Alarm AI application system based on industrial data

Abstract

The invention relates to the technical field of industrial process monitoring and fault diagnosis, and discloses an alarm AI application system based on industrial data, and the system comprises a data access and preprocessing module which is used for collecting and preprocessing the industrial data so as to generate a standardized state vector sequence; the offline model construction module is used for performing offline training and solidifying a physical channel model and a data channel model for subsequent diagnosis; the real-time diagnosis and alarm module is used for loading the model, calculating physical and mode residual errors in parallel and executing root cause diagnosis so as to generate a quantized causal score; and the diagnosis result output module is used for triggering an alarm according to the residual error and the causal score and outputting an explanatory report containing root cause positioning. According to the method, the two-channel residual error model is constructed, and the dynamic Jacobian matrix is utilized to drive the dynamic causal graph network, so that the alarm accuracy and the root cause diagnosis reliability are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of industrial process monitoring and fault diagnosis technology, specifically to an alarm AI application system based on industrial data. Background Technology

[0002] In modern process industries, such as petrochemicals, power generation, and advanced manufacturing, systems are large-scale, processes are complex, and units are tightly coupled. To ensure the safety and stability of production processes, improve product quality, and optimize operational efficiency, real-time and accurate detection of abnormal states in industrial processes and rapid identification of their root causes are of paramount technical and economic value.

[0003] In existing technologies, to improve the performance of alarm systems, various data-driven multivariate statistical process monitoring and machine learning methods have been developed, moving beyond traditional single-variable fixed-threshold alarm methods. One type of method uses multivariate statistical models such as Principal Component Analysis (PCA) or Partial Least Squares (PLS) to construct a low-dimensional statistical space under normal operating conditions and detect process anomalies based on indicators such as the sum of squared prediction error (SPE). Another type of method utilizes deep learning models such as Long Short-Term Memory (LSTM) networks or Gated Recurrent Units (GRUs) to construct autoencoders. These autoencoders learn the complex spatiotemporal correlation patterns of normal operating data in a high-dimensional space and use the magnitude of the reconstruction error as the basis for anomaly judgment. For root cause diagnosis, methods based on Bayesian networks or static causal graphs are commonly used. These methods trace the most likely source variable causing the anomaly by reasoning on a predefined graph structure.

[0004] While existing technologies have improved the level of intelligence in industrial process monitoring to some extent, some shortcomings still exist: The adaptability and accuracy of existing alarm systems need improvement. On the one hand, solutions relying on multivariate statistical models are based on linear and Gaussian distribution assumptions. However, real industrial processes generally exhibit strong nonlinearity and non-Gaussian properties, which limits the model's ability to describe complex process dynamics. When process conditions drift normally, the fixed statistical model will generate a large number of false alarms because it cannot adapt to the new conditions. On the other hand, solutions that rely purely on deep learning data reconstruction essentially learn the distribution patterns of historical data, lacking consideration of the system's inherent physical laws. This characteristic means that when faced with new conditions that have not appeared in history but conform to physical operating laws, the model may misjudge them as abnormal due to the unknown pattern; conversely, there is a risk of missing alarms for some minor early faults that violate physical dynamic laws but have unclear statistical characteristics.

[0005] Furthermore, existing root cause diagnostic methods lack accuracy and interpretability. Currently used diagnostic methods largely rely on static cause-effect diagrams built from process flow diagrams or historical data mining. However, the interrelationships (i.e., causal strength) between variables in industrial systems are not static but dynamically change with operating conditions such as load, material composition, and ambient temperature. Static diagram structures cannot capture these dynamic causal relationships that change with operating conditions; they only represent fixed, averaged influence paths between variables. Therefore, when an anomaly occurs, tracing the fault along this static path may not reflect the actual propagation path of the anomaly under specific operating conditions, leading to biased root cause localization and reduced reliability of diagnostic conclusions. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an alarm AI application system based on industrial data, which solves the problems of low alarm accuracy and inaccurate root cause location caused by the inability of existing technologies to adapt to dynamic changes in working conditions and the reliance on static causal relationships for diagnosis.

[0007] To achieve the above objectives, the present invention provides an alarm AI application system based on industrial data, the system comprising: The data access and preprocessing module is used to collect multidimensional time series data from industrial data sources and generate standardized state vector sequences. The offline model building module is used to train and solidify physical channel models and data channel models offline based on historical data; The real-time diagnosis and alarm module is used to receive the standardized state vector sequence and load the physical channel model and data channel model to generate physical residual vector and pattern residual vector in parallel. At the same time, it performs root cause diagnosis based on the physical residual vector and pattern residual vector to generate a quantified causal score. The diagnostic result output module is used to generate alarm events based on the physical residual vector, the pattern residual vector, and the causal score, and output an explanatory report containing the root cause location.

[0008] Preferably, the real-time diagnosis and alarm module includes a physical channel residual calculation unit, which is used for: Using the physical channel model, the dynamic Jacobian matrix is ​​estimated in real time based on the most recent state vector sequence; Using the dynamic Jacobian matrix, the state vector sequence before the current time, and the recent state change trend, the physical prediction value of the current state is calculated; The physical residual vector is obtained by calculating the vector difference between the actual observed state vector sequence and the physical predicted state value.

[0009] Preferably, the real-time diagnosis and alarm module includes a data channel residual calculation unit, which is used for: Using the data channel model, the real-time input state vector sequence is reconstructed to obtain a state data pattern reconstruction vector. The data channel model is a deep autoencoder model trained based on historical normal working condition data. The pattern residual vector is obtained by calculating the absolute value of the difference between the actual observed state vector sequence and the reconstructed state data pattern vector.

