Industrial knowledge question answering and fault analysis method based on large language model

Abstract

The present application relates to an industrial knowledge question answering and fault analysis method based on a large language model, and relates to the technical field of industrial intelligence, wherein the present application effectively makes up for the defect of insufficient real-time state perception of the large language model for the industrial site by performing feature dimension reduction and pattern recognition on the time sequence data of the underlying dynamic sensor and deeply fusing the quantized vector thereof with the static knowledge text vector, and performs linkage positioning based on a transparent dynamic physical topology network to automatically acquire the data of abnormal nodes and their upstream and downstream strongly correlated nodes, so that the large language model can perform global perspective cross-level reasoning along the actual physical connection link of the equipment, thereby effectively solving the technical problem that it is difficult to trace the real source in a complex pipe network or conductive fault, converting the preliminary diagnosis set generated by the large model into a control instruction, injecting the digital twin environment for reverse deduction verification, converting subjective semantic reasoning into objective physical level verification, and ensuring the industrial level rigor of the output result.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent technology, and in particular to an industrial knowledge question answering and fault analysis method based on a large language model. Background Technology

[0002] Improving the effectiveness of equipment fault diagnosis directly contributes to increased production efficiency and ensures safe equipment operation. While existing intelligent question-and-answer systems for industrial knowledge and fault analysis have made significant progress in general knowledge retrieval and basic operational guidance, they still face the following two challenges in real-world industrial applications:

[0003] I. Insufficient Capabilities in Processing and Fusion of Multi-Source Heterogeneous Data: Most current question-answering systems rely on static knowledge graphs or documents (such as equipment manuals, historical maintenance records, and expert experience), but their ability to process real-time dynamic sensor data is weak. Industrial faults often require complex data processing and feature extraction for diagnosis. Existing question-answering systems typically only operate at the semantic retrieval level, lacking the ability to directly integrate question-answering logic with underlying deep data analysis platforms, and thus cannot directly process and analyze massive amounts of operational data to support their conclusions.

[0004] Second, there is a lack of deep interaction with digital twins and underlying physical systems: many systems exist only as independent dialog boxes and cannot directly link with the factory's digital twin monitoring and simulation systems. When inquiring about the fault of a specific piece of equipment, the system cannot automatically locate the problem node in the 3D spatial mesh model, nor can it verify its fault-solving suggestions through simulation. Moreover, faults in industrial equipment are often propagating. When dealing with complex pipeline networks, microgrids (such as photovoltaic-storage-charging-discharging microgrid systems), or distribution networks, existing systems lack non-intrusive topology identification and dynamic modeling capabilities, making it difficult to accurately trace the root cause along the physical topology of the system, and often only providing superficial error explanations.

[0005] To address the aforementioned technical deficiencies, a solution is proposed. Summary of the Invention

[0006] The purpose of this invention is to: perform feature dimensionality reduction and pattern recognition on the time-series data of the underlying dynamic sensors, deeply fuse its quantized vector with the static knowledge text vector, and perform linkage positioning based on a transparent dynamic physical topology network; automatically acquire data of abnormal nodes and their upstream and downstream strongly related nodes; transform the preliminary diagnostic set generated by the large model into control commands; inject them into the digital twin environment for reverse inference verification; and transform subjective semantic reasoning into objective physical-level verification.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: an industrial knowledge question answering and fault analysis method based on a large language model, comprising the following steps:

[0008] S1. Obtain the natural language fault query command input by the user, and call the large language model to perform semantic parsing on the natural language fault query command to extract the key entity set, the key entity set including device ID, system level, and fault phenomenon description.

[0009] S2. Map the set of key entities to a transparent dynamic physical topology network of the industrial site, locate the abnormal source node and its upstream and downstream related nodes with physical relationships, and automatically obtain the dynamic sensor time series data of the abnormal source node and its upstream and downstream related nodes within a preset time window based on the location results.

[0010] The specific process for acquiring time-series data from dynamic sensors is as follows:

[0011] S21. Obtain the set of key entities, call the pre-built factory three-dimensional space mesh model, and map the standard equipment IDs contained in the set of key entities to the corresponding three-dimensional space coordinate nodes in the factory three-dimensional space mesh model as anomaly source nodes.

