A method for saving data of small PLC of electro-hydraulic actuator

Abstract

The present invention relates to the technical field of data preservation and recovery, and provides a method for preserving data of a small PLC of an electro-hydraulic actuator, comprising real-time collection of operating data generated by a small or medium-sized PLC of an electro-hydraulic actuator during debugging; pre-processing the operating data to obtain multiple operating data segments; extracting features from the operating data segments through a self-supervised learning model to obtain key features; processing the key features through a two-layer generative adversarial network to obtain a basic data structure and reconstructed data; and classifying and preserving the basic data structure and reconstructed data. The present invention improves the accuracy and efficiency of data preservation, and can dynamically adjust the storage strategy according to the importance and complexity of the data, thereby optimizing the utilization of storage resources and reducing redundant storage. This further improves the reliability and safety of the long-term operation of electro-hydraulic actuators in industrial application sites.

Description

Technical Field

[0001] The present invention relates to the technical field of data storage, and in particular to a method for storing data of a small PLC of an electro-hydraulic actuator. Background Art

[0002] With the rapid development of industrial automation and intelligence, electro-hydraulic actuators have been widely used in various industrial control systems, especially in industrial environments involving complex operations such as petrochemicals and metal smelting. Small PLCs and touch screens are widely used for the control and monitoring of electro-hydraulic actuators. In actual application, data preservation is a key issue, especially in equipment maintenance, hardware replacement or unexpected failures. How to effectively preserve data is directly related to the stability and reliability of the system.

[0003] In the prior art, data preservation methods generally rely on simple backup and redundancy strategies. Common methods include regularly backing up operating data to external storage devices or replicating data between multiple storage media through redundant storage mechanisms to improve data security and recovery capabilities. Traditional methods have obvious limitations. First, backup operations are usually triggered manually or on a timed basis, which can easily lead to untimely data backups. In the event of a sudden system failure, the latest data cannot be effectively restored, resulting in data loss or incompleteness. Secondly, although traditional redundant storage strategies can improve data availability, they are inefficient and require a large amount of storage space to store redundant data. This is particularly evident in resource-constrained situations. For embedded systems or small industrial control systems, redundant storage strategies cannot be effectively implemented. In industrial control systems, small PLCs and touch screens with electro-hydraulic actuators often face the challenges of limited data storage resources and complex environments. The data preservation methods in the prior art show significant limitations in certain scenarios. The prior art is difficult to efficiently save large amounts of operating data when storage space is limited, resulting in the risk of data loss or recovery failure during hardware replacement or recovery.

[0004] In addition, existing technologies also have shortcomings in the intelligent and automated processing of data preservation and recovery. Traditional methods usually rely on predefined rules or simple algorithms for data preservation, lack in-depth analysis and intelligent processing capabilities of data content, cannot dynamically adjust storage strategies based on the importance and complexity of the data, and cannot intelligently infer and reconstruct lost or damaged data during the data recovery process. They seem powerless when faced with data preservation and recovery needs in complex industrial environments.

[0005] In petroleum refining enterprises, electro-hydraulic actuators are widely used to control special valve equipment in catalytic cracking units. Electro-hydraulic actuator manufacturers use small PLCs and touch screens as core control and monitoring equipment for electro-hydraulic actuators. However, due to long-term operation and complex production environments, equipment often faces the risk of data loss or unrecoverable data during hardware maintenance and replacement, resulting in production line shutdowns and economic losses. Summary of the Invention

[0006] The present invention aims to address at least one of the technical problems existing in the related art. To this end, the present invention provides a data storage method for a small PLC in an electro-hydraulic actuator. This method improves the accuracy and efficiency of data storage and dynamically adjusts storage strategies based on the importance and complexity of the data, thereby optimizing storage resource utilization and reducing redundant storage. This method further enhances the reliability and safety of electro-hydraulic actuators during long-term operation in industrial applications.

