Industrial equipment fault intelligent diagnosis and root cause reasoning system based on multi-modal reasoning model

Abstract

The present application relates to the technical field of industrial equipment fault prediction, in particular to an intelligent diagnosis and root cause reasoning system for industrial equipment fault based on a multi-modal reasoning model, comprising a data acquisition and preprocessing unit, a multi-modal reasoning model reasoning unit, an intelligent fault diagnosis unit and a fault root cause reasoning unit.The present application adopts a dynamic attention weighting fusion technology based on a topological neighborhood anomaly prior constraint to determine an upstream neighborhood node set of a component, calculate a modal fault propagation interference prior coefficient, dynamically adjust a single-modal anomaly representation attention weight and realize intra-node cross-modal weighted fusion.Meanwhile, an abnormal propagation transfer matrix is constructed based on a topological graph, a fault propagation link is reasoned and a first abnormal node is screened in combination with the Bayes theorem, so that an anti-propagation interference cross-modal hierarchical reasoning of multi-modal operation data of an industrial equipment can be realized, the original abnormal representation of a component can be extracted and a fault propagation link and a first abnormal node can be accurately identified.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault prediction technology, and more specifically, to an intelligent diagnosis and root cause reasoning system for industrial equipment faults based on a multimodal reasoning model. Background Technology

[0002] As industrial equipment becomes larger and more intelligent, the multimodal monitoring data generated during its operation becomes the core basis for fault diagnosis. Traditional fault diagnosis methods struggle to efficiently process multimodal heterogeneous data and fail to analyze it in conjunction with the physical topological characteristics of equipment components. This makes it difficult to accurately distinguish between primary and propagated anomalies, resulting in fault diagnosis and root cause localization accuracy that cannot meet the actual application needs of industrial sites.

[0003] In existing technologies, relevant patents have incorporated topological structures into the research of equipment fault diagnosis and anomaly analysis. For example, invention patent CN202511160896.4 discloses a knowledge graph-based equipment fault prediction method and system. This includes preprocessing historical equipment fault data and equipment topology and attribute data; defining entity nodes and relational edges in the knowledge graph to form a structured knowledge graph; preprocessing basic data during equipment operation and extracting time-frequency and trend features; inputting these features into a residual convolutional network for feature extraction; outputting a basic feature vector; simultaneously using the residual convolutional network to predict the basic data and calculate the basic residual; and normalizing the absolute value of the basic residual to obtain the sensor anomaly score. This invention achieves integrated modeling of the semantic and data layers. For example, invention patent CN202512012890.9 discloses an anomaly analysis method that combines fault propagation path and topology, including: calculating the overall recovery time estimate under different topology structures based on business importance and recovery time target, and obtaining a set of candidate topology schemes that meet the recovery time requirements; extracting key nodes and key links for the set of candidate topology schemes, simulating fault scenarios based on component failure rate and link failure rate parameters, analyzing fault propagation paths, and evaluating the impact of faults on service quality recovery time.

[0004] Despite the design advantages of the aforementioned technical solutions, they also suffer from the following technical shortcomings: Firstly, they fail to incorporate the topological characteristics of equipment components into the multimodal operation data of industrial equipment for anti-propagation interference cross-modal hierarchical reasoning, thus hindering accurate extraction of component-level anomaly representations, fault propagation link reasoning, and identification of the first anomaly node. Secondly, invention patent CN202511160896.4 only integrates the topological structure into a knowledge graph for fault prediction, without performing component-level association mapping and cross-modal anomaly feature fusion for multimodal operation data, and without considering the interference of fault propagation on node anomaly representations, thus failing to generate anomaly representations reflecting the original state of the component itself. Thirdly, invention patent CN202512012890.9, while incorporating topological analysis of fault propagation paths, only simulates fault scenarios using fault rate parameters, without combining actual multimodal operation data of the equipment to perform actual fault propagation link reasoning under topological constraints, and thus also fails to effectively screen and identify the first anomaly node. Therefore, we propose an intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model, in order to solve the problem mentioned in the background art that the system fails to combine the topological characteristics of equipment components to carry out cross-modal hierarchical reasoning against propagation interference for multimodal operation data of industrial equipment, thus failing to achieve accurate component-level anomaly characterization extraction, fault propagation link reasoning, and initial anomaly node identification.

[0006] To address the aforementioned technical problems, the present invention aims to provide an intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model, comprising: The data acquisition and preprocessing unit acquires multimodal operating data corresponding to each monitoring component of the target industrial equipment, performs timestamp synchronization, outlier removal, numerical normalization, and feature standardization preprocessing on the acquired multimodal operating data corresponding to each monitoring component of the target industrial equipment, and outputs a standardized multimodal feature dataset that corresponds one-to-one with each monitoring component of the target industrial equipment and is aligned with the time dimension. The multimodal reasoning model reasoning unit receives a standardized multimodal feature dataset output by the data acquisition and preprocessing unit, and uses cross-modal hierarchical reasoning technology based on equipment component topology constraints to complete component-level association mapping of multimodal features, node-level cross-modal anomaly feature fusion based on topological neighborhood anomaly prior constraints, and fault propagation link reasoning under topological constraints. It outputs fault reasoning feature results including component-level anomaly representation, fault propagation link, and initial anomaly node candidate set. The fault intelligent diagnosis unit receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit, and matches and judges the component-level abnormal representations based on the pre-built industrial equipment component-level fault knowledge base and preset fault judgment thresholds. It completes the identification of fault types and the classification of fault severity levels for each monitored component of the target industrial equipment, and outputs standardized component-level fault diagnosis results that correspond one-to-one with each monitored component of the target industrial equipment. The fault root cause reasoning unit receives fault reasoning feature results output by the multimodal reasoning model reasoning unit and standardized component-level fault diagnosis results output by the fault intelligent diagnosis unit. Combining the fault propagation link, the candidate set of the first abnormal node, and the standardized component-level fault diagnosis results, it locates the first root cause node of the fault and the complete fault propagation path through fault causal propagation link tracing analysis, and outputs verifiable fault root cause location results.

[0007] As a further improvement to this technical solution, the data acquisition and preprocessing unit includes a multimodal operation data acquisition module, a multimodal data linkage preprocessing module, and a standardized feature generation and output module, wherein: The multimodal operation data acquisition module collects multimodal operation data corresponding to each monitoring component of the target industrial equipment, and outputs the collected multimodal operation data corresponding to each monitoring component of the target industrial equipment to the multimodal data linkage preprocessing module. The multimodal data linkage preprocessing module receives multimodal operation data output by the multimodal operation data acquisition module, performs timestamp synchronization, outlier removal, and numerical normalization preprocessing on the multimodal operation data in sequence, and outputs the preprocessed multimodal operation data corresponding to each monitoring component of the target industrial equipment to the standardized feature generation output module. The standardized feature generation and output module receives the preprocessed multimodal operating data, performs feature standardization preprocessing, and outputs a standardized multimodal feature dataset that corresponds one-to-one with each monitoring component of the target industrial equipment and is aligned with the time dimension.

[0008] As a further improvement to this technical solution, the multimodal reasoning model reasoning unit includes a device component topology graph construction module, a multimodal feature component-level association mapping module, a node-internal cross-modal anomaly feature fusion module, a fault propagation link reasoning module, and a reasoning result output module, wherein: The equipment component topology graph construction module constructs a directed weighted equipment component topology graph based on the component hierarchy and physical relationships of the target industrial equipment, and outputs the equipment component topology graph. The multimodal feature component-level association mapping module receives the standardized multimodal feature dataset output by the data acquisition and preprocessing unit and the equipment component topology map output by the equipment component topology map construction module, completes the component-level association mapping of multimodal features, and generates single-modal anomaly representations corresponding to each component node. The intra-node cross-modal anomaly feature fusion module receives the single-modal anomaly representation output by the multi-modal feature component-level association mapping module and the device component topology map output by the device component topology map construction module, and completes the intra-node cross-modal anomaly feature fusion based on the prior constraint of topological neighborhood anomaly, generating a component-level anomaly representation resistant to propagation interference. The fault propagation link reasoning module receives the device component topology map output by the device component topology map construction module and the anti-propagation interference component-level anomaly representation output by the node cross-modal anomaly feature fusion module, completes the fault propagation link reasoning under topological constraints, and generates the fault propagation link and the first set of anomaly node candidates. The reasoning result output module receives the anti-propagation interference component-level anomaly representation output by the intra-node cross-modal anomaly feature fusion module, the fault propagation link output by the fault propagation link reasoning module, and the initial anomaly node candidate set, and outputs the fault reasoning feature result containing the component-level anomaly representation, the fault propagation link, and the initial anomaly node candidate set.

[0009] As a further improvement to this technical solution, the directed weighted equipment component topology graph construction process of the equipment component topology graph construction module includes the following steps: S21.1 Define the equipment component topology graph as a directed weighted equipment component topology graph. ,in It is a set of nodes composed of various monitoring components of industrial equipment. It is the set of directed edges formed by the physical connections between components. This is the set of weight coefficients for each directed edge; S21.2. Based on the physical relationships of the target industrial equipment, determine the anomaly propagation capability coefficients corresponding to each directed edge in the directed weighted graph. ,in For upstream component nodes, For downstream component nodes; S21.3 Complete the construction and output of the directed weighted device component topology graph.

