Multi-parameter coupling-oriented industrial device running state evaluation system

Abstract

A kind of industrial equipment running state evaluation system for multi-parameter coupling, comprising: preprocessing module, time series reconstruction module, anomaly detection and fault diagnosis module, fault unit importance mining module, fault unit division module and evaluation report generation module, the present application is established according to equipment manual, historical maintenance record and historical operation data by causal mining fault causal diagram mining coupling between parameters and evaluates the importance of each fault unit.Online state evaluation, the real-time data of equipment detection are divided and based on time series reconstruction model adaptive anomaly detection is carried out, finally, according to fault causal diagram and fault unit importance weight, equipment is scored and fault analysis is carried out, and equipment running state comprehensive diagnostic report is generated.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of neural network applications, and particularly relates to an industrial equipment operation state evaluation system for multi-parameter coupling. BACKGROUND

[0002] Industrial equipment operation state evaluation helps to ensure stable operation of equipment, improve production efficiency and quality, however, the current industrial equipment state evaluation method has the following defects: on the one hand, large-scale industrial equipment has a complex structure mechanism, a large number of parameters and strong coupling, data usually has the characteristics of high dimension, nonlinearity and non-stationarity, the existing method is often limited to single or partial parameters, lacks consideration of the whole process state parameters, and ignores the mutual influence between parameters; on the other hand, large-scale industrial equipment has many fault manifestations, has the characteristics of uncertainty and multi-fault coupling, the evaluation guide of the existing method depends on expert experience, the subjective factor accounts for a large proportion, and it is difficult to reflect the real operation state of the equipment. SUMMARY

[0003] The application is aimed at the problems of a large number of parameters and strong coupling of the existing large-scale industrial equipment, lack of objectivity of equipment evaluation and low method interpretability, and the deficiencies of the existing supervised learning mode and dependence on high-quality labeled data, and proposes an industrial equipment operation state evaluation system for multi-parameter coupling, which establishes a fault causal diagram to mine the coupling relationship between parameters and evaluates the importance of each fault unit according to the equipment manual, historical maintenance records and historical operation data. In online state evaluation, the real-time data detected by the equipment are divided and adaptive anomaly detection is carried out based on a time series reconstruction model, finally, the equipment is scored and fault analyzed according to the fault causal diagram and the importance weight of the fault unit, and a comprehensive diagnosis report of the equipment operation state is generated.

[0004] The application is implemented by the following technical scheme:

[0005] The application relates to a multi-parameter coupling-oriented industrial equipment operation state evaluation system, comprising a preprocessing module, a time sequence reconstruction module, an anomaly detection and fault diagnosis module, a fault unit importance mining module, a fault unit division module and an evaluation report generation module, wherein: the fault unit division module performs Bert-based causal word recognition on the equipment manual and the equipment historical maintenance record in an offline stage to obtain causal prior knowledge, performs time alignment and discretization processing on the equipment historical operation data to obtain equipment historical labeled data, obtains a fault causal structure graph through fault structure search and scoring, and then generates a plurality of independent fault units through connected component mining and outputs the independent fault units to the preprocessing module, the evaluation report generation module and the fault unit importance mining module; the fault unit importance mining module performs fault cost statistics and constructs a fault cost binary weight distribution type judgment matrix according to the independent fault units in the offline stage, obtains the importance weight of the fault units by solving the weight vector of the judgment matrix, and then outputs the importance weight to the evaluation report generation module; the preprocessing module performs null value processing, time alignment and data division processing on the equipment monitoring real-time data in an online stage to obtain a complete time sequence after filling, the time sequence reconstruction module mines parameter mutual information and time characteristics through a graph neural network and a time sequence neural network according to the original input complete time sequence, and performs time sequence reconstruction to obtain a reconstructed time sequence, the anomaly detection and fault diagnosis module performs anomaly determination and fault diagnosis according to the difference between the original input and the reconstructed time sequence in the online stage to obtain an anomaly and fault type result, and the evaluation report generation module performs operation state scoring and fault analysis according to the anomaly and fault type information in the online stage to obtain an operation state evaluation report.

