Nondestructive testing and performance degradation analysis method for bridge structure material

Abstract

The invention relates to the technical field of nondestructive testing and performance analysis of engineering materials, in particular to a nondestructive testing and performance degradation analysis method for bridge structure materials. The method comprises the following steps: acquiring material nondestructive testing monitoring data, and executing unified alignment, noise portraying and quality labeling; carrying out combined noise reduction and component evaluation, and generating a noise reduction parameter snapshot; extracting a marginal spectrum and characteristic energy ratio, and establishing a material characterization characteristic index dictionary; calculating a variation index, judging material performance degradation, and constructing a supervision signal; and training the prediction network, generating a trend and a residual error, and completing early warning fusion by adopting a time neighborhood consistency strategy to obtain a material performance degradation early warning level sequence. According to the method, continuous nondestructive monitoring and quantitative characterization of material levels are realized, characteristics and parameters are traceable, cross-channel and cross-time evaluation is consistent, prediction and early warning are sensitive and stable, and the method is suitable for long-term service safety evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of non-destructive testing and performance analysis of engineering materials, and particularly relates to a non-destructive testing and performance degradation analysis method for bridge structure materials. BACKGROUND

[0002] In the field of non-destructive testing and performance analysis of engineering materials, especially for long-term performance evolution monitoring of key materials in civil engineering (such as concrete and prestressed steel), traditional methods face severe challenges. For prestressed small box girder bridges in the operation stage, key performance indicators such as micro-crack evolution of concrete materials, stress redistribution, and stress loss of prestressed steel, currently mainly rely on periodic manual detection or destructive coring experiments. These methods not only have hysteresis, dispersion and destructiveness, but also are difficult to achieve early, continuous and quantitative identification of material performance degradation, and cannot depict the dynamic whole process of material intrinsic response under complex working conditions such as traffic load and environmental disturbance.

[0003] To overcome the above limitations, the industry has gradually adopted material-level non-destructive testing methods to obtain measurement data, including but not limited to ultrasonic / acoustic emission / infrared thermography / electromagnetic eddy current / electrochemical potential material detection channel signals, and can be supplemented by cross-verification of local micro-vibration and local strain response records. However, the multi-source heterogeneous data generated thereby generally has problems such as inconsistent time base, missing data, and saturated segments; at the same time, traffic load, environmental temperature and humidity disturbance, and equipment noise introduce non-stationary background and impulse interference in the material signal, significantly affecting the stable extraction of defect indicators and material characterization parameters. Common denoising procedures (such as single threshold wavelet or empirical mode decomposition) are prone to modal aliasing and boundary artifacts, and lack standardized snapshot records of denoising parameters, making it difficult to unify cross-channel and cross-period processing strategies, and difficult to support long-term consistency analysis of material performance.

[0004] At the feature extraction level, existing methods are mostly based on energy proportion and frequency band statistics, but there is a lack of unified correlation snapshot between frequency band division, window advancement, and material physical properties (such as ultrasonic phase / amplitude changes corresponding to concrete micro-crack evolution, eddy current amplitude and phase influenced by prestressed steel corrosion, and thermal diffusion parameters under thermography), making it difficult to establish a traceable mapping from material detection features to material performance parameters (such as damage factor, equivalent elastic modulus attenuation, prestress loss indicator, and corrosion potential) in the subsequent comparison stage.

[0005] Further, the indicators for determining the material state evolution mostly rely on empirical thresholds, whose reference windows often change with batches, and are not bound with signal processing parameters and feature extraction parameters, resulting in a lack of physical basis in the threshold generation and determination process, and insufficient sensitivity in identifying early damage and micro-defects of the material. In terms of trend prediction, although algorithms such as bidirectional long short-term memory networks have been tried for inferring the material performance degradation trend, the training samples and the supervision signals used to characterize the degradation are not uniform in origin, and the supervision labels and the original material detection features lack a backtracking link, which limits the model's ability to depict and generalize the material degradation mechanism. SUMMARY

[0006] To solve the above technical problems, the present application provides a non-destructive testing and performance degradation analysis method for bridge structural materials, comprising:

[0007] Obtaining material non-destructive testing monitoring data, performing unified alignment, noise image construction and quality annotation processing, and generating a missing mask; the monitoring data includes material detection channel data reflecting the mechanical properties of the material, material micro-vibration acceleration response time history data collected by acceleration sensors arranged in the local area of the base material, material local strain response time history data collected by strain sensors arranged in the local area of the base material or the sample position, and synchronous recorded working condition information and time marker information;

[0008] Performing joint noise reduction configuration, component evaluation and noise reduction operation of the material detection signal, and generating a noise reduction parameter snapshot;

[0009] Performing marginal spectrum calculation, feature energy ratio extraction and feature index registration processing, and generating a material characterization feature index dictionary;

[0010] Performing variation index calculation, material performance degradation determination and supervision signal construction operation, and generating a material performance characterization parameter set and a supervision signal set;

[0011] Performing network training, material performance degradation trend prediction and residual record processing, and generating a predicted residual sequence;

[0012] Obtaining a trend prediction sequence, a variation index sequence and a predicted residual sequence, constructing a fusion reading list and using a time neighborhood consistency reading strategy to perform material performance degradation warning fusion, and obtaining a material performance degradation warning level sequence.

[0013] Further, the material performance degradation warning fusion specifically includes:

[0014] Taking the trend prediction sequence as the main input, and taking the variation index sequence and the predicted residual sequence as parallel inputs, the material performance degradation warning fusion is generated in the time scale consistent with the unified time base data, and a generated material performance degradation warning level sequence corresponding to the channel identifier and the section identifier is output.

[0015] Further, the generating the material performance degradation early warning fusion comprises:

[0016] The fusion reading list is constructed, the current value of the trend prediction sequence, the current item of the variation index sequence and the current difference value of the prediction residual sequence are extracted at each valid time position, and the threshold parameters, the segment boundary splicing position and the channel weight item registered in the previous step are called to form a fusion item set; for the channel dimension in the fusion item set, parallel merging or weighted aggregation is performed according to the registered channel weight, so that a single fusion reading is generated at the same time position in the cross-channel scene.

[0017] Further, the generating the material performance degradation early warning fusion further comprises:

[0018] The time neighborhood consistency reading strategy is established, and the fusion readings in the continuous time window are sequentially scanned; when the instantaneous fluctuation caused by the segment boundary, the missing mask or the window advance occurs, the neighborhood consistency rule is called for smoothing processing, so that the fusion readings near the boundary are kept consistent with the time slices on both sides.

[0019] Further, the material performance degradation early warning level sequence is obtained, which further comprises:

[0020] The material performance degradation early warning level sequence is generated as a driving input, and the noise reduction parameter snapshot is configured as a basis, a simply supported beam calculation model and a loading process consistent with the uniform time base data are constructed on a finite element simulation platform, and a simulation sample library corresponding to the channel identifier and the segment identifier is output.

[0021] Further, the simulation output is mapped according to the frequency band division scheme of the noise reduction parameter snapshot, and is windowed and sliced under the time advance consistent with the window advance mode, to form a time domain segment which can be directly input to the previous process; at the same time, each simulation segment is bound with the generated material performance degradation early warning label item, to generate a sample unit containing time scale, channel identifier, segment identifier, frequency band mapping and source label.

[0022] Further, the simulation sample library is taken as data input, and the noise reduction parameter snapshot is taken as parameter input, the parameter update and retraining are simultaneously carried out for the joint noise reduction and component credibility evaluation and the attention guided bidirectional long short-term recurrent network, and the updated process configuration is output.

[0023] Further, the new noise reduction parameter snapshot and the new network parameter are written into the process configuration dictionary, and the parameter backtracking pointer and the coverage range are recorded in the configuration dictionary; when there is a historical time period or channel that cannot be covered, the old version item is retained in the configuration dictionary and the effective interval is marked, to support phased calling.

[0024] Further, according to the loading list and the consistent mapping relationship, the boundary conditions and the action path of the simply supported beam model are configured, simulation solving is run in a time window corresponding to the generated material performance degradation early warning label entry, and multi-channel response data consistent with the high signal-to-noise ratio sequence is obtained; for the time position containing the missing mask space, simulation data is not generated, and the time index and the space reason are retained to avoid introducing inconsistent supervision information in subsequent parameter updating and retraining.

[0025] Further, in the retraining cycle, the slicing organization mode of the material performance characterization parameter set and the supervision signal set and the attention guide weight generation rule are continued to be used, the new simulation samples and the existing measured samples are optimized under the same specification, and the network parameters are optimized; when the network reaches the predetermined convergence criterion on the new samples, the new network parameters, the input specification and the weight generation rule are solidified, and a closed loop link from the generation of the material performance degradation early warning grade sequence to the simulation sample library to the updated process configuration is formed.

[0026] The following are its main beneficial effects:

[0027] (1) A material-level non-destructive and online testing and analysis closed loop for material performance degradation is constructed. The present application takes ultrasonic, acoustic emission, infrared thermal image, eddy current, electrochemical potential and material local micro-vibration / local strain as the core input, forms a complete link from time reference alignment, joint noise reduction to material degradation early warning, realizes long-term, continuous and quantitative monitoring of key engineering materials such as concrete and prestressed steel, and overcomes the hysteresis and destructiveness of coring and offline detection.