[0010] As a further improvement to this technical solution, the multimodal feature component-level association mapping and single-modal anomaly characterization generation process of the multimodal feature component-level association mapping module includes the following steps: S22.1 Obtain the standardized multimodal feature dataset output by the data acquisition and preprocessing unit; S22.2 Calculate the comprehensive matching degree between the standardized multimodal features and key component nodes. ,in To standardize multimodal features, For key component nodes; S22.3, Overall matching degree Greater than the matching threshold The standardized multimodal features are mapped to the corresponding key component nodes; S22.4. For standardized multimodal features mapped to the same key component node, generate single-modal anomaly representations according to modality type. ,in For the modal types of multimodal data, For timestamps.

[0011] As a further improvement to this technical solution, the intra-node cross-modal anomaly feature fusion module includes an upstream neighbor node set determination submodule, a fault propagation interference prior coefficient calculation submodule, a dynamic attention weighted fusion submodule, and a component-level anomaly representation standardization output submodule, wherein: The upstream neighbor node set determination submodule determines the upstream neighbor node set corresponding to the current critical component node based on the equipment component topology graph output by the equipment component topology graph construction module. The upstream neighbor node set is the set of all upstream nodes with directed edges pointing to the current critical component node, which perfectly matches the fault propagation direction. The fault propagation interference prior coefficient calculation submodule calculates the fault propagation interference prior coefficient corresponding to each mode of the current key component node based on the single-mode anomaly characterization of each node in the upstream neighboring node set, and quantifies the degree of interference of upstream fault propagation on the corresponding mode data of the current node. The dynamic attention weighted fusion submodule is based on the dynamic attention weighted fusion logic of the topological neighborhood anomaly prior constraint. With the fault propagation interference prior coefficient as the core constraint, it calculates the dynamic attention weight corresponding to each single-modal anomaly representation of the current critical component node, and performs intra-node cross-modal weighted fusion on each single-modal anomaly representation of the current critical component node based on the dynamic attention weight, to generate a component-level anomaly representation that only reflects the original anomaly state of the current node and is resistant to propagation interference. The component-level anomaly representation standardization output submodule receives the anti-propagation interference component-level anomaly representation output by the dynamic attention weighted fusion submodule, performs structured integration of the component-level anomaly representation, and outputs it to the fault propagation link inference module (24) under topological constraints.

[0012] As a further improvement to this technical solution, the dynamic attention weighted fusion process based on topological neighborhood anomaly prior constraints in the dynamic attention weighted fusion submodule includes the following steps: S23.1. Dimensionlessly normalize the single-modal anomaly characterization of the current key component nodes to obtain the normalized single-modal anomaly characterization. ; S23.2, Call the fault propagation interference prior coefficient calculation submodule to obtain the fault propagation interference prior coefficient. ; S23.3, using fault propagation interference prior coefficients To constrain the calculation, dynamic attention weights are computed. ; S23.4, Based on dynamic attention weights Characterization of normalized single-mode anomalies Weighted fusion is performed to generate component-level anomaly representations that are resistant to propagation interference. .

[0013] As a further improvement to this technical solution, the fault propagation link inference and initial abnormal node identification process under topological constraints of the fault propagation link inference module includes the following steps: S24.1 Constructing an anomaly propagation transition matrix based on the weight matrix of the equipment component topology graph. ; S24.2 Calculate the anomaly propagation consistency coefficient between adjacent component nodes. ; S24.3, The consistency coefficient of anomaly propagation Greater than the consistency threshold The propagation path was identified as a valid anomaly propagation path, and a directed subgraph of anomaly propagation was constructed. ; S24.4 Calculating the Directed Subgraph for Anomaly Propagation Based on Bayes' Theorem The posterior probability that each node is the first anomalous node ; S24.5, posterior probability Greater than the first-round probability threshold The node is identified as the first abnormal node and included in the candidate set of first abnormal nodes. At the same time, the effective abnormal propagation path is identified as the fault propagation link.

[0014] As a further improvement to this technical solution, the intelligent fault diagnosis unit includes a fault data receiving and knowledge base calling module, a fault matching and grading module, and a standardized diagnosis result output module, wherein: The fault data receiving and knowledge base calling module receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit, extracts the component-level anomaly representations, and simultaneously calls the pre-built industrial equipment component-level fault knowledge base to obtain the corresponding standard component fault features and fault severity level judgment rules. The fault matching and classification module, based on a preset fault judgment threshold, matches the extracted component-level anomaly representations with standard features in a pre-built industrial equipment component-level fault knowledge base to complete the identification of fault types and classification of fault severity levels for each monitored component of the target industrial equipment. The standardized diagnostic result output module integrates the judgment results of the fault matching and classification modules to generate and output standardized component-level fault diagnosis results that correspond one-to-one with each monitored component of the target industrial equipment.

[0015] As a further improvement to this technical solution, the fault root cause reasoning unit includes a multi-source diagnostic data receiving module, a fault causal source tracing and localization module, and a root cause result verification and output module, wherein: The multi-source diagnostic data receiving module synchronously receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit and the standardized component-level fault diagnosis results output by the fault intelligent diagnosis unit, and extracts the fault propagation link and the candidate set of the first abnormal node from the fault reasoning feature results. The fault cause tracing and localization module combines the extracted fault propagation link, the candidate set of the first abnormal node, and the standardized component-level fault diagnosis results to carry out fault cause propagation link tracing analysis and locate the first root cause node of the fault and the complete fault propagation path. The root cause result verification output module verifies the verifiability of the initial root cause node of the fault and the complete fault propagation path, and outputs verifiable fault root cause location results that meet the requirements.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a directed weighted topology graph of equipment components and uses it as the core constraint. First, it completes the component-level association mapping of multimodal features. Then, relying on dynamic attention-weighted fusion technology based on the prior constraint of topological neighborhood anomalies, it determines the upstream neighbor node set of the current component node and calculates the prior coefficient of fault propagation interference for each mode. This dynamically adjusts the attention weight of each single-mode anomaly representation and completes cross-modal weighted fusion within the node. At the same time, it constructs an anomaly propagation transition matrix based on the equipment component topology graph and combines Bayes' theorem to carry out fault propagation link reasoning and initial anomaly node screening. This enables anti-propagation interference cross-modal hierarchical reasoning of multimodal operation data of industrial equipment based on the topological characteristics of equipment components. It can extract component-level anomaly representations that only reflect the original abnormal state of the component itself, complete accurate reasoning of fault propagation links and effective identification of initial anomaly nodes, and adapt to the actual application scenarios of industrial equipment fault diagnosis. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system framework of the present invention; The meanings of the labels in the diagram are as follows: 1. Data acquisition and preprocessing unit; 11. Multimodal operation data acquisition module; 12. Multimodal data linkage preprocessing module; 13. Standardized feature generation and output module; 2. Multimodal reasoning model reasoning unit; 21. Equipment component topology diagram construction module; 22. Multimodal feature component-level association mapping module; 23. Intra-node cross-modal anomaly feature fusion module; 24. Fault propagation link reasoning module; 25. Reasoning result output module; 3. Intelligent fault diagnosis unit; 31. Fault data receiving and knowledge base retrieval module; 32. Fault matching judgment and classification module; 33. Standardized diagnostic result output module; 4. Fault root cause reasoning unit; 41. Multi-source diagnostic data receiving module; 42. Fault cause-and-effect tracing and location module; 43. Root cause result verification and output module. Detailed Implementation

[0018] The technical solutions of 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, this embodiment provides an intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model, including: The data acquisition and preprocessing unit 1 collects multimodal operational data corresponding to each monitored component of the target industrial equipment. It performs timestamp synchronization, outlier removal, numerical normalization, and feature standardization preprocessing on the collected multimodal operational data, outputting a standardized multimodal feature dataset that corresponds one-to-one with each monitored component of the target industrial equipment and is aligned with the time dimension. The data acquisition and preprocessing unit 1 includes a multimodal operational data acquisition module 11, a multimodal data linkage preprocessing module 12, and a standardized feature generation and output module 13, wherein: In this embodiment, the multimodal operation data acquisition module 11 acquires multimodal operation data corresponding to each monitoring component of the target industrial equipment, and outputs the acquired multimodal operation data corresponding to each monitoring component of the target industrial equipment to the multimodal data linkage preprocessing module 12.

[0020] Specifically, the multimodal operation data acquisition module 11 connects with the sensing and detection equipment and industrial control system of each monitoring component of the target industrial equipment through an industrial fieldbus to achieve protocol adaptation. The acquired multimodal operation data includes physical quantities of equipment operation such as vibration, temperature, pressure, current, voltage, and flow. During the acquisition process, each type of data is assigned a unique identifier for the monitoring component and an original acquisition timestamp, and the identifier corresponds one-to-one with the actual component code of the target industrial equipment. At the same time, the original data of different formats are uniformly converted into a structured data table format. The data table includes the monitoring component identifier, data mode type, original acquisition value, and original timestamp fields. After the conversion is completed, the structured multimodal operation data is output to the multimodal data linkage preprocessing module 12 in real time.

