Nuclear power intelligent isolation lock system data processing method and device based on Internet of Things

Abstract

The invention provides a nuclear power intelligent isolation lock system data processing method and device based on the Internet of Things, and the method comprises the steps: obtaining the monitoring data implicit representation of the nuclear power isolation lock Internet of Things monitoring data and the diagnosis report implicit representation of the abnormality diagnosis report based on the obtained nuclear power isolation lock Internet of Things monitoring data and the corresponding abnormality diagnosis report; the abnormal information can be automatically determined, and the abnormal information diagnosis speed is increased. And the abnormal prior mark is determined based on the data of different dimensions, so that the reliability of the target abnormal prior mark is higher. The invention further provides an intelligent isolation lock system applying the method, the system sets the operation authority for the user and limits the time range of the key authority through dual means of task authorization and equipment consistency verification, and the accuracy of an implementation object is ensured; an operator of the system can monitor the equipment isolation condition on the electronic working platform in real time, operation content can be updated in time, it is ensured that the isolation process is traceable, operation records are stored, and it is ensured that the operation records can be inquired subsequently.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a data processing method and device for a nuclear power intelligent isolation lock system based on the Internet of Things. Background Art

[0002] With the rapid development of Internet of Things technology, its application in the industrial field is becoming more and more extensive, especially in the nuclear power industry. Internet of Things technology provides strong support for improving the safety and operation efficiency of nuclear power plants. During the operation of nuclear power plants, it is necessary to monitor the status of various equipment and environmental parameters in real time to ensure nuclear safety and prevent potential accident risks. Therefore, as one of the key safety measures in nuclear power plants, the real-time and accuracy of the monitoring data of the nuclear power isolation lock system is crucial for anomaly detection, fault diagnosis and rapid response. The traditional data processing method of nuclear power isolation lock system often relies on manual inspection and regular maintenance. This method is not only inefficient, but also difficult to detect potential safety hazards in time. In addition, most traditional data processing methods are based on simple threshold judgment or rule matching, which is difficult to handle complex and changeable monitoring data, resulting in limited accuracy and reliability of anomaly detection. Summary of the invention

[0003] In view of this, the present application provides a data processing method and device for a nuclear power intelligent isolation lock system based on the Internet of Things. The technical solution of the present application is implemented as follows: On the one hand, the present application provides a data processing method for a nuclear power intelligent isolation lock system based on the Internet of Things, the method comprising: obtaining the Internet of Things monitoring data of the nuclear power isolation lock and the abnormal diagnosis report of the Internet of Things monitoring data of the nuclear power isolation lock, and performing implicit representation mining on the Internet of Things monitoring data of the nuclear power isolation lock and the abnormal diagnosis report, respectively, to obtain the implicit representation of the monitoring data and the implicit representation of the diagnosis report; merging the implicit representation of the monitoring data and the implicit representation of the diagnosis report to obtain a first fused implicit representation; obtaining abnormal information involvement features, wherein the abnormal information involvement features include implicit representations of marks corresponding to multiple abnormal prior marks in a preset abnormal information involvement topological structure; the multiple abnormal prior marks represent multiple subclassifications under the abnormal classification; For each of the abnormal prior markers, the implicit representation of the mark of the abnormal prior marker includes not only the abnormal information of the abnormal prior marker, but also the abnormal information of the abnormal prior marker connected to the corresponding abnormal prior marker in the topological structure involved in the abnormal information; the first fused implicit representation is merged with the implicit representation of each mark in the feature involved in the abnormal information to obtain the second fused implicit representations corresponding to the multiple abnormal prior markers; based on the second fused implicit representations corresponding to the multiple abnormal prior markers, the target abnormal prior marker corresponding to the nuclear power isolation lock IoT monitoring data is determined from the multiple abnormal prior markers; the target abnormal prior marker is used to indicate the abnormal information in the nuclear power isolation lock IoT monitoring data.

[0004] On the other hand, the present application provides a data processing device, including a memory and a processor, wherein the memory stores a data processing device program that can be run on the processor, and the processor implements the steps in the above method when executing the program.

[0005] The beneficial effects of the present application include at least the following: the data processing method and device of the nuclear power intelligent isolation lock system based on the Internet of Things provided by the present application, based on obtaining the Internet of Things monitoring data of the nuclear power isolation lock and the corresponding abnormal diagnosis report, obtain the monitoring data implicit representation of the Internet of Things monitoring data of the nuclear power isolation lock and the diagnostic report implicit representation of the abnormal diagnosis report, so that the implicit representation of the monitoring data and the implicit representation of the diagnostic report can be merged to obtain the first fused implicit representation. Based on the obtained first fused implicit representation, the first fused implicit representation is merged with the implicit representation of each mark in the preset abnormal information involving feature to obtain multiple second fused implicit representations, so as to determine the target abnormal prior mark corresponding to the Internet of Things monitoring data of the nuclear power isolation lock according to the multiple second fused implicit representations. The present application can automatically determine the abnormal information and improve the abnormal information diagnosis speed. And the abnormal prior mark is determined based on data of different dimensions, so that the reliability of the target abnormal prior mark is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1A schematic diagram of the implementation flow of a data processing method for a nuclear power intelligent isolation lock system based on the Internet of Things provided in an embodiment of the present application.

[0007] Figure 2 A hardware entity schematic diagram of a data processing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0008] The present application provides a data processing method for a nuclear power intelligent isolation lock system based on the Internet of Things, which can be executed by a processor of a data processing device. The data processing device can refer to a server, a laptop, a tablet computer, a desktop data processing device, a mobile device (such as a mobile phone, a portable video player, a personal digital assistant, a dedicated messaging device, a portable gaming device) and other devices with data processing capabilities.

[0009] Figure 1 A schematic diagram of the implementation flow of a data processing method for a nuclear power intelligent isolation lock system based on the Internet of Things provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method includes: Step S10: Obtain the Internet of Things monitoring data of the nuclear power isolation lock and the abnormal diagnosis report of the Internet of Things monitoring data of the nuclear power isolation lock, and perform implicit representation mining on the Internet of Things monitoring data of the nuclear power isolation lock and the abnormal diagnosis report respectively to obtain implicit representation of the monitoring data and implicit representation of the diagnosis report.

[0010] The data processing device first establishes a connection with the nuclear power isolation lock IoT monitoring system, which is widely deployed in key areas of nuclear power plants. It collects and transmits data in real time through sensors installed on various equipment and environmental parameter monitoring points. These data include but are not limited to the switch status of the isolation lock, the temperature, humidity, radiation level of the surrounding environment, and any other physical quantities related to the safe operation of the nuclear power plant. These data are recorded in the form of time series and constitute the nuclear power isolation lock IoT monitoring data. At the same time, the data processing device also accesses the abnormal diagnosis system, which performs a preliminary analysis of the IoT monitoring data based on a preset set of rules, expert experience or machine learning models, and generates an abnormal diagnosis report. These reports record in detail the abnormal events found during the monitoring cycle, such as a sudden increase in the radiation level in a certain area, an abnormal opening or closing of an isolation lock, etc. The abnormal diagnosis report not only points out the specific manifestations of the abnormality, but may also include a preliminary inference of the cause of the abnormality and recommended countermeasures.

[0011] Next, the data processing device performs implicit representation mining on the acquired nuclear power isolation lock IoT monitoring data and abnormal diagnosis report. The purpose of this step is to extract deeper information from the original data that is more useful for abnormal detection and analysis. For IoT monitoring data, the data processing device can use deep learning models, such as recurrent neural networks (RNN) or their variants, long short-term memory networks (LSTM), to process time series data. The LSTM model is particularly suitable for capturing long-term dependencies in data. For abnormal diagnosis reports, the data processing device can use natural language processing (NLP) technology to extract key information. Specifically, the report text can be encoded using a BERT model based on a Transformer structure. The BERT model has learned rich language representation capabilities through pre-training tasks and can accurately understand the semantic information in the text. When processing abnormal diagnosis reports, BERT converts the text into a series of word embedding vectors, captures the relationship between words through a self-attention mechanism, and finally generates an implicit representation vector that summarizes the core information about abnormal events in the report. Through the above implicit representation mining process, the data processing device converts complex IoT monitoring data and text diagnosis reports into structured implicit representations. These implicit representations not only retain the important information in the original data, but also remove redundancy and noise, providing a more efficient and accurate data foundation for subsequent data fusion, anomaly detection and analysis.

[0012] Step S20: Merge the monitoring data implicit representation and the diagnosis report implicit representation to obtain a first fused implicit representation.

[0013] In step S20, the data processing device designs a reasonable fusion mechanism to combine the two implicit representations. A common method is to use feature concatenation or weighted sum. However, in the specific scenario of the nuclear power intelligent isolation lock system, considering that the importance of monitoring data and diagnostic reports in anomaly detection may be different, weighted summation may be more appropriate because it allows the system to adjust the weights of different information sources according to actual application requirements.

[0014] In specific implementation, the data processing device can assign a weight coefficient (for example, w1 and w2, and w1+w2=1) to the implicit representation of monitoring data and the implicit representation of diagnostic report respectively, and then linearly combine the two implicit representation vectors according to the weight. Assume that the implicit representation of monitoring data is v monitor , the implicit representation of the diagnostic report is v report , then the first fusion implicitly represents v fused It can be calculated as: v fused =w1·v monitor+w2· vreport ; In practical applications, the weight coefficients w1 and w2 can be tuned on the training data set through cross-validation, grid search and other methods to maximize the accuracy of anomaly detection or other performance indicators.

[0015] For example, if it is found that some physical quantity changes in the monitoring data are highly correlated with abnormal events, and some keywords or sentence patterns in the diagnostic report also make important contributions to anomaly detection, then the data processing device can increase the weights of these information sources accordingly. In this way, the first fused implicit representation not only contains the key information in the monitoring data and the diagnostic report, but also reflects the difference in importance of this information in anomaly detection through weight adjustment. Finally, the first fused implicit representation v fused It will be used as input for further anomaly detection, classification or prediction in subsequent steps. In the practical application of nuclear power intelligent isolation lock system, the successful implementation of this step is of great significance to improve the accuracy and timeliness of anomaly detection.

[0016] Step S30: Acquire abnormal information-involved features, wherein the abnormal information-involved features include implicit representations of labels corresponding to multiple abnormal prior labels in a preset abnormal information-involved topological structure; multiple abnormal prior labels represent multiple subclassifications under the abnormal classification; for each abnormal prior label, the implicit representation of the label includes not only the abnormal information of the abnormal prior label, but also the abnormal information of the abnormal prior label connected to the corresponding abnormal prior label in the abnormal information-involved topological structure.

[0017] During the operation of nuclear power facilities, abnormal events often do not occur in isolation, and there may be direct or indirect connections between them. For example, the failure of a pump may cause abnormal pressure in the cooling system, which in turn affects the containment temperature of the reactor. The mutual influence and transmission between these abnormal events constitute a complex abnormal information network in the nuclear power system. The abnormal information involvement feature is used to describe the characteristic information of the relationship between abnormal events in this network. By obtaining the abnormal information involvement feature, the data processing device can better understand the overall picture of the abnormal event and identify potential chain reactions and risk paths. This is of great significance for taking preventive measures in a timely manner and avoiding the expansion of accidents.

[0018] Before step S30, the data processing device has obtained the implicit representation of monitoring data and the implicit representation of diagnosis report through steps S10 and S20, and obtained the first fused implicit representation through fusion. Now, the data processing device needs to construct the abnormal information involving topological structure based on this information and domain knowledge.

[0019] The topological structure involved in abnormal information is a directed graph, in which the nodes represent different abnormal a priori markers, and the edges represent the causal relationship or correlation between abnormal events. These abnormal a priori markers are pre-defined and used to characterize multiple sub-categories under the abnormal classification. For example, in a nuclear power system, abnormal classifications may include equipment failure, abnormal environmental parameters, improper operation, etc., and each category has multiple specific sub-categories, such as pump failure, excessive cooling water temperature, etc. In order to construct this topological structure, the data processing device may need to refer to multiple data sources such as historical abnormal records, expert experience, equipment manuals, etc. Through technical means such as data analysis, rule extraction and pattern recognition, the data processing device can identify the potential relationships between abnormal events and express these relationships in the form of a topological graph.

[0020] In the topological structure involving abnormal information, each node (i.e., abnormal prior label) requires a corresponding label implicit representation. This representation not only contains the abnormal information of the abnormal prior label itself, but also contains the abnormal information of other abnormal prior labels connected to it. This representation method helps the data processing device to comprehensively consider the impact of multiple abnormal events in subsequent analysis. The process of generating label implicit representation may involve a variety of machine learning models and techniques. For example, the data processing device can use graph embedding technology to convert nodes in the topological structure into points in a low-dimensional vector space to obtain a vector representation of each node. This process is usually achieved by optimizing an objective function that aims to maintain the consistency of the structural information of the graph in the embedding space. In addition, the data processing device can also use the attention mechanism to enhance the ability of the label implicit representation. The attention mechanism allows the model to focus on other nodes that are most relevant to each node when processing it, so as to more accurately capture the relationship between abnormal events when generating the label implicit representation.

