Multi-modal fusion causal analysis method and platform for industrial safety production

Abstract

The application provides a multi-modal fusion causal analysis method and platform for industrial safety production, which comprises the following steps: preprocessing multi-modal key data from an industrial site to generate a standardized time series data matrix; constructing an industrial safety causal graph based on the standardized time series data matrix, combining industrial physical rules and operation process semantics; optimizing the structure and edge weight of the industrial safety causal graph based on a self-supervised learning mechanism and using the standardized time series data matrix to obtain an optimized causal reasoning model; and performing path-level risk activation judgment and accident causal tracing by using the causal reasoning model and a real-time standardized time series data matrix constructed from real-time multi-modal key data, so as to generate a risk alarm and a tracing report. The application significantly improves the foresight, accuracy and interpretability of industrial safety monitoring.

Description

Technical Field

[0001] This invention belongs to the field of industrial safety production, and in particular relates to a multimodal fusion causal analysis method and platform for industrial safety production. Background Technology

[0002] In modern industrial safety production scenarios, factors such as equipment operating status, environmental fluctuations, and personnel operating behaviors interact and intertwine in a complex manner. This means that potential risks are often not triggered by a single variable, but rather accumulate gradually and trigger accidents through the continuous evolution of multiple types of data. Traditional safety monitoring systems generally rely on threshold triggering mechanisms from single-point sensors such as temperature, pressure, and current, lacking an understanding of the correlations between different modalities and struggling to identify industrial risks with strong concealment and obvious chain-like development characteristics. Furthermore, multimodal data in actual industrial systems often suffers from inconsistent sampling frequencies, inconsistent recording formats, and frequent missing points, making it difficult for traditional algorithms to perform joint analysis within a consistent data domain. This results in a break in the connection between different data sources, severely impacting the comprehensiveness and accuracy of risk identification. More critically, industrial accidents are usually accompanied by complex causal chains, such as "abnormal operation leading to equipment overload, which in turn causes temperature rise and triggers interlocking actions." Most existing technologies can only provide post-event static analysis, lacking the ability to infer causal relationships in real time and also lacking interpretability, failing to clearly demonstrate the causes, consequences, and risk evolution process of the accident to operators.

[0003] In highly dynamic, noisy, and interconnected industrial environments, this monitoring method, which lacks a causal perspective, is difficult to meet the needs of real-time safety assurance. This results in potential risks not being detected in advance, and accident tracing is also difficult to quickly locate the root cause due to the lack of systematic causal support, which seriously affects the efficiency and reliability of industrial safety production. Summary of the Invention

[0004] The purpose of this invention is to propose a multimodal fusion causal analysis method and platform for industrial safety production, thereby solving the above-mentioned problems.

[0005] To achieve the above objectives, a multimodal fusion causal analysis method for industrial safety production is provided in a first aspect of the present invention, the method comprising the following steps:

[0006] S1: Preprocess multimodal key data from industrial sites to generate a standardized time-series data matrix;

[0007] S2: Based on the standardized time-series data matrix, an industrial safety causal graph is constructed by combining industrial physical rules and operational process semantics. The industrial safety causal graph is used to represent the temporal causal relationship between industrial variables.

[0008] S3: Based on the self-supervised learning mechanism, the structure and edge weights of the industrial safety causal graph are optimized using the standardized time series data matrix to obtain the optimized causal inference model;

[0009] S4: Using the causal reasoning model and the real-time standardized time-series data matrix constructed from real-time inflow multimodal key data, perform path-level risk activation judgment and accident causal tracing to generate risk alarms and tracing reports;

[0010] The industrial safety causal graph is constructed as follows:

[0011] Based on the standardized time series data matrix, a first-order difference is performed on each variable sequence within the sliding time window. If the change exceeds the dynamic threshold, the current node is recorded as a causal event node to construct a causal node set.