[0012] S22. Using a non-intrusive topology identification method, the dynamic operating characteristics of each node in the industrial site are read in real time. By calculating the data cross-correlation function between adjacent nodes, the physical connection weights between the abnormal source node and surrounding nodes are dynamically updated, and the upstream and downstream related nodes that have a real physical linkage relationship with the abnormal source node are locked.

[0013] S23. Based on the time delay and amplitude attenuation of abnormal fluctuation propagation in historical time series data, dynamically calculate the fault propagation attenuation factor between the abnormal source node and each upstream and downstream related node. In each upstream and downstream related node, remove edge nodes whose fault propagation attenuation factor is lower than the preset effective threshold, and retain strongly related nodes to obtain a set of related nodes.

[0014] S24. Based on the extracted fault phenomenon description, match the corresponding feature frequency band in the fault feature dictionary, and adaptively calculate and set the extraction range of dynamic sensor time series data according to the Nyquist sampling frequency of the feature frequency band and the preset fault latency empirical value to obtain the preset time window.

[0015] S25. Send targeted data extraction instructions to the SCADA system in the industrial field through a unified API interface. The real-time data stream obtained undergoes preliminary data cleaning, including processing missing values, removing glitches and noise caused by sensor malfunctions, and aligning upstream and downstream sensor data with different sampling frequencies onto a unified time axis through a linear interpolation algorithm to generate high-dimensional, structured dynamic sensor time-series data.

[0016] S3. Input the time series data of the dynamic sensor into the deep data analysis platform for feature dimensionality reduction and abnormal pattern recognition to generate a quantized data analysis vector. Then, fuse the data analysis vector with the static text vector retrieved from the static knowledge base to construct a fusion prompt word that includes dynamic operating status and static domain knowledge.

[0017] S4. Input the fused prompt words into the large language model to perform fault tracing reasoning, generate a preliminary diagnosis candidate set, and at the same time, the large language model generates corresponding system control commands to drive the digital twin monitoring and simulation system, highlight the corresponding topology nodes in the three-dimensional spatial grid model, and inject the preliminary diagnosis candidate set as input conditions into the digital twin simulation environment for reverse inference verification, screen out the real root cause that best matches the actual sensor characteristics, and obtain the simulation verification results.

[0018] S5. Based on the simulation verification results, generate the final fault tracing report and repair suggestions through the large language model, and present them to the user through the digital twin interactive interface.

[0019] Furthermore, the specific process for extracting the key entity set is as follows:

[0020] S11. Obtain the natural language fault query command from the user through speech recognition conversion and text input, perform noise reduction and stop word removal on the natural language fault query command, and use a pre-trained industrial domain word embedding model to convert the processed natural language fault query command into a high-dimensional command semantic vector.

[0021] S12. Input a compound prompt word with strict format constraints into the large language model. The compound prompt word contains the following:

[0022] Role setting module: Limited to the large language model as an expert in extracting fault features of industrial equipment;

[0023] Context Injection Module: Injects the current factory's system architecture outline, which includes a hierarchical tree structure of workshop layer, production line layer, equipment layer, and component layer;

[0024] Task execution module: Maps the semantic vector of the instruction back to text, parses it based on the compound prompt words through a large language model, and outputs a coarse extracted entity set, which includes candidate device descriptions, candidate system level descriptions, and candidate fault phenomenon descriptions.

[0025] S13. The coarsely extracted entity set is forcibly aligned and standardized using a pre-built standard database of factory assets to obtain a set of key entities.

[0026] Furthermore, the specific process of forcibly aligning and standardizing the coarsely extracted entity set is as follows:

[0027] S131. Obtain candidate device descriptions from the coarsely extracted entity set, and perform text matching based on cosine similarity between the candidate device descriptions and the standard device list in the factory asset standard database. If the text matching similarity is higher than a preset similarity threshold, then map the candidate device description to a unique standard device ID.

[0028] If the match fails, a follow-up question mechanism is triggered, requiring the user to provide the accurate device name;

[0029] S132. Based on the mapped standard equipment ID, trace back the parent node to which the standard equipment ID belongs in the association topology tree of the factory asset standard database to determine the precise system level corresponding to the standard equipment ID.

