Mechanism constraint semantic encoding vibration diagnosis and evidence chain tracing method and system

Abstract

The application relates to the technical field of equipment state monitoring and intelligent operation and maintenance, and discloses a vibration diagnosis and evidence chain tracing method and system based on mechanism constraint semantic coding, which acquires equipment vibration, rotating speed and working condition parameters, generates multi-domain features through preprocessing, extracts layered mechanism atomic labels according to a fault physical model, binds information such as data sources and quality scores to construct traceable evidence objects, generates candidate fault hypotheses through joint coding, executes conflict arbitration through D-S evidence theory and a spectrum identification matrix, and outputs a diagnosis conclusion with an evidence number through mechanism consistency verification, finally generates a maintenance work order, and iteratively updates evidence weights and identification boundaries through retest results. The application suppresses diagnosis hallucination through explicit mechanism constraint, realizes full-link traceability of diagnosis, improves the robustness of non-steady-state and sensor abnormal scenes, forms a closed-loop continuous optimization diagnosis capability for operation and maintenance, and is suitable for the fields of rotating machinery state monitoring and intelligent operation and maintenance.

Description

Technical Field

[0001] This application relates to the technical field of equipment condition monitoring and intelligent operation and maintenance, and in particular to vibration diagnosis and evidence chain tracing methods and systems based on mechanism constraint semantic coding. Background Technology

[0002] Vibration diagnosis of rotating machinery is a core part of industrial equipment operation and maintenance. Traditional diagnostic methods rely on domain experts to manually analyze the spectrum, envelope spectrum, and phase relationship. Although they have strong interpretability, they are highly dependent on personnel experience, have low diagnostic efficiency, and are difficult to standardize, making them unsuitable for the real-time monitoring needs of large-scale equipment clusters.

[0003] With the development of machine learning and deep learning technologies, data-driven fault diagnosis methods have been widely used. These methods can automatically extract features from the time domain, frequency domain, or time-frequency graphs and complete fault classification, exhibiting high recognition efficiency under steady-state conditions. However, existing data-driven methods generally lack explicit constraints on the physical mechanisms of rotating machinery faults. The model learning process relies solely on the statistical distribution of data, failing to directly reflect the physical evolution of typical faults such as imbalance, misalignment, looseness, rolling bearing failure, and oil film eddy. This results in black-box diagnostic results, and the output conclusions are prone to deviations from the physical mechanisms, leading to insufficient interpretability.

[0004] Meanwhile, existing methods struggle to link diagnostic conclusions with original signals, measurement point locations, and processing chains, failing to form a complete and traceable chain of evidence and thus failing to meet the needs of operation and maintenance auditing and liability determination in industrial scenarios. Under non-steady-state conditions such as start-up and shutdown, and variable loads, as well as sensor anomaly scenarios, traditional spectrum analysis methods suffer significant accuracy degradation due to frequency smearing. Existing data-driven models lack generalization ability and cannot effectively adapt to variable speed conditions. Furthermore, they lack effective degradation diagnostic mechanisms when measurement point signals are missing or distorted, resulting in poor system robustness. In addition, existing diagnostic systems often stop at outputting fault conclusions, failing to form a complete closed loop with on-site operation and maintenance work orders and repair and testing processes. This prevents the reverse optimization of diagnostic models based on actual operation and maintenance results, making it difficult to continuously improve diagnostic accuracy with the accumulation of operational data. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a vibration diagnosis and evidence chain tracing method and system based on mechanism constraint semantic encoding.

[0006] Firstly, this application provides a vibration diagnosis and evidence chain tracing method based on mechanism constraint semantic encoding, employing the following technical solution: The vibration diagnosis and evidence chain tracing method based on mechanism-constrained semantic coding includes the following steps: Acquire vibration signals, rotational speed signals, structural parameters, and operating parameters of the target equipment; The vibration signal is preprocessed to generate time-domain, frequency-domain, order-domain, and / or envelope-domain feature representations; Based on the physical model of structural parameters, operating parameters and fault frequency, hierarchical mechanism atomic markers are extracted from the feature representation; By binding the hierarchical mechanism atomic tags with the original data source, measurement point channel, time window, transformation link, mechanism attribute, quality score, and applicable working conditions, a traceable evidence object is constructed. The hierarchical mechanism atomic markers, evidence objects, operating condition information and location information are jointly encoded to generate a diagnostic input sequence, output multiple candidate fault hypotheses and associate them with corresponding supporting evidence and conflicting evidence. Perform conflict arbitration and mechanism consistency verification on candidate fault hypotheses, and output the sorted diagnostic conclusions and traceable evidence numbers; A repair work order is generated based on the diagnostic conclusions. After the repair is completed, the evidence scoring weights and fault identification boundaries are updated based on the retest acceptance results.

