Bearing fault detection system and method based on llm feedback and physical constraints

Abstract

This disclosure provides a bearing fault detection system and method based on LLM feedback and physical constraints, relating to the field of equipment condition monitoring technology. The method acquires the one-dimensional raw vibration signal and rotational speed information of the bearing, generating physical evidence vectors, visual evidence vectors, and knowledge evidence vectors respectively. The three types of evidence are normalized and spatially aligned to generate a logical conflict matrix. When the logical conflict matrix meets preset conflict triggering conditions, a structured reflection log is generated. The structured reflection log and expert mechanism text are input into a large language model, outputting control primitives containing filter center frequency, filter bandwidth, peak-finding threshold, or search window. A digital signal processing unit parses the control primitives and resets the signal processing parameters, re-executing envelope spectrum analysis and fault feature extraction until preset convergence conditions are met, outputting the bearing fault category and interpretable diagnostic results. This disclosure improves the adaptability and traceability of bearing fault detection under high-noise conditions.

Description

Technical Field

[0001] This disclosure relates to the field of equipment condition monitoring technology, specifically to a bearing fault detection system and method based on LLM feedback and physical constraints. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] In the fields of modern heavy industrial equipment and precision agricultural machinery, the operating status of rotating components such as bearings directly determines the operational safety and power transmission efficiency of the entire machine. Bearings are typically subjected to harsh operating conditions involving variable loads, non-steady speeds, and strong random vibrations, making their early, subtle fault characteristics easily masked by high-energy background noise. Therefore, achieving highly reliable fault diagnosis in low signal-to-noise ratio (SNR) environments has always been a core technology in the field of predictive health management (PHM).

[0004] While existing bearing diagnostic methods can effectively detect faults to a certain extent, they still have the following limitations: (1) The static parameters of traditional digital signal processing (DSP) systems are difficult to adapt to time-varying operating conditions. Traditional mechanism analysis methods rely heavily on preset parameters such as filter passband and peak-finding threshold. Under non-stationary operating conditions, the mechanical resonance frequency band often drifts dynamically with rotational speed and load, while existing preprocessing modules use a fixed frequency band capture strategy. When faced with broadband weak impacts caused by early damage to rolling elements, static parameters cannot achieve adaptive tracking of frequency band shifts, resulting in the fault characteristic energy being severely diluted by background noise, and the system has poor environmental adaptability.

[0005] (2) Purely data-driven models lack physical causal constraints and are prone to statistical illusions. Most existing deep learning models are "black box" systems without mechanistic support and rely excessively on time-frequency texture features. In noisy environments, time-frequency images often undergo geometric distortion and feature aliasing, leading to misjudgments in model output that violate common sense about dynamics (for example, when time-domain indicators show significant impact features, the model still outputs a "normal" conclusion due to probability distribution deviation). In addition, existing technologies lack effective cross-modal fusion mechanisms, making it difficult to align non-parametric expert experience with parametric model inference in real time, and the diagnostic results lack logical traceability.

[0006] (3) Existing diagnostic architectures are mostly unidirectional open-loop structures, lacking adaptive control mechanisms under conflicting multi-source evidence. Existing multimodal diagnostic systems generally follow a linear process of "data input - unidirectional perception - conclusion output". When visual semantic features conflict with physical hard indicators, the system cannot exercise logical arbitration, nor does it have the ability to issue reset instructions to the underlying sensing and processing chain. This open-loop architecture cannot autonomously initiate "active detection" based on the reasoning state, making it difficult to achieve closed-loop collaboration between the underlying algorithm parameters and the top-level cognitive logic, resulting in severely insufficient decision robustness under complex interference conditions. Summary of the Invention

[0007] To address the aforementioned issues, this disclosure proposes a bearing fault detection system and method based on LLM feedback and physical constraints. By constructing a forward processing link of "perception-verification-decision" and a backward control closed loop of "feedback-recapitulation-iteration", the diagnostic logic is transformed from probability matching to physical causal driving.

[0008] According to some embodiments, the present disclosure adopts the following technical solutions: A bearing fault detection method based on LLM feedback and physical constraints includes: Acquire bearing vibration signals and corresponding operating condition information; Based on the bearing vibration signal, physical evidence vector, visual evidence vector, and knowledge evidence vector are generated respectively. A consistency check is performed on the physical evidence vector, visual evidence vector, and knowledge evidence vector to generate a logical conflict matrix. When the logical conflict matrix meets the preset conflict triggering conditions, physical priority arbitration is performed and a structured reflection log is generated; Based on the structured reflection log and the expert mechanism text corresponding to the knowledge evidence vector, control primitives for resetting signal processing parameters are generated through a large language model; The control primitives are parsed and the corresponding signal processing parameters are reset. Fault features are re-extracted from the bearing vibration signal and the physical evidence vector is updated. Iteratively perform consistency checks, parameter resets, and fault feature extractions until the preset convergence conditions are met, and then output the bearing fault category.

[0009] According to some embodiments, the present disclosure adopts the following technical solutions: A bearing fault detection system based on LLM feedback and physical constraints includes: The signal acquisition module is used to acquire bearing vibration signals and corresponding operating condition information; The perception module is used to generate physical evidence vectors, visual evidence vectors, and knowledge evidence vectors based on the bearing vibration signals, respectively. The mechanism constraint verification module is used to perform consistency verification on the physical evidence vector, visual evidence vector and knowledge evidence vector, and generate a logical conflict matrix; when the logical conflict matrix meets the preset conflict triggering conditions, physical priority arbitration is performed and a structured reflection log is generated. The adaptive diagnostic decision-making and active detection module is used to generate control primitives for resetting signal processing parameters based on the structured reflection log and the expert mechanism text corresponding to the knowledge evidence vector through a large language model; parse the control primitives and reset the corresponding signal processing parameters, re-extract fault features from the bearing vibration signal and update the physical evidence vector; iteratively perform consistency verification, parameter reset and fault feature re-extraction until the preset convergence condition is met, and output the bearing fault category.