[0030] S133. Align the candidate fault phenomenon descriptions with the standard industrial fault feature vocabulary, and transform the colloquial candidate fault phenomenon descriptions into standard quantitative description labels.

[0031] S134. Package the verified and standardized standard device ID, the precise system hierarchy, and the normalized fault feature tags to generate the final set of key entities.

[0032] Furthermore, the specific process for generating quantified data analysis vectors is as follows:

[0033] S301. Acquire dynamic sensor time series data and input it into a pre-trained temporal autoencoder in a deep data analysis platform. The encoder part of the temporal autoencoder performs multi-scale one-dimensional convolution on the input sequence to filter out redundant background information in the dynamic sensor time series data and retain abnormal fluctuation characteristics that characterize the operating status of the device.

[0034] S302. Extract the latent space state vector of the bottleneck layer of the time-series autoencoder as the quantization basis vector characterizing the dynamic anomaly features of the current industrial equipment.

[0035] S303. Construct a multilayer perceptron as a semantic projection network, input the quantized basic vector into the semantic projection network, and fine-tune the training on the historical sensor data-fault description text alignment dataset using the cross-entropy loss function, so that the semantic projection network linearly maps the quantized basic vector to the natural language semantic space of the large language model, generating a quantized data analysis vector that can be understood by the large language model. The quantized data analysis vector is mathematically equivalent to a soft prompt describing the current dynamic operation anomaly.

[0036] Furthermore, the specific process for constructing integrated prompts is as follows:

[0037] S311. Obtain the set of key entities and search for them in the local static knowledge base. The local static knowledge base includes equipment maintenance manuals, historical work orders and expert experience graphs. The search process adopts a text embedding model based on the dual-tower BERT architecture to convert the strongly related historical text paragraphs retrieved into static text vectors.

[0038] S312. Using the generated quantitative data analysis vector as the query and the static text vector as the key and value, calculate the attention weight of dynamic features on each static knowledge item.

[0039] S313. Based on the attention weights, the static text vectors are weighted and summed, and then concatenated with the quantized data analysis vectors to generate a multimodal fusion feature vector.

[0040] S314. The multimodal fusion feature vector is converted into a structured text representation through a reverse decoder to construct the final fusion prompt word.

[0041] Furthermore, the specific process for obtaining the simulation verification results is as follows:

[0042] S41. After obtaining the fusion prompt words, perform multi-step reasoning calculations, output a preliminary diagnostic candidate set, call the preset instruction compilation agent, and parse the preliminary diagnostic candidate set into a structured simulation control vector executable by the digital twin system. The structured simulation control vector includes the topology node ID to be operated, the system parameter type to be modified, and its corresponding inference assignment.

[0043] S42. Based on the ID of the topology node to be operated, coordinate system addressing is performed in the real-time three-dimensional space mesh model to lock the corresponding virtual device entity and trigger the visual rendering engine to mark the corresponding topology node with a specific highlight color.

[0044] S43. Cut off the real-time data feedforward flow of the target physical topology network region in the digital twin system, so that the digital twin system enters an independent offline simulation branch. Inject the structured simulation control vector containing the deduction assignment into the mechanism model in the offline simulation branch. Based on the injected fault parameters and combined with historical boundary conditions, perform forward differential solution and simulation operation with a preset time step to generate the simulated sensor time series data matrix that each associated node should generate under the current fault assumption.

[0045] S44. Acquire the dynamic sensor time series data and the simulated sensor time series data matrix for feature comparison. Based on the sequence alignment algorithm of Pearson correlation coefficient, calculate the morphological fit between the two time series data streams and generate a quantified simulation matching score.

[0046] S45. Traverse the simulation matching scores corresponding to the preliminary diagnosis candidate set, lock the preliminary diagnosis candidate set with the simulation matching scores as the true root cause, record the corresponding deduction parameters and comparison curves, and package them to generate a structured simulation verification result.