[0007] Optionally, when the target device is in an acceleration, deceleration or speed fluctuation condition, the vibration signal is subjected to angular domain resampling and order tracking processing to generate an order spectrum and transient feature representation; at the same time, the activation weights of various mechanism atomic markers are calculated according to the current condition vector, and the evidence objects corresponding to the markers with activation weights lower than a preset threshold are downweighted or converted to a state to be supplemented.

[0008] Optionally, the layered mechanism atomic labeling includes a three-layer structure: L1 basic quantity markers are used to characterize at least one of amplitude, frequency, order, phase, and orbital geometry parameters; L2 structural relation markers are used to characterize at least one of harmonic families, sideband groups, modulation relations, order coupling, and multichannel coherence. L3 fault semantic markers are used to characterize at least one of the following: imbalance, misalignment, looseness, rolling bearing failure, oil film whirl, and structural resonance.

[0009] Optionally, the quality score of the evidence object is calculated by weighting multiple indicators such as signal-to-noise ratio, sensor health, mechanism correlation, operating condition matching degree, and cross-channel consistency; and a time-series attenuation factor is introduced to adjust the quality score over time, with the attenuation coefficient increasing synchronously with the rate of change of rotation speed, load, and temperature.

[0010] Optionally, the conflict arbitration adopts DS evidence theory to fuse the basic probability assignments of each evidence object to obtain the trust function and similarity function of each candidate fault hypothesis; then, the candidate faults ranked first are reviewed by the spectrum feature identification matrix, and the final arbitration result is obtained by combining the trust function value and the review score; the judgment boundary of the spectrum feature identification matrix is ​​self-calibrated online by quantile statistics according to the closed-loop acceptance result.

[0011] Optionally, when an abnormality or absence of the measurement point signal is detected, the virtual sensor signal is reconstructed based on the adjacent channel signal through a time delay propagation model, the reconstruction confidence is calculated and propagated to the quality score of the corresponding evidence object; when the reconstruction confidence is lower than a preset threshold, the system switches to single-channel degradation diagnosis mode and outputs a coarse-grained fault category conclusion.

[0012] Optionally, based on the trust function value and the width of the plausible interval of the candidate fault, the diagnostic results are divided into an automatic confirmation area, an auxiliary decision-making area, and a manual takeover area; the automatic confirmation area directly generates a maintenance work order, the auxiliary decision-making area outputs an evidence summary and the best supplementary test suggestion, and the manual takeover area triggers on-site expert review.

[0013] Optionally, after the maintenance is completed, the same collection and processing process as before the maintenance is used to obtain the evidence objects after the maintenance, and the changes in key feature quantities before and after the maintenance are compared to output the acceptance conclusion. Based on the acceptance results, the weights of the evidence quality indicators are updated using the Bayesian method, and the historical case library and the baseline of the same model of equipment group are updated simultaneously.

[0014] Secondly, this application provides a vibration diagnosis and evidence chain tracing system based on mechanism constraint semantic encoding, which adopts the following technical solution: A vibration diagnosis and evidence chain tracing system based on mechanistic constraint semantic coding includes: The data acquisition and preprocessing module is used to acquire the vibration, rotational speed, structural and operating parameters of the target equipment and generate multi-domain feature representations. The mechanism labeling and evidence construction module is used to extract hierarchical mechanism atomic labels and construct traceable evidence objects with quality scores; The joint coding and diagnostic reasoning module is used to generate unified semantic codes and output candidate fault hypotheses and related evidence; The conflict arbitration and verification module is used to perform conflict arbitration and mechanism consistency verification, and output diagnostic conclusions with evidence numbers. The work order and closed-loop update module is used to generate maintenance work orders and iteratively update the diagnostic model parameters based on the retest results; The modules interact with each other through mechanism atomic tag data packets and evidence object data packets.

