Equipment fault prediction management method and device based on large model

Abstract

The embodiment of the invention provides an equipment fault prediction management method and device based on a large model, and the method and device achieve the accurate evaluation of a state through the innovative design of a multi-modal feature fusion model, data integration and feature learning. A fault diagnosis system is constructed, and a reliable root cause analysis mechanism is established in combination with semantic analysis and case retrieval. Maintenance guidance is introduced, and the feasibility of a maintenance scheme is ensured through experience precipitation and priority evaluation. According to the method, the defects of the traditional technology in the aspects of feature extraction, fault diagnosis, maintenance guidance and the like are effectively overcome, and technical guarantee is provided for equipment management.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a method and apparatus for predictive management of equipment failures based on a large model. Background Technology

[0002] Existing equipment failure prediction and management methods have significant shortcomings. Traditional systems perform poorly in data acquisition and feature extraction, failing to effectively integrate multimodal information and affecting prediction accuracy.

[0003] Furthermore, existing technologies face bottlenecks in fault diagnosis and root cause analysis. Most systems lack robust semantic understanding mechanisms and case retrieval strategies, resulting in inaccurate diagnostic results.

[0004] The existing system has technical shortcomings in maintenance guidance. It lacks in-depth analysis of maintenance experience, making it difficult to provide effective guidance through a knowledge base, thus impacting maintenance efficiency. Solving these problems is crucial for improving equipment management. Summary of the Invention

[0005] To address the problems in existing technologies, this application provides a method and apparatus for predicting and managing equipment failures based on a large model. This method can effectively solve the shortcomings of traditional technologies in feature extraction, fault diagnosis, and maintenance guidance, and provide technical support for equipment management.

[0006] To solve at least one of the above problems, this application provides the following technical solution: Firstly, this application provides a device failure prediction and management method based on a large model, including: The system collects equipment operation data, sound vibration data, and image data. It performs unified sampling and missing value imputation on the operation data, converts the sound vibration data into spectral features, extracts visual features from the image data, constructs a multimodal feature extraction network, inputs the multimodal data into the feature extraction network to generate a fused feature vector, performs word segmentation and entity recognition on the maintenance records to extract fault types and maintenance steps, constructs an equipment component association network, calculates the contrast loss between feature vectors to obtain the equipment state vector, trains an equipment anomaly detection model, and uses a digital twin method to generate fault samples to expand the training set. The device state vector is input into a large language model encoder to generate a semantic representation. A retrieval-enhanced generation model is constructed. Historical maintenance cases are read to establish a vector retrieval library. The temporal dependencies of fault events are calculated. A component-level fault propagation network is constructed. The root cause probability of the fault is calculated based on the fault propagation network. The semantic representation is compared with the retrieval library to calculate a similarity score. Maintenance records of similar cases are extracted. A fault diagnosis model is trained. The root cause localization result is determined based on the confidence score. The system monitors real-time equipment status data, inputs the status data into an anomaly detection model to calculate anomaly scores, inputs anomaly data fragments into a fault diagnosis model to generate fault causes, extracts expert maintenance experience to construct a knowledge base, calculates the priority scores of maintenance operations, establishes a dependency constraint network for maintenance steps, generates an executable maintenance guidance plan, collects maintenance feedback data to update the vector retrieval library and knowledge base, writes the maintenance feedback data into the training sample set, and performs online updates to the anomaly detection model and fault diagnosis model.

[0007] Furthermore, it also includes: collecting operating parameter data from equipment sensors, collecting vibration waveform data from acoustic sensors, collecting equipment image data from industrial cameras, resampling the operating parameter data at fixed time intervals, filling missing values ​​using cubic spline interpolation, performing short-time Fourier transform on the vibration waveform data to obtain a time-frequency spectrum, and performing convolution operation on the image data to extract edge features and texture features; A multimodal feature extraction network comprising a temporal coding layer, a spectrum analysis layer, and a visual feature layer is constructed. The running parameter data is input into the temporal coding layer, the temporal spectrum is input into the spectrum analysis layer, and the edge features and texture features are input into the visual feature layer. An attention mechanism is used to calculate the weight coefficients of the three feature layers. Based on the weight coefficients, the three feature layers are weighted and combined to generate a fused feature vector.

[0008] Furthermore, it also includes: reading historical maintenance record text, extracting professional terms and keywords using dictionary segmentation methods, labeling fault locations and maintenance actions based on named entity recognition models, constructing the fault locations as nodes, constructing the physical connection relationships between components as edges, generating a device component association network, performing sequence labeling on the maintenance actions to construct a maintenance step table, and writing the fault locations and maintenance steps into the labeled dataset; The fused feature vectors are input into the contrastive learning network to calculate the contrast loss between positive and negative sample pairs. The feature encoder is trained to generate equipment state vectors. A digital twin model is constructed based on the equipment component association network. The digital twin model is input into the physical simulation engine to simulate fault scenarios under different working conditions and generate simulation data. The simulation data is combined with real data to train the equipment anomaly detection model.

[0009] Furthermore, it also includes: inputting the equipment status vector into the large language model encoder and mapping it into the natural language space; constructing fault description text based on prompt word templates; inputting the fault description text into the large language model to generate semantic vectors; reading historical maintenance case records; performing text segmentation and key information extraction on the maintenance cases; constructing a maintenance knowledge base containing fault phenomena, root cause analysis, and handling methods; and encoding the maintenance knowledge base into vector form to establish a retrieval database. Read the equipment fault event records, extract the fault occurrence time and affected components, calculate the time interval and propagation probability between component faults, write the propagation probability into the adjacency matrix to construct a fault propagation graph, calculate the in-degree and out-degree of component nodes based on the fault propagation graph, identify fault source nodes and affected nodes, construct a component-level fault propagation network, and calculate the probability distribution of fault paths.

[0010] Furthermore, it also includes: reading the node connection relationships in the fault propagation network, calculating the conditional probability of the fault propagation path, constructing a fault probability inference model based on a Bayesian network, substituting the equipment status data into the inference model to calculate the fault probability distribution of each component, performing nearest neighbor retrieval on the semantic representation vector, calculating the cosine similarity with historical cases in the retrieval database, and extracting the maintenance record with the highest similarity score as a reference case. The fault probability distribution and reference cases are combined to construct diagnostic input features. An evidence deep learning network is used to train the fault diagnosis model. The confidence interval of the diagnosis results is calculated based on the Dirichlet distribution. The confidence interval is converted into a confidence score. The fault diagnosis results are sorted according to the confidence score, and the result with the highest score is selected as the root cause localization output.

[0011] Furthermore, it also includes: collecting equipment operating status data, segmenting the status data into time windows, performing feature extraction on the multimodal data within each time window, inputting the feature vector into an anomaly detection model to calculate residual scores, identifying abnormal data segments based on an adaptive threshold method, inputting the abnormal data segments into a fault diagnosis model, combining the fault propagation network to calculate component-level fault probabilities, and generating fault cause analysis results; The system reads expert repair case database, extracts repair operation steps and key control parameters, constructs a scoring matrix including component importance, repair difficulty, and spare parts availability, calculates the weight coefficients of the scoring indicators based on the analytic hierarchy process, multiplies the weight coefficients with the scoring matrix to obtain the comprehensive score of the repair operation, prioritizes the repair operations according to the comprehensive score, and generates a repair task sequence.

[0012] Furthermore, it also includes: reading the maintenance task sequence, extracting the preorder and postorder relationships between maintenance steps, constructing a directed acyclic graph to represent step dependency constraints, generating the execution order of maintenance steps based on a topological sorting algorithm, combining the execution order with the fault diagnosis results, querying the expert knowledge base to extract maintenance operation specifications, filling the template of the maintenance operation specifications, and generating a maintenance guidance scheme that includes a tool list, maintenance steps, and quality inspection points. The system collects performance feedback data from maintenance personnel, extracts maintenance operation time and completion status, updates the maintenance feedback data to the vector retrieval database, calculates maintenance effectiveness scores to update the expert knowledge base, labels the maintenance feedback data as training samples, updates the parameters of the anomaly detection model using an incremental learning method, and adjusts the weight coefficients of the fault diagnosis model based on an online distillation method.

[0013] Secondly, this application provides a device for predicting and managing equipment failures based on a large model, comprising: The equipment data acquisition module is used to collect equipment operation data, sound vibration data, and image data. It performs unified sampling and missing value imputation on the operation data, converts the sound vibration data into spectral features, extracts visual features from the image data, constructs a multimodal feature extraction network, inputs the multimodal data into the feature extraction network to generate a fused feature vector, performs word segmentation and entity recognition on the maintenance records to extract fault types and maintenance steps, constructs an equipment component association network, calculates the contrast loss between feature vectors to obtain the equipment state vector, trains an equipment anomaly detection model, and uses a digital twin method to generate fault samples to expand the training set. The fault root cause diagnosis module is used to input the equipment state vector into the large language model encoder to generate semantic representation, construct a retrieval enhancement generation model, read historical maintenance cases to establish a vector retrieval library, calculate the temporal dependency of fault events, construct a component-level fault propagation network, calculate the fault root cause probability based on the fault propagation network, calculate the similarity score between the semantic representation and the retrieval library, extract maintenance records of similar cases, train the fault diagnosis model, and determine the root cause location result based on the confidence score. The equipment fault prediction module is used to monitor real-time equipment status data, input the status data into an anomaly detection model to calculate anomaly scores, input anomaly data fragments into a fault diagnosis model to generate fault causes, extract expert maintenance experience to build a knowledge base, calculate the priority scores of maintenance operations, establish a dependency constraint network for maintenance steps, generate an executable maintenance guidance plan, collect maintenance feedback data to update the vector retrieval library and knowledge base, write the maintenance feedback data into the training sample set, and perform online updates to the anomaly detection model and fault diagnosis model.

[0014] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the large model-based device fault prediction and management method.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the large-model-based equipment fault prediction and management method.

[0016] Fifthly, this application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the large model-based equipment fault prediction and management method.

[0017] As described above, this application provides a method and apparatus for equipment fault prediction and management based on a large model. Through innovative design of a multimodal feature fusion model, and by integrating data and learning features, it achieves accurate status assessment. A fault diagnosis system is constructed, and a reliable root cause analysis mechanism is established by combining semantic analysis and case retrieval. Maintenance guidance is introduced, and through experience accumulation and priority evaluation, the feasibility of maintenance plans is ensured. This method effectively addresses the shortcomings of traditional technologies in feature extraction, fault diagnosis, and maintenance guidance, providing technical support for equipment management. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the equipment fault prediction and management method based on a large model in an embodiment of this application. Figure 2 This is a structural diagram of the equipment fault prediction and management device based on a large model in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device in the embodiments of this application.

[0020] Figure label: Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver storage unit 9144, antenna 9111, speaker 9131, microphone 9132. Detailed Implementation

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

[0022] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0023] In view of the problems existing in the prior art, this application provides a method and device for equipment fault prediction and management based on a large model. By innovatively designing a multimodal feature fusion model, and through data integration and feature learning, accurate status assessment is achieved. A fault diagnosis system is constructed, and a reliable root cause analysis mechanism is established by combining semantic analysis and case retrieval. Maintenance guidance is introduced, and the feasibility of maintenance plans is ensured through experience accumulation and priority evaluation. This method effectively solves the shortcomings of traditional technologies in feature extraction, fault diagnosis, and maintenance guidance, providing technical support for equipment management.