[0007] The present invention provides a method for storing data of a small PLC of an electro-hydraulic actuator, comprising:

[0008] S1: Real-time collection of operating data generated by small and medium-sized PLCs of electro-hydraulic actuators during commissioning;

[0009] S2: pre-processing the operation data to obtain multiple operation data segments;

[0010] S3: Extract features from the running data segment through a self-supervised learning model to obtain key features;

[0011] S4: Process key features through a two-layer generative adversarial network to obtain basic data structure and reconstruct data;

[0012] The two-layer generative adversarial network includes:

[0013] A first-layer generative adversarial network, comprising a first generator model and a first discriminator model, configured to generate a basic data structure based on key features and complete missing data;

[0014] A second-layer generative adversarial network, comprising a second generator model and a second discriminator model, configured to refine and optimize data details based on the basic data structure and generate reconstructed data;

[0015] S41: Input the key features into the first generator model to obtain the basic data structure and complete the missing data;

[0016] S42: Inputting the basic data structure into the first discriminator model for discrimination to obtain first discriminant data, the first discriminant data being used to guide the training of the first generator model to obtain a basic data structure that meets the accuracy requirements;

[0017] S43: Inputting the basic data structure that meets the accuracy requirements into the second generator model to obtain reconstructed data;

[0018] S44: Inputting the reconstructed data into the second discriminator model for discrimination to obtain second discriminant data, which is used to guide the training of the second generator model until the reconstructed data meets the detail requirements;

[0019] S5: Classify and save the basic data structure and reconstructed data.

[0020] Furthermore, in step S2, the preprocessing includes data cleaning, normalization and data format conversion, including:

[0021] S21: Preliminary filtering of the operating data to remove obviously abnormal or incomplete data to obtain filtered operating data;

[0022] S22: Segment the filtered operation data in seconds to obtain multiple operation data segments.

[0023] Furthermore, the basic data structure includes the main patterns and overall structure of the data; the reconstructed data includes the high-frequency details and local features of the data.

[0024] Furthermore, step S41 includes:

[0025] S411: The input layer of the first generator model maps the key features to a high-dimensional feature space through a multi-layer fully connected neural network to obtain the initial feature vector. The calculation expression is:

[0026] ;

[0027] in, For the The initial feature vector of the running data segment, is the weight matrix of the input layer, is the bias vector, is the activation function, For the Key features of each operational data segment;

[0028] S412: The initial feature vector is passed to the hidden layer of the first generator model to obtain an intermediate feature map. The calculation expression is:

[0029] ;

[0030] in, For the The intermediate feature map of the running data segment, is the convolution kernel matrix, represents the convolution operation, is the bias vector of the convolutional layer;

[0031] S413: Decode the intermediate feature map into a basic data structure through deconvolution operation. The calculation expression is:

[0032] ;

[0033] in, For the The basic data structure of a running data segment, is the basic data deconvolution kernel matrix, is the bias vector of the deconvolution layer for the base data.

[0034] Furthermore, step S42 includes:

[0035] S421: The input layer of the first discriminator model performs preliminary feature extraction on the basic data structure through a multi-layer fully connected neural network to obtain a discriminant feature vector;

[0036] S422: passing the discriminant feature vector to the hidden layer of the first discriminator model to obtain a discriminant feature map;

[0037] S423: Inputting the discriminant feature map to the output layer of the first discriminator model to convert the discriminant feature map into first discriminant data;

[0038] S424: Optimizing parameters of the first discriminator model using the first discriminant data;

[0039] S425: Use the optimized first discriminator model to discriminate the basic data structure again, and feed the discrimination result back to the first generator model to obtain a basic data structure that meets the accuracy requirements.

[0040] Furthermore, step S43 includes:

[0041] S431: The input layer of the second generator model maps the basic data structure that meets the accuracy requirements to a high-dimensional feature space through a fully connected neural network to obtain an initial detail feature vector;

[0042] S432: passing the initial detail feature vector to the hidden layer of the second generator model to obtain a detail feature map;

[0043] S433: Decode the detail feature map into reconstructed data through deconvolution operation. The calculation expression is:

[0044] ;

[0045] in, For the Reconstructed data of running data segments, For the Detailed feature map of each running data segment, To reconstruct the data deconvolution kernel matrix, is the bias vector for reconstructing the deconvolution layer.

[0046] Furthermore, step S5 includes:

[0047] S51: classify the basic data structure and the reconstructed data according to importance to obtain high-priority data and low-priority data;

[0048] S52: Using a run-length encoding algorithm to perform redundancy compression on high-priority data;

[0049] S53: storing the compressed high priority data in the EEPROM of the small PLC;

[0050] S54: storing the low-priority data in the storage area of the touch screen. The storage area of the touch screen adopts a block-level management method, and the data is divided into blocks and stored;

[0051] S55: Periodically back up the data in the touch screen storage area, and upload low-priority data to an external storage device via the network.