[0010] According to some embodiments, the present disclosure adopts the following technical solutions: A computer program product includes a computer program that, when executed by a processor, implements the aforementioned bearing fault detection method based on LLM feedback and physical constraints.

[0011] According to some embodiments, the present disclosure adopts the following technical solutions: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned bearing fault detection method based on LLM feedback and physical constraints.

[0012] According to some embodiments, the present disclosure adopts the following technical solutions: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the bearing fault detection method based on LLM feedback and physical constraints.

[0013] Compared with the prior art, the beneficial effects of this disclosure are as follows: The bearing fault detection method based on LLM feedback and physical constraints disclosed herein solves the technical problem that existing open-loop systems cannot autonomously issue control commands and dynamically reset the underlying sensing processing parameters (such as filter cutoff frequency and peak finding threshold) based on the confidence state of the upper layer inference, thus breaking the perception limitations brought about by unidirectional data flow.

[0014] This disclosed bearing fault detection method based on LLM feedback and physical constraints addresses the safety risks of high-confidence erroneous conclusions in long-tailed weak fault scenarios by introducing causal arbitration logic based on first principles of physics. It also resolves the statistical misjudgment (i.e., statistical illusion) caused by feature distortion and aliasing in a single-sensor mode under extremely strong background noise interference (e.g., SNR < -4dB), which violates the common sense of mechanical dynamics.

[0015] This disclosure presents a bearing fault detection method based on LLM feedback and physical constraints. By constructing an active cross-modal retrieval enhanced generation (RAG) mechanism, it addresses the problem that traditional systems cannot generate white-box reasoning reports covering the entire chain of "evidence tracing - mechanism verification - conclusion determination," thereby meeting the stringent requirements of transparency and accountability for diagnostic equipment in industrial TRL-8 level applications. It also addresses the issue of existing large-model diagnostic systems lacking logical transparency and failing to meet industrial safety verification needs in vertical industrial applications. This disclosure resolves the deficiency that offline maintenance manuals, engineering expert experience, and other non-parametric textual knowledge cannot be dynamically aligned with the underlying real-time diagnostic signal flow, realizing a shift in diagnostic logic from probabilistic matching to physical causal driving. Attached Figure Description

[0016] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0017] Figure 1 This is an implementation framework diagram of the bearing fault detection method based on LLM feedback and physical constraints according to an embodiment of this disclosure. Detailed Implementation

[0018] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0020] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0021] Example 1 One embodiment of this disclosure provides a bearing fault detection method based on LLM feedback and physical constraints, the method steps of which include: Step 1: Obtain the one-dimensional original vibration signal of the bearing and the corresponding rotational speed information; Step 2: Extract physical features from the one-dimensional original vibration signal, calculate the theoretical fault characteristic frequency based on the bearing geometric parameters and rotational speed information, and perform time-domain statistical analysis and envelope spectrum analysis to obtain the physical evidence vector; Step 3: Convert the one-dimensional original vibration signal into a two-dimensional time-frequency image and input it into a visual perception network with a residual geometric correction adapter to obtain a visual evidence vector; Step 4: Extract the original visual feature vector from the adapter of the pre-trained visual backbone network, and retrieve expert mechanism texts from the expert mechanism knowledge base based on the original visual feature vector to obtain the knowledge evidence vector; Step 5: Normalize and spatially align the physical evidence vector, visual evidence vector, and knowledge evidence vector, and calculate the cross-modal differences to generate a logical conflict matrix; Step Six: When the logical conflict matrix meets the preset conflict triggering conditions, perform physical priority arbitration based on the physical evidence vector and generate a structured reflection log; Step 7: Input the structured reflection log and the expert mechanism text into the large language model, so that the large language model outputs a control primitive containing at least one parameter in the filter center frequency, filter bandwidth, peak finding threshold, and search window; Step 8: The digital signal processing unit parses the control primitives and resets the corresponding signal processing parameters according to the control primitives. It then re-executes envelope spectrum analysis and fault feature extraction on the one-dimensional original vibration signal to update the physical evidence vector. Step 9: Repeat the logical conflict matrix generation, physical priority arbitration, and parameter reset until the logical conflict matrix meets the preset convergence condition, and output the bearing fault category and the corresponding interpretable diagnostic results.

[0022] As one embodiment, this disclosure presents a bearing fault detection method based on LLM feedback and physical constraints. By constructing a forward processing link of "perception-verification-decision" and a backward control closed loop of "feedback-recapitulation-iteration," the diagnostic logic is transformed from probability matching to physical causal driving. The specific execution process is as follows: Step 1: Perform multidimensional sensing on the one-dimensional original vibration signal, and perform time-domain dynamic calculation and envelope spectrum analysis, visual feature reconstruction and alignment, and cross-modal knowledge retrieval to obtain physical evidence, visual evidence, and knowledge evidence.

