Industrial fault diagnosis method and system based on natural language fault features and large model knowledge enhanced reasoning

Abstract

The invention relates to the technical field of industrial equipment state monitoring and intelligent fault diagnosis, and discloses an industrial fault diagnosis method and system based on natural language fault features and large model knowledge enhanced reasoning, and the method comprises the steps: obtaining a multi-source monitoring and state analysis result from detected equipment in an equipment operation stage; converting the features into a fault feature text of a structured natural language; carrying out vectorization coding on the fault feature text, executing similarity retrieval in a pre-constructed industrial fault knowledge base, and recalling knowledge fragments; and based on the fault feature text and the recall knowledge fragment, constructing a reasoning prompt word, inputting the reasoning prompt word into a large language model for knowledge enhanced reasoning, and generating a diagnosis result. According to the method, the problems that numerical evidence and semantic knowledge are difficult to unify, the knowledge coverage and updating cost is high, and cross-working-condition migration and conclusion consistency are insufficient in an existing intelligent diagnosis technology based on a rule base or a knowledge graph are effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment status monitoring and intelligent fault diagnosis, and in particular to an industrial fault diagnosis method and system based on natural language fault characteristics and large model knowledge enhanced reasoning. Background Art

[0002] Equipment health management (PHM) in industries such as manufacturing, energy, and transportation has gradually evolved from diagnosis based on expert experience / rule bases to an integrated monitoring-analysis-decision-making framework driven by multiple signal sources. Traditional methods focus on a single modality (such as vibration) or a small number of statistical threshold judgments, making it difficult to stably migrate under complex working conditions and across equipment models. Although data-driven methods can extract features such as spectral peaks / octaves, axis trajectory shapes, and trend slopes from signals such as vibration, temperature, and current, they are still insufficient in terms of knowledge reuse, explainable reasoning, and evidence traceability.

[0003] In recent years, the capabilities of large language models (LLMs) in natural language understanding, knowledge question answering, and causal reasoning have provided a new path for closing the industrial "data-knowledge-reasoning" loop. However, LLMs are natively oriented towards textual semantics, with limited ability to parse numerical / shape-based evidence such as spectra, trajectories, and trends. Furthermore, they lack systematic knowledge modeling and enhancement mechanisms for the industrial field. As a result, when directly used for high-reliability diagnosis, there are shortcomings in evidence constraints, cross-condition generalization, and conclusion consistency. Therefore, there is an urgent need for a technical route that can connect multi-source numerical evidence → standardized semantic representation → domain knowledge enhancement → explainable reasoning diagnosis, so as to achieve robust diagnosis and traceable decision-making for rotating equipment such as pumps, compressors, fans, and motors under real working conditions.

[0004] CN118312896A discloses a fault diagnosis expert system based on a large language model. It takes a rule base expressed in natural language as its core, extracts rule features by LLM, and then the corresponding values ​​are calculated by sensors and judged according to the rules. This solution helps alleviate the hard coding of rules, but its diagnostic link is essentially still rule → value verification. It has not established a unified transcription mechanism from numerical or shape evidence such as spectrum / axis trajectory / trend to standardized natural language fault features, nor has it disclosed the vector similarity Top-K knowledge recall and evidence splicing prompt word process for equipment type-component-working condition context, making it difficult to maintain robustness and traceability across equipment and complex working conditions.

[0005] CN117271779A discloses a fault analysis method that combines a large model with a knowledge graph. It adopts a large model + knowledge graph (KG) paradigm: keywords are extracted from the user's natural language input, and the fault type is determined and a solution is given based on KG adjacency retrieval. This method is friendly to text scenarios, but lacks support for structured access and semantic alignment of multi-source numerical signals. The KG construction and maintenance costs are high and the granularity is limited. There is a lack of vectorized cross-document recall and structured output constraints / consistency verification and fallback mechanisms for unstructured materials such as manuals / cases / experiences.

[0006] In summary, existing intelligent diagnosis technologies based on rule bases or knowledge graphs generally have problems such as difficulty in unifying numerical evidence and semantic knowledge, high cost of knowledge coverage and updating, and insufficient consistency in cross-working condition migration and conclusions. The present invention proposes a systematic solution to the key problem of semanticizing multi-source numerical / shape fault features and integrating them with vectorized knowledge bases to drive large models for knowledge-enhanced reasoning. Summary of the Invention

[0007] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract of the specification and the title of the invention of this application to avoid blurring the purpose of this section, the abstract of the specification and the title of the invention, and such simplifications or omissions cannot be used to limit the scope of the invention.

[0008] In view of the above existing problems, the present invention is proposed.

[0009] To solve the above technical problems, the present invention provides the following technical solutions: during the equipment operation phase, multi-source monitoring and status analysis results are obtained from the equipment under test, wherein the analysis results include at least vibration spectrum characteristics, axis trajectory characteristics, and operating parameter trend characteristics; Converting the features into fault feature text in structured natural language; Vectorize and encode the fault feature text, perform similarity search in a pre-built industrial fault knowledge base, and recall knowledge fragments related to equipment types, components, and working conditions; Based on the fault feature text and the recalled knowledge fragments, reasoning prompt words are constructed and input into the large language model for knowledge-enhanced reasoning to generate a diagnosis result including the fault type, affected components, diagnostic basis and disposal suggestions.