[0021] In this embodiment, the multimodal data linkage preprocessing module 12 receives the multimodal operation data output by the multimodal operation data acquisition module 11, performs timestamp synchronization, outlier removal, and numerical normalization preprocessing on the multimodal operation data in sequence, and outputs the multimodal operation data corresponding to each monitoring component of the target industrial equipment after preprocessing to the standardized feature generation output module 13.

[0022] Specifically, the implementation of timestamp synchronization, outlier removal, and numerical normalization is as follows: Timestamp synchronization uses the highest acquisition frequency of all acquisition devices as the reference time granularity, aligns the time axis of different modal data by grouping them according to the unique identifier of the monitored component, supplements the missing timestamp values ​​using linear interpolation, and uniformly converts the timestamps of all data into millisecond-level format. Outlier removal is grouped by monitoring component + modal type, and the 3σ principle is used in combination with the original factory rated operating threshold of the equipment for double judgment. Outlier values ​​that exceed the reasonable range are removed, and the neighborhood mean method is used to fill the numerical gaps after removal. Numerical normalization performs min-max normalization processing according to the data modality type, mapping the values ​​of each modality data to the [0,1] interval, eliminating the dimensional differences between different modal data, while retaining the maximum and minimum normalization parameters of each modality data. After completing all preprocessing steps, the data is output to the standardized feature generation output module 13.

[0023] In this embodiment, the standardized feature generation and output module 13 receives the preprocessed multimodal operating data, performs feature standardization preprocessing, and outputs a standardized multimodal feature dataset that corresponds one-to-one with each monitoring component of the target industrial equipment and is aligned with the time dimension. Specifically, the standardized feature generation output module 13 groups the monitoring components and modal types into units, and uses the Z-score standardization method to perform feature standardization on the preprocessed multimodal operating data. It calculates the mean and standard deviation of each group of data and transforms the data into standardized feature values ​​with a mean of 0 and a standard deviation of 1, eliminating the distribution differences of the same modality data among different monitoring components. Then, using the unique identifier of the monitoring component as the first-level index and the millisecond-level timestamp as the second-level index, it integrates all the standardized feature values ​​of each component under the same timestamp to generate a standardized multimodal feature dataset in the form of a two-dimensional structured matrix. The row dimension of the matrix is ​​the timestamp sequence, and the column dimension is the combined feature of monitoring component and modal type. Finally, the dataset is output to the multimodal inference model inference unit 2.

[0024] Multimodal reasoning model reasoning unit 2 receives the standardized multimodal feature dataset output by data acquisition and preprocessing unit 1, and adopts cross-modal hierarchical reasoning technology based on equipment component topology constraints to complete component-level association mapping of multimodal features, node-level cross-modal anomaly feature fusion based on topological neighborhood anomaly prior constraints, and fault propagation link reasoning under topological constraints. It outputs fault reasoning feature results including component-level anomaly representation, fault propagation link, and initial anomaly node candidate set. Specifically, the multimodal inference model inference unit 2 includes a device component topology graph construction module 21, a multimodal feature component-level association mapping module 22, a node-internal cross-modal anomaly feature fusion module 23, a fault propagation link inference module 24, and an inference result output module 25, wherein: In this embodiment, the equipment component topology graph construction module 21 constructs a directed weighted equipment component topology graph based on the component hierarchy and physical relationships of the target industrial equipment, and outputs the equipment component topology graph; the directed weighted equipment component topology graph construction process of the equipment component topology graph construction module 21 includes the following steps: S21.1 Define the equipment component topology graph as a directed weighted equipment component topology graph. ,in It is a set of nodes composed of various monitoring components of industrial equipment. It is the set of directed edges formed by the physical connections between components. This is the set of weight coefficients for each directed edge; this step is used to transform the physical structure of the target industrial equipment into a computable graph structure, providing objective physical constraints for subsequent cross-modal hierarchical reasoning. The specific implementation is as follows: The mathematical definition of a device component topology diagram is: ; in: This represents the set of nodes comprised of various monitoring components of industrial equipment. , The total number of monitored components for the target industrial equipment, per node For each monitoring component, its identifier corresponds one-to-one with the unique identifier of the monitoring component in the standardized multimodal feature dataset output by the data acquisition and preprocessing unit 1; This represents the set of directed edges formed by the physical connections between components. Each directed edge From upstream component nodes Pointing to downstream component nodes This indicates that the fault originated from an upstream component node. To its downstream component nodes The physical direction of transmission; This represents the set of weight coefficients for each directed edge. Each weight coefficient Corresponding to a directed edge Used to quantify upstream component nodes The anomaly is transmitted to its downstream component nodes. The strength of the transmission capability.

[0025] S21.2. Based on the physical relationships of the target industrial equipment, determine the anomaly propagation capability coefficients corresponding to each directed edge in the directed weighted graph. ,in For upstream component nodes, For downstream component nodes; this step is used to quantify the anomaly propagation strength between upstream and downstream components, improve the shortcomings of subjective weight assignment in traditional topology graphs, and provide a reproducible physical basis for subsequent feature fusion and fault propagation reasoning under topological constraints. The specific implementation is as follows: Abnormal transmission capability coefficient Used to quantize upstream component nodes The anomaly is transmitted to its downstream component nodes. The strength of the transmission capability is determined by two categories based on whether there is a physical connection between components, as described below: First, there is no direct physical connection: If upstream component node With downstream component nodes If there is no direct physical connection, the anomaly propagation capability coefficient is set to: ; in: Indicates upstream component node To its downstream component nodes The abnormal transmission capability coefficient; This represents the upstream component node of a directed edge, which is the starting component for fault propagation; This represents the downstream component node of a directed edge, which is the receiving component for fault propagation.

[0026] Secondly, there is a direct physical connection: If upstream component node With downstream component nodes If a direct physical correlation exists, the anomaly transmission capability coefficient is calculated jointly by the physical correlation tightness coefficient and the historical fault transmission frequency coefficient. The calculation formula is as follows: ; in: Indicates upstream component node To its downstream component nodes The abnormal transmission capability coefficient has a value range of (0,1]. and The values ​​must be determined based on objective data from the original equipment manufacturer's design documents and historical fault database; The coefficient represents the degree of physical connection, with a value range of (0,1]. It is determined based on the physical connection method of the target industrial equipment: direct drive connection refers to the connection method that directly transmits power or motion through mechanical transmission components such as gears, shafts, and couplings, with a value of 0.9; medium transmission connection refers to the connection method that transmits energy or signals through media such as fluids, gases, and electricity, with a value of 0.6; indirect contact connection refers to the connection method that indirectly transmits force or influence through intermediate components, housings, etc., with a value of 0.3. This represents the historical fault propagation frequency coefficient, with a value range of (0,1], obtained based on historical fault cases of the target industrial equipment. Multiple propagation refers to a fault propagation scenario that simultaneously meets both of the following conditions in statistically documented historical fault cases: Upstream component nodes The fault is transmitted to downstream component nodes. The frequency of this fault propagation is higher than the average frequency of all other fault propagation paths of this upstream node (upstream node). The frequency of fault propagation of all direct downstream nodes, i.e., statistical to each of its direct downstream nodes The frequency of fault transmission (calculated as the average of the frequency group as the baseline average frequency) is 3 times or more; In all historical fault propagation cases at the upstream node, this propagation path accounts for no less than 60% of the total, and it is a typical fault propagation path clearly defined in the original equipment manufacturer's maintenance manual; when the above conditions are met... The value is between 0.8 and 1.0; among them, the statistically significant historical failure cases (failure cases of the same type of equipment cluster) refer to complete equipment failure operation and maintenance records with a sample size of not less than 50 and a time span of not less than 12 months, to ensure that the statistical results are statistically significant.

[0027] Intermittent transmission refers to fault transmission that occurs only under a few specific operating conditions or specific fault types, without high-frequency statistical characteristics, and its value ranges from 0.2 to 0.7. Never transmitted means that there are no upstream component nodes in the historical fault database. The fault is transmitted to downstream component nodes. The statistically available cases have a value of 0.1. S21.3 Complete the construction and output of the directed weighted device component topology graph.

[0028] Specifically, the device component topology diagram construction module 21 outputs the completed directed weighted device component topology diagram in structured JSON format, and the output content includes: Node set Fields: Unique identifier for each node, corresponding monitoring component name, and associated monitoring mode type; Directed edge set Field: The starting node of each directed edge Identifier, Termination Node Identifier and physical association type description; Weight coefficient set Field: Anomaly propagation capability coefficient for each directed edge. The numerical output format is fully matched with the input interface of the multimodal feature component-level association mapping module 22, ensuring smooth data flow while retaining the editability of the topology graph, which facilitates subsequent updates of nodes, edges and weights based on equipment upgrades or monitoring range adjustments.