[0015] In summary, this application includes at least the following beneficial technical effects: By constructing a three-layer hierarchical mechanism atomic labeling system, this invention transforms the physical mechanism of vibration faults into encodeable semantic units, providing explicit mechanism constraints for the diagnostic process, fundamentally reducing the illusion risk of data-driven models and large language models, and improving the physical consistency of diagnostic results.

[0016] This invention binds diagnostic conclusions to the original data source, measurement point channel, processing link, and quality attributes across the entire evidence chain, achieving full traceability of diagnostic results and meeting the auditing and liability determination needs of industrial scenarios. Through a conflict arbitration mechanism combining DS evidence theory and a spectral feature discrimination matrix, coupled with mechanism consistency verification, it effectively improves diagnostic accuracy in complex operating conditions and scenarios with multiple conflicting pieces of evidence. Furthermore, the online self-calibration capability of the discrimination matrix can adapt to assembly differences and operating characteristics of different equipment, avoiding compatibility issues caused by fixed thresholds.

[0017] This invention utilizes a virtual sensor reconstruction and confidence propagation mechanism to maintain effective diagnosis even when measurement point signals are abnormal or missing. Combined with a single-channel degradation mode, it enhances the system's robustness and scenario adaptability. Through a closed-loop process of work order generation and retest acceptance, actual operation and maintenance results are used to update evidence weights and discrimination boundaries, achieving continuous iterative optimization of diagnostic capabilities.

[0018] This invention uses a three-level partitioning mechanism based on human-machine collaboration to ensure diagnostic efficiency in high-confidence scenarios, control the risk of misjudgment in low-confidence scenarios, and adapt to the operational decision-making needs under different confidence levels.

[0019] Overall, this invention achieves an organic combination of mechanism constraints, evidence tracing, conflict arbitration, and operation and maintenance closed loop, significantly improving the interpretability, robustness, and engineering application value of vibration diagnosis. It can be widely applied to condition monitoring and intelligent operation and maintenance scenarios of various rotating machinery such as gearboxes, motors, pumps, fans, and steam turbines. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of the vibration diagnosis and evidence chain tracing method of the present invention; Figure 2 This is a schematic diagram of the hierarchical structure of the atomic markers in the layered mechanism of this invention; Figure 3 This is a schematic diagram of the data structure of the evidence object of this invention; Figure 4 This is a schematic diagram of the conflict arbitration and mechanism consistency verification process of the present invention; Figure 5 This is a schematic diagram of the work order generation and retest acceptance closed loop of the present invention; Figure 6 This is a schematic diagram of the module architecture of the vibration diagnosis and evidence chain tracing system of the present invention. Detailed Implementation

[0021] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0022] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0023] Example 1: Overall Method Flow Reference Figure 1 The vibration diagnosis and evidence chain tracing method based on mechanism constraint semantic encoding described in this embodiment specifically includes the following steps: S1 Data Acquisition: The system acquires multi-channel vibration signals and key phase rotational speed signals from the target equipment, while simultaneously obtaining structural parameters and real-time operating parameters. The vibration signals include acceleration or velocity signals in the radial horizontal, radial vertical, and axial directions, with measurement points located at key locations such as bearing housings and the casing. The rotational speed signals are acquired by key phase sensors or speed transmitters and used for order analysis under unsteady conditions.

[0024] Structural parameters include inherent equipment parameters such as bearing type and fault characteristic frequency, number of gear teeth and meshing frequency, shaft diameter, number of blades, and rotor critical speed. Operating parameters include real-time speed, load rate, current, medium temperature, bearing temperature, and operating status (steady-state / acceleration / deceleration / start-stop).

[0025] S2 pretreatment and unsteady-state treatment: The original vibration signal undergoes conventional preprocessing such as detrending, digital filtering, and piecewise windowing. A spectrum is generated through Fourier transform, and an envelope spectrum is generated through Hilbert demodulation, yielding basic feature representations in the time, frequency, and envelope domains. When the equipment is detected to be in a non-steady-state condition (i.e., the rate of change of rotational speed is greater than 5 r / s²), a non-steady-state processing procedure is initiated: based on the key phase signal, the vibration signal sampled at equal intervals in the time domain is converted to angular domain samples at equal intervals, eliminating frequency smearing caused by rotational speed fluctuations. The interpolation method for angular domain resampling is adaptively selected according to the rate of change of rotational speed: linear interpolation is used when the rate of change of rotational speed is less than 2 r / s², cubic spline interpolation is used when the rate of change of rotational speed is 2~10 r / s², and sinc interpolation is used when the rate of change of rotational speed is greater than 10 r / s², balancing computational efficiency and resampling accuracy. After resampling, order spectra, envelope order spectra, and acceleration / deceleration transient characteristic curves are generated.

