A single steel strand prestress identification method and system based on physical constraint generative data enhancement

Abstract

The application discloses a single steel strand prestress identification method and system based on physical constraint generative data enhancement. The method collects multi-source acoustic response signals of an anchoring system under impact excitation, forms a fusion feature vector through pretreatment and cross-channel consistent feature screening; trains a generative data enhancement network with steel strand dynamics physical constraints with real samples, obtains virtual acoustic samples; trains a Gaussian process regression identification model with real samples and virtual samples, and performs rejection and drift incremental update according to the prediction variance, thereby supporting long-term prestress detection.

Description

Technical Field

[0001] This application belongs to the field of intelligent detection and health monitoring technology of civil engineering structures, and relates to non-destructive testing of prestressed steel strands, and particularly to a method and system for identifying prestress in a single steel strand based on physical constraint generative data augmentation. Background Technology

[0002] Prestressed steel strands are load-bearing components in bridges, buildings, and underground structures, and their effective prestress affects the structural load-bearing state. The anchorage ends of in-service structures are often sealed, making direct measurement difficult; therefore, acoustic non-destructive testing becomes a feasible approach. Existing methods typically rely on a small number of labeled samples to train the recognition model, and acoustic samples generated by ordinary generative models tend to deviate from the dynamic response of the steel strands, resulting in augmented data that cannot stably support the recognition model. During long-term operation, changes in relaxation, temperature, and anchorage condition can also cause acoustic feature drift, gradually causing the fixed model to deviate from the actual object being tested. Summary of the Invention

[0003] To address the shortcomings of existing technologies, such as insufficient labeled samples, lack of physical constraints in generated acoustic samples, and acoustic feature drift during long-term operation, this application provides a method and system for identifying prestressed strands of steel strands based on physically constrained generative data augmentation. This method utilizes multi-source acoustic acquisition, cross-channel consistent feature screening, physically constrained generative augmentation, Gaussian process regression prediction variance assessment, and drift increment updates to create a closed loop of training sample expansion, result rejection, and long-term adaptive updates. This reduces the impact of sample scarcity and time-varying drift on prestressed strand identification. Specifically, this application provides the following technical solutions.

[0004] In a first aspect, this application provides a method for identifying prestress in a single steel strand based on physically constrained generative data augmentation, comprising: acquiring multi-source acoustic response signals of the anchoring system of the steel strand under test under standardized impact excitation; preprocessing the multi-source acoustic response signals to obtain effective impact response samples; extracting multi-scale acoustic features from the effective impact response samples, and screening key features according to cross-channel consistency to form a fused feature vector; using the effective impact response samples as real samples, training a generative data augmentation network with physical constraints of steel strand dynamics to obtain virtual acoustic samples; training a Gaussian process regression identification model using the real samples and the virtual acoustic samples, and obtaining prestress identification results and prediction variance based on the fused feature vector under test; evaluating the confidence of the prestress identification results based on the prediction variance, and triggering rejection when the evaluation fails; continuously monitoring the temporal changes of the prediction variance, generating drift samples when feature drift is detected, and incrementally updating the Gaussian process regression identification model.

[0005] Optionally, the acquisition of the multi-source acoustic response signal includes: applying impact excitation to the anchoring system using a rebound spring; simultaneously acquiring the response signals of the cochlear sensor, the external sound pressure sensor, and the internal sensor of the rebound spring; and associating the response signals of multiple channels under the same impact excitation as the same detection sample.

[0006] Optionally, the preprocessing includes: performing downsampling and high-pass filtering on the original response signal; performing pre-emphasis and peak alignment on the filtered signal; removing abnormal impulses according to sample quality rules; and truncating impulse response segments of fixed length.

[0007] Optionally, the extraction and screening of the multi-scale acoustic features includes: extracting time-domain features, frequency-domain features, and cepstral features from each valid impact response sample; evaluating the correlation between candidate features of each channel and the prestress label using the maximum mutual information coefficient; selecting candidate features that satisfy the correlation condition for multiple channels as key features, and concatenating them to form the fused feature vector.