[0029] In this embodiment, the multimodal feature component-level association mapping module 22 receives the standardized multimodal feature dataset output by the data acquisition and preprocessing unit 1 and the equipment component topology map output by the equipment component topology map construction module 21, completes the component-level association mapping of multimodal features, and generates single-modal anomaly representations corresponding to each component node; the multimodal feature component-level association mapping and single-modal anomaly representation generation process of the multimodal feature component-level association mapping module 22 includes the following steps: S22.1 Obtain the standardized multimodal feature dataset output by data acquisition and preprocessing unit 1 to provide complete and standardized basic input data for subsequent comprehensive matching degree calculation, ensuring the compatibility of feature data with device component topology diagram. The specific implementation is as follows: The multimodal feature component-level association mapping module 22 reads the standardized multimodal feature dataset output by the data acquisition and preprocessing unit 1 through a preset standardized data interface. This dataset uses "unique identifier of monitoring component" as the first-level index and "millisecond-level timestamp" as the second-level index, and contains multimodal standardized feature values ​​such as vibration, temperature, and pressure. After reading, the dataset is checked for completeness to confirm that all timestamps and monitoring component identifiers correspond one-to-one with the node identifiers of the equipment component topology map output by the equipment component topology map construction module 21. If there are missing or mismatched items, an alarm message is output. After the verification is passed, the process proceeds to the next step.

[0030] S22.2 Calculate the comprehensive matching degree between standardized multimodal features and key component nodes. ,in To standardize multimodal features, This step identifies key component nodes. It quantifies the physical correlation and numerical similarity between multimodal features and their corresponding component nodes, providing a quantifiable basis for subsequent feature mapping. The specific implementation is as follows: Overall matching degree The calculation formula is: ; in: Indicates the first The standardized multimodal feature and the first The overall matching degree of each key component node, with a value range of [0,1]; The feature-component physical association weights, with values ​​ranging from (0,1), are determined based on the equipment component topology diagram output by the equipment component topology diagram construction module 21: if the first... The multimodal feature acquisition sensor is directly installed on the first Each key component node, If the sensor is installed at the same location as the first... Components that are directly physically associated with key component nodes, then If the sensor is installed on an indirectly connected component, then ; Indicates the first A standardized multimodal feature With the Reference features of key component nodes The Pearson correlation coefficient, after taking the absolute value, is mapped to the [0,1] interval to quantify the numerical similarity between the feature and the reference feature of the component; Indicates the first Time series feature values ​​of standardized multimodal features; Indicates the first The time series of full-scale standardized multimodal features of key component nodes under historical normal operation conditions, with the time length corresponding to the standardized multimodal features to be matched. To maintain consistency, this reference feature is statistically derived from the standardized multimodal feature dataset under historical normal operating conditions output by data acquisition and preprocessing unit 1. The dataset must meet the following requirements: Duration: Not less than the total duration of 30 complete operating cycles of continuous normal operation of the equipment, covering the stable operation phase of the entire life cycle of the equipment; Operating conditions are limited to: all conventional production and operation conditions, including rated load conditions, variable load transition conditions, and normal start-up and shutdown conditions, excluding characteristic data under unstable production conditions such as equipment debugging, maintenance, no-load trial operation, and fault shutdown. Data validity: Remove invalid segments with missing data or anomalous jumps, ensuring that the continuous valid duration of the dataset accounts for no less than 95%; A unique identifier representing standardized multimodal features; The unique identifier representing the key component node, and the set of device component topology graph nodes output by device component topology graph construction module 21. The node identifiers in the table correspond one-to-one.

[0031] S22.3, Overall matching degree Greater than the matching threshold The standardized multimodal features are mapped to the corresponding key component nodes. This step is used to achieve accurate association between multimodal features and component nodes, avoiding abnormal representation distortion caused by feature mismatch. The specific implementation is as follows: Multimodal feature component-level association mapping module 22 preset matching degree threshold This threshold is determined by performing ROC curve analysis on historical feature mapping cases to select the optimal threshold that balances the false alarm rate and the missed alarm rate, or it can be set to 0.6 based on the equipment monitoring accuracy requirements. An empirical value between 0.8; iterate through all combinations of standardized multimodal features and key component nodes, and if the overall matching degree... Then the first The standardized multimodal feature is mapped to the first... For each key component node, establish a unique "feature-component node" association; if the comprehensive matching degree of the same standardized multimodal feature with multiple key component nodes is greater than 100%, then... Then select the overall matching degree. The key component nodes corresponding to the maximum values ​​are mapped to ensure that each standardized multimodal feature belongs to only one key component node.

[0032] S22.4. For standardized multimodal features mapped to the same key component node, generate single-modal anomaly representations according to modality type. ,in For the modal types of multimodal data, This is a timestamp. This step is used to provide a unified format of component-level single-modal anomaly input for the intra-node cross-modal anomaly feature fusion module 23, specifically implemented as follows: Single-mode anomaly characterization The calculation formula is: ; in: Indicates the first The key component node, the first The modal type, the first The single-modal anomaly representation corresponding to each timestamp, with values ​​ranging from [value range missing]. ; Indicates mapping to the first The first key component node The modal type, the first The standardized multimodal feature values ​​corresponding to each timestamp; Indicates the first The key component node, the first The historical mean of features for each modality type is obtained by statistically analyzing the standardized multimodal feature dataset under historical normal operating conditions output by data acquisition and preprocessing unit 1, excluding feature data under fault conditions. Among them, fault condition refers to the equipment operating condition that meets any of the following conditions: fault alarm condition or protection shutdown condition clearly stated in the original equipment manufacturer's operation and maintenance manual; fault operating condition that has been confirmed in the historical operation and maintenance records to have caused component damage, functional failure, or abnormal shutdown; abnormal operating condition where the single modal feature value of the corresponding component exceeds the 3σ interval of the historical normal operating condition feature for more than 10 consecutive sampling periods. Indicates the first The key component node, the first The historical standard deviation of each modality type is obtained by statistical analysis of the standardized multimodal feature dataset under historical normal operating conditions output by data acquisition and preprocessing unit 1, excluding feature data under fault conditions; A unique identifier representing a critical component node; It indicates the mode type of multimodal data, such as vibration mode, temperature mode, pressure mode, etc. The timestamp corresponds one-to-one with the millisecond-level timestamps in the standardized multimodal feature dataset output by data acquisition and preprocessing unit 1.

[0033] when Time (of which) The threshold is determined based on the physical properties of the equipment, targeting vibration characteristics. Targeting temperature characteristics , The values ​​are determined from the original equipment manufacturer's maintenance manual or the physical limit calibration at the industrial site. Cut off as ;when When that happens, retain the original calculated value.

[0034] After generation, the multimodal feature component-level association mapping module 22 represents all single-modal anomalies. The data is integrated into a structured dataset according to the dimensions of "key component node - modality type - timestamp" and output to the cross-modal anomaly feature fusion module 23 within the node.

[0035] In this embodiment, the intra-node cross-modal anomaly feature fusion module 23 receives the single-modal anomaly representation output by the multi-modal feature component-level association mapping module 22 and the device component topology map output by the device component topology map construction module 21, and completes the intra-node cross-modal anomaly feature fusion based on the prior constraint of topological neighborhood anomalies, generating a component-level anomaly representation resistant to propagation interference; the intra-node cross-modal anomaly feature fusion module 23 includes an upstream neighbor node set determination submodule, a fault propagation interference prior coefficient calculation submodule, a dynamic attention weighted fusion submodule, and a component-level anomaly representation standardization output submodule, wherein: The upstream neighbor node set determination submodule determines the upstream neighbor node set corresponding to the current critical component node based on the equipment component topology graph output by the equipment component topology graph construction module 21. The upstream neighbor node set is the set of all upstream nodes that have directed edges pointing to the current critical component node (only direct upstream nodes are included, and indirect upstream nodes are excluded), which is completely matched with the fault propagation direction.

[0036] Specifically, the upstream neighbor node set determination submodule traverses the device component topology graph output by the device component topology graph construction module 21, which is the set of directed edges. For the current key component nodes Filter all edges that satisfy "directed edge" downstream nodes For current key component nodes "directed edge The upstream nodes corresponding to these directed edges Summarize and form the current key component nodes. upstream neighbor node set ,in The number of upstream neighboring nodes, each upstream node All are sets of device component topology graph nodes output by the device component topology graph construction module 21. The nodes in, and related to the current critical component nodes. There are fault propagation paths that are directly physically related.

[0037] The fault propagation interference prior coefficient calculation submodule calculates the fault propagation interference prior coefficients corresponding to each mode of the current critical component node based on the single-mode anomaly characterization of each node in the upstream neighboring node set, and quantifies the degree of interference of upstream fault propagation on the corresponding mode data of the current node.

[0038] Specifically, the fault propagation interference prior coefficient calculation submodule targets the current critical component nodes. , No. The modal type, the first A timestamp, based on the set of upstream neighboring nodes. Each upstream node Single-mode anomaly characterization The anomaly propagation capability coefficient output by the device component topology construction module 21 Calculate the prior coefficient of fault propagation interference. The calculation formula is: ; in: Indicates the current critical component node , No. The modal type, the first The prior coefficient of fault propagation interference corresponding to each timestamp has a value range of [0,+∞). Indicates the current critical component node The upstream neighboring nodes belong to the set of upstream neighboring nodes. ; This represents the upstream node output by the device component topology graph construction module 21. To current critical component nodes The abnormal transmission capability coefficient has a value range of [0,1]. The upstream node represented by the multimodal feature component-level association mapping module 22 , No. The modal type, the first Single-modal anomaly representation corresponding to each timestamp; A unique identifier representing the current critical component node; Indicates the modality type of multimodal data; Represents a timestamp.