[0021] For example, suppose that in the nuclear power intelligent isolation lock system, two abnormal a priori tags are concerned: "pump failure" and "cooling water temperature is too high". In the abnormal information involvement topology, these two tags are connected by a directed edge, indicating that pump failure may cause the cooling water temperature to be too high.

[0022] Basic abnormal information includes the operating status of the pump, fault code, occurrence time, etc. The related abnormal information obtains the information of other abnormal prior tags connected to "pump failure" through the edges in the topological structure. For example, if "pump failure" is connected to "power fluctuation", the abnormal information of "power fluctuation" will also be included in the implicit representation of the tag of "pump failure". The vector representation uses graph embedding technology to convert the above information into a low-dimensional vector, such as a 128-dimensional floating-point array. This vector can capture the complex relationship between "pump failure" and its related abnormal events.

[0023] Basic abnormal information includes the real-time temperature of cooling water, historical temperature change trends, time exceeding the threshold, etc. Related abnormal information also obtains information about other abnormal prior tags connected to "cooling water temperature is too high" through the topological structure. For example, "pump failure" is one of the possible causes of "cooling water temperature is too high", and its information will be integrated into the implicit representation of the tag "cooling water temperature is too high". Graph embedding technology is also used to generate a low-dimensional vector representation.

[0024] After obtaining the implicit representation of each anomaly prior label, the data processing device can use these features to perform more in-depth anomaly detection and analysis. For example: By comparing the similarity between the implicit representation of the current monitoring data and the implicit representation of each abnormal prior marker, the data processing device can predict the upcoming abnormal event and issue a warning in advance. When an abnormal event occurs, the data processing device can infer the root cause and propagation path of the abnormal event by analyzing the topological structure involved in the abnormal information and the implicit representation of the relevant markers. Based on the analysis results of the characteristics involved in the abnormal information, the data processing device can provide targeted decision-making suggestions for operation and maintenance personnel to quickly respond to and handle abnormal events.

[0025] Step S40: Merge the first fused implicit representation with the implicit representations of each mark in the abnormal information-related feature to obtain second fused implicit representations corresponding to multiple abnormal prior marks.

[0026] Step S40 is a key step to further explore the potential association of data and accurately identify the abnormal type. At this stage, the data processing device deeply integrates the abnormal information involvement features (i.e., the implicit representation of the labels of multiple abnormal prior labels) obtained in the previous step (S30) with the first fused implicit representation generated in step S20, aiming to analyze the intrinsic connection between the current monitoring data and the known abnormal pattern from a multi-dimensional perspective.

[0027] In step S40, the data processing device first merges the first fused implicit representation, which combines the essence of monitoring data and diagnostic reports and is regarded as a comprehensive overview of the current system status, with each labeled implicit representation in the feature library involved in the abnormal information. The "merge" operation here is actually achieved through a certain fusion algorithm (such as feature concatenation, weighted summation, attention mechanism, etc.). The specific algorithm to be selected depends on the complexity of the system design and the characteristics of the target task. Taking feature concatenation as an example, assuming that the first fused implicit representation is a D-dimensional vector v fused , and the label of a certain abnormal prior label "cooling system fault" is implicitly represented as another E-dimensional vector vcooling_fault. The data processing device merges these two vectors into a new vector through feature concatenation, that is, the second fusion implicit representation: v second_fused =[v fused ;v cooling_fault ]; Among them, [·;·] represents the vector concatenation operation, and the resulting vector v second_fused The dimension will be D+E.

[0028] Through step S40, the data processing device deeply integrates the first fused implicit representation with each labeled implicit representation in the abnormal information-related feature library to generate multiple second fused implicit representations. These representations not only contain the real-time status information of the monitoring data, but also incorporate the knowledge and wisdom of abnormal events that have occurred in history, providing strong data support for tasks such as abnormal classification and risk assessment in subsequent steps. In the nuclear power intelligent isolation lock system.

[0029] Step S50: Determine a target abnormal priori mark corresponding to the nuclear power isolation lock IoT monitoring data from among the multiple abnormal prior marks according to the second fused implicit representations corresponding to the multiple abnormal prior marks; the target abnormal prior mark is used to indicate abnormal information in the nuclear power isolation lock IoT monitoring data.

[0030] In step S50, the data processing device uses a classification algorithm or decision mechanism to evaluate the match between each second fused implicit representation and the current monitoring data. This evaluation process may involve the reasoning step of a machine learning model, such as using a support vector machine (SVM), a random forest, or a deep neural network (DNN). These models have been trained with a large amount of historical data and have learned how to predict the output category (i.e., the abnormal prior label) based on the input feature vector (i.e., the second fused implicit representation).

[0031] Taking a deep neural network as an example, suppose that a trained classification neural network is deployed inside the system. The network takes the second fused implicit representation as input, and the output layer contains neurons equal to the number of abnormal prior markers. The output value of each neuron represents the probability that the input data belongs to the corresponding abnormal category. The data processing device will input all the second fused implicit representations into the network in sequence and obtain the predicted probability of each abnormal prior marker. Subsequently, the data processing device compares these predicted probabilities and selects the abnormal prior marker with the highest probability value as the target abnormal prior marker. This marker is the abnormal type that the system considers to be the most matching with the current nuclear power isolation lock IoT monitoring data, and is used to indicate abnormal information in the monitoring data. Assume that there are three abnormal prior markers: "A-Cooling system failure", "B-Abnormal increase in radiation level", and "C-Safety system response delay", and their corresponding second fused implicit representations are v A_fused 、v B_fused 、v C_fusedAfter the inference of the classification neural network, the data processing device obtains the following prediction probabilities: P(A)=0.2; P(B)=0.7; P(C)=0.1.

[0032] Since P(B) has the highest probability, the data processing device determines that "B-abnormal increase in radiation level" is a target abnormality priori mark, and issues a corresponding alarm or processing instruction accordingly.

[0033] Through step S50, the data processing device identifies the target abnormal prior marker that best matches the current nuclear power isolation lock IoT monitoring data from multiple abnormal prior markers. This process not only relies on the high-quality second fusion implicit representation generated by the previous step, but also relies on advanced machine learning models and efficient classification algorithms. In practical applications, such a mechanism can significantly improve the abnormal identification capability of the nuclear power intelligent isolation lock system and provide strong protection for the safe operation of nuclear power plants.

[0034] As an implementation mode, before obtaining the abnormal information involving features, the method further includes a process of determining the abnormal information involving features, specifically including: Step S201: Acquire an abnormal information-involved topological structure; the abnormal information-involved topological structure includes a plurality of abnormal a priori markers, and connections are made between abnormal a priori markers of involved situations.

[0035] The abnormal information involvement topology is a graphical representation method used to intuitively show the association and mutual influence between different abnormal types (i.e., abnormal prior markers). In this structure, each node represents a specific abnormal prior marker, such as "cooling system failure", "radiation leakage", "electrical system failure", etc. These abnormal prior markers are predefined based on historical data and expert knowledge, and are used to quickly identify and classify abnormal events. The edges between nodes represent the involvement relationship between abnormal prior markers. This relationship can be a direct causal relationship, a spatial adjacent relationship, or a temporal successive relationship. For example, if historical data shows that "cooling system failure" often leads to "reactor temperature increase", then in the topology structure, there will be a directed edge connecting the "cooling system failure" node and the "reactor temperature increase" node, and the arrow points to the direction of the influence.

[0036] In order to construct such a topological structure, the data processing device adopts a series of data processing and analysis technologies. First, the collected abnormal information is preprocessed, including data cleaning, deduplication, formatting and other operations to ensure the accuracy and consistency of the data. Then, the association rules or patterns between abnormal prior markers are extracted from the preprocessed data using data mining and pattern recognition algorithms (such as association rule mining, causal reasoning algorithms, etc.). After extracting the association rules, the data processing device will construct the topological structure involved in the abnormal information according to these rules. This process may involve relevant algorithms and data structures in graph theory, such as adjacency lists, adjacency matrices, etc., which are used to efficiently store and query the connection relationship between nodes.

[0037] Finally, the data processing device obtains a topological structure containing multiple abnormal a priori markers and their connection relationships. This structure not only reflects the various abnormal situations that may occur during the operation of the nuclear power plant, but also reveals the internal connections and mutual influences between these abnormal situations, providing strong support for subsequent data fusion, anomaly detection and response.

[0038] For example, in a nuclear power intelligent isolation lock system, if the data processing device detects an abnormally high radiation level in a certain area (corresponding to the "radiation leakage" abnormal prior marker), it can use the abnormal information involving the topological structure to quickly infer other systems or areas that may be affected (such as cooling systems, electrical systems, etc.), thereby taking more comprehensive and effective response measures.

[0039] Step S202: construct a connection relationship array and a feature embedding array according to the topological structure involved in the abnormal information; the feature embedding array includes feature embedding results corresponding to multiple abnormal prior markers, and the connection relationship array is used to describe the connection between multiple abnormal prior markers.

[0040] In step S202, the data processing device further constructs two core data structures based on the previously acquired abnormal information involving topological structure (output of step S201): connection relationship array and feature embedding array. These two arrays provide a basis for subsequent data processing, analysis and abnormality identification.

[0041] The connection relationship array is used to describe the connection between each abnormal a priori mark in the topological structure involved in the abnormal information. In the nuclear power system, these abnormal a priori marks may include "cooling system failure", "electrical system failure", "radiation leakage", etc. The connection relationship between them reflects the mutual influence and transmission path between abnormal events. In order to construct the connection relationship array, the data processing device traverses the nodes (i.e., abnormal a priori marks) and edges (i.e., connection relationships) in the topological structure involved in the abnormal information. For each node, the data processing device checks the connection status between it and all other nodes. If there is a connection between two nodes, the corresponding position in the connection relationship array is marked as 1 (or other values ​​indicating "there is a connection"); if there is no connection, it is marked as 0. For example, suppose there are three abnormal a priori marks: A (cooling system failure), B (electrical system failure), C (radiation leakage), and it is known that there is a connection between A and B, and there is also a connection between B and C, but there is no direct connection between A and C. Then the connection relationship array may be as follows (rows and columns represent different abnormal a priori marks): ; This array clearly shows the connection between the three abnormal prior markers A, B, and C.

[0042] The feature embedding array contains the feature embedding results of each abnormal prior marker. Feature embedding is a technology that maps discrete objects (such as abnormal prior markers) to continuous vector space, so that these objects can be processed and analyzed in the form of numerical vectors in the data processing device. In the nuclear power intelligent isolation lock system, the feature embedding results may contain a variety of feature information related to the abnormal prior marker, such as historical occurrence frequency, scope of impact, severity, etc. In order to construct the feature embedding array, the data processing device uses machine learning techniques (such as word embedding, autoencoder, etc.) to extract and encode the features of the abnormal prior marker.

[0043] Assuming that the feature embedding result of each abnormal prior marker is a 10-dimensional vector, then for the three abnormal prior markers A, B, and C mentioned above, the feature embedding array may be as follows (each row represents the feature embedding vector of an abnormal prior marker): ; Among them, v A1 ,v A2 ,…,v A10 are the elements in the feature embedding vector of the abnormal prior marker A. Similarly, v B and v C Represent the feature embedding vector elements of B and C respectively.

[0044] Through step S202, the data processing device converts the topological structure involved in the abnormal information into two specific data structures: a connection relationship array and a feature embedding array. These two arrays provide strong data support for subsequent data fusion, abnormality detection and identification.

[0045] Step S203: Merge the connection relationship array and the feature embedding array to obtain a first-generation integrated array.

[0046] In step S203, the data processing device merges the connection relationship array with the feature embedding array to generate an integrated array that contains both the connection relationship information between abnormal priori labels and the feature information of each label, which serves as the basis for subsequent iterative updates.

[0047] First, reviewing the output of step S202, the data processing device has obtained two core arrays: a connection relationship array and a feature embedding array. The connection relationship array is a two-dimensional array used to describe the connection between abnormal prior markers, and its elements usually indicate whether there is a connection between the two markers (such as 0 for no connection and 1 for connection). The feature embedding array is a two-dimensional array, in which each row represents a feature embedding vector of an abnormal prior marker, containing multiple feature information of the marker. In order to merge the two arrays, the data processing device can adopt a variety of strategies, but a common and effective method is to use the idea of ​​Graph Neural Network (GNN) for information aggregation. However, in order to simplify the description and fit the actual scenario of the nuclear power intelligent isolation lock system, a more intuitive matrix operation method is used here to merge.