[0010] Preferably, the real-time diagnosis and alarm module is further used for: Each component of the physical residual vector is concatenated with the corresponding component of the weighted pattern residual vector to generate a hybrid residual initial feature vector for each node in the system.

[0011] Preferably, the real-time diagnosis and alarm module includes a dynamic causal diagnosis unit, which is used for: The initial feature vectors of the hybrid residuals generated for each node in the system are combined into an initial node feature matrix, which is then used as input to the dynamic causal graph network. In the dynamic causal graph network, the dynamic Jacobian matrix estimated in real time is used as the weighted adjacency matrix for information propagation between nodes. The initial node feature matrix is ​​updated through multi-layer information propagation to obtain the final node feature matrix. The final node feature matrix is ​​processed by a readout function to calculate the causal score characterizing the likelihood of each node as the root cause of an anomaly.

[0012] Preferably, the offline model building module is used for: A recurrent neural network is trained as the physical channel model to achieve a nonlinear mapping from a sequence of state vectors to a dynamic Jacobian matrix; An autoencoder based on a recurrent neural network is trained as the data channel model to learn and reproduce high-dimensional data association patterns under normal operating conditions.

[0013] Preferably, the diagnostic result output module includes an alarm generation and management unit, used to trigger the alarm event when the physical residual magnitude of any node exceeds a first preset threshold or the pattern residual magnitude exceeds a second preset threshold, or when the causal score of any node exceeds a global score threshold.

[0014] Preferably, the alarm generation and management unit is further configured to adaptively adjust the first preset threshold and the second preset threshold based on the statistical distribution of normal data residuals observed within a sliding time window.

[0015] Preferably, the diagnostic result output module further includes: an interpretive report generation unit, used to generate a diagnostic report containing the following content when the alarm event is triggered: The node with the highest causal score is highlighted as the root cause of the anomaly; Specify whether the residual type that initially triggered the alarm was a physical residual overrun or a mode residual overrun.

[0016] Preferably, the explanatory report generation unit is also used to provide a visual representation of the fault propagation path by performing the following operations: On the system topology graph, starting from the affected node, the path is traced back and highlighted to the root cause node along the connection with a large weight defined by the dynamic Jacobian matrix at the alarm time.

[0017] This invention provides an alarm AI application system based on industrial data, which has the following beneficial effects: This invention improves the accuracy and adaptability of alarms by constructing a dual-channel residual generation mechanism that combines physical and data channels. The physical channel model uses a dynamic Jacobian matrix to predict the physical dynamic behavior of the system, effectively identifying and adapting to normal operating condition changes that conform to physical laws, thus reducing false alarms. The data channel model, on the other hand, captures complex nonlinear patterns in historical data through deep learning, enabling the detection of early anomalies with weak statistical characteristics. The combination of the two residuals provides a more comprehensive basis for anomaly judgment, thereby achieving more sensitive and reliable detection of various faults.

[0018] This invention introduces a dynamic causal graph network, which directly uses the dynamic Jacobian matrix estimated in real time by the physical channel as the weighted adjacency matrix for graph information propagation. This design ensures that fault tracing analysis strictly follows the real physical influence path and intensity of the system under the current operating conditions, thereby improving the accuracy of root cause diagnosis.

[0019] In this invention, the system can not only clearly identify the root cause node, but also visualize the abnormal propagation path traced based on the dynamic Jacobian matrix on the process topology diagram through the diagnostic result output module. This visualization method clearly shows the evolution process of the fault from the source to the affected point, which improves the operator's trust in the system and helps them to quickly formulate countermeasures. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the internal structure of the data access and preprocessing module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the internal structure of the offline model building module in an embodiment of the present invention; Figure 5 This is a schematic diagram of the internal structure of the real-time diagnosis and alarm module according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the internal structure of the diagnostic result output module in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 As shown, Figure 1 This is a schematic diagram of a system structure according to an embodiment of the present invention. The embodiment of the present invention provides an alarm AI application system based on industrial data, which may include: The data access and preprocessing module 100 is used to collect multidimensional time series data from various data sources in the industrial field, and perform data cleaning, time alignment and standardization processing to generate standard format data for use by subsequent modules.

[0023] The offline model building module 200 is used to complete the offline training and parameter solidification of the system's core algorithm model using historical datasets before system deployment, including the construction of physical channel models and data channel models.

[0024] The real-time diagnosis and alarm module 300 is used to receive real-time data streams, load offline trained models, perform parallel computation of dual-channel residuals, fusion of hybrid residuals, and root cause diagnosis based on dynamic causal graph networks when the system is running online.

[0025] The diagnostic result output module 400 is used to convert the quantitative analysis results generated by the real-time diagnostic and alarm module 300 into a visual diagnostic report for operators, which includes clear alarm signals, root cause location, and fault propagation path.

[0026] like Figure 2 As shown, Figure 2 This is a schematic diagram of a method flow according to an embodiment of the present invention. The present invention provides a method for an alarm AI application system based on industrial data, comprising the following steps: S1 performs data acquisition and preprocessing, obtaining raw multidimensional time series data from industrial data sources, and completing data cleaning, alignment and standardization to generate a standardized state vector sequence.

[0027] S2 executes parallel dual-channel model construction, using historical normal operating condition data to train offline the dynamic Jacobian matrix estimation model of the physical channel and the spatiotemporal pattern reconstruction model of the data channel.