[0026] Optionally, a multi-granularity time-frequency attention auxiliary module can be added: multi-resolution time-frequency maps are generated through short-time Fourier transform and continuous wavelet transform, and a sparse attention mechanism is used to locate high-energy anomaly regions, dense sideband regions, and impact clusters in the time-frequency plane. Their location, bandwidth, and energy parameters are used as supplementary candidate sources for L1 basic quantity markers to improve the ability to capture weak fault symptoms.

[0027] S3 mechanism atom labeling extraction: Reference Figure 2 Based on the equipment structural parameters and the physical model of the failure frequency, hierarchical mechanism atomic markers are extracted from multi-domain features, and the markers are divided into three levels: L1, L2, and L3.

[0028] L1 basic quantities are low-level physical quantities that can be directly extracted from signals, including: amplitude-related quantities such as passband amplitude, amplitude of each order, 0.5X order amplitude, 1X order amplitude, 2X order amplitude, and bearing fault characteristic frequency amplitude; frequency and order-related quantities such as peak frequency and characteristic order; phase-related quantities such as fundamental frequency phase, phase of each order, and phase difference between channels; track-related quantities such as track major axis, track minor axis, track ellipticity, and centerline offset; as well as statistical quantities such as signal-to-noise ratio and number of impact pulses.

[0029] L2 structural relationship markers are structural features formed by the combination of multiple L1 markers, including: harmonic family features such as the number of harmonic chains and the harmonic amplitude attenuation law; sideband group features such as sideband spacing, sideband symmetry, and sideband energy ratio; modulation relationships such as amplitude modulation depth and frequency modulation degree; order coupling relationships such as sum frequency and difference frequency components; multi-channel correlation features such as inter-channel coherence coefficient, axial / radial energy ratio, and time delay difference; as well as track shape features and critical speed crossing characteristics, etc.

[0030] L3 fault semantic tags are high-level semantic labels corresponding to specific fault modes, including imbalance, misalignment, base looseness, dynamic-static friction, shaft bending, rolling bearing inner ring fault, rolling bearing outer ring fault, rolling element fault, gear pitting, gear tooth breakage, cavitation, oil film whirl, oil film oscillation, structural resonance, and subsynchronous instability. To reduce noise interference introduced by irrelevant tags, this embodiment sets an adaptive activation gating based on operating conditions, dynamically adjusting the participation weight of various tags according to the current operating conditions.

[0031] The activation weight of the r-th mechanism atom label at time t is calculated using the following formula: ; in, The value ranges from 0 to 1, with higher values ​​indicating higher diagnostic value under the current operating conditions. The current operating condition vector contains m-dimensional features, including rotational speed, rate of change of rotational speed, load rate, temperature, and steady-state characteristics. The embedding vector of the r-th class label is obtained by pre-training from historical fault samples, and represents the degree of adaptation of the label to different working conditions; and The weight vector and bias parameters of the gating network are obtained through training with labeled samples. When the activation weight is below 0.3, the quality score of the evidence object generated by the corresponding label is reduced by 50%; when the activation weight is below 0.1, the corresponding evidence object is put into a state of pending supplementation and does not participate in automatic conflict arbitration. For example, under steady-state low-speed conditions, the activation weights of oil film whirl and critical speed crossing related labels are reduced; under acceleration conditions, the activation weights of the subsynchronous component and forward / backward components are increased compared to the related labels.