[0048] Specifically, the data processing device may first keep the feature embedding array unchanged as the main part of the integration array because it contains rich feature information. Then, for the connection relationship array, the data processing device may regard it as a weighted matrix to guide the propagation and fusion of feature information. However, in direct matrix operations, the elements (0 or 1) of the connection relationship array may not be sufficient to express complex weighted relationships, so in practical applications, the elements of the connection relationship array may be adjusted according to the strength or importance of the connection relationship (although it is kept as 0 or 1 in this simplified example). Next, the data processing device applies the connection relationship array to the feature embedding array through matrix multiplication or a custom aggregation function to update the feature representation of each abnormal prior marker, thereby taking into account the influence of other markers connected to it. However, since matrix multiplication usually requires the operands to meet specific dimensional requirements and may not be the most direct method in this scenario, a simplified "weighted summation" idea is adopted here to simulate the merging process.

[0049] Assume that the feature embedding array is F∈Rn×d , where n is the number of abnormal prior tags, d is the dimension of the feature vector of each tag; the connection relationship array is A∈R n×n , where the element aij represents the connection relationship between the tag i and the tag j (here simplified to 0 or 1). In order to simulate the effect of weighted summation, instead of directly performing matrix multiplication, the data processing device can design an iterative or traversal process to update the feature representation of each tag according to the connection relationship. However, to simplify the explanation, it is assumed that there is an "integration function" g(·) that can take the connection relationship array and the feature embedding array as input and output the first generation of the integrated array Z(0).

[0050] In actual operation, since direct application may be relatively complicated, a simplified example can be imagined: for each tag i, the data processing device traverses all tags j (i.e., a) connected to it in the connection relationship array. ij =1), and then update the feature representation of tag i according to some rule (such as averaging, summing or weighted summing). However, in this highly simplified description, the form of the first-generation integrated array Z(0) is directly given, which may be a transformation or preservation of the feature embedding array F, while implicitly considering the influence of the connection relationship (although this influence is not explicitly expanded in this description). Ultimately, the first-generation integrated array Z(0) contains both the original feature embedding information and the preliminary fusion of the connection relationship information between the abnormal prior tags, providing a starting point for iterative updates in subsequent steps. In the subsequent iterative process (such as steps S204 to S205), the data processing device will further mine and integrate richer information based on this integrated array.

[0051] It should be noted that the “integration function” g(·) in the above description is a highly abstract concept used to schematically illustrate the merging process. In practical applications, this function may be implemented by a complex graph neural network layer, attention mechanism, or other advanced machine learning models.

[0052] Step S204: In the current generation starting from the second generation after the first generation, the integration array output by the previous generation is merged with the connection relationship array to obtain the integration array output by the current generation.

[0053] In step S203, the data processing device has generated a first generation of integrated arrays based on the connection relationship array and the feature embedding array. This integrated array preliminarily integrates the feature information of the abnormal prior markers and the connection relationship information between them. However, due to the complexity of the nuclear power system, the interaction between abnormal information is often multi-level and nonlinear, so it is necessary to gradually reveal these deep relationships through an iterative process.

[0054] In step S204, the data processing device performs a similar merging operation for each generation starting from the second generation. Specifically, for the current generation (starting from the second generation), the data processing device merges the integrated array output by the previous generation with the connection relationship array to generate a new integrated array output by the current generation. This process is repeated until a certain stop condition is met (such as the number of iterations reaches a preset upper limit, the change in the integrated array is less than a certain threshold, etc.).

[0055] The specific implementation of the merging operation may vary depending on the application scenario and the selected technology. In the nuclear power intelligent isolation lock system, since the relationship between abnormal information is complex and dynamically changing, it may be necessary to adopt a machine learning model that can process graph structured data, such as a graph neural network (GNN).

[0056] This process may involve complex mathematical operations and the application of machine learning models, but a simplified example can be given here to illustrate the concept. Assuming that weighted average is used as the aggregation function, the updated feature representation of the current node i may be expressed as: ; in, is the feature representation of node i in the second generation integrated array, Neighbors(i) is the set of neighbor nodes of node i in the connection relationship array A, a ij is the strength of the connection between node i and node j in the connection relationship array (0 or 1 in this simple example), is the feature representation of node j in the previous generation integrated array.

[0057] Through the iterative merging process of step S204, the data processing device can gradually deepen its understanding of the complex relationship between abnormal information. Each generation of integrated arrays contains richer and more detailed information than the previous generation, providing a solid foundation for subsequent abnormal detection and response.

[0058] Step S205: Determine the next generation as the current generation, jump to the step of merging the integrated array output by the previous generation with the connection relationship array to iterate until the first stop judgment requirement is met. In step S204, the data processing device has completed the generation of the current generation integrated array, which is obtained by merging the previous generation integrated array with the connection relationship array. Step S205 is responsible for controlling the progress of the entire iteration process to ensure that the system can continuously repeat the operation in step S204 according to the set logic until a certain termination condition is met. Specifically, at the end of each iteration, the data processing device will regard the current generation as the "previous generation" and prepare to enter the next iteration. In this process, the "next generation" does not actually exist because it will be created in the next iteration. However, in order to control the continuity of the process, the "next generation to be entered" can be regarded as the logical "next generation" and set as the current "current generation". Then, the data processing device checks whether the first stop judgment requirement is met. This requirement is usually based on a series of preset conditions, such as whether the number of iterations has reached the preset maximum value, whether the change in the integrated array is less than a certain threshold, or whether a certain performance indicator (such as the accuracy and recall rate of anomaly detection) has been achieved. These conditions can be customized according to specific application scenarios and performance requirements.

[0059] If the first stop judgment requirement is met, the iteration process will terminate, and the data processing device will use the last generation of integrated arrays to perform subsequent abnormality detection, classification or response tasks. If the stop condition is not met, the iteration process will continue, that is, the current generation of integrated arrays will be regarded as the previous generation, and the process will jump to step S204 to repeat the operation of merging the previous generation of integrated arrays with the connection relationship array.

[0060] Assume that in the nuclear power intelligent isolation lock system, a stop condition with a maximum number of iterations of 10 is set. After each iteration, the data processing device will check the number of iterations that have been completed. If the number of iterations is less than 10, the next iteration will continue; if the number of iterations reaches 10, or the change in the integrated array is small enough (for example, less than a preset threshold of 0.01), the iteration process will be terminated.

[0061] In this process, the amount of change in the integrated array can be measured by calculating a certain distance or similarity between two adjacent generations of integrated arrays. For example, the Euclidean distance, cosine similarity, or any other suitable metric can be used to evaluate the change in the integrated array. However, since the integrated array usually contains high-dimensional vectors or complex graph structure data, more complex similarity measurement methods or specialized graph similarity evaluation algorithms may be required in actual calculations.

[0062] It should be noted that although the maximum number of iterations and the amount of change in the integrated array are given as examples of stopping conditions, in actual applications, the choice of stopping conditions should be determined based on specific application scenarios and performance requirements. For example, in some cases, performance indicators such as the accuracy or recall rate of anomaly detection may be more concerned, so these indicators can be used as part of the stopping condition.

[0063] Step S206: Determine the features involved in the abnormal information according to the integrated array output by the last generation.

[0064] After completing multiple iterations (steps S204 to S205), the data processing device obtains the last generation of integrated arrays. This array integrates the connection relationship between the initial feature embedding results (generated in step S202) and the abnormal priori tags (represented by the connection relationship array), and gradually deepens the understanding of the complex relationship between abnormal information through the information aggregation and update mechanism in the iterative process.

[0065] In step S206, the data processing device extracts the features involved in the abnormal information from the last generation of integrated arrays. These features not only contain the core attributes of each abnormal priori marker, but also incorporate their mutual influence and dependency with other markers. To achieve this goal, the data processing device may use a variety of data processing and analysis techniques, including but not limited to cluster analysis, principal component analysis (PCA), feature selection, or advanced feature extraction methods based on graph neural networks (GNN).

[0066] Taking the graph neural network (GNN) as an example, assume that during the iteration process, the data processing device uses a certain form of graph neural network to process the integrated array. GNN can use the node features and edge information in the graph structure data to update the node representation through the message passing mechanism, thereby capturing the complex relationship between the nodes in the graph. After multiple iterations, the node representation output by GNN not only contains the feature information of the node itself, but also implies the position of the node in the graph, the features of the neighboring nodes, and the interactions between them. In step S206, the data processing device can use the final node representation output by GNN as the abnormal information involving features. These features can be directly used in subsequent anomaly detection models (such as classifiers, anomaly detectors, etc.) to help the system identify potential abnormal events and the complex relationship network behind them.

[0067] As an implementation method, the rows in the feature embedding array are matched one by one with the abnormal priori marks, and the rows in the connection relationship array are matched one by one with the abnormal priori marks; in the feature embedding array and the connection relationship array, the abnormal priori marks corresponding to the rows of the feature embedding array and the rows of the connection relationship array with the same number are consistent. Based on this, step S203, merging the connection relationship array with the feature embedding array to obtain the first generation of integrated array, may include: Step S2031: performing array fusion on the connection relationship array and the feature embedding array to obtain a fused array; Step S2032: Add the fusion array and the feature embedding array to obtain a merged array; Step S2033: performing a standardization operation on the merged array to obtain a first generation integrated array.

[0068] In step S203, the data processing device effectively merges the connection relationship array with the feature embedding array to generate a first-generation integrated array. This step not only requires the integration of two different types of data structures, but also requires ensuring that the integrated array can accurately reflect the connection relationship between the abnormal prior markers and their feature information. To this end, step S203 is subdivided into three sub-steps, S2031, S2032, and S2033, which respectively perform array fusion, addition, and standardization operations.

[0069] In step S2031, it is assumed that the connection relationship array and the feature embedding array have been obtained through steps S201 and S202. The connection relationship array describes the connection between abnormal prior markers, such as the causal relationship between "cooling system failure" and "reactor temperature increase"; the feature embedding array contains the feature information of each abnormal prior marker, such as historical occurrence frequency, impact range, etc. In order to integrate these two types of information together, the array fusion method is not limited, such as matrix multiplication.

[0070] Specifically, for each feature vector (i.e., each row) in the feature embedding array, the value of the corresponding row in the connection relationship array (0 or 1) determines whether to combine it with other feature vectors in some form. In order to be closer to the actual matrix operation process, an extended operation can be imagined, that is, using the connection relationship array as a mask to filter or adjust the combination of feature vectors. However, since directly applying this filtering may not be intuitive mathematically, a simplified alternative will be adopted.

[0071] After completing the array fusion of step S2031 (although the "fusion" here is not matrix multiplication in the traditional sense), a fused intermediate array is obtained. This array reflects the influence of the connection relationship on the feature information to some extent. In step S2032, this intermediate array (although its specific form is not clearly given in the above description) is added to the original feature embedding array. The purpose of this step is to combine the fused feature information with the original feature information to retain the importance of the original features while introducing the influence of the connection relationship on the features. In step S2032, a merged array is obtained by the addition operation. However, since there may be differences in the numerical range of the fused array and the feature embedding array, direct addition may cause the merged array to lose numerical balance. In order to solve this problem, the merged array needs to be standardized to ensure that its numerical range is within a reasonable range, which is convenient for processing and analysis in subsequent steps.

[0072] Standardization operations usually include multiple methods, such as Min-Max standardization, Z-Score standardization, etc. Here, you can choose a method that is suitable for the characteristics of the nuclear power intelligent isolation lock system data for standardization. Since the merged array contains a comprehensive reflection of the feature information and the connection relationship information, it is necessary to ensure that the standardization process does not destroy the relative importance of these two types of information.

[0073] For example: Assume that the Min-Max normalization method is selected, which scales each element in the array to a specified minimum and maximum value (usually 0 and 1). For each element zij′ in the merged array Z′ (where i represents the row index and j represents the column index), the following formula can be used for normalization: zij′′=max(Zj′)−min(Zj′)zij′−min(Zj′); Among them, min(Zj′) and max(Zj′) represent the minimum and maximum values ​​of the jth column of the merged array respectively. The standardized array is recorded as Z′′∈R n×d .

[0074] The Min-Max normalization method may not be robust enough when dealing with data containing outliers. If there are extreme outliers in the merged array, this method can cause the normalization results of other data points to be compressed into a very small range. To avoid this, you can consider using a more robust normalization method, such as Z-Score normalization (although it is usually used for data that conforms to a normal distribution), or choose appropriate scaling factors and translation amounts based on the actual distribution of the data.

[0075] As an implementation method, the process of mining implicit representations of nuclear power isolation lock IoT monitoring data includes: Step S11: Determine a monitoring data implicit representation mining network; the monitoring data implicit representation mining network includes a plurality of sequentially connected skip connection components.

