A composite fault diagnosis method and system of large language model guided multi-model collaborative neural architecture search

Abstract

This invention belongs to the field of industrial fault diagnosis and intelligent operation and maintenance technology. It discloses a composite fault diagnosis method and system guided by a large language model and using a multi-model collaborative neural architecture search. The steps are as follows: (1) Divide the multivariate time series signal of the industrial system into subsystems; (2) Extract local features using the encoder of the corresponding subsystem and realize system-level fault prediction through the global fusion module; (3) Generate a structured performance report based on the system-level accuracy, component-level accuracy and loss on the validation set; (4) Use a large language model to generate architecture optimization actions and iteratively update the model configuration under single-step constraints and acceptance rules. This invention can reduce the cost of manual architecture design, improve the combined generalization ability and stability of composite fault diagnosis in complex industrial systems, and is suitable for system-level fault identification in multi-subsystem coupling scenarios.

Description

Technical Field

[0001] This invention belongs to the field of industrial fault diagnosis and intelligent operation and maintenance technology, specifically involving a composite fault diagnosis method and system based on a multi-model collaborative neural architecture search guided by a large language model. Background Technology

[0002] As the complexity of industrial systems continues to increase and the interactions between subsystems strengthen, intelligent fault diagnosis, especially the generalized diagnosis of complex fault combinations, faces higher demands. In recent years, intelligent fault diagnosis has gradually shifted from feature engineering processes to deep learning methods that can directly learn discriminative features from raw sensor signals. While these end-to-end modeling methods have achieved good results in rotating machinery tasks such as bearings, gears, and motors, real-world industrial scenarios typically require system-level fault diagnosis for the joint operation of multiple heterogeneous subsystems, where faults often appear in coupled or complex forms. Meanwhile, although neural architecture search can reduce manual trial and error to some extent, model structure design and fusion mechanism selection remain highly complex in heterogeneous subsystem scenarios.

[0003] However, existing methods still have the following problems:

[0004] (1) Existing research focuses on isolated diagnosis of single components, which is difficult to apply directly to system-level complex fault scenarios. Most existing deep learning fault diagnosis methods are mainly focused on single objects such as bearings, gears, and motors. However, real industrial systems are usually operated by multiple heterogeneous subsystems. The diagnosis task needs to consider the joint behavior between multiple components and system-level fault categories at the same time. Therefore, the traditional modeling method oriented towards single components is difficult to meet the needs of system-level complex fault diagnosis.

[0005] (2) The large differences in signals and strong cross-component coupling between heterogeneous subsystems make it easy for fault-sensitive patterns to be masked, and the absence of combined faults further increases the difficulty of diagnosis. The signals of different subsystems usually have significant differences in statistical characteristics, fault sensitivity, and representation forms; at the same time, cross-component interactions may distort or mask fault-related patterns, causing the effective response of weakly faulty components to be drowned out by components with strong signals. In particular, combined faults often appear in combination forms that have not appeared in the training phase, which places more stringent requirements on the model's combinatorial generalization ability.

[0006] (3) Existing methods generally rely on expert experience to manually design static architectures or on rule-driven controllers for searching, making it difficult to simultaneously and effectively optimize local encoder selection and system-level fusion configuration under the dynamic conditions of heterogeneous subsystems. For heterogeneous industrial systems, architecture design not only requires selecting appropriate local models for each subsystem, but also determining the system-level fusion mechanism and related training parameters. This results in high costs for manual trial and error and poor engineering portability. Although existing search methods based on rule or heuristic controllers can explore the limited search space, they are difficult to incorporate empirical diagnostic knowledge, structured verification feedback, and cross-iteration error patterns into the update logic in a controllable manner. In particular, it is difficult to achieve coordinated optimization of encoder selection and fusion structure under the conditions of heterogeneous subsystems.