[0052] The above one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects:

[0053] The present invention realizes intelligent processing of data preservation and recovery by introducing a self-supervised learning model and a two-layer generative adversarial network. In the data preservation stage, the self-supervised learning model is used to conduct in-depth analysis and feature extraction of the data. Through the training process of the first generator model and the first discriminator model, the system automatically identifies and extracts key features in the data, and generates basic data structures and completes missing data based on the features. This not only improves the accuracy and efficiency of data preservation, but also can dynamically adjust the storage strategy according to the importance and complexity of the data, thereby optimizing the utilization of storage resources and reducing the demand for redundant storage.

[0054] The present invention optimizes the adversarial loss function between the generator and the discriminator, so that the first generator model and the second generator model can gradually improve the quality of generated data during the training process. By introducing regularization terms and gradient penalty mechanisms, the stability of adversarial training and the authenticity of generated data are optimized, effectively avoiding the problems of overfitting and mode collapse, and ensuring that the system can provide high-quality data recovery results during hardware replacement or system fault recovery.

[0055] In terms of data storage and management, the present invention uses classified storage and redundant compression algorithms to classify and store data according to its importance, and preferentially stores important data in the EEPROM of the small PLC, while storing secondary data in the storage area of the touch screen. In addition, a lossless compression algorithm is used to compress and store the data, ensuring the security and recoverability of key data, optimizing the utilization efficiency of storage space, and avoiding the resource waste problem caused by redundant storage in traditional methods.

[0056] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0058] Figure 1 The present invention provides a flow chart of a method for storing data of a small PLC of an electro-hydraulic actuator. DETAILED DESCRIPTION

[0059] To make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below. Obviously, the embodiments described are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0060] In the description of the embodiments of the present invention, it should be noted that the terms "first", "second" and "third" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiment of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0061] The following combination Figure 1 The present invention describes a method for storing data of a small PLC of an electro-hydraulic actuator.

[0062] like Figure 1 As shown, a method for storing data of a small PLC of an electro-hydraulic actuator includes:

[0063] S1: Real-time collection of operating data generated by small and medium-sized PLCs of electro-hydraulic actuators during commissioning;

[0064] In some specific embodiments of the present invention, data preservation and recovery operations are performed on a key electro-hydraulic actuator on a device of a petroleum refining enterprise. During daily operation, a small PLC is responsible for real-time acquisition and processing of data from multiple sensors, including voltage, current, temperature, displacement and pressure signals. At the same time, a touch screen is used to display and store operation data. The system automatically saves the operation data for subsequent analysis and maintenance.

[0065] During the debugging process of the small PLC of the electro-hydraulic actuator, the operating data of the small PLC is collected in real time, including voltage sensor signals, current sensor signals, temperature sensor signals, displacement sensor signals, pressure sensor signals and control instruction data of the small PLC. Among them, each set of collected operating data is marked with a timestamp.

[0066] S2: pre-processing the operation data to obtain multiple operation data segments;

[0067] Preprocessing includes data cleaning, normalization, and data format conversion, including:

[0068] S21: Preliminary filtering of the operating data to remove obviously abnormal or incomplete data to obtain filtered operating data;

[0069] In some specific embodiments of the present invention, the real-time collected operation data is preliminarily filtered to remove obviously abnormal or incomplete data and retain complete and continuous operation data to obtain filtered operation data. The calculation expression of the filtered operation data D is:

[0070] ;

[0071] in, is the complete running data at the first time point, This is the complete running data at the second time point. For the Complete operation data at each time point, including sensor signals and small PLC instruction data;

[0072] S22: Segment the filtered running data in seconds to obtain multiple running data segments. The calculation expression is:

[0073] ;

[0074] in, For the Operation data segment, For the The first data in a running data segment, For the The second data in a running data segment, For the The mth data in an operation data segment, each operation data segment covers a complete operation cycle or time segment.

[0075] S3: Extract features from the running data segment through a self-supervised learning model to obtain key features;

[0076] A self-supervised learning model is used to analyze sensor signals in the time and frequency domains to extract key features related to the control behavior of small PLCs:

[0077] ;

[0078] in, For the The key features of each operational data segment are: For the The first feature of the running data segment, For the The second feature of the running data segment, For the The pth feature of a running data segment includes signal amplitude, frequency, phase and noise characteristics.