[0047] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are as follows:

[0048] The industrial knowledge Q&A and fault analysis method based on the large language model effectively compensates for the deficiency of the large language model in perceiving the real-time state of the industrial site by performing feature dimensionality reduction and pattern recognition on the underlying dynamic sensor time series data and deeply fusing its quantization vectors with static knowledge text vectors. Based on the transparent dynamic physical topology network for linkage positioning, it automatically obtains the data of abnormal nodes and their strongly associated upstream and downstream nodes, enabling the large language model to perform cross-level reasoning from a global perspective along the actual physical connection link of the device, thus effectively solving the technical problem of difficult to trace the real source in complex pipe networks or conductive faults. It converts the preliminary diagnosis set generated by the large model into control instructions, injects them into the digital twin environment for reverse deduction verification, and converts subjective semantic reasoning into objective physical-level verification, ensuring the industrial-level rigor of the output results. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It shows the overall method flow diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0051] Embodiment:

[0052] As Figure 1 shown, the industrial knowledge Q&A and fault analysis method based on the large language model includes the following steps:

[0053] S1. Obtain the natural language fault query instruction input by the user, and call the large language model to perform semantic parsing on the natural language fault query instruction to extract the key entity set, and the key entity set includes the device ID, system level, and fault phenomenon description;

[0054] The specific process of extracting the key entity set is as follows:

[0055] S11. Obtain the natural language fault query command (e.g., the temperature control system in Workshop No. 3 is not cooling down and there seems to be an abnormal sound from the main pipeline) by the user through speech recognition conversion and text input. Perform noise reduction and stop word removal on the natural language fault query command, and use a pre-trained industrial domain word embedding model to convert the processed natural language fault query command into a high-dimensional command semantic vector.

[0056] S12. Input a compound prompt word with strict format constraints into the large language model. The compound prompt word contains the following:

[0057] Role setting module: Limited to the large language model as an expert in extracting fault features of industrial equipment;

[0058] Context Injection Module: Injects the current factory's system architecture outline, which includes a hierarchical tree structure of workshop layer, production line layer, equipment layer, and component layer;

[0059] Task execution module: Maps the semantic vector of the instruction back to text, parses it based on the compound prompt words using a large language model, and outputs a coarse extraction entity set. The coarse extraction entity set includes candidate equipment descriptions (such as main pipeline), candidate system level descriptions (such as workshop No. 3, temperature control system), and candidate fault phenomenon descriptions (such as temperature not dropping, abnormal sound).

[0060] S13. The coarsely extracted entity set is forcibly aligned and standardized using a pre-built standard database of factory assets to obtain a set of key entities.

[0061] The specific process of forcibly aligning and standardizing the coarsely extracted entity set is as follows:

[0062] S131. Obtain candidate device descriptions from the coarsely extracted entity set, and perform text matching based on cosine similarity between the candidate device descriptions and the standard device list in the factory asset standard database. If the text matching similarity is higher than a preset similarity threshold, then map the candidate device description to a unique standard device ID (for example, map the main pipeline to the standard ID PIPE-MAIN-WS03).

[0063] If the match fails, a follow-up question mechanism is triggered, requiring the user to provide the accurate device name;

[0064] S132. Based on the mapped standard equipment ID, trace back the parent node to which the standard equipment ID belongs in the association topology tree of the factory asset standard database to determine the precise system level corresponding to the standard equipment ID (such as Workshop No. 3 - Cooling Pipeline Network - Main Node).

[0065] S133. Align the candidate fault phenomenon descriptions with the standard industrial fault feature vocabulary library, and transform the colloquial candidate fault phenomenon descriptions into standard quantitative description labels (such as temperature exceeding limits, high frequency vibration / abnormal noise).

[0066] S134. Package the verified and standardized standard device ID, the precise system hierarchy, and the normalized fault feature tags to generate the final set of key entities.

[0067] S2. Map the set of key entities to a transparent dynamic physical topology network of the industrial site, locate the abnormal source node and its upstream and downstream related nodes with physical relationships, and automatically obtain the dynamic sensor time series data of the abnormal source node and its upstream and downstream related nodes within a preset time window based on the location results.

[0068] The specific process for acquiring time-series data from dynamic sensors is as follows:

[0069] S21. Obtain the set of key entities, call the pre-built factory three-dimensional space mesh model, and map the standard equipment IDs contained in the set of key entities to the corresponding three-dimensional space coordinate nodes in the factory three-dimensional space mesh model as anomaly source nodes.