[0028] S3 executes real-time dual-channel residual generation. When the system is running online, the real-time state vector sequence is input in parallel into the two pre-trained channel models to calculate the physical residual vector and the mode residual vector respectively.

[0029] S4 performs hybrid residual fusion and root cause diagnosis, fusing the two residual vectors into a hybrid residual feature, and using it as the initial node state of the dynamic causal graph network. The causal score of each node is calculated through network propagation.

[0030] S5 executes interpretive alarms and diagnostic results output, triggers alarms based on residual values ​​and causal scores, and generates a structured diagnostic report containing root cause location and abnormal propagation path.

[0031] The embodiments of the present invention are as follows: By constructing two parallel analysis channels—physical and data—the degree of deviation from the system state can be quantified from different dimensions.

[0032] In the physical channel, the system uses a trained model to estimate the dynamic Jacobian matrix, which characterizes the instantaneous influence relationships between system variables, based on the most recent system state sequence. This dynamic Jacobian matrix, combined with the state at the previous moment and the recent state change trend, makes a physical prediction of the system state at the current moment. The physical residual is the vector difference between the actual observed state and the physical prediction state, and its components directly quantify the degree of deviation of the corresponding measurement point in the physical dynamic behavior.

[0033] In the data channel, the system uses a deep autoencoder model trained on historical normal data to reconstruct the real-time system state. The mode residual is the absolute value of the difference between the real state and the model reconstructed state. Each component quantifies the degree to which the corresponding measurement point deviates from the learned high-dimensional nonlinear data pattern.

[0034] Then, in this embodiment of the invention, the physical residual and the weighted mode residual are concatenated to form a two-dimensional hybrid residual feature vector, which serves as the initial state of each node in the graph.

[0035] Furthermore, through the information propagation mechanism of the dynamic causal graph network, the weight of the propagation and aggregation process of node features in this dynamic causal graph network is not determined by the static adjacency matrix, but is directly driven by the dynamic Jacobian matrix generated in real time by the physical channel. This design forces the initial mixed residual signal to propagate along the real physical influence path under the current operating conditions of the system in the dynamic causal graph network. The calculation of each layer of the network simulates the process of abnormal signals being transmitted and converged according to the actual influence intensity between variables.

[0036] After multiple layers of propagation, the system uses a readout function to transform the final aggregated feature information from each node into a scalar causal score. The node with the highest score is identified as the root cause of the system's cascading anomalies. Through this systematic and structured approach, this embodiment of the invention achieves accurate and interpretable alarms and diagnoses of industrial process anomalies.

[0037] like Figure 3 As shown, Figure 3 This is a schematic diagram of the internal structure of a data access and preprocessing module according to an embodiment of the present invention. The data access and preprocessing module 100 in this embodiment of the invention acts as a bridge between the system and the industrial field data source, transforming raw, heterogeneous industrial data into a structured, standardized data stream for use by subsequent analysis modules. Specifically, the data access and preprocessing module 100 may include: a data acquisition unit 110, a data cleaning and alignment unit 120, and a data standardization unit 130.

[0038] The data acquisition unit 110 is used to establish a stable and reliable communication connection with one or more industrial data sources. In a specific implementation, the data acquisition unit 110 establishes a communication connection directly with the real-time or historical database of a distributed control system (DCS) or supervisory control and data acquisition (SCADA) system by configuring a standard industrial communication protocol interface, such as OPC Unified Architecture (OPCUA), Modbus TCP / IP, or other communication protocols suitable for industrial sites.

[0039] The data acquisition unit 110 periodically or based on an event-triggered list of preset variable tags to acquire raw data records containing timestamps, variable tags, and variable values. These acquired data records form the basis for all subsequent analyses. The acquired raw data records are then transferred to the data cleaning and alignment unit 120 for further processing.

[0040] The data cleaning and alignment unit 120 is used to handle quality issues in the raw data records received from the data acquisition unit 110. Due to network latency, sensor malfunction, or maintenance, the raw data records may contain missing or invalid values. For missing values, the data cleaning and alignment unit 120 can be configured to employ various imputation strategies, such as linear interpolation, nearest neighbor interpolation, or forward / backward imputation. Linear interpolation is suitable for continuously changing variables with gradual changes, while forward imputation is suitable for step variables that remain constant over a period of time.

[0041] In industrial settings, the sampling periods or data upload times of different sensors may vary, resulting in inconsistent timestamps at the same logical point in time. To address this issue, the data cleaning and alignment unit 120 establishes a unified, fixed time grid, for example, one sampling point per second. Subsequently, the data cleaning and alignment unit 120 resamples the data sequences of all variables onto this time grid. During the resampling process, methods such as taking the nearest value or time-weighted averaging can be used to generate data frames where all variables have valid values ​​at each unified timestamp. The processed data frame sequence is then passed to the data normalization unit 130.

[0042] The data standardization unit 130 is used to eliminate the influence of different process variables due to differences in physical units and numerical ranges, so as to improve the convergence and stability of subsequent model training.

[0043] Finally, the data normalization unit 130 independently performs Z-score normalization processing on each variable in the data frame sequence. Specifically, for the first variable... Variables in Observations at time Its standardized value Calculated using the following formula: ; In the formula, For the first Variables in The value obtained after Z-score standardization at each time step; For the first Variables in The original observation value at time 1, which is the value after data cleaning and alignment steps; To utilize historical normal operating condition datasets to pre-calculate and solidify data belonging to the first... The mean of each variable; To utilize historical normal operating condition datasets to pre-calculate and solidify data belonging to the first... The standard deviation of each variable.