[0028] (2) A traceable material characterization system closely related to material physical properties is established. The marginal spectrum and characteristic energy ratio are extracted according to the noise reduction parameter snapshot, and the analysis parameters such as frequency band division, window advancement are bound to the material mechanism quantity (such as ultrasonic echo phase / amplitude change, eddy current amplitude and phase, thermal diffusivity parameter, corrosion potential) through a material characterization feature index dictionary, forming a structured mapping of features-characterization parameters, and ensuring stable and traceable characterization of features to material state changes.

[0029] (3) The consistency of the material performance degradation index is quantitatively evaluated. Under the constraint of unified parameter snapshot and index dictionary, the variation index and defect indicator are calculated, the influence of batch difference and working condition disturbance on the reference benchmark is weakened, the consistency comparison across time and channels is supported, and the complex material signal is converted into sensitive and robust measurement of micro-crack initiation, prestress loss and corrosion process.

[0030] (4) The data-driven material degradation trend prediction and early warning capability is improved. The material degradation determination result and threshold parameter are converted into a supervised signal to train an attention-guided time sequence network. The long-term degradation trend is inferred while keeping the material detection means unchanged. The short-term anomaly is captured through residual analysis to provide multi-evidence fusion early warning support for material service safety.

[0031] (5) A self-verification closed loop of material parameter inversion-model calibration is formed to ensure the accuracy of long-term analysis. The application is based on the material characterization parameter set to perform material parameter inversion modeling (optional finite element or analytical inversion) to obtain calibrated material performance evaluation results, and accordingly update the noise reduction and prediction model parameters to realize continuous self-calibration and rolling optimization for material level, and ensure the reliability and sustainability of long-term evaluation results. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 A flowchart of a nondestructive testing and performance degradation analysis method for bridge structure materials is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0033] The table is intended to clarify how the method of the present application intelligently selects and combines nondestructive testing channels according to different bridge structure materials, ensuring the relevance and effectiveness of the method.

[0034] Table 1: Material method comparison table

[0035] Parameter templating and adaptive selection: The application has pre-set acquisition parameter templates (such as incident method, excitation frequency band, sampling rhythm, etc.) bound to the material types and detection methods in the above table in the specific implementation part. The system will automatically select at least two core detection channels according to the automatically identified material type label, and load the corresponding parameter templates, thereby realizing precise and adaptive matching from materials to methods.

[0036] Consistency of closed loop process: Regardless of the material and the combination of detection methods, the data generated is connected to the unified closed loop process (S100-S600) proposed in the present application, including unified alignment, joint noise reduction, feature extraction, degradation determination, trend prediction and early warning fusion. The consistency of the core processing logic and evaluation standards of the method is ensured when the method is applied across materials.

[0037] Bridge component materials: steel, steel cable, steel box girder, steel-concrete composite joint; reinforced concrete, prestressed concrete, box girder / T girder, pier column and pile cap; composite material and FRP reinforcement layer (sheet, cloth, plate). Material type label is used to indicate the above material categories as the index key of subsequent channel selection, parameter template, feature dictionary and degradation model.

[0038] Steel / steel cable / steel box girder: ultrasonic phased array, magnetic leakage, eddy current, acoustic emission, infrared thermal imaging; pay attention to cracks, corrosion thinning, inclusions, weld defects.

[0039] Reinforced concrete / prestressed: ultrasonic pulse, impact echo, ground penetrating radar, infrared thermal imaging, half potential; pay attention to hollowing, cracks, steel corrosion, pore defects.

[0040] Composite materials / FRP reinforcement: infrared thermal imaging, ultrasonic guided wave / surface scanning, laser ultrasonic; pay attention to debonding, delamination, inclusions.

[0041] Parameter template points (automatically brought out with materials): incident mode, excitation frequency band and step, coupling / isolation mode, sampling rhythm, thermal excitation / cooling timing, etc.

[0042] Material-channel selection rules:

[0043] a) Identify the material type label;

[0044] b) Automatically select at least two detection channels according to the principle of complementarity (penetration ability complementarity / sensitive mechanism complementarity / environmental interference complementarity);

[0045] c) Load the acquisition and excitation parameter template bound with the material+channel;

[0046] d) Perform unified calibration and space-time registration;

[0047] e) Call the feature dictionary corresponding to the material, and enter the subsequent fusion and degradation evaluation.

[0048] Example one: Steel box girder weld and plate corrosion screening

[0049] Method selection: ultrasonic phased array + magnetic leakage (add infrared thermal imaging if necessary);

[0050] Parameter points: coupling medium and array angle, magnetization intensity and scanning step, thermal imaging observation window;

[0051] Pay attention to features: echo amplitude attenuation / side wall reflection, magnetic flux distortion profile, thermal anomaly boundary;

[0052] Output: crack / inclusion location and length estimation, corrosion area thinning range, degradation level and disposal suggestion.

[0053] Example two: prestressed concrete box beam hole defect / corrosion evaluation

[0054] Selection method: ground penetrating radar + impact echo (if necessary, add half potential);

[0055] Parameter points: antenna arrangement and wiring, impact energy and incident point, reference electrode layout;

[0056] Focus features: reflection interface and time delay change, frequency dispersion characteristics and peak shift, potential field gradient;

[0057] Output: hole hollowing / grouting defect positioning, corrosion risk zoning, degradation level and disposal suggestion.

[0058] Example three: FRP reinforcement layer debonding rapid investigation

[0059] Selection method: infrared thermography + ultrasonic guided wave / surface scanning;

[0060] Parameter points: thermal excitation / cooling rhythm, guided wave frequency band and path planning, surface scanning coverage;

[0061] Focus features: thermal anomaly boundary and time delay, guided wave energy dissipation ratio and echo attenuation map;

[0062] Output: debonding range and depth estimation, delamination tendency, degradation level and disposal suggestion.

[0063] Example four, in the field of nondestructive testing and performance analysis technology of engineering materials, according to Figure 1 It is a flowchart of a nondestructive testing and performance degradation analysis method for bridge structure materials provided by the embodiment of the present application, which can at least include steps S100-S600:

[0064] S100, obtain material nondestructive testing monitoring data, perform unified alignment, noise image construction and quality labeling processing, and generate missing mask;

[0065] S200, perform joint noise reduction configuration, component evaluation and noise reduction operation of material detection signal, and generate noise reduction parameter snapshot;

[0066] S300, perform marginal spectrum calculation, feature energy ratio extraction and feature index registration processing, and generate material characterization feature index dictionary;

[0067] S400, perform variation index calculation, material performance degradation determination and supervision signal construction operation, and generate material performance characterization parameter set and supervision signal set;

[0068] S500, perform network training, material performance degradation trend prediction and residual record processing, and generate prediction residual sequence;

[0069] S600, acquire the trend prediction sequence, the variation index sequence and the prediction residual sequence, construct a fusion read list and adopt a time neighborhood consistency reading strategy to generate a material performance degradation early warning fusion to obtain a material performance degradation early warning level sequence.

[0070] The step S100 at least comprises steps S110-S130:

[0071] S110, acquire monitoring data, perform unified alignment and obtain unified time base data;

[0072] The monitoring data specifically comprises original data acquired from each sensor channel of the bridge existing monitoring system, mainly including material detection channel data for reflecting material mechanical properties, material micro-vibration acceleration response time history data collected by acceleration sensors arranged in local areas of the parent material, material local strain response time history data collected by strain sensors arranged in local areas of the parent material or sample positions, and synchronous recorded working condition information and time marker information; wherein the material local strain response time history data is used to inverse the stress state of the concrete material, the working condition information covers traffic load state, environmental temperature and humidity change and special event record, and the time marker information includes the start and end time of each channel data collection, sampling frequency and absolute time stamp corresponding to each data point; these multi-channel, multi-physical quantity monitoring data serve as the original input of the method and provide a data basis for subsequent unified alignment, noise image construction and quality labeling. The bridge serves as a platform for providing load and environment, and the concrete box girder and the prestressed steel tendon serve as the material samples to be tested and analyzed.

[0073] Specifically, the present step takes the output of the existing monitoring system of the prestressed small box girder bridge as input, specifically, selects the original monitoring data of the acceleration channel and the strain channel of the midspan and the key section, and synchronously reads the working condition record and the time marker information. First, according to the time marker information of the monitoring system, the channel data of different sampling frequencies and different collection start and end points are aligned to establish a consistent time scale and a unified calculation point across channels. In the alignment process, the time scale sequence of each channel is first reconstructed by difference and interval checked, and abnormal repeated time markers and reverse time markers are eliminated to ensure the continuity of the time scale. Secondly, under the premise of not changing the original effective waveform form, the window resampling and channel alignment strategy is adopted to unify the channel data to the same sampling interval, and smooth splicing is implemented at the window boundary to avoid boundary artifacts to the subsequent frequency band analysis. Thirdly, combined with the working condition record, the day and night period, the traffic load change period and the environmental disturbance period are labeled to form a time period index set that can be referenced by subsequent steps, which is used to call in subsequent construction of noise image parameters and quality labeling. Subsequently, the aligned channel data is subjected to amplitude range verification and sensor range verification, and the out-of-range segments and saturated segments are identified and labeled on the time scale for uniform processing in subsequent quality labeling and missing data identification. After alignment, the unified time base data is generated, and the unified time base data is ensured to retain the channel identification, measurement unit, installation position and direction information of the original channel, which facilitates the configuration and selection of component sets in the subsequent joint denoising and component credibility evaluation steps. To ensure that the unified time base data can be directly used by subsequent steps, the present step generates an alignment log and a time period index set while outputting the unified time base data, and passes the unified time base data, the alignment log and the time period index set to the step of constructing noise image from the unified time base data for use, and uses them as one of the inputs in the joint denoising and component credibility evaluation steps. Thus, the input of the present step is the original monitoring data and the time marker information, and the output is the unified time base data and the time period index set, and the unified time base data is directly input to the steps of constructing noise image from the unified time base data and labeling quality and identifying missing data of the unified time base data, and is also a necessary input for configuring component sets in the joint denoising and component credibility evaluation steps.