[0006] Technical effects

[0007] The application discloses an equipment fault unit division method based on K2 score causal structure mining, an adaptive anomaly detection method based on the combination of a graph attention network and an LSTM, and a fault diagnosis method based on fault causal graph walking; an adaptive threshold value method based on an optimal F-measure is adopted to determine the threshold value of anomaly determination, compared with a traditional method based on expert experience or a fixed threshold value of a POT theory, the threshold value can be dynamically adjusted according to real determination records, and the adaptability is good; a fault cause and potential harm are mined based on the causal graph walking to generate an equipment evaluation report, which is beneficial to troubleshooting and repair by operation and maintenance personnel, and the method has strong interpretability; the above fault unit division, time sequence reconstruction, anomaly detection and equipment evaluation report generation module are dominated by data, the degree of manual participation and the degree of dependence on expert experience are low, and the method has good portability. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 The method framework diagram of the application;

[0009] Figure 2 System structure diagram for the embodiment of the present application. DETAILED DESCRIPTION

[0010] As Figure 1 shown, the embodiment of the present application relates to a multi-parameter coupling-oriented industrial equipment operation state evaluation system, which comprises a preprocessing module, a time series reconstruction module, an anomaly detection and fault diagnosis module, a fault unit importance mining module, a fault unit division module and an evaluation report generation module.

[0011] The fault unit division module comprises a causal predicate identification unit, a fault structure search and scoring unit and a connected component mining unit, wherein: after the causal predicate identification unit performs preprocessing operations such as text data cleaning and replacement, sentence segmentation, word segmentation and part-of-speech tagging, it extracts <subject, predicate, object> triples composed of two nouns and one verb according to word and syntax pattern matching, and finally identifies the causal relationship; the fault structure search and scoring unit initializes the nodes in the graph and the priority order between the nodes based on causal prior knowledge, greedily searches for the causal graph structure, makes it have the optimal structure score, and outputs the causal structure graph with the highest score; the connected component mining unit converts the fault causal structure graph into an undirected graph, performs depth-first traversal on the entire graph until all nodes are visited, divides the nodes in the same batch into the same fault unit during the process, and finally restores the directed edge relationship between the nodes and outputs the fault unit set that is independent of each other.

[0012] The text data comprises equipment manuals and historical maintenance records.

[0013] The causal predicate identification specifically comprises: selecting some known causal predicates and relational words as the causal word set. On the one hand, the causal relationship between the subject and the predicate in each triple is considered, the vectorization representation of the predicate verb in the triple is learned based on Bert, and the cosine similarity between the predicate verb and the words in the causal relationship word set is calculated. If the similarity between the word and each word in the word set is less than a threshold value, the triple is considered to have no causal relationship, otherwise the triple has a causal relationship, and the word is added to the causal relationship word set. On the other hand, the causal relationship between adjacent triples is considered. Similarly, the causal relationship words between two triples are learned and identified based on Bert. If there is a causal relationship between the two triples, the two triples are combined into two events respectively, and the new triple <event 1, cause, event 2> is formed by connecting the cause. Finally, the causal predicate identification unit based on Bert outputs the <subject, predicate, object> triple containing the causal relationship as the causal prior knowledge.

[0014] The structure scoring refers to a scoring function

[0015] wherein: D represents the data set after discretization processing, G represents the causal structure, n represents the total number of parameters, i is used to traverse the parameters, q i represents the value number of the i-th parameter parent node, j is used to traverse all values of a certain parameter parent node, r i represents the value number of the i-th parameter, k is used to traverse all values of a certain parameter, m ijk represents the parameter X i whose value is v ik and whose parent node is the j-th value, the total number of records in the data set, represents the total number of records containing the parameter X i all values and the parent node is the j-th value.

