Controller Loop Self-Diagnosis Method, Device and Equipment

Abstract

The present invention provides a controller loop self-diagnosis method, device and equipment. The method includes: extracting multi-level phase space features from multiple sensor signals in the control system of the target device to construct a feature time series matrix; constructing a functional dependence graph of the controller loop based on the feature time series matrix and calculating time-varying causal relationships, and obtaining a masking effect decoupling matrix by analyzing abnormal flow distributions; constructing a fault interaction potential field by combining the feature time series matrix and the masking effect decoupling matrix, analyzing the fault interaction characteristics on different time scales, and generating a fault interaction feature map; inputting the fault interaction feature map into a preset working condition perception conditional probability network, and applying a hybrid inference framework to output a comprehensive diagnosis result of multiple faults. The present invention can effectively decouple the masking effect between multiple faults, accurately identify and diagnose various existing fault types, avoid the triggering of dangerous false alarms, and improve the safety and reliability of equipment operation.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment diagnosis, and particularly to a controller loop self-diagnosis method, device, and equipment. Background Art

[0002] Controller loop self-diagnosis technology has been widely applied in the field of industrial automation, especially in the equipment control systems of key industries such as power and chemical industries. Existing controller loop self-diagnosis methods are mainly designed based on a single fault model, and identify and locate abnormalities in the control loop through methods such as observer models, spectrum analysis, or statistical feature matching. However, in the actual industrial environment, the simultaneous occurrence of multiple faults is not uncommon. For example, in a condenser liquid level control system, a water level gauge blockage may coexist with a control valve failure. When multiple faults occur simultaneously, the symptoms they produce may mask or modify each other, resulting in the failure of traditional diagnosis methods based on a single fault assumption. Some fault signals are masked by other faults, and the system may be in a dangerous "false healthy" state, or trigger false fault alarms, affecting the safe operation and production efficiency of the equipment. Summary of the Invention

[0003] The main objective of the present invention is to solve the technical problem of the failure or misdiagnosis of existing controller loop self-diagnosis methods when facing multiple faults due to the mutual masking effect between faults;

[0004] The first aspect of the present invention provides a controller loop self-diagnosis method, and the controller loop self-diagnosis method includes:

[0005] Extract multi-level phase space features from multiple sensor signals in the control system of the target device, and construct a feature time series matrix based on the extracted phase space features;

[0006] Based on the feature time series matrix, construct a functional dependency graph of the controller loop of the target device and calculate the time-varying causal relationship between components in the functional dependency graph. By analyzing the abnormal flow distribution in the time-varying causal relationship, obtain a masking effect decoupling matrix;

[0007] Based on the feature time series matrix and the masking effect decoupling matrix, construct a fault interaction potential field and analyze the fault interaction characteristics of the fault interaction potential field on different time scales to generate a fault interaction feature map;

[0008] Input the fault interaction feature map into a preset working condition perception conditional probability network, and apply a hybrid inference framework to output a comprehensive diagnosis result of multiple faults.

[0009] Optionally, in the first implementation manner of the first aspect of the present invention, the multi-level phase space feature extraction of multiple sensor signals in the control system of the target device and the construction of the feature time series matrix according to the extracted phase space features include:

[0010] Apply the delay coordinate method to reconstruct the phase space of the time series signals of the multiple sensors to obtain a high-dimensional phase space trajectory;

[0011] Calculate the non-linear dynamics index for the high-dimensional phase space trajectory, and use the adaptive sliding window technique to calculate the non-linear dynamics index segment by segment to obtain a time-varying non-linear dynamics index sequence;

[0012] Take the high-dimensional phase space trajectory, the non-linear dynamics index, and the time-varying non-linear dynamics index sequence as phase space features, and construct a feature time series matrix according to the extracted phase space features.

[0013] Optionally, in the second implementation manner of the first aspect of the present invention, based on the feature time series matrix, construct a functional dependence graph of the controller loop of the target device and calculate the time-varying causal relationship between components in the functional dependence graph. By analyzing the abnormal traffic distribution in the time-varying causal relationship, the masking effect decoupling matrix is obtained, including:

[0014] Analyze the phase space feature correlation of each sensor in the feature time series matrix, and combine the physical structure of the target device and the schematic diagram of the controller loop to construct an initial functional dependence graph;

[0015] Based on the time-varying sequence of the non-linear dynamics index in the feature time series matrix, use the improved PC algorithm of the conditional mutual information criterion to determine the causal association direction and strength between components, and optimize the initial functional dependence graph based on the causal association direction and strength;

[0016] Based on the optimized functional dependence graph, divide the feature time series matrix into multiple time periods, apply the time-varying Granger causality analysis method to each time period, establish a vector autoregressive model and calculate the statistic to obtain the time-varying causal relationship representing the change of each component over time;

[0017] Regard the causal strength in the time-varying causal relationship as the abnormal traffic distribution, and calculate the abnormal absorption rate for each component according to the principle of abnormal traffic conservation to construct a masking effect decoupling matrix.

[0018] Optionally, in the third implementation manner of the first aspect of the present invention, the step of regarding the causal strength in the time-varying causal relationship as the abnormal traffic distribution, calculating the abnormal absorption rate for each component according to the principle of abnormal traffic conservation, and constructing the masking effect decoupling matrix includes:

[0019] Establish a directed weighted graph structure for the time-varying causal relationship, where nodes represent components of the target device, the direction of the edges represents the direction of the causal relationship, and the weight of the edges represents the causal intensity. Define the causal intensity as the abnormal traffic distribution between components.

[0020] For each node in the directed weighted graph, calculate the sum of the abnormal traffic distribution flowing into the node and the sum of the abnormal traffic distribution flowing out of the node to obtain the abnormal absorption rate of each node.

[0021] Construct a masking effect decoupling matrix based on the abnormal absorption rate and the causal intensity. The rows and columns of the masking effect decoupling matrix correspond to components, and the matrix element values are determined according to the abnormal absorption rate of the node and the relative causal intensity between two nodes.

[0022] Optionally, in the fourth implementation manner of the first aspect of the present invention, constructing a fault interaction potential field based on the feature time series matrix and the masking effect decoupling matrix and analyzing the fault interaction characteristics of the fault interaction potential field on different time scales to generate a fault interaction feature map includes:

[0023] Reorganize the feature time series matrix into a third-order tensor according to three dimensions: the number of sensors, the number of time windows, and the feature dimension, and apply a constrained Tucker decomposition algorithm to the third-order tensor to extract a spatial pattern matrix representing the spatial pattern of the sensors and a core tensor representing pattern interaction.

[0024] Calculate the Hadamard product of the spatial pattern matrix and the masking effect decoupling matrix. The Hadamard product is used to adjust the influence weights of the components in the spatial pattern matrix. Combine the adjusted spatial pattern matrix with the core tensor to define a fault interaction potential field.

[0025] Obtain the measured values of the thermodynamic parameters of the target device, calculate the thermodynamic compensation factor according to the thermodynamic model of the device, and compensate and adjust the fault interaction potential field according to the thermodynamic compensation factor to obtain a compensated fault interaction potential field.

[0026] Apply wavelet transform to the compensated fault interaction potential field for multi-scale decomposition, extract the fault interaction components at each scale, and map the fault interaction components to a fault interaction feature map.

[0027] Optionally, in the fifth implementation manner of the first aspect of the present invention, applying wavelet transform to the compensated fault interaction potential field for multi-scale decomposition, extracting the fault interaction components at each scale, and mapping the fault interaction components to a fault interaction feature map includes:

[0028] Apply discrete wavelet transform to the compensated fault interaction potential field according to a preset wavelet basis function to decompose the fault interaction potential field into wavelet decomposition coefficients at different scales.

[0029] Calculate the energy distribution of the wavelet decomposition coefficients at different scales, obtain the energy contribution rate of each scale layer, and determine the time scale at which the main fault interaction occurs according to the energy contribution rate;

[0030] Extract the phase information of the time scale at which the main fault interaction occurs, and identify the time delay characteristics and synchronization characteristics of the fault interaction between different components through cross-scale phase coherence analysis;

[0031] Convert the energy distribution, phase information, time delay characteristics and synchronization characteristics as fault interaction components into a graph structure representation, and apply a community detection algorithm to the graph structure representation to map it into a fault interaction feature map.