[0074] S120, constructing a noise image to obtain noise image parameters;

[0075] The step takes the unified time base data and the time period index set as input. Specifically, the unified time base data is statistically characterized in multiple time periods, multiple frequency bands, and multiple channels around the background disturbance of the bridge under different environmental and load conditions to form noise image parameters that can be directly referenced by subsequent joint noise reduction and component credibility evaluation. The bridge serves as a platform for providing load and environment, while the concrete box girder and prestressed steel tendon are used as material samples for testing and analysis. First, based on the time period index set, the reference period with weak load and environmental disturbance is extracted, the baseline distribution, mutation probability, and pulse disturbance proportion of each channel in the reference period are calculated, and the background level curve of the time period-channel is formed on the unified time scale. At the same time, for the traffic load concentration period and the environmental disturbance significant period, the energy concentration interval and the short-time amplitude fluctuation interval of each channel are calculated to characterize the non-stationary background. Second, the multi-scale frequency band energy distribution and time-frequency texture features are extracted from the unified time base data, and the mapping relationship of frequency band division and energy proportion is established for each channel to determine the energy proportion, dominant frequency band range, interference frequency band range, and common interference interval across channels in different frequency bands, which serves as the basis for subsequent component set screening and improved threshold scheduling. Third, combined with the alignment log, the out-of-range segments and saturated segments registered during the alignment process are expanded and supplemented in the time period to improve the interference event list and mutation position index, so as to avoid directly including interference events into the component set as a reliable basis during joint noise reduction configuration. Subsequently, the baseline distribution, energy proportion, dominant frequency band range, interference frequency band range, mutation position index, and time period-channel background level curve are structured and summarized to generate noise image parameters. The noise image parameters clearly record the channel identification, time period type, frequency band division, and corresponding energy proportion, and are organized in a unified time scale to ensure one-to-one correspondence with the unified time base data. Finally, the noise image parameters and the unified time base data are transmitted to the joint noise reduction and component credibility evaluation step to set the number, center frequency band, and penalty strategy of the component set during joint noise reduction configuration, and serve as the prior source for subsequent weighting of marginal spectrum and frequency band merging of feature energy ratio vectors. At the same time, the noise image parameters are registered in the record table for parameter snapshot recording of high signal-to-noise ratio sequences to facilitate review and traceback during the generation of material performance degradation warning fusion and finite element for generating material benchmark response data under specific working conditions. The input of this step is the unified time base data and the time period index set, and the output is the noise image parameters, which can be referenced for quality labeling and missing data identification of the unified time base data to distinguish between interruptions caused by background disturbance and real missing data during missing data identification.

[0076] S130, quality labeling and missing data identification are performed to generate a missing data mask.

[0077] The above-mentioned unified time base data and noise image parameters are input, specifically, for the consistency, continuity and integrity of cross-channel data, the quality annotation and missing identification process is constructed, and the missing identification result is formed into a missing mask in a consistent manner with the unified time scale. First, according to the channel identifier and time scale of the unified time base data, the data integrity of each channel is scanned, the long interval breakpoint, short interval breakpoint and intermittent missing are identified, and the missing position and duration identified are registered as a to-be-confirmed list. In order to avoid misjudgment of short-time abnormality caused by background disturbance as missing, further, the interference event list and mutation position index in the noise image parameter are called to cross-check the to-be-confirmed list, if the breakpoint is located in the time range covered by the interference event list, it is marked as interference-related abnormality; if the breakpoint is not located in the range and meets the minimum duration and cross-channel consistency conditions, it is marked as real missing. Secondly, the amplitude range verification and adjacent window continuity verification are performed on the unified time base data, the out-of-range segments, saturated segments and discontinuous splicing boundary segments are uniformly marked, and the segment extension is performed on the consistent time scale with the time period index set to form the quality annotation item which can be called by the downstream. Thirdly, according to the real missing and interference-related abnormality marks, the missing mask is generated; the missing mask takes the unified time scale as the index, takes the channel as the dimension, uses the segment marking method to record the effective segment which can be used for joint noise reduction configuration and the segment which needs to be masked, ensuring that the subsequent joint noise reduction and component reliability evaluation steps can directly call the missing mask when configuring the component set and solving the parameters, avoiding unnecessary processing in the missing and saturated segments. Subsequently, the quality annotation item and the missing mask are attached to the unified time base data to form the attached information which can be referenced, and are archived together with the noise image parameter, so as to be used for window selection and frequency band weighting when calculating the marginal spectrum based on the high signal-to-noise ratio sequence and the noise reduction parameter snapshot, for masking the missing segment when calculating the variation index sequence from the feature energy ratio vector and the material characterization feature index dictionary, and for sample screening, weight distribution and residual registration when training the attention-guided bidirectional long short-term recurrent network according to the high signal-to-noise ratio sequence, the material performance characterization parameter set and the supervised signal set. Finally, the missing mask and the quality annotation are transmitted to the joint noise reduction and component reliability evaluation step through the data interface, which is used as the input item for joint noise reduction configuration, ensuring that the step skips the missing segment when setting the component set, performs weight reduction processing on the low-quality segment when generating the high signal-to-noise ratio sequence, and synchronously records the masking segment information when recording the parameter snapshot of the high signal-to-noise ratio sequence, to support the subsequent generation of material performance degradation warning fusion and finite element for generating the benchmark response data of the material under specific working conditions.Thus, the input of this step is the unified time-based data and the noise image parameters, and the output is the missing data mask and the quality label; the missing data mask is a necessary input for joint noise reduction and component reliability evaluation, and the quality label is a reference input for subsequent calculation of marginal spectrum, extraction of feature energy ratio vector, calculation of variation index sequence and training of attention-guided bidirectional long short-term recurrent network, which ensures that the technical path from aligned monitoring data to noise image parameters to quality label and missing data recognition is continuously called and reused in subsequent links, forming a closed loop through data acquisition, feature construction, identification and trend prediction.

[0078] In summary, S110, S120 and S130 complete the alignment of unified time-based data, the construction of noise image parameters and the generation of missing data mask, and pass the unified time-based data, noise image parameters and missing data mask as explicit inputs for subsequent steps, so that joint noise reduction and component reliability evaluation can be directly referenced when configuring component set and generating high signal-to-noise ratio sequence, so that the calculation of marginal spectrum and the extraction of feature energy ratio vector based on high signal-to-noise ratio sequence and noise reduction parameter snapshot can be consistently processed when window selection and frequency band weighting are performed, and the calculation of variation index sequence from feature energy ratio vector and material characterization feature index dictionary and the training of attention-guided bidirectional long short-term recurrent network from high signal-to-noise ratio sequence and material performance characterization parameter set and supervised signal set can maintain consistency when sample screening and weight allocation are performed, thereby forming a traceable, reviewable and backfillable execution basis in the noise reduction, identification and prediction process of the present application.

[0079] Step S200 includes at least steps S210-S230:

[0080] S210, based on the missing data mask, performing joint noise reduction configuration of the material detection signal to obtain a component set;

[0081] The step is based on the input of the unified time base data, and synchronously reads the noise image parameters and the missing mask, and carries out the joint noise reduction configuration of the material detection signal for the cross-channel time consistency and frequency band consistency, to generate a component set for subsequent processing. Specifically, first, in the time scale consistent with the unified time base data, the missing sections, saturated sections and unusable sections of each channel are marked and shielded according to the missing mask, and an available section index set is constructed; on this basis, the available section index set is matched with the background level, dominant frequency band range and interference frequency band range in the noise image parameters, to form a section-level frequency band constraint list, so that the subsequent decomposition and reconstruction only occur in the available section and follow the frequency band constraint. Further, according to the energy proportion and mutation position index in the noise image parameters, the number of components, center frequency band and penalty strength at the channel level and section level are determined, and a component candidate rule for distinguishing structural response and background disturbance is established; the above candidate rule includes preferential allocation to the dominant frequency band, suppression allocation to the interference frequency band and trade-off allocation to the overlapping frequency band, while giving the splicing strategy and window advancing strategy of the section boundary, to ensure the time domain continuity of the reconstruction process at the section splicing place. Subsequently, the channel identifier, section identifier, component number, center frequency band, penalty strength, window length, step interval and section boundary splicing strategy are summarized as a configuration record, and a component set is generated in the time scale consistent with the unified time base data; the component set lists the frequency band position, energy assignment and reconstruction priority of each candidate component in the available section, for component credibility evaluation of the component set and joint noise reduction of the material detection signal. Understandably, the input of the step is the unified time base data, noise image parameters and missing mask, and the output is the component set and configuration record; the component set is directly passed to the component credibility evaluation of the component set and joint noise reduction of the material detection signal, to generate a high signal-to-noise ratio sequence in the subsequent; the configuration record is referenced when the parameter snapshot record of the high signal-to-noise ratio sequence is performed in the subsequent, to maintain the consistency across steps, and is used as the source of frequency band constraint when the marginal spectrum is calculated based on the high signal-to-noise ratio sequence and the noise reduction parameter snapshot in the subsequent, so as to maintain the consistent connection of terms and data structures with the sequentially performed marginal spectrum calculation, feature energy ratio extraction, variation index calculation and trend prediction.