[0016] The fault unit importance mining module comprises a fault cost statistics unit, a fault cost matrix construction unit and a weight vector solving unit, wherein: the fault cost statistics unit takes the equipment maintenance record text data as input, extracts the average repair cost of each fault unit from three aspects of time, money and manpower according to a pre-designed keyword set, and is respectively denoted as Cost time，i , Cost money，i , Cost manpower，i , and outputs the average fault cost of each fault unit after normalization processing The fault cost matrix construction unit constructs a K*K binary distribution type matrix M according to the average repair cost Cost average，i of each fault unit; and the weight vector solving unit obtains a weight vector by establishing an optimization objective function and solving according to the adjusted binary distribution type judgment matrix M.

[0017] In the keyword identification, the time aspect keywords include {maintenance duration, downtime duration, response time, occurrence time, repair time, impact duration, time consumption, hour, minute}; the money aspect keywords include {maintenance fee, equipment depreciation, spare parts fee, labor cost, expenditure, cost, yuan}; and the manpower aspect keywords include {manpower input, personnel configuration, number of maintenance personnel, person}.

[0018] The normalization processing is specifically: wherein: type can take values {time, money, manpower}, Cost' type，i represents the normalized repair cost of the i-th fault unit type, MIN type is 0 by default, and MAX type represents the maximum cost of all fault units of this type.

[0019] In the binary distribution type matrix M: K represents the total number of fault units, and each element in the matrix is The relative importance of the i-th fault unit and the j-th fault unit was measured;

[0020] The optimization objective function Where: w i w represents the importance weight of the i-th fault unit. j The weight vector W = {w1, ..., wj} represents the importance weight of the j-th fault unit. i , ...w K}

[0021] The data preprocessing module includes a null value processing unit, a time alignment unit, and a data partitioning unit. The null value processing unit addresses the issue of missing values ​​at certain time points in the original monitoring data by using previous value filling, i.e., filling the missing monitoring values ​​at the previous time point with the values ​​from the previous time point. The time alignment unit standardizes the time scale and determines the monitoring value for each time point, i.e., taking the most recent monitoring value of each monitoring parameter before that time point. The data partitioning unit, based on the specific parameter nodes contained within each fault unit in the support product, divides the aligned monitoring data of each parameter into K independent fault units, serving as the data input for subsequent spatiotemporal information reconstruction of each unit.

[0022] The unified time scale is specifically defined as follows: the latest start time, the earliest end time, and the smallest time step are selected to unify the time axis.

[0023] The time series reconstruction module includes K parallel and similar reconstruction units corresponding to K independent fault units. Each reconstruction unit is a network model containing a parametric mutual information mining layer, a temporal feature extraction layer, and a temporal decoding reconstruction layer. Specifically, the parametric mutual information mining layer uses a graph attention neural network, which receives a k*T multi-parameter time series matrix I and outputs a k*T matrix O. The temporal feature extraction layer uses a long short-term memory recurrent neural network to map the mined feature matrix to a specific latent vector space Z. The temporal decoding reconstruction layer uses a long short-term memory network to decode and reconstruct the latent vectors to generate a new multi-parameter time series matrix I'.

[0024] In the graph attention neural network described above: k represents the number of parameters contained in the unit, T represents the length of the time series, and each element in the output matrix O... This represents the feature value of the i-th parameter at time j after fusing information from other parameters, where σ is the activation function and α is the eigenvalue. im These are network parameters.

[0025] The overall loss function of the reconstruction unit in: This term represents the log-likelihood of reconstructing the original time series I from the latent vector space Z. Essentially, this term can be represented by the reconstruction error, i.e., Difference(I, I'). The difference between the vector space Z mapped from the input I and the prior distribution of Z is measured, typically assuming that Z follows a standard normal distribution. During training, the reconstruction unit uses only normal time series as training samples and minimizes the loss. After training, the reconstruction unit can effectively map normal samples to the latent vector space and then recover normal samples from the latent vector space. During prediction, the entire time series reconstruction module receives time series data from K faulty units and outputs time series data reconstructed from K faulty units.