[0032] Optionally, in the sixth implementation manner of the first aspect of the present invention, inputting the fault interaction feature map into a preset working condition perception conditional probability network and applying a hybrid inference framework, the comprehensive diagnosis result of multiple faults includes:

[0033] By monitoring the key operating parameters of the target device, map the key operating parameters into a predefined working condition space by combining the fuzzy clustering algorithm to determine the current working condition category;

[0034] According to the determined current working condition category, call the corresponding conditional probability network model from the pre-trained conditional probability network library, convert the fault interaction feature map into a network input feature vector, input it into the conditional probability network, and calculate the posterior probabilities of each fault type and the combination of fault types through forward inference;

[0035] Process the posterior probabilities using the Dempster-Shafer evidence theory framework, define a basic probability assignment function for each fault type and combination, and calculate the uncertainty factor according to the operating state of the target device;

[0036] Use the basic probability assignment function and the uncertainty factor to calculate the comprehensive trust degree through the Dempster combination rule, and generate a comprehensive diagnosis result in combination with the interaction characteristics in the fault interaction feature map.

[0037] The second aspect of the present invention provides a controller loop self-diagnosis device, and the controller loop self-diagnosis device includes:

[0038] A feature extraction module for hierarchically collecting the direct thermal characteristic parameters, indirect thermal characteristic parameters and coolant quality characteristic parameters of the air compressor of the fuel cell to obtain a multi-dimensional thermal data set of the air compressor;

[0039] A causality analysis module, configured to extract features from the multi-dimensional thermal data set, input the extracted feature data set into a neural network model based on deep transfer learning, and obtain a corresponding heat dissipation efficiency decay prediction model;

[0040] An interaction recognition module, configured to establish a multi-objective optimization equation of the air compressor based on the multi-dimensional thermal data set and the heat dissipation efficiency decay prediction model, and solve the multi-objective optimization equation to obtain a heat dissipation control strategy set;

[0041] A fault diagnosis module, configured to dynamically partition the heat dissipation system of the air compressor, and coordinately adjust the flow rate and flow direction of the coolant in each area of the dynamic partition according to the heat dissipation control strategy set, so as to achieve the heat dissipation of the air compressor.

[0042] The third aspect of the present invention provides a controller loop self-diagnosis device, including: a memory and at least one processor, instructions are stored in the memory, and the memory and the at least one processor are interconnected by a line; the at least one processor calls the instructions in the memory, so that the controller loop self-diagnosis device executes the steps of the above-mentioned controller loop self-diagnosis method.

[0043] The above-mentioned controller loop self-diagnosis method, device and equipment extract multi-level phase space features from multiple sensor signals in the target device control system to construct a feature time series matrix; construct a functional dependence graph of the controller loop based on the feature time series matrix and calculate the time-varying causality relationship, and obtain a masking effect decoupling matrix by analyzing the abnormal flow distribution; construct a fault interaction potential field by combining the feature time series matrix and the masking effect decoupling matrix, analyze the fault interaction characteristics on different time scales, and generate a fault interaction feature map; input the fault interaction feature map into a preset working condition perception conditional probability network, and apply a hybrid inference framework to output a comprehensive diagnosis result of multiple faults. The present invention can effectively decouple the masking effect between multiple faults, accurately identify and diagnose various existing fault types, avoid the triggering of dangerous false alarms, and improve the safety and reliability of equipment operation.

[0044] Other features and advantages of the present invention will be described in the following description, and, in part, will be obvious from the description, or will be understood by implementing the present invention. The objectives and other advantages of the present invention are achieved and obtained by the structures specifically pointed out in the description, the claims and the drawings.

[0045] In order to make the above-mentioned objectives, features and advantages of the present invention more obvious and understandable, the following specifically gives preferred embodiments and cooperates with the attached drawings to make a detailed description as follows. Description of the Drawings

[0046] Figure 1Schematic diagram of the first embodiment of the controller loop self-diagnosis method in the embodiments of the present invention;

[0047] Figure 2 Schematic diagram of an embodiment of the controller loop self-diagnosis device in the embodiments of the present invention;

[0048] Figure 3 Schematic diagram of an embodiment of the controller loop self-diagnosis equipment in the embodiments of the present invention. Detailed implementation manners

[0049] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0050] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but optionally further includes other unlisted steps or units, or optionally further includes other steps or units inherent to these processes, methods, products or devices.

[0051] To facilitate the understanding of this embodiment, a controller loop self-diagnosis method disclosed in the embodiments of the present invention will be introduced in detail first. As Figure 1 shown, the method includes the following steps:

[0052] 101. Extract multi-level phase space features from multiple sensor signals in the control system of the target device, and construct a feature time series matrix based on the extracted phase space features;

[0053] In an embodiment of the present invention, the extracting multi-level phase space features from multiple sensor signals in the control system of the target device and constructing a feature time series matrix based on the extracted phase space features includes: reconstructing the phase space of the time series signals of the multiple sensors by using the delay coordinate method to obtain a high-dimensional phase space trajectory; calculating non-linear dynamics indexes for the high-dimensional phase space trajectory, and segmentally calculating the non-linear dynamics indexes by using an adaptive sliding window technique to obtain a time-varying non-linear dynamics index sequence; using the high-dimensional phase space trajectory, the non-linear dynamics indexes and the time-varying non-linear dynamics index sequence as phase space features, and constructing a feature time series matrix based on the extracted phase space features.

[0054] Specifically, in this embodiment, taking the condenser liquid level control system as an example, the implementation process of the feature extraction module in the controller loop self-diagnosis method is described in detail. The condenser liquid level control system typically includes a variety of sensors, such as differential pressure liquid level gauges, capacitance liquid level gauges, ultrasonic liquid level gauges, as well as temperature, pressure, and flow sensors, etc. When processing the time series signals of these sensors, it is first necessary to collect raw data of sufficient length, usually collecting data for 30 minutes to 2 hours at the system sampling frequency to capture the dynamic characteristics of the system under different working conditions. The collected signals are preprocessed, including operations such as removing outliers, filtering, and normalization, to ensure data quality. Subsequently, the delay coordinate method is applied to reconstruct the phase space for the preprocessed time series signals of multiple sensors. This method is based on the Takens embedding theorem and can reconstruct the high-dimensional phase space of the system from a single-variable time series, revealing the internal dynamic characteristics of the system. In specific implementation, for the time series x(t) of each sensor, a vector [x(t), x(t+τ), x(t+2τ),..., x(t+(m-1)τ)] is constructed, where τ is the time delay and m is the embedding dimension. In the condenser liquid level control system, the time delay τ is usually determined by calculating the first minimum value of the mutual information function of the time series, while the embedding dimension m is determined by the false nearest neighbor analysis method. For the liquid level signal, the typical embedding dimension is 3 to 5. Through the above process, the one-dimensional time series signals are mapped into an m-dimensional phase space to form high-dimensional phase space trajectories, and these trajectories can reflect the evolution law and dynamic characteristics of the system state.

[0055] Specifically, nonlinear dynamic indices are calculated for the obtained high-dimensional phase space trajectories. These indices can quantify the complex behavior characteristics of the system and provide an important basis for fault diagnosis. First, the correlation dimension is calculated. This index quantifies the complexity of the trajectory by measuring the distribution characteristics of the distances between point pairs in the phase space. Healthy systems and faulty systems usually exhibit different correlation dimension values. Then, the largest Lyapunov exponent is calculated. This index measures the separation rate of adjacent trajectories in the phase space, reflects the chaotic degree of the system and its sensitivity to initial conditions. In a liquid level control loop, valve faults usually lead to a significant increase in this exponent. Then, indices such as determinism, laminarity, and entropy are calculated through recurrence plot analysis. The recurrence plot projects the phase space trajectory onto a two-dimensional plane and analyzes the proximity between trajectory points. Among them, determinism measures the predictability of system behavior, laminarity reflects the tendency of the system state to stay in a specific region, and entropy quantifies the complexity of system dynamics. When calculating these nonlinear dynamic indices, to capture the evolution characteristics of the system state over time, an adaptive sliding window technique is used to calculate the nonlinear dynamic indices in segments. In the condenser liquid level control system, the window width is adaptively adjusted according to the dynamic characteristics of the system, usually ranging from 500 to 2000 data points, corresponding to 1 to 5 minutes of actual time. When it is detected that the system operating conditions change rapidly, the window width is automatically reduced to improve the time resolution; when the system is relatively stable, the window width is increased to improve the statistical reliability of index calculation. The overlap rate of adjacent windows is set to 50%, that is, half of the window width, which not only ensures time continuity but also avoids waste of computing resources. Through sliding window processing, the static nonlinear dynamic indices are transformed into time-varying sequences, forming time-varying nonlinear dynamic index sequences. These sequences can reflect the evolution process of the system's dynamic behavior, especially the system response characteristics when the condenser load changes or external disturbances exist.