[0082] S220, component credibility evaluation and joint noise reduction of the component set, to generate a high signal-to-noise ratio sequence;

[0083] The step takes the component set as input, and carries out joint noise reduction processing of component credibility evaluation and material detection signal in the available segment corresponding to the unified time base data, to obtain a high signal-to-noise ratio sequence directly usable for subsequent feature construction. Specifically, first, according to the given channel identifier, segment identifier and frequency band position in the component set, the decomposition and initial reconstruction of each candidate component are performed one by one, and the energy distribution and time continuity metric are constructed within the segment to characterize the representativeness of the candidate component to the structural response; at the same time, combined with the background level and interference frequency band range in the noise image parameter, the energy concentration, frequency band matching and cross-channel consistency of the candidate component are comprehensively evaluated to form the component credibility evaluation value of this step, and the priority is sorted and screened out between the candidate components in the same frequency band or adjacent frequency band. Further, the joint noise reduction processing of the material detection signal is implemented on the candidate components selected by screening: within the segment, a time-advancing threshold scheduling strategy is adopted, so that the threshold is adaptively adjusted according to the background level and mutation position index of the noise image parameter; at the segment boundary, a splicing transition strategy is adopted to keep the continuity of amplitude and phase of the reconstructed signal at the window junction; between channels, a weight fusion strategy is adopted to make the available segments of different channels have consistent noise reduction strength and relative weight under the same time scale. Again, the missing mask is called to force mask the unavailable segments, so that the joint noise reduction of the material detection signal is only carried out in the available segments, avoiding the introduction of abnormal components due to missing or saturation; at the same time, the time index of the masked segment is reserved for synchronization with the segment splicing strategy when recording the subsequent parameters. Subsequently, the components after screening and joint noise reduction are weighted and reconstructed, and connected at the cross-segment according to the splicing transition strategy, to obtain a high signal-to-noise ratio sequence with the same time scale as the unified time base data; in this process, the component credibility evaluation value, threshold scheduling trajectory, segment splicing position and channel weight are arranged together as the noise reduction process record for subsequent record parameter and review call. Understandably, the input of this step is the component set, the noise image parameter and the missing mask, and the output is the high signal-to-noise ratio sequence and the noise reduction process record; the high signal-to-noise ratio sequence is directly transmitted to the parameter snapshot record of the high signal-to-noise ratio sequence to complete the arrangement of the noise reduction parameter snapshot, and used as the only time domain input when calculating the marginal spectrum based on the high signal-to-noise ratio sequence and the noise reduction parameter snapshot; the noise reduction process record is used for review of segment, frequency band and weight setting when generating material performance degradation warning fusion and finite element for generating benchmark response data of the material under specific working conditions, and the contrast information formed by recording the residual of the trend prediction sequence is used for parameter update and retraining.

[0084] S230, parameter snapshot record of the high signal-to-noise ratio sequence is performed to generate the noise reduction parameter snapshot;

[0085] The present step takes high signal-to-noise ratio sequence as input, and combines the foregoing configuration record and noise reduction process record to structure the key parameters involved in the joint noise reduction link of the material detection signal. A noise reduction parameter snapshot consistent with the time scale, channel identification and section identification is generated. Specifically, first, in the time scale consistent with the unified time-based data, the number of components, center frequency band, penalty strength, threshold scheduling track, channel weight, window length, step interval, section boundary splicing position and masking index involved in the foregoing configuration record and noise reduction process record are collected section by section, and a parameter index table is established in units of channels, so that any time position can be located to the corresponding parameter entry. Further, the parameter entries within the section are sequentially organized, so that the splicing position between the sections can be traced at the parameter level, and the masking index outside the section and the available section index are stored synchronously to ensure that the masking relationship and the reconstructed boundary can be accurately restored during review and retraining. Thirdly, to support subsequent calculation of marginal spectrum and extraction of feature energy ratio vector based on high signal-to-noise ratio sequence and noise reduction parameter snapshot, the correspondence between the frequency band division scheme and the window advancing scheme within the section is explicitly recorded in the parameter snapshot, and the source of the component reliability evaluation value and the reconstruction weight is marked in the channel index table, so that the frequency band weighting and window selection can be reproduced in the subsequent steps. Subsequently, the noise reduction parameter snapshot and the high signal-to-noise ratio sequence are bound to form a data and parameter pair that can be played back; the binding relationship includes time scale mapping, channel mapping and section mapping, which is used to directly read the frequency band division and window advancing scheme when calculating the marginal spectrum based on the high signal-to-noise ratio sequence and the noise reduction parameter snapshot, and is referenced as masking basis and weighting basis when calculating the variation index sequence from the feature energy ratio vector and the material characterization feature index dictionary. Finally, the noise reduction parameter snapshot and the foregoing binding relationship are registered to the parameter entry of the generated material performance degradation warning fusion and finite element for generating the benchmark response data loop of the material under specific working conditions, so that when generating samples and reviewing parameters according to the generated material performance degradation warning level sequence, the number of components, center frequency band and threshold scheduling can be updated consistently, and the updated results can be returned to the joint noise reduction configuration and component reliability evaluation link for retraining and reconstruction. Understandably, the input of the present step is high signal-to-noise ratio sequence, configuration record and noise reduction process record, and the output is noise reduction parameter snapshot and data parameter binding relationship; the noise reduction parameter snapshot and the high signal-to-noise ratio sequence are directly input to the subsequent marginal spectrum calculation and feature energy ratio extraction, and are also the review input of the generated material performance degradation warning fusion and finite element for generating the benchmark response data loop of the material under specific working conditions, and participate in parameter updating and retraining together with the residual record of trend prediction.In summary, the present step starts from the joint noise reduction configuration of the unified time base data, noise image parameters and missing mask, and finally forms a replayable parameter and data pair when recording the parameter snapshot of the high signal-to-noise ratio sequence through component reliability evaluation and joint noise reduction of the component set, realizes the integrated archiving of time scale, frequency band division, section splicing and masking relationship, and further provides consistent input and traceable parameter basis for subsequent feature construction, identification and trend prediction. The above steps realize the segmented constraint configuration, reliability driven noise reduction and parameter level binding of the unified time base data as a whole, so that the high signal-to-noise ratio sequence and the noise reduction parameter snapshot remain consistent and reproducible in the subsequent calling.

[0086] The step S300 at least includes steps S310-S330:

[0087] S310, based on the noise reduction parameter snapshot, marginal spectrum calculation is performed to obtain marginal spectrum data;

[0088] The step takes high signal-to-noise ratio sequence as the main input, and carries out marginal spectrum calculation based on the configuration of the noise reduction parameter snapshot, and outputs the marginal spectrum data with the same scale as the unified time base data. Specifically, first, in the time scale consistent with the unified time base data, read the channel identifier, section identifier, frequency band division scheme, window advancing scheme, reconstruction weight, masking index and section boundary splicing position in the noise reduction parameter snapshot, and perform section positioning and windowing processing on the high signal-to-noise ratio sequence. The windowing processing follows the window advancing scheme to ensure that the adjacent windows are continuously covered in time domain and compatible with the section boundary splicing position, and at the same time, the unusable section is shielded as invalid samples according to the masking index, to avoid frequency domain estimation at the missing and saturated positions. Subsequently, in each valid window, the frequency band division scheme and the reconstruction weight in the noise reduction parameter snapshot are called to implement the time-frequency analysis process, to obtain the local phase quantity and the local energy quantity, and to accumulate the energy of the frequency band point by point with the window as the granularity, so that the energy estimates of different windows can be spliced together in the unified time scale. In the process, the energy contribution of different components is weighted by the reconstruction weight to ensure consistency with the formation source of the high signal-to-noise ratio sequence. Further, the continuity of the energy track across the windows is corrected: when the window boundary coincides with the section boundary, a gradual transition strategy is adopted according to the section boundary splicing position to smooth the energy difference of adjacent windows at the boundary; when the window crosses the section marked by the masking index, the time domain position of the section is recorded as an energy gap, and the energy drop points are marked on both sides of the boundary for subsequent feature energy ratio extraction. Thirdly, for multi-channel data, an energy matrix aligned with the channels is established in the same time scale, the time dimension and the frequency band dimension of the matrix are uniformly constrained by the noise reduction parameter snapshot, and the channel dimension is assigned by the channel identifier; in the multi-channel merging scene, the energy is merged between channels according to the reconstruction weight to form a cross-channel consistent energy representation, and spliced into a continuous marginal spectrum data sequence in time sequence. Finally, the above energy matrix and its time scale, frequency band index, channel identifier and section identifier are uniformly packaged as marginal spectrum data, and the binding relationship with the noise reduction parameter snapshot is written in the data header to ensure that the frequency band division scheme and the window advancing scheme can be directly called for subsequent feature energy ratio extraction of the marginal spectrum data; at the same time, to maintain consistency with the subsequent calculation of the variation index sequence from the feature energy ratio vector and the material characterization feature index dictionary, the masking index is simultaneously registered in the marginal spectrum data to maintain consistent exclusion rules in the missing section for subsequent calculation. Through the above processing, the marginal spectrum data, as the output of the step, is transmitted to the feature energy ratio extraction of the marginal spectrum data, and is also one of the review bases for the subsequent generation of material performance degradation warning fusion and finite element for generating the benchmark response data of the material under a specific working condition, and together with the aforementioned noise reduction parameter snapshot, it forms a replayable time-frequency parameter pair.