[0012] Based on the semantics of industrial physical rules and operational processes between any two events, a corresponding regularization term is constructed, and an edge weight matrix is ​​generated to represent the strength of causal influence between all variables under industrial logic and data-driven conditions.

[0013] By combining the causal node set and the edge weight matrix, an industrial safety causal graph is generated.

[0014] Furthermore, the preprocessing includes:

[0015] Collect heterogeneous data from sensors, control systems, and production management systems; perform integrity verification and cleaning on all data; and complete any missing data.

[0016] Data from different sampling frequencies are uniformly mapped to the same reference time axis, and all variables are standardized to form a standardized time series data matrix indexed by a unified time step.

[0017] Furthermore, the heterogeneous data includes at least continuous physical quantities, high-frequency state quantities, and operation event logs.

[0018] Furthermore, the regularization terms constructed based on the industrial physical rules and operational process semantics between any two events include physical feasibility constraints and operational process consistency constraints.

[0019] The physical feasibility constraint is used to exclude causal relationships that violate the physical principles of industrial equipment, and the operational process consistency constraint is used to ensure that causal relationships conform to the preset process operation sequence.

[0020] Furthermore, S3 specifically includes:

[0021] A graph neural network model is constructed, using multimodal time-series data fragments from the previous time window as node features and the initial industrial safety causal graph as the initial graph structure.

[0022] Design a self-supervised loss function, which includes a structural perturbation consistency loss and a causal sparsity preservation regularization term;

[0023] By minimizing the self-supervised loss function, the graph neural network model is trained, and the edge direction and edge strength of the initial graph structure are dynamically adjusted to obtain an optimized causal graph and corresponding causal inference model that is structurally concise and conforms to industrial logic.

[0024] Furthermore, the causal sparsity preservation regularization term penalizes redundant causal edges based on at least one of industrial segment logic, physical isolation boundaries, and temporal independence between variables.

[0025] Furthermore, the steps for determining path-level risk activation include:

[0026] The real-time standardized time-series data matrix is ​​input into the optimized causal reasoning model to obtain the current state embedding representation of the variables;

[0027] Based on the preset high-risk causal path, the current state embedding representation, the optimized causal edge strength, and the dynamic response delay penalty factor, the real-time activation score of each path is calculated. When the real-time activation score of a certain path exceeds its corresponding threshold, it is a risky path, and it is activated and an alarm is triggered.

[0028] Furthermore, the dynamic response delay penalty factor is calculated based on the degree of deviation between the actual response delay of the variable pair and its historical average response delay; the greater the deviation, the stronger the penalty.

[0029] Furthermore, the steps for tracing the cause of the accident include:

[0030] After the risk path is activated, the complete causal chain from the outcome variable to the cause variable is traced in reverse based on the optimized causal graph and the current state embedding of the variables;

[0031] Calculate the causal thrust of each edge in the complete causal chain, which integrates the upstream variable state, the causal edge strength, and the difference in state embedding between the downstream and upstream nodes;

[0032] Based on the causal reasoning, the relevant variables are ranked in order of responsibility, and a source tracing report containing causal paths and responsibility weights is generated.

[0033] In a second aspect, the present invention provides a multimodal fusion causal analysis platform for industrial safety production, including a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the method described in any one of the above.

[0034] The beneficial technical effects of the present invention are at least as follows:

[0035] This invention proposes a multimodal fusion causal analysis method and platform for industrial safety production, forming a complete technical closed loop from data preprocessing, causal chain construction, self-supervised causal optimization to real-time risk identification and accident tracing. Addressing the issues of mixed formats and varying sampling frequencies in industrial multimodal data, this invention integrates key variables such as temperature, pressure, current, vibration, and operating status into highly consistent time-series data suitable for causal analysis through a unified time axis and standardization mechanism, providing stable input for causal inference. Recognizing that causal relationships in industrial scenarios depend on equipment physical characteristics and process constraints, this invention embeds mutation event detection results, industrial physical rules, and operational process semantics into a knowledge graph to construct a causal chain structure that accurately reflects the operational logic of industrial systems. To address the lack of labeled samples and the susceptibility of initial causal chain construction to noise in industrial scenarios, this invention further introduces a self-supervised graph structure optimization mechanism, enabling dynamic correction of causal edge weights within continuously running industrial data, filtering redundant paths and strengthening critical paths to form a causal expression that better reflects the true logic of the production process.