[0010] As a preferred solution of the industrial fault diagnosis method based on natural language fault signatures and large model knowledge enhanced reasoning according to the present invention, the acquisition of multi-source monitoring and status analysis results from the equipment under test includes: Perform fast Fourier transform on the original vibration signal, extract representative spectrum peaks based on a preset peak prominence threshold, merge similar peaks according to frequency resolution, and determine the fundamental frequency and its multiples; The X and Y displacements are collected to form an axis trajectory point set. The trajectory is translated to align with the main direction, the curvature disturbance and distance disturbance are calculated, and the number of trajectory intersections is detected. Based on this, the trajectory is determined to be an 8-shaped, oblate ellipse, banana-shaped, or normal ellipse. The mean, standard deviation, and trend slope of slowly varying parameters such as temperature, current, and vibration amplitude are calculated within a sliding time window. Dynamic threshold intervals are then used to determine over-limit, volatility anomalies, and trend anomalies to generate structured state analysis results.

[0011] As a preferred solution of the industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning described in the present invention, the generation of the fault feature text includes: Extract the main peak frequency, octave / sideband interval and amplitude ratio of the spectrum features, and generate sentences according to the preset word model; Obtain the trajectory shape of the axis trajectory feature and generate sentences according to the preset word model; The mean, volatility and trend slope of the parameter trend characteristics are calculated, and the quantitative results are mapped to level labels and synthesized into a multi-sentence structured text description.

[0012] As a preferred solution of the industrial fault diagnosis method based on natural language fault signatures and large model knowledge-enhanced reasoning described in the present invention, the similarity search is performed in a pre-built industrial fault knowledge base to recall knowledge fragments related to equipment types, components, and operating conditions, including: Encoding the fault feature text into a query vector via an embedding model; Unstructured knowledge from maintenance manuals, historical cases, and expert experience is segmented and embedded into codes based on equipment type, fault component, and fault characteristics to create a vector index. Perform Top-K search based on cosine similarity and filter by device model, working condition label and timestamp to return the most relevant knowledge fragments and their source identifiers for the current scenario.

[0013] As a preferred solution of the industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning described in the present invention, the industrial fault knowledge base includes structured data and unstructured documents, wherein: The structured data includes parameter threshold tables, component hierarchical structures, and common fault cause tables; Unstructured documents include maintenance manuals, historical failure cases, and expert experience entries; Unstructured documents are segmented into paragraphs and identifiers and encoded into vectors using an embedding model. A vector index is established to support Top-K retrieval and filtering based on device model, working condition label, and timestamp.

[0014] As a preferred solution of the industrial fault diagnosis method based on natural language fault features and large model knowledge-enhanced reasoning described in the present invention, a reasoning prompt word is constructed based on the fault feature text and the recalled knowledge fragment, and is input into the large language model for knowledge-enhanced reasoning, including: Construct inference prompt words based on prompt word templates of roles, tasks, constraints, and formats; The fault feature text and the Top-K knowledge fragments are spliced ​​and input into the large language model, and the fault type, fault manifestation, fault analysis and reasoning process, diagnosis result confidence and maintenance suggestions are output according to structured fields; When the matching degree between the fault feature text and the Top-K knowledge fragment is lower than the threshold, the fault type is output as no fault, the confidence level of the diagnosis result is output as low, and the remaining fields are left blank.

[0015] As a preferred solution of the industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning described in the present invention, the construction of the reasoning prompt words adopts a four-element template of role, task, constraint, and format. The four-element template includes at least: The role is set as an industrial fault diagnosis expert; The task is to combine fault feature text and knowledge fragments to perform causal analysis and provide executable disposal suggestions; The constraints include unified units, standardized component names, and conclusions traceable to specific evidence items; The format requires output in structured fields.

[0016] As a preferred solution of the industrial fault diagnosis system based on natural language fault characteristics and large model knowledge enhanced reasoning described in the present invention, it includes: Multi-source sensor module, used to collect vibration, temperature, sound, and electrical quantity signals of the device under test and output them through standard industrial interfaces; The data acquisition and analysis unit is electrically connected to the multi-source sensor module and is used to perform signal conditioning, analog-to-digital conversion, filtering and standardization on the collected signals and generate status analysis results and fault characteristic data; The fault diagnosis and reasoning unit is in communication with the data acquisition and analysis unit, and is used to transcribe the fault feature data into fault feature text, perform vector retrieval on the fault knowledge base, perform large model reasoning and diagnosis, and output structured diagnosis results; A result display service module is in communication with the fault diagnosis and reasoning unit, and is used to receive and provide diagnostic results, evidence fragments and original / derived data in a service manner; A display unit, for displaying a spectrum diagram, axis trajectory, parameter trend, alarm information and the structured diagnosis results in a graphical interface, and supporting interactive query; External communication and data interface module, used to realize data interaction and remote synchronization with the host computer or PLC system; The power management module is used to supply power to the above units and modules and provide overvoltage, undervoltage and overcurrent protection.

[0017] As a preferred solution of the industrial fault diagnosis system based on natural language fault signatures and large model knowledge enhanced reasoning according to the present invention, the system further includes one or more processors; A memory storing operable instructions, wherein when the instructions are executed by the one or more processors, the one or more processors are caused to perform operations, wherein the operations include the process of the aforementioned industrial fault diagnosis method based on natural language fault characteristics and large model knowledge enhanced reasoning.