[0023] As one embodiment, this disclosure is responsible for performing physical feature extraction, visual representation reconstruction, and cross-modal knowledge association tasks in this step. The specific implementation details are as follows: Step 11: Obtain the one-dimensional original vibration signal of the bearing under strong noise conditions (variable load); The specific process for obtaining the one-dimensional original vibration signal of the bearing under strong noise conditions (variable load) in this disclosure includes: (1) Hardware connection and deployment: 1) Sensor installation: A piezoelectric accelerometer is arranged radially in the bearing housing and rigidly fixed by a magnetic mount or bolts to capture transient impacts caused by bearing damage.

[0024] 2) Signal conditioning: The analog charge signal output by the sensor is connected to a charge amplifier or signal conditioning module for gain amplification, and then filtered out high-frequency interference noise by an anti-aliasing low-pass filter.

[0025] (2) Key acquisition parameters: 1) Sampling frequency: In order to cover the high-frequency resonance band of bearing failure (usually 3kHz-10kHz) and satisfy the sampling theorem, the system sampling frequency is set to no less than 20kHz.

[0026] Digital conversion: Analog signals are converted to digital signals through a high-precision 16-bit A / D converter, and the sampling depth must be sufficient to distinguish subtle changes in the amplitude of early faults.

[0027] As one embodiment, this disclosure uses an accelerometer arranged radially at the bearing housing to continuously acquire vibration sequences affected by variable loads at a sampling frequency of not less than 20 kHz. The acquired analog current signals are amplified and converted to digital signals using a 16-bit A / D converter to form a one-dimensional original vibration vector containing time-series characteristics. X= { x 1 ,x 2 ,…x n} 。 The numerical sequence was then fed into the multidimensional perception module, serving as the common input basis for temporal calculus, visual reconstruction, and knowledge retrieval.

[0028] Step 12: Perform multidimensional sensing on the one-dimensional original vibration signal, and execute time-domain dynamic calculation and envelope spectrum analysis.

[0029] Specifically, a one-dimensional raw vibration signal is received. The expert-level digital signal processing (DSP) subunit dynamically calculates the theoretical fault characteristic frequency using formulas based on preset bearing geometric parameters and real-time rotational speed input. Simultaneously, time-domain physical quantity calculations and envelope spectrum analysis are performed to calculate the overlap of energy distribution between the measured spectral peaks and theoretical frequencies, outputting a mechanism matching score matrix. This step achieves deep alignment between the measured signal and the physical mechanism through the following specific algorithm flow: (1) Calculate the time-domain statistical index of the one-dimensional original vibration signal and construct a multi-dimensional time-domain physical quantity feature vector, including: Simultaneously calculate N time-domain statistical indices of the one-dimensional original vibration signal to construct an feature vector:

[0030] Where RMS is the root mean square value, used to assess overall vibration energy; K is kurtosis, used to capture transient impact characteristics in the signal; C is the crest factor, used to measure the severity of the impact; and S is the impulse index. The system presets a baseline vector under normal conditions. Calculate the current and The deviation, as a preliminary fault, has a confidence weight. .

[0031] (2) Perform envelope spectrum analysis, and use an adaptive window function mapping algorithm to calculate the energy overlap ratio, obtaining the overlap ratio between the energy integral of the measured envelope spectrum within the observation window and the theoretical pulse energy within the window, including: After performing envelope spectrum analysis, the system does not simply compare peak values. Instead, it uses an adaptive window function mapping algorithm to calculate the energy integral of the measured envelope spectrum in the theoretical fault characteristic frequency and its harmonic neighborhood, and the matching ratio between this energy and the corresponding theoretical reference pulse energy. This quantifies the degree of consistency between the measured signal and the theoretical fault mechanism.

[0032] in, To measure the envelope spectrum power density, For the first k Reference energy distribution corresponding to the theoretical fault pulse. For the first i Theoretical characteristic frequencies of this type of fault k =1,2,3 represent the first three harmonics. For half the width of the observation window, To prevent extremely small constants with a denominator of zero.

[0033] (3) Combining the time-domain confidence score and the scores in each frequency domain, a structured mechanism matching score matrix is ​​output as physical evidence, including: Comprehensive time-domain confidence Scores in each frequency domain Output a structured mechanism matching score matrix :

[0034] in, (Inner ring fault score) represents the self-consistent probability between the envelope spectrum characteristics of the measured vibration signal and the theoretical mechanism of bearing inner ring fault (such as BPFI and its harmonics). (Outer ring fault score) represents the self-consistent probability between the envelope spectrum characteristics of the measured vibration signal and the theoretical mechanism of bearing outer ring fault (such as BPFO and its harmonics). (Rolling element fault score) represents the self-consistent probability between the envelope spectrum characteristics of the measured vibration signal and the theoretical mechanism of bearing rolling element faults (such as BSF and its harmonics). The (normal state score) represents the probability that the measured signal has no significant energy distribution in each fault characteristic frequency band and that the time-domain index is within the baseline range, reflecting the confidence level that the bearing is in a healthy state. Each element in the matrix represents the mechanism self-consistency probability of the corresponding fault mode. For example, a higher Sinner indicates that the energy distribution of the envelope spectrum at the inner ring characteristic frequency is highly consistent with the physical mechanism deduction. This matrix will serve as the core criterion in the subsequent mechanism constraint verification module. w 1 (Time Domain Confidence Weight) is calculated by using the currently acquired time domain feature vector. (Including root mean square value, kurtosis, peak factor, etc.) and the preset normal state baseline vector The deviation is obtained and used to provide a preliminary confidence assessment of the presence of the fault from the perspective of overall vibration energy. (Frequency domain mechanism locking score) characterizes the measured envelope spectrum at theoretical fault characteristic frequencies. The overlap ratio between the energy integral within the narrowband observation window centered on the theoretical pulse energy is quantized using an adaptive window function mapping algorithm to determine the frequency domain mechanism matching degree.