[0024] To effectively address the shortcomings of traditional technologies in feature extraction, fault diagnosis, and maintenance guidance, and to provide technical support for equipment management, this application provides an embodiment of an equipment fault prediction and management method based on a large model. See [link to embodiment]. Figure 1 The equipment failure prediction and management method based on a large model specifically includes the following: Step S101: Collect equipment operation data, sound vibration data, and image data; perform unified sampling and missing value imputation on the operation data; convert the sound vibration data into spectral features; extract visual features from the image data; construct a multimodal feature extraction network; input the multimodal data into the feature extraction network to generate a fused feature vector; perform word segmentation and entity recognition on the maintenance records to extract fault types and maintenance steps; construct an equipment component association network; calculate the contrast loss between feature vectors to obtain the equipment state vector; train the equipment anomaly detection model; and use the digital twin method to generate fault samples to expand the training set. Optionally, this embodiment focuses on a speed reducer and bearing assembly for a continuous rolling line, centered around S101.

[0025] First, the data acquisition end captures speed, torque, current, oil temperature, shell temperature, and lubrication pressure from the PLC. Since the sampling frequencies vary, this embodiment maps them uniformly to a dual-track time grid of 1 Hz and 10 Hz: slow variables are entered into the 1 Hz sequence, and fast power-related variables are entered into the 10 Hz sequence. Cubic splines are used for interpolation within the gap; if the gap exceeds a set value, a mask is applied to retain the "unknown" semantics, avoiding false smoothing. The acoustic / vibration channel uses an accelerometer and microphone array to acquire the raw time-domain signal. DC removal, pre-emphasis, and Hanning window framing are performed, followed by a short-time Fourier transform to obtain the time spectrum and Mel-energy map. Considering that equipment speed drift can drag the modulation sideband, this embodiment performs a speed-normalized order transformation on the spectrum, mapping the spectral axis to the order domain, preventing subsequent features from being diluted by small fluctuations in operating conditions. On the image side, an industrial camera acquires images of the transmission end and oil window. Distortion correction, brightness equalization, and region masking are performed, retaining only the area of ​​interest to avoid interference from background pipelines.

[0026] This embodiment constructs a multimodal feature extraction network, structurally divided into three branches: the temporal coding layer employs a hybrid of gated temporal convolution and lightweight Transformer to capture slow thermal drift and instantaneous impacts; the spectral analysis layer uses 2D convolution to extract order spectral texture, supplemented by channel attention to highlight impact resonance and sideband pairs; the visual layer uses shallow convolution to extract edge cracks, oil surface fluctuations, and leakage traces, with spatial pyramid pooling added at the back end to accommodate changes in viewing distance. The outputs of the three branches enter the cross-modal attention fusion module, which adaptively calculates weights based on mutual information estimation, synthesizing the key signals of flow, sound, and image into a fused feature vector z. The weight learning follows physical intuition: when load pulsation is prominent, the weight of the temporal branch increases; when periodic sidebands appear, the proportion of the spectral branch increases; when oil stains expand in the image, the weight of the visual branch temporarily increases. The model is not "guessed," but rather these weights change with statistical indicators through joint training.

[0027] This embodiment reads maintenance records and inspection reports. First, it uses a domain dictionary and rule-based word segmentation to extract component names, fault phenomena, and maintenance actions. Then, a named entity recognition model identifies "part-symptom-action" triples, and sequence labeling is used to organize the step order and necessary conditions, such as "shutdown → depressurization → cover removal → seal replacement → reassembly → lubrication." Based on this, a network of equipment component associations is constructed. Nodes are components, and edges are physical connections or functional couplings. Edge attributes include transmission paths and lubrication sharing relationships, used to constrain subsequent feature alignment: vibration anomalies at bearing nodes are more likely to propagate along the shaft edge to gear nodes, and rarely jump directly to the motor housing.

[0028] In this embodiment, the fused feature vector z is input into the contrastive learning network to construct positive and negative sample pairs. Positive samples are synchronous slices of multimodal events or different time slices of similar minor faults under the same operating condition, while negative samples come from events in different operating conditions or different locations. Temperature-adaptive contrastive loss is used to bring samples that are "same cause but different appearances" closer together and to separate samples that are "seemingly similar but different locations," resulting in the equipment state vector h. The physical meaning here is clear: h should encode a separate representation of "speed, load, and temperature background" and "fault mode," facilitating the subsequent model's sensitivity to faults under different operating conditions. To mitigate scene bias, this embodiment introduces a constraint regularization of the component association network within the contrast space. If the component nodes corresponding to two samples are close in distance on the graph, they are penalized for excessive separation.

[0029] This embodiment uses an h-trained anomaly detection model, employing a combination of an autoregressive reconstructor and density estimator. The former reconstructs multimodal features, while the latter estimates the probability density of state vectors on a healthy manifold. The reconstruction residuals reflect "unintelligible" new patterns, and the density values ​​reflect "unseen" rare states; both are calibrated to become anomaly scores. The model is reasonable because the mechanical system's statistical regularity is stable during the healthy period, and its energy distribution and geometric texture follow normal patterns; these two are the first to be disrupted when disturbances occur. To cover more aspects of the training set, this embodiment introduces digital twins to generate fault samples: interpretable physical twins are superimposed on the component association network, with parameters derived from manufacturer manuals and field calibration. This drives a simulation engine to simulate scenarios such as bearing pitting, tooth surface uneven loading, and insufficient lubrication, outputting torque waveforms, acceleration signals, and temperature rise curves. The rendering pipeline then synthesizes oil stains and hot spots from the images. The simulated samples are not directly copied into the model but rather the background is randomized through domain randomization to avoid overfitting the simulated texture.

[0030] This embodiment further illustrates the alignment logic between simulation and actual measurements. Early pitting corrosion in bearings manifests in the order spectrum as an increase in octave sidelobes and envelope peaks; the energy stacking in the simulation output should be consistent with the measured order position. Insufficient lubrication results in ripples and a darker color in the oil window of the image, and the oil temperature slowly rises over time. In this embodiment, these indicators that conform to natural laws are used as weak supervisory signals to guide contrastive learning to focus on the stable mapping between "cause and symptom" rather than being led by environmental noise. During training, a long window during the healthy period is used as the primary method, supplemented by a short window during the fault period, to prevent threshold drift caused by an abnormal proportion of faults in the dataset.

[0031] In its engineering implementation, this embodiment also considers prototype differences and sensor anomalies. To address the noise floor differences between different batches of equipment, this embodiment performs channel layer normalization at the input end for each device. For sensors that have experienced packet loss or drift, this embodiment monitors mutual information and energy consistency. If a contradiction occurs—such as a sharp drop in sound intensity while vibration remains high—the channel is marked as low confidence and its weight is reduced in the fusion module to avoid false triggering. For nighttime segments with many missing values, interpolation is only used for context completion by the timing encoder; the density branch for anomaly detection still reads the mask.

[0032] This embodiment provides a field scenario to connect the above key points. During the night shift, the load decreased slightly, and the speed fluctuation was minimal. However, a peak consistent with the inner ring defect appeared on the vibration spectrum. No leakage was observed in the oil window image, but the oil temperature rose slightly. The fusion weights automatically reduced the visual branch and increased the spectral branch. The density of the state vector in the healthy manifold decreased, the reconstructed residual increased, and the anomaly score exceeded the adaptive threshold. A review of the maintenance records revealed that a previous case of "inner ring pitting—bearing replacement—normal retest" for the same model in this load range was identified as a nearest neighbor. The component association network also limited the propagation path to within the shaft system. Although the root cause diagnosis step has not yet been initiated, the state representation in stage S101 has provided a reliable signal of "what's wrong," and the relevant evidence has been packaged into training pairs for subsequent online updates.

[0033] To facilitate understanding, the formula for synthesizing outlier scores is illustrated here: S = α·R + (1−α)·(1−P), Where S is the final anomaly score, R is the normalized value of the reconstructed residual, P is the probability of the state vector on the healthy density, and α is the weighting coefficient. R characterizes the "difficulty of model reproduction", P characterizes the "healthy prior confidence", and α is calibrated by the validation set during the healthy period and undergoes short-term adaptive regression when the ambient temperature changes drastically. The physical meaning and data sources of these quantities have been explained above.

[0034] Ultimately, this embodiment streamlines the information flow along the acquisition-fusion-comparison-twin-training chain, solving the long-standing problems of difficulty in uniformly characterizing multi-source heterogeneous data, the obscuring of minor early faults by noise, and the unbalanced nature of the model due to sample scarcity, thus achieving the goal of being more sensitive to early anomalies and more robust to operating conditions.

[0035] Step S102: Input the device state vector into the large language model encoder to generate a semantic representation, construct a retrieval enhancement generation model, read historical maintenance cases to establish a vector retrieval library, calculate the temporal dependency of fault events, construct a component-level fault propagation network, calculate the root cause probability of the fault based on the fault propagation network, calculate the similarity score between the semantic representation and the retrieval library, extract maintenance records of similar cases, train the fault diagnosis model, and determine the root cause localization result based on the confidence score; Optionally, in this embodiment, the equipment state vector h obtained in S101 is used to perform S102 for fault analysis of the rolling mill reducer and bearing assembly.

[0036] In this embodiment, h is first input into the front projection layer of the large language model encoder. Using a set of prompt templates aligned with the semantics of the scene, the key signs on the sensing side are mapped into short text fragments, such as "sideband enhancement near order 2.9, slow oil temperature rise, shell vibration Y-axis head-up". These texts are fed into the encoder along with h, and the semantic representation e is output. The reason for this processing is that engineers' knowledge has long been recorded in language. If numerical embedding is directly used to align with the text library, it would be awkward. Adding a semantic bridge makes it easier for the retrieval and generation ends to "understand" the state they are referring to.

[0037] Based on this, this embodiment constructs a retrieval-enhanced generation (RAG) model, which is divided into two paths: one path uses e to perform nearest neighbor recall in the vector retrieval library, and the other path concatenates the recalled fragments with e and enters the generator decoder to form an interpretable diagnostic draft.

[0038] The retrieval database is built from years of maintenance cases, weekly inspection reports, and manufacturer manuals. Before being added to the database, it is segmented, retaining structured fields of "phenomenon-location-treatment-verification," and synonym mapping is performed for industry terms to avoid splitting similarity between the terms "bearing outer ring / outer ring." To prevent long texts from diluting the key points, in this embodiment, the information density score is calculated for each segment. Descriptions with low density only retain keywords, which are then vectorized and placed into the FAISS index to ensure that online retrieval is completed within milliseconds.

[0039] Optionally, this embodiment reads the historical fault event stream of the equipment, archives the occurrence time and impact range according to component nodes, and estimates the temporal dependencies between events: if the bearing inner ring anomaly often precedes the tooth surface off-center loading, and the time interval is concentrated in several operating condition windows, in this embodiment, the condition distribution is recorded on the edges of component pairs. This constructs a component-level fault propagation network G, where nodes are components, edges are propagation channels, and parameters include propagation probability and time delay. To ensure G conforms to physical laws, propagation is only established along "reachable edges" of force flow, heat flow, lubrication loops, or control coupling, and unfounded crossings are not allowed. The observed symptoms at the current moment are projected onto G, and combined with the time delay distribution of the edges, a probabilistic inference is run to obtain the root cause probability π. Each node has a score that is interpreted as "more like the source or implicated."

[0040] This embodiment integrates semantic retrieval with propagation reasoning. First, cosine similarity scores are calculated using e in the retrieval database. Then, the "root cause-treatment-verification" segments of several high-scoring cases are taken as evidence E. Finally, π is used as the prior strength and fed into the diagnostic model.

[0041] The diagnostic model employs an evidence-based deep learning framework. Its input features are composed of "current symptom encoding e, propagation prior π, and evidence fragment embedding vE," and its output is the probability distribution and uncertainty parameters of component-level root causes. The evidence network was chosen because field data is often incomplete, requiring the model to clearly distinguish between "uncertainty" and "low probability," which aligns with the conservative nature of engineering decision-making.