[0008] Optionally, the training of the generative data augmentation network includes: performing batch normalization on the extracted real samples; introducing a physical constraint loss consisting of the spectral peak, decay rate, and bandwidth energy into the adversarial training loss; and fusing the various losses using a weighting strategy adjusted according to the training stage and the generation quality feedback.

[0009] Optionally, the generation of the virtual acoustic samples includes: organizing real samples according to the strand hole or prestress level; generating multi-channel synchronous virtual signals using a trained generative data augmentation network; and combining the virtual signals with the real samples to form an augmentation training set.

[0010] Optionally, the training of the Gaussian process regression identification model includes: inputting the enhanced training set into the Gaussian process regression model; obtaining the model hyperparameters using kernel functions and cross-validation; and outputting the prestress identification result and the corresponding prediction variance.

[0011] Optionally, the confidence assessment includes: forming an absolute variance evaluation based on the predicted variance; forming a relative interval evaluation based on the relative width of the confidence interval; and outputting a rejection flag and prompting for re-collection or manual review when any evaluation exceeds the corresponding threshold.

[0012] Secondly, this application provides a prestress identification system for a single steel strand based on physical constraint generative data augmentation, comprising: a signal acquisition unit for acquiring multi-source acoustic response signals of the anchoring system of the steel strand under test under standardized impact excitation; a sample processing unit for preprocessing the multi-source acoustic response signals and forming a fused feature vector; a generative augmentation unit for training a generative data augmentation network with physical constraints of steel strand dynamics and obtaining virtual acoustic samples; an identification evaluation unit for training a Gaussian process regression identification model based on real samples and virtual acoustic samples, and outputting prestress identification results, prediction variance, and rejection labels; and a drift update unit for monitoring the temporal change of prediction variance and generating drift samples and updating the Gaussian process regression identification model when feature drift is detected.

[0013] Optionally, the system further includes a sample storage unit, which is connected to the sample processing unit, the generation and enhancement unit, the identification and evaluation unit, and the drift update unit, respectively, for storing real samples, virtual acoustic samples, rejected samples, and recent samples with high confidence, and providing incremental update data to the drift update unit.

[0014] The beneficial effects of this application are as follows: the multi-source acoustic response signal and the cross-channel consistent feature screening work together to make the input features more reflective of the common dynamic response of the steel strand anchoring system; the physical constraint generative data augmentation network is simultaneously constrained by the real waveform distribution and dynamic physical constraints when generating samples, which can reduce the risk of virtual samples deviating from the real detection scenario; the prediction variance of the Gaussian process regression recognition model is used for confidence assessment and drift monitoring, so that the rejection of abnormal inputs, the generation of drift samples and incremental updates form a continuous feedback path, thereby improving the recognition stability in long-term detection. Attached Figure Description

[0015] Figure 1 A flowchart of the method for identifying prestress in a single steel strand, which is supplementary to this application.

[0016] Figure 2 A block diagram of a single strand prestress identification system to supplement this application.

[0017] Figure 3 This is a schematic diagram of environmental noise time-domain analysis provided in an embodiment of this application.

[0018] Figure 4 This is a schematic diagram of environmental noise frequency domain analysis provided in an embodiment of this application.

[0019] Figure 5 This is a schematic diagram comparing the time-domain signals before and after pre-emphasis, provided in an embodiment of this application.

[0020] Figure 6This is a schematic diagram of the data quality control results of the cochlear sensor provided in an embodiment of this application.

[0021] Figure 7 This is a schematic diagram of the data quality control results of the external sound pressure sensor provided in the embodiments of this application.

[0022] Figure 8 This is a schematic diagram of the data quality control results of the internal sensor of the rebound hammer provided in an embodiment of this application.

[0023] Figure 9 This is a schematic diagram comparing the frequency domain responses of different sensors provided in the embodiments of this application.

[0024] Figure 10 A schematic diagram showing the maximum mutual information coefficient screening results of the cochlear sensor, external sound pressure sensor, and internal sensor of the rebound device provided in the embodiments of this application.

[0025] Figure 11 This is a schematic diagram of the prediction results for working condition III provided in the embodiments of this application.

[0026] Figure 12 A schematic diagram showing the comparison between predicted prestress and actual prestress under working condition III, provided for an embodiment of this application.