[0035] As one implementation, the expert-level digital signal processing (DSP) subunit can be manifested at the physical hardware level as a processor core with computing capabilities. Specifically, it can be a dedicated digital signal processing chip, the core of an embedded microprocessor, or a high-speed computing module running in an industrial control computer. This hardware processes the digitized vibration vectors in real time through pre-stored algorithm programs, outputs a mechanism matching score matrix, and supports parameter resetting based on upper-level feedback instructions.

[0036] Step 13: Perform visual feature reconstruction and alignment on the one-dimensional original vibration signal to obtain visual evidence.

[0037] Specifically, the signal is mapped to a two-dimensional time-frequency image using continuous wavelet transform. This two-dimensional time-frequency image is then input into a sensing network integrating a residual geometric correction adapter. This part consists of two components: a backbone sensing network and a built-in residual geometric correction adapter. (1) Backbone: A pre-trained deep visual neural network ViT is used to extract high-order semantic features from time-frequency images.

[0038] (2) Residual Geometric Correction Adapter: Employs a lightweight bottleneck structure, embedded between layers of the backbone network. Internally, it includes: Linear 1 (Dimension Reduction Layer): Projects the high-dimensional input feature x into a low-dimensional space to compress the feature and capture the core geometric skeleton.

[0039] Nonlinear activation layer (σ): Utilizes the ReLU function to enhance the nonlinear representation of features and filter out small-amplitude clutter caused by strong noise.

[0040] Linear2 layer: Remaps the processed features back to the original dimension so that the residuals can be summed with the original features.

[0041] Dynamic gain control module: A calculation unit that calculates the gain coefficient α (SNR) in real time based on the measured signal-to-noise ratio (SNR).

[0042] This adapter corrects the texture geometric distortion caused by strong noise through residual terms, and its forward propagation formula is:

[0043] in, This refers to the dynamic residual gain coefficient calculated based on the environmental noise energy distribution. To ensure a smooth transition of the system under different noise gradients, this disclosure constructs a dynamic activation strategy based on the Sigmoid function, the calculation formula of which is:

[0044] in, The maximum residual weight threshold (preferred range is 0.4 to 0.8), and SNR is the signal-to-noise ratio evaluation value of the real-time acquired signal. The preset noise tolerance threshold, This is a smoothing sensitivity adjustment factor. This dynamic mapping model ensures that when the signal-to-noise ratio drops sharply under operating conditions, The value can quickly approach the upper limit, activating the strongest feature geometric correction; while in a high signal-to-noise ratio steady state, The value decays to near zero, preserving the original high-fidelity texture features.

[0045] Ultimately, the perceptual network's final output is visual evidence, specifically manifested as a visual soft confidence vector. This vector consists of the probabilities of each fault category (which sum to 1), representing the system's preliminary probability judgment of the current bearing status (such as normal, inner race fault, outer race fault, etc.) from the perspective of visual features.

[0046] As one example, the final output of the perceptual network is a visual soft confidence vector. This vector contains the probability distributions of C preset fault categories. The acquisition process first reconstructs the one-dimensional signal into a two-dimensional time-frequency image using continuous wavelet transform; subsequently, the residual geometric correction adapter integrated in the sensing network dynamically adjusts the gain coefficient based on the measured signal-to-noise ratio. The image features are subjected to nonlinear geometric correction; finally, the network output layer outputs the probability distributions for each category, providing visual evidence support for the subsequent mechanism constraint verification module. The complete logic of transforming a signal from a one-dimensional sequence into visual evidence is as follows: (1) Time-frequency mapping stage: The one-dimensional original vibration signal is converted into a two-dimensional time-frequency image using continuous wavelet transform (CWT). This image can intuitively present the energy distribution of the transient impact excited by the bearing failure on the time-frequency axis.

[0047] (2) Feature input and splitting stage: Time-frequency image input to the sensing network. When the feature stream x passes through a specific network layer, it is synchronously input into the residual geometric correction adapter path.

[0048] (3) Environmental adaptive calibration stage: The system evaluates the environmental signal-to-noise ratio in real time.

[0049] Under high noise conditions (SNR is below the preset threshold τ), the dynamic activation strategy causes α(SNR) to increase rapidly and approach the maximum weight threshold. This enhances the adapter's ability to correct image texture distortion.

[0050] Under high signal-to-noise ratio steady state, α(SNR) tends to zero, and the system automatically turns off the correction function to preserve the original high-fidelity texture features.

[0051] (4) Residual Feature Fusion Stage: The correction term calculated by the adapter is multiplied by the dynamic gain coefficient, and the result is obtained by applying the residual formula. It is fused with the original features to correct feature shifts caused by background noise.

[0052] (5) Evidence generation stage: After correction, the features continue to propagate forward in the perceptual network, and finally the probability distribution of each fault category is output through the Softmax function to form visual evidence.

[0053] Step 14: Perform cross-modal knowledge retrieval on the one-dimensional original vibration signal to obtain knowledge evidence.