[0032] S4 Evidence Object Construction: Reference Figure 3 The hierarchical mechanism atomic tags are bound to end-to-end attributes to form standardized evidence objects. Each evidence object contains the following core fields: source_id: Identifier of the original data source, associated with the original data file and the acquisition device; channel_id: Measurement point channel number, recording the physical location of the signal; window_id: Time window or corner window number, marking the time or corner range corresponding to the feature; transform_id: Transform link identifier, which records all processing steps from the original signal to this tag; mech_level: Mechanism level, with values ​​L1 / L2 / L3; mech_label: Specific mechanism semantic label; support_flag: Evidence attribute, with values ​​of support, conflict, and need to be supplemented; quality_score: Evidence quality score, with a value range of 0 to 1; `applicable_condition`: Applicable operating conditions range, recording the effective rotational speed and load range for this evidence. Optional extended fields include collection timestamp, evidence freshness, evidence source pattern, and confidence propagation path. The basic quality score of the evidence object is calculated using a multi-indicator weighted fusion formula: ; in, The baseline quality score for the i-th evidence object; The normalized score of the i-th piece of evidence on the k-th quality index, with a value of 0 to 1; The initial weight for the k-th indicator is set by expert experience. There are six quality indicators: signal-to-noise ratio, sensor health, mechanism relevance, operating condition matching, cross-channel consistency, and non-steady-state reliability. Considering the timeliness of evidence, a time-series decay factor is introduced to adjust the quality score, using the following formula: ; in, The time difference between the evidence collection time and the current diagnosis time, in hours; The attenuation coefficient has an adaptive adjustment formula as follows: ; In the formula The baseline attenuation coefficient is set to 0.01. , , The influence coefficients for rotational speed, load, and temperature change rate are respectively taken as 0.005, 0.003, and 0.002; The rate of change of rotational speed, For load change rate, This represents the rate of temperature change. The faster the operating conditions change, the larger the attenuation coefficient, and the lower the weight of earlier collected evidence. When the quality score after time-adjusted period is below 0.2, the evidence object is placed in a state of pending supplementation, used only for interpretation and display, and does not participate in automatic arbitration. When abnormal or missing measurement point signals are detected, the virtual sensor reconstruction mechanism is activated.

[0033] Based on historical normal data from adjacent channels, cross-spectral analysis is used to estimate the transfer coefficient and time delay parameters. A time delay transfer model is then employed to reconstruct the signal of the missing channel. The reconstruction formula is as follows: ; in, The signal of the measurement point to be reconstructed; The signal of the j-th adjacent associated channel; This is the amplitude transfer coefficient; This represents the signal propagation delay. The parameters are estimated using the cross-spectral phase difference and amplitude ratio under historical normal operating conditions. The reconstruction confidence level is calculated after reconstruction is complete. Based on residual confidence Coherence confidence Template matching confidence Order consistency confidence The weighted fusion was performed with weights of 0.4, 0.3, 0.2, and 0.1, respectively. The formula for calculating the residual confidence level is as follows: ; In the formula To reconstruct the mean square error between the signal and the true signal, The noise variance under healthy operating conditions. To prevent extremely small constants with a denominator of zero, the final quality score of the evidence object generated based on the reconstructed signal is adjusted as follows: ; in To minimize the retention factor, a value of 0.3 is chosen to ensure that the evidence for the reconstructed signal is not completely erased. When the reconstruction confidence level... At this time, all corresponding evidence objects are switched to a state of pending supplementation, the system switches to single-channel downgrade diagnosis mode, and only outputs coarse-grained fault categories such as rotor-type anomalies and bearing-type anomalies, without giving fine-grained component-level conclusions.

[0034] S5 Joint Encoding and Candidate Fault Generation: The hierarchical mechanism atomic tags are converted into pre-trained embedding vectors. The quality score, applicable working conditions, and location information of the evidence objects are converted into normalized feature vectors. The working condition parameters are converted into working condition feature vectors. These are concatenated in a preset order to form a fixed-dimensional diagnostic input sequence, completing joint encoding. The diagnostic input sequence is then fed into the diagnostic inference model, which employs a 3-layer Transformer encoder structure. The output layer is a multi-label classification layer, outputting the initial confidence score of each candidate fault hypothesis. The model automatically associates each candidate fault hypothesis with corresponding supporting and conflicting evidence objects and generates a globally unique evidence number, achieving a one-to-one binding between conclusions and evidence.

[0035] S6 Conflict Arbitration and Interpretation Output: Reference Figure 4 This step sequentially performs DS evidence fusion, identification matrix verification, and mechanism consistency check, and finally outputs the arbitration result.