[0089] S320, extracting feature energy ratio from marginal spectrum data to generate feature energy ratio vector;

[0090] This step takes marginal spectrum data as input and, under the constraints of the frequency band division scheme and window advancement scheme specified in the noise reduction parameter snapshot, extracts feature energy ratio vectors that reflect the energy distribution structure, while maintaining a time index consistent with the unified time base data. Specifically, based on the frequency band index in the marginal spectrum data, the energy values ​​at each time position are first aggregated into a frequency band energy list by frequency band, and the reconstruction weights in the noise reduction parameter snapshot are used to weight the frequency band energy list to obtain a frequency band energy sequence with consistent weights. When a time position hits the masking index, that position is marked as an invalid sample and excluded from subsequent extraction. Subsequently, within each valid time position, ratio calculations and merging operations are performed according to the order of the frequency band energy list to form a set of feature energy ratio entries with a fixed order. To ensure consistency between different channels and different segments, this step packages the set of feature energy ratio entries together with the channel identifier, segment identifier, and frequency band index to form a feature recording unit at the time position. Furthermore, to enhance the stability of feature recording units during cross-channel processing, this step employs a neighborhood consistency correction strategy for the temporal location of short-term energy anomalies: provided that adjacent time locations are continuously valid, the relative order of feature energy ratio entries is compared within the temporal neighborhood. If a single-point mutation is found and the neighborhood consistency condition is not met, the entry at that location is weakly interpolated or weakly removed based on the segment boundary splicing position in the denoising parameter snapshot, preventing instantaneous distortion caused by boundary splicing from entering subsequent calculations. Next, after constructing the single-point feature recording unit, the recording units are spliced ​​into a feature energy ratio vector sequence in chronological order. In multi-channel scenarios, if subsequent calculations of the variation index sequence from the feature energy ratio vector and the material characterization feature index dictionary require cross-channel calculations, this step establishes a channel-aligned temporal index mapping without changing the content of the recording units, enabling different channel recording units at the same time location to be read synchronously in subsequent calculations. Subsequently, to maintain consistency with the input specifications of the subsequent attention-guided bidirectional long-short-term recurrent network, this step structurally encapsulates the feature energy ratio vector with time index, channel identifier, and segment identifier to generate a feature energy ratio vector with a metadata header. The metadata header also includes a binding relationship with the denoising parameter snapshot, ensuring consistency between the frequency band division scheme and the window advancement scheme when the trend prediction sequence and the variation index sequence are used in conjunction. Finally, the feature energy ratio vector is passed as the output of this step to the feature index registration process. It also serves as direct input for subsequent variation index calculations based on the feature energy ratio vector and the material characterization feature index dictionary. Furthermore, it can be used as a candidate source for the material performance characterization parameter set and the supervision signal set in guiding the construction of the trend prediction sequence.

[0091] S330. Register the feature index of the feature energy ratio vector and generate a material characterization feature index dictionary;

[0092] The step takes the feature energy ratio vector as input, and completes the feature index registration around the consistent search and traceability requirements across time positions, channels and sections, to generate a material characterization feature index dictionary supporting subsequent multi-link reuse. Specifically, first, according to the metadata header of the feature energy ratio vector, an index skeleton is established under the time scale of unified time-based data, which includes four levels of time index, channel index, section index and frequency band index; On the skeleton, the feature energy ratio entry set corresponding to each time position is written in a predetermined order, and the version information of the frequency band division scheme and the window advancing scheme is repeatedly registered in the entry header, so as to maintain multiple batches of data in parallel in the future. Subsequently, in order to support the calculation of the variation index sequence from the feature energy ratio vector and the material characterization feature index dictionary, the step constructs the mapping rule of baseline reference channel and reference section in the material characterization feature index dictionary: when subsequent comparison is needed within a specific time window, the reference entry set can be quickly located according to the mapping rule, thereby reducing the cost of repeated scanning on large-scale data; At the same time, the invalid sample positions in the masking index are recorded in the form of time period as entries that cannot participate in comparison, ensuring that subsequent comparison is carried out only within the effective sample range. Further, in order to support the training of the attention-guided bidirectional long short-term recurrent network according to the high signal-to-noise ratio sequence, the material performance characterization parameter set and the supervision signal set, the step adds an input mapping list to the material characterization feature index dictionary, which describes the sampling step length of the feature energy ratio entry in the time dimension, the superposition rule in the channel dimension and the splicing order in the section dimension, so that the subsequent network can directly obtain the input slice conforming to the network structure when reading the feature energy ratio vector; When multi-channel merging or section merging is enabled, the step gives the corresponding superposition weight according to the reconstruction weight recorded in the noise reduction parameter snapshot, ensuring that the network input is consistent with the previous processing. Thirdly, considering the need to generate material performance degradation warning fusion and finite element for generating material benchmark response data under specific working conditions, the step sets a parameter backtracking pointer for each time position in the material characterization feature index dictionary, which points to the corresponding entry in the noise reduction parameter snapshot, to quickly restore the frequency band division scheme, window advancing scheme, channel weight and section boundary splicing position in subsequent sample generation and parameter review; At the same time, in order to facilitate multiple iterative updates, the step sets version identification and effective time period identification for the material characterization feature index dictionary, so that different batches of feature registration can coexist under the same engineering object without overlapping each other. Finally, the material characterization feature index dictionary is output as the step, which is directly provided to the variation index calculation from the feature energy ratio vector and the material characterization feature index dictionary, and is used as the retrieval entrance for the generation of material performance degradation warning fusion and finite element for generating material benchmark response data under specific working conditions, and is also used as the input mapping basis and feature retrieval list for the attention-guided bidirectional long short-term recurrent network.In summary, S310 completes the construction of marginal spectrum data and binds with the noise reduction parameter snapshot, S320 completes the extraction of feature energy ratio vector and consistent packaging with time index, S330 completes the registration of material characterization feature index dictionary and establishes the retrieval and backtracking relationship with the subsequent steps, the three together form a stable channel from the high signal-to-noise ratio sequence to the feature energy ratio vector, so that the subsequent calculation of the variation index sequence, the material performance degradation judgment result and the trend prediction sequence have consistent time-frequency alignment basis and traceable parameter basis.

[0093] Step S400 at least contains steps S410-S430:

[0094] S410, based on the material characterization feature index dictionary, the variation index is calculated to obtain a variation index sequence;

[0095] This step uses the feature energy ratio vector as the main input and the material characterization feature index dictionary as the retrieval and backtracking basis to conduct a consistent comparison across time positions and channels, generating a variation index sequence with a scale consistent with the unified time base data. Specifically, firstly, under the time scale consistent with the unified time base data, the feature energy ratio vector is located and read according to the time index, channel index, and segment index in the material characterization feature index dictionary, and the baseline reference mapping rules in the material characterization feature index dictionary are called to determine the reference time window and reference channel set corresponding to the current time position; when the current time position hits the masking index, it is marked as an invalid sample and remains empty in this step, not participating in the calculation. Subsequently, for each valid time position, the frequency band order of the current entry set and the reference entry set is read, and a one-to-one comparison is performed according to the frequency band division scheme. During the comparison process, the version information of the frequency band division scheme and window advancement scheme of the noise reduction parameter snapshot is called to ensure the consistency of cross-batch data. For positions with segment boundary splicing, a gradual transition is performed based on the segment boundary information recorded in the material characterization feature index dictionary, so that the comparison adopts neighborhood consistency processing near the boundary to avoid abnormal deviations introduced by splicing distortion. Furthermore, to take into account the consistency of multiple channels, this step performs channel alignment on the feature recording units of different channels within the same time position. The alignment rules are explicitly assigned by the channel index and superposition weight given by the material characterization feature index dictionary. After the alignment is completed, the comparison results of different channels are merged according to weight, and a neighborhood consistency correction strategy is adopted for short-term isolated anomalies, retaining stable differences only under the condition that adjacent time positions are continuously valid, thus forming the difference measurement entry for the current time position. Next, the aforementioned difference measurement items are concatenated in chronological order into a sequence consistent with a unified time scale, retaining invalid sample gaps within the sequence to maintain the correspondence with the masking index. For sequence concatenation across segments, a gradual transition is performed along the segment boundary concatenation position to ensure the continuity of the value domains of the time slices on both sides. Finally, the time scale, channel identifier, segment identifier, and difference measurement items are encapsulated into a variation index sequence, and a binding relationship with the material characterization feature index dictionary and the denoising parameter snapshot is written at the beginning of the sequence. This allows for direct retrieval of reference and parameter entries when performing material performance degradation determination on the variation index sequence. Thus, the input of this step is the feature energy ratio vector and the material characterization feature index dictionary, and the output is the variation index sequence. This sequence is then passed as the direct input to the material performance degradation determination of the variation index sequence, forming the basis for generating the material performance degradation determination result and threshold parameters.