[0044] After this step, the module outputs the standardized state vector sequence, where each component approximately follows a distribution with a mean of 0 and a standard deviation of 1.

[0045] Figure 4 This is a schematic diagram of the internal structure of an offline model building module according to an embodiment of the present invention. Figure 4 As shown, the offline model building module 200 in this embodiment of the invention completes the training and construction of all analysis models using historical datasets before the system is formally deployed, and solidifies the trained model parameters for use in the online diagnostic phase. Specifically, the offline model building module 200 may include: a physical topology modeling unit 210, a physical channel model training unit 220, and a data channel model training unit 230.

[0046] The physical topology modeling unit 210 is used to abstract actual industrial production processes into mathematically tractable graph structures. The inputs to this physical topology modeling unit 210 are industrial process and instrumentation diagrams, equipment layout diagrams, and relevant domain expert knowledge, such as piping and instrumentation diagrams (P&ID).

[0047] In the specific implementation of the physical topology modeling unit 210, the operator or automated script defines the key monitoring points (e.g., temperature, pressure, flow sensors), control units (e.g., valves, pumps), and core equipment (e.g., reaction vessels, separation towers) in the system as a set of nodes in the graph. ,in This represents the total number of nodes. This is the set of all nodes in the system. Then, based on the physical material flow direction, energy transfer path, or logical relationship of control signals, edges are established between the corresponding node pairs to form an edge set. For example, if material flows from device A to device B, a directed edge is created between the nodes representing A and B.

[0048] The output of the physical topology modeling unit 210 is a static topology graph. This graph structure itself does not contain dynamic information, but it provides the basic node definitions and connection framework for the subsequent construction of dynamic causal graph networks.

[0049] The physical channel model training unit 220 is used to train a neural network model that can estimate the local dynamic characteristics of the system in real time based on the current operating state of the system. These dynamic characteristics are characterized by the dynamic Jacobian matrix of the system.

[0050] In this embodiment, the physical channel model training unit 220 employs a recurrent neural network (RNN) as the estimation model for the Jacobian matrix, specifically a long short-term memory network (LSTM) or a gated recurrent unit (GRU). These types of models excel at handling and remembering long-term dependencies in time-series data.

[0051] The estimation model is denoted as... The input is the system state sequence within the most recent time window, and the output is the estimated value of the dynamic Jacobian matrix at the current moment. The principle is that the local dynamics of the system (i.e., the interactions between variables) are not static, but change with the system's operating conditions (e.g., load, catalyst activity). By learning from historical data, a nonlinear mapping relationship is established from a state sequence representing the current operating conditions to its corresponding local dynamics (Jacobi matrix).

[0052] The physical channel model training unit 220 utilizes a time-series dataset under historical normal operating conditions to train the model through self-supervised learning. parameters Optimization is performed. The training process is as follows: First, the estimation model Based on the input state sequence Calculate the time Jacobian matrix estimate : ; In the formula, For at any time The estimated system dynamic Jacobian matrix is ​​a × A 3D matrix whose elements reflect the instantaneous influence relationships between variables; The current moment; For a parameterized model used to estimate the Jacobian matrix, such as a neural network, the trainable parameters are: ; For the model The set of trainable parameters; A time series window, containing data from time 1 to 2. At the time Continuous A sequence of state vectors; This represents the length of the time series window.

[0053] The prediction is based on time. The current state and recent trends in state changes: ; In the formula, For physical channel model time System status The prediction vector represents the desired state of the system under the constraints of physical laws; The current moment; For at any time Representing the true state of the system 3D standardized observation vector; For the physical channel model at time Estimated × A dynamic Jacobian matrix of dimension 1, whose elements represent the instantaneous influence of a change in one variable on another under the current operating conditions; This is the multiplication operator for matrices and vectors; For at any time Representing the true state of the system Dimensionally standardized observation vector.

[0054] Finally, the model parameters are updated by minimizing the mean squared error between the predicted and actual values. Loss function Defined as: ; In the formula, This represents the loss function value of the physical channel model; This is a physical channel model, specifically a dynamic Jacobian matrix estimation model. The set of all trainable parameters contained therein; The total time step of the historical data samples used for model training is a positive integer. The current moment; For at any time Representing the true state of the system 3D standardized observation vector; For physical channel model time System status dimensional prediction vector; This is an operator that calculates the square of the L2 norm (i.e., Euclidean distance) of the vector difference, used to quantify the difference between the true and predicted values.

[0055] The data channel model training unit 230 is used to train a deep learning model, the purpose of which is to learn and reproduce the high-dimensional data association patterns between multidimensional data under normal working conditions.

[0056] High dimension refers to the large number of monitoring variables in an industrial process, which in this embodiment means N nodes. At any given time, the state vector composed of the standardized observations of all N nodes is a data point in the high-dimensional data space.

[0057] A correlation pattern refers to the implicit and complex interrelationships that these high-dimensional data points follow under normal system operation. This relationship is not merely a simple linear correlation, but encompasses the following two levels of meaning: Spatial correlation patterns between variables refer to the inherent constraints and synergistic changes between different variable components at the same time. For example, in a certain chemical unit, an increase in feed flow rate is inevitably accompanied by a specific trend of change in temperature and pressure inside the reactor. This coupling relationship between multiple variables, determined by the process mechanism, constitutes a static data pattern.

