluster. When training the basic model, it not only integrates the parameters of the marginal model with the sample device data, but also uses "the basic model learns the correlation operation mode of the devices to be calibrated through the correlation factors" as the deep learning convergence condition, and finally generates the central calibration model. The intrinsic correlation features between the data of the devices to be calibrated are extracted through the correlation factors, and the convergence condition forces the basic model to learn the correlation operation mode, so that the central calibration model has the ability to optimize the calibration results based on the data correlation. This solves the defect of traditional calibration methods that ignore the correlation of multi-device data, and reduces the calibration error caused by not considering the data correlation from the root.

[0102] Furthermore, the communication module synchronously acquires the data of the device to be calibrated and the parameters of the edge device model through real-time communication with the edge end. The edge device model has already completed the preliminary calibration of the data of the device to be calibrated. On the one hand, this application moves the data calibration work that does not need to consider the correlation between devices to the edge end. On the other hand, it puts the secondary calibration of the device based on the correlation operation mode into the computer system and completes it by the central calibration model. In this way, the relatively simple calibration tasks with relatively low requirements for computing resources can be completed by the edge end, reducing the requirements for the computing power and performance of the edge end. At the same time, the relatively complex secondary calibration with relatively high requirements for computing resources can be completed by the relatively high-performance computer system, which can optimize the overall allocation of computing resources and performance and improve the efficiency of the overall calibration process.

[0103] Furthermore, the edge-end preliminary correction of this application focuses on efficiency and quickly completes the preliminary correction processing of device data; the central correction model, having learned the associated operating mode of the devices, can focus on the accuracy level and optimize the preliminary correction results through associated features. The correction method of this application can perform targeted preliminary correction for device data at the edge end, while the common correction work is completed by the central correction model, thus achieving a balance between efficiency and accuracy in complex multi-source heterogeneous data correction scenarios.

[0104] Finally, the training module introduces edge device model parameters during the basic model training process, so that the model training not only relies on sample device data, but also integrates the model parameters of actual operation at the edge, avoiding the disconnect between the central model and the actual device operation scenario caused by relying solely on sample data for training; combined with the device association operation mode learned by the association factor, it further ensures that the secondary correction data output by the central correction model is highly matched with the actual operating state of the device to be corrected, which can significantly improve the reliability of the correction results.

[0105] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A computer system based on a deep learning model, characterized in that, The computer system is deployed with a basic model; The computer system includes a communication module, a clustering module, a training module, and a correction module; wherein, The communication module is used to acquire, in real time, data of multiple devices to be calibrated and edge model parameters of the edge device model deployed on the edge terminal based on communication with at least one edge terminal; the edge device model is used to perform preliminary calibration processing on the data of the devices to be calibrated to obtain preliminary calibration parameters. The clustering module is used to perform clustering processing on the data of the device to be corrected, and obtain the correlation factor corresponding to each data cluster based on the clustering results. The training module is used to acquire sample device data; train the base model using the edge model parameters and the sample device data; and obtain a center correction model when the base model meets the preset deep learning convergence conditions; wherein, the deep learning convergence conditions include the base model learning the associated operating mode of the device to be corrected through the association factor. The correction module is used to input the preliminary correction parameters into the central correction model and obtain secondary correction data based on the output of the central correction model. The clustering module performs clustering processing on the data of the device to be corrected, and obtains the correlation factors corresponding to each data cluster based on the clustering results, including: The clustering module performs a difference operation on the data of the device to be calibrated and the preliminary calibration parameters to obtain parameter deviation characteristics; Obtain the equipment operating status label, and construct a multimodal feature matrix based on the equipment operating status label, the type of the dominant source of deviation, the parameter deviation characteristics, and the data of the equipment to be corrected; The multimodal feature matrix is ​​reconstructed using a preset adversarial autoencoder to obtain enhanced feature vectors; The enhanced feature vectors are sequentially subjected to density peak clustering and fuzzy C-means clustering to obtain the multidimensional correlation factor for each data cluster; the multidimensional correlation factor includes intra-cluster correlation factor, inter-cluster correlation factor and equipment operating status correlation factor. The training module trains the base model using the edge model parameters and the sample device data. When the base model meets the preset deep learning convergence conditions, a center correction model is obtained, including: The training module extracts a summary of association factors based on the parameters of the edge model, and determines the weights of each edge end in the basic model process based on the matching degree between the summary of association factors and the association factors corresponding to the data clusters. Therefore, based on the weights corresponding to each edge end, the basic model is trained using the edge model parameters and the sample device data.