[0026] The aforementioned anomaly detection and fault diagnosis module includes an anomaly determination unit and a fault diagnosis unit, wherein: the anomaly determination unit achieves anomaly determination through reconstruction error assessment and threshold adaptive update; the fault diagnosis unit performs fault classification and updates the fault sample library based on the data determined to be abnormal.

[0027] The aforementioned reconstruction error assessment refers to: calculating the difference between the original time-series data and the reconstructed time-series data in the time domain.

[0028] The aforementioned adaptive threshold update refers to: establishing an optimization objective function F with respect to the threshold τ. β =(1+β) 2 )×P×R / (β 2 P+R), where P represents precision, R represents recall, and β is a control coefficient. When 0 < β < 1, precision is more important; when β = 1, precision and recall are equally important; and when β > 1, recall is more important. By maximizing F... β To obtain the optimal threshold τ.

[0029] The aforementioned fault classification refers to: calculating the difference between the current time series and the time series of the built-in historical fault sample database using DTW, and then classifying faults based on KNN using this difference. The diagnosed results are then manually confirmed and added to the historical sample database.

[0030] The assessment report generation module includes an operation status scoring unit and a fault analysis unit. The operation status scoring unit gives a comprehensive score for the entire device based on the reconstruction error and importance weight of each fault unit. The fault analysis unit moves through the fault cause-effect graph according to the possible fault types to generate an operation status assessment report.

[0031] The aforementioned comprehensive score in:

[0032] εi reconstruction error of the i-th faulty unit, μ i mean of the reconstruction error of the i-th faulty unit calculated based on normal samples, σ i standard deviation of the reconstruction error of the i-th faulty unit calculated based on normal samples, s i score of the i-th faulty unit, w i importance weight of the i-th faulty unit, Score represents the running state score of the entire device.

[0033] The said wandering filling refers to: starting from the current fault type node to perform a depth-first traversal, for each accessed node, if the node type is a parameter representation, a report text "may cause + <node content> + phenomenon" is generated, if the node type is a fault anomaly, a report text "may cause + <node content> + harm" is generated, according to the access order of the node, the texts are assembled and spliced into the potential harm part of the report; then reverse the causal graph, here from the node to perform a depth-first traversal, for each accessed node, when the node type is a parameter representation, a report text "may have appeared + <node content> + phenomenon" is generated, if the node type is a fault anomaly, a report text "may have occurred + <node content> + fault" is generated, and the texts are assembled and filled into the fault reason part of the report according to the access order. The finally generated report contains the following parts: device name, evaluation time, running state score, fault record, wherein the fault record records the fault diagnosis result of the current device in table form, each row contains fault type, fault reason and potential harm fields.

[0034] As Figure 2 shown, it is a specific application scenario of the embodiment, including a Web application, the above-mentioned industrial equipment running state evaluation system and an infrastructure layer, wherein: the Web application uses a React framework to build a front-end page, renders and displays incoming real-time data, and periodically calls a back-end service interface in a timer manner to evaluate the running state of the equipment, and exports the running report generated by the back-end in the form of a file; the industrial equipment running state evaluation system responds to the call of the upper interactive layer, and meets the data processing and model dependence of internal functions by calling a structured data transmission MySQL interface (Pymysql encapsulation implementation), a graph data transmission Neo4j interface (py2neo encapsulation implementation) and a file I / O interface; the infrastructure layer is used for persistent storage of data, which is divided into three parts according to different data types, wherein the MySQL data is used to store historical running data and historical fault cases; Neo4J is used as a graph database to store the causal structure graph of each faulty unit; the local file system is used to store the equipment manual, the historical maintenance record of the equipment and the trained time series reconstruction model.

[0035] The comparison of the system with the prior art is shown in Table 1.