[0079] S4: Process key features through a two-layer generative adversarial network to obtain basic data structure and reconstruct data;

[0080] The two-layer generative adversarial network includes:

[0081] A first-layer generative adversarial network, comprising a first generator model and a first discriminator model, configured to generate a basic data structure based on key features and complete missing data;

[0082] The second-layer generative adversarial network includes a second generator model and a second discriminator model, which is used to refine and optimize the details of the data on the basic data structure to generate reconstructed data.

[0083] The reconstructed data contains high-frequency details and local features of the data;

[0084] S41: Inputting the running data features into the first generator model to obtain a basic data structure;

[0085] The basic data structure contains the main patterns and overall structure of the data;

[0086] S411: The input layer of the first generator model maps the key features to a high-dimensional feature space through a multi-layer fully connected neural network to obtain the initial feature vector. The calculation expression is:

[0087] ;

[0088] in, For the The initial feature vector of the running data segment, is the weight matrix of the input layer, is the bias vector, is the activation function, For the The key features of each operational data segment are:

[0089] The initial eigenvector is a high-dimensional representation of the key eigenvalues of the input, capturing the nonlinear characteristics of the input running data;

[0090] S412: The initial feature vector is passed to the hidden layer of the first generator model to obtain an intermediate feature map. The calculation expression is:

[0091] ;

[0092] in, For the The intermediate feature map of the running data segment, is the convolution kernel matrix, represents the convolution operation, is the bias vector of the convolutional layer;

[0093] The hidden layers of the first generator model consist of multi-layer convolutional neural networks to capture local correlations and spatial patterns among key eigenvalues;

[0094] S413: Decode the intermediate feature map into a basic data structure through deconvolution operation. The calculation expression is:

[0095] ;

[0096] in, For the The basic data structure of a running data segment, is the basic data deconvolution kernel matrix, is the bias vector of the basic data deconvolution layer; Restored to a structure similar to the input data, generating a low-resolution , where global information and main patterns are preserved;

[0097] By optimizing the first generator model Parameters, minimize the loss function of the first generator The calculation expression is:

[0098] ;

[0099] in, For the expected operation, is the first discriminator pair The judgment result of is the first regularization coefficient, It is the matching degree between the generated basic data structure and the key features.

[0100] S42: Inputting the basic data structure into the first discriminator model for discrimination to obtain first discriminant data, the first discriminant data being used to guide the training of the first generator model to obtain a basic data structure that meets the accuracy requirements;

[0101] S421: The input layer of the first discriminator model performs preliminary feature extraction on the basic data structure through a multi-layer fully connected neural network to obtain a discriminant feature vector;

[0102] The discriminant feature vector is a high-dimensional representation of the input underlying data structure;

[0103] S422: passing the discriminant feature vector to the hidden layer of the first discriminator model to obtain a discriminant feature map,

[0104] The hidden layer of the first discriminator model includes a multi-layer convolutional neural network to capture The local features and spatial patterns in the dataset are used to generate discriminant feature maps through convolution operations, where the discriminant feature maps reflect the authenticity characteristics of the underlying data structure.

[0105] S423: Inputting the discriminant feature map to the output layer of the first discriminator model to convert the discriminant feature map into first discriminant data;

[0106] The output layer of the first discriminator model is a fully connected layer, which converts the discriminant feature map into a single discriminant value, namely the first discriminant data. The first discriminant data is used to measure the similarity between the input basic data structure and the real data structure. The closer the first discriminant data is to 1, the closer the basic data structure is to the real data structure.

[0107] S424: Optimizing parameters of the first discriminator model using the first discriminant data;

[0108] By optimizing the parameters of the first discriminator model, the loss function of the first discriminator is minimized The calculation expression is:

[0109] ;

[0110] in, is the real data structure, For the expected operation, is the result of the first discriminator on the real data structure, is the second regularization coefficient, for and The contrast loss term between is used to guide the discriminator to learn more effective distinguishing features;

[0111] S425: Use the optimized first discriminator model to discriminate the basic data structure again, and feed the discrimination result back to the first generator model to obtain a basic data structure that meets the accuracy requirements;

[0112] The optimized first discriminator model is used to discriminate the basic data structure again, and the discrimination result is fed back to the first generator model. The calculation expression is:

[0113] ;

[0114] in, are the parameters of the optimized first generator model, are the parameters of the first generator model, is the learning rate, is the gradient of the loss function, is the weight of the regularization term, is the regularization loss term of the first generator.