[0036] Ultimately, based on the optimized causal graph and model, this invention proposes a path-level risk activation mechanism and a multi-hop causal thrust assessment method. It incorporates the current state, causal edge credibility, and historical response delays into a real-time scoring framework, enabling the identification of chain risks and the automatic generation of clear traceability responsibility chains. This provides timely, interpretable, and actionable risk alarms and accident root cause analysis for industrial sites. The methodology of this invention achieves end-to-end innovation from multimodal data to real-time causal inference and then to actionable alarm output, significantly improving the foresight, accuracy, and interpretability of industrial safety monitoring. Attached Figure Description

[0037] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0038] Figure 1 This is a flowchart of the multimodal fusion causal analysis method for industrial safety production according to the present invention. Detailed Implementation

[0039] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0040] like Figure 1 As shown in the embodiment of the present invention, a multimodal fusion causal analysis method for industrial safety production is provided, the method comprising:

[0041] S1: Preprocess multimodal key data from industrial sites to generate a standardized time-series data matrix.

[0042] Specifically, the input data mainly comes from two aspects: first, a list of key variables identified in the previous stage analysis and included in the causal modeling scope; and second, raw data streams acquired from the industrial acquisition systems already deployed on-site. This includes: continuous physical quantity data such as temperature, pressure, and flow rate, collected from analog input modules in PLC or DCS systems, with a sampling frequency typically of 1Hz or 10Hz; high-frequency state data such as vibration and current, from dedicated acquisition equipment (such as edge acquisition modules connected to vibration sensors), with a sampling frequency up to 1kHz; personnel operation records and equipment operation logs, from event databases in SCADA or MES systems, recorded as unstructured text or time event pairs; and work environment information (such as access control status, location information, etc.), which may originate from RFID or Bluetooth beacon systems. During the acquisition process, all data is uniformly uploaded to the cache queue on the main control server via industrial Ethernet as the raw input for this step. The system assigns a unique identifier to each data channel (such as Sensor_T01, Log_OPR03, etc.) to ensure consistency in variable naming.

[0043] First, the integrity of all data channels needs to be verified, eliminating any data that is interrupted or has abnormal waveforms. Taking a temperature sensor as an example, if the temperature data remains constant at 0 for a certain period, or if the sampling interval jumps significantly, the system will automatically determine that it is a device malfunction or communication anomaly, and directly mark that segment of data as unusable. Then, the data cleaning and missing data completion stage begins. For continuous variables (such as temperature and pressure), noise suppression is performed using the moving window mean, and Z-score anomaly detection is used for abrupt changes. If a sample deviates from the moving window mean by more than 3 times the standard deviation, the point is marked as an anomaly and replaced with the window mean. For operation log data, because the records are sparse and discontinuous, the system first normalizes and numbers all time events, then constructs a time series, and automatically fills in the missing "hold" states by using the preceding and following states. For example, if the "feeding start" event occurs at 10:05 and the "feeding stop" event occurs at 10:15, the system automatically generates a "feeding in progress" state for every second between 10:05 and 10:15.

[0044] The time synchronization process employs a unified time axis alignment method. The system constructs a reference time axis (e.g., a timeline with 1-second intervals), and all variables with different sampling frequencies are mapped to this time axis. High-frequency variables (such as vibrations) are averaged and compressed to a low-frequency step size using a window; low-frequency events (such as operation commands) are labeled and expanded according to the most recent occurrence time of the event. In this way, each time point corresponds to a complete set of variable states.