[0036] Table 1 Comparison of technical characteristics

[0037]

[0038]

[0039] Compared with the prior art, the fault unit division module in the application adopts causal structure mining based on K2 score to obtain independent fault units, and the anomaly detection and fault determination of each unit do not affect each other and can be independently and parallelly run, so that the flexibility and execution efficiency of the method are higher; the time series reconstruction module adopts a model based on the combination of graph attention network and LSTM, fully considers the spatiotemporal feature interaction of the original input, and makes the feature extraction of the method more rich and complete; the anomaly detection module adopts an adaptive threshold method based on the optimal F-measure to determine the threshold of anomaly determination, compared with the traditional method of fixed threshold based on expert experience or statistical POT theory, the threshold can be dynamically adjusted according to the real determination record, and the method has better applicability; the evaluation report generation module adopts a method based on causal graph walking to mine fault causes and potential hazards to generate a report, which is beneficial to the troubleshooting and repair of operation and maintenance personnel, and the method has better interpretability; the above fault unit division, time series reconstruction, anomaly detection and equipment evaluation report generation modules are mainly based on data, and the degree of artificial participation and the degree of dependence on expert experience are low, and the method has better portability.

[0040] The above specific implementations can be adjusted in different ways by those skilled in the art without departing from the principles and purposes of the application, the protection scope of the application is subject to the claims and is not limited by the above specific implementations, and each implementation scheme within the scope is subject to the constraints of the application.

Claims

1. A system for evaluating the operational status of industrial equipment oriented towards multi-parameter coupling, characterized in that, include: The system comprises a preprocessing module, a time series reconstruction module, an anomaly detection and fault diagnosis module, a fault unit importance mining module, a fault unit partitioning module, and an evaluation report generation module. Specifically: the fault unit partitioning module performs BERT-based causal word recognition on equipment manuals and historical maintenance records offline to obtain causal prior knowledge; it performs time alignment and discretization on historical equipment operation data to obtain historical labeled data; it then obtains a fault causal structure graph through fault structure search and scoring; and finally generates several independent fault units through connected component mining, which are output to the preprocessing module, the evaluation report generation module, and the fault unit importance mining module, respectively. The fault unit importance mining module performs fault cost statistics based on basic keyword recognition for independent fault units offline and constructs a binary weighted allocation of fault costs. The judgment matrix is ​​used to determine the importance weight of the faulty unit by solving the judgment matrix weight vector. The weight is then output to the evaluation report generation module. In the online stage, the preprocessing module performs null value processing, time alignment, and data partitioning based on the real-time monitoring data of the equipment to obtain a complete time series after filling. The time series reconstruction module performs parametric mutual information and time feature mining through graph neural networks and temporal neural networks based on the complete time series of the original input to obtain a reconstructed time series. In the online stage, the anomaly detection and fault diagnosis module performs anomaly determination and fault diagnosis based on the differences between the original input and the reconstructed time series to obtain anomaly and fault type results. In the online stage, the evaluation report generation module performs operation status scoring and fault analysis based on the anomaly and fault type information to obtain an operation status evaluation report.

2. The industrial equipment operation status assessment system oriented towards multi-parameter coupling according to claim 1, characterized in that, The fault unit partitioning module includes: a causal predicate recognition unit, a fault structure search and scoring unit, and a connected component mining unit. Specifically: the causal predicate recognition unit performs preprocessing operations such as cleaning and replacing text data, sentence segmentation, word segmentation, and part-of-speech tagging. Based on part-of-speech and syntactic pattern matching, it extracts <subject, predicate, object> triples consisting of two nouns and one verb, and finally identifies causal relationships. The fault structure search and scoring unit initializes the nodes and their priority order in the graph based on causal prior knowledge, then greedily searches for the causal graph structure to achieve the optimal structure score and outputs the causal structure graph with the highest score. The connected component mining unit transforms the fault causal structure graph into an undirected graph, performs a depth-first traversal of the entire graph until all nodes are visited. Nodes visited in the same batch are classified into the same fault unit. Finally, it restores the directed edge relationships between nodes and outputs independent fault unit sets.