[0039] The dynamic attention weighted fusion submodule is based on the dynamic attention weighted fusion logic of topological neighborhood anomaly prior constraints. Taking the fault propagation interference prior coefficient as the core constraint, it calculates the dynamic attention weights corresponding to the single-modal anomaly representations of the current critical component nodes. The dynamic attention weighted fusion process of the dynamic attention weighted fusion submodule based on topological neighborhood anomaly prior constraints includes the following steps: S23.1. Dimensionlessly normalize the single-modal anomaly characterization of the current key component nodes to obtain the normalized single-modal anomaly characterization. .

[0040] Specifically, normalized single-modal anomaly characterization The calculation formula is: ; in: Indicates the current critical component node , No. The modal type, the first The normalized single-modal anomaly representation corresponding to each timestamp has a value range of [0,1]. This represents the current key component node output by the multimodal feature component-level association mapping module 22. , No. The modal type, the first Single-modal anomaly representation corresponding to each timestamp; Indicates the current critical component node In the The maximum absolute value of the single-modal anomaly representation for all modal types under each timestamp; A unique identifier representing the current critical component node; Indicates the modality type of multimodal data; Represents a timestamp; Indicates the current critical component node In timestamp All modal types under.

[0041] S23.2, Call the fault propagation interference prior coefficient calculation submodule to obtain the fault propagation interference prior coefficient. .

[0042] Specifically, this step only applies to the current critical component nodes. The set of direct upstream neighboring nodes Within the nodes, the fault propagation interference prior coefficient calculation submodule is invoked, based on each upstream node. Single-mode anomaly characterization The anomaly propagation capability coefficient output by the device component topology construction module 21 Perform a weighted summation calculation to obtain the current key component node. , No. The modal type, the first The prior coefficient of fault propagation interference corresponding to each timestamp The specific calculation formula and parameter definition of this coefficient are completely consistent with the calculation submodule of the prior coefficient of fault propagation interference.

[0043] S23.3, using fault propagation interference prior coefficients To constrain the calculation, dynamic attention weights are computed. .

[0044] Specifically, dynamic attention weights The calculation employs Softmax weighted logic based on topological neighborhood anomaly prior constraints, using fault propagation interference prior coefficients. As a core constraint, the weights of the normalized single-mode anomaly representations for each mode are adaptively adjusted to ensure that modes with higher levels of disturbance have lower corresponding weights. The calculation formula is as follows: ; in: Indicates the current critical component node , No. The modal type, the first The dynamic attention weights corresponding to each timestamp take values ​​in the range [0,1], and satisfy the following conditions: ; This represents the interference suppression coefficient, with a value range of (0, +∞). It is determined based on the fault propagation intensity and monitoring sensitivity requirements of the target industrial equipment: the higher the fault propagation intensity, the lower the required monitoring sensitivity. The larger the value, the lower the fault propagation intensity and the higher the required monitoring sensitivity. The smaller the value, the more specific the quantification rules are as follows: Scenarios requiring high fault propagation intensity and low monitoring sensitivity: This applies to strongly coupled component groups such as mechanical direct-drive transmission chains and media series pipelines, with a historical fault propagation rate ≥70%, and where the monitored object is a non-core safety component, allowing for short-term delayed alarms. The value ranges from 4.0 to 5.0; Medium fault propagation intensity and medium monitoring sensitivity requirements: This applies to component groups with indirect contact connections and indirect media transmission, with a historical fault propagation rate of 30%–70%, and where the monitoring targets are conventional production components and have conventional alarm requirements. The value ranges from 2.0 to 3.0; Scenarios requiring low fault propagation intensity and high monitoring sensitivity: These scenarios correspond to component groups with no direct physical connection but only spatial coupling, a historical fault propagation rate of <30%, and where the monitoring targets are core safety components, critical transmission components, and require early warning. The value ranges from 1.0 to 2.0; Among them, the historical fault propagation rate refers to the ratio of the number of historical fault propagations of the corresponding component group to the total number of faults of the upstream components; Indicates the current critical component node In the All modal types under each timestamp The sum of these is used to normalize the attention weights; S23.4, Based on dynamic attention weights Characterization of normalized single-mode anomalies Weighted fusion is performed to generate component-level anomaly representations that are resistant to propagation interference. .

[0045] Specifically, component-level anomaly characterization to resist propagation interference. By analyzing the current key component nodes Normalized single-modal anomaly representation for all modal types According to dynamic attention weights We obtain the result by weighted summation, using the following formula: ; in: Indicates the current critical component node , No. The component-level anomaly representation corresponding to each timestamp, with a value range of [0,1]; The dynamic attention-weighted fusion submodule represents the component-level anomalies of all key component nodes. The dataset is integrated into a structured dataset according to the "key component node - timestamp" dimension and output to the fault propagation link reasoning module 24.

[0046] The component-level anomaly representation standardization output submodule receives the anti-propagation interference component-level anomaly representation output by the dynamic attention weighted fusion submodule, performs structured integration of the component-level anomaly representation, and outputs it to the fault propagation link inference module 24 under topological constraints.

[0047] Specifically, the component-level anomaly representation standardization output submodule receives, in real time, the anti-propagation interference component-level anomaly representations of all key component nodes output by the dynamic attention-weighted fusion submodule through a preset standardized data interface. ; Furthermore, the component-level anomaly representation standardization output submodule traverses all received... The data is structured and integrated according to the two-dimensional dimension of "key component node - timestamp" to form a standardized dataset: Row dimension: in millisecond timestamps The sequence is arranged in chronological order and is perfectly aligned with the timestamp sequence output by data acquisition and preprocessing unit 1. Column dimension: identified by key component nodes For the sequence, and the set of device component topology graph nodes output by device component topology graph construction module 21. The node identifiers in the data are in the exact same order; Matrix element: the line, number The elements of the column correspond to the current timestamp. Key component nodes Component-level anomaly characterization .

[0048] The integrated dataset is in the form of a two-dimensional matrix, ensuring that the data structure is fully compatible with the input interface of the fault propagation link reasoning module 24 under topological constraints.

[0049] Finally, after the structured integration is completed, the component-level anomaly representation standardization output submodule outputs the standardized dataset in a preset format (such as JSON, CSV or binary matrix format) to the fault propagation link reasoning module 24 under topological constraints; the output data includes a complete timestamp sequence, a key component node identifier sequence and the corresponding component-level anomaly representation values, ensuring that downstream modules can directly read and execute fault propagation link reasoning operations without additional data preprocessing.

[0050] In this embodiment, the fault propagation link reasoning module 24 receives the device component topology map output by the device component topology map construction module 21 and the anti-propagation interference component-level anomaly representation output by the intra-node cross-modal anomaly feature fusion module 23, completes the fault propagation link reasoning under topological constraints, and generates the fault propagation link and the candidate set of the first abnormal node. The fault propagation link reasoning and first abnormal node identification process under topological constraints of the fault propagation link reasoning module 24 includes the following steps: S24.1 Constructing an anomaly propagation transition matrix based on the weight matrix of the equipment component topology graph. .

[0051] Specifically, the anomaly propagation transition matrix The anomaly propagation capability coefficient weight matrix output by the device component topology diagram construction module 21 Normalization yields the result, used to quantify anomalies originating from upstream component nodes. to each downstream component node Probability of propagation, elements of the transition matrix The calculation formula is: ; in: This indicates an anomaly originating from an upstream component node. Downstream component nodes The propagation transition probability takes values ​​in the range [0,1], and is valid for nodes with downstream nodes. ,satisfy ; This represents the upstream component nodes output by the device component topology graph construction module 21. Downstream component nodes The abnormal transmission capability coefficient has a value range of [0,1]. Indicates the connection between the device component topology graph and the upstream component node. A set of downstream component nodes that have direct physical connections; Indicates upstream component node All direct downstream nodes.

[0052] S24.2 Calculate the anomaly propagation consistency coefficient between adjacent component nodes. Consistency coefficient of anomaly propagation Used to quantize adjacent component nodes (Upstream) and (Downstream) at timestamp The degree of matching between the abnormal state and the probability of abnormal propagation and transition is calculated using the following formula: ; in: Indicates adjacent component nodes and In timestamp The anomaly propagation consistency coefficient is set to [0,1]. A higher value indicates greater consistency. abnormal direction The stronger the consistency of the communication; For extremely small positive numbers, the value must be much less than 1 to avoid affecting... / The ratio, typically taking the value of This is used to avoid extreme operating conditions where the denominator is 0.

[0053] S24.3, The consistency coefficient of anomaly propagation Greater than the consistency threshold The propagation path was identified as a valid anomaly propagation path, and a directed subgraph of anomaly propagation was constructed. .

[0054] Specifically, a preset consistency threshold is set. The value range is (0,1), determined based on the fault diagnosis accuracy requirements of the target industrial equipment and historical fault propagation cases. The optimal threshold, where the false negative rate and false positive rate are balanced, can be selected through ROC curve analysis of historical fault cases; a typical value is 0.4-0.6. All directed edges of the equipment component topology graph are traversed. ,like Then the directed edge Valid anomaly propagation paths were identified; based on all valid anomaly propagation paths, a directed subgraph of anomaly propagation was constructed. ,in: The set of nodes representing the directed subgraph of anomaly propagation consists of all component nodes that participate in the valid anomaly propagation path. ; Let the set of directed edges of the directed subgraph of anomaly propagation be defined by all edges satisfying the condition... Effective abnormal propagation path The structure is such that each directed edge retains its original direction and corresponding orientation. As edge weight; The original device component topology graph retains its directionality, and only includes nodes and edges with valid anomaly propagation, for subsequent inference of the first anomaly node.