[0045] Furthermore, the standardization phase employs a zero-mean, unit-variance approach, ensuring that all variables are modeled on the same scale. The processing method is as follows:

[0046]

[0047] in, Indicates the first Variables in time The raw sampled values ​​come from sensors or logging systems; The historical average of this variable over the most recent complete production cycle (e.g., the last 24 hours) is calculated using a windowed cumulative method. This represents the standard deviation of the variable within the same time window. This represents the standardized variable values, used as input for subsequent modeling. Structurally, the final result is a time-series matrix indexed by a uniform time step, with each row representing a time point and each column representing a standardized variable. For example, vibration sensor data is collected 1000 times per second by accelerometers connected to edge nodes. The data is first averaged and compressed per second, then synchronized with temperature and pressure data to form a record of one row per second. Operator feeding operations are recorded by the MES system, showing the event time and operation command. The system expands this into one operation status per second and merges them into the same time-series structure.

[0048] S2: Based on the standardized time-series data matrix, an industrial safety causal graph is constructed by combining industrial physical rules and operational process semantics. The industrial safety causal graph is used to represent the temporal causal relationship between industrial variables.

[0049] Specifically, the input is the standardized time-series data matrix output in step 1. ,in To standardize the number of time steps, This represents the total number of standardized variables. (Each column...) Indicates the first Variables in time The values ​​are derived from standardized device sensor data (such as temperature, pressure, and current) or discretized human operating states (such as start, stop, and other commands). The goal of causal graphing is to generate a directed graph structure. ,in For a set of causal nodes, For a weighted adjacency matrix, each edge Representing variables Changes on variables The strength of the causal impact of the change over time.

[0050] First, before constructing an industrial safety causal graph, potential causal event nodes need to be identified. This is achieved through... Finding local mutation regions within a sliding time window For each variable sequence, perform a first-order difference; if its change exceeds a dynamic threshold... If a point is identified as a causal event node, then that point is recorded as such. The specific judgment rules are as follows:

[0051]

[0052] in, Indicate whether to include time The Each variable is considered as a valid causal event. For variables The mutation threshold, determined by its position within the historical window. Internal standard deviation Multiplier setting (e.g.) To accommodate differences in sensitivity under different units of measurement; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. These are the input variables that have already been normalized in step 1, and do not require repeated processing. The innovation of this judgment mechanism lies in its reliance on the standard deviation of a local dynamic window rather than a global threshold, which enhances the ability to identify sudden, small-scale events (such as operational jumps, current spikes, etc.), and is particularly suitable for high-frequency, low-varying industrial signals.

[0053] Furthermore, the event nodes obtained based on the above event recognition mechanism constitute a set. Used to generate node sets in a graph structure. To generate edge sets Furthermore, the chronological relationships between events are analyzed in a focused manner, and physical and logical constraints are designed based on industrial knowledge to serve as a constraint mechanism for causal inference. The overall mapping optimization objective is defined as follows:

[0054]

[0055] in, It is a similarity loss function built based on event time delay statistics, used to quantify events. Prior to the event Frequency of occurrence, such as events in a sliding window and If the time difference is less than a certain threshold, then increase. The strength; This is a physical feasibility regularization term used to penalize edges that violate basic industrial principles (such as "temperature cannot rise before power"). For consistency of operation process, regularization is used to ensure that the generated edge relationships conform to the process operation sequence. For example, "equipment start-up → current increase → main process start" is an allowed path, while reverse paths are suppressed. To control the weight hyperparameters of the two regularization terms, they need to be set based on experience in real-world scenarios. The innovation lies in the fact that the design of the two regularization terms does not rely on large-scale labeled data, but is achieved through symbolic rules derived from industrial safety standards and flowchart conversions. For example, in the control of a high-temperature reactor, there exists a strict sequence of "start stirring → start heating → reach the set temperature → start cooling," and these structures can be defined and built into a state machine. This significantly reduces the occurrence of physical anomalies such as "motor malfunction affecting stirring control" in causal mapping. Based on the physical characteristics of the equipment and the law of conservation of energy, settings such as "input electrical power cannot be driven by vibration signals" are used to eliminate meaningless paths.