[0054] Specifically, this step is implemented by the Knowledge Retrieval Subunit (RAG) using an implicit semantic anchoring mechanism based on source domain features, and the specific implementation process is as follows: (1) Construction and extraction of high-dimensional feature vectors: from frozen pre-trained visual backbone networks (such as CLIP) The original visual feature vector is truncated after the penultimate normalization layer (LayerNorm) of the model (ViT-B / 32 or a self-supervised learning-based MAE backbone). This visual feature vector is typically a D-dimensional dense tensor. Since the truncation point is located before the residual geometric correction adapter, this feature only contains the original texture distribution information of the time-frequency map, effectively avoiding semantic compression caused by the adapter during fine-tuning in specific fault tasks.

[0055] Specifically, the original visual feature vector is obtained by truncating the output of the penultimate normalized layer of the frozen pre-trained visual backbone network (CLIP ViT-B / 32). This feature is a D-dimensional dense tensor, and because its truncation position is before the residual geometric correction adapter, it can completely preserve the original geometric features of the time-frequency image, avoiding interference with semantic information in subsequent task fine-tuning, and providing an accurate semantic anchoring basis for cross-modal knowledge retrieval.

[0056] (2) Sub-millisecond retrieval logic based on FAISS: The local vectorized knowledge base pre-converts the expert mechanism text into embeddings of the same dimension using a Text Encoder and stores them in an index structure based on FAISS. During retrieval, the cosine similarity between the query vector and the vectors in the database is calculated, and the top K text blocks with the highest similarity scores are recalled in real time using the K-nearest neighbor algorithm.

[0057] (3) Structure the expert mechanism text blocks: Each recalled text block is not a jumble of text, but rather structured data containing the following three core fields: a. Prior values ​​of physical indicators: theoretical characteristic frequencies (such as inner ring fault frequency multiplication fBPFI) and expected values ​​of theoretical kurtosis corresponding to the fault modes; b. Causal mechanism logic: Describe the failure evolution path (e.g., pitting → periodic pulse → high-frequency resonance); c. Search range constraints: Provide the initial search step size and frequency domain search window for dynamically reset parameters.

[0058] (4) Mapping mechanism of physical boundary constraints: The recalled Top-K text blocks are used as the dynamic context for DeepSeek-V3 model inference. LLM extracts the corresponding physical hard index threshold range from the text blocks through the semantic understanding capability of the large language model (for example, the bandpass filter frequency of the inner ring fault of a certain bearing model at 2000rpm should be 220Hz±15Hz). This range constitutes the physical boundary constraint for subsequent diagnostic decisions. If the calculation result of the underlying DSP deviates from this range, it will be judged as a logical conflict and reported.

[0059] Step 2: Quantitatively assess the consistency of physical evidence, visual evidence, and knowledge evidence at the level of physical mechanism, and generate a logical conflict matrix.

[0060] Specifically, the consistency of the multimodal evidence at the physical mechanism level is quantitatively assessed, generating a logical conflict matrix. The specific algorithmic implementation details for this step are as follows: (1) Normalized spatial alignment of multi-source evidence: Extract evidence vectors of the same dimension from three modal evidences (assuming the system has C preset fault categories): A. Visual soft confidence vector The probability of each fault category output by the visual perception network (satisfying...) ).

[0061] B. Physical Hard Mechanism Vector Mechanism matching score matrix output by expert-level DSP (expert-level digital signal processing) subunit Obtained after normalization.

[0062] C. Physical boundary constraint vector Derived from the knowledge retrieval subunit (RAG). When the measured physical properties of the DSP (such as kurtosis, envelope spectrum peak value) meet the requirements of a specific fault... i When the expert theory threshold range is reached, ,otherwise .

[0063] (2) Calculation of cross-modal component difference measurement. A multimodal cross-residual calculation method is used to quantify the degree of orientation divergence between pairs of modes, including: 1) Conflict between visual models and physical calculations: Calculate the vector of absolute differences between visual probability and physical score:

[0064] in, is the scale alignment factor.

[0065] 2) Conflict between visual model and prior knowledge: The Hadamard product is used to calculate the penalty for visual probabilities exceeding physical boundaries.

[0066] If vision predicts a certain fault with high probability, but physical indicators violate prior common sense ( If ), then the difference value at the corresponding position will increase dramatically.

[0067] 3) Logical Conflict Matrix Generation and state determination: Concatenate the above difference feature vectors to construct a dimension of... Logical conflict matrix :

[0068] To achieve quantitative evaluation, a conflict threshold matrix is ​​introduced. Perform a binarization activation operation to generate a state matrix. (in (This is an indicator function).

[0069] At this point, uniformity can be quantified by calculating the matrix norm or the rank of the state matrix. When the number of non-zero elements (or the rank) of the matrix is ​​greater than 0, a logical directional deviation is determined, and this conflict matrix... It will be directly transmitted to the multi-level logic arbitration executor to guide the veto or trigger the active detection loop.

[0070] Step 3: Set up a multi-level logic arbitration executor, and guide a veto or trigger the multi-level logic arbitration executor to actively detect based on the logic conflict matrix.