[0007] In summary, there is an urgent need for a method and system that can address system-level complex fault diagnosis scenarios, integrate heterogeneous subsystem modeling with system-level interaction, and utilize large language models to perform constrained optimization of the architecture based on structured performance feedback. Summary of the Invention

[0008] To address the problems in existing system-level composite fault diagnosis technologies, such as large signal differences in heterogeneous subsystems, complex cross-component coupling, difficulty in identifying unseen fault combinations, and reliance on manual trial and error in model structure design, this invention provides a composite fault diagnosis method and system guided by a large language model for multi-model collaborative neural architecture search. This method employs a two-layer optimization framework: the inner loop performs subsystem collaborative learning under a fixed architecture configuration, while the outer loop generates architecture optimization actions based on structured performance feedback from the validation set, iteratively updating the configuration under single-step constraints and acceptance rules.

[0009] To achieve the above objectives, this invention provides a composite fault diagnosis method for multi-model collaborative neural architecture search guided by a large language model, comprising the following steps:

[0010] (1) Obtain the multivariate time series signal of the industrial system and divide it into several non-overlapping signal subsets according to the subsystem;

[0011] (2) Assign corresponding local encoders to each subsystem signal subset, extract subsystem representations and obtain local prediction results;

[0012] (3) Input the representations of each subsystem into the global fusion module, model the cross-subsystem interaction relationship, and obtain the system-level fault prediction results;

[0013] (4) Complete model training under a fixed configuration and generate a structured performance report based on the validation set results;

[0014] (5) Based on the structured performance report, the large language model generates architecture optimization actions and maps them to candidate edits in the search space;

[0015] (6) Perform constraint projection on the candidate edits to obtain candidate configurations;

[0016] (7) Train and validate the candidate configurations. If the validation target meets the preset acceptance conditions, update the current configuration.

[0017] (8) Repeat steps (4) to (7) until the stopping condition is met, and output the optimized composite fault diagnosis model.

[0018] Furthermore, the subsystem representation extraction and local prediction described in step (2) are implemented using a subsystem-specific encoder and a local predictor, and the formula is as follows:

[0019]

[0020] in, For the first Each subsystem corresponds to a subset of signals. For the first The encoder corresponding to each subsystem For the first Characterization of each subsystem For local predictors, For the first Local prediction results of individual subsystems The number of subsystems.

[0021] Furthermore, the global fusion module described in step (3) is configured using fusion structure parameters, and its fusion search space and specific configuration are expressed as follows:

[0022]

[0023] in, To integrate embedded dimensions, For the number of attention heads, For the number of fusion layers, This is the dropout rate.

[0024] Furthermore, the system-level fault prediction in step (3) is achieved using a relational attention fusion mechanism, the formula of which is:

[0025]

[0026] in, For learnable projection matrices, Embedded into a learnable subsystem For the initial token sequence, For fusion encoders, This provides system-level fault prediction results. The design is able to highlight critical components relevant to faults and capture subsystem couplings beyond simple averaging or voting.

[0027] Furthermore, the model training described in step (4) is optimized using a composite objective function, the formula of which is:

[0028]

[0029] in, For system-level loss prediction, For the first The local loss corresponding to each subsystem These are the trainable parameters of the model. and These are the weighting coefficients. After completing training with a fixed configuration, the inner loop evaluates the model on the validation subset and summarizes the system-level metrics, component-level metrics, and validation loss into a structured performance report, which serves as a feedback signal for the outer loop optimization.

[0030] Furthermore, the outer loop search in step (5) adopts a unified configuration representation consisting of the subsystem encoder allocation space, the global fusion configuration space, and the training-related hyperparameter space, and its formula is:

[0031]

[0032] in, Indicates the space allocated to the encoder of the subsystem. Indicates the integrated configuration space. This represents the training-related hyperparameter space. Indicates the current candidate configuration. Indicates assignment to the first The encoder of each subsystem Indicates global fusion configuration. This represents the training-related hyperparameter configuration. This unified representation can support subsystem-specific modeling and system-level aggregation while ensuring compatibility with the embedding interface.