[0056] Organize the above three levels of phase space features, namely high-dimensional phase space trajectories, nonlinear dynamics indicators, and time-varying nonlinear dynamics indicator sequences, into a multi-dimensional feature time series matrix, which serves as the basic data structure for subsequent fault diagnosis. The construction process of the feature time series matrix is as follows: First, determine the number of rows and columns of the matrix. The number of rows is equal to the number of sliding windows, representing the time dimension; the number of columns is equal to the total number of phase space features of all sensors, representing the feature dimension. For each sensor, select the key parameters (such as the main components or statistical characteristics of the trajectory) in its high-dimensional phase space trajectory, the calculated nonlinear dynamics indicators (such as correlation dimension, maximum Lyapunov exponent, etc.), and the time-varying sequence values of these indicators, and arrange them in the columns of the matrix in a predefined order. In the actual application of the condenser liquid level control system, a typical feature time series matrix may contain 20 to 50 time windows (rows) and 30 to 100 features (columns). The value of each element in the matrix represents the calculation result of a specific feature within a specific time window. For example, the element in the i-th row and j-th column may represent the maximum Lyapunov exponent value of the liquid level sensor signal within the i-th time window. Standardize each column of the matrix to make features with different dimensions comparable. The feature time series matrix constructed in this way not only retains the dynamic characteristics of the original signal but also contains the deep-level nonlinear system behavior characteristics, and can effectively capture the complex system response in the case of multiple faults.

[0057] 102. Based on the feature time series matrix, construct a functional dependency graph of the controller loop of the target device and calculate the time-varying causal relationships between the components in the functional dependency graph. By analyzing the abnormal flow distribution in the time-varying causal relationships, obtain the masking effect decoupling matrix.

[0058] In an embodiment of the present invention, the constructing a functional dependency graph of the controller loop of the target device based on the feature time series matrix and calculating the time-varying causal relationships between the components in the functional dependency graph, and obtaining the masking effect decoupling matrix by analyzing the abnormal flow distribution in the time-varying causal relationships includes: analyzing the phase space feature correlations of the sensors in the feature time series matrix, and combining the physical structure of the target device and the schematic diagram of the controller loop to construct an initial functional dependency graph; based on the time-varying sequences of the nonlinear dynamics indicators in the feature time series matrix, use the improved PC algorithm based on the conditional mutual information criterion to determine the causal association directions and strengths between the components, and optimize the initial functional dependency graph based on the causal association directions and strengths; based on the optimized functional dependency graph, divide the feature time series matrix into multiple time periods, apply the time-varying Granger causality analysis method to each time period, establish a vector autoregressive model and calculate the statistics to obtain the time-varying causal relationships representing the changes of the components over time; regard the causal strengths in the time-varying causal relationships as the abnormal flow distribution, and according to the principle of abnormal flow conservation, calculate the abnormal absorption rate for each component to construct the masking effect decoupling matrix.

[0059] Specifically, after obtaining the feature time series matrix, it is necessary to construct a controller loop functional dependency graph of the target device based on this matrix to reveal the mutual influence relationships among system components. First, analyze the phase space feature correlations of each sensor in the feature time series matrix. In this process, calculate the Pearson correlation coefficient, Spearman rank correlation coefficient, and mutual information value for each pair of features in the matrix to form a correlation matrix of features. In the condenser liquid level control system, the phase space features of the liquid level signal usually have a high correlation with parameters such as the cooling water flow rate, turbine load, and condenser pressure, which reflects the physical coupling relationship within the system. At the same time, combine the physical structure of the target device and the schematic diagram of the control loop to identify the direct physical connections or signal transmission paths among the components in the system, and clarify the functional boundaries and interface definitions of each component. For example, in the condenser liquid level control loop, there are physical connections from the liquid level sensor to the signal conditioning circuit, from the conditioning circuit to the controller, from the controller to the actuator (such as a control valve), and from the actuator to the controlled object (liquid level). By integrating the results of the phase space feature correlation analysis and the physical structure knowledge, construct an initial functional dependency graph. This graph is represented in the form of a directed graph, where the nodes represent system components (such as various sensors, controllers, actuators, controlled objects, etc.), and the edges represent the physical or signal connection relationships among the components. The initial functional dependency graph reflects the expected connection relationships during system design, but the causal relationships revealed by the actual operation data have not been considered and need to be further optimized.

[0060] Based on the time-varying sequences of non-linear dynamic indicators in the feature time series matrix, the improved PC algorithm using the conditional mutual information criterion is employed to determine the causal association direction and strength between components. The PC algorithm is originally a causal graph learning algorithm based on conditional independence testing, but it has limitations in dealing with industrial process data with non-linear and non-Gaussian distributions. Therefore, in this embodiment, the improved PC algorithm is adopted, changing the conditional independence test from the linear correlation coefficient to conditional mutual information, which is more suitable for capturing complex causal relationships in non-linear systems. The specific implementation process is as follows: First, each column of data in the feature time series matrix is regarded as a random variable, representing the phase space characteristics of a certain component. Then, a fully connected undirected graph is constructed, where each node in the graph corresponds to a component. Next, the conditional independence between any two nodes is judged through conditional mutual information calculation. When the conditional mutual information value is less than the preset threshold, the two nodes are considered conditionally independent, and the corresponding edge is removed. In the condenser system, the threshold is usually set between 0.05 and 0.1, depending on the system noise level and data quality. Subsequently, the directions of the remaining undirected edges are determined to form a directed acyclic graph. This algorithm uses the kernel density estimation method to calculate the high-dimensional mutual information during the edge case processing, avoiding the curse of dimensionality problem in traditional methods. Through this process, the causal association direction and strength between components are obtained, where the direction represents the transmission direction of causal influence, and the strength is represented by the value of conditional mutual information. Based on these causal association information, the initial functional dependency graph is optimized, including adding causal links that are not considered in the physical design but indicated by the data, removing links that are physically connected but actually have no significant causal relationship, and adjusting the edge weights to reflect the causal strength. The optimized functional dependency graph more accurately reflects the actual operating state of the system and the mutual influence relationship between components.

[0061] Based on the optimized functional dependency graph, the feature time series matrix is divided into multiple time periods to analyze the dynamic characteristics of the system at different operating stages. The division of time periods takes into account the system operating conditions and operating cycles. In the condenser liquid level control system, typical division methods include the start-up stage, stable operation stage, load change stage, and disturbance response stage, etc. The time-varying Granger causality analysis method is applied to each time period. This method is based on the prediction principle, that is, if the past values of variable X can help predict the future values of variable Y, then X is considered to Granger cause Y. In the specific operation, for each pair of causally connected nodes in the functional dependency graph, a vector autoregressive model is established. Taking node Y and all its possible parent nodes as an example, first, a complete model including all parent nodes is established, representing the relationship between the current value of Y and the past values of itself and all parent nodes; then, a model excluding a specific The restricted model. By comparing the prediction errors of the complete model and the restricted model, the F-statistic is calculated, which reflects the predictive contribution of the excluded variables to the target variable. The calculation of the F-statistic takes into account model complexity and sample size to ensure statistical significance. This process is carried out separately for each time period, and finally a time-varying causal relationship characterizing the changes over time among components is obtained. The time-varying causal relationship is represented in matrix form, and the matrix element represents the strength of the causal influence of component i on component j at time t. In the condenser liquid level control system, this matrix reveals the dynamic changes in the causal relationships among system components under different operating conditions. For example, when the load changes sharply, the causal influence of the turbine load on the liquid level will increase significantly.