[0018] As a preferred embodiment of the computer-readable medium for storing software described in the present invention, the software includes instructions that can be executed by one or more computers, and the instructions enable the one or more computers to perform operations through such execution, and the operations include the process of the industrial fault diagnosis method based on natural language fault characteristics and large model knowledge enhanced reasoning as mentioned above.

[0019] The beneficial effects of the present invention are as follows: the present invention improves the portability and evidence traceability across working conditions and models through structured collection and textual expression of multi-source data; through vector retrieval and label filtering, unstructured manuals / cases are effectively incorporated into the reasoning context, alleviating the problems of high knowledge coverage and update costs; through controlled prompt words and structured output, the causal chain and basis references are strengthened, risks are reduced, and on-site execution is facilitated; at the same time, full-link interface standardization is achieved, which facilitates edge deployment and system integration, has good maintenance convenience, and significantly improves the reliability of diagnostic decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort. Among them: Figure 1Schematic diagram of the process of the industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning shown in the present invention; Figure 2 This is an example diagram of spectrum data shown in the present invention; Figure 3 This is the axis trajectory diagram shown in the present invention; Figure 4 This is a schematic diagram of the module structure distribution of the industrial fault diagnosis system based on natural language fault characteristics and large model knowledge enhanced reasoning shown in the present invention. DETAILED DESCRIPTION

[0021] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0022] Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without making any creative work should fall within the scope of protection of the present invention.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] According to an embodiment of the present invention, Figure 1 The flowchart shown is an industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning, which specifically includes the following steps: S1. During the equipment operation phase, obtain multi-source monitoring and status analysis results from the equipment under test, the analysis results including at least vibration spectrum characteristics, axis trajectory characteristics, and operating parameter trend characteristics; S2, converting the features into fault feature text in structured natural language; S3. Vectorize and encode the fault feature text, perform similarity search in the pre-built industrial fault knowledge base, and recall knowledge fragments related to equipment types, components, and working conditions; S4. Construct reasoning prompt words based on the fault feature text and recalled knowledge fragments, input them into the large language model for knowledge-enhanced reasoning, and generate diagnostic results that include fault type, affected components, diagnostic basis, and disposal suggestions.

[0025] It should be noted that compared with the existing technology, the diagnostic method provided by the embodiment of the present invention converts the multi-source monitoring and status analysis results (spectral characteristics, axis trajectory characteristics, parameter trend characteristics) into a structured natural language fault feature text, then performs vectorized encoding and performs similarity retrieval in the pre-built industrial fault knowledge base to recall knowledge fragments related to equipment types, components, and working conditions. Then, the LLM is driven by a prompt word template in the role-task-constraint format to complete knowledge enhancement reasoning, and outputs a structured conclusion including the fault type, affected components, diagnostic basis and disposal suggestions, thereby establishing a transferable, verifiable and traceable closed loop between numerical evidence-semantic knowledge-model reasoning.

[0026] Preferably, step S1 forms structured features covering the frequency domain, trajectory morphology and slowly changing parameters by synchronously acquiring and analyzing multiple types of signals during the equipment operation phase, thereby achieving comprehensive characterization and evidence retention of fault signs; thereby reducing the risk of missed detection of a single signal, improving the sensitivity and stability of early anomaly identification, and providing basic data with clear sources and traceability for subsequent semantic processing and reasoning.

[0027] Preferably, step S2 achieves consistency and comparability of feature expressions under different equipment and different working conditions by transcribing quantitative results such as spectrum peaks, octaves / sidebands, trajectory shapes and parameter trends into fault feature texts with controlled vocabulary and unified sentences; thereby facilitating subsequent retrieval and model consumption, reducing ambiguity caused by differences in manual descriptions, and retaining the correspondence with the original data, which is convenient for review and tracking.

[0028] Preferably, step S3 vectorizes the fault feature text and performs similarity retrieval in the industrial fault knowledge base organized by equipment type, component and working condition to achieve rapid association with historical cases, maintenance manuals and expert experience; thereby improving the relevance and pertinence of the retrieval results to the current scenario, reducing interference from irrelevant information, shortening the time to locate the cause and find disposal suggestions, and providing citation-capable evidence fragments for subsequent reasoning.

[0029] Preferably, step S4 inputs the fault feature text and the recall knowledge fragment into the big model according to the preset prompt word template, and outputs a structured result including the fault type, affected components, diagnostic basis and disposal suggestions, thereby realizing causal chain reasoning and interpretable conclusion generation under the support of external knowledge, thereby improving the accuracy and consistency of diagnosis.

[0030] The following combination Figure 2 、 Figure 3 The schematic diagrams shown and some preferred or optional examples of the present invention more specifically describe the implementation process and / or effects of certain examples of the present invention.

[0031] It should be noted that in order to achieve the conversion of structured fault features into natural language descriptions, three key features from the state analysis module are used as input data: vibration spectrum features, axis trajectory features, and parameter trend features. The data format is as follows: Vibration spectrum: By analyzing the original vibration velocity or acceleration signal Perform fast Fourier transform (FFT) to obtain the frequency domain amplitude spectrum , whose structure is a set of key-value pairs of frequencies and corresponding amplitudes, namely:

[0032] in, represents the frequency axis, is the corresponding amplitude sequence.