[0036] Firstly, a preliminary fusion is performed based on the DS evidence theory: the identification framework is a set of all candidate fault hypotheses, and each evidence object is assigned a basic probability value for each candidate fault. The assignment rules are generated based on the degree of matching between the mechanistic semantics of the evidence and the failure mode, combined with the evidence quality score. The Dempster combination rule is used to fuse the basic probability assignments of all evidence to obtain the trust function for each candidate failure. Similarity function , seemingly true interval The width reflects the uncertainty of the diagnostic results.

[0037] The spectral feature discrimination matrix is ​​then used for verification. The rows of the discrimination matrix correspond to each candidate fault type, and the columns correspond to nine core discrimination features, including axial / radial amplitude ratio, inter-channel phase difference, 0.5X / 1X amplitude ratio, 1X / 2X amplitude ratio, sideband symmetry, multi-channel coherence coefficient, orbital ellipticity, forward / backward component ratio, and subsynchronous energy percentage. Each fault type corresponds to a fault determination interval for each feature. The normalized distance between the current feature value and the center of the corresponding fault interval is calculated to obtain the matrix verification score. The value ranges from 0 to 1. The final arbitration score is calculated using the following formula: ; in This is the fusion coefficient, with an initial value of 0.6, which can be adjusted online based on historical closed-loop results.

[0038] Candidate faults are sorted from highest to lowest based on their final arbitration scores. To achieve adaptive optimization of the identification boundary, this embodiment sets up an online self-calibration mechanism: a historical sample sliding window of size 100 is maintained for each identification feature of each fault type. After each repair acceptance and fault type confirmation, the corresponding feature value is written into the sliding window. The updated identification boundary is calculated using the quantile method. ; ; in, and These are the updated lower and upper boundaries of the j-th feature of the r-th fault, respectively. For the updated sliding window sample set; and These are the 5th percentile and 95th percentile, respectively. Through continuous iteration, the identification boundary can be gradually adapted to the operating characteristics of specific devices.

[0039] To suppress the hallucinatory output of large language models or diagnostic models, a mechanism consistency check is performed. The check rules include four items: First, whether the output fault semantics have corresponding L2 structural relation tags to support them; second, whether key diagnostic terms are associated with at least one valid evidence object number; third, whether the output fault order is consistent with the conflict arbitration result; and fourth, whether fault conclusions involving half-harmonics and subsynchronous faults have 0.5X order correlation tags to support them. If any item is not satisfied, the output result is downweighted, rejected, or regenerated.

[0040] Optionally, a causal link tracing module can be added: based on a pre-set fault causal directed graph, nodes correspond to L3 fault semantics, and edges correspond to the causal relationship and propagation weight between faults; the trust function value of each fault is used as the initial activation strength of the node, the activation value is propagated along the causal edge, and after convergence, the root cause node and the secondary node are distinguished according to the causal strength of the node, and the root cause fault is processed first when the work order is generated.

[0041] This embodiment sets up a three-level human-machine collaboration partition, based on the trust function value of the highest-scoring candidate fault and the width of the plausible interval. Division: Automatic confirmation area: Satisfied and The diagnostic results have high confidence and automatically generate repair work orders. Decision support area: meets ,or but It generates an interactive evidence summary, provides the optimal supplementary testing suggestion to eliminate ambiguity, and pushes it to the operation and maintenance experts for confirmation; Manual control area: meets the requirements If there is high conflicting evidence or low confidence in reconstruction, an automatic work order will be issued, triggering on-site retesting or expert review.

[0042] S7 Work Order Generation and Closed-Loop Update: Reference Figure 5 Based on the arbitration results and root cause analysis, a maintenance action tree is generated, and then a standardized work order is generated based on the maintenance action tree. The work order adopts a two-level field design: the core fields include equipment identifier, target maintenance action, associated evidence number, risk level, and retesting requirements; the extended fields include spare parts requirements, tool requirements, execution team, suggested time window, estimated duration, downtime impact level, safety tips, and references to similar historical cases.

[0043] Work order status is categorized into six types: automatically generated and pending execution, pending expert confirmation, pending supplementary testing, pending manual takeover, accepted and closed, and upgraded re-inspection. The initial status is automatically matched based on the diagnostic zoning results. After repair, the same acquisition parameters, measurement point locations, preprocessing procedures, and marker extraction methods as before repair are used to acquire post-repair vibration signals and generate post-repair evidence objects. Changes in key fault characteristic quantities before and after repair are compared: a decrease of over 70% in fault characteristic amplitude indicates a resolved issue, a decrease of 30% to 70% indicates a partially resolved issue, and a decrease of less than 30% indicates an unresolved issue. Based on the acceptance conclusion, the work order is closed, upgraded, or regenerated.