[0058] The time-related patterns of variables themselves and between them refer to the dynamic behavior and dependencies exhibited by one or more variables in a time series. This includes time-series response curves under specific operations, periodic fluctuations, and delay relationships between different variables. For example, after the opening degree of valve A changes, the downstream flow B always starts to respond after a delay of several time steps. This time-series dependency is a dynamic data pattern.

[0059] Therefore, the high-dimensional data association pattern is a comprehensive embodiment of the above spatial and temporal association patterns, and it constitutes a holistic data behavior pattern that can uniquely represent the normal operating state of the system.

[0060] In this embodiment, the data channel model training unit 230 adopts an autoencoder architecture based on a recurrent neural network. The model consists of an encoder. and a decoder It is configured in series, where both the encoder and decoder can be implemented by multi-layer LSTM or GRU networks.

[0061] The working principle of the data channel model training unit 230 is as follows: Encoder The input time series window is compressed into a low-dimensional, fixed-length sequence of hidden state vectors that capture the core information of the input sequence. Decoder The system then attempts to reconstruct the original input time series using the hidden state vector sequence. Since this offline model building module 200 is trained only on normal operating condition data, it learns to accurately reconstruct normal data patterns. When the input data deviates from these learned normal patterns, the model will be unable to reconstruct effectively, resulting in a large reconstruction error.

[0062] The data channel model training unit 230 uses historical normal operating condition data to train model parameters by minimizing the difference between the input and the reconstructed output. and .

[0063] For the input time series window The model first calculates its reconstructed output. : ; In the formula, The output sequence is the reconstruction of the input time series window by the data channel model; This is the decoder part in the data channel model, whose function is to reconstruct the data sequence based on a sequence of hidden state vectors; For decoder The set of all trainable parameters included; This is the encoder part in the data channel model, whose function is to compress the input data sequence into a hidden state vector sequence; For encoder The set of all trainable parameters included; From time At the time The standardized observation sequence, representing the true state of the system, is used as the input to the model; The current moment; The length of the time series window is a positive integer that defines the length of the sequence processed by the model at one time.

[0064] training loss function Defined as the average reconstruction error over the entire sequence window and all time steps: ; In the formula, The loss function value for the data channel model is a scalar whose magnitude measures the model's reconstruction error of the input sequence. and These are the sets of trainable parameters for the encoder and decoder, respectively. This represents the total number of sample windows used for model training. The length of the time series window; This is the index of the sample window; This is the index of the time step within the window; In the first Within the sample window, the first The actual state observation vector at each point in time; For the model pair The reconstructed output vector.

[0065] Please see Figure 5 , Figure 5 This is a schematic diagram of the internal structure of a real-time diagnosis and alarm module according to an embodiment of the present invention. The real-time diagnosis and alarm module 300 in this embodiment is used to analyze real-time data streams and perform anomaly detection and root cause diagnosis. Specifically, the real-time diagnosis and alarm module 300 may include: a dual-channel residual generation unit 310, a hybrid residual fusion unit 320, and a dynamic causal diagnosis unit 330.

[0066] The dual-channel residual generation unit 310 is used to receive the real-time standardized state vector sequence from the data access and preprocessing module 100, load the model parameters trained by the offline model building module 200, and calculate the physical residual and mode residual in parallel.

[0067] When performing calculations, the physical channel residual calculation unit first defines the current moment for diagnosis. This unit will extract a value at the current moment. Previously, with a preset length The standardized state vector sequence within the time window is used as the model input. Subsequently, the physical channel model uses this sequence to estimate the current time. The dynamic Jacobian matrix.

[0068] When calculating physical residuals, the benchmark used for comparison is the current time. The standardized state vector sequence generated in real time by the data access and preprocessing module 100 This vector is the sequence of actual observed state vectors. Ultimately, the physical residual vector is defined as... The difference between the physical state predictions calculated by the model and the actual state predictions.

[0069] Subsequently, using the Jacobian matrix and historical states, the current state physics prediction is calculated. Finally, the physical residual vector is obtained by calculating the vector difference between the actual observed values ​​and the physical prediction values. : ; In the formula, For a moment The physical residual vector is a vector containing A column vector of components, each component corresponding to the physical deviation of a system node; For a moment The sequence of real observed state vectors is a collection of... The column vector of each component is provided by the data access and preprocessing module 100; For a moment The state-physics prediction vector is a vector containing The column vectors of each component are calculated from the physical channel model; The current moment; This represents the total number of nodes monitored in the system.

[0070] The dual-channel residual generation unit 310 generates the real-time state vector sequence in parallel. Input to the pre-written encoder and decoder In the constructed autoencoder model, the reconstructed values ​​of the data are obtained. : ; In the formula, For a moment The state data pattern reconstruction vector is a vector containing The column vector of each component is calculated from the data channel model; For the decoder part of the data channel model, its set of trainable parameters is: ; For the encoder part of the data channel model, its trainable parameter set is as follows: .

[0071] Then, by calculating the absolute value of the difference between the true observations and the reconstructed values, the mode residual vector is obtained. : ; In the formula, For a moment The mode residual vector is a vector containing A column vector of components, each component corresponding to the degree of deviation of the data pattern of a system node; For a moment The state data pattern reconstruction vector is a vector containing The column vector of each component is calculated from the data channel model; For a moment The actual observed state vector sequence; This represents the operation of taking the absolute value of each element.

[0072] The magnitude of each component of the residual vector quantifies the degree to which the corresponding measuring point deviates from the complex spatiotemporal data pattern it exhibits under historical normal operating conditions.