[0115] S43: Inputting the basic data structure that meets the accuracy requirements into the second generator model to obtain reconstructed data;

[0116] S431: The input layer of the second generator model maps the basic data structure that meets the accuracy requirements to a high-dimensional feature space through a fully connected neural network to obtain an initial detail feature vector;

[0117] The initial detail feature vector is a high-dimensional representation of the underlying data structure;

[0118] S432: passing the initial detail feature vector to the hidden layer of the second generator model to obtain a detail feature map;

[0119] The hidden layer of the second generator model includes a multi-layer convolutional neural network, which is used to extract high-frequency detail features in the basic data structure and generate detail feature maps through convolution operations, wherein the detail feature maps represent high-frequency details and local features of the data;

[0120] S433: Decode the detail feature map into reconstructed data through deconvolution operation. The calculation expression is:

[0121] ;

[0122] in, For the Reconstructed data of running data segments, For the Detailed feature map of each running data segment, To reconstruct the data deconvolution kernel matrix, is the bias vector of the deconvolution layer for reconstructing data;

[0123] The deconvolution operation restores the detail feature map to a structure similar to the input data, generating reconstructed data containing high-frequency details and local features;

[0124] Perform detail enhancement and optimization on the generated reconstructed data, use adaptive enhancement methods to adjust the contrast and sharpness of the reconstructed data, and optimize the visual quality and detail performance of the reconstructed data;

[0125] By optimizing the parameters of the second generator model to minimize the loss function of the second generator

[0126] ;

[0127] in, For the expected operation, is the discrimination result of the second discriminator on the reconstructed data, is the first weight parameter, is the second weight parameter, for and The degree of match between It is a perceptual loss term used to optimize the perceptual quality and detail restoration of the reconstructed data.

[0128] S44: Input the reconstructed data into the second discriminator model for discrimination to obtain second discriminant data, which is used to guide the training of the second generator model until the reconstructed data meets the detail requirements.

[0129] In the first generator model and the first discriminator model During the training process, the optimized adversarial loss function Expressed as:

[0130] ;

[0131] in, is the distribution of real data structures, is the distribution of the input key eigenvalues, is the third regularization coefficient, For the first discriminator model to the i-th input data gradient;

[0132] In the second generator model and the second discriminator model During the training process, the optimized adversarial loss function Expressed as:

[0133] ;

[0134] in, The distribution of the generated basic data structure, is the second regularization coefficient, For the second discriminator model, the jth generated data block gradient;

[0135] The stochastic gradient descent algorithm is used to update the parameters of the generator and discriminator. The update rules are as follows:

[0136] ;

[0137] ;

[0138] in, For the updated , For the updated is the second learning rate, is the gradient of the loss function with respect to the parameters of the first generator model, is the gradient of the loss function to the parameters of the second generator model, by gradually updating the parameters and , continuously optimize the performance of the generator model so that the data structure it generates is closer to the real data;

[0139] In the adversarial loss function and When converging to the preset threshold, the final parameters are determined;

[0140] ;

[0141] in, is the first preset convergence threshold, is the second preset convergence threshold, are the first generator model parameters, are the second generator model parameters, is the first discriminator model parameter, are the second discriminator model parameters, For the first generator model tth and tth The change of loss function in iterations, For the second generator model t-th and The change of the loss function in the iteration is less than the preset threshold, and the model is considered to have converged and the final training parameters are locked.

[0142] S5: Classify and save the basic data structure and reconstructed data;

[0143] S51: classify the basic data structure and the reconstructed data according to importance to obtain high-priority data and low-priority data;

[0144] S52: Using a run-length encoding algorithm to perform redundancy compression on high-priority data;

[0145] S53: storing the compressed high priority data in the EEPROM of the small PLC;

[0146] S54: storing the low-priority data in the storage area of the touch screen. The storage area of the touch screen adopts a block-level management method, and the data is divided into blocks and stored;

[0147] S55: Periodically back up the data in the touch screen storage area, and upload low-priority data to an external storage device via the network.

[0148] During a scheduled maintenance, the small PLC of the electro-hydraulic actuator needed to undergo hardware replacement. To ensure that the replaced equipment could function normally and be restored to its original state, the technicians first used the method of the present invention to save data on the equipment. Specifically, the system first preprocessed the raw data collected by the small PLC, including data cleaning and normalization, to remove noise and abnormal data. Then, through a self-supervised learning model, feature extraction was performed on the preprocessed data to generate a feature representation representing the operating status of the equipment.