[0071] The actuator incorporates a hierarchical arbitration rule based on the confidence level of physical evidence. Specifically, when the physical evidence vector meets a preset strong confidence condition, physical evidence verified by physical boundary constraints is prioritized as the arbitration baseline. When the physical evidence vector is within a preset ambiguity range, visual evidence vectors are allowed to participate in the fusion judgment according to the visual gain coefficient, triggering an active detection loop to avoid misjudgments caused by single-modal evidence. The specific logical hierarchy is as follows: (1) Physically Strong Constraint Priority Arbitration: When the fault probability confidence level output by the visual perception network deviates directionally from the physical evidence vector calculated in real time by the digital signal processing unit, the system first determines whether the physical evidence vector meets the preset strong confidence conditions. The preset strong confidence conditions include: the highest score in the physical evidence vector is greater than the preset physical confidence threshold, and the energy overlap at the corresponding fault feature frequency is greater than the preset energy threshold, or the time-domain impact index exceeds the corresponding fault threshold, and the physical evidence vector is consistent with the physical boundary constraints in the knowledge evidence vector. When the preset strong confidence conditions are met, the arbitrator uses the fault category corresponding to the physical evidence vector as the diagnostic baseline for this round and suppresses visual prediction results that are inconsistent with the physical evidence; at the same time, a structured reflection log is generated for subsequent large language model to generate control primitives and triggers the digital signal processing unit to perform parameter reset and secondary feature extraction.

[0072] As one example, if the visual model determines it to be "normal", but the physical evidence vector shows that the non-Gaussian impact energy exceeds the standard, and the envelope spectrum has significant energy overlap near the theoretical fault characteristic frequency, then the arbitrator does not adopt the "normal" conclusion of the visual model, but uses the fault category pointed to by the physical evidence as the diagnostic baseline for this round; if the visual model determines it to be fault A, but the envelope spectrum characteristic frequency and knowledge evidence both point to fault B, then the arbitrator uses fault B as the diagnostic baseline for this round and triggers subsequent active detection closed loop.

[0073] (2) Trigger boundary for visual gain compensation: The blur zone is defined in real time through the following quantization logic, and the degree of intervention of the visual modality is dynamically controlled: Step 1: Quantification of Signal Quality and Mechanism Reliability. The system monitors the signal-to-noise ratio (SNR) of the input raw signal in real time through the underlying sensing processing unit. When it detects... And the theoretical fault frequency in the envelope spectrum When the peak signal-to-noise ratio (PSNR) is below 1.2, the system determines that it has entered the physical feature flooding zone.

[0074] Step 2: Defining the critical interval of physical indicators. The system synchronously extracts time-domain physical indicators and determines whether they are within the fuzzy boundary: A. Kurtosis determination: When the kurtosis index K is in the range of [3.0, 4.5] (i.e. higher than the baseline value but not exceeding the significant fault threshold); B. Energy matching score determination: The mechanism matching score matrix described in Section 2.1.1 Highest score item hour.

[0075] Step 3: Visual Feature Strengthening Mechanism for the Chain of Evidence. The arbitration executor is only allowed to activate the visual gain coefficient when both of the above conditions are simultaneously met. ( At this point, the geometric distribution features captured by the visual perception network are converted into probability components. The system performs evidence chain fusion calculations:

[0076] in, It increases moderately as SNR decreases. This mechanism utilizes weakly correlated textures captured by the visual modality for logical reinforcement when physical features are not obvious, thereby improving the overall sensitivity of the diagnostic system. The physical evidence vector is the physical diagnostic evidence formed by the combination of time-domain statistical indicators and envelope spectrum mechanism matching scores.

[0077] Step 4: Forced fusion based on first principles of physics. Once the hard physical indicators provide a clear conclusion (such as...) Confirmed or (After troubleshooting), the arbitrator immediately executes a physical logic circuit breaker, forcibly... Reset to 0. This operation ensures that the visual compensation mechanism automatically fails in deterministic scenarios, preventing statistical illusions from large language models from interfering with qualitative physical conclusions.

[0078] Step 4: When the logical conflict matrix meets the preset conflict triggering conditions, the feature offsets of the conflict samples are automatically parsed to generate a structured reflection log containing timestamps, deviations of key physical parameters, and candidate mechanism causes. The structured reflection log and expert mechanism text are input into the large language model, which outputs a control primitive containing at least one parameter in the filter center frequency, filter bandwidth, peak finding threshold, and search window. The digital signal processing unit parses the control primitive and resets the corresponding signal processing parameters according to the control primitive. The envelope spectrum analysis and fault feature extraction are re-executed on the one-dimensional original vibration signal to update the physical evidence vector. The logical conflict matrix generation, physical priority arbitration, and parameter reset are repeated until the logical conflict matrix meets the preset convergence conditions, and the bearing fault category and corresponding interpretable diagnostic results are output.

[0079] Specifically, upon receiving the reflection log, forward diagnostics are suspended, and a "feedback-recapture-iteration" backward control loop is initiated. Unlike traditional static rule mapping, the decision center of this system constructs a Chain of Thought (CoT) inference workflow based on a large model LLM. LLM extracts "structured reflection logs" based on a dynamic context window, using missing frequency band features and the mechanism text retrieved by RAG as joint prompt trigger words. Subsequently, LLM strictly follows the structured Chain of Thought (CoT) template of "time-frequency anomaly analysis → fault mechanism anchoring → compensation parameter deduction" for deduction. The specific computer implementation steps of this deduction process are as follows: (1) Time-frequency anomaly analysis stage: LLM first analyzes the feature offsets in the reflective log. Specifically, the model extracts the actual kurtosis values ​​from the log. K real Compared with the theoretical prior kurtosis threshold K prior The difference, and the frequency bands in the envelope spectrum where the peak corresponding to the theoretical fault frequency does not appear.