[0044] After acceptance, the weights of the evidence quality indicators are updated using a Bayesian method based on the acceptance results. The update formula is as follows: ; in, This represents the weight of the k-th quality indicator before the update. For the evidence corresponding to the k-th indicator in the acceptance results The likelihood probability is set to 0.9 when the evidence corresponds to a correct diagnosis and 0.2 when it corresponds to a wrong diagnosis. The updated weights are written back to the evidence scoring module to achieve continuous optimization of the scoring rules.

[0045] For multiple devices of the same model, a cluster baseline library is established, storing the statistical distribution of mechanism atomic markers, evidence quality distribution, typical fault semantic distribution, and key order amplitude distribution. When diagnosing a single device, it can be compared horizontally with the cluster baseline to identify individual anomalies and common deviations. High-confidence evidence that has been validated through closed-loop verification of multiple devices can be marked as reused evidence to improve the cold start diagnostic accuracy of new devices of the same model.

[0046] Example 2 System Architecture: Reference Figure 6 The vibration diagnosis and evidence chain tracing system based on mechanism constraint semantic coding described in this embodiment includes five core modules and underlying support components.

[0047] The data acquisition and preprocessing module interfaces with vibration sensors, speed sensors, and operating condition acquisition devices to complete the acquisition, preprocessing, and non-steady-state angular domain resampling of raw signals, and outputs standardized multi-domain features.

[0048] The mechanism labeling and evidence construction module extracts hierarchical mechanism atomic labels based on the fault physical model and equipment structural parameters, binds them to full-link attributes to generate evidence objects, and completes quality scoring and time-series decay calculation.

[0049] The joint encoding and diagnostic reasoning module completes the joint semantic encoding of multi-source information and outputs candidate fault hypotheses and related evidence through the diagnostic reasoning model.

[0050] The conflict arbitration and verification module performs DS evidence fusion, identification matrix review, and mechanism consistency verification, and outputs sorted diagnostic conclusions, evidence numbers, and explanatory notes, while managing human-machine collaboration partitioning logic.

[0051] The work order and closed-loop update module generates maintenance work orders, manages the entire lifecycle of work orders, performs post-repair retesting and acceptance, and completes weight updates, boundary calibration, and case library iteration.

[0052] The modules interact with each other through standardized mechanism atomic tag data packets and evidence object data packets. The data packet fields correspond one-to-one with the core fields of the evidence object, ensuring the consistency and traceability of data flow.

[0053] The underlying support components of the system include a historical case library, a cluster baseline library, an identification matrix library, and a migration adaptation layer, which provide data and algorithm support for the upper-layer modules.

[0054] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A vibration diagnosis and evidence chain tracing method based on mechanism-constrained semantic coding, characterized in that, Includes the following steps: Acquire vibration signals, rotational speed signals, structural parameters, and operating parameters of the target equipment; The vibration signal is preprocessed to generate time-domain, frequency-domain, order-domain, and / or envelope-domain feature representations; Based on the physical model of structural parameters, operating parameters and fault frequency, hierarchical mechanism atomic markers are extracted from the feature representation; By binding the hierarchical mechanism atomic tags with the original data source, measurement point channel, time window, transformation link, mechanism attribute, quality score, and applicable working conditions, a traceable evidence object is constructed. The hierarchical mechanism atomic markers, evidence objects, operating condition information and location information are jointly encoded to generate a diagnostic input sequence, output multiple candidate fault hypotheses and associate them with corresponding supporting evidence and conflicting evidence. Perform conflict arbitration and mechanism consistency verification on candidate fault hypotheses, and output the sorted diagnostic conclusions and traceable evidence numbers; A repair work order is generated based on the diagnostic conclusions. After the repair is completed, the evidence scoring weights and fault identification boundaries are updated based on the retest acceptance results.