[0062] Regarding the causal strength in the time-varying causal relationship as an abnormal flow distribution, the masking effect of multiple faults is analyzed through the principle of abnormal flow conservation. In this step, the system is regarded as an abnormal signal propagation network, and the causal strength represents the flux of abnormal signals between nodes. Ideally, based on the principle of flow conservation, the total amount of abnormal signals flowing into a node should be equal to the total amount of abnormal signals flowing out of the node, unless the node absorbs or amplifies the abnormal signals. For each time point in the time-varying causal relationship matrix C(t), a directed weighted graph is constructed. The nodes represent system components, the direction of the edges represents the direction of the causal relationship, and the weight of the edges represents the causal strength, that is, the abnormal flow distribution. For each node i, the total sum of the abnormal flow In(i) flowing into the node is calculated, which is the sum of the weights of all the edges pointing to node i; at the same time, the total sum of the abnormal flow Out(i) flowing out of the node is calculated, which is the sum of the weights of all the edges pointing out from node i. According to the principle of abnormal flow conservation, the abnormal absorption rate of node i is defined as (In(i) - Out(i)) / In(i). A positive abnormal absorption rate indicates that the node absorbs part of the abnormal signal, that is, the fault signal is masked; a negative value indicates that the node amplifies the abnormal signal. In the condenser system, when the water level gauge is blocked and the control valve fails simultaneously, the abnormal absorption rate of the control valve node is usually high, indicating that the abnormal signal generated by the water level gauge fault is partially masked by the control valve fault. Based on the abnormal absorption rates of each node and the relative causal strengths between nodes, a masking effect decoupling matrix M is constructed, and the matrix element represents the degree of masking or amplification of the abnormal signal of component i on component j. Specifically, A value close to 1 indicates a strong masking effect, close to 0 indicates no significant masking, and a negative value indicates an amplification effect.

[0063] Further, regarding the causal strength in the time-varying causal relationship as the abnormal traffic distribution, according to the principle of abnormal traffic conservation, calculating the abnormal absorption rate for each component and constructing the masking effect decoupling matrix includes: establishing a directed weighted graph structure for the time-varying causal relationship, where the nodes represent the components of the target device, the direction of the edges represents the direction of the causal relationship, and the weights of the edges represent the causal strength, and defining the causal strength as the abnormal traffic distribution between components; for each node in the directed weighted graph, calculating the sum of the abnormal traffic distribution flowing into the node and the sum of the abnormal traffic distribution flowing out of the node to obtain the abnormal absorption rate of each node; constructing the masking effect decoupling matrix according to the abnormal absorption rate and the causal strength, corresponding the rows and columns of the masking effect decoupling matrix to the components, and determining the matrix element values according to the abnormal absorption rate of the node and the relative causal strength between the two nodes.

[0064] Specifically, in the condenser liquid level control system, the time-varying causal relationship matrix C(t) contains the causal strength information between components at different time points, and these information need to be mapped into a directed weighted graph. In the specific operation, first determine the nodes of the graph, and each node represents a physical component or functional unit in the system, such as a liquid level sensor, a pressure sensor, a controller, a control valve, a water pump, etc. Then establish the connection relationship of the edges according to the causal relationship matrix, and the non-zero elements in the matrix indicate that there is a causal relationship from component i to component j at time t, and accordingly establish a directed edge between the corresponding nodes, and the direction of the edge points from i to j, indicating the transmission direction of the causal influence. At the same time, take the numerical value of the matrix element as the weight of the edge, representing the strength of the causal relationship. In the actual implementation, usually set a threshold θ (such as 0.05), and only establish a connection when the causal strength is greater than θ to filter out weak or noise-induced pseudo-causal relationships. In addition, for the time-varying causal relationship, the causal matrix at a specific time point can be selected for analysis, or the average causal strength over a period of time can be calculated. In the condenser system, a typical directed weighted graph contains 10 to 20 nodes and 30 to 50 edges, forming a complex network structure. The key innovation lies in defining the causal strength as the abnormal traffic distribution between components, and this definition is based on the following understanding: faults or abnormal states in the system will generate signal disturbances through the causal chain, and these disturbances propagate in the system like fluids. The greater the causal strength, the smoother the propagation channel of the abnormal signal, and the corresponding abnormal traffic is larger. In this way, the abstract causal relationship is transformed into the concept of abnormal traffic distribution with physical meaning, laying a foundation for the subsequent masking effect analysis.

[0065] Specifically, for the constructed directed weighted graph, it is necessary to calculate the abnormal traffic distribution of each node, so as to quantify the absorption or amplification effect of each component on the abnormal signal. For each node i in the graph, first calculate the total sum of the abnormal traffic distribution flowing into this node, In(i), that is, the sum of the weights of all edges pointing to node i. This calculation process is formally expressed as the accumulation of all weight values w(j, i) that satisfy the existence of the edge j→i. In the condenser liquid level control system, for the controller node, the abnormal traffic flowing in comes from multiple signal sources such as liquid level sensors and pressure sensors; for the control valve node, it mainly comes from the output signal of the controller. Similarly, calculate the total sum of the abnormal traffic distribution flowing out of node i, Out(i), that is, the sum of the weights of all edges pointing out from node i, representing the total amount of abnormal signals generated by node i spreading to other components. After obtaining the total sum of the abnormal traffic flowing in and out, calculate the abnormal absorption rate α(i) = (In(i) - Out(i)) / In(i) of node i. The abnormal absorption rate is a value in the range of [-∞, 1], and its physical meaning is the processing characteristic of the node for the abnormal signal flowing through it: when α(i) is close to 1, it means that node i absorbs almost all the abnormal signals flowing in, that is, a strong masking effect; when α(i) is close to 0, it means that node i transmits the abnormal signal almost without loss, without significant masking or amplification; when α(i) is negative, it means that node i amplifies the abnormal signal and the output exceeds the input. In the condenser system, this situation is common in the control link where self-excited oscillation occurs. It should be noted that when the total sum of the abnormal traffic flowing in, In(i), is 0, the abnormal absorption rate cannot be directly calculated, and at this time, α(i) is set to 0, indicating that this node does not participate in the processing of the abnormal signal. By calculating the abnormal absorption rate of each component in the system, a quantitative representation of the masking or amplification characteristics of each component for the fault signal is obtained, which is the basic data for constructing the masking effect decoupling matrix.

[0066] Specifically, based on the calculated abnormal absorption rate and causal strength, constructing the masking effect decoupling matrix is the key link to achieve multiple fault separation. Both the rows and columns of this matrix M correspond to the components in the system, and the matrix element represents the masking or amplification degree of component i for the abnormal signal of component j. The specific calculation of the matrix element considers two factors: the abnormal absorption rate of the source node and the relative causal strength between the two nodes. First, set = α(i), that is, the diagonal elements of the matrix are equal to the abnormal absorption rate of the corresponding node, representing the processing characteristic of the node for its own fault signal. For the non-diagonal elements (i≠j), they are calculated in the following way: if there is a direct causal path from i to j, then =α(i)×(w(i,j) / Out(i)), where w(i,j) is the edge weight from i to j, and this formula distributes the abnormal absorption rate of node i according to the relative causal strength from i to j; if there is no direct path, but there is an indirect path i→k→j passing through an intermediate node k, then = × , representing the composite masking effect transmitted along the path; if there is neither a direct path nor an indirect path, then =0, indicating that i has no masking effect on j. In actual calculations, the maximum length of the indirect path is usually limited (such as 2 or 3) to control the computational complexity and avoid weak effects on overly long paths. For a condenser liquid level control system, a typical size of the masking effect decoupling matrix is 10×10 to 20×20, depending on the number of system components. The masking effect decoupling matrix constructed in this way comprehensively describes the abnormal signal modulation relationship between components in the system. Larger positive values in the matrix indicate strong masking regions that need special attention because faults in these regions are easily masked by other faults and not detected; while larger negative values represent signal amplification regions, and small faults in these regions may be misjudged as serious problems.