3. The industrial equipment operation status assessment system oriented towards multi-parameter coupling according to claim 2, characterized in that, The text data includes: equipment manuals and historical maintenance records; The aforementioned causal predicate recognition specifically includes: selecting a subset of known causal predicates and relation words as a causal word set; considering the causal relationship between the subject and predicate in each triplet, learning the vectorized representation of the predicate verb in the triplet based on BERT, and calculating the cosine similarity between the predicate verb and the words in the causal word set; if the similarity between the word and every word in the word set is less than a threshold, then the triplet is considered to have no causal relationship; otherwise, the triplet has a causal relationship, and the word is added to the causal word set; considering the causal relationship between adjacent triplets, similarly learning to identify causal relation words between two triplets based on BERT; if there is a causal relationship between two triplets, then merging the two triplets into two events, and connecting them with the predicate cause to form a new triplet <event 1, cause, event 2>; finally, the BERT-based causal predicate recognition unit outputs a <subject, predicate, object> triplet containing causal relationship as causal prior knowledge. The aforementioned structural scoring refers to: scoring function Where: D represents the discretized dataset, G represents the causal structure, n represents the total number of parameters, and i is used to iterate through the parameters. This represents the number of values ​​that the parent node of the i-th parameter can take, and j is used to iterate through all the values ​​that the parent node of a given parameter can take. This represents the number of possible values ​​for the i-th parameter, and k is used to iterate through all possible values ​​for a given parameter. Representing parameters Values And when its parent node is the j-th value, the total number of records in the dataset, , indicating that parameters are included The total number of records when all values ​​are taken and the parent node is the j-th value.

4. The industrial equipment operation status assessment system oriented towards multi-parameter coupling according to claim 1, characterized in that, The fault unit importance mining module includes: a fault cost statistics unit, a fault cost matrix construction unit, and a weight vector solution unit. The fault cost statistics unit takes equipment maintenance record text data as input and extracts the average repair cost of each fault unit from three aspects: time, money, and manpower, based on a pre-designed keyword set. These costs are denoted as follows: , , After normalization, the average failure cost of each faulty unit is output. The fault cost matrix construction unit is based on the average repair cost of each fault unit. Construct a K*K binary allocation matrix M; the weight vector solving unit obtains the weight vector by establishing and solving the optimization objective function based on the adjusted binary allocation judgment matrix M.

5. The industrial equipment operation status assessment system oriented towards multi-parameter coupling according to claim 4, characterized in that, In the keyword recognition described above, time-related keywords include {maintenance duration, downtime, response time, occurrence time, repair time, impact duration, time consumed, hours, minutes}; money-related keywords include {maintenance cost, equipment depreciation, spare parts cost, labor cost, expenditure, expense, yuan}; and human resource-related keywords include {human resource input, personnel allocation, number of maintenance personnel, people}. The normalization process specifically refers to: Where: type can take the values ​​{time, money, manpower}. This represents the normalized repair cost for the i-th faulty unit type. The default value is 0. This represents the maximum cost of all faulty units within this category of costs; In the binary allocation matrix M, K represents the total number of faulty units, and each element in the matrix is ​​denoted as K. This measures the relative importance of the i-th fault unit and the j-th fault unit. The optimization objective function ,in: This represents the importance weight of the i-th faulty unit. The weight vector represents the importance weight of the j-th fault unit.

6. The industrial equipment operation status assessment system oriented towards multi-parameter coupling according to claim 1, characterized in that, The data preprocessing module includes a null value processing unit, a time alignment unit, and a data partitioning unit. The null value processing unit addresses the issue of missing values ​​at certain time points in the original monitoring data by using previous values ​​to fill in the missing monitoring values ​​at the previous time point. The time alignment unit standardizes the time scale and determines the monitoring value for each time point, i.e., it takes the most recent monitoring value of each monitoring parameter before that time point. The data partitioning unit divides the aligned monitoring data of each parameter into K independent fault units based on the specific parameter nodes contained in each fault unit in the support product, serving as the data input for subsequent spatiotemporal information reconstruction of each unit. The unified time scale is specifically defined as follows: the latest start time, the earliest end time, and the smallest time step are selected to unify the time axis.