[0042] This embodiment employs two types of supervision to collaboratively advance the training phase. The first type comes from historical samples that have already closed the loop, where the root cause has been repaired and verified, serving as a hard label. The second type consists of weak labels, derived from digital twins and expert review, informing the model that certain symptoms are strongly but not uniquely associated with a particular component. During training, this embodiment uses the edge weights of the propagation network as regularization terms. If the root cause path provided by the model deviates significantly from G, it will be penalized; conversely, inferences along high-confidence edges will be rewarded.

[0043] This embodiment provides a confidence score during the inference output stage. In this embodiment, the Dirichlet parameters of the diagnostic output are mapped to confidence intervals, and the total score C is constructed by combining the coverage and consistency of the retrieved evidence. If multiple candidate root cause scores are close, this embodiment generates an ordered checklist, such as first measuring abrasive particles in the oil sample, then collecting vibration data near the tooth surface, and then performing a low-speed frequency sweep to distinguish them at the lowest possible cost.

[0044] For example, a night shift data point triggers an anomaly. The comparison database retrieves a case of "early inner ring pitting—envelope energy increase—maintain load observation," but G indicates a "insufficient lubrication" path with a significant π because the oil temperature background is rising. The diagnostic model gives similar probabilities for both, with C slightly favoring pitting. In this example, oil samples are prepared first according to the review checklist. The abrasive level is normal, and after recalculating C, the inner ring pitting score increases, the root cause is identified, and the subsequent maintenance instructions in S103 can be generated accordingly. This example illustrates that RAG emphasizes experience, while the propagation diagram grasps the mechanism; combining both provides a more accurate assessment of workshop conditions.

[0045] This embodiment also considers the migration of equipment families. Gearboxes of the same series but different rated power have similar skeletons in their G-graphs, but different edge weights. In this embodiment, a hierarchical prior is used to freeze the "skeleton," and the edge weights are quickly corrected with a small sample size after the new equipment comes online. The search library is also tagged and filtered by machine type, oil type, and load spectrum to avoid bringing in irrelevant cases that interfere with decision-making. Ultimately, the model's input is h and textual symptoms, and the output is the root cause probability and confidence level, connecting the state representation of S101 with subsequent maintenance scheduling. This solves the pain point of "numerical symptoms and linguistic experience being disconnected, and phenomena and mechanisms being out of sync." The technical effect is reflected in more stable and interpretable root cause judgments, clear verification paths, and convergence of false alarms and erroneous repairs.

[0046] Step S103: Monitor the real-time status data of the equipment, input the status data into the anomaly detection model to calculate the anomaly score, input the anomaly data fragment into the fault diagnosis model to generate the fault cause, extract expert maintenance experience to build a knowledge base, calculate the priority score of maintenance operations, establish a dependency constraint network of maintenance steps, generate an executable maintenance guidance plan, collect maintenance feedback data to update the vector retrieval library and knowledge base, write the maintenance feedback data into the training sample set, and perform online updates to the anomaly detection model and the fault diagnosis model.

[0047] Optionally, this embodiment deploys a field-connected system for the rolling mill reducer and bearing assembly, constructing a closed loop around S103 from "monitoring data - judgment - order placement - review".

[0048] On the data front, the system monitors speed, torque, current, casing temperature, oil temperature, vibration triaxial acceleration, and envelope signal in a second-level rolling window. The image captures the oil window and end cap area every 30 seconds. Within each window, time synchronization and mask alignment are performed first, and gaps are labeled instead of being hard-interpolated. Then, a set of physically relevant health indicators are calculated: specific energy per unit, temperature rise slope, vibration envelope energy, order spectrum sideband ratio, and image oil surface texture change rate. The indicator vector is interfaced with the anomaly detection model trained in stage S101. The model outputs the reconstruction residual and health density, which are then combined to form an anomaly score. When the score exceeds an adaptive threshold, the window is labeled as an "abnormal fragment" and merged with the adjacent context to ensure that subsequent diagnosis sees a sufficiently long cause-and-effect relationship.

[0049] In this embodiment, the abnormal segment is input into the fault diagnosis model trained by S102. The input is packaged into three parts: First, the embedding of multimodal symptoms; Second, the root cause prior of the component-level fault propagation network in the current time sequence; Third, retrieve similar case evidence vectors recalled by the enhanced retrieval model.

[0050] The model returns the root cause probability, path explanation, and uncertainty for each component. To avoid "forced judgment," the system reads the uncertainty; if the distribution is high-entropy, it switches to a "pending verification" state and triggers a low-cost retest suggestion (e.g., adding low-speed frequency sweep and oil sample spectrum) for secondary inference. The underlying relationships here are quite simple: vibration order is proportional to mechanical failure, oil temperature and lubrication state have a delayed coupling, and changes in oil level in the image are accompanied by leakage. The reason the model can make relatively stable judgments on probabilities is that the physical relationships between these indicators have been solidified during training through contrastive learning and propagation graph regularization.

[0051] This embodiment extracts expert repair experience from a knowledge accumulation perspective to build a structured knowledge base. Knowledge items are described by a six-tuple: "Component—Symptom—Diagnostic Points—Treatment Steps—Risk Points—Verification Method," sourced from historical work orders, oral notes, and manufacturer manuals. To avoid only collecting "conclusions," the system intentionally retains counterexamples of failed handling and records triggering conditions for subsequent priority evaluation. Before being stored in the database, knowledge items undergo semantic deduplication and consistency checks, ensuring that chains like "replacing bearing—adjusting preload—retesting envelope decrease" can be directly referenced by the model in diagnostic interpretations.

[0052] This embodiment calculates the priority score of maintenance operations based on diagnostic output and a knowledge base. Three scoring factors are defined: risk (safety, probability of propagation, downtime cost), feasibility (availability of spare parts, available workstations, required skills), and diagnostic gain (whether the operation can significantly reduce uncertainty). A hierarchical analysis is performed on these three factors to determine their weights, and on-site constraints (such as the current shift's personnel qualification matrix) are overlaid. To ensure transparency in the evaluation, the system provides a breakdown of the score composition, allowing engineers to see the logical order of "why oil samples are taken first, then bearings are disassembled and inspected."

[0053] Before generating the maintenance guidance plan, this embodiment first establishes a dependency constraint network for the maintenance steps. Steps in the knowledge base are extracted as nodes, and preconditions and parallel mutual exclusion relationships are used to construct a directed acyclic graph. Node attributes are written with tools, torque standards, cleaning requirements, and quality inspection points. After obtaining an executable sequence through topological sorting, the system inserts buffers and verification points for critical steps based on equipment status and production plans. For example, after "replacing the bearing," three inspection nodes are forcibly added: "10-minute no-load break-in—retesting the order spectrum—checking oil temperature drop," to prevent hasty resumption of production from leaving hidden defects. The plan is finally filled with a natural language template, including a tool list, torque table, precautions, and rejection criteria, facilitating implementation by work teams.

[0054] This embodiment incorporates feedback into the maintenance execution process in a closed loop. Mobile work orders record start / end times, whether execution was sequential, key measurement values, and anomaly notes. Upon completion, a "maintenance effectiveness score" is generated, based on changes in retest indicators, recurrence rates, and operational deviations. Feedback updates the vector retrieval library: effective handling segments are given higher recall weights, while misleading steps are downgraded and given contextual constraints. It also updates knowledge base entries, correcting critical conditions and risk points in documents. All feedback is written into the training sample set, sliced ​​by time window, and used for two types of online updates: the anomaly detection model performs small-step incremental learning, biased towards the health distribution of recent operating conditions; the fault diagnosis model uses online distillation, pushing the latest "diagnosis-handling-verification" chain into the student model, making it more sensitive to new patterns without forgetting existing capabilities.

[0055] For example, during a night shift, the abnormal score continued to exceed the limit under conditions of slow oil temperature rise and slight vibration. The diagnostic tool gave similar probabilities for "insufficient lubrication" and "early pitting corrosion of the inner ring," prioritizing low-cost oil replenishment and oil sample analysis. The network-generated solution first arranged for cleaning the oil passage and checking the nozzles, and then set two retest nodes. Feedback showed that after cleaning the nozzle blockage, the envelope energy decreased, the uncertainty converged, and the search library increased the weight of cases with "nozzle blockage - temperature rise first." Two weeks later, similar symptoms reappeared, and the system converged to the same treatment path in a shorter time. The engineer's explanation of the cause aligned with the model output, reducing communication costs.

[0056] This embodiment controls the relationship between anomaly score, diagnostic probability, and priority within an interpretable range, using a simple composite quantity to drive alarm level escalation: L = β1·S + β2·(1−U) + β3·R, In the formula, L represents the alarm level score, S represents the anomaly score, U represents the diagnostic uncertainty, R represents the maintenance risk factor, and β1, β2, and β3 are weighting coefficients calibrated through historical playback. S reflects "whether it is abnormal," U reflects "the degree of uncertainty," and R reflects "the cost of delay." The combined trend of these three factors is consistent with the urgency of on-site handling, avoiding the incongruity of "high-level alarms despite high uncertainty." In summary, this embodiment links monitoring, diagnosis, order generation, and learning into a closed loop, solving a series of problems such as "disorganized actions after anomaly detection, lack of constraints in maintenance execution, and difficulty in accumulating and reusing experience." The technical effects are reflected in more restrained alarms, more methodical handling, and continuous growth of model capabilities with on-site updates.

[0057] As described above, the equipment fault prediction and management method based on a large model provided in this application can achieve accurate status assessment through innovative design of a multimodal feature fusion model, data integration, and feature learning. It constructs a fault diagnosis system, combining semantic analysis and case retrieval to establish a reliable root cause analysis mechanism. Furthermore, it introduces maintenance guidance, ensuring the feasibility of maintenance plans through experience accumulation and priority evaluation. This method effectively addresses the shortcomings of traditional technologies in feature extraction, fault diagnosis, and maintenance guidance, providing technical support for equipment management.

[0058] In one embodiment of the equipment failure prediction and management method based on a large model in this application, it may further include the following: Step S201: Collect operating parameter data from equipment sensors, collect vibration waveform data from acoustic sensors, and collect equipment image data from industrial cameras. Resample the operating parameter data at fixed time intervals, fill missing values ​​using cubic spline interpolation, perform short-time Fourier transform on the vibration waveform data to obtain a time-frequency spectrum, and perform convolution operation on the image data to extract edge and texture features. Step S202: Construct a multimodal feature extraction network comprising a temporal coding layer, a spectrum analysis layer, and a visual feature layer. Input the running parameter data into the temporal coding layer, input the temporal spectrum into the spectrum analysis layer, and input the edge features and texture features into the visual feature layer. Calculate the weight coefficients of the three feature layers using an attention mechanism. Based on the weight coefficients, weight the three feature layers to generate a fused feature vector.

[0059] Optionally, this embodiment is deployed in the main drive area of ​​the cold rolling line of the steel plant, taking the reducer-bearing-coupling as the object, and carrying out multimodal signal acquisition and feature construction around S201 and S202.

[0060] On the data acquisition side, clocks are synchronized on the PLC and edge gateway, and raw sequences are input from sensors such as speed, current, torque, oil temperature, housing temperature, lubrication pressure difference, and axial displacement. The acoustic and vibration channels use triaxial accelerometers and condenser microphones, sampling above 20 kHz to cover the characteristic bands of gear meshing and bearing failure. The industrial camera is fixed at the transmission end and oil window, with shutter and exposure locked, and a polarizer is used to reduce reflection. Due to the different time bases of each channel, this embodiment sets a unified time axis: operating parameters are resampled by a fixed Δt (e.g., 100 ms), cubic spline interpolation is used for short-segment missing measurements, and long empty windows retain masks to prevent false smoothing from masking sudden changes. Vibration waveforms are framed with overlapping windows, and short-time Fourier transforms are performed after adding Hanning windows to obtain the time spectrum. Frame shift and window length are determined by back-calculation from the target order resolution. The image is first subjected to distortion correction, white balance, and ROI cropping, and then fed into shallow convolution to extract edges and textures. Edges are used for crack and leakage contours, and textures are used for judging lubrication film and stain diffusion.