[0149] After generating the feature representation, the system is further trained using a two-layer generative adversarial network model. First, the first-layer generative adversarial network generates the basic data structure and completes the missing or incomplete data. Then, the second-layer generative adversarial network further optimizes the details of the data and generates reconstructed data. Finally, the system classifies and stores the data according to its importance. Important feature data and reconstructed data are preferentially stored in the EEPROM of the small PLC, while the completed data and secondary data are stored in the storage area of the touch screen and undergo redundant compression.

[0150] After the hardware replacement was complete, technicians performed a recovery operation using the stored data. The system first read the basic data structure and reconstruction data from the EEPROM and read the supplementary data from the touchscreen's storage area. This data was then fed into a two-layer generative adversarial network model. The system used the first-layer generator model to restore the basic data structure and generate complete operational data based on the supplementary data. The second-layer generator model then refined the basic data structure in detail, ultimately restoring the complete reconstructed data. The entire recovery process was swift and accurate, allowing the restored device to seamlessly integrate into the production line and continue operation.

[0151] To verify the effectiveness of the proposed method, we compared the performance of traditional data preservation and recovery methods with that of the proposed method in different scenarios. The test location was a company's production workshop, and the test samples were 20 electro-hydraulic actuators on the company's production line.

[0152] In the first test, the traditional redundant storage strategy and the method of the present invention were used to save data for the small PLC of the electro-hydraulic actuator, and the data recovery process after hardware replacement was simulated. In the traditional method, the redundantly stored data was directly restored by copying, while the method of the present invention performed intelligent recovery through a two-layer generative adversarial network model. The test results showed that when faced with a normally operating electro-hydraulic actuator, both the traditional method and the method of the present invention could successfully restore data, and the recovery time was not much different. The average recovery time of the traditional method was 2.5 hours, while that of the method of the present invention was 2.1 hours.

[0153] The results were significantly different in the second test, which introduced data loss or corruption. By deleting some sensor data to simulate data loss, the results showed that the recovery success rate using the traditional method dropped significantly, with only 60% of the tested devices successfully recovering complete data, and the recovery time extended to an average of 4.8 hours. In contrast, the method of the present invention demonstrated greater robustness, with a recovery success rate of 95% and an average recovery time of 2.6 hours, significantly outperforming the traditional method.

[0154] In the third test, the performance of the two methods under complex hardware failures was further tested. In the simulated hardware failure, the device not only lost some data, but also suffered data corruption. The test results showed that the traditional method was almost unable to cope with this complex situation, with a recovery success rate of only 30% and a recovery time of more than 6 hours. In contrast, the method of the present invention maintained an 80% recovery success rate and controlled the recovery time within 3.5 hours, demonstrating its superiority in complex situations.

[0155] In addition, to verify the efficiency of storage space utilization, the experimenters compared the performance of the two methods in terms of storage resources. The traditional method redundantly stores data, and the average storage requirement for each device is 1.8GB. However, the method of the present invention only requires 0.9GB of storage space through feature extraction and redundant compression, which greatly saves resources.

[0156] In summary, the method of the present invention not only outperforms traditional methods in terms of data recovery accuracy and success rate, but also has obvious advantages in storage resource utilization efficiency and recovery time. When dealing with complex data damage and loss, the method of the present invention exhibits higher robustness and adaptability, fully demonstrating its feasibility and effectiveness in actual industrial environments.

[0157] The present invention realizes intelligent processing of data preservation and recovery by introducing a self-supervised learning model and a two-layer generative adversarial network. In the data preservation stage, the self-supervised learning model is used to conduct in-depth analysis and feature extraction of the data. Through the training process of the first generator model and the first discriminator model, the system automatically identifies and extracts key features in the data, and generates basic data structures and completes missing data based on the features. This not only improves the accuracy and efficiency of data preservation, but also can dynamically adjust the storage strategy according to the importance and complexity of the data, thereby optimizing the utilization of storage resources and reducing the demand for redundant storage.