[0076] The data processing device needs to determine a suitable neural network model as the basis for mining the implicit representation of monitoring data based on the characteristics of the Internet of Things monitoring data of nuclear power isolation locks. In this scenario, it is particularly appropriate to select a network structure containing multiple sequentially connected jump connection components (such as ResNet, i.e., residual network). Jump connection solves the problem of gradient vanishing or gradient explosion in deep neural network training by introducing jump connection (also known as residual connection). This connection allows input information to be passed directly to deeper layers of the network, which helps to retain shallow features while promoting the learning of deep features. For example, suppose that the data processing device decides to use a ResNet network with multiple residual blocks as the implicit representation mining network for monitoring data. Each residual block consists of several convolutional layers, batch normalization layers, and nonlinear activation functions, and passes the input directly to the output through a jump connection.

[0077] Step S12: Based on the first jump connection component in the monitoring data implicit representation mining network, the nuclear power isolation lock IoT monitoring data is implicitly mined to obtain the monitoring data mining implicit representation.

[0078] In step S12, the data processing device inputs the nuclear power isolation lock IoT monitoring data into the first skip connection component (i.e., the first residual block) in the monitoring data implicit representation mining network. This component usually consists of several key parts: convolution layer, batch normalization layer, nonlinear activation function (such as ReLU), and a key skip connection.

[0079] The convolution layer is the core of feature extraction. It uses a set of learnable convolution kernels to perform local perception and parameter sharing operations on the input data. In the scenario of nuclear power isolation lock IoT monitoring data, these convolution kernels can capture local patterns in sensor readings, such as trends in temperature changes, frequencies of pressure fluctuations, etc. Through the convolution operation, the raw data is converted into a series of feature maps, each of which represents an abstract representation of the input data at different scales or angles. Subsequently, the output of the convolution layer is processed by a batch normalization layer. The purpose of batch normalization is to accelerate the training process of the neural network and reduce the impact of internal covariate shift. In this step, the data of each channel of each feature map is standardized to have zero mean and unit variance, thereby improving the stability and convergence speed of the model. Next, the batch normalized data is activated by a nonlinear activation function such as ReLU. The ReLU function introduces nonlinear characteristics by setting negative values ​​to zero and retaining only positive values. This step is crucial for building complex models because it allows the network to learn nonlinear relationships between input and output. Finally, the first skip connection component adds the input directly to the output processed by the convolution layer and the activation function through a skip connection. This design helps alleviate the gradient vanishing problem in deep network training and allows information to flow more smoothly in the network. After being processed by the first skip connection component, the data processing device obtains an implicit representation for monitoring data mining. This representation is a high-dimensional vector or tensor (depending on the specific design of the network) that captures the key features and information in the original IoT monitoring data. These features may include time series patterns of sensor readings, correlations between different sensors, and changing trends in device status.

[0080] Step S13: determining the current skip connection component corresponding to the current generation in the mining network implicitly representing the monitoring data in the current generation starting from the second generation after the first generation.

[0081] Step S13 involves determining the corresponding network components for processing according to the current number of iterations (or "generations") in the multi-layer neural network structure. Specifically, in step S13, the data processing device needs to determine the current jump connection components (such as residual blocks in ResNet) in the mining network corresponding to the monitoring data implicit representation corresponding to each generation starting from the second generation after completing the processing of the first generation (i.e., the initial layer or the first layer).

[0082] When building a monitoring data implicit representation mining network, multiple layers or stages are usually designed, each of which contains one or more skip connection components (such as residual blocks in ResNet). These components are connected in sequence to form a deep neural network structure. When the network system receives the nuclear power isolation lock IoT monitoring data, it starts processing from the first component (i.e., the first generation) and gradually goes deeper into the network.

[0083] In step S13, the task of the data processing device is to locate and activate the corresponding skip connection components according to the current iteration number (or "generation"). This process is automated and relies on the predefined and iterative control logic of the network structure.

[0084] Next, in step S13, when the iteration enters the second generation, the data processing device selects and activates the second residual block as the current skip connection component. This process does not require manual intervention, but is automatically completed by the network structure and iteration control logic. Similarly, when the iteration enters the third generation, the data processing device selects and activates the third residual block as the current component.

[0085] In this process, each residual block receives the output from the previous residual block (which may be one or more feature vectors, matrices, or higher-dimensional tensors) and performs further feature extraction and transformation on it. For example, the output of the first residual block may be a matrix containing multiple feature vectors, each of which represents the feature representation of the input data at different dimensions or perspectives. When this matrix is ​​passed to the second residual block (i.e., the current skip connection component in the second generation), it further processes these feature vectors, extracts deeper features, and may retain some of the original information through skip connections.

[0086] Step S14: Based on the current skip connection component, the implicit representation mining of the monitoring data output by the previous skip connection component is performed to obtain the implicit representation of the transition monitoring data.

[0087] In step S14, the data processing device first obtains the monitoring data mining implicit representation output by the previous jump connection component (i.e., the previous component of the current component). This representation is usually a high-dimensional vector, matrix or tensor that captures the key features and information in the original nuclear power isolation lock IoT monitoring data. Then, the data processing device passes this implicit representation as input to the current jump connection component. The current component contains a series of processing units, such as convolutional layers, batch normalization layers, nonlinear activation functions, etc., which work together on the input data to extract deeper features. In this processing process, the convolution layer performs local perception and parameter sharing operations on the input data to generate a new feature map; the batch normalization layer standardizes the output of the convolution layer to accelerate the training process and improve the generalization ability of the model; nonlinear activation functions (such as ReLU) introduce nonlinear characteristics, enabling the model to learn complex data representations.

[0088] Finally, after being processed by the current skip connection component, the data processing device will obtain a new implicit representation, namely the implicit representation of transition monitoring data. Compared with the output of the previous component, this representation contains deeper features and information and is closer to the essential characteristics of the data.

[0089] In step S14, the data processing device obtains the implicit representation of monitoring data mining output by residual block A (i.e., the previous jump connection component). This representation may be a matrix containing multiple feature vectors, each of which represents the characteristics of the original monitoring data in different dimensions or perspectives. Then, the data processing device passes this matrix as input to residual block B. Inside residual block B, the convolution layer performs a convolution operation on the input matrix to generate a new feature map; the batch normalization layer standardizes the feature map; and the ReLU activation function performs nonlinear activation on the standardized feature map. After this series of processing, residual block B outputs a new implicit representation, namely, the implicit representation of transition monitoring data. This representation may be a new feature matrix, in which the feature vectors not only contain the feature information output by residual block A, but also incorporate deeper features extracted by residual block B through operations such as convolution, normalization, and activation.

[0090] Step S15: Add the monitoring data mining implicit representation output by the previous skip connection component to the transition monitoring data implicit representation to obtain the monitoring data mining implicit representation output by the current skip connection component.

[0091] In step S15, the data processing device obtains two implicit representations: one is the implicit representation of monitoring data mining output by the previous jump connection component (i.e., the previous component of the current component), and the other is the implicit representation of transition monitoring data after being processed by the current jump connection component. These two representations represent the feature abstraction of data at different network levels. Subsequently, the data processing device performs an addition operation to perform element-level addition operations on the two implicit representations. This addition operation is not a simple numerical addition, but an addition between corresponding elements, requiring the two implicit representations to have the same shape (i.e., the same dimension and size). If their shapes are not exactly the same, appropriate adjustments (such as cropping, padding, or interpolation) may be required to match the shapes. However, in the design of the residual network, due to the existence of the jump connection, it is usually guaranteed that the two implicit representations have the same shape. The jump connection directly passes the output of the previous component as input to the next component, and processes it inside the component (such as convolution, normalization, and activation) to finally generate a transition implicit representation. Due to the existence of the jump connection, the transition implicit representation can be directly added to the output of the previous component. The result of the addition operation is the implicit representation of the monitoring data mining output of the current skip connection component. This new implicit representation not only contains the new features extracted by the current component through nonlinear transformation, but also retains the useful information in the output of the previous component, avoiding information loss.

[0092] In step S15, the data processing device adds the output of the first residual block to the transition representation generated by the second residual block at the element level. Since the two have the same shape, the addition operation can be performed directly, and the result is still an implicit representation with a shape of (batch_size, num_features, height, width). This new implicit representation, as the output of the second residual block, not only contains the new features processed by the second residual block, but also retains the useful information in the output of the first residual block. This design enables the deep network to learn the feature representation of the data more effectively, improving the performance and generalization ability of the model.

[0093] Step S16: Take the next generation as the current generation, jump to the step of determining the current jump connection component corresponding to the current generation in the monitoring data implicit representation mining network, and iterate until the second stop judgment requirement is met.

[0094] In step S16, the data processing device first sets the "next generation" as the new "current generation". The "next generation" here refers to the level where the next component immediately after the currently processed component is located in the network structure. For example, if the first residual block (i.e., the first generation) has been processed, then the "next generation" is the level where the second residual block is located (i.e., the second generation).

[0095] After setting the new "current generation", the data processing device will jump to the step of determining the current skip connection component (similar to step S13), and find the corresponding skip connection component according to the new "current generation". This component will be used as the target of a new round of processing to continue the implicit representation mining process.

[0096] Next, the operations from step S12 to S15 are repeated: the input data is processed based on the current skip connection component (if the current generation is not the first generation, the input data is the output of the previous component), an implicit representation of the transition monitoring data is generated, and it is added to the output of the previous component (through the skip connection) to obtain the output of the current component.

[0097] This process is repeated, and each iteration advances the processing to the next level in the network until the second preset stopping judgment requirement is met. This stopping condition can be based on a limit on the number of iterations (for example, all preset levels are processed) or on the evaluation of some performance indicator (for example, a certain metric implicitly represented by monitoring data mining reaches a stable or optimal state).

[0098] After completing the processing of the first generation (i.e., steps S12 to S15), enter step S16, set the second generation as the new current generation, and jump to the step of determining the current skip connection component. At this time, the second residual block becomes the new current skip connection component, and the implicit representation mining process continues, using the output of the second residual block as the implicit representation of the transition monitoring data, and adding it to the output of the first residual block (through the skip connection) to obtain the final result of the second generation processing.

[0099] This process will continue until the third generation of processing is completed. In each iteration, it will be checked whether the second stop judgment requirement is met. For example, if the preset processing level is only three, then when the third generation of processing is completed, the data processing device will naturally stop iterating because all preset levels have been processed. In addition, if the system sets a stop condition based on performance indicators (such as the change in the monitored data features is less than a certain threshold), the indicator will be evaluated after each iteration, and the iteration will end early if the condition is met.

[0100] Step S17: mining implicit representations based on the monitoring data output by the last generation to determine implicit representations of monitoring data of the nuclear power isolation lock IoT monitoring data.

[0101] After completing the iterative processing of all preset levels (step S16), the data processing device obtains the monitoring data mining implicit representation output by the last generation of jump connection components (such as the last residual block in ResNet). This representation is a high-level feature vector or matrix that is gradually extracted and transformed through a multi-layer network structure, and it contains the most critical and abstract information in the original monitoring data.

[0102] Assume that the implicit representation mining network of the monitoring data of the Internet of Things of the nuclear power isolation lock consists of three residual blocks, each of which is responsible for extracting features at different levels in the input data. After completing the processing of all three residual blocks (i.e., iterating to the third generation and ending), the data processing device obtains the implicit representation of the monitoring data mining output by the third residual block. This representation may be a high-dimensional feature vector or matrix, and its dimension depends on the specific parameters of the network design and the characteristics of the input data. For example, if the input data is multi-channel time series data (such as readings from different sensors), then the final feature representation may be a multidimensional array that contains the feature information of the time series data at different time scales and sensor channels.

[0103] The data processing device directly regards this multidimensional array as an implicit representation of the monitoring data of the nuclear power isolation lock IoT monitoring data. This representation not only captures the basic statistical characteristics of the original data (such as mean, variance, etc.), but also contains deeper structural information and dynamic characteristics (such as trend, periodicity, correlation, etc.). These features are crucial for subsequent anomaly detection and analysis.

[0104] For example, in an anomaly detection task, the data processing device can use this implicit representation as input to train a classifier or anomaly detection model. When new monitoring data arrives, the data processing device first performs the same implicit representation mining process on it to obtain a new implicit representation. The data processing device then compares this new representation with the trained model to determine whether there is an abnormal behavior or pattern.

[0105] As an implementation method, the implicit representation mining process of the abnormal diagnosis report includes: Step S10A: For each diagnostic event in the abnormal diagnosis report, embed the component tags of the current diagnostic event to obtain an event embedding representation corresponding to the current diagnostic event.