[0055] S24.4 Calculating the Directed Subgraph for Anomaly Propagation Based on Bayes' Theorem The posterior probability that each node is the first anomalous node .

[0056] Specifically, according to Bayes' theorem, nodes The formula for calculating the posterior probability of the first anomalous node is: ; in: This indicates the observed anomalous propagation of the directed subgraph. Under the premise that the node The posterior probability of the first abnormal node is denoted as [0,1]. Represents a node The prior probability of the first abnormal node is determined based on historical fault statistics of the target industrial equipment. If historical data is unavailable, a uniform distribution is used. ( for (Number of nodes); Indicates hypothetical nodes When the initial abnormal node is observed, a directed subgraph of abnormal propagation is observed. The likelihood probability is calculated using the following formula: ; In the formula: Represents a node Directed subgraphs in anomalous propagation The in-degree in the middle, that is, the direction of the direction. The number of effective propagation edges; Directed subgraph representing anomaly propagation The maximum in-degree of all nodes in the set; Directed subgraph representing anomaly propagation All from Starting from the beginning, the set of effective propagation edges reachable along directed edges. This represents the total number of elements in the set, i.e., from The number of effective propagation edges that can be reached from the starting point; Directed subgraph representing anomaly propagation The set consisting of all valid propagation edges in the middle. This represents the total number of elements in the set, i.e., the directed subgraph for anomaly propagation. The total number of all valid propagation edges in the process; express The arithmetic mean of the anomaly propagation consistency coefficients of all edges in the equation; if If it is an empty set, then Value is a node Its own component-level anomaly representation ; It is a very small positive number, and its value is [value]. This is used to avoid extreme operating conditions where the denominator is 0; As a normalization factor, it ensures that the sum of the posterior probabilities of all nodes is 1; Directed subgraph representing anomaly propagation Unique identifier for nodes within; Represents a timestamp.

[0057] S24.5, posterior probability Greater than the first-round probability threshold The node is identified as the first abnormal node and included in the candidate set of first abnormal nodes. At the same time, the effective abnormal propagation path is identified as the fault propagation link.

[0058] Specifically, a preset first-launch probability threshold is set. The value ranges from (0,1) and is determined based on the confidence requirements for fault diagnosis of the target industrial equipment. It can be selected by analyzing the Precision-Recall curves of historical first-time anomaly node identification cases to find the optimal threshold that meets the diagnostic confidence requirements; a typical value is 0.6-0.8. The directed subgraph of anomaly propagation is then traversed. All nodes within ,like Then It was identified as the first abnormal node and included in the candidate set of first abnormal nodes. ; At the same time, all valid anomaly propagation paths determined in step S24.3 (i.e. The directed edges in the data are used to determine the current timestamp. The fault propagation link below fully reflects the path of the anomaly from the initial abnormal node to other component nodes; finally, the fault propagation link inference module 24 will select the initial abnormal node candidate set. The fault propagation link outputs to the fault diagnosis and decision-making module for subsequent fault handling and early warning.

[0059] In this embodiment, the inference result output module 25 receives the anti-propagation interference component-level anomaly representation output by the intra-node cross-modal anomaly feature fusion module 23, the fault propagation link and the first anomaly node candidate set output by the fault propagation link inference module 24, and outputs the fault inference feature result containing the component-level anomaly representation, the fault propagation link and the first anomaly node candidate set.

[0060] Specifically, the inference result output module 25 receives three types of core input data through a preset standardized data interface and performs integrity checks: Receive component-level anomaly characterization data: Obtain the anti-propagation interference component-level anomaly characterization output by the cross-modal anomaly feature fusion module 23 within the node. The data is a two-dimensional structured matrix with the dimension of "key component node - timestamp". The row dimension is a millisecond-level timestamp sequence, and the column dimension is a unique identifier of the key component node. Receive fault propagation link data: Obtain the fault propagation link output by the fault propagation link inference module 24. This data is a set of directed edges, and each directed edge contains the identifier of its upstream node. Downstream node identifier Consistency coefficient with the corresponding anomaly propagation ; Receive the candidate set of the first abnormal node: Obtain the candidate set of the first abnormal node output by the fault propagation link inference module 24. This data is a set of key component node identifiers, containing the unique identifiers of all component nodes that were identified as the first abnormal node; Verification rules: Verify that the timestamp units and key component node identifiers of all input data are fully aligned with the output of the front-end module. If there are missing data, incorrect format, or mismatched identifiers, immediately discard the current batch of data and output an exception verification log until the data verification passes before proceeding with the subsequent processing.

[0061] Furthermore, the inference result output module 25 uses timestamps. Using a unique core index, and following the rule of processing frame by frame with a single timestamp and continuous batch output, the three types of data under the same timestamp are dimension-bound and structurally integrated to construct a unified fault reasoning feature dataset. The integration rules are as follows: Time dimension alignment: Traverse all consecutive millisecond-level timestamps , will the same corresponding The association is a single fault reasoning feature entry, ensuring that the time series data is continuous and without gaps; Structured field encapsulation: Component-level anomaly representation field: Encapsulates the anomaly status of each key component node under the current timestamp, stored in key-value pairs, with unique component node identifiers, and the anomaly representation value range is limited to [0,1]. Fault propagation link field: Encapsulates all valid anomaly propagation paths under the current timestamp, stored in the form of triples, including upstream component node identifier, downstream component node identifier, and corresponding anomaly propagation consistency coefficient, maintaining the topology propagation direction; First abnormal node candidate set field: Encapsulates the unique identifier of all first abnormal nodes under the current timestamp, stored in the form of a one-dimensional array, with no duplicate identifiers; Sequence continuity guarantee: The integrated dataset is kept completely synchronized with the timestamp sequence output by the data acquisition and preprocessing unit 1, with no time duplication or temporal disorder.

[0062] Finally, the inference result output module 25 encapsulates the integrated fault inference feature dataset into an industrial-grade general-purpose structured data format, the format definition of which includes the following core fields: Timestamp field: Records the millisecond-level timestamp corresponding to the current data, which is completely consistent with the timestamp of the input data; Component-level anomaly characterization field: Stores the unique identifier of each key component node and its corresponding anomaly characterization value; Fault propagation link field: Stores the node identifiers and consistency coefficients of all valid anomaly propagation paths; First abnormal node candidate set field: Stores a unique list of all first abnormal nodes.

[0063] After encapsulation, the inference result output module 25 outputs the standardized fault inference feature results to the preset downstream functional modules according to the rule of continuous frame-by-frame output. The downstream functional modules include a fault visualization module, a fault decision and early warning module, and a historical data storage module. Each module is an essential component of system fault analysis and operation and maintenance decision-making, and their specific positions are as follows: The fault visualization module is used to transform the fault reasoning feature results into a visualization interface, rendering component-level anomaly heatmaps, fault propagation topology maps, and initial anomaly node markers to achieve an intuitive display of fault status and propagation path; The fault decision and early warning module triggers hierarchical alarms and generates corresponding operation and maintenance instructions based on the abnormal representation values ​​in the fault reasoning feature results and the first node candidate set. The historical data storage module is used to persistently store the fault reasoning feature results to the time series database, providing data support for subsequent fault backtracking, root cause analysis and algorithm model iteration optimization.

[0064] The inference result output module 25 outputs the standardized fault inference feature results to the aforementioned preset downstream functional modules according to the rule of continuous frame-by-frame output. This ensures complete compatibility with the input interfaces of each downstream module, guaranteeing smooth data flow and efficient utilization within the system. Specifically: It interfaces with the fault visualization module to provide core data support for its visualization rendering; It interfaces with the fault decision-making and early warning module to provide quantitative data for its hierarchical alarm and operation and maintenance instruction generation; It interfaces with the historical data storage module to provide a standardized input format for the persistence of fault data and subsequent analysis.

[0065] The intelligent fault diagnosis unit 3 receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit 2, and, based on a pre-built industrial equipment component-level fault knowledge base and preset fault judgment thresholds, matches and judges the component-level anomaly representations to complete the identification of fault types and fault severity levels of each monitored component of the target industrial equipment, and outputs standardized component-level fault diagnosis results corresponding one-to-one with each monitored component of the target industrial equipment. The intelligent fault diagnosis unit 3 includes a fault data receiving and knowledge base calling module 31, a fault matching and judgment and classification module 32, and a standardized diagnosis result output module 33, wherein: The fault data receiving and knowledge base calling module 31 receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit 2, extracts the component-level anomaly representations, and simultaneously calls the pre-built industrial equipment component-level fault knowledge base to obtain the corresponding standard component fault features and fault severity level judgment rules.