[0056] Finally, the initial causal edge weight matrix This represents the strength of causal influence among all variables under industrial logic and data-driven conditions, and can be used as input for the causal inference model structure in the next step. Causal node set With the initial causal edge weight matrix Together they form a cause-and-effect graph of industrial safety It features a clear structure, orderly path, and reasonable side logic and physics, which can effectively support the needs of risk prediction and accident tracing in industrial safety production.

[0057] S3: Based on the self-supervised learning mechanism, the structure and edge weights of the industrial safety causal graph are optimized using the standardized time series data matrix to obtain the optimized causal inference model.

[0058] Specifically, this step aims to refine the industrial safety causal graph constructed in step 2. By optimizing the structure and parameters, the causal network not only possesses statistical relevance but also reflects real-world causal relationships within the context of physical and operational constraints in industrial safety production scenarios. This step innovatively introduces a training mechanism combining "self-supervised optimization of graph structure" and "preservation of industrial causal semantics," enabling adaptive adjustment of edge directions and strengths in the graph structure based on actual data without external annotations. This mechanism significantly improves the accuracy and interpretability of the causal graph structure in subsequent risk reasoning and accident tracing processes.

[0059] The input is the industrial safety cause-effect graph output from step 2. ,in This is a set of event nodes, representing the state points where industrial variables undergo abrupt changes in a time series. The initial causal edge weight matrix between variables is derived from the fusion of information on event sequence, physical feasibility, and operational procedures. Additionally, the input includes the standardized multimodal time series matrix output from step 1. Each column represents the standardized value of the variable, with a time step of [value missing]. The total number of variables is The goal of this step is to adjust the initial causal edge weight matrix. By simplifying the structure and redistributing the weights, unnecessary paths are eliminated and the expression of critical paths is enhanced while maintaining the accuracy of the causal chain.

[0060] Furthermore, to achieve the above objectives, a lightweight graph neural network model (causal reasoning model) was constructed. This model is used to simulate state transfer between variables in a causal graph. It consists of a two-layer graph convolutional network, each layer containing an adjacency matrix weighted aggregation module and a linear transformation module. The input node features are... The corresponding variable in the past The observation vectors within each time step are calculated using model propagation as follows:

[0061]

[0062] in, For the first Features of each node in a layered graph convolutional network For the first Features of each node in a layered graph convolutional network Indicates the initial node characteristics. yes In the past Step data segment; In the initial causal edge weight matrix The edge weight matrix is ​​obtained by adding self-loops and normalizing the matrix. For the first The linear transformation weight matrix of the layer; This is the activation function. This structure allows information to propagate along the graph structure from the dependent variable to the effect variable, enabling the model to learn the role of the graph structure in state evolution.

[0063] During the training phase, a self-supervised loss function based on "structure-invariant prediction preservation" is introduced. This is used to determine whether the current graph structure stably represents a causal propagation path. The method involves using the initial causal edge weight matrix... The structure is perturbed to generate a set of contrasting structures. Each Changes are generated through edge sampling or edge reversal. Using the same input... In each The image is propagated upwards, and its output is compared with the original image output:

[0064]

[0065] Where M represents the causal reasoning model constructed in step 3; This represents the regularization weight coefficients. The first term is the structural perturbation consistency loss, which ensures that the current graph structure can still stably output the same causal signal path under perturbation. For the number of perturbations, The first term is the Frobenius norm; the second term is the innovative causal sparsity-preserving regularization term. ,in Represents variable pairs The redundancy score is determined by the following engineering rule of thumb: If and Edges belonging to different work sections, without process flow relationships, or whose data do not have long-term time series dependencies are considered redundant edges, and their penalty weight is increased. The variables obtained in step 2 to variable The initial causal edge weights are used to characterize the variables. For variables The strength of the causal influence.