[0033] Furthermore, in steps (5) and (6), the training objective and validation objective under the fixed configuration are respectively expressed as:

[0034]

[0035] On the outer ring In this iteration, the controller receives information including the current configuration and system-level precision. Component-level precision and performance reports verifying the loss The large language model generates natural language optimization actions and maps them to specific edits in the search space. .

[0036] Furthermore, the candidate configuration described in step (6) is updated using a single-step constrained projection, and its formula is:

[0037]

[0038] in, This indicates projection onto the set of feasible configurations. This means that only one architectural decision or hyperparameter decision can be modified in each iteration. This single-step constraint helps improve the stability and interpretability of the optimization process.

[0039] Furthermore, the acceptance determination of the candidate configuration in step (7) adopts an acceptance rule based on the verification target, the formula of which is:

[0040]

[0041] in, This is a tolerance threshold; updates are only accepted if the performance of the candidate configuration on the same validation set improves by at least a preset margin.

[0042] This invention also provides a composite fault diagnosis system for multi-model collaborative neural architecture search guided by a large language model, including a data input and subsystem partitioning module, a local encoder allocation and feature extraction module, a global fusion prediction module, a structured performance report generation module, a large language model optimization control module, a constraint projection module, and an acceptance judgment module; wherein, each module cooperates with the other to implement the above-mentioned composite fault diagnosis method.

[0043] Compared with the prior art, the present invention has the following main advantages:

[0044] 1. By dividing the input signals by subsystem and configuring dedicated encoders, cross-component interference is reduced, which is beneficial for learning more suitable representations for signals from heterogeneous subsystems;

[0045] 2. By modeling cross-subsystem interaction relationships through a global fusion module, the system-level composite fault identification capability is improved;

[0046] 3. The outer loop constrained optimization guided by a large language model reduces the engineering costs caused by traditional manual design and repeated trial and error;

[0047] 4. On the BJTU-RAO system-level bogie dataset, compared with random search NAS, the combined generalization accuracy was improved from 83.24% to 87.42%, and it showed more stable diagnostic performance in multiple repeated experiments. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall framework used in this invention;

[0049] Figure 2 A distribution of generalization accuracy for combinations of different models;

[0050] Figure 3 Box plots of combined generalization accuracy for five repeated experiments with different models;

[0051] Figure 4 The flowchart shows a composite fault diagnosis method and system for multi-model collaborative neural architecture search guided by a large language model, which is provided by the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] This invention provides a composite fault diagnosis method based on a large language model-guided multi-model collaborative neural architecture search. For example... Figure 1 As shown, this invention comprises two main parts: an inner-loop collaborative learning system and an outer-loop constraint optimization system. The inner loop is used to perform feature extraction of multi-subsystem signals, cross-subsystem information fusion, and system-level fault prediction under a given architecture configuration. The outer loop is used to iteratively optimize the architecture configuration based on the verification results and complete the update when preset acceptance conditions are met. In this embodiment, the method specifically includes the following three steps.

[0054] Step 1: Construct an inner-loop collaborative learning model for multi-heterogeneous subsystems to achieve local feature extraction and system-level fault prediction.

[0055] like Figure 1 As shown, the multivariate time series input signal of the industrial system is first acquired. Let the input be...

[0056]

[0057] in, Indicates the length of time. This indicates the number of signal channels. Further, all channels are divided according to subsystems or key components. A set of non-overlapping signals Each signal subset corresponds to a different subsystem to preserve the differences in physical structure and signal distribution between the different subsystems, providing a basis for subsequent differential modeling.

[0058] Regarding the first Each subsystem uses a dedicated encoder. Extract subsystem characterization and through local

[0059] Partial predictor The formula for obtaining the local prediction result is as follows:

[0060]

[0061] in, For the first Each subsystem corresponds to a subset of signals. For the encoder assigned to this subsystem, For subsystem characterization, For local predictors, This is a local prediction result.