[0027] Figure 13 This is a schematic diagram of the predicted residual for operating condition III provided in an embodiment of this application.

[0028] Figure 14 This is a schematic diagram of the sample prediction results of the three-sensor fusion model provided in the embodiments of this application.

[0029] Figure 15 This is a schematic diagram comparing the predicted prestress and the actual prestress provided by the three-sensor fusion model in the embodiments of this application.

[0030] Figure 16 This is a schematic diagram of the prediction residual of the three-sensor fusion model provided in the embodiments of this application. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0033] Example 1 like Figures 1-2 As shown, this embodiment provides a method and system for identifying prestress in a single steel strand based on physically constrained generative data augmentation. This method obtains multi-source acoustic response signals through standardized impact excitation, forms a fused feature vector through preprocessing and cross-channel consistent feature screening, expands the training samples through a physically constrained generative data augmentation network, outputs the prestress identification results and prediction variance through a Gaussian process regression identification model, and performs rejection and drift incremental updates based on the prediction variance. The system consists of a signal acquisition unit, a sample processing unit, a generative augmentation unit, an identification evaluation unit, a drift update unit, and a sample storage unit working collaboratively to address the problems of limited labeled samples of in-service single steel strands, insufficient physical consistency of virtual samples, and acoustic feature drift during long-term testing. Figure 1 As shown in the flowchart, the method includes multi-source acoustic response signal acquisition, effective impact response sample acquisition, fusion feature vector formation, physical constraint generative enhancement, Gaussian process regression identification, confidence assessment and rejection, feature drift monitoring, and incremental model update. Specifically: S100 collects multi-source acoustic response signals of the steel strand anchoring system under test under standardized impact excitation.

[0034] In this embodiment, the object to be measured is a 7Φ5-15.2 steel strand, which is inserted into a shear wall duct. The shear wall dimensions are 7.5 m × 1.5 m × 10 m. One end of the steel strand serves as the tensioning end, where a hydraulic jack, pressure sensor, anchor plate, 40Cr alloy steel single-hole anchor, and clamp are installed sequentially. The other end of the steel strand serves as the data acquisition end, where an anchor plate, single-hole anchor, and clamp are installed. The hydraulic jack has a range of 300 kN, and the pressure sensor has a range of 300 kN. The tension force is set to five levels: 40 kN, 80 kN, 120 kN, 160 kN, and 200 kN, with progressive loading starting from 40 kN. This arrangement provides a traceable prestressing label at the tensioning end and a repeatable excitation and sensor placement anchoring system at the data acquisition end.

[0035] Exemplary instrumentation includes an INV9206 sound pressure sensor, an INV3062T cloud-based data acquisition unit, an HS6020 sound level calibrator, a ZC3-A rebound hammer, an 1A116E accelerometer, and a DH5922N dynamic signal analysis system. The INV9206 sound pressure sensor has a frequency range of 20 Hz to 20 kHz and a dynamic range of 20 dB to 146 dB. The INV3062T data acquisition unit has a maximum sampling frequency of 51.2 kHz and is used to acquire synchronous signals from multiple channels. The HS6020 sound level calibrator outputs a 94 dB and 1000 Hz calibration sound source for sensor consistency calibration before acquisition. This instrument combination covers the main frequency components of the impulse response and maintains consistent sampling time references across different channels.

[0036] S110 uses a rebound hammer to apply impact excitation to the anchoring system.

[0037] The rebound hammer applies impact excitation to the side of the anchor at the acquisition end, with the rebound position selected on the side of the anchor surface furthest from the anchor plate. This position reduces the direct influence of the anchor plate boundary on the local response at the impact point, allowing the impact energy to be transmitted along the anchor, wedges, and steel strands. Multiple impact acquisitions are performed at each tension level, and the trigger status of the rebound hammer and the recording status of the acquisition instrument are checked after each impact. If the impact trigger is too weak, the rebound position deviates from the set area, or the channel recording shows obvious saturation, then that record is not used as a subsequent candidate sample.