[0106] In step S10A, the data processing device first identifies each diagnostic event in the abnormal diagnosis report. Each diagnostic event is usually described by one or more component tags, which concisely summarize the essential characteristics of the event. For example, a diagnostic event may be labeled as "cooling system failure", where "cooling system" and "failure" are the component tags of the event. Next, the data processing device uses a pre-trained word embedding model (such as Word2Vec, GloVe, FastText, or a Transformer-based model such as the embedding layer of BERT) to process these labels. These models learn to map words or phrases to points in a high-dimensional vector space through unsupervised learning on a large-scale text corpus, so that semantically similar words are also located similarly in the vector space.

[0107] For the specific implementation in step S10A, assume that the BERT model is used to obtain the embedding representation of the event label. The BERT model first converts the text input into a series of tokens, and then embeds each token to generate the corresponding embedding vector. However, here, since the entire event label is processed instead of a single word, one of the following strategies can be adopted: Whole-label embedding: The event label is considered as a whole and directly used as the input of the BERT model (special start and end tags may need to be added). The model outputs the embedding vector of the [CLS] tag corresponding to the label as the event embedding representation. This method is suitable for situations where the label length is moderate and the model supports long text input.

[0108] Label segmentation embedding: If the event label is too long or the model does not support directly processing the entire label, the label can be segmented into shorter segments and each segment can be embedded separately. Then, the embedding vectors of these segments can be merged by averaging, concatenation, or more complex aggregation methods (such as attention mechanisms) to obtain the overall embedding representation of the event.

[0109] Aggregation after embedding of label components: For event labels composed of multiple words or phrases, each component element can be embedded first, and then weighted summation or concatenation can be performed according to the semantic importance of the element (such as determined by weights or attention mechanisms) to finally obtain the event embedding representation.

[0110] Step S10B: for each event data item in the current diagnosis event, determine the distribution order of the current event data item in the current diagnosis event, and embed the distribution order to obtain a distribution embedding representation corresponding to the current event data item.

[0111] In step S10B, the data processing device traverses all event data items in the current diagnostic event. Each event data item may contain different types of information, such as timestamp, numerical measurement results, descriptive text, etc., which together constitute a detailed description of the diagnostic event.

[0112] For each event data item, the data processing device performs the following operations: Determine the distribution order: The system needs to determine the position or order of the data item in the event description. This is usually determined by the natural order of appearance of the data items in the event text, but it may also be identified based on specific tags or structures (such as lists, tables, etc.). For example, in the "cooling system failure" event, if the data items are listed in the order of "failure occurrence time", "failure impact range", "measures taken", etc., the data processing device can identify the specific location of each data item.

[0113] Embed the distribution order: Once the distribution order of the data items is determined, the data processing device needs to convert this order information into a numerical representation. This is usually achieved through an embedding layer, which maps the sequential index (such as the first data item is 0, the second is 1, and so on) into a vector space of fixed dimension. This vector, the distribution embedding representation, captures the relative position information of the data items in the event description. It is worth noting that the embedding layer here is not necessarily pre-trained through machine learning; instead, it may be a simple lookup table where each index corresponds to a predefined vector. However, in some advanced applications, more complex methods based on attention mechanisms or position encodings (such as position embedding in the Transformer model) may also be used to generate distribution embedding representations.

[0114] In this way, step S10B not only considers the content information of the event data items (although this part of the content is processed in step S10C), but also additionally captures their structural position information in the event description. This dual representation helps the data processing device to more accurately understand the complex relationships and data patterns in the abnormal diagnosis report.

[0115] Step S10C: embed the current event data item to obtain a data item embedding representation corresponding to the current event data item, and obtain a feature embedding result corresponding to the current event data item based on the event embedding representation corresponding to the current diagnostic event, the distribution embedding representation corresponding to the current event data item, and the data item embedding representation.

[0116] In step S10C, the data processing device first performs data item embedding processing on each current event data item. Data item embedding is a technology that converts text, numerical or other types of data into fixed-dimensional vector representations, which helps to maintain the original information of the data and facilitate calculation and comparison in subsequent processing.

[0117] For text-type data items (such as descriptive text, error messages, etc.), the data processing device can use word embedding models in natural language processing (NLP) technology (such as Word2Vec, GloVe, BERT, etc.) to obtain vector representations of each word or phrase. These vectors can then be aggregated into an embedded representation of the entire data item by averaging, weighted summation, concatenation, or using more complex sequence models (such as LSTM, GRU, or Transformer).

[0118] For numerical data items (such as temperature, pressure and other measurements), the data processing device can simply standardize or normalize them and use them directly as embedding representations (in some cases, they can also be converted to one-hot encoding or binary encoding, but this method may not be practical when the feature dimension is high). After obtaining the data item embedding representation, the data processing device combines the event embedding representation and distribution embedding representation generated in the previous step to generate a feature embedding result. This process usually involves a fusion step in which different embedding representations are combined in some way (such as concatenation, weighted summation, bilinear transformation, etc.).

[0119] Data item embedding representation: If the data item is in text form, the data processing device can use the BERT model to obtain the embedding vector of each word and aggregate these vectors through the Transformer encoder to obtain the embedding representation of the entire data item. For example, assume that the embedding representation of "cooling tower area" is a 256-dimensional vector v_data.

[0120] Event embedding representation: In step S10A, the data processing device has generated an event embedding representation for the event "cooling system failure", which is assumed to be a 768-dimensional vector v_event.

[0121] Distributed embedding representation: In step S10B, if “fault impact range: cooling tower area” is the second data item in the event, then its distributed embedding representation may be a simple position encoding vector, such as a 5-dimensional vector v_position, which is used to represent the order of the data items in the event.

[0122] To generate the feature embedding result, the data processing device can concatenate these three vectors together to form a new vector. For example: Feature embedding result = [v_event; v_position; v_data]; Where [;] represents a vector concatenation operation. If v_event is 768-dimensional, v_position is 5-dimensional, and v_data is 256-dimensional, then the feature embedding result will be a 1030-dimensional vector.

[0123] However, in practical applications, directly concatenating high-dimensional vectors can lead to dimensionality curse and computational inefficiency. Therefore, data processing devices can adopt more complex fusion strategies such as attention mechanism, bilinear transformation or use more compact embedding space to reduce the dimension.

[0124] Step S10D: Determine the diagnostic report implicit representation of the abnormal diagnosis report according to the feature embedding results corresponding to each event data item in the abnormal diagnosis report.

[0125] In step S10D, the data processing device reviews the feature embedding results generated for each event data item in the previous steps (S10A to S10C). These feature embedding results not only contain the semantic information of the data item itself (obtained through data item embedding), but also incorporate the contextual information of the data item in the event (realized through event embedding and distribution embedding).

[0126] Next, the data processing device designs a strategy to integrate these feature embedding results to generate an overall implicit representation of the diagnosis report. This strategy can be flexibly selected according to the specific application scenario and data characteristics. Common methods include but are not limited to: average / weighted average, sequence model, attention mechanism, and graph neural network (GNN).

[0127] Assume that in the diagnostic report of "cooling system failure", there are multiple event data items, such as "fault occurrence time", "fault impact scope", "measures taken", etc., and each data item is processed through steps S10A to S10C to obtain the corresponding feature embedding results.

[0128] If a sequence model (such as LSTM) is used as the integration strategy, the data processing device inputs all feature embedding results into the LSTM network in the natural order of the event data items in the report. The LSTM network processes each feature embedding result step by step through its internal state transfer mechanism, and after processing all inputs, it outputs a fixed-length hidden state vector as the overall implicit representation of the diagnostic report. This implicit representation vector captures the essence of all key information in the diagnostic report, including the type of event, the specific content of the data items, and the order and relationship between them. In subsequent tasks (such as anomaly classification, trend prediction, etc.), the data processing device can directly use this implicit representation vector as input without having to process the original text or numerical data.

[0129] As an implementation manner, step S20, merging the monitoring data implicit representation and the diagnosis report implicit representation to obtain a first fused implicit representation, may specifically include: Step S21: Acquire a contribution parameter set corresponding to the implicit representation of the monitoring data and a contribution parameter set corresponding to the implicit representation of the diagnostic report; each contribution parameter set is composed of a reduced-dimensional contribution parameter array obtained by performing singular value decomposition on a corresponding preset contribution parameter array; Step S22: merging each dimension-reduced contribution parameter array in the contribution parameter set corresponding to the implicit representation of the monitoring data with the implicit representation of the monitoring data to obtain a plurality of weighted implicit representations of the monitoring data; Step S23: merging each dimension-reduced contribution parameter array in the contribution parameter set corresponding to the implicit representation of the diagnosis report with the implicit representation of the diagnosis report to obtain a plurality of weighted implicit representations of the diagnosis report; Step S24: Merge multiple monitoring data weighted implicit representations and multiple diagnosis report weighted implicit representations to obtain a first fused implicit representation.

[0130] In step S21, the data processing device obtains the corresponding contribution parameter sets for the implicit representation of monitoring data and the implicit representation of diagnostic report respectively. These contribution parameter sets are composed of reduced dimension parameter arrays obtained by performing singular value decomposition (SVD) on the preset contribution parameter array. SVD is a powerful matrix decomposition technology that can decompose the original matrix into three specific matrix products, which contain important feature information of the original data.

[0131] Assume that the monitoring data is implicitly represented as an m×n matrix V, and the diagnostic report is implicitly represented as a p×q matrix R. For the monitoring data, the data processing device first defines a preset contribution parameter array A that is similar in shape to V (but may contain more redundant information). Then, A is decomposed by SVD to obtain three matrices: UA, ΣA, and VAT (where UA and VA are orthogonal matrices, and ΣA is a diagonal matrix whose diagonal elements are singular values). By selecting the first k largest singular values ​​and their corresponding left and right singular vectors, a reduced-dimensional contribution parameter array can be constructed. Similarly, the preset contribution parameter array of the diagnostic report is processed in the same way.

[0132] For example, assume that the size of the implicit representation matrix V of the monitoring data is 100×50 (100 time points, 50 features at each time point), and the size of the preset contribution parameter array A is 100×100 (for simplicity, it is assumed that the number of features is the same as the number of time points, but it may be different in actual applications). After SVD decomposition, select the top 10 largest singular values ​​and their corresponding left and right singular vectors to construct a reduced-dimensional contribution parameter array. In this way, each contribution parameter array is a 100×10 matrix, which contains the main contribution information of the implicit representation of the monitoring data.

[0133] In step S22, the data processing device traverses each dimension-reduced contribution parameter array in the contribution parameter set, and merges (or fuses) them with the implicit representation of the monitoring data. The "merge" operation here may involve matrix multiplication, weighted summation, or other forms of linear combination, depending on the specific forms of the contribution parameter array and the implicit representation of the monitoring data.

[0134] For each dimension-reduced contribution parameter array Wi (assuming its size is m×k) in the contribution parameter set, the data processing device merges it with the monitoring data implicit representation matrix V (size is m×n). A simple method is to multiply Wi with the transposed VT of V (note that the dimension may need to be adjusted to match the operation rules here), and then perform some form of normalization or rearrangement on the result to obtain the weighted implicit representation. Use Wi as a weight matrix and perform element-by-element weighted summation or more complex nonlinear transformation with V.

[0135] Continuing with the previous example, assume that the contribution parameter array W1 is a 100×10 matrix, which represents the weights of the features corresponding to 100 time points in the implicit representation of the monitoring data. The data processing device combines W1 with V in some way (such as element-wise multiplication followed by summation) to generate a weighted implicit representation matrix V1′. This process is repeated until all the reduced-dimensional contribution parameter arrays in the contribution parameter set are used, thereby generating multiple weighted implicit representations of the monitoring data V1′, V2′, …, Vt′ (where t is the number of contribution parameter arrays).

[0136] Step S23 is similar to step S22, but the operation object is the implicit representation of the diagnostic report. The data processing device traverses each reduced-dimensional contribution parameter array in the contribution parameter set corresponding to the implicit representation of the diagnostic report, and merges them with the implicit representation of the diagnostic report respectively to generate multiple weighted implicit representations of the diagnostic report. Similar to the processing of monitoring data, for the diagnostic report implicit representation matrix R (size is p×q) and each reduced-dimensional contribution parameter array Xj (assuming its size is p×l), the data processing device combines them by some transformation. This transformation may involve matrix multiplication, weighted summation or other merging strategies applicable to text or sequence data. Since the implicit representation of the diagnostic report usually contains more complex text or symbolic information, the merging process may require the use of natural language processing (NLP) technology or a specialized sequence processing model. However, in order to simplify the description and maintain consistency with the processing of monitoring data, it can be assumed that there is a similar weighted summation process to generate the weighted implicit representation of the diagnostic report. That is, for each contribution parameter array Xj, the data processing device generates a corresponding weighted implicit representation matrix Rj′. Assuming that the size of the diagnostic report implicit representation matrix R is 50×100 (50 diagnostic items, 100 features or word embedding dimensions for each item), the preset contribution parameter array corresponding to the diagnostic report is decomposed by SVD to obtain multiple reduced-dimensional contribution parameter arrays. For one of the arrays X1 (size 50×5), the data processing device generates a weighted implicit representation matrix R1′ by weighted summation or other merging strategies. This process is repeated until all contribution parameter arrays are processed, generating multiple weighted implicit representations of diagnostic reports R1′, R2′,…, Rs′ (where s is the number of contribution parameter arrays).