[0062] By adopting the above approach, different subsystems can use coding models adapted to their own signal characteristics, thereby reducing mutual interference caused by signal heterogeneity when multi-subsystem hybrid modeling, and providing a local feature basis for subsequent system-level fusion.

[0063] To capture system-level composite failure modes formed by cross-component coupling, local subsystem characteristics are... Further input to the global fusion module. The fusion search space is defined as:

[0064] in, To integrate embedded dimensions, For the number of attention heads, For fusion layer

[0065] number, This refers to the dropout rate; specific fusion configurations are denoted as:

[0066]

[0067] This parameterization method enables the outer ring to search for the fused structure under a unified representation.

[0068] Furthermore, the global fusion module first maps the features of each subsystem to construct a fusion input sequence, the formula of which is:

[0069]

[0070] in, For learnable projection matrices, Embedded into a learnable subsystem; let This represents the initial token sequence. Then, it is obtained through a fusion encoder:

[0071]

[0072] The final system-level prediction results are expressed as follows:

[0073]

[0074] in, For fusion encoders, This is the result of system-level fault prediction.

[0075] This design can highlight fault-related critical components and characterize cross-subsystem coupling relationships that go beyond simple averaging or voting.

[0076] Under a fixed configuration, the collaborative model is trained using a composite objective function, the formula of which is:

[0077]

[0078] in, For system-level loss prediction, For the first The local loss corresponding to each subsystem These are the trainable parameters of the model. and These are the weighting coefficients. After completing training with a fixed configuration, the model is evaluated on a validation subset, and the system-level metrics, component-level metrics, and validation loss are summarized into a structured performance report, which serves as feedback input for outer-loop optimization.

[0079] Step two: Construct a unified outer-ring search space and generate candidate architectures based on structured performance reports.

[0080] like Figure 1 As shown, to achieve unified optimization of the local encoder, global fusion structure, and training parameters, an overall search space is constructed, consisting of the subsystem encoder allocation space, the fusion configuration space, and the training-related hyperparameter space. Its formula is:

[0081]

[0082] in, Indicates the space allocated to the encoder of the subsystem. This indicates the global fusion configuration space. This represents the training-related hyperparameter space.

[0083] For those with For an industrial system with heterogeneous subsystems, a candidate configuration is represented as:

[0084]

[0085] in, Indicates the current candidate configuration. This represents the encoder allocation set for each subsystem. Indicates assignment to the first The encoder of each subsystem Indicates global fusion configuration. This represents the training-related hyperparameter configuration. This unified configuration allows different subsystems to use different types of local models while maintaining consistency in the downstream fusion interface.

[0086] Given configuration Then, the inner loop for continuous parameters The optimization objective is expressed as follows:

[0087]

[0088] Furthermore, the current configuration is evaluated by verifying the objective function:

[0089]

[0090] in, Indicates in configuration The parameter results obtained from the training are as follows. Indicates configuration The corresponding verification and evaluation results.

[0091] Furthermore, in the During the second outer loop iteration, the current configuration and its corresponding system-level accuracy, component-level accuracy, and validation loss are input into the large language model optimization module. The optimization module then generates an optimization action for the current configuration and maps this optimization action to candidate edits in the search space. .

[0092] By adopting the above method, joint optimization of subsystem local encoders, system-level fusion structures, and key training parameters can be achieved simultaneously under a unified configuration representation, thereby improving the flexibility and adaptability of the overall architecture search.

[0093] Step 3: Use single-step constraint projection and acceptance rules to complete the configuration update and output the optimized composite fault diagnosis model.

[0094] like Figure 1 As shown, to ensure the stability and interpretability of the optimization process, only one single-step edit is allowed per outer loop iteration. Candidate configurations are constructed as follows:

[0095]

[0096] in, Indicates the first Current configuration at the next iteration This represents the candidate edits generated and mapped by the large language model optimization module. Indicates candidate configurations. This indicates that feasible domain projection is performed on the candidate configuration. This indicates a configuration update operation. This means that only one architectural decision or hyperparameter decision can be modified in each iteration.