[0096] S420. Determine the material performance degradation of the variation index sequence and generate the material performance degradation determination result and threshold parameters.

[0097] The above-mentioned variation index sequence is input into this step, the variation index is used to represent the change of the effective prestress of the prestressed steel material, the reference rule of the material characteristic index dictionary and the version information of the noise reduction parameter snapshot are combined, and the material performance degradation judgment and threshold parameter generation are carried out under the premise of unified time scale and channel alignment. Specifically, first, read the reference window and channel set information bound to each time position during sequence traversal, and perform availability check on the variation index of the current time position according to the time coverage of the reference window and the invalid sample position of the masking index; for the position hitting the invalid sample, this step records the skip identifier and keeps the empty position, so as to maintain the time alignment relationship with the high signal-to-noise ratio sequence and the feature energy ratio vector in the subsequent supervision signal construction. Subsequently, for the time positions passing the availability check, the baseline reference entries and the superposition weight in the material characteristic index dictionary are combined to generate a threshold candidate list; the threshold candidate list considers the comprehensive constraints of cross-channel merging, cross-section splicing and frequency band order consistency in the same time position, to ensure that the generation of the threshold can maintain parameter-level consistency with the previous feature construction. Further, in order to take into account the differences of different work points and different operation periods, this step groups the threshold candidate list according to the section information and window advancing scheme version recorded in the noise reduction parameter snapshot, to form the intra-section threshold and the cross-section connection threshold; within the same group, the time neighborhood consistency strategy is used to sequentially select the threshold candidates to obtain the final threshold parameter, and the threshold parameter is compared with the variation index of the current time position to generate the material performance degradation judgment result. Again, in order to facilitate the subsequent generation of material performance degradation warning fusion and supervision signal construction, this step structures and encapsulates the material performance degradation judgment result and the threshold parameter according to the time scale, and attaches the channel identifier, the section identifier and the version identifier, so that they can be directly retrieved by the attention-guided bidirectional long-short-term recurrent network trained by the high signal-to-noise ratio sequence and the material performance characteristic parameter set and the supervision signal set; for the time positions containing empty positions, continue to keep the skip identifier and write the empty position reason, which is consistent with the masking index. Finally, this step outputs the material performance degradation judgment result and the threshold parameter, and establishes a retrieval entry in the form consistent with the key-value rule of the material characteristic index dictionary, so that the subsequent supervision signal construction of the material performance degradation judgment result can directly trace back to the time slice and channel slice corresponding to the original feature energy ratio vector and the high signal-to-noise ratio sequence, and complete the alignment and disk storage of labels and samples.

[0098] S430, supervision signal construction is performed on the material performance degradation judgment result to generate a material performance characteristic parameter set and a supervision signal set;

[0099] The step takes the material performance degradation judgment result and the threshold parameter as input, and constructs the supervised signal under the unified time scale, channel identifier and section identifier, outputs the material performance characterization parameter set and the supervised signal set which can be directly called by the downstream network. Specifically, first, according to the time index and channel index of the material characterization feature index dictionary, the material performance degradation judgment result and the threshold parameter are aligned to the time slice corresponding to the feature energy ratio vector and the high signal-to-noise ratio sequence. For the empty position reserved in the previous step, this step continues to use the masking index for masking marking, and does not write the label to the time slice, but keeps the time scale and reason explanation, so as to automatically skip when the network reads. Subsequently, in each valid time slice, a label item is generated according to the material performance degradation judgment result, and is structurally bound with the threshold parameter, the feature energy ratio item and the high signal-to-noise ratio segment corresponding to the time slice; the binding content includes time scale, channel identifier, section identifier, feature item index, threshold parameter version and parameter backtracking pointer, which points to the corresponding item in the noise reduction parameter snapshot, so as to restore the frequency band division scheme and window advancing scheme at that time when review or playback is needed. Further, in order to match the input specification of the attention-guided bidirectional long-short-term recurrent network trained according to the high signal-to-noise ratio sequence, the material performance characterization parameter set and the supervised signal set, this step slices the structured bound samples according to the input mapping list in the material characterization feature index dictionary, generates sequence slices and label slices suitable for the network; in the multi-channel merging or section merging scene, sample superposition is performed according to the superposition weight recorded in the noise reduction parameter snapshot, and the label after superposition still maintains the one-to-one correspondence with the time scale. Again, in order to meet the review needs of the subsequent generation of material performance degradation warning fusion and the closed loop of finite element for generating the benchmark response data of the material under specific working conditions, this step records the source information and empty information of each sequence slice in the material performance characterization parameter set and the supervised signal set, and writes the binding relationship with the material performance degradation judgment result, the threshold parameter, the feature energy ratio vector and the high signal-to-noise ratio sequence in the header of the set, so that the set can be referenced by the generation of material performance degradation warning fusion and parameter review. Finally, the material performance characterization parameter set and the supervised signal set are output in a standardized data structure, the key value rules of which are consistent with the material characterization feature index dictionary, and can be directly read by the downstream network, and when necessary, the noise reduction parameter item and the frequency band division item at that time are restored through the parameter backtracking pointer, so that the consistent source of samples and labels is maintained in the training and inference links.In summary, S410 completes the construction of the sequence of variation indicators bound to the index dictionary of material characterization features, S420 generates the material performance degradation judgment result and threshold parameter under a unified time scale, S430 binds the material performance degradation judgment result and threshold parameter to the feature energy ratio vector and high signal-to-noise ratio sequence and outputs the material performance characterization parameter set and the supervision signal set, all of which are consistent in time index, channel identifier and section identifier and establish a traceable relationship with the noise reduction parameter snapshot, forming a closed-loop path from features to indicators, from indicators to judgments, and from judgments to supervision. In summary, the above steps realize the stable generation of the indicator sequence, the structured landing of the judgment rule, and the standardized packaging of the supervision signal, providing consistent, traceable and reviewable input basis for subsequent network training and generation of material performance degradation early warning fusion.

[0100] Step 500 at least contains steps S510-S530:

[0101] S510, based on the material performance characterization parameter set and the supervision signal set, training the attention-guided bidirectional long short-term recurrent network to obtain a bidirectional long short-term recurrent network;

[0102] The step takes high signal-to-noise ratio sequences as the main sample input, and takes the material performance characterization parameter set and the supervised signal set as the label input, and completes the training of the attention-guided bidirectional long short-term recurrent network under the time scale consistent with the unified time-based data, and outputs the pre-trained attention-guided bidirectional long short-term recurrent network. Specifically, first, read the time scale, channel identifier, section identifier and label slice in the material performance characterization parameter set and the supervised signal set, window the slice of the corresponding time slice, and use the record of the window advancing mode and the section splicing position in the previous processing, so that the sample slice and the label slice are aligned in the time dimension and the section boundary; if the time slice hits the missing mask, skip the segment according to the null marker of the material performance characterization parameter set and the supervised signal set, and ensure that the training only occurs in the available segment. Subsequently, according to the threshold parameter and the material performance degradation determination result recorded in the material performance characterization parameter set and the supervised signal set, an attention-guided weight is generated for each sample slice, and the attention-guided weight gives a higher proportion to the key segment related to damage and a lower proportion to the segment related to background disturbance in the time dimension; in a multi-channel scene, referring to the channel superposition relationship and weight item in the material performance characterization parameter set and the supervised signal set, the sample slices of different channels are fused or input in parallel according to the channel weight, and the attention-guided weight is loaded synchronously at the input end, so that the network can pay attention to the sequence information and the key segment weight at the same time when traversing in both directions. Further, to maintain consistency between training and inference, the sample slice is subjected to amplitude normalization and time standardization processing consistent with the noise reduction stage, and the processing parameters are recorded for subsequent inference; in the training cycle, the network first encodes the historical information in the forward direction, and then encodes the future information in the backward direction, and introduces the attention-guided weight at the bidirectional convergence point, and weights and aggregates the state vectors of important segments to form an output with the same length as the label slice. The output is aligned with the label at each time point in the material performance characterization parameter set and the supervised signal set, and the gradient is returned and the parameters are updated through the loss function until the convergence stopping criterion is reached. During the entire training process, the source information and parameter backtracking pointer of the material performance characterization parameter set and the supervised signal set are recorded in the training log, which is used to review the mapping relationship from the sample to the parameter in the subsequent generation of material performance degradation warning fusion and finite element for generating the benchmark response data of the material under a specific working condition. After training, the network parameters, input specifications, window advancing mode and attention-guided weight generation rules are solidified together to generate the pre-trained attention-guided bidirectional long short-term recurrent network, which is directly input to the next step.