[0067] 103. Based on the characteristic time series matrix and the masking effect decoupling matrix, construct a fault interaction potential field and analyze the fault interaction characteristics of the fault interaction potential field on different time scales to generate a fault interaction characteristic map;

[0068] In an embodiment of the present invention, the constructing a fault interaction potential field based on the characteristic time series matrix and the masking effect decoupling matrix and analyzing the fault interaction characteristics of the fault interaction potential field on different time scales to generate a fault interaction characteristic map includes: restructuring the characteristic time series matrix into a third-order tensor according to three dimensions of the number of sensors, the number of time windows, and the feature dimension, and applying a constrained Tucker decomposition algorithm to the third-order tensor to extract a spatial pattern matrix representing the spatial pattern of the sensors and a core tensor representing pattern interaction; calculating the Hadamard product of the spatial pattern matrix and the masking effect decoupling matrix, where the Hadamard product is used to adjust the influence weights of each component in the spatial pattern matrix, combining the adjusted spatial pattern matrix with the core tensor to define a fault interaction potential field; obtaining the measured thermodynamic parameters of the target device, calculating a thermodynamic compensation factor according to the thermodynamic model of the device, and compensating and adjusting the fault interaction potential field according to the thermodynamic compensation factor to obtain a compensated fault interaction potential field; applying wavelet transform to the compensated fault interaction potential field for multi-scale decomposition, extracting fault interaction components at each scale, and mapping the fault interaction components into a fault interaction characteristic map.

[0069] Specifically, in the condenser liquid level control system, the characteristic time series matrix contains various phase space characteristics of each sensor in different time windows. Usually, the rows represent the time windows and the columns represent the characteristics. To better capture the interaction relationships between sensors, this two-dimensional structure needs to be reorganized into a third-order tensor. When specifically implementing, it is first necessary to determine the three dimensions of the tensor after reorganization: the number of sensors, the number of time windows, and the characteristic dimension of each sensor. For the sensor dimension, there are 5 to 10 types in the system, including liquid level, pressure, temperature, flow rate, etc.; the number of time windows depends on the duration of data acquisition and the setting of the sliding window, usually there are 20 to 50; the characteristics of each sensor include the key parameters of phase space reconstruction, nonlinear dynamics indicators, etc., usually 5 to 10. When performing dimension reorganization, first group the columns of the original matrix according to the sensor type, and then rearrange each group of data according to the time window and characteristic type. In specific operations, write a data reorganization algorithm, sequentially read each element in the original matrix, and place it in the corresponding position of the third-order tensor according to the predefined mapping relationship to ensure that different characteristics of the same sensor are continuously stored in the tensor. After the reorganization is completed, perform a constrained Tucker decomposition on the tensor. This algorithm first decomposes the tensor by the alternating least squares method, setting the initial rank to the minimum value that can retain more than 85% of the information. In the condenser system, it is usually 3 to 5, and then iteratively optimize the decomposition result until convergence. In each iteration, fix two factor matrices to solve the third one, and use the Lagrange multiplier method to handle the constraint conditions. The constraint conditions include the non-negativity and sparsity of the spatial mode. The iteration stop condition is that the relative error is less than the preset threshold (such as 0.001) or the maximum number of iterations is reached (such as 100 times). After the decomposition is completed, extract the matrix representing the sensor spatial mode and the core tensor representing the mode interaction. The former describes the cooperative relationship between sensors, and the latter characterizes the interaction relationship of these spatial modes at different times and characteristics.

[0070] Specifically, calculating the Hadamard product of the spatial pattern matrix and the masking effect decoupling matrix is the core step in constructing the fault interaction potential field. This process integrates the data-driven characteristic patterns with the masking effect based on the physical model. The Hadamard product performs an element-wise multiplication operation on two matrices of the same dimension, and each element of the resulting matrix is the product of the corresponding elements of the original matrices. When implementing this calculation process, first ensure that the dimensions of the two matrices match. In the condenser system, the number of rows of the spatial pattern matrix (i.e., the number of sensors) is usually 5 to 10, while the masking effect decoupling matrix is a square matrix of the same dimension. When the dimensions do not match, adjust the matrix size by row expansion or compression. Specific methods include duplicating rows, deleting rows, or merging multiple rows into one row. After the adjustment, write a loop to traverse each corresponding element of the two matrices, multiply them, and store the result in the same position of the resulting matrix. The computational complexity of this process is the square of the number of sensors. The calculated Hadamard product is actually an adjustment of the original spatial pattern, correcting the pattern distortion caused by fault masking. Next, combine the adjusted spatial pattern matrix with the core tensor to define the fault interaction potential field. In implementation, a tensor-matrix product operation is used to calculate the product of the adjusted spatial pattern matrix and the first dimension of the core tensor. The specific algorithm is as follows: Initialize an output tensor with the same dimension as the core tensor, and then for each slice of the core tensor, multiply it by the adjusted spatial pattern matrix and store the result in the corresponding position of the output tensor. In the code implementing this calculation, a triple nested loop is used to traverse the elements of the tensor, and matrix multiplication is used to calculate the updated value of each slice. The computational complexity of this operation depends on the dimension of the tensor and the number of spatial patterns, and usually requires thousands to tens of thousands of basic arithmetic operations in the condenser system. Through the combined operations of the Hadamard product and the tensor-matrix product, a multi-dimensional potential field reflecting the fault interaction characteristics between components is successfully constructed.

[0071] Specifically, obtaining the measured values of the thermodynamic parameters of the target device and performing compensation and adjustment based on the thermodynamic model is an important link to ensure the accuracy of the fault interaction potential field. This process solves the problem of the modulation of fault characteristics by the change of the system's thermodynamic state. In the condenser liquid level control system, the acquisition of thermodynamic parameters is achieved through sensors distributed throughout the system, including temperature sensors and pressure sensors. The temperature sensors use PT100 or thermocouples and transmit temperature data through 4-20mA current signals; the pressure sensors use piezoresistive or capacitive sensors and also transmit through 4-20mA signals. These original signals are converted into actual physical quantity values after A / D conversion and linear calibration by the data acquisition card. During the data acquisition process, median filtering is used to remove interference signals, and the median of 10 sampling points within one second is taken as the current measured value to reduce the influence of instantaneous fluctuations. The calculation of the thermodynamic compensation factor is based on the thermodynamic model of the condenser, which is obtained by training with historical data during system deployment. In specific implementation, first, the deviation between the current thermodynamic parameters and the reference state is calculated. The reference state is the average value under the system design conditions or during historical normal operation. Then, these deviations are converted into compensation factors through polynomial fitting or look-up table methods. This factor is a value between 0 and 1, representing the similarity between the current thermodynamic state and the reference state. During the calculation process, temperature and pressure have different weights, which are determined according to system characteristics and expert experience and are stored in the calculation program in the form of constants. The implementation of thermodynamic compensation uses a multiplication operation, multiplying each element of the fault interaction potential field by the same compensation factor. This operation is achieved by traversing all elements of the tensor through multiple loops, and the time complexity is proportional to the size of the tensor. In the compensated fault interaction potential field, the abnormal characteristics caused by the change of the thermodynamic state are appropriately weakened, while the characteristics caused by real faults remain at the original intensity, thus enhancing the pertinence of fault recognition.

[0072] Specifically, applying wavelet transform to the compensated fault interaction potential field for multi-scale decomposition is an important means to identify fault interaction characteristics, and this process reveals fault interaction patterns at different time scales. In the specific implementation, first, a wavelet basis function suitable for industrial process data analysis is selected. The 4th-order Daubechies wavelet is usually adopted for the condenser system, which has good signal reconstruction characteristics and computational efficiency. The program implementation of wavelet transform uses the pyramid algorithm to process each one-dimensional signal sequence of the fault interaction potential field along the time dimension. This algorithm first applies a low-pass filter and a high-pass filter to the signal to obtain the approximation coefficients and detail coefficients respectively, and then decimates the approximation coefficients by a factor of two to form the input of the next layer, and so on recursively until the preset decomposition level is reached. In the implementation, the pre-computed filter coefficient array is used to obtain the coefficients of each layer through convolution operations. The convolution uses the boundary extension method to process the signal edges to avoid distortion. For the condenser system, usually 3 to 5 layers of decomposition are performed to capture the dynamic characteristics at different time scales. Subsequently, the energy distribution of the detail coefficients of each layer is calculated. The method is to solve the sum of the squares of the coefficients and divide it by the total energy to obtain the energy proportion of each scale. The time scale at which the main fault interaction occurs is determined according to the energy distribution. Usually, the scale layer with an energy proportion exceeding 10% is selected as the focus of attention. Then, the characteristic coefficients are extracted from each important scale layer. The specific method is to retain the part of the coefficient whose absolute value exceeds the threshold, and at the same time consider its time position and sign. Finally, the fault interaction components of each scale extracted are mapped into a fault interaction feature map. In the implementation, a graphical representation method is adopted, where each system component is represented as a node in the graph, and the interaction relationship between components is represented as an edge between nodes. The attributes of the nodes (size, color) encode the severity of the fault, and the attributes of the edges (thickness, color) represent the interaction intensity and direction. The graph construction algorithm first initializes an empty graph structure, then sets the attributes of the nodes and edges according to the characteristic values of the fault interaction components, and finally converts the graph structure into a visual format for output, such as a network graph or a heat map.