2. The computer system based on a deep learning model as described in claim 1, characterized in that, The edge device model is used to perform preliminary correction processing on the data of the device to be corrected, and to obtain preliminary correction parameters, including: The edge device model uses a mutual information calculation method to filter out associated device data from a preset historical device database that have a mutual information value greater than a preset association threshold with the data of the device to be corrected. A knowledge graph is constructed by using the associated device data as nodes and the mutual information value as edge weights; the historical drift rate of the nodes is labeled, and the associated feature carrier is obtained based on the labeled knowledge graph; wherein, the historical drift rate is calculated based on the average change of parameters of the associated devices over multiple operating cycles; Joint feature extraction is performed on the data of the device to be corrected and the associated feature carrier to obtain fused features; wavelet threshold denoising is then applied to the fused features. Preliminary correction parameters are obtained by processing the fused features after wavelet threshold denoising using a preset deviation-aware correction model.

3. A computer system based on a deep learning model as described in claim 2, characterized in that, The fusion features include the deviation value of the data of the device to be calibrated, the deviation value of the data of the associated device, and the real-time operating temperature of the device to be calibrated. The edge device model processes the fused features after wavelet threshold denoising using a preset deviation-aware correction model to obtain preliminary correction parameters, including: The edge device model inputs the fused features into a preset deviation-aware correction model; the preset deviation-aware correction model includes a deviation attributor and a correction predictor; the deviation attributor uses a decision tree, and the correction predictor uses a long short-term memory network; The bias attributor is used to identify the fused features and output the dominant source of the bias. The correction predictor is used to determine the corresponding deviation-dominant attention weight based on the deviation-dominant source, and to predict the correction value corresponding to the fused feature using the deviation-dominant attention weight in combination with a preset attention coefficient to obtain preliminary correction parameters; the preset attention coefficient includes aging deviation coefficient, operating condition interference coefficient and coupling deviation coefficient.

4. A computer system based on a deep learning model as described in claim 2, characterized in that, The edge device model performs wavelet threshold denoising on the fused features, including: The edge device model uses Sym8 wavelet basis functions to denoise the fused features. During the denoising process, the fluctuation frequency of the data of the device to be corrected is calculated in real time, and the wavelet threshold is dynamically adjusted according to the fluctuation frequency.

5. A computer system based on a deep learning model as described in claim 1, characterized in that, The intra-cluster correlation factor is calculated by learning a causal graph of intra-cluster parameters through a Bayesian network structure.

6. A computer system based on a deep learning model as described in claim 1, characterized in that, The inter-cluster correlation factor is obtained through the inter-cluster correlation degree matrix; the inter-cluster correlation degree matrix is ​​generated based on the nonlinear correlation degree combined with the connection weight determined based on the physical connection relationship of the devices; the nonlinear correlation degree is calculated by using the maximum information coefficient to determine the correlation degree between the enhanced feature vectors of each data cluster.

7. A computer system based on a deep learning model as described in claim 1, characterized in that, The equipment operating status correlation factor is obtained by slicing the data cluster according to the equipment operating status and using Granger causality test to calculate the lag effect coefficient between adjacent slices.

8. A computer system based on a deep learning model as described in any one of claims 1 to 7, characterized in that, The edge terminal communicates with at least one device to be calibrated, and the edge terminal is an edge gateway.

Images