[0070] S22. Using a non-intrusive topology identification method, the dynamic operating characteristics (such as current fluctuations and flow rate change rates) of each node in the industrial site are read in real time. By calculating the data cross-correlation function between adjacent nodes, the physical connection weights between the abnormal source node and surrounding nodes are dynamically updated, and the upstream and downstream related nodes that have a real physical linkage relationship with the abnormal source node are locked.

[0071] S23. Based on the time delay and amplitude attenuation of abnormal fluctuation propagation in historical time series data, dynamically calculate the fault propagation attenuation factor between the abnormal source node and each upstream and downstream related node. In each upstream and downstream related node, remove edge nodes whose fault propagation attenuation factor is lower than the preset effective threshold, and retain strongly related nodes to obtain a set of related nodes.

[0072] S24. Based on the extracted fault phenomenon description, match the corresponding feature frequency band in the fault feature dictionary. Based on the Nyquist sampling frequency of the feature frequency band and the preset fault latency empirical value, adaptively calculate and set the extraction range of dynamic sensor time series data to obtain the preset time window. For example, for thermodynamic slow-change faults, the system automatically extends the time window to 24 hours before the fault occurs; for electrical transient faults, it shortens and intensively extracts the high-frequency sampling data of the first 10 minutes.

[0073] S25. Send targeted data extraction instructions to the SCADA system in the industrial field through a unified API interface (such as the OPCUA protocol). The real-time data stream obtained undergoes preliminary data cleaning, including processing missing values, removing glitches and noise caused by sensor malfunctions, and aligning upstream and downstream sensor data with different sampling frequencies onto a unified time axis through a linear interpolation algorithm to generate high-dimensional, structured dynamic sensor time-series data.

[0074] S3. Input the time series data of the dynamic sensor into the deep data analysis platform for feature dimensionality reduction and abnormal pattern recognition to generate a quantized data analysis vector. Then, fuse the data analysis vector with the static text vector retrieved from the static knowledge base to construct a fusion prompt word that includes dynamic operating status and static domain knowledge.

[0075] The specific process for generating quantized data analysis vectors is as follows:

[0076] S301. Acquire dynamic sensor time-series data and input it into a pre-trained temporal autoencoder in a deep data analysis platform (e.g., constructed based on a long short-term memory network LSTM-Autoencoder or a temporal convolutional network TCN architecture). The encoder part of the temporal autoencoder performs multi-scale one-dimensional convolution on the input sequence to filter out redundant background information in the dynamic sensor time-series data and retain abnormal fluctuation characteristics that characterize the device's operating state.

[0077] S302. Extract the latent space state vector of the bottleneck layer of the time-series autoencoder as a quantization basis vector to characterize the dynamic anomaly features of the current industrial equipment. The quantization basis vector realizes extreme dimensionality reduction and high-density feature condensation of massive complex monitoring data.

[0078] S303. Construct a multilayer perceptron as a semantic projection network, input the quantized basic vector into the semantic projection network, and fine-tune the training on the historical sensor data-fault description text alignment dataset using the cross-entropy loss function, so that the semantic projection network linearly maps the quantized basic vector to the natural language semantic space of the large language model, generating a quantized data analysis vector that can be understood by the large language model. The quantized data analysis vector is mathematically equivalent to a soft prompt describing the current dynamic operation anomaly.

[0079] The specific process of constructing integrated prompts is as follows:

[0080] S311. Obtain the set of key entities and search for them in the local static knowledge base. The local static knowledge base includes equipment maintenance manuals, historical work orders and expert experience graphs. The search process adopts a text embedding model based on the dual-tower BERT architecture to convert the strongly related historical text paragraphs retrieved into static text vectors.

[0081] S312. Using the generated quantitative data analysis vector (representing the current real-time abnormal state) as the query and the static text vector (representing historical static knowledge) as the key and value, calculate the attention weight of dynamic features on each static knowledge item.

[0082] S313. Based on the attention weights, the static text vectors are weighted and summed, and then concatenated with the quantized data analysis vectors to generate a multimodal fusion feature vector.

[0083] S314. The multimodal fusion feature vector is converted into a structured text representation through a reverse decoder to construct the final fusion prompt word. An example of the structure of the fusion prompt word is: [System prompt: The current device's dynamic running feature vector mapping shows a periodic oscillation at frequency X (i.e., the translation of data analysis vectors); combined with historical static knowledge (i.e., the translation of static text vectors), the common cause of this phenomenon in this model of device is Y; please trace the fault based on the above comprehensive input].