[0137] After obtaining multiple weighted implicit representations of monitoring data and diagnostic reports respectively, the data processing device needs to further integrate them to generate a comprehensive first fused implicit representation. This representation will contain both real-time measurement information of monitoring data and expert analysis information of diagnostic reports, providing comprehensive data support for subsequent anomaly detection, classification or prediction tasks. In step S24, the data processing device adopts a certain merging strategy to integrate multiple weighted implicit representations of monitoring data and multiple weighted implicit representations of diagnostic reports. This merging strategy may involve simple concatenation, weighted averaging, weighted summation based on attention mechanism or other more complex graph fusion methods. The merged result is the first fused implicit representation, which will become the input data for subsequent steps.

[0138] The specific implementation of the merging operation depends on the specific form of the weighted implicit representation and the requirements of the subsequent tasks. If the weighted implicit representations are all numerical matrices and the dimensions are compatible, then simple concatenation or weighted averaging may be sufficient. However, in practical applications, due to the essential differences between monitoring data and diagnostic reports (one is a real-time measurement value matrix and the other is an expert analysis of text or symbolic representation), direct merging may not be the best choice. Therefore, the data processing device may need to design a more complex fusion network or model to process these two different types of data. One possible solution is to use a neural network model based on the attention mechanism. The model can automatically learn the relative importance between different weighted implicit representations and perform weighted summation on them according to this importance. Specifically, the model can assign an attention weight to each weighted implicit representation and then combine them into a single fusion representation based on these weights. Assume that after the previous steps, the data processing device obtains t weighted implicit representations of monitoring data V1′, V2′, …, Vt′ and s weighted implicit representations of diagnostic reports R1′, R2′, …, Rs′. In order to generate the first fused implicit representation, the data processing device adopts a neural network model based on an attention mechanism.

[0139] As an implementation method, the contribution parameter set corresponding to the implicit representation of the monitoring data is obtained by using the following steps: Step S211: obtaining a preset contribution parameter array corresponding to the implicit representation of the monitoring data; Step S212: determining the linear independence of the preset contribution parameter array corresponding to the implicit representation of the monitoring data; the linear independence of the preset contribution parameter array corresponding to the implicit representation of the monitoring data represents the dimension of the preset contribution parameter array corresponding to the implicit representation of the monitoring data; Step S213: split the preset contribution parameter array corresponding to the implicit representation of the monitoring data into multiple reduced-dimensionality contribution parameter arrays with linear independence equal to 1; the number of reduced-dimensionality contribution parameter arrays obtained by splitting is consistent with the linear independence of the preset contribution parameter array corresponding to the implicit representation of the monitoring data.

[0140] In step S211, the data processing device obtains a preset contribution parameter array corresponding to the implicit representation of the monitoring data from a predefined or trained model, rule base or expert system. This array can be a multidimensional tensor whose shape and structure match the implicit representation of the monitoring data for subsequent weighting or fusion operations.

[0141] Assume that the implicit representation of the monitoring data is a two-dimensional array of shape (time_steps, features), where time_steps represents the time step (i.e., the number of sampling times of the monitoring data), and features represents the number of features at each time step. In order to evaluate the importance of these features, the data processing device may obtain a preset contribution parameter array of the same shape from a pre-trained feature importance evaluation model. Each element of this array represents the importance weight of the feature at the corresponding position. In practical applications, the shape of the preset contribution parameter array may not completely match the implicit representation of the monitoring data. For example, it may be a larger tensor that contains additional information (such as time correlation, spatial dependence, etc.) in addition to the direct feature weights. In this case, the data processing device needs to use some method (such as slicing, reshaping, dimensionality reduction, etc.) to extract the part directly related to the implicit representation of the monitoring data. However, for the sake of simplicity, it is assumed that the preset contribution parameter array is a two-dimensional array of shape (time_steps, features), whose elements directly represent the importance weights of the corresponding monitoring data features.

[0142] In step S212, the data processing device calculates the linear independence of the preset contribution parameter array by mathematical methods. This calculation process may involve linear algebra operations such as singular value decomposition (SVD) of the matrix, elementary transformation of rows (columns), etc. Linear independence not only helps to understand the structural characteristics of the array, but also provides a basis for subsequent dimensionality reduction processing. In practical applications, directly calculating the linear independence of a large tensor may be computationally intensive and not always necessary. Therefore, in some cases, the data processing device may use heuristic methods or rules based on domain knowledge to estimate linear independence, or simply set a fixed, experience-based linear independence value. Assume that the data processing device accurately calculates the linear independence of the preset contribution parameter array by SVD decomposition. SVD decomposition decomposes the matrix into the product of three matrices: an orthogonal matrix, a diagonal matrix (elements on the diagonal are called singular values), and the transpose of another orthogonal matrix. The number of non-zero singular values ​​in the diagonal matrix is ​​equal to the linear independence of the original matrix. For example, continuing the previous example, assume that the preset contribution parameter array is a two-dimensional array of shape (time_steps, features). The data processing device performs SVD decomposition on the array to obtain a singular value matrix. By checking the number of non-zero elements on the diagonal of the singular value matrix, the data processing device can determine the linear independence (rank) of the preset contribution parameter array. For example, if there are k non-zero singular values ​​in the singular value matrix (k <= min(time_steps, features)), the linear independence of the preset contribution parameter array is k.

[0143] After determining the linear independence of the preset contribution parameter array, the data processing device needs to split the array into multiple smaller, reduced-dimensional contribution parameter arrays with linear independence equal to 1. The advantage of doing this is that the original high-dimensional contribution parameter space can be decomposed into a series of low-dimensional subspaces, each of which contains a portion of independent information of the original data. This splitting helps the subsequent weighted fusion operation to be more refined and flexible. In step S213, the data processing device splits the preset contribution parameter array according to the result of linear independence. Since the linear independence equal to 1 means that each row (or column) of the matrix is ​​linearly independent, the splitting process is essentially looking for all linearly independent row (or column) combinations in the original array. However, in most cases, it may be unrealistic to directly find all such combinations because their number may be very large and many combinations have no practical significance in practical applications.

[0144] Therefore, the data processing device usually adopts a more practical method: extract the key information (i.e., rows or columns with larger singular values) from the original array through SVD decomposition or other dimensionality reduction techniques, and use them as the dimensionality reduction contribution parameter array. Since the SVD decomposition has sorted the rows (columns) according to the size of the singular values, the data processing device can simply select the rows (columns) corresponding to the first k singular values ​​to form the dimensionality reduction contribution parameter array (where k is the linear independence).

[0145] However, it should be noted that the "rows" or "columns" mentioned here are not the rows or columns of the original array in the literal sense, but the rows (or column vectors) in the orthogonal matrix obtained after SVD decomposition. These vectors are the projections of the original data in the subspace corresponding to the singular values, and they retain the main features and information of the original data. In order to simplify the description and maintain consistency with the shape of the original array, it can be assumed that the data processing device generates a dimensionality reduction contribution parameter array with a shape close to the original array but a lower dimension in some way (such as selecting the row or column vector with the largest singular value and appropriately expanding or copying it). Alternatively, the data processing device does not directly generate a dimensionality reduction contribution parameter array with a shape exactly the same as the original array, but generates a series of array sets with smaller shapes and lower dimensions. Each array represents the projection or representation of the original data in a specific subspace.

[0146] As an implementation, the target anomaly prior label is output based on a data analysis neural network, which is calibrated by the following operations to obtain: Step S110: Acquire a calibration sample database; the calibration sample database includes a plurality of nuclear power isolation lock IoT monitoring training data, and a plurality of nuclear power isolation lock IoT monitoring training data respectively corresponding to the abnormal diagnosis report samples and abnormal supervision information; Step S120: for each nuclear power isolation lock IoT monitoring training data in a plurality of nuclear power isolation lock IoT monitoring training data, perform implicit representation mining on the corresponding nuclear power isolation lock IoT monitoring training data and the corresponding abnormal diagnosis report sample to obtain the implicit representation of the inference monitoring data and the implicit representation of the inference diagnosis report; Step S130: merging the implicit representation of the inference monitoring data and the implicit representation of the inference diagnosis report to obtain a first fused inference implicit representation; Step S140: Acquire features involved in the abnormal information sample, wherein the features involved in the abnormal information sample include implicit representations of sample labels corresponding to a plurality of abnormal a priori labeled samples in the abnormal information-involved topological structure sample; for each abnormal a priori labeled sample, the implicit representation of the sample label of the abnormal a priori labeled sample includes not only the abnormal information of the abnormal a priori labeled sample, but also the abnormal information of the abnormal a priori labeled samples connected to the corresponding abnormal a priori labeled sample in the abnormal information-involved topological structure sample; Step S150: merging the first fused reasoning implicit representation with the implicit representation of each sample label in the feature involved in the abnormal information sample, to obtain second fused reasoning implicit representations corresponding to multiple abnormal a priori label samples; Step S160: Determine the inferred abnormal prior labeled samples corresponding to the nuclear power isolation lock IoT monitoring training data among the multiple abnormal prior labeled samples based on the second fused reasoning implicit representations respectively corresponding to the multiple abnormal prior labeled samples, and optimize the network parameters of the data analysis neural network based on the error between the inferred abnormal prior labeled samples and the abnormal supervision information corresponding to the corresponding nuclear power isolation lock IoT monitoring training data.

[0147] Before calibrating the data analysis neural network, we first build a calibration sample database containing rich samples. This database should cover various monitoring data and corresponding abnormal diagnosis reports of the nuclear power isolation lock system under normal and abnormal conditions, as well as abnormal supervision information annotated by experts (i.e., labels for training).

[0148] The data processing device collects historical monitoring data, abnormal diagnosis reports and expert annotations of the nuclear power isolation lock system to build a calibration sample database. These sample data are divided into training sets, validation sets and test sets, which are used for network training, parameter adjustment and performance evaluation respectively.

[0149] Assume that the calibration sample database contains 1,000 samples, each of which consists of three parts: nuclear power isolation lock IoT monitoring training data (such as real-time readings of sensors such as temperature, pressure, and radiation level), corresponding abnormal diagnosis report samples (including detailed analysis and evaluation of system status), and abnormal supervision information (such as labels such as "cooling system failure" and "electrical system failure"). These samples cover a variety of situations such as normal system operation, single failure, and multiple failures.

[0150] During the calibration process, implicit representation mining is performed on each nuclear power isolation lock IoT monitoring training data and its corresponding abnormal diagnosis report sample to extract deep features in the data. These features will serve as input to the data analysis neural network.

[0151] The data processing device processes the monitoring data and the diagnostic report using a previously defined implicit representation mining network (such as a deep learning model including a skip connection component). This step is similar to step S10 mentioned above, but here it is for the data in the calibration sample database.

[0152] For a sample in the calibration sample database, the data processing device first uses the implicit representation mining network to process the monitoring data to obtain the implicit representation of the inferred monitoring data. Then, the corresponding abnormal diagnosis report sample is processed in the same way to obtain the implicit representation of the inferred diagnosis report. These implicit representations capture the key features and potential patterns in the data and reports.

[0153] In step S130, after obtaining the inference monitoring data implicit representation and the inference diagnosis report implicit representation, they need to be merged into a comprehensive first fusion inference implicit representation. This step helps to integrate information from different data sources and improve the accuracy of data analysis.

[0154] The data processing device adopts a method similar to the above step S20 to merge the implicit representation of the inference monitoring data and the implicit representation of the inference diagnosis report, which usually involves operations such as determination of contribution parameters and weighted fusion.

[0155] In step S140, in order to identify abnormal events in the monitoring data, the data processing device needs to obtain the features involved in the abnormal information samples. These features describe the complex relationship between the abnormal information and help improve the accuracy of abnormality detection. For example, the data processing device generates a sample label implicit representation for each abnormal prior label sample based on the abnormal information involving the topological structure sample. These representations not only contain the feature information of the abnormal prior label itself, but also incorporate the information of other abnormal prior labels associated with it.

[0156] In step S150, after obtaining the first fused reasoning implicit representation and the features involved in the abnormal information sample, they need to be further merged to generate multiple second fused reasoning implicit representations. These representations will be used for subsequent abnormal classification or recognition tasks.