[0066] It is important to note that this regularization term is a major innovation of this invention: it is set based on three sources: industrial division of labor logic, physical isolation boundaries, and timing independence. The values ​​are then integrated into causal graph structure learning, solving the problem of "strong logical constraints and weak labeling support" in industrial applications that traditional GNN structure learning struggles to handle. For example, in chemical engineering scenarios, if variables... Indicates the liquid level in the raw material tank, variable This represents the packaging temperature. Although there is a correlation in the data (such as changes in production batches at the same time), there is no operational logic path between the two. Therefore, it should be identified as a redundant connection and eliminated in the model.

[0067] The final output of the model is the edge weight matrix after structural optimization. This matrix maintains the same shape as the initial causal edge weights. While maintaining consistent node definitions and dimensions, the model underwent structural reconstruction in edge direction selection and edge strength distribution. After structural optimization, the causal paths in the model are more sparse, clear, and consistent with industrial logic, and possess a certain degree of resilience against faulty edge perturbations, providing a solid foundation for real-time path activation detection and risk path identification in the next step. Furthermore, the trained model... It can be deployed as a high-frequency risk prediction module to the edge gateway to continuously monitor the evolution trend of causal paths in the production process, further improving the practicality and response speed of the invention platform.

[0068] S4: Using the causal reasoning model and the real-time standardized time-series data matrix constructed from real-time inflow multimodal key data, perform path-level risk activation judgment and accident causal tracing to generate risk alarms and tracing reports.

[0069] Specifically, this step involves optimizing the cause-effect graph structure in step S3. With graph neural network models Based on this, combined with standardized multimodal time-series data streams This system enables real-time monitoring and activation assessment of risk paths in industrial production processes, and traces the root cause of events when anomalies occur, forming directly executable alarm and tracing solutions. Unlike traditional safety early warning systems based on single variables or threshold triggers, this step introduces "path-level causal activation measurement" and "segmented causal response intensity assessment" to achieve overall perception, joint judgment, and segmented interpretation of the accident risk chain. This mechanism is particularly suitable for the target scenario of this invention—a multimodal driven industrial safety production system, characterized by: numerous variable types, diverse modes, long state trigger chains, complex interactions, and many potential risks not caused by isolated variables but evolving from the interaction of multiple variables.

[0070] The input section consists of three main structures: first, the continuous data stream obtained from the normalization preprocessing in step S1. ,in For the number of variables, Indicates at time The first step is to determine the normalized state of each industrial variable; the second step is to construct the causal graph structure in step 2 and optimize it in step 3. It contains all variable nodes. With edge weight matrix Finally, there is the trained causal reasoning model. The model consists of a two-layer graph convolutional network structure. Each layer uses a normalized adjacency matrix for feature propagation, the activation function is ReLU, the inter-layer embedding dimension is fixed, and the model parameters are optimized through self-supervised learning.

[0071] Furthermore, the core operation process begins with the extraction of consecutive time segments. Every other sliding window... Extracting data of length from the most recent historical data. window sequence Input to the model Obtain the final node embedding ,in , The embedding dimension represents the graph semantic representation of the variable in its current state. To detect whether potentially high-risk paths are activated, the system pre-defines several sets of high-risk paths. Each path This represents a multivariate causal sequence.

[0072] For each one Construct the following path activation scoring function:

[0073]

[0074] in, Score the path activation level. For time On the time path The input values ​​of each node come from ; The total number of causal nodes. It is the Sigmoid nonlinear mapping function; The optimized causal edge strength; This is an important innovative design element used to introduce dynamic response latency penalty, which is calculated as follows:

[0075]

[0076] in, For dynamic response delay penalty, This indicates that in the current window, from the variable State change to variable The reaction delay (which can be modeled) The graph propagation yields the implicit activation time. are variable pairs The historical average response latency is derived from the statistical values ​​learned in the training set; This is the penalty coefficient (set between 0.5 and 1.5), used to control the sensitivity to deviations from historical responses. Its function is to automatically reduce the activation score of a path when the response delay significantly deviates from the normal pattern, thereby avoiding false alarms or misjudgments caused by atypical activation patterns.