[0084] S4. Input the fused prompt words into the large language model to perform fault tracing reasoning, generate a preliminary diagnosis candidate set, and at the same time, the large language model generates corresponding system control commands to drive the digital twin monitoring and simulation system, highlight the corresponding topology nodes in the three-dimensional spatial grid model, and inject the preliminary diagnosis candidate set as input conditions into the digital twin simulation environment for reverse inference verification, screen out the real root cause that best matches the actual sensor characteristics, and obtain the simulation verification results.

[0085] The specific process for obtaining the simulation verification results is as follows:

[0086] S41. After obtaining the fusion prompt words, perform multi-step reasoning calculations and output a preliminary diagnostic candidate set (e.g., assumption A: valve internal leakage; assumption B: sensor drift). Call the preset instruction compilation agent to parse the preliminary diagnostic candidate set into a structured simulation control vector executable by the digital twin system. The structured simulation control vector includes the topology node ID to be operated, the type of system parameter to be modified, and its corresponding inference assignment (e.g., modify the leakage rate parameter of the valve node in the digital twin space from 0 to a preset fault experience value).

[0087] S42. Based on the topology node ID to be operated, coordinate system addressing is performed in the real-time three-dimensional spatial mesh model to lock the corresponding virtual device entity and trigger the visual rendering engine to mark the corresponding topology node with a specific highlight color, so as to realize the intuitive spatial positioning of the underlying data anomaly at the front-end UI level.

[0088] S43. Cut off the real-time data feedforward flow of the target physical topology network region in the digital twin system, so that the digital twin system enters an independent offline simulation branch. Inject the structured simulation control vector containing the deduction assignment into the mechanism model in the offline simulation branch. Based on the injected fault parameters and combined with historical boundary conditions (such as the load state before the fault occurred), perform forward differential solution and simulation operation with a preset time step to generate the simulated sensor time series data matrix that each associated node should generate under the current fault assumption.

[0089] S44. Acquire the dynamic sensor time series data and the simulated sensor time series data matrix for feature comparison. Based on the sequence alignment algorithm of Pearson correlation coefficient, calculate the morphological fit between the two time series data streams and generate a quantified simulation matching score.

[0090] S45. Traverse the simulation matching scores corresponding to the preliminary diagnostic candidate set, lock the preliminary diagnostic candidate set of the simulation matching scores as the true root cause, record the corresponding inference parameters and comparison curves, and package them to generate structured simulation verification results.

[0091] S5. Based on the simulation verification results, generate the final fault tracing report and repair suggestions through the large language model, and present them to the user through the digital twin interactive interface.

[0092] This invention effectively compensates for the shortcomings of large language models in perceiving real-time status in industrial sites by performing feature dimensionality reduction and pattern recognition on time-series data from underlying dynamic sensors, and deeply fusing its quantized vectors with static knowledge text vectors. Based on a transparent dynamic physical topology network, it performs linkage positioning and automatically acquires data on abnormal nodes and their upstream and downstream strongly related nodes. This enables the large language model to perform cross-level reasoning from a global perspective along the actual physical connection links of the equipment, thereby effectively solving the technical problem of tracing the true source in complex pipeline networks or conductive faults. The preliminary diagnostic set generated by the large model is transformed into control commands and injected into a digital twin environment for reverse inference verification, transforming subjective semantic reasoning into objective physical-level verification, ensuring the industrial-grade rigor of the output results.

[0093] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0094] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0095] In the embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways; for example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed; another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces.