[0097] For candidate configurations Retrain the inner loop and evaluate it again on the same validation split. The acceptance rules for candidate configurations are as follows:

[0098]

[0099] in, An update is only accepted if the performance of the candidate configuration on the same validation set improves by at least a preset margin, as a tolerance threshold; otherwise, the current configuration is retained and the process proceeds to the next iteration. The paper also points out that to ensure the operability of the refinement, a fixed training budget is used throughout the process, and the same validation partition is always used, ensuring that the accepted updates truly reflect the performance improvement brought about by the configuration itself, rather than fluctuations caused by changes in the evaluation protocol.

[0100] In one specific embodiment, system-level bogie data, comprising four key components—traction motor, gearbox, left axle box, and right axle box—was used as the experimental subject. This data was collected synchronously from vibration signals of each component via distributed sensors, encompassing 21 types of local faults and 51 types of system-level faults. Sample construction employed a sliding window of length 1024 points. Model training was conducted within a deep learning framework, with 8 training rounds for the inner loop and a large language model used for the outer loop optimization module.

[0101] Furthermore, the candidate local encoders in this embodiment include RNN-CNN, TFSCL, DS-CNN, CFSPT, TCN, BiLSTM-Attention, Inc-ResCNN, GatedCNN, and LightweightTransformer; the fusion configuration space includes parameters such as fusion embedding dimension, number of attention heads, number of fusion layers, and dropout rate; the training-related space includes local network structure variables and key training hyperparameters.

[0102] like Figure 2 As shown, under the combined generalization accuracy metric, the method of this invention achieves better results than various comparative methods, reaching 87.42%, which is higher than the 83.24% of the random search architecture search method. Figure 3As shown, in repeated experiments, the combined generalization accuracy of the method of the present invention fluctuates little, indicating that the present invention has good stability in system-level composite fault diagnosis tasks.

[0103] like Figure 3 As shown, in five repeated experiments, the combined generalization accuracy of the proposed method fluctuated less, indicating that the proposed distributed collaborative architecture exhibits more stable diagnostic performance under different test fault categories. The paper further points out that this distributed architecture also includes a degradation case: when all components share the same sub-model, it can degenerate into homogeneous full-channel modeling. Therefore, the proposed method maintains compatibility while possessing greater architectural flexibility to adapt to the dynamics of heterogeneous subsystems.

[0104] In summary, this invention combines local coding modeling of multiple heterogeneous subsystems, system-level fusion prediction, and outer-loop constraint optimization guided by a large language model to achieve automatic architecture optimization for complex industrial system composite fault diagnosis tasks. This can effectively reduce manual design costs and improve the accuracy and stability of system-level fault identification.

[0105] It will be readily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A composite fault diagnosis method based on multi-model collaborative neural architecture search guided by a large language model, characterized in that: The method includes the following steps: (1) Obtain the multivariate time series signal of the industrial system and divide it into several non-overlapping signal subsets according to the subsystem; (2) Assign corresponding local encoders to each subsystem signal subset, extract subsystem representations and obtain local prediction results; (3) Input the representations of each subsystem into the global fusion module, model the cross-subsystem interaction relationship, and obtain the system-level fault prediction results; (4) Complete model training under a fixed configuration and generate a structured performance report based on the validation set results; (5) Based on the structured performance report, the large language model generates architecture optimization actions and maps them to candidate edits in the search space; (6) Perform constraint projection on the candidate edits to obtain candidate configurations; (7) Train and validate the candidate configurations. If the validation target meets the preset acceptance conditions, update the current configuration. (8) Repeat steps (4) to (7) until the stopping condition is met, and output the optimized composite fault diagnosis model.

2. The composite fault diagnosis method for multi-model collaborative neural architecture search guided by a large language model as described in claim 1, characterized in that: The subsystem representation extraction and local prediction are implemented using a dedicated encoder and local predictor, and the formula is as follows: In the formula, For the first The signal subsets corresponding to each subsystem; For the first The encoders corresponding to each subsystem; For the first Characterization of each subsystem; For local predictors; For the first Local prediction results for each subsystem; The number of subsystems.