[0077] The innovation lies in the fact that this mechanism no longer assesses risk solely through the current variable magnitude or edge weights, but integrates three elements: the current state magnitude, the reliability of structural edges, and the consistency of historical delays, to construct a more stable and noise-resistant path activation mechanism. Especially in the context of high-frequency equipment interference, where data such as vibration and current are easily affected by instantaneous spikes, this mechanism can significantly suppress short-term spurious activation phenomena.

[0078] When any path of Exceeding its set threshold (Determined by the training process or defined by experts), the system determines that the risk path has been activated, triggers a real-time alarm, and transmits this path as the core trigger link to the source tracing module. Upon entering the source tracing phase, the system will use the cause-effect graph... and current embedding We trace the complete causal path from the end node to the starting dependent variable in reverse, and calculate the actual "causal push" on each side, as defined below:

[0079]

[0080] in, For time Time variable The standardized state value, The causal graph optimized in step 3 consists of variables Pointer variable The causal edge weights are used to characterize the strength of the causal influence between variables. For variables The node embedding results in the last layer of the graph neural network model For variables Node embedding results in the last layer of the graph neural network model;

[0081] This causal impetus It quantifies the degree to which a variable, in its current state, influences the state of its downstream variables. A larger value indicates a stronger impact of the variable's change on the system's state, thus carrying a higher "accident responsibility weight." The squared term emphasizes significant changes in the reasoning state, making it more helpful in identifying paths with abrupt shifts.

[0082] The final output contains two structured results: the first is the alarm event, which includes the activation path identifier and the score. The first item is the trigger time, the sequence of variables involved, and the recommended action level; the second item is the source tracing report, which includes a list of causal paths and the push for each edge. The output includes values, response timing information for each node, and responsibility order. These outputs can be directly pushed to industrial control systems, security platforms, and visual monitoring terminals via standard REST API interfaces, enabling closed-loop early warning and comprehensively improving the accident response capabilities of industrial systems.

[0083] This invention also provides a multimodal fusion causal analysis platform for industrial safety production, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the method described in any one of the above embodiments.