[0073] The hybrid residual fusion unit 320 is used to integrate the two heterogeneous residual signals from the dual-channel residual generation unit 310 to provide a comprehensive initial feature containing multi-dimensional anomaly information for subsequent causal diagnosis.

[0074] In this embodiment, for the first in the system topology diagram Nodes The hybrid residual fusion unit 320 fuses its corresponding physical residual components. Compared with weighted mode residual components The features are concatenated to form a two-dimensional initial feature vector. : ; In the formula, For nodes At any moment The two-dimensional initial feature vector is used as the input to the dynamic causal diagnostic unit, with superscript... This indicates that this is the 0th layer (input layer) of the graph network; The index of the node, with values ​​ranging from 1 to... ; For a moment node The magnitude of the physical residual; For a moment node The residual amplitude of the model; A preset scalar fusion weight is used to balance the nodes. The dimensions or importance of the two residuals; This represents a vector concatenation operation, which concatenates two scalars into a single two-dimensional vector.

[0075] All The initial feature vectors of each node are stacked to form a... The initial feature matrix in 2×2 dimensions It is then transmitted to the dynamic causal diagnosis unit 330.

[0076] The dynamic causal diagnosis unit 330 is the diagnosis engine of this invention. The dynamic causal diagnosis unit 330 uses a graph neural network structure to propagate and analyze the fused residual signal in order to locate the root cause of the anomaly.

[0077] The dynamic causal diagnostic unit 330 constructs a dynamic causal graph network (DCGN). Unlike traditional graph neural networks that use static adjacency matrices, the information propagation weights in this network are dynamically changing. Specifically, the dynamic causal diagnostic unit 330 directly uses the dynamic Jacobian matrix calculated in real time by the physical channels. This serves as a weighted adjacency matrix for information propagation between graph nodes. This design ensures that the analysis of abnormal signal propagation strictly follows the actual physical impact path and intensity under the current system operating conditions.

[0078] This dynamic causal graph network is composed of multiple stacked information propagation layers. In the... Layer, node feature matrix The update process is as follows: ; In the formula, For the network Layer at time The node feature matrix has dimensions of . ; It is a non-linear activation function, such as the ReLU function; For a moment The estimated dynamic Jacobian matrix has dimension . × It is generated in real time by the physical channel model; For the network Layer at time The node feature matrix; For the network A learnable weight matrix for the layer; This refers to the layer index of a graph network.

[0079] The physical meaning of this propagation rule is that the features of each node in the next layer are obtained by weighting and aggregating the features of its neighboring nodes in the current layer according to the physical influence strength defined by the Jacobian matrix, and then undergoing a feature transformation.

[0080] go through After the information is propagated through the layers, the final node feature matrix is ​​obtained. This matrix captures the combined effects of the initial residuals propagating and converging throughout the system.

[0081] The dynamic causal diagnostic unit 330 then passes through a readout function This maps the high-dimensional final feature vector of each node to a single scalar value, namely the causal score of that node.

[0082] ; In the formula, The last layer (the first layer) of the dynamic causal graph network The node feature matrix output by the layer; This represents the total number of layers in the dynamic cause-effect graph network. For learnable readout functions, such as a simple multilayer perceptron; To include all Causal scores of each node Dimensional vector.

[0083] At any time Causal score vector The node corresponding to the component with the largest median value is determined to be the most likely root cause of the anomaly at that moment. This vector will be passed to the diagnostic result output module 400.

[0084] like Figure 6 As shown, Figure 6 This is a schematic diagram of the internal structure of a diagnostic result output module according to an embodiment of the present invention. The diagnostic result output module 400 in this embodiment serves as an interaction interface between the system and the operator, converting the quantitative analysis results output by the real-time diagnostic and alarm module 300 into intuitive and interpretable alarm information and diagnostic reports. Specifically, the diagnostic result output module 400 may include: an alarm generation and management unit 410, and an interpretive report generation unit 420.

[0085] The alarm generation and management unit 410 is used to evaluate the real-time analysis results according to preset logical rules to determine whether to generate an alarm event. The alarm generation and management unit 410 ensures the sensitivity and accuracy of alarms by setting composite alarm triggering conditions and dynamically adjusting alarm thresholds.

[0086] This alarm generation and management unit 410 uses composite conditions to trigger alarms, avoiding the limitations of a single judgment criterion. An alarm event is triggered when any of the following conditions are met: The residual exceedance condition refers to the situation where the magnitude of the physical residual or the mode residual at any node exceeds its respective preset threshold. This condition is used to quickly detect any significant deviations from the system state, and its mathematical expression is: ; In the formula, It is a logical existential quantifier; For the index of the node; This refers to the total number of variables or nodes monitored in the system. The current moment; For a moment node The magnitude of the physical residual; This represents the operation of taking the absolute value of each element; For nodes The physical residual alarm threshold; For logical OR operator; For a moment node The residual amplitude of the model; For nodes The residual alarm threshold for the mode.

[0087] The causal score exceeding the threshold condition refers to a situation where the causal score of any node is significantly higher than that of other nodes, exceeding a global score threshold. This condition is used to trigger an alarm when the root cause is clearly identified. Its mathematical expression is: ; In the formula, It is a logical existential quantifier; For the index of the node; For a moment node The causal score; This is the global causal score alarm threshold.

[0088] To accommodate operating condition drift during normal industrial process operation, this alarm generation and management unit 410 sets residual alarm thresholds. and An adaptive adjustment mechanism is employed. These thresholds are dynamically calculated based on the statistical characteristics of historical normal operation data.