[0066] Specifically, the fault data receiving and knowledge base calling module 31 receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit 2 through a preset standardized data interface, extracts three types of core data: millisecond-level timestamps, unique identifiers of key components, and anti-propagation interference component-level anomaly representations, and simultaneously stores the fault propagation link and the first abnormal node candidate set for subsequent auxiliary tracing. This part of the data does not participate in the current fault judgment process. Simultaneously, a pre-built industrial equipment component-level fault knowledge base is invoked. This knowledge base uses the unique identifier of the component as the core index and stores it in key-value pairs. It is built based on the historical fault data of the target equipment and the component operation and maintenance manual. It stores the standard fault types, common abnormal feature ranges and fault severity level judgment rules corresponding to each component. The fault data receiving and knowledge base invocation module 31 retrieves the matching knowledge base data according to the extracted component identifier and transmits it to the fault matching judgment and classification module 32.

[0067] The fault matching and classification module 32, based on a preset fault judgment threshold, matches the extracted component-level anomaly representations with the standard features in the pre-built industrial equipment component-level fault knowledge base to complete the identification of fault types and the classification of fault severity levels for each monitored component of the target industrial equipment.

[0068] Specifically, first, preset the threshold configuration: Global fault determination threshold The value range is (0,1), with a typical value of 0.3. 0.5, determined based on the monitoring sensitivity requirements of the target industrial equipment, can be calibrated through ROC curve analysis of historical fault samples; Fault severity level classification threshold: (Minor fault lower limit, typically 0.3) (Lower limit for general faults, typically 0.5) (Lower limit of severe fault, typically 0.7) (Critical fault lower limit, typically 0.9), all thresholds are stored in a dynamically configurable parameter file and can be adjusted online.

[0069] Furthermore, for each key component node timestamp Execute the following matching logic: like : Directly determine the component In timestamp If there is no fault, mark the fault type as "normal". like Traversing the knowledge base Corresponding standard fault type set ,Will Standard abnormal characteristic threshold ranges for each fault type Perform interval matching, with the following matching rules: ,in To avoid a denominator of 0, select The most significant fault type will be used as the final matching result; if multiple fault types exist... Faults of the same type are grouped and labeled as "compound faults".

[0070] Finally, based on component-level anomaly representation Based on preset grading thresholds, the severity level of the fault is classified: Minor fault: ; Common faults: ; Critical fault: ; Critical fault: ; like If so, the severity level will be marked as "none".

[0071] Encapsulation of judgment results: Encapsulating each key component node timestamp Corresponding "part identification" timestamp Abnormal characteristics Fault type, fault severity level, and matching confidence level The result is encapsulated as a temporary judgment entry and passed to the standardized diagnostic result output module 33.

[0072] The standardized diagnostic result output module 33 integrates the judgment results of the fault matching and classification module 32, generates and outputs standardized component-level fault diagnosis results that correspond one-to-one with each monitored component of the target industrial equipment. Specifically, the standardized diagnostic result output module 33 uses the timestamp-component unique identifier as the unique association index, integrates the judgment data of the fault matching and classification module 32, removes redundant information, and generates standardized fault diagnosis results that correspond one-to-one with each monitored component of the target industrial equipment. Each result includes core fields such as timestamp, component identifier, abnormal characterization value, fault type, and fault severity level.

[0073] After integration, standardized fault diagnosis results will be output for subsequent operation and maintenance display, and stored in the historical fault database to provide data support for the iterative optimization of the fault knowledge base and the verification of fault source. The output format will be adapted to the subsequent data docking requirements of the system to ensure normal data flow.

[0074] The fault root cause reasoning unit 4 receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit 2 and the standardized component-level fault diagnosis results output by the fault intelligent diagnosis unit 3. Combining the fault propagation link, the candidate set of initial abnormal nodes, and the standardized component-level fault diagnosis results, it locates the initial root cause node and the complete fault propagation path through fault causal propagation link tracing analysis, and outputs verifiable fault root cause location results. The fault root cause reasoning unit 4 includes a multi-source diagnostic data receiving module 41, a fault causal tracing and location module 42, and a root cause result verification and output module 43, wherein: The multi-source diagnostic data receiving module 41 synchronously receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit 2 and the standardized component-level fault diagnosis results output by the fault intelligent diagnosis unit 3, and extracts the fault propagation link and the candidate set of the first abnormal node from the fault reasoning feature results.

[0075] Specifically, the multi-source diagnostic data receiving module 41 serves as the unit data entry point, synchronously connecting to upstream data from both sources to complete data reception, core extraction, and standardized alignment. First, it receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit 2, accurately extracting two types of core tracing data: the directed fault propagation link and the candidate set of the first abnormal node. Second, it receives the standardized component-level fault diagnosis results output by the fault intelligent diagnosis unit 3, extracting the unique identifier of the key component, the corresponding fault type, the fault severity level, and the component-level abnormal characterization value. Using the millisecond-level timestamp-key component unique identifier as the unique association index, the two types of upstream data are aligned and bound frame by frame. If there is invalid data with mismatched timestamps, missing component identifiers, or data gaps, it is directly removed and marked as abnormal, ensuring that the data input to the tracing module is complete, consistent, and valid. The processed data is synchronously transmitted to the fault causal tracing and positioning module 42.

[0076] The fault cause tracing and localization module 42 combines the extracted fault propagation link, the candidate set of the first abnormal node, and the standardized component-level fault diagnosis results to carry out fault cause propagation link tracing analysis and locate the first root cause node of the fault and the complete fault propagation path.

[0077] Specifically, the fault cause tracing and localization module 42 is based on the fault cause topology backtracking logic and combined with quantitative screening rules to complete the root cause localization and path sorting, which fully conforms to the physical cause and effect law of industrial fault propagation. Taking the candidate set of the first abnormal node as the initial source tracing basis, it backtracks along the directed link of fault propagation and screens the first root cause node according to the preset quantitative priority: the candidate node with the highest fault severity level and the largest component-level abnormality characterization value is selected first. If the level and score are consistent, the initial node at the upstream of the fault propagation link is selected, and downstream nodes that are only affected by fault propagation and have no original high abnormality characteristics are excluded. Finally, the unique first root cause node of the fault is locked. Simultaneously, starting from the root cause node, trace forward along the directed fault propagation link to connect all faulty components affected by the propagation, forming a continuous and complete fault propagation path; if the initial abnormal node candidate set is empty, directly determine the component with the highest fault severity level and the largest abnormal characterization value as the temporary root cause node; if there is no effective fault propagation link, only the single-node root cause result is output, covering various extreme working conditions.

[0078] The root cause result verification output module 43 verifies the verifiability of the initial root cause node of the fault and the complete fault propagation path obtained from the location, and outputs verifiable fault root cause location results that meet the requirements.

[0079] Specifically, the root cause result verification output module 43 performs multi-dimensional standardized verification of the location results to ensure that the root cause results are verifiable and reliable. The verification rules include three core components: First, causal level verification, the severity level of the fault at the initial root cause node is not lower than the level of all downstream faulty components; second, abnormal value verification, the component-level abnormal characterization value of the root cause node is higher than the corresponding value of all downstream related faulty components; third, data matching verification, the fault type of the root cause node matches the component-level abnormal characterization range and the standard features of the pre-built fault knowledge base. If all three verifications pass, the result is deemed valid. If any verification fails, the result is marked as pending review and additional verification anomalies are added. Finally, by integrating valid verification conclusions, initial root cause node information of the fault, and complete fault propagation path, standardized and verifiable fault root cause location results are generated. The results are synchronously output to the operation and maintenance display module and the historical fault database. This is used for both the visualization of fault root causes and the subsequent iteration and optimization of the fault knowledge base and the retrospective analysis of historical faults. The output format is fully compatible with the downstream module interface, ensuring smooth data flow.

[0080] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0081] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. An intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model, characterized in that, include: The data acquisition and preprocessing unit (1) acquires multimodal operation data corresponding to each monitoring component of the target industrial equipment, performs time stamp synchronization, outlier removal, numerical normalization, and feature standardization preprocessing on the acquired multimodal operation data corresponding to each monitoring component of the target industrial equipment, and outputs a standardized multimodal feature dataset that corresponds one-to-one with each monitoring component of the target industrial equipment and is aligned with the time dimension. The multimodal reasoning model reasoning unit (2) receives the standardized multimodal feature dataset output by the data acquisition and preprocessing unit (1), and adopts cross-modal hierarchical reasoning technology based on equipment component topology constraints to complete the component-level association mapping of multimodal features, the fusion of cross-modal anomaly features within nodes based on topological neighborhood anomaly prior constraints, and the fault propagation link reasoning under topological constraints. It outputs fault reasoning feature results containing component-level anomaly representation, fault propagation link, and first-time anomaly node candidate set. The fault intelligent diagnosis unit (3) receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit (2), and performs matching and judgment on the component-level abnormal representation based on the pre-built industrial equipment component-level fault knowledge base and the preset fault judgment threshold, completes the fault type identification and fault severity level classification of each monitored component of the target industrial equipment, and outputs standardized component-level fault diagnosis results that correspond one-to-one with each monitored component of the target industrial equipment. The fault root cause reasoning unit (4) receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit (2) and the standardized component-level fault diagnosis results output by the fault intelligent diagnosis unit (3). Combining the fault propagation link, the candidate set of the first abnormal node and the standardized component-level fault diagnosis results, the fault root cause location is located through fault causal propagation link tracing analysis, and verifiable fault root cause location results are output.