[0084] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0085] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A multimodal fusion causal analysis method for industrial safety production, characterized in that, The method includes the following steps: S1: Preprocess multimodal key data from industrial sites to generate a standardized time-series data matrix; S2: Based on the standardized time-series data matrix, an industrial safety causal graph is constructed by combining industrial physical rules and operational process semantics. The industrial safety causal graph is used to represent the temporal causal relationship between industrial variables. S3: Based on the self-supervised learning mechanism, the structure and edge weights of the industrial safety causal graph are optimized using the standardized time series data matrix to obtain the optimized causal inference model; S4: Using the causal reasoning model and the real-time standardized time-series data matrix constructed from real-time inflow multimodal key data, perform path-level risk activation judgment and accident causal tracing to generate risk alarms and tracing reports; The industrial safety causal graph is constructed as follows: Based on the standardized time series data matrix, a first-order difference is performed on each variable sequence within the sliding time window. If the change exceeds the dynamic threshold, the current node is recorded as a causal event node to construct a causal node set. Based on the semantics of industrial physical rules and operational processes between any two events, a corresponding regularization term is constructed, and an edge weight matrix is ​​generated to represent the strength of causal influence between all variables under industrial logic and data-driven conditions. By combining the aforementioned causal node set and edge weight matrix, an industrial safety causal graph is generated. S3 specifically includes: A graph neural network model is constructed, using multimodal time-series data fragments from the previous time window as node features and an initial industrial safety causal graph as the initial graph structure. Design a self-supervised loss function, which includes a structural perturbation consistency loss and a causal sparsity preservation regularization term; By minimizing the self-supervised loss function, the graph neural network model is trained, and the edge directions and edge strengths of the initial graph structure are dynamically adjusted to obtain a simplified optimized causal graph and a corresponding causal inference model that conforms to industrial logic; wherein, the self-supervised loss function Represented as: ； Where M represents the causal reasoning model; It is a standardized time series data matrix In the past Step data segment; For the initial causal edge weight matrix The structure is perturbed to generate a set of contrast structures; represents the regularization weight coefficients; the first term Represents the regularization weight coefficients; the first term The structural perturbation consistency loss is used to ensure that the current graph structure can still stably output the same causal signal path under perturbation. For the number of perturbations, Norm; second term To maintain regularity for causal sparsity, ,in, Represents variable pairs The redundancy score is determined by the following engineering rule of thumb: If and Edges belonging to different work sections, without process flow relationships, or without long-term time series dependencies are considered redundant edges, and their penalty weight is increased. For variables to variable The initial causal edge weights are used to characterize the variables. For variables The strength of the causal influence; The causal sparsity preservation regularization term applies a penalty to redundant causal edges based on at least one of the following: industrial division logic, physical isolation boundaries, and temporal independence between variables.

2. The multimodal fusion causal analysis method for industrial safety production according to claim 1, characterized in that, The preprocessing includes: Collect heterogeneous data from sensors, control systems, and production management systems; perform integrity verification and cleaning on all data; and complete any missing data. Data from different sampling frequencies are uniformly mapped to the same reference time axis, and all variables are standardized to form a standardized time series data matrix indexed by a unified time step.

3. The multimodal fusion causal analysis method for industrial safety production according to claim 2, characterized in that, The heterogeneous data includes at least continuous physical quantities, high-frequency state quantities, and operation event logs.

4. The multimodal fusion causal analysis method for industrial safety production according to claim 1, characterized in that, The regularization terms corresponding to the industrial physical rules and operational process semantics between any two events are constructed, including physical feasibility constraints and operational process consistency constraints. The physical feasibility constraint is used to exclude causal relationships that violate the physical principles of industrial equipment, and the operational process consistency constraint is used to ensure that causal relationships conform to the preset process operation sequence.

5. The multimodal fusion causal analysis method for industrial safety production according to claim 1, characterized in that, The steps for determining path-level risk activation include: The real-time standardized time-series data matrix is ​​input into the optimized causal reasoning model to obtain the current state embedding representation of the variables; Based on the preset high-risk causal path, the current state embedding representation, the optimized causal edge strength, and the dynamic response delay penalty factor, the real-time activation score of each path is calculated. When the real-time activation score of a certain path exceeds its corresponding threshold, it is a risky path, and it is activated and an alarm is triggered.

6. The multimodal fusion causal analysis method for industrial safety production according to claim 5, characterized in that, The dynamic response delay penalty factor is calculated based on the degree of deviation between the actual response delay of the variable pair and its historical average response delay; the greater the deviation, the stronger the penalty.

7. The multimodal fusion causal analysis method for industrial safety production according to claim 5, characterized in that, The steps for tracing the cause of the accident include: After the risk path is activated, the complete causal chain from the outcome variable to the cause variable is traced in reverse based on the optimized causal graph and the current state embedding of the variables; Calculate the causal thrust of each edge in the complete causal chain, which integrates the upstream variable state, the causal edge strength, and the difference in state embedding between the downstream and upstream nodes; Based on the causal inference, the relevant variables are ranked in order of responsibility, and a source tracing report containing causal paths and responsibility weights is generated.

8. A multimodal fusion causal analysis platform for industrial safety production, characterized in that: It includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the method as described in any one of claims 1 to 7.

Images