[0089] In one specific implementation, the threshold can be set based on the statistical distribution of the normal data residuals observed within a sliding time window, for example, using the 3-sigma principle: ; In the formula, For a moment The dynamic threshold; This is the mean of the residuals calculated from the historical normal data window; The standard deviation of the residuals is calculated from the historical normal data window; This is a configurable sensitivity coefficient, for example, a value of 3. This mechanism allows the alarm system to adapt to normal process changes, reducing false alarms caused by normal operations such as switching operating conditions.

[0090] When the alarm generation and management unit 410 triggers an alarm event, the interpretive report generation unit 420 is activated to generate a structured diagnostic report containing rich contextual information to help the operator understand the fault and take action.

[0091] The report generated by the explanatory report generation unit 420 first clearly points out the root cause of the inferred anomaly, the specific content of which includes: Root cause identification refers to the report clearly highlighting the node with the highest causal score. This node was identified as the most likely source of the failure. The determination method was as follows: ; In the formula, The node that is determined to be the root cause; It is a node in the system; It is the set of all nodes in the system; For a moment node The causal score; Anomaly type analysis refers to the report indicating the type of residual that initially triggered the alarm. If the physical residual exceeds the limit first, the report will suggest that the anomaly may be related to a deviation from the system's physical laws; if the mode residual exceeds the limit first, it suggests that the anomaly may be a new, complex failure mode that has not been seen in the past.

[0092] Deviation from the physical laws of the system refers to a significant inconsistency between the real-time operating state of the industrial system and the physical model learned through the physical channel model of the present invention, which can characterize its local dynamic behavior.

[0093] Specifically, its technical principles are explained as follows: In this embodiment of the invention, the physical laws of the system (such as mass balance, energy balance, chemical reaction kinetics, etc. under specific operating conditions) are not pre-embedded in the form of analytical equations, but are implicitly and data-drivenly captured by learning massive amounts of historical normal operating condition data through a physical channel model (such as a recurrent neural network).

[0094] The result learned by this physical channel model is that it can characterize the system at any time. The dynamic Jacobian matrix of the local dynamic characteristics Each element of this matrix All variables were quantified. For variables The instantaneous physical influence relationship.

[0095] The physical channel model uses its learned physical laws (i.e., dynamic Jacobian matrix) and recent historical states to predict the system state at the current moment, obtaining a physical state prediction value. This predicted value represents the state that the system should exhibit, given its inherent physical relationships.

[0096] When the standardized state vector is actually measured Compared with the above-mentioned state physical prediction values When significant differences arise, substantial physical residuals are produced. Therefore, a large physical residual magnitude is a direct, quantitative manifestation of the deviation of the system's physical laws from the expected value in terms of technology.

[0097] Suppose a temperature sensor malfunctions, causing its readings to drift inaccurately. At this time, other relevant system variables (such as pressure, flow rate, and valve position) remain normal. Based on the physical correlations exhibited by these normal variables, the physical channel model of this embodiment predicts a correct temperature value. However, because the actual reading of the faulty sensor differs significantly from this physically-based prediction, the difference (i.e., the physical residual) will far exceed the normal fluctuation range. This situation is identified by this embodiment as an anomaly related to a deviation from the system's physical laws, and its root cause is likely a physical fault at the instrument level.

[0098] In summary, the deviation of the physical laws of the system in this embodiment of the invention is defined as a statistically significant and non-negligible deviation between the real-time observed values ​​of one or more nodes and the expected values ​​predicted by a data-driven physical model based on the coordinated state of other related nodes in the entire system.

[0099] Quantitative evidence presentation refers to the report displaying the nodes with the highest causal scores and their scores in the form of lists or charts, as well as the change curves of the physical residuals and pattern residuals of the relevant nodes before and after the alarm time, providing quantitative data support for the diagnostic conclusions.

[0100] To help operators intuitively understand the evolution of a fault, this explanatory report generation unit 420 provides a visualization function for the fault propagation path. This function is implemented on a user interface that loads the system topology diagram generated by the offline model building module 200. This function is a specific technical implementation for tracing back and highlighting the path to the node corresponding to the root cause. This reverse tracing process is achieved through the following steps: First, this unit determines the starting and ending points of the tracing path as follows: The ending point of the path is the root cause node that is determined to have the highest causal score, denoted as RC_node; the starting point of the path is the node that is determined to be the first node in the system to trigger the residual over-limit alarm, that is, the affected node that first shows obvious abnormal signs, denoted as Start_node.

[0101] Then, this explanatory report generation unit 420 is configured to execute an iterative reverse tracing algorithm starting from Start_node, which utilizes the alarm time... The dynamic Jacobian matrix estimated in real time by the physical channel model As a basis for physical causal relationships.

[0102] In any iteration step of the tracing process, if the currently known path node is (Initial time) If =Start_node), then its upstream affected nodes This is determined by solving the following optimization problem: ; In the formula, This formula is used to calculate the value of the current node in the reverse tracing path. The next node upstream; In the reverse tracing path, the first In this iteration, the known nodes that are the current objects of analysis; These are mathematical operators whose function is to find and return expressions within parentheses that make the expression within the parentheses appear normal. The node index that yields the maximum value ; This is an index variable used to traverse all nodes in the system. To be at the moment of alarm Dynamic Jacobian matrix One of the elements, whose value specifically represents the node. When the state variable undergoes a unit change, for the node The degree of instantaneous impact of the state variables; The current moment when the alarm event is triggered and root cause diagnosis is performed; A pre-defined scalar tracing threshold is used to filter strong causal relationships. Its function is to only consider the instantaneous influence strength between two nodes. Only when the value is greater than this threshold is the connection considered a valid path with outstanding physical significance and traceability.