[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for storing data of a small PLC of an electro-hydraulic actuator, characterized in that: include: S1: Real-time collection of operating data generated by small and medium-sized PLCs of electro-hydraulic actuators during commissioning; S2: pre-processing the operation data to obtain multiple operation data segments; S3: Extract features from the running data segment through a self-supervised learning model to obtain key features; S4: Process key features through a two-layer generative adversarial network to obtain basic data structure and reconstruct data; The two-layer generative adversarial network includes: A first-layer generative adversarial network, comprising a first generator model and a first discriminator model, configured to generate a basic data structure based on key features and complete missing data; A second-layer generative adversarial network, comprising a second generator model and a second discriminator model, configured to refine and optimize data details based on the basic data structure and generate reconstructed data; S41: Input the key features into the first generator model to obtain the basic data structure and complete the missing data; S42: Inputting the basic data structure into the first discriminator model for discrimination to obtain first discriminant data, the first discriminant data being used to guide the training of the first generator model to obtain a basic data structure that meets the accuracy requirements; S43: Inputting the basic data structure that meets the accuracy requirements into the second generator model to obtain reconstructed data; S44: Inputting the reconstructed data into the second discriminator model for discrimination to obtain second discriminant data, which is used to guide the training of the second generator model until the reconstructed data meets the detail requirements; S5: Classify and save the basic data structure and reconstructed data; S51: classify the basic data structure and the reconstructed data according to importance to obtain high-priority data and low-priority data; S52: Using a run-length encoding algorithm to perform redundancy compression on high-priority data; S53: storing the compressed high priority data in the EEPROM of the small PLC; S54: storing the low-priority data in the storage area of the touch screen. The storage area of the touch screen adopts a block-level management method, and the data is divided into blocks and stored; S55: Periodically back up the data in the touch screen storage area, and upload low-priority data to an external storage device via the network.

2. The method for storing data of a small PLC of an electro-hydraulic actuator according to claim 1, characterized in that: In step S2, the preprocessing includes data cleaning, normalization and data format conversion, including: S21: Preliminary filtering of the operating data to remove obviously abnormal or incomplete data to obtain filtered operating data; S22: Segment the filtered operation data in seconds to obtain multiple operation data segments.

3. The method for storing data of a small PLC of an electro-hydraulic actuator according to claim 1, characterized in that: The basic data structure contains the main patterns and overall structure of the data, and the reconstructed data contains the high-frequency details and local features of the data.

4. The method for storing data of a small PLC of an electro-hydraulic actuator according to claim 1, characterized in that: Step S41 includes: S411: The input layer of the first generator model maps the key features to a high-dimensional feature space through a multi-layer fully connected neural network to obtain the initial feature vector. The calculation expression is: ； in, For the The initial feature vector of the running data segment, is the weight matrix of the input layer, is the bias vector, is the activation function, For the Key features of each operational data segment; S412: The initial feature vector is passed to the hidden layer of the first generator model to obtain an intermediate feature map. The calculation expression is: ； in, For the The intermediate feature map of the running data segment, is the convolution kernel matrix, represents the convolution operation, is the bias vector of the convolutional layer; S413: Decode the intermediate feature map into a basic data structure through deconvolution operation. The calculation expression is: ； in, For the The basic data structure of a running data segment, is the basic data deconvolution kernel matrix, is the bias vector of the deconvolution layer for the base data.

5. The method for storing data of a small PLC of an electro-hydraulic actuator according to claim 1, characterized in that: Step S42 includes: S421: The input layer of the first discriminator model performs preliminary feature extraction on the basic data structure through a multi-layer fully connected neural network to obtain a discriminant feature vector; S422: passing the discriminant feature vector to the hidden layer of the first discriminator model to obtain a discriminant feature map; S423: Inputting the discriminant feature map to the output layer of the first discriminator model to convert the discriminant feature map into first discriminant data; S424: Optimizing parameters of the first discriminator model using the first discriminant data; S425: Use the optimized first discriminator model to discriminate the basic data structure again, and feed the discrimination result back to the first generator model to obtain a basic data structure that meets the accuracy requirements.

6. The method for storing data of a small PLC of an electro-hydraulic actuator according to claim 1, characterized in that: Step S43 includes: S431: The input layer of the second generator model maps the basic data structure that meets the accuracy requirements to a high-dimensional feature space through a fully connected neural network to obtain an initial detail feature vector; S432: passing the initial detail feature vector to the hidden layer of the second generator model to obtain a detail feature map; S433: Decode the detail feature map into reconstructed data through deconvolution operation. The calculation expression is: ； in, For the Reconstructed data of running data segments, For the Detailed feature map of each running data segment, To reconstruct the data deconvolution kernel matrix, is the bias vector of the deconvolution layer for reconstructing data.

Images