[0103] S520, based on the bidirectional long short-term recurrent network, generating a material performance degradation trend prediction sequence and an uncertainty estimate;

[0104] The pre-trained attention-guided bidirectional long short-term recurrent network is used as the inference model in this step. The to-be-processed monitoring sequence is inferred under the time scale and channel identifier consistent with the high signal-to-noise ratio sequence, and the material performance degradation trend prediction sequence and uncertainty estimation are output. Specifically, first, the to-be-processed monitoring sequence is pre-processed in the same way as in the training phase, including amplitude normalization according to the amplitude normalization rules recorded in the training log, slicing according to the window advancing method, and calling the missing mask during slicing to mask the unavailable segments, so that the input slice only contains available samples; for cross-section slices, smooth connection is performed at the section boundary splicing position to ensure the continuity of the slice at the boundary. Then, the processed slice is input into the pre-trained attention-guided bidirectional long short-term recurrent network according to the channel fusion or parallel mode determined in the training phase; the attention-guided weight is generated synchronously at the input end, and the generation rule of the weight is consistent with that in the training phase. When the material performance characterization parameter set and the supervision signal set are updated online, the latest material performance degradation determination result and the threshold parameter are used to ensure that the weight assignment matches the current working condition. The network encodes the slice in both forward and backward directions, and gives the prediction value corresponding to the input time position at the output end; to obtain the uncertainty estimation, a random disturbance mechanism or multiple repeated inference strategy is introduced in this step during the inference process, and the multiple outputs are collected and statistically summarized at the time position to form interval or quantile information corresponding to the prediction value, which is used as the uncertainty estimation. Again, to facilitate the subsequent generation of material performance degradation warning fusion and parameter review, this step aligns the material performance degradation trend prediction sequence and the uncertainty estimation with the time scale of the unified time-based data immediately after generation, and adds channel identifier, section identifier and input slice source information to form a structured prediction record; the positions that hit the missing mask are also kept empty, and the reason for the empty position is written, so that the subsequent residual record can correctly skip this position. Finally, the material performance degradation trend prediction sequence and the uncertainty estimation are output as the output of this step, which are directly transmitted to the step of recording the residual of the material performance degradation trend prediction sequence, and are written in parallel channels to generate the material performance degradation warning fusion and the input buffer of the finite element for generating the benchmark response data of the material under a specific working condition, providing inference basis for subsequent material performance degradation warning level calculation and sample generation.

[0105] S530, recording the residual of the material performance degradation trend prediction sequence to generate a prediction residual sequence;

[0106] The step takes the material performance degradation trend prediction sequence as the prediction input, and takes the high signal-to-noise ratio sequence with the same time scale as the observation input, completes point-by-point comparison and difference registration at the same time position, and outputs the prediction residual sequence. Specifically, first, according to the time scale, channel identifier and section identifier in the material performance degradation trend prediction sequence, the corresponding time slice and channel slice are retrieved in the high signal-to-noise ratio sequence to construct a one-to-one comparison pair; when the time position hits the missing mask, skip the comparison according to the empty position mark in the input buffer, and only record the empty position information and reason explanation. Subsequently, at each valid time position, the observed segment is normalized according to the amplitude normalization and time standardization parameters fixed in the training phase, eliminating the offset caused by the scale difference; after the processing is completed, the point-by-point comparison is performed at the same time position and the difference value is registered, and the difference value and the uncertainty estimation are recorded in parallel, so as to distinguish between model inherent fluctuations and real deviations in subsequent generation of material performance degradation warning fusion. Further, in order to avoid the influence of instantaneous fluctuations at the boundary splicing position on the registration result, the step performs neighborhood smoothing strategy on the difference value near the section boundary, and simultaneously references the section boundary splicing position and window advancing mode to sequentially splice the difference records across the windows, forming a continuous prediction residual sequence; in a multi-channel scene, referring to the channel fusion relationship in the training log, the difference records of different channels are saved in parallel or aggregated according to the weight, and the channel weight item is written at the head of the sequence for subsequent use. Thirdly, in order to support the subsequent generation of material performance degradation warning fusion and the finite element for generating the benchmark response data of the material under specific working conditions, the step writes the source binding relationship between the prediction residual sequence and the material performance degradation trend prediction sequence, the uncertainty estimation, and the high signal-to-noise ratio sequence into the residual archive. The residual archive includes time scale mapping, channel mapping, section mapping, and parameter backtracking pointer, which points to the training log and noise reduction parameter item, facilitating review of the corresponding relationship during parameter update and retraining. Finally, the prediction residual sequence is output as the output of the step, which is directly provided to the step of obtaining the material performance degradation trend prediction sequence, the variation index sequence, and the prediction residual sequence to generate the material performance degradation warning fusion, and is simultaneously stored in the database to support sample generation and parameter review of the finite element for generating the benchmark response data of the material under specific working conditions according to the material performance degradation warning level sequence. In summary, after the processing of the step, the material performance degradation trend prediction sequence and the high signal-to-noise ratio sequence are compared point by point on the unified time scale, the prediction residual sequence forms a difference record that is traceable, reviewable and parallel to the uncertainty estimation, providing a consistent data basis and parameter backtracking basis for subsequent generation of material performance degradation warning fusion, threshold revision and retraining.

[0107] Step S600 includes at least steps S610-S630:

[0108] S610, acquire the prediction residual sequence, perform generation material performance degradation early warning fusion, and obtain a generation material performance degradation early warning level sequence;

[0109] This step takes the material performance degradation trend prediction sequence as the main input, and takes the variation index sequence and the prediction residual sequence as parallel inputs. The generation material performance degradation early warning fusion is implemented under the same time scale as the unified time base data, and the generation material performance degradation early warning level sequence corresponding to the channel identifier and the section identifier is output. Specifically, first, the three types of inputs are time-aligned on the unified time scale, the channel identifier, the section identifier and the window advancing mode in the previous record are read, and the boundary position of the cross-section window is processed to make the three types of inputs at the same time position continuous mapping at the section splicing position; when a certain time position hits the missing mask, the position is set to an unparticipating state according to the registered empty position marker, and an empty position is reserved in the generation material performance degradation early warning level sequence to maintain the index consistency. Subsequently, a fusion reading list is constructed, the current value of the material performance degradation trend prediction sequence, the current entry of the variation index sequence and the current difference value of the prediction residual sequence are extracted at each valid time position, and the threshold parameters, section boundary splicing positions and channel weight entries registered in the previous step are called to form a fusion entry set for subsequent judgment level. For the channel dimension in the fusion entry set, parallel merging or weighted aggregation is performed according to the registered channel weight to generate a single fusion reading at the same time position in the cross-channel scene. Further, a time neighborhood consistency reading strategy is established to sequentially scan the fusion readings in the continuous time window, and when a transient fluctuation caused by the section boundary, the missing mask or the window advancing occurs, the neighborhood consistency rule is called for smoothing processing to ensure the continuity of the fusion readings near the boundary and the two sides of the time slice. Subsequently, according to the relationship between the fusion reading and the threshold parameter, a generation material performance degradation early warning label entry is generated at each valid time position, and the entry is structurally bound with the time scale, the channel identifier and the section identifier; to support the subsequent finite element for the reference response data closed loop of the generation material under specific working conditions, this step synchronously writes the parameter backtracking pointer, which points to the frequency band division scheme and the window advancing scheme entry in the noise reduction parameter snapshot, so that the simulation generation sample can be aligned according to the same frequency band and window configuration. Finally, the generation material performance degradation early warning level sequence is output in a standardized data structure, including the time-sequenced generation material performance degradation early warning label entry and the bound index information, as the direct input of the next step, and the generation material performance degradation early warning fusion log is written in parallel channels to record the specific values of the empty position reason, the boundary position and the channel weight, ensuring that the source of the simulation sample library and the parameter update can be traced back.

[0110] S620, according to the generated material performance degradation warning level sequence, the finite element is used for generating the benchmark response data closed loop of the material under specific working conditions, and the simulation sample library is obtained;

[0111] This step is driven by the generated material performance degradation warning level sequence and is configured according to the noise reduction parameter snapshot. A simply supported beam calculation model and a loading process consistent with the uniform time-based data scale are constructed on the finite element simulation platform, and a simulation sample library corresponding to the channel identifier and the section identifier is output. Specifically, first, the time scale, channel identifier and section identifier in the generated material performance degradation warning level sequence are read, and a simulation loading list is established according to the order of the generated material performance degradation warning label entries. The loading list indicates the working condition intensity, action history and time length that need to be reproduced in each time section. To ensure that the simulation generated samples can be directly replaced or supplemented in the subsequent steps, this step synchronously calls the frequency band division scheme, window advancing mode, channel weight and section boundary splicing position in the noise reduction parameter snapshot to establish a consistent mapping relationship between simulation output and measured input. Subsequently, according to the loading list and the consistent mapping relationship, the boundary conditions and action paths of the simply supported beam model are configured, and the simulation solution is run in the time window corresponding to the generated material performance degradation warning label entry to obtain multi-channel response data consistent with the high signal-to-noise ratio sequence. For time positions containing missing mask spaces, this step does not generate simulation data, but retains the time index and space reason to avoid introducing inconsistent supervision information in subsequent parameter updating and retraining. Further, to maintain parameter consistency with the previous joint denoising and component credibility evaluation, this step maps the simulation output according to the frequency band division scheme of the noise reduction parameter snapshot, and slices the window according to the time advancement consistent with the window advancing mode to form time domain segments that can be directly input into the previous process. At the same time, each simulation segment is bound with the generated material performance degradation warning label entry from which it comes, and a sample unit containing time scale, channel identifier, section identifier, frequency band mapping and source label is generated. Again, to meet the needs of multi-condition and multi-intensity stratification, this step groups the sample units according to different levels of the generated material performance degradation warning level sequence, forms a grouping list and writes a parameter backtracking pointer in the list header, so that each group of samples can be traced back to specific frequency band division scheme, window advancing mode and channel weight entries. Finally, the above sample units are structured and packaged in time sequence and grouping list to generate a simulation sample library; the simulation sample library records the triple mapping relationship of sample to model, sample to parameter and sample to generated material performance degradation warning label, and is directly input into the next step, while writing the benchmark response data log of the material under specific working conditions in parallel channels for subsequent quick review and playback when parameter updating fails or deviation is too large.