[0073] Further, applying wavelet transform to the compensated fault interaction potential field for multi-scale decomposition, extracting fault interaction components at each scale, and mapping the fault interaction components to a fault interaction feature map includes: applying discrete wavelet transform to the compensated fault interaction potential field according to a preset wavelet basis function, decomposing the fault interaction potential field into wavelet decomposition coefficients at different scales; calculating the energy distribution of the wavelet decomposition coefficients at different scales, obtaining the energy contribution rate of each scale layer, and determining the time scale at which the main fault interaction occurs according to the energy contribution rate; extracting the phase information of the time scale at which the main fault interaction occurs, and identifying the time delay characteristics and synchronization characteristics of fault interaction between different components through cross-scale phase coherence analysis; converting the energy distribution, phase information, time delay characteristics and synchronization characteristics into a graph structure representation as fault interaction components, and applying a community detection algorithm to the graph structure representation to map it to a fault interaction feature map.

[0074] Specifically, in the condenser liquid level control system, DB4 (4th order Daubechies wavelet) is usually selected as the preset basis function. In specific implementation, the compensated fault interaction potential field is first expanded into a one-dimensional signal sequence according to the time dimension. Then, the Mallat fast wavelet decomposition algorithm is executed to create arrays of low-pass and high-pass filter coefficients, perform convolution operations on each signal sequence and perform downsampling by a factor of two to obtain approximation coefficients and detail coefficients. For the approximation coefficients, the above decomposition process is repeated, usually performing 3 to 5 layers of decomposition. Symmetric extension method is used for boundary processing to avoid signal edge distortion. After decomposition, the detail coefficients of each layer and the approximation coefficients of the final layer are obtained. These coefficients represent the characteristics of the fault interaction potential field at different time scales. In the condenser system, the first layer of detail coefficients corresponds to fast fluctuations of 5 - 10 seconds, the second layer corresponds to fluctuations of 10 - 20 seconds, and higher layers correspond to slower changes over longer time periods.

[0075] Specifically, calculating the energy distribution of the wavelet decomposition coefficients at different scales to identify the dominant scale of fault interaction. In implementation, the energy values are calculated for the detail coefficients of each layer and the approximation coefficients of the final layer respectively by summing the squares of the coefficient sequences. To eliminate the influence of signal length differences, the energy of each layer is divided by the total energy to obtain the normalized energy contribution rate. In specific operations, an array with the same length as the number of decomposition layers is created to store the energy of each layer. By looping through the coefficients of each layer and accumulating the squared values, and finally performing normalization. According to the calculated energy contribution rate, the threshold method is used to determine the main fault interaction scale, usually selecting the scale layer with a contribution rate exceeding 15%. In the condenser system, when the water level gauge is blocked and the control valve fails simultaneously, the energy contribution rate of the medium time scale (10 - 20 seconds) is often higher; while the cooling water pump failure is reflected in the energy of the longer time scale (30 - 60 seconds). This energy distribution characteristic helps to distinguish different types of multiple faults.

[0076] Specifically, the phase information of the main fault interaction time scale is extracted to analyze the timing characteristics of the fault signal. The implementation method is to perform a Hilbert transform on the wavelet coefficients of the main fault interaction scale. First, perform a fast Fourier transform (FFT), set the negative frequency part to zero, multiply the positive frequency part by 2, and then perform an inverse FFT to obtain the analytic signal. From the analytic signal, the phase sequence is obtained by calculating arctan(imaginary part / real part), and phase unwrapping is performed to obtain the continuous phase change. Subsequently, cross-scale phase coherence analysis is carried out. The phase difference is calculated for the phase sequences of any two components, and the average value and standard deviation of the phase difference are analyzed. The average value of the phase difference is divided by the center frequency to estimate the time delay, and the standard deviation reflects the stability of the phase relationship. Through this method, the propagation order and speed of the fault signal between different components are determined, and a causal relationship network of the fault is established, which is crucial for identifying the fault source and understanding the fault propagation mechanism.

[0077] Specifically, the fault interaction components are converted into a graph structure representation and a community detection algorithm is applied to map them into a fault interaction feature map. The implementation process first creates a graph data structure, such as an adjacency matrix, defines the nodes to represent system components, and the edges to represent the interaction relationships between components. Set the node attributes including ID, type, size (determined by the energy ratio), and color (indicating the fault type); the edge attributes include weight (determined by the energy correlation), direction (determined by the phase analysis), and delay value (calculated according to the phase difference). After the graph construction is completed, the Louvain community detection algorithm is applied to identify the closely connected node clusters. The algorithm first assigns each node to an independent community, and then iteratively moves the nodes to the neighbor community that can maximize the modularity gain until the modularity no longer increases. Finally, the community structure is integrated with the original graph to generate the fault interaction feature map, with different colors marking each community, the thickness of the edges representing the interaction intensity, and the direction arrow indicating the fault propagation path. In the condenser system, this map visually shows the fault interaction patterns among components such as the water level gauge, control valve, and cooling system, providing visual support for multiple fault diagnosis.

[0078] 104. Input the fault interaction feature map into a preset condition-aware conditional probability network and apply a hybrid inference framework to output the comprehensive diagnosis results of multiple faults.

[0079] In an embodiment of the present invention, the process of inputting the fault interaction feature map into a preset working condition perception conditional probability network and applying a hybrid inference framework to output a comprehensive diagnosis result for multiple faults includes: by monitoring key operating parameters of the target device, mapping the key operating parameters into a predefined working condition space in combination with a fuzzy clustering algorithm to determine the current working condition category; according to the determined current working condition category, calling the corresponding conditional probability network model from a pre-trained conditional probability network library, converting the fault interaction feature map into a network input feature vector, inputting it into the conditional probability network, and calculating the posterior probabilities of each fault type and the combination of fault types through forward inference; applying the Dempster-Shafer evidence theory framework to process the posterior probabilities, defining a basic probability assignment function for each fault type and combination, and calculating the uncertainty factor according to the operating state of the target device; using the basic probability assignment function and the uncertainty factor, calculating the comprehensive belief degree through the Dempster combination rule, and combining the interaction characteristics in the fault interaction feature map to generate a comprehensive diagnosis result.

[0080] Specifically, the process of determining the current working condition category by monitoring key operating parameters of the target device and mapping the key operating parameters into a predefined working condition space in combination with a fuzzy clustering algorithm starts from data collection. In the condenser liquid level control system, the key operating parameters include turbine load, condenser vacuum, inlet and outlet temperatures of cooling water, ambient temperature, and system start-stop state, etc. These parameters are collected in real time through a fieldbus network (such as PROFIBUS or Modbus), and the sampling period is usually 100 milliseconds to 1 second. The collected parameter data is preprocessed, including operations such as outlier removal, filtering, and standardization. Outlier removal uses the 3σ criterion, and data points outside the range of the mean ± 3 times the standard deviation are identified as outliers and replaced with the previous valid value. Filtering uses a low-pass filter to eliminate high-frequency noise, and the cut-off frequency is set to 1 Hz, which is sufficient to retain the main dynamic characteristics of the system while filtering out measurement noise. Standardization processing converts each parameter to the [0,1] interval for subsequent processing. The processed parameter data is used as the input of the fuzzy clustering algorithm, which uses the fuzzy C-means (FCM) method to softly assign data points in the multi-dimensional parameter space to predefined working condition categories. When implementing the FCM algorithm, the clustering centers are first initialized, usually selected based on typical working condition points in historical data; then the membership degrees of data points and the clustering centers are iteratively updated until the objective function converges or reaches the maximum number of iterations (such as 100 times). In the condenser system, the working condition space usually includes 5 to 8 categories such as full-load operation, part-load operation, start-up process, shutdown process, and rapid load change. After clustering, calculate the membership degrees of the current parameters to each working condition center, and take the category with the highest membership degree as the current working condition. This working condition identification method takes into account the fuzziness of parameters and the transitional characteristics between working conditions, and is suitable for judging the operating state of industrial systems.