[0096] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An industrial knowledge question answering and fault analysis method based on a large language model, characterized in that, Includes the following steps: S1. Obtain the natural language fault query command input by the user, and call the large language model to perform semantic parsing on the natural language fault query command to extract the key entity set, the key entity set including device ID, system level, and fault phenomenon description. S2. Map the set of key entities to a transparent dynamic physical topology network of the industrial site, locate the abnormal source node and its upstream and downstream related nodes with physical relationships, and automatically obtain the dynamic sensor time series data of the abnormal source node and its upstream and downstream related nodes within a preset time window based on the location results. S3. Input the time series data of the dynamic sensor into the deep data analysis platform for feature dimensionality reduction and abnormal pattern recognition to generate a quantized data analysis vector. Then, fuse the data analysis vector with the static text vector retrieved from the static knowledge base to construct a fusion prompt word that includes dynamic operating status and static domain knowledge. S4. Input the fused prompt words into the large language model to perform fault tracing reasoning, generate a preliminary diagnosis candidate set, and at the same time, the large language model generates corresponding system control commands to drive the digital twin monitoring and simulation system, highlight the corresponding topology nodes in the three-dimensional spatial grid model, and inject the preliminary diagnosis candidate set as input conditions into the digital twin simulation environment for reverse inference verification, screen out the real root cause that best matches the actual sensor characteristics, and obtain the simulation verification results. S5. Based on the simulation verification results, generate the final fault tracing report and repair suggestions through the large language model, and present them to the user through the digital twin interactive interface.

2. The industrial knowledge question answering and fault analysis method based on a large language model according to claim 1, characterized in that, The specific process for extracting the key entity set is as follows: S11. Obtain the natural language fault query command from the user through speech recognition conversion and text input, perform noise reduction and stop word removal on the natural language fault query command, and use a pre-trained industrial domain word embedding model to convert the processed natural language fault query command into a high-dimensional command semantic vector. S12. Input a compound prompt word with strict format constraints into the large language model. The compound prompt word contains the following: Role setting module: Limited to the large language model as an expert in extracting fault features of industrial equipment; Context Injection Module: Injects the current factory's system architecture outline, which includes a hierarchical tree structure of workshop layer, production line layer, equipment layer, and component layer; Task execution module: Maps the semantic vector of the instruction back to text, parses it based on the compound prompt words through a large language model, and outputs a coarse extracted entity set, which includes candidate device descriptions, candidate system level descriptions, and candidate fault phenomenon descriptions. S13. The coarsely extracted entity set is forcibly aligned and standardized using a pre-built standard database of factory assets to obtain a set of key entities.

3. The industrial knowledge question answering and fault analysis method based on a large language model according to claim 2, characterized in that, The specific process of forcibly aligning and standardizing the coarsely extracted entity set is as follows: S131. Obtain candidate device descriptions from the coarsely extracted entity set, and perform text matching based on cosine similarity between the candidate device descriptions and the standard device list in the factory asset standard database. If the text matching similarity is higher than a preset similarity threshold, then map the candidate device description to a unique standard device ID. If the match fails, a follow-up question mechanism is triggered, requiring the user to provide the accurate device name; S132. Based on the mapped standard equipment ID, trace back the parent node to which the standard equipment ID belongs in the association topology tree of the factory asset standard database to determine the precise system level corresponding to the standard equipment ID. S133. Align the candidate fault phenomenon descriptions with the standard industrial fault feature vocabulary, and transform the colloquial candidate fault phenomenon descriptions into standard quantitative description labels. S134. Package the verified and standardized standard device ID, the precise system hierarchy, and the normalized fault feature tags to generate the final set of key entities.

4. The industrial knowledge question answering and fault analysis method based on a large language model according to claim 1, characterized in that, The specific process for acquiring time-series data from dynamic sensors is as follows: S21. Obtain the set of key entities, call the pre-built factory three-dimensional space mesh model, and map the standard equipment IDs contained in the set of key entities to the corresponding three-dimensional space coordinate nodes in the factory three-dimensional space mesh model as anomaly source nodes. S22. Using a non-intrusive topology identification method, the dynamic operating characteristics of each node in the industrial site are read in real time. By calculating the data cross-correlation function between adjacent nodes, the physical connection weights between the abnormal source node and surrounding nodes are dynamically updated, and the upstream and downstream related nodes that have a real physical linkage relationship with the abnormal source node are locked. S23. Based on the time delay and amplitude attenuation of abnormal fluctuation propagation in historical time series data, dynamically calculate the fault propagation attenuation factor between the abnormal source node and each upstream and downstream related node. In each upstream and downstream related node, remove edge nodes whose fault propagation attenuation factor is lower than the preset effective threshold, and retain strongly related nodes to obtain a set of related nodes. S24. Based on the extracted fault phenomenon description, match the corresponding feature frequency band in the fault feature dictionary, and adaptively calculate and set the extraction range of dynamic sensor time series data according to the Nyquist sampling frequency of the feature frequency band and the preset fault latency empirical value to obtain the preset time window. S25. Send targeted data extraction instructions to the SCADA system in the industrial field through a unified API interface. The real-time data stream obtained undergoes preliminary data cleaning, including processing missing values, removing glitches and noise caused by sensor malfunctions, and aligning upstream and downstream sensor data with different sampling frequencies onto a unified time axis through a linear interpolation algorithm to generate high-dimensional, structured dynamic sensor time-series data.