[0112] S630, based on the simulation sample library, parameter updating and retraining are performed to generate an updated process configuration;

[0113] This step takes the simulation sample library as the data input and the noise reduction parameter snapshot as the parameter input. Parameter updating and retraining are simultaneously performed for the joint noise reduction and component credibility evaluation and the attention-guided bidirectional long short-term recurrent network, and an updated process configuration is output. Specifically, first, on the time scale consistent with the unified time-based data, the simulation sample units are read according to the grouping list, the frequency band mapping and window advance information recorded in the sample units are called, and the sample units are sequentially sent to the joint noise reduction and component credibility evaluation. In this process, according to the component number, center frequency band, penalty strength and channel weight entries in the noise reduction parameter snapshot, the component set and segment splicing strategy are reconfigured, so that the simulation sample undergoes consistent decomposition, screening and reconstruction process as the measured data. The high signal-to-noise ratio segments and process records generated by each sample unit are summarized, and the component credibility evaluation value, threshold scheduling track and segment boundary splicing position are compared to determine the parameters that need to be updated, form a new noise reduction parameter snapshot and record its version and effective time period. Subsequently, on the network side, the sample units in the simulation sample library are generated into training slices and label slices according to their binding relationship with the material performance degradation warning label, the existing input specification and slice length are called, and the training slices are sent to the attention-guided bidirectional long short-term recurrent network for retraining. In the retraining cycle, the slice organization method and attention-guided weight generation rule of the material performance characterization parameter set and the supervision signal set are continued to be used, so that the new samples and existing samples optimize the network parameters under the same specification. When the network reaches the predetermined convergence criteria on the new samples, the new network parameters, input specification and weight generation rule are solidified. Further, to form a consistent configuration across processes, this step writes the new noise reduction parameter snapshot and the new network parameters into the process configuration dictionary, and records the parameter backtracking pointer and coverage range in the configuration dictionary; when there are historical time periods or channels that cannot be covered, the old version entries are retained in the configuration dictionary and the effective interval is marked to support phased invocation. Under the premise of keeping the material detection means unchanged, inverse modeling (optional finite element or analytical inversion) for material parameter calibration is performed based on the material performance characterization parameter set to generate calibrated material performance evaluation results; and based on this, the joint noise reduction and component credibility evaluation and the attention-guided bidirectional long short-term recurrent network are compared and the parameter updating is performed to generate the process configuration dictionary and the updated process configuration.

[0114] Finally, the updated process configuration is output in a standardized form, including the retrained attention-guided bidirectional long short-term recurrent network, the updated parameters for joint denoising and component confidence evaluation, and the mapping relationships between them and the time scale, channel identifier, and section identifier. This process configuration is immediately distributed to each processing node after output, used to replace or incrementally cover the existing configuration, and written synchronously into the parameter entry of the material performance degradation warning fusion and finite element for generating material benchmark response data under specific working conditions, providing a consistent starting point for the next round of closed-loop execution. In summary, through this step, parameter updating and retraining are completed, which enables the simulation sample library and denoising parameter snapshot to be digested and solidified into an executable configuration under the same data and parameter framework, forming a closed-loop link from the sequence of generated material performance degradation warning levels to the simulation sample library to the updated process configuration. Overall, it achieves the integration of sample generation driven by generated material performance degradation warning fusion, sample-driven parameter correction, and retraining after parameter correction, providing a basis for subsequent processing with consistent timing, parameter traceability, and configuration reusability.

Claims

1. A method for non-destructive testing and performance degradation analysis of bridge structural materials, characterized in that, include: Acquire non-destructive testing monitoring data of materials, perform unified alignment, noise profile construction and quality labeling processing, and generate missing measurement mask; The monitoring data includes material testing channel data reflecting the mechanical properties of the material, material micro-vibration acceleration response time history data collected by accelerometers deployed in local areas of the parent material, material local strain response time history data collected by strain sensors deployed in local areas of the parent material or at the sample location, and synchronously recorded working condition information and time stamp information. Perform joint noise reduction configuration, component evaluation and noise reduction operation on material detection signals, and generate a snapshot of noise reduction parameters; Perform marginal spectrum calculation, characteristic energy ratio extraction and characteristic index registration processing to generate a material characterization characteristic index dictionary; Perform operations such as calculating variation indices, determining material performance degradation, and constructing monitoring signals to generate a set of material performance characterization parameters and a set of monitoring signals; Perform network training, predict material property degradation trends, and process residual records to generate a predicted residual sequence; The trend prediction sequence, variation index sequence, and prediction residual sequence are obtained. A fusion reading list is constructed and a time neighborhood consistency reading strategy is adopted to generate a material performance degradation early warning fusion, resulting in a material performance degradation early warning level sequence.

2. The method according to claim 1, characterized in that, The fusion of early warning for material performance degradation specifically includes: Using the trend prediction sequence as the main input and the variation index sequence and prediction residual sequence as parallel inputs, the material performance degradation early warning fusion is generated under the time scale consistent with the unified time base data, and the output is a material performance degradation early warning level sequence that corresponds one-to-one with the channel identifier and the segment identifier.

3. The method according to claim 2, characterized in that, The generated material performance degradation early warning fusion includes: A fusion read list is constructed, and at each valid time position, the current value of the trend prediction sequence, the current entry of the variation index sequence, and the current difference value of the prediction residual sequence are extracted simultaneously. The threshold parameters, segment boundary splicing positions, and channel weight entries registered in the previous steps are called to form a fusion entry set. For the channel dimension in the fusion entry set, parallel merging or weighted aggregation is performed according to the registered channel weights, so that a single fusion read is generated at the same time position in the cross-channel scenario.

4. The method according to claim 3, characterized in that, The fusion of early warning for material performance degradation also includes: A time neighborhood consistency reading strategy is established to sequentially scan the fused readings within a continuous time window. When instantaneous fluctuations occur due to segment boundaries, missing measurement masks, or window advancement, the neighborhood consistency rule is invoked to handle the gradual change, ensuring that the fused readings near the boundary remain consistent with the time slices on both sides.

5. The method according to claim 1, characterized in that, The material performance degradation early warning level sequence also includes: Using the generated material performance degradation early warning level sequence as the driving input and the noise reduction parameter snapshot as the configuration basis, a simple supported beam calculation model and loading process with the same scale as the unified time base data are constructed on the finite element simulation platform, and the simulation sample library corresponding to the channel identifier and section identifier is output.

6. The method according to claim 5, characterized in that, include: The simulation output is frequency band mapped according to the frequency band division scheme of the noise reduction parameter snapshot, and windowed slices are performed under the time advancement consistent with the window advancement method to form time domain segments that can be directly input into the preceding process; at the same time, each simulation segment is bound to the material performance degradation warning label entry of its source to generate sample units containing time scale, channel identifier, segment identifier, frequency band mapping and source label.

7. The method according to claim 6, characterized in that, include: Using a simulation sample library as data input and a snapshot of denoising parameters as parameter input, the system simultaneously performs parameter updates and retraining on a bidirectional long short-term recurrent network that combines joint denoising, component reliability assessment, and attention guidance, and outputs the updated process configuration.

8. The method according to claim 1, characterized in that, include: The new noise reduction parameter snapshot and the new network parameters are written into the process configuration dictionary, and the parameter backtracking pointer and coverage range are recorded in the configuration dictionary. When there are historical time periods or channels that cannot be covered, the old version entries are retained in the configuration dictionary and the effective range is marked to support phased calls.

9. The method according to claim 1, characterized in that, include: Based on the loading list and the uniform mapping relationship, the boundary conditions and action paths of the simply supported beam model are configured, and the simulation is run within the time window corresponding to the generated material performance degradation warning label entries to obtain multi-channel response data that is isomorphic to the high signal-to-noise ratio sequence. For time positions containing missing measurement mask gaps, simulation data is not generated, but the time index and the reason for the gap are retained to avoid introducing inconsistent supervision information in subsequent parameter updates and retraining.

10. The method according to claim 1, characterized in that, In the retraining loop, the slicing organization method of the material property characterization parameter set and the supervision signal set and the attention-guided weight generation rule are continued to be used, so that the new simulation samples and the existing measured samples jointly optimize the network parameters under the same specification. When the network reaches the predetermined convergence criterion on the new samples, the new network parameters, input specifications and weight generation rules are solidified, forming a closed-loop link from generating the material property degradation early warning level sequence to the simulation sample library and then to the updated process configuration.

Images