[0157] For example, the data processing device merges the first fused reasoning implicit representation with the sample label implicit representation of each abnormal prior label sample (such as weighted summation, splicing, etc.) to obtain multiple second fused reasoning implicit representations.

[0158] In step S160, after obtaining multiple second fusion reasoning implicit representations, the data processing device needs to optimize the parameters of the data analysis neural network according to these representations and abnormal supervision information. This step adjusts the network weights by minimizing the prediction error to improve the generalization ability and accuracy of the network.

[0159] For example, the data processing device first determines the inference anomaly priori labeled sample (i.e., the anomaly type predicted by the network) based on the second fusion inference implicit representation. Then, the difference between the inference anomaly priori labeled sample and the anomaly supervision information is compared to calculate the value of the loss function (such as the cross entropy loss). Finally, the back propagation algorithm and optimizer (such as SGD, Adam, etc.) are used to update the network parameters to reduce the value of the loss function.

[0160] Through the above steps, the data processing device successfully adjusted the data analysis neural network, enabling it to accurately identify abnormal events in the nuclear power isolation lock IoT monitoring data. This process not only improves the performance of the network, but also provides strong support for the safe operation and maintenance of the nuclear power intelligent isolation lock system.

[0161] As an implementation method, the data analysis neural network includes a monitoring data implicit representation mining network. Based on this, before performing implicit representation mining on the nuclear power isolation lock IoT monitoring training data in the calibration sample database, the method may also include a calibration process of the monitoring data implicit representation mining network, which may specifically include: Step S1: Determine the database capacity of the calibration sample database and the number of labels of multiple abnormal prior labeling samples; Step S2: determining the adaptive adjustment complexity of the pre-trained network according to the number of labels of multiple abnormal prior labeled samples; Step S3: obtaining an initial monitoring data implicit representation mining network, and determining trainable components in the initial monitoring data implicit representation mining network according to the database capacity of the calibration sample database and the adaptive calibration complexity of the pre-trained network; Step S4: Based on the calibration sample database, the initial monitoring data implicit representation mining network is calibrated by knowledge distillation to optimize the network parameters of the trainable components in the initial monitoring data implicit representation mining network, and the monitoring data implicit representation mining network calibrated by knowledge distillation is obtained; wherein the monitoring data implicit representation mining network calibrated by knowledge distillation is used to perform implicit representation mining on the nuclear power isolation lock Internet of Things monitoring training data in the calibration sample database.

[0162] In step S1, before calibration begins, the data processing device first needs to clarify the specific size of the calibration sample database and the types of abnormal prior markers contained therein. The database capacity determines the number of samples available during the training process, and the number of abnormal prior markers reflects the diversity of abnormal types that the system needs to identify. For example, the data processing device traverses the calibration sample database and counts the total number of samples and the number of occurrences of various abnormal prior markers. These data provide important references for subsequent network design and calibration strategies.

[0163] In step S2, the diversity of abnormal prior labels requires the pre-trained network to have sufficient complexity and generalization ability to accurately identify different types of abnormalities. Therefore, the calibration complexity needs to be adjusted according to the number of abnormal prior labels to ensure that the network can fully learn the characteristics of various types of abnormalities without overfitting due to excessive complexity. For example, the data processing device evaluates the adaptive calibration complexity of the pre-trained network according to the number of abnormal prior labels. Generally speaking, the more labels there are, the larger the feature space that the network needs to learn, and the calibration complexity increases accordingly. To this end, it may be necessary to increase the depth and width of the network or adopt a more complex structure (such as introducing an attention mechanism, residual connection, etc.). For example, for a calibration sample database containing 10 abnormal prior labels, the data processing device can select a convolutional neural network (CNN) with appropriate depth and width as the basic model. In order to further enhance the adaptability of the network, several fully connected layers can be added after the convolution layer, and the ReLU activation function can be used to increase nonlinearity. In addition, in order to alleviate the gradient disappearance problem, a batch normalization layer can also be introduced into the network.

[0164] In step S3, after determining the adaptive tuning complexity of the pre-trained network, the data processing device needs to build an initial monitoring data implicit representation mining network and clarify which components (such as network layers, parameters, etc.) are trainable. This step is the basis of the tuning process and determines the upper limit of the network's learning ability.

[0165] For example, the data processing device designs an initial monitoring data implicit representation mining network based on the evaluation result of step S2. The network may include multiple convolutional layers, pooling layers, fully connected layers and other components. During the construction process, the data processing device will mark which components are trainable (i.e., their parameters will be optimized during the tuning process) and which are not trainable (such as embedded layers with fixed parameters).

[0166] For example, assume that the initial monitoring data implicit representation mining network adopts a CNN structure consisting of three convolutional layers, two pooling layers and one fully connected layer. Among them, the parameters of the convolutional layer and the fully connected layer are trainable, while the parameters of the pooling layer (such as pooling window size, step size, etc.) are fixed. In addition, in order to introduce nonlinear transformations, a ReLU activation function layer is connected to each convolutional layer. In step S4, knowledge distillation is a model compression and transfer learning technique that allows the knowledge of a large model (teacher model) to be transferred to a smaller model (student model). In the calibration of the monitoring data implicit representation mining network, knowledge distillation can be used to optimize the network parameters and improve the generalization ability and computational efficiency of the network.

[0167] For example, the data processing device first selects a teacher model with good performance (which can be a pre-trained large neural network), and then uses the calibration sample database to perform knowledge distillation calibration on the initial monitoring data implicit representation mining network (student model). During the calibration process, the output of the teacher model will be used as soft targets to guide the training of the student model, so that the student model can reduce the number of parameters and calculations while maintaining accuracy.

[0168] Specific adjustment steps may include: Data preprocessing: Perform preprocessing operations such as standardization and normalization on the samples in the calibration sample database to ensure the consistency of the input data.

[0169] Forward propagation: The preprocessed samples are input into the initial monitoring data implicit representation mining network to calculate the output of the student model.

[0170] Calculate the loss: Use the output of the teacher model as a soft target and calculate the loss (such as cross entropy loss) between the output of the student model and the output of the teacher model. At the same time, you can also add a loss term for the hard target (i.e. the true label) to balance the weight between the teacher's guidance and the true label.

[0171] Back propagation: Back propagation is performed based on the loss value to update the parameters of the trainable components in the student model. This process may take multiple iterations to reach convergence.

[0172] Evaluation and Adjustment: Regularly evaluate the performance of the student model (such as accuracy and recall on the validation set) during the tuning process, and adjust hyperparameters (such as learning rate, batch size, etc.) or network structure based on the evaluation results.

[0173] For example, assume that the teacher model is a deep residual network (ResNet) with powerful feature extraction capabilities, and the student model is the initial monitoring data implicit representation mining network constructed in step S3. During the calibration process, the data processing device inputs the samples in the calibration sample database into the teacher model and the student model to obtain the teacher output and the student output, respectively. Then, the cross entropy loss function is used to calculate the difference between the student output and the teacher output, and the parameters of the student model are optimized by the back propagation algorithm. After multiple iterations of calibration, the performance of the student model gradually approaches or even exceeds the teacher model, while the number of parameters and the amount of calculation are greatly reduced.

[0174] Through the above steps, the data processing device successfully completed the calibration process of the monitoring data implicit representation mining network. The calibrated network not only has stronger feature extraction capabilities and higher recognition accuracy, but also can reduce the amount of calculation and storage requirements while maintaining performance. This provides a solid foundation and reliable data support for the subsequent nuclear power intelligent isolation lock system.

[0175] As an implementation method, step S4, performing knowledge distillation calibration on the initial monitoring data implicit representation mining network based on the calibration sample database, may include: Step S4a: for each nuclear power isolation lock IoT monitoring training data in the calibration sample database, determine the abnormal supervision information of the corresponding nuclear power isolation lock IoT monitoring training data, the abnormal supervision information of the nuclear power isolation lock IoT monitoring training data is a label sequence; each label in the label sequence corresponds to an abnormal priori labeled sample, and indicates whether the corresponding nuclear power isolation lock IoT monitoring training data matches the corresponding abnormal priori labeled sample; Step S4b: Determine the label update strategy corresponding to each label according to the value of each label in the label sequence; Step S4c: updating each label in the label sequence according to the label update strategy corresponding to each label, and obtaining updated abnormal supervision information; Step S4d: perform knowledge distillation and calibration on the initial monitoring data implicit representation mining network according to the abnormal supervision information of the label update corresponding to each nuclear power isolation lock IoT monitoring training data in the calibration sample database.

[0176] Step S4 involves knowledge distillation tuning of the initial network based on the tuning sample database. Knowledge distillation is an effective model compression and transfer learning technique that extracts knowledge from a large teacher model (usually a model with excellent performance but complex computation) and passes it to a small student model (i.e., the initial monitoring data implicit representation mining network being tuned), thereby reducing the computational burden of the student model while maintaining high accuracy. The following is a detailed explanation of each sub-step in step S4, explained in conjunction with specific application scenarios, and examples are used to assist understanding.

[0177] In step S4a, each piece of nuclear power isolation lock IoT monitoring training data is accompanied by a set of abnormal supervision information, which exists in the form of a label sequence. Each element (label) in the label sequence is a numerical value (such as 0 or 1) that indicates whether the monitoring data contains a specific type of abnormality. For example, if 10 abnormal prior markers are defined in the system (such as "cooling system failure", "electrical system failure", etc.), each piece of monitoring data will correspond to a label sequence of length 10, and each element in the sequence corresponds to whether an abnormal prior marker exists.

[0178] For example, the data processing device traverses each piece of nuclear power isolation lock IoT monitoring training data in the calibration sample database and reads its corresponding abnormal supervision information (label sequence). These label sequences are usually labeled by domain experts based on the actual performance of the monitoring data, ensuring the accuracy and reliability of the labels.

[0179] In step S4b, in order to improve the generalization ability and robustness of the student model, it is sometimes necessary to adjust or smooth the output of the teacher model. This adjustment is usually achieved by modifying the values ​​in the label sequence, that is, applying an update strategy to each label to generate a softer target distribution. This "soft" target distribution can provide more information about the similarities between categories and help the student model learn a smoother decision boundary. For example, the data processing device determines a corresponding label update strategy based on the value (0 or 1) of each label in the label sequence. This strategy may include converting hard labels (0 or 1) into soft labels (i.e., probability values ​​between 0 and 1), or applying specific weighting strategies to specific types of labels. The specific form of the label update strategy depends on the requirements of the calibration task and the characteristics of the teacher model.

[0180] In step S4c, after determining the update strategy corresponding to each label, the data processing device updates the label sequence of all monitoring data in the calibration sample database according to these strategies. The updated label sequence will be used as the target distribution of the student model training in the knowledge distillation process.

[0181] For example, the data processing device traverses each monitoring data in the calibration sample database and its corresponding original label sequence. For each label value in the sequence, the value is replaced or adjusted according to its update strategy. The update process may involve simple value replacement (such as the above-mentioned label smoothing strategy) or more complex calculations (such as weighted processing based on the output of the teacher model).

[0182] In step S4d, after obtaining the updated abnormal supervision information, the data processing device will use this knowledge to perform knowledge distillation tuning on the student model (i.e., the initial monitoring data implicit representation mining network). The goal of knowledge distillation tuning is to optimize the parameters of the student model by minimizing the difference between the output of the student model and the output of the teacher model, while maintaining or improving the performance of the student model.

[0183] For example, the data processing device runs the teacher model to predict each monitoring data in the calibration sample database and obtains the output distribution of the teacher model (usually a probability vector). Then, the updated label sequence corresponding to each monitoring data is used as the "soft target" of the teacher model to compare with the predicted output of the student model. The difference between the two is measured by defining a loss function (such as cross entropy loss), and the back propagation algorithm is used to optimize the parameters of the student model. In the process of knowledge distillation, in order to balance the guiding role of the teacher model and the learning ability of the student model itself, the loss between the predicted output of the student model and the true label (ie, hard target) is usually taken into consideration. In this way, the student model will not only learn useful knowledge from the teacher model, but also be constrained by the true label, thereby avoiding overfitting the specific bias of the teacher model.

[0184] In order to balance the weight between soft targets and hard targets, the data processing device can also include the loss between the predicted output of the student model and the true label sequence (i.e., the original hard label sequence) into the loss function. In this way, the student model will be guided by the output of the teacher model and constrained by the true label during the training process. After multiple iterations of training, the performance of the student model will gradually improve and approach the level of the teacher model or even exceed the teacher model (in some cases).

[0185] An embodiment of the present application provides a data processing device, including a memory and a processor, wherein the memory stores a data processing device program that can be run on the processor, and when the processor executes the program, some or all of the steps in the above method are implemented.