[0080] Example of an LLM inference instruction: "According to the reflection log, the current bearing speed is 2000 rpm, the theoretical outer ring fault frequency is 120 Hz, but the measured envelope spectrum has no significant peak in the 100-150 Hz range, and the background noise energy in the high-frequency band (3 kHz-5 kHz) accounts for more than 85%. Please analyze the physical cause of the characteristic frequency being submerged." (2) Fault Mechanism Anchoring Stage: Based on the time-frequency anomaly analysis results, LLM combined with expert mechanism texts recalled by RAG is used to construct a multiple hypothesis tree. The model eliminates hypotheses that contradict the current working conditions through contextual logic matching, and anchors the most core physical mechanism cause.

[0081] LLM inference process: The model deduces that the high-frequency noise submersion is due to the improper setting of the cutoff frequency of the initial bandpass filter, which fails to effectively cover the system resonance frequency band induced by the weak bearing fault, resulting in an extremely low signal-to-noise ratio before envelope demodulation. The LLM is responsible for receiving the structured reflection log containing the deviation of physical parameters, and combining it with the physical boundary constraints retrieved through RAG, using the Chain of Thought (CoT) to deduce and generate control instructions for resetting the underlying DSP parameters.

[0082] (3) Compensation parameter derivation stage: After anchoring the cause, LLM enters the parameter solution step. At this time, LLM forcibly calls the external physics algorithm library. LLM uses the rotational speed and abnormal frequency band in stage one as constraints, instructing the external algorithm to re-search for the resonance band corresponding to the maximum kurtosis within the constraint range.

[0083] (4) Parameter calculation: The external algorithm returns the center frequency of the optimal resonance band. fc and bandwidth B w LLM encapsulates these purely numerical results into machine-executable instructions based on preset control primitive templates.

[0084] Following the rigorous CoT derivation described above, the LLM ultimately generates JSON-formatted control primitives containing precise action instructions. The underlying DSP sub-unit directly parses these JSON instructions, performs directional bandpass filter parameter resets and secondary refined peak finding, until the updated feature matrix is ​​fed back to achieve logical consistency.

[0085] Furthermore, the underlying physical signal processing matrix, arbitration decision logic, and final diagnostic conclusion are structurally mapped to output a white-box report following the architecture of "[physical evidence source]-[causal reasoning chain]-[final judgment result]". This report achieves a structured mapping from the underlying multi-source heterogeneous sensor data state space to the high-level semantic space, ensuring the physical traceability of the diagnostic results.

[0086] Example 2 One embodiment of this disclosure provides a bearing fault detection system based on LLM feedback and physical constraints, including: The signal acquisition module is used to acquire bearing vibration signals and corresponding operating condition information; The perception module is used to generate physical evidence vectors, visual evidence vectors, and knowledge evidence vectors based on the bearing vibration signals, respectively. The mechanism constraint verification module is used to perform consistency verification on the physical evidence vector, visual evidence vector and knowledge evidence vector, and generate a logical conflict matrix; when the logical conflict matrix meets the preset conflict triggering conditions, physical priority arbitration is performed and a structured reflection log is generated. The adaptive diagnostic decision-making and active detection module is used to generate control primitives for resetting signal processing parameters based on the structured reflection log and the expert mechanism text corresponding to the knowledge evidence vector through a large language model; parse the control primitives and reset the corresponding signal processing parameters, re-extract fault features from the bearing vibration signal and update the physical evidence vector; iteratively perform consistency verification, parameter reset and fault feature re-extraction until the preset convergence condition is met, and output the bearing fault category.

[0087] As one example, such as Figure 1 As shown, solid arrows indicate the flow of multimodal features from underlying data acquisition and multidimensional perception to logical arbitration; red dashed arrows indicate the closed-loop control command flow issued from the LLM decision center to the underlying digital signal processing subunit for resetting algorithm parameters. The figure highlights the topological relationship of the cross-modal retrieval line before the residual geometric correction adapter (Adapter).

[0088] As one embodiment, the main modules and topology logic are described below: 1. Multidimensional Perception Module: This module includes an expert-level digital signal processing (DSP) unit, a visual feature reconstruction unit, and a cross-modal knowledge retrieval unit. The visual feature reconstruction unit integrates a residual geometric correction adapter to correct image distortion under strong noise conditions.

[0089] 2. Cross-modal knowledge retrieval topology: The system extracts uncontaminated original feature vectors from the frozen visual backbone network as retrieval indexes, and performs FAISS nearest neighbor search before the adapter performs specific task projection. This topological order ensures semantic alignment between the retrieved expert mechanism text and the original physical signal.

[0090] 3. Mechanism Constraint Verification Module: Performs logical consistency evaluation of multimodal evidence. When the visual soft probability distribution deviates from or conflicts logically with the physical hard indicators, the multi-level logic arbitration executor enforces the arbitration rule that physical mechanisms take precedence and generates a structured reflection log.

[0091] 4. LLM Decision Center: Receives reflection logs and suspends forward diagnostics. The LLM infers compensation parameters based on the Chain of Thought (CoT) inference template, autonomously generates JSON-formatted instructions containing control primitives, and feeds them back to the underlying DSP unit via the red dotted line path.

[0092] Active detection feedback closed loop: After the underlying DSP unit parses the JSON instructions, it dynamically resets the filter passband and peak finding threshold, performs secondary refined feature extraction, until logical self-consistency is achieved and a white-box interpretable diagnostic report is generated.

[0093] Example 3 One embodiment of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned bearing fault detection method based on LLM feedback and physical constraints.

[0094] Example 4 One embodiment of this disclosure provides a non-transitory computer-readable storage medium for storing computer instructions. When these computer instructions are executed by a processor, they implement the bearing fault detection method based on LLM feedback and physical constraints.