[0061] This embodiment employs a rudimentary physical judgment regarding the trade-offs between resampling and interpolation. The peaks in torque and current are correlated with mechanical shocks; if the interpolation span exceeds twice the control cycle, interpolation is discontinued, preferring to retain NaN values ​​and transmit the "uncertain" state as a mask in subsequent networks. Gradual variables such as oil temperature can be supplemented using splines to maintain temporal continuity. The selection of STFT parameters follows a trade-off between energy conservation and resolution: gear meshing frequencies and their sidebands require high frequency resolution, while bearing envelopes require temporal resolution. This embodiment utilizes multi-scale spectrum stacking, with the main scale covering the meshing region and the secondary scale covering the envelope band; these are then stitched together in subsequent networks using channel dimensions. The image-side convolution kernels do not prioritize depth; the focus is on extracting the edge gradient direction histogram and micro-texture statistics to avoid being biased by strong textures in the background pipelines and labels.

[0062] This embodiment enters S202 to construct a multimodal feature extraction network. The three branches of the network are respectively connected to three types of data: the temporal coding layer receives the resampled sequence of operating parameters, and uses gated temporal convolution (TCN) superimposed with a lightweight Transformer. The TCN is responsible for local mode and delay compensation, and the Transformer captures thermal drift and load cycles across minutes; the spectrum analysis layer takes in the time spectrum and multi-scale envelope spectrum, and uses 2D convolution + residual block to extract spectral texture and order bandwidth changes, and channel attention emphasizes the coupling of sideband pairs and formants; the visual feature layer takes in the edge map and texture map, and uses front-end small kernel convolution to preserve details, and back-end dilated convolution to expand the field of view to cover the leakage spread area. The outputs of the three branches converge on the aligned timestamps. To avoid cross-modal delay, vibration and image features are finely aligned according to the causal relationship identified by the system. For example, if the torque impact leads the shell temperature rise, this embodiment applies hysteresis compensation to the temperature features before fusion.

[0063] This embodiment employs an attention mechanism to calculate the weight coefficients of the three feature layers in each time slice. The weights are not constant but are adaptively provided by a small gating network based on the operating conditions. The input to the weight estimation includes confidence metrics for each branch: residual stationarity and autocorrelation decay on the temporal side, signal-to-noise ratio and modulation index on the spectral side, and edge continuity and texture stability on the visual side. Thus, when the confidence of low-light images decreases, the visual weights automatically decrease; when order sidebands appear and the SNR is high, the spectral weights increase; during periods of rapid load changes, the temporal weights dominate. The weight changes conform to natural laws and are consistent with the "health-fault" separability objective extracted by contrastive learning in S101, ensuring that the fused vector is neither biased by single-modal noise nor loses key features.

[0064] This embodiment generates a fused feature vector z by weighted combination of three features. The vector composition preserves interpretable partitions: the statistical sub-vector represents the steady-state background (rotation speed, temperature, load), the dynamic sub-vector captures transient shocks and spectral sidebands, and the visual sub-vector characterizes leakage and surface anomalies. To suppress spurious correlations, this embodiment adds a cross-modal mutual information regularization term to the fusion header, encouraging alignment of historically frequently co-occurring signs (e.g., "torque pulse ↑—envelope energy ↑") while weakening irrelevant coincidences (e.g., "shift lighting change—current fluctuation"). The training process employs multi-task objectives: predicting health reconstruction at the next time step, estimating key KPIs (specific energy per unit, vibration RMS), and weakly supervised classification of labeled fault segments, allowing z to share representations across the three tasks and reducing dependence on a single label.

[0065] During on-site verification in this embodiment, two typical scenarios were observed. The first scenario is early bearing inner ring defects: narrow sidebands appear in the spectral layer at specific orders, the impact index in the temporal layer rises slightly, and there is no obvious leakage in the visual layer. Attention allocation is biased towards the spectral and temporal layers, the dynamic sub-vector of the fusion vector is activated, and subsequent anomaly detection gives this segment a high anomaly score. The second scenario is insufficient lubrication: the temporal temperature rises slowly, the high-frequency noise band in the spectral spectrum widens, the visual texture shows oil surface ripples and dark spots, the weights of the three branches are more balanced, and the visual and steady-state sub-vectors in z rise synchronously. The distinction between the two modes is not based on single-point thresholds, but on the geometric separation of multimodal evidence in the fusion space.

[0066] This embodiment provides a calculation diagram of weight normalization at the formula level for ease of understanding: w_i = f(q_i) / Σ_j f(q_j), Where w_i is the weight coefficient of the i-th branch, q_i is the comprehensive score of the confidence index vector of this branch after linear transformation, and f(·) is the soft threshold function with temperature factor; the physical meaning of w_i is "the reliable proportion of the i-th modal information under the current working condition". This set of weights is used to weight the three-branch features to obtain z, which will serve as the input base for the subsequent large language model encoding and retrieval enhancement in S102.

[0067] Finally, two implementation details are added to this embodiment. First, the trigger delay of the camera and accelerometer is in the millisecond range, but it can still affect order alignment under high-speed conditions. In this embodiment, phase correction is performed at the edge using a tooth frequency reference signal to ensure that the temporal semantics of the spectral peaks and image hotspots are consistent. Second, for artifacts introduced by non-equipment factors such as nighttime light fluctuations and sensor loosening, the confidence index will drop rapidly, and the weights will automatically shrink to avoid misleading fusion. By connecting S201 and S202 in this way, the system establishes a stable and interpretable fusion representation in the noisy background of the original signal, providing a reliable entry point for subsequent root cause diagnosis and maintenance decisions.

[0068] In one embodiment of the equipment failure prediction and management method based on a large model in this application, it may further include the following: Step S301: Read the historical maintenance record text, extract professional terms and keywords using dictionary segmentation, label the fault location and maintenance action based on the named entity recognition model, construct the fault location as a node, construct the physical connection relationship between components as an edge, generate the equipment component association network, perform sequence labeling on the maintenance action to construct the maintenance step table, and write the fault location and maintenance step into the labeled dataset. Step S302: Input the fused feature vector into the contrastive learning network, calculate the contrast loss between positive and negative sample pairs, train the feature encoder to generate the device state vector, construct a digital twin model based on the device component association network, input the digital twin model into the physical simulation engine, simulate fault scenarios under different working conditions to generate simulation data, and combine the simulation data with real data to train the device anomaly detection model.

[0069] Optionally, this embodiment is based on the equipment family of cold rolling mill reducer-bearing-coupling, focusing on bridging the data from S301 and S302 to the model. In this embodiment, maintenance records and inspection logs from the past five years are read first. The text format is mixed, including form fields and free descriptions. This embodiment first performs cleaning and template, unifying the mapping of timestamps, machine positions, and component names / aliases, and then proceeds to dictionary segmentation: the dictionary is expanded from the manufacturer's manual, spare parts list, and on-site spoken language, covering synonyms such as "outer ring / outer ring, tooth surface pitting / pockmarks, preload, nozzle, oil return". After segmentation, mutual information filtering is performed on candidate terms to remove irrelevant and frequently used terms. Subsequently, a named entity recognition model is loaded, labeled with five categories of tags: "location, symptom, action, tool, parameter", ensuring that key chains such as "inner ring peeling - bearing replacement - preload 18Nm - retest envelope decrease" are fully exposed. Based on physical entities, a network of interconnected equipment components is constructed: components (bearings, gears, couplings, lubrication stations, nozzles, oil lines, etc.) are treated as nodes, and edges are established based on the coupling of force flow, heat flow, lubrication, and control. Edge attributes are written with connection type, transmission direction, and experience propagation delay. Sequence annotation is performed on maintenance actions, extracting pre- and post-requirement relationships and parallel / mutually exclusive constraints to form a step table, such as "shutdown → depressurization → cover removal → seal replacement → reassembly → no-load break-in → vibration retest," and the necessary tools and acceptance criteria for each step are bound together. The parts and steps are uniformly written into the annotation dataset, enabling both supervised learning and providing prior knowledge for simulation and diagnosis.

[0070] This embodiment then processes the fused feature vector z from S201 / S202 and feeds it into the contrastive learning network. Positive sample pairs are defined as combinations of "same root cause, different loads, or different modal segments," such as time-frequency slices and local images of early pitting corrosion of bearings under light and medium loads. Negative sample pairs select segments of "similar loads but different locations or positions" to avoid mistaking similar operating conditions for the same cause. The contrastive loss uses temperature-scaled InfoNCE, constructing multiple contrasts within a small batch. The network trains the feature encoder f(·) to output the device state vector h, making "same cause, different appearance" close together and "different causes easily confused" separated. To make the geometry closely resemble physics, this embodiment introduces graph regularization for the component association network: if the components corresponding to two samples are close on the graph path and have a high propagation probability, then the distance between h is allowed to be close but not overlapping; if they are not reachable on the graph, a separation penalty is added to avoid learning spurious similarities across components. The logic of this step is straightforward—to make the proximity relationship in the data space conform to the reachability of the mechanical system.

[0071] This embodiment builds a digital twin model based on h to expand the fault samples and fill in the gaps. The twin skeleton unfolds along the component association network, and physical sub-models are instantiated for bearings, gears, and couplings respectively: bearings use a defect-carrying model with parameters including the number of rolling elements, pitch circle diameter, and defect location and size; gears use a meshing stiffness periodic disturbance model; and lubrication uses a simplified equation coupled with rheology and heat. These sub-models are connected to the physical simulation engine, inputting the operating condition boundaries (speed, torque, ambient temperature, oil viscosity), and outputting torque pulsation, vibration order spectrum, and temperature rise curve. Controllable leakage and oil surface ripple textures are superimposed on the image pipeline. To bridge the gap with pure simulation, this embodiment uses two types of calibration: one is phase / frequency alignment based on the measured order position, and the other is amplitude scaling based on the energy statistical distribution, so that the simulation samples fall within the same order of magnitude as the field in key indicators without requiring point-to-point consistency. The fault segments generated in this way are tagged according to the "location-step-acceptance" metadata marked in S301 to ensure that semantic noise is not introduced into subsequent training.

[0072] This embodiment trains the device anomaly detection model using a mixture of simulated and real data. The model employs a dual-head structure of reconstructor + density estimation, with inputs of z or h and outputs of reconstruction error and health probability. The hybrid strategy is not a simple patchwork; the health segment primarily uses real data, while the fault segment uses a combination of "small amounts of real data + diverse simulated data" to cover multiple early signs. During training, simulated samples are randomized (noise, illumination, micro-frequency drift) to prevent the network from memorizing simulated textures. To make anomaly detection more aware of "where to be sensitive," this embodiment embeds a component association network into the loss function: on reachable paths, anomaly scores are allowed to have "propagation trails," while anomalies on unreachable paths are suppressed, reducing the spread of false alarms.

[0073] This embodiment illustrates a specific segment to connect the links. During the night shift, a unit experienced a slight increase in specific energy, a rise in the steady-state subvector in z, and narrow sidebands in the inner ring feature order of the spectral layer, while no leakage was observed in the image layer. Comparing the encoder output h with the historical "early inner ring pitting" cluster, the anomaly detection gave a moderately high score. In this embodiment, reviewing the S301 step table, the correlation network showed a low probability of bearing → gear propagation under this load. Therefore, the simulation engine selected "inner ring defect only" scenario enhancement training. After online small-step updates, the model's sensitivity to similar segments became more focused on the bearing node, preventing the anomaly from extending to the tooth surface and reducing unnecessary disassembly and inspection.