[0033] Axis trajectory: The two-dimensional motion trajectory in the rotor cross-sectional plane is obtained by a dual-channel axial displacement sensor, and the sequence of sampled trajectory points is expressed as:

[0034] in, For the The horizontal displacement of the sampling points, For the The vertical displacement of the sampling points constitutes a continuous trajectory graph. After normalization, it forms a standard coordinate point array input for the shape recognition model to determine the trajectory category (such as ellipse, figure 8, eccentricity, etc.). The input form is a two-dimensional vector set and no additional label information is required.

[0035] Parameter characteristics: At least include slow-changing monitoring signals such as temperature, current, vibration amplitude, etc., and record their changes in the form of time series , and extracts the following key statistics as input: average value:

[0036] Volatility (Standard Deviation):

[0037] Trend slope (least squares method):

[0038] in, At the sampling time The value of the monitored parameter (such as temperature, current or vibration amplitude), For the The sampling time point, is the total number of sampling points.

[0039] Furthermore, the feature generation model generates a text description of the fault features that can be understood by the large language model. This text description serves as input for subsequent diagnostic reasoning, including the generation of vibration spectrum fault feature descriptions, axis trajectory fault feature descriptions, and parameter anomaly descriptions: (1) Generation of vibration fault feature description Input data graph example Figure 2 As shown in the figure, the vibration velocity or acceleration signal is collected under operating conditions, and the spectrum data is obtained through fast Fourier transform (FFT), which is expressed as a mapping relationship between a set of amplitude sequences and the corresponding frequency axis:

[0040] in, represents the frequency axis, is the corresponding amplitude sequence; In order to extract key spectral features, the significant peak extraction threshold prominence is first calculated based on the dynamic range of the spectrum amplitude. :

[0041] in, It is an adjustable coefficient (the default value is 0.2) used to control the frequency peak sensitivity; On this basis, spectrum peak search is performed to identify all peaks with a prominence greater than The peak point is recorded as:

[0042] in, For the The frequency corresponding to the spectrum peak is is the set of all peak points; In order to reduce the impact of noise and merge similar frequency components, the similar peak merging rule is introduced, and the frequency spacing filter setting value is:

[0043] in, is the frequency resolution; The frequency spacing is less than The multiple peaks of are merged into their mean frequency to obtain the final representative frequency peak set ,in After the merger The representative peak frequencies, is the final number of representative frequency peaks; By default, the first frequency peak is taken as the fundamental frequency. , the remaining frequency peaks are classified according to their ratio to the fundamental frequency: like ,in , it is considered to be Frequency doubling; The remaining frequencies are recorded as “other frequency peaks”; Generate spectral feature description text of natural language structure, for example: Output starts: "The fundamental frequency of a vibration measurement point is 100Hz, the second harmonic frequency is 200Hz, and the other frequency peaks are 1000Hz." The vibration spectrum feature description is used as the input of the reasoning model to assist in subsequent fault causal analysis and diagnosis conclusion generation.

[0044] (2) Generation of axis trajectory fault feature description An example of an input example is Figure 3 As shown in the figure, when the equipment is running, the axial displacement signals in the X and Y directions are collected to construct a two-dimensional trajectory point set:

[0045] in, For the The horizontal displacement of the sampling points, For the The vertical displacement of each sampling point; To enhance the versatility and robustness of the analysis, the trajectories are first normalized, including centroid translation and principal direction alignment; Calculate the coordinates of the center (center of mass) of the trajectory:

[0046] The trajectory points after translation are:

[0047] Perform principal component analysis (PCA) on the trajectory points and extract the variance ratio of the first principal component to the second principal component as the criterion for the degree of ellipse flattening:

[0048] in, is the eigenvalue corresponding to the main direction; Calculate the curvature sequence of the trajectory after translation:

[0049] in, No. The curvature of a sampling point is used to measure the curvature of the trajectory at that point. is the first-order derivative of the trajectory in the x and y directions after translation, representing the velocity component, is the second-order derivative of the trajectory in the x and y directions after translation, representing the acceleration component; Extract the extreme value and median of the curvature and define the curvature perturbation ratio:

[0050] in, is the curvature of all sampling points A collection of sequences; The Euclidean distance from each point to the centroid of the trajectory:

[0051] Calculate its mean and standard deviation to get the distance disturbance ratio:

[0052] in, is the Euclidean distance from all sampling points to the trajectory centroid A collection of sequences; For trajectories at multiple rotation angles Reprojection is performed to detect whether there are intersections between line segments and count the maximum number of intersection points The specific process is as follows: For any rotation angle , perform a two-dimensional rotation transformation on the trajectory points around the origin to obtain the rotated coordinates:

[0053] Define a set of line segments consisting of continuous trajectory points:

[0054] For any two non-adjacent line segments (line segments , ), determine whether it meets the line segment intersection condition (using cross product judgment): judge:

[0055] The total number of intersections at this angle is ; For multiple angles (Equally spaced in Repeat the above rotation and intersection detection operations, and finally take the maximum number of intersection points as one of the shape features of the trajectory:

[0056] Based on the above multiple shape feature indicators, the trajectories are classified into the following typical types and natural language description results are generated: like , identified as: 8-shaped; like and and , identified as: flat oval; like and and , identified as: banana shape; Otherwise, it is identified as a normal ellipse.