[0186] Figure 2 A hardware entity diagram of a data processing device provided in an embodiment of the present application, such as Figure 2As shown, the hardware entity of the data processing device 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a data processing device program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

[0187] The embodiment of the present application also provides an intelligent isolation lock system that applies the above-mentioned data processing method. The system can support multi-terminal collaborative access from PC, mobile, and control cabinet host, covering key authentication and authorization, lock status management, equipment storage / return / scrapping / loss, isolation sharing, emergency unlocking, operation log, data cockpit and other functions. It can play a role in enforcing the implementation of safety production responsibilities and is suitable for online and offline multi-scenario use. The intelligent isolation locker included in the system supports three ways to receive and return smart devices: face recognition, card swiping, and email verification. It can also monitor the status of the device through RFID technology (radio frequency identification technology); the system has a corresponding smart key, which is powered by a rechargeable lithium battery, has a built-in Bluetooth module, and has a unique ID, which is convenient for identification. Through Bluetooth connection with a tablet computer, remote authorization is achieved, and precise control and efficient management of the device status are achieved.

[0188] The intelligent isolation lock system uses dual means of task authorization and equipment consistency verification, and sets operating permissions for users and limits the time range of key permissions to ensure the accuracy of implementation objects; the system operator can monitor the equipment isolation status in real time on the electronic work platform, and the operation content can be updated in time. It can also ensure that the isolation process is traceable and save operation records, including time, personnel and other information, to ensure subsequent query and traceability.

Claims

1. A data processing method for a nuclear power intelligent isolation lock system based on the Internet of Things, characterized in that: The method comprises: Obtaining the Internet of Things monitoring data of the nuclear power isolation lock and the abnormal diagnosis report of the Internet of Things monitoring data of the nuclear power isolation lock, and performing implicit representation mining on the Internet of Things monitoring data of the nuclear power isolation lock and the abnormal diagnosis report, respectively, to obtain the implicit representation of the monitoring data and the implicit representation of the diagnosis report; Merging the monitoring data implicit representation and the diagnosis report implicit representation to obtain a first fused implicit representation; Acquire abnormal information involving features, wherein the abnormal information involving features include implicit representations of labels corresponding to a plurality of abnormal a priori labels in a preset abnormal information involving topological structure; the plurality of abnormal a priori labels represent a plurality of sub-categories under an abnormal classification; for each of the abnormal a priori labels, the implicit representation of the labels of the abnormal a priori labels includes not only the abnormal information of the abnormal a priori labels, but also the abnormal information of the abnormal a priori labels connected to the corresponding abnormal a priori labels in the abnormal information involving topological structure; Merging the first fused implicit representation with the implicit representations of each mark in the feature involved in the abnormal information, respectively, to obtain second fused implicit representations corresponding to the multiple abnormal prior marks; According to the second fused implicit representations respectively corresponding to the multiple abnormal prior marks, a target abnormal prior mark corresponding to the nuclear power isolation lock IoT monitoring data is determined from the multiple abnormal prior marks; the target abnormal prior mark is used to indicate abnormal information in the nuclear power isolation lock IoT monitoring data.

2. The method according to claim 1, characterized in that Before obtaining the abnormal information involving features, the method further includes: Acquiring an abnormal information-involved topological structure; the abnormal information-involved topological structure includes a plurality of abnormal a priori markers, and there are connections between the abnormal a priori markers of the involved situations; Constructing a connection relationship array and a feature embedding array according to the topological structure involved in the abnormal information; the feature embedding array includes feature embedding results corresponding to the multiple abnormal prior markers respectively, and the connection relationship array is used to describe the connection between the multiple abnormal prior markers; Merging the connection relationship array with the feature embedding array to obtain a first generation integrated array; In a current generation starting from the second generation after the first generation, the integrated array output by the previous generation is merged with the connection relationship array to obtain the integrated array output by the current generation; Determine the next generation as the current generation, jump to the step of merging the integration array output by the previous generation with the connection relationship array, and iterate until the first stop judgment requirement is met; Based on the integrated array output by the last generation, the features involved in the abnormal information are determined.

3. The method according to claim 2, characterized in that The rows in the feature embedding array are matched one by one with the abnormal a priori marks, and the rows in the connection relationship array are matched one by one with the abnormal a priori marks; in the feature embedding array and the connection relationship array, the abnormal a priori marks corresponding to the rows of the feature embedding array and the rows of the connection relationship array having the same number are consistent; the connection relationship array and the feature embedding array are merged to obtain a first generation integrated array, including: Performing array fusion on the connection relationship array and the feature embedding array to obtain a fused array; Adding the fused array to the feature embedding array to obtain a merged array; A standardization operation is performed on the merged array to obtain a first generation integrated array.

4. The method according to claim 1, characterized in that The process of mining implicit representations of the nuclear power isolation lock IoT monitoring data includes: Determine a monitoring data implicit representation mining network; the monitoring data implicit representation mining network includes a plurality of sequentially connected skip connection components; Based on the first jump connection component in the monitoring data implicit representation mining network, the nuclear power isolation lock IoT monitoring data is subjected to implicit representation mining to obtain the monitoring data mining implicit representation; Determine, in a current generation starting from a second generation after the first generation, a current skip connection component in a mining network corresponding to the current generation in which the monitoring data implicitly represents the mining network; Performing implicit representation mining on the monitoring data mining implicit representation output by the previous skip connection component based on the current skip connection component to obtain an implicit representation of the transition monitoring data; Adding the monitoring data mining implicit representation output by the previous skip connection component to the transition monitoring data implicit representation to obtain the monitoring data mining implicit representation output by the current skip connection component; Taking the next generation as the current generation, jumping to the step of determining the current jump connection component corresponding to the current generation in the monitoring data implicit representation mining network, and iterating until the second stop judgment requirement is met; According to the monitoring data outputted by the last generation, an implicit representation is mined to determine the monitoring data implicit representation of the nuclear power isolation lock IoT monitoring data.

5. The method according to claim 1, characterized in that The implicit representation mining process of the abnormal diagnosis report includes: For each diagnostic event in the abnormal diagnostic report, embedding the component tags of the current diagnostic event to obtain an event embedding representation corresponding to the current diagnostic event; For each event data item in the current diagnosis event, determine the distribution order of the current event data item in the current diagnosis event, and embed the distribution order to obtain a distribution embedding representation corresponding to the current event data item; Performing data item embedding on the current event data item to obtain a data item embedding representation corresponding to the current event data item, and obtaining a feature embedding result corresponding to the current event data item based on the event embedding representation corresponding to the current diagnosis event, the distribution embedding representation corresponding to the current event data item, and the data item embedding representation; The diagnostic report implicit representation of the abnormal diagnostic report is determined according to the feature embedding results respectively corresponding to each event data item in the abnormal diagnostic report.

6. The method according to claim 1, characterized in that The step of merging the monitoring data implicit representation and the diagnosis report implicit representation to obtain a first fused implicit representation includes: Acquire a contribution parameter set corresponding to the implicit representation of the monitoring data and a contribution parameter set corresponding to the implicit representation of the diagnostic report; each contribution parameter set is composed of a reduced-dimensional contribution parameter array obtained by performing singular value decomposition on a corresponding preset contribution parameter array; Merging each dimension-reduced contribution parameter array in the contribution parameter set corresponding to the implicit representation of the monitoring data with the implicit representation of the monitoring data to obtain a plurality of weighted implicit representations of the monitoring data; Merging each dimension-reduced contribution parameter array in the contribution parameter set corresponding to the implicit representation of the diagnosis report with the implicit representation of the diagnosis report to obtain a plurality of weighted implicit representations of the diagnosis report; The multiple monitoring data weighted implicit representations and the multiple diagnosis report weighted implicit representations are merged to obtain a first fused implicit representation.

7. The method according to claim 6, characterized in that The contribution parameter set corresponding to the implicit representation of the monitoring data is obtained by using the following steps: Obtaining a preset contribution parameter array corresponding to the implicit representation of the monitoring data; Determining the linear independence of the preset contribution parameter array corresponding to the implicit representation of the monitoring data; the linear independence of the preset contribution parameter array corresponding to the implicit representation of the monitoring data represents the dimension of the preset contribution parameter array corresponding to the implicit representation of the monitoring data; The preset contribution parameter array corresponding to the implicit representation of the monitoring data is split into multiple reduced-dimensionality contribution parameter arrays with linear independence equal to 1; the number of reduced-dimensionality contribution parameter arrays obtained by the splitting is consistent with the linear independence of the preset contribution parameter array corresponding to the implicit representation of the monitoring data.

8. The method according to any one of claims 1 to 7, characterized in that: The target abnormal a priori marker is output based on a data analysis neural network, and the data analysis neural network is adjusted by the following operations to obtain: Acquire a calibration sample database; the calibration sample database includes a plurality of nuclear power isolation lock IoT monitoring training data, and abnormal diagnosis report samples and abnormal supervision information corresponding to the plurality of nuclear power isolation lock IoT monitoring training data; For each nuclear power isolation lock IoT monitoring training data in the plurality of nuclear power isolation lock IoT monitoring training data, performing implicit representation mining on the corresponding nuclear power isolation lock IoT monitoring training data and the corresponding abnormal diagnosis report sample, and obtaining the implicit representation of the inference monitoring data and the implicit representation of the inference diagnosis report; Merging the inference monitoring data implicit representation and the inference diagnosis report implicit representation to obtain a first fused inference implicit representation; Acquire features involved in the abnormal information sample, wherein the features involved in the abnormal information sample include implicit representations of sample labels corresponding to a plurality of abnormal a priori labeled samples in the abnormal information-involved topological structure sample; for each of the abnormal a priori labeled samples, the implicit representation of the sample labels of the abnormal a priori labeled sample includes not only the abnormal information of the abnormal a priori labeled sample, but also the abnormal information of the abnormal a priori labeled samples connected to the abnormal a priori labeled sample in the abnormal information-involved topological structure sample; Merging the first fused reasoning implicit representation with the implicit representation of each sample label in the feature involved in the abnormal information sample, respectively, to obtain second fused reasoning implicit representations corresponding to the multiple abnormal a priori label samples; According to the second fused reasoning implicit representations respectively corresponding to the multiple abnormal prior labeled samples, the inferred abnormal prior labeled samples corresponding to the nuclear power isolation lock Internet of Things monitoring training data are determined among the multiple abnormal prior labeled samples, and according to the error between the inferred abnormal prior labeled samples and the abnormal supervision information corresponding to the corresponding nuclear power isolation lock Internet of Things monitoring training data, the network parameters of the data analysis neural network are optimized.

9. The method according to claim 8, characterized in that The data analysis neural network includes a monitoring data implicit representation mining network; Before performing implicit representation mining on the nuclear power isolation lock IoT monitoring training data in the calibration sample database, the method further includes: Determining the database capacity of the calibration sample database and the number of labels of the plurality of abnormal priori labeled samples; Determining the adaptive calibration complexity of the pre-trained network according to the number of labels of the plurality of abnormal prior labeled samples; Acquire an initial monitoring data implicit representation mining network, and determine trainable components in the initial monitoring data implicit representation mining network according to the database capacity of the calibration sample database and the adaptive calibration complexity of the pre-trained network; Based on the calibration sample database, the initial monitoring data implicit representation mining network is calibrated by knowledge distillation to optimize the network parameters of the trainable components in the initial monitoring data implicit representation mining network, and obtain the monitoring data implicit representation mining network calibrated by knowledge distillation; The monitoring data implicit representation mining network calibrated by the knowledge distillation is used to perform implicit representation mining on the nuclear power isolation lock IoT monitoring training data in the calibration sample database; The step of performing knowledge distillation calibration on the initial monitoring data implicit representation mining network based on the calibration sample database includes: For each nuclear power isolation lock IoT monitoring training data in the calibration sample database, determine the abnormal supervision information of the corresponding nuclear power isolation lock IoT monitoring training data, wherein the abnormal supervision information of the nuclear power isolation lock IoT monitoring training data is a label sequence; each label in the label sequence corresponds to an abnormal priori labeled sample, and indicates whether the corresponding nuclear power isolation lock IoT monitoring training data matches the corresponding abnormal priori labeled sample; Determine a label update strategy corresponding to each label according to the value of each label in the label sequence; According to the label update strategy corresponding to each label, each label in the label sequence is updated to obtain updated abnormal supervision information; According to the abnormal supervision information of the label update corresponding to each nuclear power isolation lock IoT monitoring training data in the calibration sample database, the initial monitoring data implicit representation mining network is calibrated for knowledge distillation.

10. A data processing device, comprising a memory and a processor, wherein the memory stores a data processing device program that can be run on the processor, characterized in that: When the processor executes the program, the steps in the method according to any one of claims 1 to 9 are implemented.

Images