2. The intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model according to claim 1, characterized in that, The data acquisition and preprocessing unit (1) includes a multimodal operation data acquisition module (11), a multimodal data linkage preprocessing module (12), and a standardized feature generation and output module (13), wherein: The multimodal operation data acquisition module (11) acquires multimodal operation data corresponding to each monitoring component of the target industrial equipment and outputs the acquired multimodal operation data corresponding to each monitoring component of the target industrial equipment to the multimodal data linkage preprocessing module (12). The multimodal data linkage preprocessing module (12) receives the multimodal operation data output by the multimodal operation data acquisition module (11), performs time stamp synchronization, outlier removal, and numerical normalization preprocessing on the multimodal operation data in sequence, and outputs the multimodal operation data corresponding to each monitoring component of the target industrial equipment after preprocessing to the standardized feature generation output module (13). The standardized feature generation and output module (13) receives the preprocessed multimodal operation data and performs feature standardization preprocessing to output a standardized multimodal feature dataset that corresponds one-to-one with each monitoring component of the target industrial equipment and is aligned with the time dimension.

3. The intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model according to claim 1, characterized in that, The multimodal reasoning model reasoning unit (2) includes a device component topology graph construction module (21), a multimodal feature component-level association mapping module (22), a node-internal cross-modal anomaly feature fusion module (23), a fault propagation link reasoning module (24), and a reasoning result output module (25), wherein: The equipment component topology construction module (21) constructs a directed weighted equipment component topology based on the component hierarchy and physical relationship of the target industrial equipment, and outputs the equipment component topology. The multimodal feature component-level association mapping module (22) receives the standardized multimodal feature dataset output by the data acquisition and preprocessing unit (1) and the equipment component topology map output by the equipment component topology map construction module (21), completes the component-level association mapping of multimodal features, and generates the single-modal anomaly representation corresponding to each component node; The intra-node cross-modal anomaly feature fusion module (23) receives the single-modal anomaly representation output by the multi-modal feature component-level association mapping module (22) and the device component topology map output by the device component topology map construction module (21), and completes the intra-node cross-modal anomaly feature fusion based on the prior constraint of topological neighborhood anomaly, generating a component-level anomaly representation resistant to propagation interference. The fault propagation link reasoning module (24) receives the equipment component topology map output by the equipment component topology map construction module (21) and the anti-propagation interference component-level anomaly representation output by the node cross-modal anomaly feature fusion module (23), completes the fault propagation link reasoning under topological constraints, and generates the fault propagation link and the first set of anomaly node candidates. The reasoning result output module (25) receives the anti-propagation interference component-level anomaly representation output by the intra-node cross-modal anomaly feature fusion module (23), the fault propagation link output by the fault propagation link reasoning module (24), and the first abnormal node candidate set, and outputs the fault reasoning feature result containing the component-level anomaly representation, the fault propagation link, and the first abnormal node candidate set.

4. The intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model according to claim 3, characterized in that, The directed weighted equipment component topology graph construction process of the equipment component topology graph construction module (21) includes the following steps: S21.1 Define the equipment component topology graph as a directed weighted equipment component topology graph. ,in It is a set of nodes composed of various monitoring components of industrial equipment. It is the set of directed edges formed by the physical connections between components. This is the set of weight coefficients for each directed edge; S21.

2. Based on the physical relationships of the target industrial equipment, determine the anomaly propagation capability coefficients corresponding to each directed edge in the directed weighted graph. ,in For upstream component nodes, For downstream component nodes; S21.3 Complete the construction and output of the directed weighted device component topology graph.

5. The intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model according to claim 3, characterized in that, The multimodal feature component-level association mapping and single-modal anomaly representation generation process of the multimodal feature component-level association mapping module (22) includes the following steps: S22.1 Obtain the standardized multimodal feature dataset output by the data acquisition and preprocessing unit (1); S22.2 Calculate the comprehensive matching degree between standardized multimodal features and key component nodes. ,in To standardize multimodal features, For key component nodes; S22.3, Overall matching degree Greater than the matching threshold The standardized multimodal features are mapped to the corresponding key component nodes; S22.

4. For standardized multimodal features mapped to the same key component node, generate single-modal anomaly representations according to modality type. ,in For the modal types of multimodal data, For timestamps.

6. The intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model according to claim 3, characterized in that, The intra-node cross-modal anomaly feature fusion module (23) includes an upstream neighbor node set determination submodule, a fault propagation interference prior coefficient calculation submodule, a dynamic attention weighted fusion submodule, and a component-level anomaly representation standardization output submodule, wherein: The upstream neighbor node set determination submodule determines the upstream neighbor node set corresponding to the current critical component node based on the equipment component topology graph output by the equipment component topology graph construction module (21). The upstream neighbor node set is the set of all upstream nodes with directed edges pointing to the current critical component node, which is completely matched with the fault propagation direction. The fault propagation interference prior coefficient calculation submodule calculates the fault propagation interference prior coefficient corresponding to each mode of the current key component node based on the single-mode anomaly characterization of each node in the upstream neighboring node set, and quantifies the degree of interference of upstream fault propagation on the corresponding mode data of the current node. The dynamic attention weighted fusion submodule is based on the dynamic attention weighted fusion logic of the topological neighborhood anomaly prior constraint. With the fault propagation interference prior coefficient as the core constraint, it calculates the dynamic attention weight corresponding to each single-modal anomaly representation of the current critical component node, and performs intra-node cross-modal weighted fusion on each single-modal anomaly representation of the current critical component node based on the dynamic attention weight, to generate a component-level anomaly representation that only reflects the original anomaly state of the current node and is resistant to propagation interference. The component-level anomaly representation standardization output submodule receives the anti-propagation interference component-level anomaly representation output by the dynamic attention weighted fusion submodule, performs structured integration of the component-level anomaly representation, and outputs it to the fault propagation link inference module under topological constraints.

7. The intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model according to claim 6, characterized in that, The dynamic attention weighted fusion process based on topological neighborhood anomaly prior constraints in the dynamic attention weighted fusion submodule includes the following steps: S23.

1. Dimensionlessly normalize the single-modal anomaly characterization of the current key component nodes to obtain the normalized single-modal anomaly characterization. ; S23.2, Call the fault propagation interference prior coefficient calculation submodule to obtain the fault propagation interference prior coefficient. ; S23.3, using fault propagation interference prior coefficients To constrain the calculation, dynamic attention weights are computed. ; S23.4, Based on dynamic attention weights Characterization of normalized single-mode anomalies Weighted fusion is performed to generate component-level anomaly representations that are resistant to propagation interference. .

8. The intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model according to claim 3, characterized in that, The fault propagation link reasoning and first abnormal node identification process under topological constraints of the fault propagation link reasoning module (24) includes the following steps: S24.1 Constructing an anomaly propagation transition matrix based on the weight matrix of the equipment component topology graph. ; S24.2 Calculate the anomaly propagation consistency coefficient between adjacent component nodes. ; S24.3, The consistency coefficient of anomaly propagation Greater than the consistency threshold The propagation path was identified as a valid anomaly propagation path, and a directed subgraph of anomaly propagation was constructed. ; S24.4 Calculating the Directed Subgraph for Anomaly Propagation Based on Bayes' Theorem The posterior probability that each node is the first anomalous node ; S24.5, posterior probability Greater than the first-round probability threshold The node is identified as the first abnormal node and included in the candidate set of first abnormal nodes. At the same time, the effective abnormal propagation path is identified as the fault propagation link.

9. The intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model according to claim 1, characterized in that, The intelligent fault diagnosis unit (3) includes a fault data receiving and knowledge base calling module (31), a fault matching and classification module (32), and a standardized diagnosis result output module (33), wherein: The fault data receiving and knowledge base calling module (31) receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit (2), extracts the component-level anomaly representation, and calls the pre-built industrial equipment component-level fault knowledge base to obtain the corresponding standard component fault features and fault severity level judgment rules. The fault matching and classification module (32) matches the extracted component-level anomaly representation with the standard features in the pre-built industrial equipment component-level fault knowledge base based on the preset fault judgment threshold, and completes the identification of fault type and classification of fault severity level of each monitored component of the target industrial equipment. The standardized diagnostic result output module (33) integrates the judgment results of the fault matching and classification module (32) to generate and output standardized component-level fault diagnosis results that correspond one-to-one with each monitored component of the target industrial equipment.

10. The intelligent fault diagnosis and root cause reasoning system for industrial equipment based on a multimodal reasoning model according to claim 1, characterized in that, The fault root cause reasoning unit (4) includes a multi-source diagnostic data receiving module (41), a fault cause-and-effect tracing and localization module (42), and a root cause result verification and output module (43), wherein: The multi-source diagnostic data receiving module (41) synchronously receives the fault reasoning feature results output by the multimodal reasoning model reasoning unit (2) and the standardized component-level fault diagnosis results output by the fault intelligent diagnosis unit (3), and extracts the fault propagation link and the candidate set of the first abnormal node from the fault reasoning feature results; The fault cause tracing and localization module (42) combines the extracted fault propagation link, the first abnormal node candidate set and the standardized component-level fault diagnosis results to carry out fault cause propagation link tracing analysis and locate the first root cause node of the fault and the complete fault propagation path. The root cause result verification output module (43) verifies the verifiability of the initial root cause node of the fault and the complete fault propagation path obtained by the location, and outputs the verifiable fault root cause location result that meets the requirements.

Images