[0057] For example, the final axis trajectory shape description output is: Output starts: "The axis trajectory is in the shape of an 8"; Output starts: "The axis trajectory is a flat ellipse"; Output begins: "The trajectory shape is a normal ellipse" and other standard descriptive text.

[0058] (3) Parameter exception description For slowly changing monitoring parameters (such as temperature, current, and vibration amplitude) during the operation of industrial equipment, statistical feature extraction and trend analysis are performed within a set time window to identify typical abnormal behaviors, including: increased parameter volatility, exceeding reference thresholds, and rapid trend increases. Standardized natural language fault description text is generated. Assume that the parameter time series is:

[0059] To ensure the adaptability of anomaly determination, a reference model is constructed for each type of parameter during the equipment commissioning or stable operation phase, including the following benchmark indicators, which are used as a reference for parameter anomaly generation; Historical average:

[0060] Historical standard deviation:

[0061] Dynamic threshold range:

[0062] in, For the The parameter value of a sampling period, is the number of historical samples, is the historical mean of the parameter, which is used to describe the long-term average level. is the historical standard deviation of the parameter, which is used to measure the fluctuation range of the parameter. are the dynamic lower and upper thresholds of the parameters, Is the relaxation coefficient, used to adjust the threshold range, generally taken The value between , which is used to dynamically determine whether the parameter is abnormal and can adapt to the fluctuation characteristics of different parameters.

[0063] In an optional embodiment, the abnormal feature extraction and generation includes: a. Abnormal volatility Calculate the standard deviation of the sampled data in the current time window:

[0064] in, is the mean of the parameter values ​​in the current time window; If the following conditions are met:

[0065] It is determined that the parameter volatility has increased significantly, and a natural language description of the parameter characteristics is generated: Output begins: "This parameter fluctuates widely and shows an unstable trend." in, Is the relaxation coefficient, used to adjust the threshold range, generally taken The value between .

[0066] b. Over-limit abnormality For each sampling point Execute threshold judgment: or

[0067] If true, the current parameters are considered to have exceeded the safe operating range, and the parameter characteristics generate a text description: The output begins: "The current value of this parameter has exceeded the set threshold range, and there is a potential risk." c. Abnormal trend To evaluate the trend of parameters over time, the system Perform a linear fit on the point set and calculate the trend slope:

[0068] If satisfied:

[0069] It is considered that the parameter shows a rapid upward trend, and a natural language description of the parameter characteristics is generated: The output begins: "This parameter has risen rapidly in a short period of time and may have a trend of continued abnormality." in, Is the relaxation coefficient, used to adjust the threshold range, generally taken The value between .

[0070] When multiple features meet abnormal conditions simultaneously, the system automatically combines multiple pieces of information to form a complete, structured description of the fault characteristics. For example, "The temperature parameter has risen rapidly over the past 10 minutes, with a slope of 2.4°C / min, significantly higher than the historical average rate of increase of 1.0°C / min. The current value of 58.2°C has exceeded the upper threshold of 55°C, and the volatility has increased significantly." (4) Fault feature description synthesis Combine the natural language description results of various fault characteristics such as vibration spectrum, axis trajectory, parameter trend, etc. to generate a unified fault feature text description , as one of the inference inputs:

[0071] in, Descriptive text generated for spectrum analysis (e.g., the fundamental frequency of a measurement point is 100Hz, there is a 2X frequency multiplication, and an abnormal frequency peak of 1000Hz appears). is the result of axis trajectory analysis (e.g. axis trajectory is in the shape of figure 8), For trend and over-limit analysis results (such as rapid temperature rise, the current value of 58.2℃ has exceeded the upper threshold of 55℃), the final synthesis As the main input field of the fault knowledge retrieval and reasoning diagnosis module, it is the result of natural language splicing in structure and the semantic representation of structured features in content.

[0072] It should be noted that fault knowledge base retrieval uses the text description of fault features as the query condition to retrieve structured and unstructured fault knowledge, supporting causal chain completion and professional support in large model reasoning tasks, including: (1) Fault vector knowledge base The knowledge base content is derived from industrial field data, including but not limited to equipment manuals, maintenance manuals, historical failure case records, expert experience documents and question and answer records; All knowledge texts are segmented according to the structure of “equipment type + fault component + fault characteristics”. The knowledge fragments are:

[0073] By embedding the model Convert text to a high-dimensional vector:

[0074] All embedded vectors are stored in the vector database to build an index.

[0075] (2) Query and similarity retrieval Fault feature text description (For example, the axis trajectory is in the shape of an 8, and the vibration spectrum has a 2X frequency multiplication) It is also converted into a query vector through the embedding model:

[0076] in, is a text description of the fault characteristics, For text embedding models, Described by text The converted query vector; Perform vector similarity search and calculate query vector Vector collection of knowledge fragments Similarity:

[0077] in, is the vector representation of the knowledge fragment (document fragment), is the vector dot product operation, 、 is the L2 norm (length) of the vector, is the cosine similarity between the query vector and the knowledge fragment vector, with a value range of , the closer it is to 1, the higher the correlation; Perform Top-K search by similarity and select the most relevant Knowledge snippets:

[0078] in, The most relevant A collection of knowledge fragments, For the Pieces of knowledge, After sorting by similarity value from high to low, take the first The operation of the bar, the number of knowledge fragments returned for the search; Finally return the knowledge fragment collection The content includes the maintenance knowledge, case descriptions, common fault causes and handling suggestions that are most relevant to the current fault characteristics, which are used to splice with the inference model prompt words.