[0095] Example 5 One embodiment of this disclosure provides an electronic device, including a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the bearing fault detection method based on LLM feedback and physical constraints.

[0096] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. 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, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0097] 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.

[0098] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A bearing fault detection method based on LLM feedback and physical constraints, characterized in that, include: Acquire bearing vibration signals and corresponding operating condition information; Based on the bearing vibration signal, physical evidence vector, visual evidence vector, and knowledge evidence vector are generated respectively. A consistency check is performed on the physical evidence vector, visual evidence vector, and knowledge evidence vector to generate a logical conflict matrix. When the logical conflict matrix meets the preset conflict triggering conditions, physical priority arbitration is performed and a structured reflection log is generated; Based on the structured reflection log and the expert mechanism text corresponding to the knowledge evidence vector, control primitives for resetting signal processing parameters are generated through a large language model; The control primitives are parsed and the corresponding signal processing parameters are reset. Fault features are re-extracted from the bearing vibration signal and the physical evidence vector is updated. Iteratively perform consistency checks, parameter resets, and fault feature extractions until the preset convergence conditions are met, and then output the bearing fault category.

2. The method for bearing fault detection based on LLM feedback and physical constraints as claimed in claim 1, wherein, A physical evidence vector is generated based on the bearing vibration signal, including: Calculate the time-domain statistical index of the one-dimensional original vibration signal and construct a multi-dimensional time-domain physical quantity feature vector; Envelope spectrum analysis is performed, and an adaptive window function mapping algorithm is used to calculate the energy integral of the measured envelope spectrum in the theoretical fault characteristic frequency and its harmonic neighborhood, and the matching ratio between the energy of the corresponding theoretical reference pulse; By combining the time-domain confidence score and the scores in each frequency domain, a structured mechanism matching score matrix is ​​output as physical evidence.

3. The method for bearing fault detection based on LLM feedback and physical constraints as claimed in claim 1, wherein, Based on the bearing vibration signal, a visual evidence vector is generated, including: The one-dimensional original vibration signal is mapped to a two-dimensional time-frequency image using continuous wavelet transform. Two-dimensional time-frequency images are input into a perception network that integrates a residual geometric correction adapter. The residual term corrects the texture geometric distortion caused by strong noise. The probability of each fault category is output based on a dynamic activation strategy using the Sigmoid function to obtain visual evidence.

4. The bearing fault detection method based on LLM feedback and physical constraints as described in claim 1, characterized in that, Based on the bearing vibration signal, a knowledge evidence vector is generated, including: Extract the original visual feature vectors from the penultimate normalization layer of the frozen pre-trained visual backbone network; The local vectorized knowledge base pre-converts expert mechanism texts into embeddings of the same dimension using a Text Encoder and stores them in a FAISS-based index structure. During retrieval, the cosine similarity between the query vector and the vectors in the database is calculated. The K-nearest neighbor algorithm is used to recall the top K text blocks with the highest similarity scores in real time, construct expert mechanism text blocks, and extract the corresponding physical hard index threshold range as knowledge evidence.

5. The bearing fault detection method based on LLM feedback and physical constraints as described in claim 1, characterized in that, The consistency of the physical evidence vector, visual evidence vector, and knowledge evidence vector is checked to generate a logical conflict matrix, including: Normalized spatial alignment is performed on physical evidence, visual evidence, and knowledge evidence to obtain physical hard mechanism vector, visual soft confidence vector, and physical boundary constraint vector of the same dimension, respectively. A multimodal cross-residual calculation method is used to quantify the degree of directional divergence between pairwise modal vectors and obtain the difference vector. The difference vectors are concatenated to obtain the logical conflict matrix.

6. The bearing fault detection method based on LLM feedback and physical constraints as described in claim 1, characterized in that, The multi-level logic arbitration executor has built-in hierarchical arbitration rules based on the confidence level of physical evidence: when the physical evidence vector meets the preset strong confidence condition, the fault category corresponding to the physical evidence vector is used as the diagnostic baseline for this round; when the physical evidence vector is in the preset fuzzy range, the visual evidence vector is allowed to participate in the fusion judgment according to the visual gain coefficient and trigger the active detection closed loop; the specific logic hierarchy is divided into: physical strong constraint priority arbitration and visual gain compensation trigger boundary.

7. A bearing fault detection system based on LLM feedback and physical constraints, characterized in that, include: The signal acquisition module is used to acquire bearing vibration signals and corresponding operating condition information; The perception module is used to generate physical evidence vectors, visual evidence vectors, and knowledge evidence vectors based on the bearing vibration signals, respectively. The mechanism constraint verification module is used to perform consistency verification on the physical evidence vector, visual evidence vector and knowledge evidence vector, and generate a logical conflict matrix. When the logical conflict matrix meets the preset conflict triggering conditions, physical priority arbitration is performed and a structured reflection log is generated; The adaptive diagnostic decision-making and active detection module is used to generate control primitives for resetting signal processing parameters based on the structured reflection log and the expert mechanism text corresponding to the knowledge evidence vector through a large language model; parse the control primitives and reset the corresponding signal processing parameters, re-extract fault features from the bearing vibration signal and update the physical evidence vector; iteratively perform consistency verification, parameter reset and fault feature re-extraction until the preset convergence condition is met, and output the bearing fault category.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the bearing fault detection method based on LLM feedback and physical constraints as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the bearing fault detection method based on LLM feedback and physical constraints as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the bearing fault detection method based on LLM feedback and physical constraints as described in any one of claims 1-6.

Images