[0074] This embodiment adds two more points regarding engineering details.

[0075] First, the reliability of text-side NLP annotations affects network edge weights. In this embodiment, confidence scores from manual sampling are introduced, and low-confidence entries contribute half to edge weights. Second, negative sample sampling in contrastive learning is stratified by machine location and operating condition to prevent an excessively high proportion of easy negatives from leading to loose discrimination boundaries.

[0076] Ultimately, S301 embeds linguistic knowledge into a structured component-step-constraint diagram, and S302 pulls perceptual features onto the coordinate system of the physical world. The technical effects are reflected in more stable state representation, more focused anomaly identification, and more reliable data for subsequent diagnosis and maintenance.

[0077] In one embodiment of the equipment failure prediction and management method based on a large model in this application, it may further include the following: Step S401: Input the device status vector into the large language model encoder and map it into the natural language space. Construct fault description text based on prompt word templates. Input the fault description text into the large language model to generate semantic vectors. Read historical maintenance case records. Perform text segmentation and key information extraction on the maintenance cases. Construct a maintenance knowledge base containing fault phenomena, root cause analysis, and handling methods. Encode the maintenance knowledge base into vector form to establish a retrieval database. Step S402: Read the equipment fault event record, extract the fault occurrence time and affected components, calculate the time interval and propagation probability between component faults, write the propagation probability into the adjacency matrix to construct a fault propagation graph, calculate the in-degree and out-degree of component nodes based on the fault propagation graph, identify the fault source node and affected nodes, construct a component-level fault propagation network, and calculate the probability distribution of fault paths.

[0078] Optionally, this embodiment inherits the multimodal perception and state representation of the cold rolling line reducer-bearing-coupling in S201-S302, translates the "numerical state" into "language knowledge" around S401 and S402, and sinks the event sequence into a computable propagation network.

[0079] The starting point is the device state vector h output by S302. In this embodiment, h is not directly fed to the language model. Instead, a semantic alignment layer is pre-applied to a Large Language Model Encoder (LLM-Encoder) to map the interpretable sub-dimensions of h (steady-state background, dynamic impact, visual leakage) into several semantic slots. Combining operating condition metadata and component labels, structured short sentences are constructed using prompt word templates, such as "Bearing - inner side, envelope energy rises, sidebands appear near order 2.9, oil temperature rises slowly by 0.x, no leakage is observed in the image." These sentences, along with h, are input into the LLM-Encoder, which outputs a semantic vector e located in the natural language embedding space.

[0080] This embodiment reads historical maintenance case records in batches, including work orders, weekly inspection reports, and manufacturer technical bulletins. First, it segments the data into paragraphs, breaking down lengthy descriptions into "phenomenon, test, diagnosis, treatment, and verification." Then, a key information extraction model is used to label components, symptom terms, treatment methods, and verification indicators, such as "inner ring pitting—envelope energy improvement—bearing replacement—retesting and order decline." To reduce the disconnect between colloquial language and technical terms, a thesaurus (outer ring / outer circle, back clearance / gap) is pre-created, and ambiguous terms (such as "jumping") are categorized into displacement or vibration based on context. The extracted triples are aligned with the evidence sentences and organized into atomic entries for the maintenance knowledge base, each entry including source, applicable operating conditions, and risk points. Finally, a text encoder encodes the entries into vectors, establishing a vector retrieval database and recording metadata indexes (engine model, oil type, rated speed range) to filter search results based on conditions and avoid mistakenly recalling irrelevant cases.

[0081] In this embodiment, the fault event records of the equipment are read from anomaly detection triggers, manual confirmations, and maintenance closed-loop tags. For each record, the occurrence time, the initial component, and the affected components are extracted and archived as an event stream along a timeline. For each pair of components (u, v), the conditional frequency of v occurring within a time interval Δt after u occurs is statistically analyzed to obtain the empirical propagation probability p(v|u, Δt). To ensure this probability conforms to physical laws, this embodiment limits the statistics to reachable edges in the component association network (force flow, heat flow, lubrication loops, control coupling), and separate slots are created for extreme operating conditions to avoid mixing seasonal deviations such as "high incidence of anomalies during maintenance season" into the general probability.

[0082] In this embodiment, the aforementioned propagation probabilities are written into the adjacency matrix A, where A[u,v] stores the propagation strength and time delay parameters across u to v, constructing a fault propagation graph. Basic graph theory analysis is then performed on the graph to calculate the in-degree, out-degree, and weighted centrality of each component node. However, this is only a coarse screening; true source identification requires consideration of the temporal consistency of current observations.

[0083] In this embodiment, for a given observation window, the observed abnormal signs are marked on the relevant components according to time. Using the time delay distribution in A, the time consistency score of each possible path is calculated, and then multiplied by the static propagation intensity to obtain the path probability.

[0084] This embodiment merges the semantic and propagation sides into the "evidence aggregation" for fault localization. When the new state vector h is mapped to e, this embodiment performs nearest neighbor recall in the retrieval database, selecting several highly similar entries and matching them with the current aircraft model / fuel type to form an evidence set E. On the other hand, the source probability calculated from the propagation graph is used as the prior π, representing "where it is more likely to originate mechanistically." Then, the embeddings of e and E are concatenated with π and input into the subsequent diagnostic model, which provides the root cause probability and explanation.

[0085] This embodiment takes into account the inconsistent quality of historical records and sets up quality control for knowledge entry into the database. Before an entry is added, a consistency review is performed: the verification indicators of the same fault-handling combination are checked to see if they are consistent in different cases. If there is a large discrepancy, the credibility weight of the entry is lowered or "applicable conditions" are added. After being added to the database, the entries are dynamically corrected in the maintenance feedback stage of S103. Entries with poor performance are gradually downgraded, while those with good performance are upgraded.

[0086] This embodiment provides a calculation illustration on the propagation network: for the path P=u→…→v, its probability can be written as: Π_{(i,j)∈P} p(j|i)·g(Δt_{ij}), Where p(j|i) is the conditional probability of propagation from component i to component j, Δt_{ij} is the observation time difference of this propagation edge, and g(·) is the time consistency function based on the historical time delay distribution. The closer the output value is to 1, the more consistent it is with the usual time sequence. The physical meaning of these parameters is clear: p(j|i) captures whether the propagation will occur, and Δt_{ij} and g capture when the propagation will occur. The product of the two is the credible path recognized by the field engineer. This set of path probabilities is summarized at the nodes to obtain the component-level root cause distribution, and cross-validated with linguistic evidence from the e-E side.

[0087] In one embodiment of the equipment failure prediction and management method based on a large model in this application, it may further include the following: Step S501: Read the node connection relationship in the fault propagation network, calculate the conditional probability of the fault propagation path, construct a fault probability inference model based on the Bayesian network, substitute the equipment status data into the inference model to calculate the fault probability distribution of each component, perform nearest neighbor search on the semantic representation vector, calculate the cosine similarity with historical cases in the search library, and extract the maintenance record with the highest similarity score as a reference case. Step S502: Combine the fault probability distribution with reference cases to construct diagnostic input features, train the fault diagnosis model using an evidence deep learning network, calculate the confidence interval of the diagnosis results based on the Dirichlet distribution, convert the confidence interval into a confidence score, sort the fault diagnosis results according to the confidence score, and select the result with the highest score as the root cause localization output.

[0088] Optionally, this embodiment inherits the semantic representation and propagation network of S401-S402 on the cold rolling line reducer-bearing-coupling, and provides root cause localization around S501 and S502.

[0089] In this embodiment, the node connection relationships and edge attributes are first read from the fault propagation network. The edge attributes include propagation probability and time delay distribution, and the nodes are accompanied by component health priors and observable symptom lists. In order to move from the graph to computable probabilities, this embodiment parameterizes the propagation graph as a Bayesian network: each component node X_i is a binary or multi-level fault state variable, the parent node set Pa(X_i) is determined by the incoming edges of the propagation graph, and the conditional probability table P(X_i|Pa(X_i)) is derived from the propagation probability estimated by S402 and Bayesian correction is performed using maintenance closed-loop data.

[0090] This embodiment incorporates the fused state data from S201-S202 and the semantic representation generated in S102 into the inference. On the numerical side, this embodiment maps time-series indicators (envelope energy, order sideband ratio, temperature rise slope, etc.) to node observation evidence E_i. The strength of this evidence is related to the magnitude and stability of the indicator's deviation from the healthy baseline, avoiding a single spike overwhelming the evidence. On the semantic side, the confidence levels of component-related keywords in e are converted into soft evidence, serving as weights for the observation likelihood. Inference uses variable elimination or belief propagation to calculate the posterior failure probability distribution π_i for each component. The output includes not only point estimates but also a sensitivity assessment to conflicting evidence: if two parent nodes provide contradictory evidence, the posterior will carry higher uncertainty, and subsequent ranking will lower its credibility.

[0091] For the vector retrieval library constructed by S401, in this embodiment, semantic vector e is used for nearest neighbor recall, cosine similarity is calculated, and several high-scoring cases are selected. To prevent false recalls due to cross-classification, filtering is first performed by engine type, oil type, and speed range, and then the matching degree between the "verification steps" in the case and the currently available measurement points is examined. Each reference case is given in a structured block: typical symptoms, root cause, treatment, and verification. In this embodiment, the single case with the highest similarity is not taken as the "answer," but rather the Top-k cases are weighted and summed according to similarity to form an evidence set E, preserving diversity and avoiding overfitting to a single statement.

[0092] In this embodiment, π and E are fused into diagnostic input features, which are then fed into the S502 evidence-based deep learning network. The input consists of three types of vectors: first, a summary of the node posterior and path probability of π, representing the contraction of the mechanism prior under current observation; second, the embedding v_E of e and retrieved evidence, representing the directionality of linguistic experience; and third, a low-dimensional generalization z' of multimodal symptoms, capturing the current "likeness". The network uses multi-head attention to align these three vectors, prioritizing combinations that are "path-consistent and symptom-matching". If e conflicts with π, the network will not force a compromise but will output high uncertainty.

[0093] To allow for a concrete explanation of the uncertainty, this embodiment adds a Dirichlet evidence layer after the classification head. The network predicts the evidence strength of each component, which is then converted into the Dirichlet parameter α, and the probability mean and variance are obtained.

[0094] This embodiment transforms confidence intervals into ranking confidence scores. In this embodiment, the score C_i is defined as a composite of three parts: the mean diagnostic probability μ_i, the consistency of evidence ξ_i (log-likelihood from multi-head attention alignment), and the uncertainty u_i (a monotonic mapping of Dirichlet variance). The synthesis strategy prioritizes scores with high μ_i, low u_i, and high ξ_i. This can be summarized by the following formula: C_i = γ1·μ_i + γ2·ξ_i − γ3·u_i, Where γ1, γ2, and γ3 are weighting coefficients calibrated from historical playback, μ_i is the mean diagnostic probability of component i, ξ_i is the consistency score of multi-source evidence, and u_i is the uncertainty measure. The physical meanings of the three quantities are intuitive: μ_i indicates "likeliness," ξ_i indicates "consistency of evidence," and u_i indicates "degree of certainty." Finally, the results are sorted by C_i, with the highest score used as the root cause localization output, and the rest retained as alternatives with review suggestions.