[0079] Furthermore, the structured fault feature description A collection of knowledge fragments related to recall The unified semantic prompt words are spliced ​​together as the input of the large language model, and the inference model is used to perform the tasks of fault type judgment, causal chain construction and maintenance suggestion generation; (1) Fault reasoning input construction : Generated text description of fault features (natural language expression); : Semantic vector retrieval recall Fault knowledge snippets; : The prompt word template structure used for reasoning; For example, the input context of the inference model is represented as:

[0080] For example, the prompt word concatenation structure include: You are an expert in intelligent industrial equipment diagnosis with experience in industrial equipment operation and maintenance. Please perform professional fault diagnosis and analysis based on the fault characteristics described below and the provided "Fault Knowledge Base."

[0081] # Diagnostic requirements If the fault characteristics do not match the fault knowledge well, please output "no fault" in the fault type, output "low" in the diagnosis result confidence, and output empty values ​​in other fields.

[0082] The description of fault manifestations must strictly follow the fault characteristics, and fabricated descriptions are not allowed. # Output format Please output a structured troubleshooting report including: - "Fault Type": Match the most likely fault name - "Fault manifestation": Describe the intuitive phenomenon of the fault in combination with the fault characteristics - "Fault Analysis Reasoning Process": Detailed description of the diagnosis cause, citing parameter trends, spectrum, vibration, axis characteristics, etc. The analysis process should be no less than 100 words. - "Confidence of diagnosis result": The value is "High", "Medium" or "Low". The confidence of the diagnosis result matches the fault type feature description as strictly as possible. High means a high match, medium means a fair match, and low means a low match. - "Maintenance Suggestions": Provide professional repair or maintenance suggestions # Fault symptoms: {Fault feature text description Q} # Fault knowledge base: {Related fault knowledge K top} (2) Inference model function The inference model is a large language model , whose parameters are , which can be deployed for local models (such as Deepseek-7B, Qwen3-4B):

[0083] This function completes the mapping from fault phenomena and knowledge background to diagnostic conclusions.

[0084] (3) Output structure definition The model output is a structured diagnostic report:

[0085] Specifically expressed as: Fault type: ; Fault description: natural language text; Reasoning process: logical chain combining features such as spectrum, trend, trajectory, etc. Diagnostic confidence: ; Recommended actions: case-based maintenance recommendations; in, Indicates the fault type label identified by the system, It is a set of predefined fault types, such as typical fault categories such as bearing wear, rotor imbalance, and coupling misalignment.

[0086] Exemplarily, the fault reasoning diagnosis result output includes: Fault type: coupling misalignment Fault symptoms: The equipment bearing temperature rose from 20°C to 53°C, the vibration value reached 75 mm / s (far exceeding the 20 mm / s threshold), the vibration spectrum showed obvious 2X frequency components, and the axis trajectory was in the shape of an "8".

[0087] Fault analysis and reasoning process: Based on the sensor characteristic data, it was first observed that the equipment vibration seriously exceeded the standard (75 mm / s), far exceeding the fault threshold of 20 mm / s, indicating that the equipment had a significant mechanical problem. A clear 2X frequency multiplication feature appeared in the vibration spectrum, which is one of the typical signal characteristics of coupling misalignment. Secondly, the axis trajectory showed an "8" shape, which is a hallmark trajectory characteristic of coupling misalignment in mechanical fault diagnosis. In addition, the bearing temperature rose rapidly from 20°C to 53°C. Although it is not an extreme temperature, combined with the severe vibration, it indicates that the rotor system has a serious misalignment or imbalance problem.

[0088] Compared with the coupling misalignment item in the fault knowledge base, its typical characteristics are: "spectrum 2X frequency is prominent", "axis trajectory is 8-shaped", and "main shaft and motor vibration is increased", which are highly consistent with the current characteristics. Other fault types such as "bearing damage", "shaft bending", "internal friction", etc., although there are some increased vibration or temperature rise manifestations, they are not matched in terms of spectrum frequency doubling and trajectory characteristics.

[0089] Therefore, comprehensive analysis shows that the most likely fault is coupling misalignment.

[0090] Diagnosis confidence: high Maintenance recommendations: 1. After stopping the machine, focus on checking the alignment of the shafts on both sides of the coupling and use a laser alignment instrument to recalibrate the coupling installation. 2. Check the coupling body for wear or structural defects such as loose sleeves, and replace parts if necessary.

[0091] 3. Check for auxiliary damage caused by misalignment, such as bearing wear, shaft seal deformation, etc., and simultaneously investigate and deal with them 4. After troubleshooting, be sure to retest the vibration and temperature baselines before restarting to ensure the problem has been eradicated. 5. It is recommended to strengthen the online monitoring and trend analysis of 2X frequency multiplication during operation and establish an early warning mechanism for changes in centering status.