2. The vibration diagnosis and evidence chain tracing method based on mechanism-constrained semantic encoding according to claim 1, characterized in that, When the target equipment is in acceleration, deceleration or speed fluctuation conditions, the vibration signal is subjected to angular domain resampling and order tracking processing to generate order spectrum and transient feature representation; Simultaneously, the activation weights of various mechanism atomic markers are calculated based on the current working condition vector, and the evidence objects corresponding to markers with activation weights lower than the preset threshold are either downweighted or converted to a pending supplementation state.

3. The vibration diagnosis and evidence chain tracing method based on mechanism-constrained semantic encoding according to claim 1, characterized in that, The layered mechanism atomic labeling includes a three-layer structure: L1 basic quantity markers are used to characterize at least one of amplitude, frequency, order, phase, and orbital geometry parameters; L2 structural relation markers are used to characterize at least one of harmonic families, sideband groups, modulation relations, order coupling, and multichannel coherence. L3 fault semantic markers are used to characterize at least one of the following: imbalance, misalignment, looseness, rolling bearing failure, oil film whirl, and structural resonance.

4. The vibration diagnosis and evidence chain tracing method based on mechanism-constrained semantic encoding according to claim 1, characterized in that, The quality score of the evidence object is calculated by weighting multiple indicators such as signal-to-noise ratio, sensor health, mechanism correlation, operating condition matching degree, and cross-channel consistency; and a time-series attenuation factor is introduced to adjust the quality score over time, with the attenuation coefficient increasing synchronously with the rate of change of rotation speed, load, and temperature.

5. The vibration diagnosis and evidence chain tracing method based on mechanism-constrained semantic encoding according to claim 1, characterized in that, The conflict arbitration adopts the DS evidence theory to fuse the basic probability assignments of each evidence object to obtain the trust function and the likelihood function of each candidate fault hypothesis; then, the candidate faults ranked first are reviewed by the spectrum feature identification matrix, and the final arbitration result is obtained by combining the trust function value and the review score; the judgment boundary of the spectrum feature identification matrix is ​​self-calibrated online by quantile statistics according to the closed-loop acceptance result.

6. The vibration diagnosis and evidence chain tracing method based on mechanism-constrained semantic encoding according to claim 1, characterized in that, When an abnormality or missing signal is detected at the measurement point, the virtual sensor signal is reconstructed based on the signal from the adjacent channel through a time delay propagation model. The reconstruction confidence is calculated and propagated to the quality score of the corresponding evidence object. When the reconstruction confidence is lower than a preset threshold, the system switches to a single-channel degradation diagnosis mode and outputs a coarse-grained fault category conclusion.

7. The vibration diagnosis and evidence chain tracing method based on mechanism-constrained semantic encoding according to claim 1, characterized in that, Based on the trust function value and the width of the plausible interval of the candidate fault, the diagnostic results are divided into an automatic confirmation area, an auxiliary decision-making area, and a manual takeover area. The automatic confirmation area directly generates a maintenance work order, the auxiliary decision-making area outputs an evidence summary and the best supplementary test suggestion, and the manual takeover area triggers on-site expert review.

8. The vibration diagnosis and evidence chain tracing method based on mechanism-constrained semantic coding according to claim 1, characterized in that, After the maintenance is completed, the same collection and processing procedures as before the maintenance are used to obtain the evidence objects after the maintenance. The changes in key feature quantities before and after the maintenance are compared to output the acceptance conclusion. Based on the acceptance results, the weights of the evidence quality indicators are updated using the Bayesian method, and the historical case library and the baseline of the same type of equipment group are updated simultaneously.

9. A vibration diagnosis and evidence chain tracing system based on mechanism-constrained semantic coding, employing the vibration diagnosis and evidence chain tracing method based on mechanism-constrained semantic coding as described in any one of claims 1-8, characterized in that, include: The data acquisition and preprocessing module is used to acquire the vibration, rotational speed, structural and operating parameters of the target equipment and generate multi-domain feature representations. The mechanism labeling and evidence construction module is used to extract hierarchical mechanism atomic labels and construct traceable evidence objects with quality scores; The joint coding and diagnostic reasoning module is used to generate unified semantic codes and output candidate fault hypotheses and related evidence; The conflict arbitration and verification module is used to perform conflict arbitration and mechanism consistency verification, and output diagnostic conclusions with evidence numbers. The work order and closed-loop update module is used to generate maintenance work orders and iteratively update the diagnostic model parameters based on the retest results; The modules interact with each other through mechanism atomic tag data packets and evidence object data packets.

Images