[0095] This embodiment does not aim for full supervision when training the evidence network. The sources are divided into three categories: one is hard-labeled samples after the maintenance loop is closed, which are few in number but reliable; another is weak labels from digital twins plus expert review, which tell the model the strength of the association between certain symptoms and components; and the third is the "path feasibility" constraint implied by the propagation network, which is added to the loss as graph regularization to constrain the network not to output conclusions that contradict high-confidence paths.

[0096] In one embodiment of the equipment failure prediction and management method based on a large model in this application, it may further include the following: Step S601: Collect equipment operating status data, segment the status data according to time windows, perform feature extraction on the multimodal data in each time window, input the feature vector into the anomaly detection model to calculate the residual score, identify abnormal data segments based on the adaptive threshold method, input the abnormal data segments into the fault diagnosis model, combine the fault propagation network to calculate the component-level fault probability, and generate fault cause analysis results. Step S602: Read the expert repair case library, extract the repair operation steps and key control parameters, construct a scoring matrix including component importance, repair difficulty and spare parts availability, calculate the weight coefficients of the scoring indicators based on the analytic hierarchy process, multiply the weight coefficients with the scoring matrix to obtain the comprehensive score of the repair operation, prioritize the repair operations according to the comprehensive score, and generate a repair task sequence.

[0097] Optionally, in this embodiment, the reducer and bearing assembly deployed in the main drive area of ​​the cold rolling line form a closed loop for online monitoring of maintenance task scheduling around S601 and S602.

[0098] First, data is aggregated from the PLC and edge acquisition nodes, including speed, torque, current, oil temperature, casing temperature, lubrication pressure difference, triaxial acceleration and envelope of vibration, and image streams of the oil window and end cap. In this embodiment, the data is segmented using sliding windows: short windows capture transients (2–5 s, 50% overlap), and long windows depict the background (60–300 s). After segmentation, the time axis is unified, and mask bits are preserved to avoid excessive interpolation to match the time. Physically relevant features are extracted within each window: on the time-series side, specific energy, temperature rise slope, impact index, and torque pulsation factor are calculated; on the spectrum side, sideband ratio, envelope energy, and full width at half maximum (FWHM) of the resonance peaks are extracted from the order time-frequency plot; on the visual side, edge connectivity, oil surface texture directionality, and hot spot area (if infrared is available) are extracted within the ROI. These features are first normalized by the channel layer of each machine and then written into the feature vector z_t.

[0099] In this embodiment, z_t is input to the anomaly detection model that has been jointly trained with real-world and Siamese algorithms in the S302 stage. The model consists of a reconstructor and a density estimate branch, and outputs the reconstruction residual R_t and the health probability P_t. Considering the non-stationarity of on-site noise, this embodiment does not use a fixed threshold, but rather an adaptive thresholding method: the threshold baseline is updated within a rolling health reference window, taking into account the gradual changes in intraday temperature and load cycles. To avoid false alarms triggered by a single isolated spike, this embodiment applies a temporal stability constraint to the score. Only when the intersection of R_t continuously exceeding the limit and P_t decreasing remains for several frames is the window labeled as an anomaly. In this way, the anomaly segments screened out will contain context for several seconds.

[0100] In this embodiment, the abnormal segment is input into the fault diagnosis model. The input consists of three parts: first, the multimodal symptom embedding of the current segment; second, the prior π of the fault propagation network constructed by S402 for the current component (obtained by matching observed events with time delays); and third, the embedding of similar case evidence retrieved from the S401 / S102 sides. The diagnostic model adopts an evidence-based deep learning framework and outputs the component-level fault probability distribution and uncertainty. The underlying logic here is not mysterious: vibration order is strongly correlated with mechanical defects, oil temperature and lubrication state are coupled with time delay, and oil surface texture in the image appears together with leakage. If these three are consistent with the propagation path in time, the model will increase the posterior probability of the component; if the evidence is contradictory, it will give a higher uncertainty.

[0101] In generating the "fault cause analysis results," this embodiment not only provides the name of the component with the highest probability but also explains the path and evidence. For example, in a night shift segment, the unit specific energy increases, the envelope energy rises, and the 2.9 order sideband is clear, but no leakage is seen in the oil window image. The propagation network has some prior knowledge of the "nozzle → bearing" path under this temperature background, but the time delay does not match (the oil temperature is just starting). The diagnostic output is more biased towards "early defects in the bearing inner ring" and includes a sub-step of "suggesting low-speed frequency sweep and oil sample verification."

[0102] In step S602 of this embodiment, maintenance operation steps and key control parameters are extracted from the expert maintenance case library and manufacturer manuals to form structured entries. Each entry includes the target component, operation steps, tools and torque standards, necessary pre-operation and quality inspection points, as well as possible risk points and alternative solutions. Next, a scoring matrix S is constructed, covering dimensions such as component importance (safety and cascading effects), maintenance difficulty (working hours, skills, disassembly and assembly complexity), spare parts availability (in stock, delivery time), diagnostic gain (contribution to reducing uncertainty), and downtime cost (production planning window). These indicators are not uniform. Importance is given based on the centrality of the component in the propagation network and historical severity statistics. Difficulty is estimated by the depth of the step-dependent network and the parallel mutual exclusion relationship. Spare parts status comes from the warehousing system, and diagnostic gain depends on the current uncertainty and the verification strength corresponding to the operation.

[0103] This embodiment uses the analytic hierarchy process (AHP) to calculate the weight coefficient w of each scoring indicator. The pairwise comparison matrix is ​​initially set by the process manager upon deployment and then fine-tuned during quarterly replays to ensure it aligns with company preferences. To avoid excessive subjectivity, this embodiment incorporates data-driven correction: if the actual benefits (a combination of reduced uncertainty and lower recurrence rate) of a certain type of operation have consistently exceeded expectations in the past, its weight is increased provided the consistency check passes. The final comprehensive score can be written as a simple linear synthesis. :Score(op) = Σ_k w_k · S_k(op), Where Score(op) is the overall score for a maintenance operation, w_k is the weighting coefficient of the k-th indicator, and S_k(op) is the standardized score of the operation on the k-th indicator. The physical meanings of each parameter are "trade-off preference" and "performance of the operation in a certain aspect," respectively. The higher the score, the more worthy it is to be prioritized.

[0104] This embodiment prioritizes candidate maintenance operations and generates a maintenance task sequence. During generation, the step-dependency network is topologically sorted and resource constraints are checked: the same lifting device cannot be used simultaneously, and steps involving hazardous energy isolation must be completed first; rapid verifications with "low cost and high diagnostic gain" (oil samples, frequency sweeps) are prioritized to reduce uncertainty before disassembly and assembly. The task list also inserts time windows adapted to the current production rhythm, such as utilizing roll change intervals for short-term inspections to avoid affecting the production line cycle time.

[0105] In one embodiment of the equipment failure prediction and management method based on a large model in this application, it may further include the following: Step S701: Read the maintenance task sequence, extract the preorder and postorder relationships between maintenance steps, construct a directed acyclic graph to represent the step dependency constraints, generate the execution order of maintenance steps based on the topology sorting algorithm, combine the execution order with the fault diagnosis results, query the expert knowledge base to extract maintenance operation specifications, fill the template of the maintenance operation specifications, and generate a maintenance guidance plan that includes a tool list, maintenance steps, and quality inspection points. Step S702: Collect the execution feedback data of maintenance personnel, extract the maintenance operation time and completion status, update the maintenance feedback data to the vector retrieval library, calculate the maintenance effect score to update the expert knowledge base, label the maintenance feedback data as training samples, update the parameters of the anomaly detection model using the incremental learning method, and adjust the weight coefficients of the fault diagnosis model based on the online distillation method.

[0106] Optionally, this embodiment follows the output of S601-S602, targeting the cold rolling line reducer-bearing maintenance site, and translates "what to do" into "how to do it step by step".

[0107] The system reads candidate operations and their prerequisites from the maintenance task sequence generated by S602, such as "power off and lock → depressurize → remove cover → replace bearing → reassemble → no-load break-in → frequency sweep retest → release". To obtain strict execution relationships, this embodiment parses the constraint phrases of the task description and knowledge entries one by one, extracts the preorder / postorder, parallel mutual exclusion, and resource occupation (lifting tool, electrician, vibration meter) tags, and constructs a directed acyclic graph, where nodes are steps, edges are dependency constraints, and edge attributes are written with necessary conditions and minimum waiting time. For occasional cycles (multiple source texts can easily lead to contradictions of "disassemble before inspect, inspect before disassemble"), the system backtracks to the original entry, cuts a low-weight edge according to the priority of safety rules, and retains hard constraints such as "power off, depressurize, and tag" in the upstream of the graph.

[0108] This embodiment performs topology sorting on a DAG to obtain one or more feasible execution sequences. Considering limited on-site resources and tight production cycles, this embodiment overlays resource scheduling on the topology layer: staggering the timing of steps that occupy the same resources, and inserting concurrent segments for parallel lightweight verifications (such as oil sampling and spare parts pre-inspection). After the sorting is generated, the system references the sequence together with the root cause probabilities of S501-S502. If the diagnostic uncertainty is still high, the "low-cost, high-gain" verification steps are prioritized to be moved forward to converge conclusions as quickly as possible and avoid blind disassembly. After the execution order is determined, the system queries the expert knowledge base to retrieve the operating specifications for the corresponding components and processes, filters irrelevant entries by machine type, oil type, and operating condition, and extracts tools, torque standards, precautions, and quality inspection points.

[0109] This embodiment uses a template to fill in the operating procedures, generating an executable maintenance guidance plan. Template fields include: tool list (specifications, quantity, calibration date), safety isolation and energy release steps, key operating points (torque range, preload sequence, lubrication amount), intermediate quality control points (surface roughness, axial clearance, vibration RMS threshold), retesting methods (low-speed frequency sweep, envelope assessment, oil sample indicators), release conditions, and recurrence observation items. To enhance traceability, each quality control point is bound to a "measurement method—qualified range—abnormal handling branch." For example, after replacing a bearing, the plan is forced to execute "10 minutes of no-load break-in → order 2.9 bandwidth ≤ threshold → oil temperature drop slope ≤ threshold." If this is not met, the plan automatically branches to "recheck preload force—retest—if still unqualified, disassemble and re-inspect." The plan also inserts a "production changeover interval execution" time stamp based on the time window constraints in the task sequence to reduce the impact on cycle time.

[0110] In this embodiment, step S702 involves recording maintenance execution feedback from mobile work orders and edge data acquisition, including start / end times, whether the work was performed in sequence, key measurement values, anomaly notes, and cause classification. The system first analyzes the actual time consumed at each step and decomposes the deviation from the planned time, distinguishing between "waiting for resources," "insufficient personnel," and "technical difficulties." Then, it calculates a maintenance effectiveness score, covering dimensions such as: short-term effectiveness (whether retest indicators meet standards, whether anomaly scores have decreased), medium-term effectiveness (whether recurrence occurs within two weeks), and process compliance (quality inspection point coverage rate, number of violations). The score is not a single value but a set of weighted items, with weights agreed upon by the safety and production managers and adjustable quarterly.

[0111] This embodiment writes feedback data back to the vector retrieval library and the knowledge base. On the vector library side, segments of "phenomenon-treatment-verification-effect" are encoded and stored. High-scoring segments have increased recall weight, while low-scoring or "unverified" segments have decreased weight and are subject to application restrictions to prevent misleading future searches. On the knowledge base side, the "applicable conditions" and "risk points" of operational guidelines are corrected. For example, the clause "cooling buffer needs to be increased during high-temperature seasons" will be solidified into the template after several consecutive positive feedbacks. All feedback is labeled with timestamps and camera location tags as training samples and enters the online learning channel.