3. The composite fault diagnosis method for multi-model collaborative neural architecture search guided by a large language model as described in claim 1, characterized in that: The global fusion module is configured using fusion structure parameters, the formula of which is: In the formula, Configuration for the global fusion module; To integrate embedded dimensions; For the number of attention heads; Number of fusion layers; This is the dropout rate.

4. The composite fault diagnosis method for multi-model collaborative neural architecture search guided by a large language model as described in claim 3, characterized in that: The system-level fault prediction is achieved using a relational attention fusion mechanism, the formula of which is: In the formula, For learnable projection matrix; Embed vectors for subsystems; For the first The initial token corresponding to each subsystem; This is the initial token sequence; For fusion encoder; This is the representation sequence after the fusion layer; This is the result of system-level fault prediction.

5. The composite fault diagnosis method for multi-model collaborative neural architecture search guided by a large language model as described in claim 1, characterized in that: The model training employs a composite objective function for optimization, the formula of which is: In the formula, This is the total loss function; For system-level loss prediction; For the first Local losses corresponding to each subsystem; System-level target labels; This is a system-level prediction result; This is a local prediction result; These are the trainable parameters of the model; and These are the weighting coefficients.

6. The composite fault diagnosis method for multi-model collaborative neural architecture search guided by a large language model as described in claim 1, characterized in that: The outer loop search adopts a unified configuration representation consisting of subsystem encoder allocation, global fusion configuration, and training-related hyperparameters, and its formula is: In the formula, For the overall search space; Allocate space for the subsystem encoder; Provide a global fusion configuration space; To train the relevant hyperparameter space; For one candidate configuration; Assign sets to the encoders of each subsystem; To be assigned to the The encoder of each subsystem; Configure for global integration; Configure the relevant hyperparameters for training.

7. The composite fault diagnosis method for multi-model collaborative neural architecture search guided by a large language model as described in claim 1, characterized in that: The inner loop training results are evaluated using the training objective and the validation objective, and the formula is as follows: In the formula, For the current configuration; In configuration The optimal parameters obtained from the training; For training objective function; To verify the objective function; To verify the loss function.

8. The composite fault diagnosis method for multi-model collaborative neural architecture search guided by a large language model as described in claim 1, characterized in that: The candidate configuration is updated using a single-step constrained projection, and the formula is as follows: In the formula, For the first The current configuration for the next iteration; Candidate edits are generated and mapped from a large language model; Candidate configuration; A constraint projection operator for the set of feasible configurations; This indicates a configuration update operation; This means that only one architectural decision or hyperparameter decision can be modified in each iteration.

9. The composite fault diagnosis method for multi-model collaborative neural architecture search guided by a large language model as described in claim 1, characterized in that: The acceptance determination of the candidate configuration adopts an acceptance rule based on the verification target, and the formula is as follows: In the formula, For the first Configuration after the next iteration; Candidate configuration; To verify the objective function; The acceptance threshold.

10. A composite fault diagnosis system based on a large language model-guided multi-model collaborative neural architecture search, characterized in that, include: The data input and subsystem partitioning module is used to acquire multivariate time series signals of industrial systems and partition them into subsystems; the local encoder allocation and feature extraction module is used to extract the characteristics of each subsystem and obtain local prediction results. The global fusion prediction module is used to complete cross-subsystem interactive modeling and output system-level fault prediction results; The structured performance report generation module is used to generate structured performance reports based on the verification results. The large language model optimization control module is used to generate architecture optimization actions based on the structured performance report; The constraint projection module is used to perform constraint projection on candidate edits; The acceptance determination module is used to determine whether to accept the candidate configuration update based on the verification target; wherein, each module cooperates with the other to execute the method described in any one of claims 1 to 9.

Images