[0081] Calling the corresponding condition-specific conditional probability network model from the pre-trained conditional probability network library according to the determined current condition category is the core link for realizing condition-aware diagnosis. Before system deployment, a dedicated conditional probability network model for each predefined condition is trained. These models adopt the Bayesian network structure and are learned from historical fault data. The training process includes two stages: structure learning and parameter learning. Structure learning uses the scoring search algorithm to determine the network topology, and the scoring function adopts the Bayesian Information Criterion (BIC) to balance the model complexity and fitting degree. Parameter learning uses the maximum likelihood estimation method to determine the values of the conditional probability table. The trained models are stored in the network library, and each model contains a node definition file, a structure file, and a parameter file. When the current condition is determined, the system loads the corresponding model from the library according to the condition ID. The loading process includes reading files, parsing the network structure, and initializing the conditional probability table. At the same time, the fault interaction feature map is converted into a network input feature vector. The conversion method is to extract key features from the map, such as the fault energy of the main nodes, the interaction intensity of the edges, and the community structure features. In the condenser system, the typical length of the feature vector is 20 to 50, which contains the main information of the fault map. The extracted feature vector is input into the conditional probability network, set as the observed evidence of the corresponding node, and then the forward inference of the network is executed. The inference process adopts the variable elimination algorithm or the junction tree algorithm, and the calculation process is bottom-up message passing, and finally the posterior probabilities of each fault type and combination are obtained. This condition-specific diagnosis method takes into account the influence of the system state on the fault characteristics and improves the diagnosis accuracy in the case of multiple faults.

[0082] Applying the Dempster-Shafer evidence theory framework to the posterior probability is a key step in solving diagnostic uncertainty. Different from traditional probability theory, the Dempster-Shafer theory allows the assignment of belief to a set of propositions rather than a single proposition, making it more suitable for dealing with situations of incomplete and partially conflicting evidence. The implementation process first defines the frame of discernment Θ, which includes all possible fault types and their combinations. In a condenser system, it usually includes 10 to 20 basic fault types, such as level gauge faults, control valve faults, etc. Then, the posterior probability output by the conditional probability network is converted into a basic probability assignment function (BPA). The conversion method is to determine the degree of belief of each fault set according to the probability value and threshold. In specific implementation, a series of thresholds (such as 0.1, 0.3, 0.5, 0.7) are set, and fault types with posterior probabilities falling in different intervals are assigned different basic probability values. At the same time, a part of the probability mass is assigned to the "unknown" state to represent the uncertainty of the evidence. Next, the uncertainty factor is calculated according to the operating state of the target device. This factor takes into account factors affecting diagnostic reliability such as operating duration, load change rate, and environmental interference. In a condenser system, the uncertainty factor is relatively high at the initial stage of startup or during rapid load changes, indicating that the reliability of the diagnostic results is relatively low at this time; while in the stable operation stage, the uncertainty factor is relatively low. The uncertainty factor is calculated using a weighted sum method, and the weights of each influencing factor are determined through expert experience or historical data analysis. This processing method based on evidence theory takes into account the uncertainty in the diagnostic process and provides richer information than a simple probability output.

[0083] Using the basic probability assignment function and the uncertainty factor, the combined belief degree is calculated through Dempster's combination rule, and combined with the interaction characteristics in the fault interaction feature map, generating the comprehensive diagnosis result is the final diagnosis step. Dempster's combination rule is used to fuse information from different evidence sources and is implemented in the system as pairwise calculation of the orthogonal sum and then normalization of the result. In the specific calculation process, the basic probability assignments of multiple feature dimensions (such as energy features, phase features, interaction features, etc.) in the system are regarded as independent evidence sources, and their combined results are calculated through nested loops. For the common 3 - 5 evidence sources in the condenser system, the time complexity of the combined calculation is proportional to the number of evidence sources and the size of the recognition frame, and it can usually be completed in milliseconds on modern processors. The calculated combined belief degree represents the comprehensive belief degree for each fault hypothesis, considering the support and conflict situations of all evidence. Subsequently, the combined belief degree is combined with the uncertainty factor to appropriately reduce the belief level for the diagnosis results under high - uncertainty working conditions. Finally, combined with the interaction characteristics in the fault interaction feature map, especially the community structure and propagation path information, the types, locations, severities, and causal relationships of multiple faults are determined. The generated comprehensive diagnosis result includes: the identification result of the main fault type and combination (such as "water level gauge blockage + control valve failure"); the severity assessment of each fault, represented by 0 - 100%; the confidence interval of the diagnosis result, reflecting the reliability range of the result; the prediction of the fault evolution trend, estimating the fault development direction and rate based on time - series analysis; and the causal relationship analysis of the fault, clarifying which are the root faults and which are the derivative faults.

[0084] In this embodiment, through multi - level phase - space feature extraction of multiple sensor signals in the control system of the target device, a feature time - series matrix is constructed; based on the feature time - series matrix, a functional dependence graph of the controller loop is constructed and the time - varying causal relationship is calculated, and the masking effect decoupling matrix is obtained by analyzing the abnormal flow distribution; combining the feature time - series matrix and the masking effect decoupling matrix to construct a fault interaction potential field, analyzing the fault interaction characteristics on different time scales, and generating a fault interaction feature map; inputting the fault interaction feature map into a preset condition - aware conditional probability network, and applying a hybrid inference framework to output the comprehensive diagnosis result of multiple faults. The present invention can effectively decouple the masking effect between multiple faults, accurately identify and diagnose various existing fault types, avoid triggering dangerous false alarms, and improve the safety and reliability of equipment operation.

[0085] The self - diagnosis method of the controller loop in the embodiment of the present invention has been described above. Next, the self - diagnosis device of the controller loop in the embodiment of the present invention will be described. The self - diagnosis device of the controller loop please refer to Figure 2 , an embodiment of the self - diagnosis device of the controller loop in the embodiment of the present invention includes:

[0086] The feature extraction module 201 is configured to hierarchically collect the direct thermal feature parameters, indirect thermal feature parameters, and coolant quality feature parameters of the air compressor of the fuel cell to obtain a multi-dimensional thermal data set of the air compressor;

[0087] The causal analysis module 202 is configured to extract features from the multi-dimensional thermal data set, and input the extracted feature data set into a neural network model based on deep transfer learning to obtain a corresponding heat dissipation efficiency decay prediction model;

[0088] The interaction recognition module 203 is configured to establish a multi-objective optimization equation of the air compressor based on the multi-dimensional thermal data set and the heat dissipation efficiency decay prediction model, and solve the multi-objective optimization equation to obtain a heat dissipation control strategy set;

[0089] The fault diagnosis module 204 is configured to dynamically partition the heat dissipation system of the air compressor, and coordinately adjust the flow rate and flow direction of the coolant in each region of the dynamic partition according to the heat dissipation control strategy set to achieve the heat dissipation of the air compressor.

[0090] In an embodiment of the present invention, the controller loop self-diagnosis device runs the above-mentioned controller loop self-diagnosis method. The controller loop self-diagnosis device extracts multi-level phase space features from multiple sensor signals in the target device control system to construct a feature time series matrix; constructs a functional dependence graph of the controller loop based on the feature time series matrix and calculates the time-varying causal relationship, and obtains a masking effect decoupling matrix by analyzing the abnormal flow distribution; combines the feature time series matrix and the masking effect decoupling matrix to construct a fault interaction potential field, analyzes the fault interaction characteristics on different time scales, and generates a fault interaction feature map; inputs the fault interaction feature map into a preset working condition perception conditional probability network, and applies a hybrid reasoning framework to output a comprehensive diagnosis result of multiple faults. The present invention can effectively decouple the masking effect between multiple faults, accurately identify and diagnose various existing fault types, avoid the triggering of dangerous false alarms, and improve the safety and reliability of equipment operation.