5. The industrial knowledge question answering and fault analysis method based on a large language model according to claim 1, characterized in that, The specific process for generating quantized data analysis vectors is as follows: S301. Acquire dynamic sensor time series data and input it into a pre-trained temporal autoencoder in a deep data analysis platform. The encoder part of the temporal autoencoder performs multi-scale one-dimensional convolution on the input sequence to filter out redundant background information in the dynamic sensor time series data and retain abnormal fluctuation characteristics that characterize the operating status of the device. S302. Extract the latent space state vector of the bottleneck layer of the time-series autoencoder as the quantization basis vector characterizing the dynamic anomaly features of the current industrial equipment. S303. Construct a multilayer perceptron as a semantic projection network, input the quantized basic vector into the semantic projection network, and fine-tune the training on the historical sensor data-fault description text alignment dataset using the cross-entropy loss function, so that the semantic projection network linearly maps the quantized basic vector to the natural language semantic space of the large language model, generating a quantized data analysis vector that can be understood by the large language model. The quantized data analysis vector is mathematically equivalent to a soft prompt describing the current dynamic operation anomaly.

6. The industrial knowledge question answering and fault analysis method based on a large language model according to claim 1, characterized in that, The specific process of constructing integrated prompts is as follows: S311. Obtain the set of key entities and search for them in the local static knowledge base. The local static knowledge base includes equipment maintenance manuals, historical work orders and expert experience graphs. The search process adopts a text embedding model based on the dual-tower BERT architecture to convert the strongly related historical text paragraphs retrieved into static text vectors. S312. Using the generated quantitative data analysis vector as the query and the static text vector as the key and value, calculate the attention weight of dynamic features on each static knowledge item. S313. Based on the attention weights, the static text vectors are weighted and summed, and then concatenated with the quantized data analysis vectors to generate a multimodal fusion feature vector. S314. The multimodal fusion feature vector is converted into a structured text representation through a reverse decoder to construct the final fusion prompt word.

7. The industrial knowledge question answering and fault analysis method based on a large language model according to claim 1, characterized in that, The specific process for obtaining the simulation verification results is as follows: S41. After obtaining the fusion prompt words, perform multi-step reasoning calculations, output a preliminary diagnostic candidate set, call the preset instruction compilation agent, and parse the preliminary diagnostic candidate set into a structured simulation control vector executable by the digital twin system. The structured simulation control vector includes the topology node ID to be operated, the system parameter type to be modified, and its corresponding inference assignment. S42. Based on the ID of the topology node to be operated, coordinate system addressing is performed in the real-time three-dimensional space mesh model to lock the corresponding virtual device entity and trigger the visual rendering engine to mark the corresponding topology node with a specific highlight color. S43. Cut off the real-time data feedforward flow of the target physical topology network region in the digital twin system, so that the digital twin system enters an independent offline simulation branch. Inject the structured simulation control vector containing the deduction assignment into the mechanism model in the offline simulation branch. Based on the injected fault parameters and combined with historical boundary conditions, perform forward differential solution and simulation operation with a preset time step to generate the simulated sensor time series data matrix that each associated node should generate under the current fault assumption. S44. Acquire the dynamic sensor time series data and the simulated sensor time series data matrix for feature comparison. Based on the sequence alignment algorithm of Pearson correlation coefficient, calculate the morphological fit between the two time series data streams and generate a quantified simulation matching score. S45. Traverse the simulation matching scores corresponding to the preliminary diagnostic candidate set, lock the preliminary diagnostic candidate set of the simulation matching scores as the true root cause, record the corresponding inference parameters and comparison curves, and package them to generate structured simulation verification results.

Images