[0112] This embodiment employs incremental learning to update the parameters of the anomaly detection model. To prevent short-term biases from distorting the healthy distribution, a strategy of "small step size for new samples, replaying old samples" is used within the training batch. For drift-sensitive channels (such as temperature), seasonal binning parameters are used to avoid the threshold shifting upwards across the board during seasonal changes. The fault diagnosis model uses online distillation: using the current large model as the teacher, soft labels are generated from the latest closed-loop "symptom-diagnosis-treatment-verification" chain; the student model aligns with the teacher's output distribution without accessing the old full dataset, but maintains higher adaptability to newly emerging symptom patterns.

[0113] To effectively address the shortcomings of traditional technologies in feature extraction, fault diagnosis, and maintenance guidance, and to provide technical support for equipment management, this application provides an embodiment of a large-model-based equipment fault prediction management device for implementing all or part of the aforementioned large-model-based equipment fault prediction management method. See [link to embodiment]. Figure 2 The equipment fault prediction and management device based on the large model specifically includes the following components: The equipment data acquisition module 10 is used to collect equipment operation data, sound vibration data, and image data. It performs unified sampling and missing value filling on the operation data, converts the sound vibration data into spectral features, extracts visual features from the image data, constructs a multimodal feature extraction network, inputs the multimodal data into the feature extraction network to generate a fused feature vector, performs word segmentation and entity recognition on the maintenance record to extract fault types and maintenance steps, constructs an equipment component association network, calculates the contrast loss between feature vectors to obtain the equipment state vector, trains the equipment anomaly detection model, and uses a digital twin method to generate fault samples to expand the training set. The fault root cause diagnosis module 20 is used to input the equipment state vector into the large language model encoder to generate semantic representation, construct a retrieval enhancement generation model, read historical maintenance cases to establish a vector retrieval library, calculate the temporal dependency of fault events, construct a component-level fault propagation network, calculate the fault root cause probability based on the fault propagation network, calculate the similarity score between the semantic representation and the retrieval library, extract maintenance records of similar cases, train the fault diagnosis model, and determine the root cause location result based on the confidence score. The equipment fault prediction module 30 is used to monitor real-time equipment status data, input the status data into the anomaly detection model to calculate the anomaly score, input the anomaly data fragment into the fault diagnosis model to generate the fault cause, extract expert maintenance experience to build a knowledge base, calculate the priority score of maintenance operations, establish a dependency constraint network of maintenance steps, generate an executable maintenance guidance plan, collect maintenance feedback data to update the vector retrieval library and knowledge base, write the maintenance feedback data into the training sample set, and perform online updates to the anomaly detection model and the fault diagnosis model.

[0114] As described above, the equipment fault prediction and management device based on a large model provided in this application can achieve accurate status assessment through innovative design of a multimodal feature fusion model, data integration, and feature learning. It constructs a fault diagnosis system, combining semantic analysis and case retrieval to establish a reliable root cause analysis mechanism. Furthermore, it introduces maintenance guidance, ensuring the feasibility of maintenance plans through experience accumulation and priority evaluation. This method effectively addresses the shortcomings of traditional technologies in feature extraction, fault diagnosis, and maintenance guidance, providing technical support for equipment management.

[0115] From a hardware perspective, in order to effectively address the shortcomings of traditional technologies in feature extraction, fault diagnosis, and maintenance guidance, and to provide technical support for equipment management, this application provides an embodiment of an electronic device for implementing all or part of the aforementioned large-model-based equipment fault prediction and management method. The electronic device specifically includes the following components: The system comprises a processor, memory, a communications interface, and a bus; wherein the processor, memory, and communications interface communicate with each other via the bus; the communications interface is used to realize information transmission between the large-model-based equipment fault prediction management device and core business systems, user terminals, and related databases and other related devices; the logic controller can be a desktop computer, tablet computer, or mobile terminal, etc., and this embodiment is not limited to these. In this embodiment, the logic controller can be implemented with reference to the embodiments of the large-model-based equipment fault prediction management method and the large-model-based equipment fault prediction management device, the contents of which are incorporated herein by reference, and repeated details will not be described again.

[0116] It is understood that the user terminal may include smartphones, tablet computers, network set-top boxes, portable computers, desktop computers, personal digital assistants (PDAs), in-vehicle devices, smart wearable devices, etc. Among these, the smart wearable devices may include smart glasses, smartwatches, smart bracelets, etc.

[0117] In practical applications, parts of the device fault prediction and management method based on large models can be executed on the electronic device side as described above, or all operations can be completed in the client device. The choice can be made based on the processing power of the client device and the limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are completed in the client device, the client device may further include a processor.

[0118] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0119] Figure 3 This is a schematic block diagram illustrating the system configuration of the electronic device 9600 according to an embodiment of this application. Figure 3 As shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It is worth noting that... Figure 3 This is an example; other types of structures can also be used to supplement or replace this structure to achieve telecommunications functions or other functions.

[0120] In one embodiment, the equipment fault prediction and management method based on a large model can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control: Step S101: Collect equipment operation data, sound vibration data, and image data; perform unified sampling and missing value imputation on the operation data; convert the sound vibration data into spectral features; extract visual features from the image data; construct a multimodal feature extraction network; input the multimodal data into the feature extraction network to generate a fused feature vector; perform word segmentation and entity recognition on the maintenance records to extract fault types and maintenance steps; construct an equipment component association network; calculate the contrast loss between feature vectors to obtain the equipment state vector; train the equipment anomaly detection model; and use the digital twin method to generate fault samples to expand the training set. Step S102: Input the device state vector into the large language model encoder to generate a semantic representation, construct a retrieval enhancement generation model, read historical maintenance cases to establish a vector retrieval library, calculate the temporal dependency of fault events, construct a component-level fault propagation network, calculate the root cause probability of the fault based on the fault propagation network, calculate the similarity score between the semantic representation and the retrieval library, extract maintenance records of similar cases, train the fault diagnosis model, and determine the root cause localization result based on the confidence score; Step S103: Monitor the real-time status data of the equipment, input the status data into the anomaly detection model to calculate the anomaly score, input the anomaly data fragment into the fault diagnosis model to generate the fault cause, extract expert maintenance experience to build a knowledge base, calculate the priority score of maintenance operations, establish a dependency constraint network of maintenance steps, generate an executable maintenance guidance plan, collect maintenance feedback data to update the vector retrieval library and knowledge base, write the maintenance feedback data into the training sample set, and perform online updates to the anomaly detection model and the fault diagnosis model.

[0121] As described above, the electronic device provided in this application, through the innovative design of a multimodal feature fusion model, achieves accurate state assessment via data integration and feature learning. A fault diagnosis system is constructed, combining semantic analysis and case retrieval to establish a reliable root cause analysis mechanism. Maintenance guidance is introduced, ensuring the feasibility of maintenance plans through experience accumulation and priority evaluation. This method effectively addresses the shortcomings of traditional technologies in feature extraction, fault diagnosis, and maintenance guidance, providing technical support for equipment management.

[0122] In another embodiment, the large-model-based equipment failure prediction management device can be configured separately from the central processing unit 9100. For example, the large-model-based equipment failure prediction management device can be configured as a chip connected to the central processing unit 9100, and the large-model-based equipment failure prediction management method function can be implemented through the control of the central processing unit.

[0123] like Figure 3 As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily need to include these components. Figure 3 All components shown; in addition, the electronic device 9600 may also include Figure 3 For components not shown, please refer to existing technologies.

[0124] like Figure 3 As shown, the central processing unit 9100, sometimes also referred to as a controller or operating control, may include a microprocessor or other processor device and / or logic device, which receives inputs and controls the operation of various components of the electronic device 9600.

[0125] The memory 9140 may be, for example, one or more of a cache, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices. It may store the aforementioned failure-related information, and also store a program for executing that information. The central processing unit 9100 may execute the program stored in the memory 9140 to perform information storage or processing, etc.

[0126] Input unit 9120 provides input to central processing unit 9100. Input unit 9120 may be, for example, a keypad or touch input device. Power supply 9170 provides power to electronic device 9600. Display 9160 displays images and text. Display may be, for example, an LCD display, but is not limited thereto.

[0127] The memory 9140 can be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is off, can be selectively erased, and contains more data; examples of this type of memory are sometimes referred to as EPROMs. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142 for storing application programs and function programs or processes for executing the operation of the electronic device 9600 via the central processing unit 9100.

[0128] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for the electronic device for communication functions and / or for performing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0129] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processing unit 9100 to provide input signals and receive output signals, which is the same as in a conventional mobile communication terminal.

[0130] Based on different communication technologies, multiple communication modules 9110 can be configured in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless LAN modules. The communication module 9110 (transmitter / receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide audio output via the speaker 9131 and receive audio input from the microphone 9132, thereby realizing typical telecommunications functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Additionally, the audio processor 9130 is coupled to a central processing unit 9100, enabling on-device recording via the microphone 9132 and on-device playback of stored audio via the speaker 9131.

[0131] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the large-model-based device fault prediction and management method with a server or client as the execution subject in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the large-model-based device fault prediction and management method with a server or client as the execution subject in the above embodiments. For example, when the processor executes the computer program, it implements the following steps: Step S101: Collect equipment operation data, sound vibration data, and image data; perform unified sampling and missing value imputation on the operation data; convert the sound vibration data into spectral features; extract visual features from the image data; construct a multimodal feature extraction network; input the multimodal data into the feature extraction network to generate a fused feature vector; perform word segmentation and entity recognition on the maintenance records to extract fault types and maintenance steps; construct an equipment component association network; calculate the contrast loss between feature vectors to obtain the equipment state vector; train the equipment anomaly detection model; and use the digital twin method to generate fault samples to expand the training set. Step S102: Input the device state vector into the large language model encoder to generate a semantic representation, construct a retrieval enhancement generation model, read historical maintenance cases to establish a vector retrieval library, calculate the temporal dependency of fault events, construct a component-level fault propagation network, calculate the root cause probability of the fault based on the fault propagation network, calculate the similarity score between the semantic representation and the retrieval library, extract maintenance records of similar cases, train the fault diagnosis model, and determine the root cause localization result based on the confidence score; Step S103: Monitor the real-time status data of the equipment, input the status data into the anomaly detection model to calculate the anomaly score, input the anomaly data fragment into the fault diagnosis model to generate the fault cause, extract expert maintenance experience to build a knowledge base, calculate the priority score of maintenance operations, establish a dependency constraint network of maintenance steps, generate an executable maintenance guidance plan, collect maintenance feedback data to update the vector retrieval library and knowledge base, write the maintenance feedback data into the training sample set, and perform online updates to the anomaly detection model and the fault diagnosis model.

[0132] As described above, the computer-readable storage medium provided in this application, through an innovative design of a multimodal feature fusion model, achieves accurate state assessment via data integration and feature learning. It constructs a fault diagnosis system, combining semantic analysis and case retrieval to establish a reliable root cause analysis mechanism. Furthermore, it introduces maintenance guidance, ensuring the feasibility of maintenance plans through experience accumulation and priority evaluation. This method effectively addresses the shortcomings of traditional technologies in feature extraction, fault diagnosis, and maintenance guidance, providing technical support for equipment management.