[0092] Reference Figure 4 Other aspects disclosed in the embodiments of the present invention further provide an industrial fault diagnosis system based on natural language fault signatures and large model knowledge enhanced reasoning, including: Multi-source sensor module, used to collect vibration, temperature, sound, and electrical quantity signals of the device under test and output them through standard industrial interfaces; The data acquisition and analysis unit is electrically connected to the multi-source sensor module and is used to perform signal conditioning, analog-to-digital conversion, filtering and standardization on the collected signals and generate status analysis results and fault characteristic data; It should be noted that the data acquisition and analysis unit consists of a sensor interface module, a signal conversion and processing module, a data cache module, a state recognition module, and an analysis model module; The fault diagnosis and reasoning unit is in communication with the data acquisition and analysis unit, and is used to transcribe the fault feature data into fault feature text, perform vector retrieval on the fault knowledge base, perform large model reasoning and diagnosis, and output structured diagnosis results; Exemplarily, the fault diagnosis reasoning unit includes an embedded AI chip, an edge reasoning module, a local knowledge base cache structure, an embedded thermal database, a model reasoning engine, and an interface processing module for fault feature text generation, fault knowledge base retrieval, and fault reasoning; A result display service module is connected to the fault diagnosis reasoning unit for receiving and providing diagnosis results, evidence fragments and original / derived data in a service manner; It should be noted that the result display service module includes a page display service module, a data interface service module and an agent workflow module, wherein the agent workflow module is used to trigger knowledge retrieval, question-answer linkage and alarm linkage processes based on the running status or user request; A display unit, connected to the fault diagnosis and reasoning unit and / or the result display service module, is used to display spectrum diagrams, axis trajectories, parameter trends, alarm information and structured diagnosis results in a graphical interface, and supports interactive query; The external communication and data interface module is connected to the fault diagnosis and reasoning unit and the result display service module to realize data interaction and remote synchronization with the host computer or PLC system; Exemplarily, the external communication and data interface module includes a USB interface, an RS485 serial port, an HDMI interface, an Ethernet interface, and an audio output interface; A power management module, which is used to supply power to the above units and modules and provide overvoltage, undervoltage and overcurrent protection; Among them, the multi-source sensor module transmits the processed feature data to the fault diagnosis and reasoning unit through the data acquisition and analysis unit. The fault diagnosis and reasoning unit provides the structured diagnosis results to the result display service module and the display unit for presentation, and outputs them to the outside through the external communication and data interface module.

[0093] The system also includes one or more processors and memory.

[0094] The memory is used to store operable instructions, which, when executed by the one or more processors, cause the one or more processors to perform operations, including the process of the industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning in the aforementioned embodiment, especially Figure 1 The process of the method shown.

[0095] Other aspects disclosed in the embodiments of the present invention further provide a computer-readable medium storing software, wherein the software includes instructions that can be executed by one or more computers, and the execution of these instructions causes the one or more computers to perform operations, including the process of the industrial fault diagnosis method based on natural language fault characteristics and large model knowledge enhanced reasoning of the aforementioned embodiment, especially Figure 1 The process of the method shown.

[0096] It should be appreciated that embodiments of the present invention can be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable memory.

[0097] The method may be implemented in a computer program using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program, wherein the storage medium so configured causes a computer to operate in a specific and predefined manner.

[0098] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with a computer system, however, the program can be implemented in assembly or machine language if desired.

[0099] In any case, the language may be a compiled or interpreted language.

[0100] Furthermore, the program can be run on an application specific integrated circuit programmed for this purpose.

[0101] The processes described herein (or variations and / or combinations thereof) may be performed under the control of one or more computer systems configured with executable instructions and may be implemented as code (e.g., executable instructions, one or more computer programs, or one or more applications) that collectively executes on one or more processors, by hardware, or a combination thereof. The computer program includes a plurality of instructions that can be executed by one or more processors.

[0102] Further, the method may be implemented in any type of computing platform operably connected to a suitable computer, including but not limited to a personal computer, a minicomputer, a mainframe, a workstation, a network or distributed computing environment, a separate or integrated computer platform, or in communication with a charged particle tool or other imaging device.

[0103] Aspects of the present invention may be implemented in machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optical read and / or write storage media, RAM, ROM, etc., such that it can be read by a programmable computer and, when the storage medium or device is read by the computer, can be used to configure and operate the computer to perform the processes described herein.

[0104] Additionally, the machine-readable code, or portions thereof, can be transmitted over a wired or wireless network.

[0105] The invention described herein includes these and other various types of non-transitory computer-readable storage media when such media include instructions or programs that implement the steps described above in conjunction with a microprocessor or other data processor.

[0106] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning, characterized by: include: During the equipment operation phase, multi-source monitoring and status analysis results are obtained from the equipment under test, wherein the analysis results include at least vibration spectrum characteristics, axis trajectory characteristics, and operating parameter trend characteristics; Converting the features into fault feature text in structured natural language; Vectorize and encode the fault feature text, perform similarity search in a pre-built industrial fault knowledge base, and recall knowledge fragments related to equipment types, components, and working conditions; Based on the fault feature text and the recalled knowledge fragments, reasoning prompt words are constructed and input into the large language model for knowledge-enhanced reasoning to generate a diagnosis result including the fault type, affected components, diagnostic basis and disposal suggestions.