[0091] Above Figure 2 The controller loop self-diagnosis device in the embodiment of the present invention is described in detail from the perspective of modular functional entities. Below, the controller loop self-diagnosis device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0092] Figure 3FIG. 0 is a schematic structural diagram of a controller loop self-diagnosis device provided by an embodiment of the present invention. The controller loop self-diagnosis device 300 may vary greatly due to different configurations or performances, and may include one or more processors (central processing units, CPUs) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 for storing application programs 333 or data 332 (for example, one or more mass storage device terminals). Among them, the memory 320 and the storage media 330 may be transient storage or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations on the controller loop self-diagnosis device 300. Further, the processor 310 may be configured to communicate with the storage media 330 and execute a series of instruction operations in the storage media 330 on the controller loop self-diagnosis device 300 to implement the steps of the above-mentioned controller loop self-diagnosis method.

[0093] The controller loop self-diagnosis device 300 may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, and so on. Those skilled in the art can understand that Figure 3 the shown structural diagram of the controller loop self-diagnosis device does not limit the controller loop self-diagnosis device provided by the present invention, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0094] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described system or device and unit can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0095] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0096] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A controller loop self-diagnosis method, characterized in that: The controller loop self-diagnosis method comprises: Reconstructing the phase space by applying the delayed coordinate method to the time series signals of multiple sensors in the control system of the target device to obtain a high-dimensional phase space trajectory; calculating the nonlinear dynamic index for the high-dimensional phase space trajectory, and using the adaptive sliding window technology to perform segmented calculation on the nonlinear dynamic index to obtain a time-varying nonlinear dynamic index sequence; using the high-dimensional phase space trajectory, the nonlinear dynamic index and the time-varying nonlinear dynamic index sequence as phase space features, and constructing a feature time series matrix based on the extracted phase space features; Based on the characteristic timing matrix, a functional dependency graph of the controller loop of the target device is constructed and the time-varying causal relationship between the components in the functional dependency graph is calculated, and a masking effect decoupling matrix is ​​obtained by analyzing the abnormal traffic distribution in the time-varying causal relationship; Based on the characteristic timing matrix and the masking effect decoupling matrix, a fault interaction potential field is constructed and the fault interaction characteristics of the fault interaction potential field at different time scales are analyzed to generate a fault interaction characteristic map; The fault interaction feature map is input into a preset working condition perception conditional probability network, and a hybrid reasoning framework is applied to output a comprehensive diagnosis result of multiple faults.

2. The controller circuit self-diagnosis method according to claim 1, characterized in that: Based on the characteristic timing matrix, a functional dependency graph of the controller loop of the target device is constructed and the time-varying causal relationship between the components in the functional dependency graph is calculated. By analyzing the abnormal traffic distribution in the time-varying causal relationship, a masking effect decoupling matrix is ​​obtained, including: Analyze the phase space feature correlation of each sensor in the feature timing matrix, and construct an initial functional dependency graph by combining the physical structure of the target device and the controller loop schematic diagram; Based on the time-varying sequence of nonlinear dynamic indicators in the characteristic time series matrix, the improved PC algorithm of the conditional mutual information criterion is used to determine the direction and strength of the causal association between components, and the initial functional dependency graph is optimized based on the direction and strength of the causal association; Based on the optimized functional dependency graph, the characteristic time series matrix is ​​divided into multiple time periods, and the time-varying Granger causality analysis method is applied to each time period, a vector autoregression model is established and statistics are calculated to obtain the time-varying causal relationship that characterizes the time-varying changes between the components; The causal strength in the time-varying causal relationship is regarded as the abnormal traffic distribution. According to the abnormal traffic conservation principle, the abnormal absorption rate is calculated for each component to construct the masking effect decoupling matrix.

3. The controller circuit self-diagnosis method according to claim 2, characterized in that: The causal strength in the time-varying causal relationship is regarded as abnormal traffic distribution, and according to the abnormal traffic conservation principle, the abnormal absorption rate is calculated for each component to construct the masking effect decoupling matrix, which includes: A directed weighted graph structure is established for the time-varying causal relationship, wherein nodes represent components of a target device, directions of edges represent directions of causal relationships, and weights of edges represent causal strength, wherein the causal strength is defined as abnormal traffic distribution between components; For each node in the directed weighted graph, the sum of abnormal flow distribution of the inflow node and the sum of abnormal flow distribution of the outflow node are calculated to obtain the abnormal absorption rate of each node; A masking effect decoupling matrix is ​​constructed according to the abnormal absorption rate and the causal strength, rows and columns of the masking effect decoupling matrix correspond to components, and matrix element values ​​are determined according to the abnormal absorption rate of the node and the relative causal strength between two nodes.

4. The controller circuit self-diagnosis method according to claim 1, characterized in that: The constructing of a fault interaction potential field based on the characteristic timing matrix and the masking effect decoupling matrix and analyzing the fault interaction characteristics of the fault interaction potential field at different time scales to generate a fault interaction feature map comprises: Reorganize the feature time series matrix into a third-order tensor according to the three dimensions of the number of sensors, the number of time windows, and the feature dimension, and apply the constrained Tucker decomposition algorithm to the third-order tensor to extract the spatial pattern matrix representing the sensor spatial pattern and the core tensor representing the pattern interaction; Calculating the Hadamard product of the spatial mode matrix and the masking effect decoupling matrix, wherein the Hadamard product is used to adjust the influence weights of each component in the spatial mode matrix, combining the adjusted spatial mode matrix with the core tensor, and defining a fault interaction potential field; Acquire thermodynamic parameter measurement values ​​of the target device, calculate a thermodynamic compensation factor according to a thermodynamic model of the device, and perform compensation adjustment on the fault interaction potential field according to the thermodynamic compensation factor to obtain a compensated fault interaction potential field; The compensated fault interaction potential field is subjected to multi-scale decomposition by applying wavelet transform, fault interaction components of each scale are extracted, and the fault interaction components are mapped into a fault interaction feature spectrum.

5. The controller circuit self-diagnosis method according to claim 4, characterized in that: The applying wavelet transform to the compensated fault interaction potential field to perform multi-scale decomposition, extracting fault interaction components of each scale, and mapping the fault interaction components into a fault interaction feature map comprises: According to a preset wavelet basis function, a discrete wavelet transform is applied to the compensated fault interaction potential field to decompose the fault interaction potential field into wavelet decomposition coefficients of different scales; Calculating energy distribution of wavelet decomposition coefficients of different scales to obtain energy contribution rate of each scale layer, and determining the time scale of occurrence of major fault interactions according to the energy contribution rate; Extracting the phase information of the time scale at which the main fault interaction occurs, and identifying the time delay characteristics and synchronization characteristics of the fault interaction of different components through cross-scale phase coherence analysis; The energy distribution, phase information, time delay characteristics and synchronization characteristics are converted into a graph structure representation as fault interaction components, and a community detection algorithm is applied to the graph structure representation to map it into a fault interaction feature spectrum.

6. The controller circuit self-diagnosis method according to claim 1, characterized in that: The inputting of the fault interaction feature map into a preset working condition perception conditional probability network and applying a hybrid reasoning framework to output a comprehensive diagnosis result of multiple faults includes: By monitoring the key operating parameters of the target equipment and combining the fuzzy clustering algorithm to map the key operating parameters to the predefined operating condition space, the current operating condition category is determined; According to the determined current working condition category, the corresponding conditional probability network model is called from the pre-trained conditional probability network library, and the fault interaction feature map is converted into a network input feature vector, which is input into the conditional probability network, and the posterior probability of each fault type and the combination of fault types is calculated through forward reasoning; Applying the Demster-Schaffer evidence theory framework to the posterior probability, defining a basic probability distribution function for each fault type and combination, and calculating an uncertainty factor based on the operating status of the target device; The basic probability distribution function and the uncertainty factor are used to calculate the comprehensive trust through the Demster combination rule, and the comprehensive diagnosis result is generated by combining the interaction characteristics in the fault interaction feature map.

7. A controller circuit self-diagnosis device, characterized in that: The controller loop self-diagnosis device comprises: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instructions in the memory to enable the controller loop self-diagnosis device to perform the steps of the controller loop self-diagnosis method according to any one of claims 1 to 6.

Images