[0103] This iterative process continues, calculating the upstream node at each step. Add it to the path and use it as the current node for the next iteration, until the node traced is the known path endpoint RC_node.

[0104] Finally, the sequence of nodes with clear physical meaning, consisting of Start_node, all intermediate influencing nodes found by the above algorithm, and RC_node, along with the connecting edges between them, will be presented on the system topology graph in a highlighted color, bold lines, or other eye-catching visual manner.

[0105] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An alarm AI application system based on industrial data, characterized in that, The system includes: The data access and preprocessing module is used to collect multidimensional time series data from industrial data sources and generate standardized state vector sequences. The offline model building module is used to train and solidify physical channel models and data channel models offline based on historical data; The real-time diagnosis and alarm module is used to receive the standardized state vector sequence, load the physical channel model and the data channel model, generate physical residual vector and pattern residual vector in parallel, and perform root cause diagnosis based on the physical residual vector and the pattern residual vector to generate a quantified causal score. The diagnostic result output module is used to generate an alarm event based on the physical residual vector, the pattern residual vector, and the causal score, and output an explanatory report containing the root cause location.

2. The alarm AI application system based on industrial data according to claim 1, characterized in that, The real-time diagnosis and alarm module includes a physical channel residual calculation unit, which is used for: Using the physical channel model, the dynamic Jacobian matrix is ​​estimated in real time based on the state vector sequence within a time window of a preset length prior to the current moment; Using the dynamic Jacobian matrix, the state vector sequence within the time window, and the state change trend, the physical prediction value of the state at the current moment is calculated; The physical residual vector is obtained by calculating the vector difference between the standardized state vector generated by the data access and preprocessing module at the current time and the state physical prediction value. The absolute value of each component of the physical residual vector is defined as the physical residual amplitude of the corresponding node.

3. The alarm AI application system based on industrial data according to claim 1, characterized in that, The real-time diagnosis and alarm module includes a data channel residual calculation unit, which is used for: Using the data channel model, a state vector sequence within a time window of a preset length prior to the current moment is reconstructed to obtain a state data pattern reconstruction vector corresponding to the current moment. The data channel model is a deep autoencoder model trained based on historical normal operating condition data. The pattern residual vector is obtained by calculating the absolute value of the difference between the standardized state vector generated by the data access and preprocessing module at the current time and the state data pattern reconstruction vector. Each component of the pattern residual vector is defined as the pattern residual magnitude of the corresponding node.

4. The alarm AI application system based on industrial data according to claim 1, characterized in that, The real-time diagnosis and alarm module further includes a hybrid residual fusion unit, which is used for: Each component of the physical residual vector is concatenated with the corresponding component of the weighted pattern residual vector to generate a hybrid residual initial feature vector for each node in the system.

5. The alarm AI application system based on industrial data according to claim 4, characterized in that, The real-time diagnosis and alarm module further includes a dynamic causal diagnosis unit, which is used for: The initial feature vectors of the hybrid residuals generated for each node in the system are combined into an initial node feature matrix, which is then used as input to the dynamic causal graph network. In the dynamic causal graph network, the dynamic Jacobian matrix estimated in real time is used as the weighted adjacency matrix for information propagation between nodes. The initial node feature matrix is ​​updated through multi-layer information propagation to obtain the final node feature matrix. The final node feature matrix is ​​processed by a readout function to calculate the causal score characterizing the likelihood of each node as the root cause of an anomaly.

6. The alarm AI application system based on industrial data according to claim 2, characterized in that, The offline model building module is used for: Each dimension of the multidimensional time series data is defined as a node in a graph, and edges are established based on the known physical, energy, or signal flow relationships between the nodes to generate a system topology graph. A recurrent neural network is trained as the physical channel model to achieve a nonlinear mapping from the state vector sequence to the dynamic Jacobian matrix; An autoencoder based on a recurrent neural network is trained as the data channel model to learn and reproduce high-dimensional data association patterns under normal operating conditions.

7. The alarm AI application system based on industrial data according to claim 6, characterized in that, The diagnostic result output module includes: The alarm generation and management unit is used to trigger the alarm event when the physical residual magnitude of any node exceeds a first preset threshold or the pattern residual magnitude exceeds a second preset threshold, or when the causal score of any node exceeds a global score threshold.

8. The alarm AI application system based on industrial data according to claim 7, characterized in that, The alarm generation and management unit is also used to adaptively adjust the first preset threshold and the second preset threshold based on the statistical distribution of normal data residuals observed within a sliding time window.

9. The alarm AI application system based on industrial data according to claim 8, characterized in that, The diagnostic result output module also includes: The explanatory report generation unit is used to generate a diagnostic report containing the following content when the alarm event is triggered: The node with the highest causal score is highlighted as the root cause of the anomaly; Determine whether the residual type that initially triggered the alarm was a physical residual exceeding the limit or a mode residual exceeding the limit.

10. The alarm AI application system based on industrial data according to claim 9, characterized in that, The explanatory report generation unit is also used to provide a visual representation of the fault propagation path by performing the following operations: On the system topology map, starting from the affected node, the path is traced back and highlighted to the node corresponding to the root cause, following the connection defined by the dynamic Jacobian matrix at the alarm time and whose weight exceeds the preset tracing threshold.

Images