2. The industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning according to claim 1 is characterized in that: The obtaining of multi-source monitoring and status analysis results from the device under test includes: Perform fast Fourier transform on the original vibration signal, extract representative spectrum peaks based on a preset peak prominence threshold, merge similar peaks according to frequency resolution, and determine the fundamental frequency and its multiples; The X and Y displacements are collected to form an axis trajectory point set. The trajectory is translated to align with the main direction, the curvature disturbance and distance disturbance are calculated, and the number of trajectory intersections is detected. Based on this, the trajectory is determined to be an 8-shaped, oblate ellipse, banana-shaped, or normal ellipse. The mean, standard deviation, and trend slope of slowly varying parameters such as temperature, current, and vibration amplitude are calculated within a sliding time window. Dynamic threshold intervals are then used to determine over-limit, volatility anomalies, and trend anomalies to generate structured state analysis results.

3. The industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning according to claim 1 or 2 is characterized in that: The generation of the fault feature text includes: Extract the main peak frequency, octave / sideband interval and amplitude ratio of the spectrum features, and generate sentences according to the preset word model; Obtain the trajectory shape of the axis trajectory feature and generate sentences according to the preset word model; The mean, volatility and trend slope of the parameter trend characteristics are calculated, and the quantitative results are mapped to level labels and synthesized into a multi-sentence structured text description.

4. The industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning according to claim 3 is characterized in that: The similarity search is performed in the pre-built industrial fault knowledge base to recall knowledge fragments related to equipment types, components and working conditions, including: Encoding the fault feature text into a query vector via an embedding model; Unstructured knowledge from maintenance manuals, historical cases, and expert experience is segmented and embedded into codes based on equipment type, fault component, and fault characteristics to create a vector index. Perform Top-K search based on cosine similarity and filter by device model, working condition label and timestamp to return the most relevant knowledge fragments and their source identifiers for the current scenario.

5. The industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning according to claim 4 is characterized in that: The industrial fault knowledge base includes structured data and unstructured documents, wherein: The structured data includes parameter threshold tables, component hierarchical structures, and common fault cause tables; Unstructured documents include maintenance manuals, historical failure cases, and expert experience entries; Unstructured documents are segmented into paragraphs and identifiers and encoded into vectors using an embedding model. A vector index is established to support Top-K retrieval and filtering based on device model, working condition label, and timestamp.

6. The industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning according to claim 1 is characterized in that: Constructing reasoning prompt words based on the fault feature text and the recalled knowledge fragments, and inputting them into the large language model for knowledge-enhanced reasoning, including: Construct inference prompt words based on prompt word templates of roles, tasks, constraints, and formats; The fault feature text and the Top-K knowledge fragments are spliced ​​and input into the large language model, and the fault type, fault manifestation, fault analysis and reasoning process, diagnosis result confidence and maintenance suggestions are output according to structured fields; When the matching degree between the fault feature text and the Top-K knowledge fragment is lower than the threshold, the fault type is output as no fault, the confidence level of the diagnosis result is output as low, and the remaining fields are left blank.

7. The industrial fault diagnosis method based on natural language fault features and large model knowledge enhanced reasoning according to claim 6 is characterized in that: The construction of the reasoning prompt word adopts a four-element template of role, task, constraint, and format, and the four-element template includes at least: The role is set as an industrial fault diagnosis expert; The task is to combine fault feature text and knowledge fragments to perform causal analysis and provide executable disposal suggestions; The constraints include unified units, standardized component names, and conclusions traceable to specific evidence items; The format requires output in structured fields.

8. An industrial fault diagnosis system based on natural language fault features and large model knowledge enhanced reasoning, characterized by: include: Multi-source sensor module, used to collect vibration, temperature, sound, and electrical quantity signals of the device under test and output them through standard industrial interfaces; The data acquisition and analysis unit is electrically connected to the multi-source sensor module and is used to perform signal conditioning, analog-to-digital conversion, filtering and standardization processing on the collected signals and generate status analysis results and fault characteristic data; The fault diagnosis and reasoning unit is in communication with the data acquisition and analysis unit, and is used to transcribe the fault feature data into fault feature text, perform vector retrieval on the fault knowledge base, perform large model reasoning and diagnosis, and output structured diagnosis results; A result display service module is in communication with the fault diagnosis and reasoning unit, and is used to receive and provide diagnostic results, evidence fragments and original / derived data in a service manner; A display unit, for displaying a spectrum diagram, axis trajectory, parameter trend, alarm information and the structured diagnosis results in a graphical interface, and supporting interactive query; External communication and data interface module, used to realize data interaction and remote synchronization with the host computer or PLC system; The power management module is used to supply power to the above units and modules and provide overvoltage, undervoltage and overcurrent protection.

9. The industrial fault diagnosis system based on natural language fault features and large model knowledge enhanced reasoning according to claim 8 is characterized in that: The system further includes one or more processors; A memory storing operable instructions, wherein the instructions, when executed by the one or more processors, cause the one or more processors to perform operations, wherein the operations include the process of the industrial fault diagnosis method based on natural language fault characteristics and large model knowledge enhanced reasoning as described in any one of claims 1 to 7.

10. A computer-readable medium storing software, characterized in that: The software includes instructions that can be executed by one or more computers, and the instructions, through such execution, enable the one or more computers to perform operations, and the operations include the process of the industrial fault diagnosis method based on natural language fault characteristics and large model knowledge enhanced reasoning as described in any one of claims 1 to 7.

Images