7. The industrial equipment operation status assessment system oriented towards multi-parameter coupling according to claim 1, characterized in that, The time series reconstruction module includes K parallel and similar reconstruction units corresponding to K independent fault units. Each reconstruction unit is a network model containing a parametric mutual information mining layer, a temporal feature extraction layer, and a temporal decoding reconstruction layer. Specifically, the parametric mutual information mining layer uses a graph attention neural network, which receives a k*T multi-parameter time series matrix I and outputs a k*T matrix O. The temporal feature extraction layer uses a long short-term memory recurrent neural network to map the mined feature matrix to a specific latent vector space Z. The temporal decoding reconstruction layer uses a long short-term memory network to decode and reconstruct the latent vectors to generate a new multi-parameter time series matrix. .

8. The industrial equipment operation status assessment system oriented towards multi-parameter coupling according to claim 7, characterized in that, In the graph attention neural network described above: k represents the number of parameters contained in the unit, T represents the length of the time series, and each element in the output matrix O... This represents the feature value of the i-th parameter at time j after fusing information from other parameters, where It is an activation function. These are network parameters; The overall loss function of the reconstruction unit ,in: Let represent the log-likelihood of reconstructing the original time series I from the latent vector space Z. Essentially, this term is represented by the reconstruction error, i.e., Difference(I, ); The difference between the vector space Z mapped by the input I and the prior distribution of Z is measured. It is usually assumed that Z follows a standard normal distribution. During training, the reconstruction unit only uses normal time series as training samples and minimizes the loss. After training, the reconstruction unit can map normal samples to the latent vector space well and then recover normal samples from the latent vector space. During prediction, the entire time series reconstruction module receives the time series data of K fault units and outputs the time series data reconstructed by K fault units.

9. The industrial equipment operation status assessment system oriented towards multi-parameter coupling according to claim 1, characterized in that, The aforementioned anomaly detection and fault diagnosis module includes an anomaly determination unit and a fault diagnosis unit, wherein: the anomaly determination unit achieves anomaly determination through reconstruction error assessment and threshold adaptive update; the fault diagnosis unit performs fault classification and updates the fault sample library based on the data determined to be abnormal; The aforementioned reconstruction error assessment refers to: calculating the difference between the original time-series data and the reconstructed time-series data in the time domain. ; The aforementioned threshold adaptive update refers to: establishing a threshold... Optimization objective function ,in: Indicates accuracy rate. Indicates recall rate, It is a control factor, when When accuracy is more important, it indicates that... At that time, accuracy and recall are equally important. At that time, recall is more important, by maximizing To obtain the most threshold ; The aforementioned fault classification refers to: calculating the difference between the current time series and the time series of the built-in historical fault sample library through DTW, and classifying fault categories based on KNN based on this difference. The results of the diagnosis are then manually confirmed and added to the historical sample database.

10. An application system based on the evaluation system described in any one of claims 1-9, characterized in that, include: The system comprises a web application, an evaluation system, and an infrastructure layer. The web application uses the React framework to build the front-end page, rendering and displaying the incoming real-time data. It periodically calls the backend service interface via a timer to evaluate the equipment's operating status and exports the generated operating report as a file. The industrial equipment operating status evaluation system responds to calls from the interaction layer and calls the structured data transmission MySQL interface, the graph data transmission Neo4j interface, and the file I / O interface to meet the data processing and model dependencies of its internal functions. The infrastructure layer is used for persistent data storage and is divided into three parts based on data type: MySQL data stores historical operating data and historical fault cases; Neo4j, as a graph database, stores the causal structure graph of each fault unit; and the local file system stores equipment manuals, historical equipment maintenance records, and trained time-series reconstruction models.

Images