[0133] Embodiments of this application also provide a computer program product capable of implementing all steps of the large-model-based equipment fault prediction and management method described above, where the execution subject is a server or client. When executed by a processor, this computer program / instruction implements the steps of the large-model-based equipment fault prediction and management method. For example, the computer program / instruction implements the following steps: Step S101: Collect equipment operation data, sound vibration data, and image data; perform unified sampling and missing value imputation on the operation data; convert the sound vibration data into spectral features; extract visual features from the image data; construct a multimodal feature extraction network; input the multimodal data into the feature extraction network to generate a fused feature vector; perform word segmentation and entity recognition on the maintenance records to extract fault types and maintenance steps; construct an equipment component association network; calculate the contrast loss between feature vectors to obtain the equipment state vector; train the equipment anomaly detection model; and use the digital twin method to generate fault samples to expand the training set. Step S102: Input the device state vector into the large language model encoder to generate a semantic representation, construct a retrieval enhancement generation model, read historical maintenance cases to establish a vector retrieval library, calculate the temporal dependency of fault events, construct a component-level fault propagation network, calculate the root cause probability of the fault based on the fault propagation network, calculate the similarity score between the semantic representation and the retrieval library, extract maintenance records of similar cases, train the fault diagnosis model, and determine the root cause localization result based on the confidence score; Step S103: Monitor the real-time status data of the equipment, input the status data into the anomaly detection model to calculate the anomaly score, input the anomaly data fragment into the fault diagnosis model to generate the fault cause, extract expert maintenance experience to build a knowledge base, calculate the priority score of maintenance operations, establish a dependency constraint network of maintenance steps, generate an executable maintenance guidance plan, collect maintenance feedback data to update the vector retrieval library and knowledge base, write the maintenance feedback data into the training sample set, and perform online updates to the anomaly detection model and the fault diagnosis model.

[0134] As described above, the computer program product provided in this application, through the innovative design of a multimodal feature fusion model, achieves accurate state assessment through data integration and feature learning. It constructs a fault diagnosis system, combining semantic analysis and case retrieval to establish a reliable root cause analysis mechanism. Furthermore, it introduces maintenance guidance, ensuring the feasibility of maintenance plans through experience accumulation and priority evaluation. This method effectively addresses the shortcomings of traditional technologies in feature extraction, fault diagnosis, and maintenance guidance, providing technical support for equipment management.

[0135] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0136] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0137] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0138] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0139] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A large model-based device failure prediction management method, characterized by, The method comprises: Collecting equipment operation data, sound vibration data and image data, performing uniform sampling and missing value filling on the operation data, converting the sound vibration data into spectral features, extracting visual features from the image data, constructing a multi-modal feature extraction network, inputting the multi-modal data into the feature extraction network to generate a fusion feature vector, performing word segmentation and entity recognition on the maintenance records to extract fault types and maintenance steps, constructing a device component association network, calculating the contrast loss between the feature vectors to obtain a device state vector, training an equipment anomaly detection model, and generating a fault sample to expand the training set using a digital twin method; Inputting the device state vector into a large language model encoder to generate a semantic representation, constructing a retrieval enhancement generation model, reading historical maintenance cases to establish a vector retrieval library, calculating the temporal dependence of fault events, constructing a component-level fault propagation network, calculating the fault root cause probability based on the fault propagation network, calculating the similarity score between the semantic representation and the retrieval library, extracting the maintenance records of similar cases, training a fault diagnosis model, and determining the root cause positioning result based on the confidence score; Monitoring real-time state data of the equipment, inputting the state data into the anomaly detection model to calculate an anomaly score, inputting the abnormal data segment into the fault diagnosis model to generate a fault cause, extracting expert maintenance experience to construct a knowledge base, calculating a priority score of the maintenance operation, establishing a dependency constraint network of the maintenance steps, generating an executable maintenance guidance scheme, collecting maintenance feedback data to update the vector retrieval library and the knowledge base, and writing the maintenance feedback data into the training sample set to perform online updating of the anomaly detection model and the fault diagnosis model.

2. The large model-based device failure prediction management method of claim 1, wherein, The collecting equipment operation data, sound vibration data and image data, performing uniform sampling and missing value filling on the operation data, converting the sound vibration data into spectral features, extracting visual features from the image data, constructing a multi-modal feature extraction network, inputting the multi-modal data into the feature extraction network to generate a fusion feature vector, comprises: Collecting operation parameter data from equipment sensors, collecting vibration waveform data from acoustic sensors, and collecting equipment image data from industrial cameras, resampling the operation parameter data at fixed time intervals, filling missing values using a cubic spline interpolation method, performing short-time Fourier transform on the vibration waveform data to obtain a time-frequency spectrogram, and performing convolution operation on the image data to extract edge features and texture features; Constructing a multi-modal feature extraction network comprising a time series encoding layer, a spectral analysis layer and a visual feature layer, inputting the operation parameter data into the time series encoding layer, inputting the time-frequency spectrogram into the spectral analysis layer, inputting the edge features and texture features into the visual feature layer, calculating the weight coefficients of the three feature layers using an attention mechanism, and generating a fusion feature vector by weighting and combining the three layers of features based on the weight coefficients. 3.The large model-based device failure prediction management method of claim 1, wherein, The maintenance record is segmented and entity recognition is performed to extract fault types and maintenance steps, a device component association network is constructed, a comparison loss between feature vectors is calculated to obtain a device state vector, a device anomaly detection model is trained, and a digital twin method is used to generate fault samples to expand the training set, including: Read the historical maintenance record text, extract professional terms and keywords using dictionary segmentation method, annotate fault parts and maintenance actions based on named entity recognition model, construct the fault parts as nodes, and the physical connection relationship between components as edges to generate a device component association network, perform sequence labeling on the maintenance actions to construct a maintenance step table, and write the fault parts and maintenance steps into the annotation dataset; Input the fusion feature vector into the contrast learning network, calculate the contrast loss between the positive sample pair and the negative sample pair, train the feature encoder to generate the device state vector, construct the digital twin model based on the device component association network, input the digital twin model into the physical simulation engine to simulate different working condition fault scenarios and generate simulation data, and combine the simulation data and real data to train the device anomaly detection model. 4.The large model-based device failure prediction management method of claim 1, wherein, The device state vector is input into the large language model encoder to generate semantic representation, construct a retrieval enhancement generation model, read historical maintenance cases to establish a vector retrieval library, calculate the time-dependent relationship of fault events, and construct a component-level fault propagation network, including: Input the device state vector into the large language model encoder to map it to the natural language space, construct fault description text based on the prompt word template, input the fault description text into the large language model to generate a semantic vector, read historical maintenance case records, perform text segmentation and key information extraction on the maintenance cases, construct a maintenance knowledge base containing fault phenomena, root cause analysis, and disposal methods, encode the maintenance knowledge base into vector form to establish a retrieval database; Read the device fault event record, extract the fault occurrence time and affected components, calculate the time interval and propagation probability between component faults, write the propagation probability into an adjacency matrix to construct a fault propagation graph, calculate the in-degree and out-degree of component nodes based on the fault propagation graph, identify the fault source node and affected node, and construct a component-level fault propagation network to calculate the probability distribution of the fault path. 5.The large model-based device failure prediction management method of claim 1, wherein, The fault root cause probability is calculated based on the fault propagation network, the semantic representation and the retrieval library are calculated to obtain a similarity score, the maintenance records of similar cases are extracted, the fault diagnosis model is trained, and the root cause positioning result is determined based on the confidence score, including: Read the node connection relationship in the fault propagation network, calculate the conditional probability of the fault propagation path, construct a fault probability reasoning model based on the Bayesian network, substitute the device state data into the reasoning model to calculate the fault probability distribution of each component, perform nearest neighbor retrieval on the semantic representation vector, calculate the cosine similarity with the historical cases in the retrieval library, and extract the maintenance record with the highest similarity score as the reference case; The fault probability distribution is combined with the reference case to construct a diagnostic input feature, an evidence deep learning network is used to train a fault diagnosis model, a Dirichlet distribution is used to calculate a confidence interval of a diagnosis result, the confidence interval is converted into a reliability score, the fault diagnosis result is sorted according to the reliability score, and the result with the highest score is selected as the root cause positioning output. 6.The large model-based device failure prediction management method of claim 1, wherein, The monitoring device real-time state data, inputs the state data into an anomaly detection model to calculate an anomaly score, inputs an abnormal data segment into a fault diagnosis model to generate a fault cause, extracts expert maintenance experience to construct a knowledge base, and calculates a priority score of a maintenance operation, including: Collecting equipment operating state data, segmenting the state data according to a time window, performing feature extraction on multi-modal data in each time window, inputting the feature vector into an anomaly detection model to calculate a residual score, identifying an abnormal data segment based on an adaptive threshold method, inputting the abnormal data segment into a fault diagnosis model, combining a fault propagation network to calculate a component-level fault probability, and generating a fault cause analysis result; Reading an expert maintenance case library, extracting maintenance operation steps and key control parameters, constructing a scoring matrix including component importance, maintenance difficulty, and spare part availability, calculating weight coefficients of scoring indicators based on an analytic hierarchy process, multiplying the weight coefficients and the scoring matrix to obtain a comprehensive score of the maintenance operation, prioritizing the maintenance operation according to the comprehensive score, and generating a maintenance task sequence. 7.The large model-based device failure prediction management method of claim 1, wherein, The establishment of the dependency constraint network of the maintenance steps generates an executable maintenance guidance scheme, collects maintenance feedback data to update the vector retrieval library and the knowledge base, writes the maintenance feedback data into a training sample set, and performs online updates on the anomaly detection model and the fault diagnosis model, including: Reading a maintenance task sequence, extracting the precedence and sequence relationship between maintenance steps, constructing a directed acyclic graph to represent step dependency constraints, generating an execution order of the maintenance steps based on a topological sorting algorithm, combining the execution order and the fault diagnosis result, querying the expert knowledge base to extract maintenance operation specifications, template filling the maintenance operation specifications, and generating a maintenance guidance scheme including a tool list, maintenance steps, and quality inspection points; Collecting execution feedback data of maintenance personnel, extracting maintenance operation time consumption and completion status, updating the maintenance feedback data to the vector retrieval library, calculating a maintenance effect score to update the expert knowledge base, labeling the maintenance feedback data as a training sample, updating parameters of the anomaly detection model using an incremental learning method, and adjusting weight coefficients of the fault diagnosis model based on an online distillation method.

8. A large model-based device failure prediction management apparatus characterized by comprising: The device includes: The device data acquisition module is used for collecting device operation data, sound vibration data and image data, performing uniform sampling and missing value filling on the operation data, converting the sound vibration data into frequency spectrum features, extracting visual features from the image data, constructing a multi-modal feature extraction network, inputting the multi-modal data into the feature extraction network to generate a fusion feature vector, performing word segmentation and entity recognition on the maintenance records to extract fault types and maintenance steps, constructing a device component association network, calculating a comparison loss between the feature vectors to obtain a device state vector, training a device anomaly detection model, and generating a fault sample to expand a training set by using a digital twin method. The fault root cause diagnosis module is used for inputting the device state vector into a large language model encoder to generate a semantic representation, constructing a retrieval enhancement generation model, reading historical maintenance cases to establish a vector retrieval library, calculating a time sequence dependency relationship of fault events, constructing a component-level fault propagation network, calculating a fault root cause probability based on the fault propagation network, calculating a similarity score between the semantic representation and the retrieval library, extracting maintenance records of similar cases, training a fault diagnosis model, and determining a root cause positioning result based on a confidence score. The device fault prediction module is used for monitoring real-time state data of a device, inputting the state data into an anomaly detection model to calculate an anomaly score, inputting an anomaly data segment into a fault diagnosis model to generate a fault cause, extracting expert maintenance experience to construct a knowledge base, calculating a priority score of a maintenance operation, establishing a dependency constraint network of a maintenance step, generating an executable maintenance guidance scheme, collecting maintenance feedback data to update the vector retrieval library and the knowledge base, writing the maintenance feedback data into a training sample set, and performing online updating on the anomaly detection model and the fault diagnosis model.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the device fault prediction management method based on a large model according to any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the device fault prediction management method based on a large model according to any one of claims 1 to 7.

Images