[0038] For example, 43 records were collected at the 40 kN tension level of channel 1, 48 records at the 80 kN tension level, 46 records at the 120 kN tension level, 51 records at the 160 kN tension level, and 60 records at the 200 kN tension level, resulting in a total of 248 records for channel 1. For channel 2, 45, 43, 53, 41, and 45 records were collected at the five tension levels, respectively, resulting in a total of 227 records. The collection process for channel 3 can be performed in the same manner, and together with channels 1 and 2, it forms the basis of the real samples for subsequent model training.

[0039] S120 simultaneously acquires response signals from the cochlear sensor, external sound pressure sensor, and internal sensor of the rebound device.

[0040] A cochlear sensor is attached directly above the anchorage to collect the solid-borne acoustic response within the anchoring system; an external sound pressure sensor is positioned 20 cm from the anchorage to collect the airborne acoustic response; and an internal sensor in the rebound hammer collects the mechanical vibration response of the rebound hammer itself. All three channels record synchronously under the same impact, forming a multi-source acoustic response signal for the same impact event. By combining solid-borne, airborne, and mechanical vibration signals from the excitation source, the response differences caused by variations in the prestress of the steel strand can be observed from different propagation paths.

[0041] For example, the initial sampling frequency of the data acquisition instrument is set to 51.2 kHz, and the time-domain waveforms of three channels are recorded simultaneously with each impact. If a certain impact event produces a clear impact peak in the cochlear sensor, a corresponding sound pressure peak in the external sound pressure sensor, and a corresponding mechanical peak in the rebound hammer's internal sensor, then the event is associated as a multi-source detection sample. For the same impact event, the sample records the sensor channel number, tension label, channel number, acquisition time, and impact sequence number, facilitating subsequent grouping and processing according to channel and tension level.

[0042] S130 associates the response signals of multiple channels under the same impact excitation as the same detection sample.

[0043] After synchronous acquisition is completed, a sample index is created for each impact. The sample index is used to bind waveform segments, prestress labels, and test conditions for the three channels. This binding relationship does not depend on the response strength of a single channel, but rather on the consistency of the acquisition instrument's trigger record and timestamp. If any channel is lost, saturated, or triggers abnormally in a certain impact, the records of the three channels corresponding to that impact are completely removed to avoid channel mismatch during subsequent feature stitching.

[0044] For example, the sample index can be written as "Channel 1-120 kN-18th impact". This index corresponds to a segment of solid-borne sound waveform from the cochlear sensor, a segment of air-borne sound waveform from the external sound pressure sensor, and a segment of mechanical vibration waveform from the internal sensor of the rebound device. The three waveforms together correspond to a 120 kN tension label, and subsequent preprocessing and feature extraction use this index as the smallest processing unit.

[0045] S200, preprocess the multi-source acoustic response signal to obtain effective impact response samples.

[0046] After downsampling, high-pass filtering, pre-emphasis, sample quality control, peak alignment, and fixed-window truncation, the multi-source acoustic response signals form effective impact response samples. Preprocessing does not change the prestress label of the samples; it only processes noise, trigger time offset, and abnormal impacts in the waveform, ensuring that subsequent feature extraction targets stable impact response segments. For example... Figures 3 to 9As shown, environmental noise, pre-emphasis effect, quality control, and channel frequency response can all be verified through visualization results.

[0047] S210 performs downsampling and high-pass filtering on the original response signal.

[0048] The raw response signal was first subjected to spectral analysis. After performing Fast Fourier Transform on the basic test conditions at five tension levels (40 kN, 80 kN, 120 kN, 160 kN, and 200 kN), the frequencies to be analyzed were mainly within 20 kHz. Therefore, without losing information in the target frequency band, the sampling frequency was adjusted from 51.2 kHz to 40 kHz. This processing reduced the data length of each sample and kept the input size of the subsequent fixed-window truncation and generation network consistent.

[0049] For example, the time-domain and frequency-domain analysis of environmental noise from different sensors (sensor 1, sensor 2, sensor 3) is as follows: Figures 3-4 As shown, where, Figure 3 Time-domain plots for different sensors. Figure 4 The frequency domain diagrams for different sensors are shown. Sensor 1 is a cochlear sensor, sensor 2 is an external sound pressure sensor, and sensor 3 is an internal sensor of the rebound hammer. Low-frequency interference exists in the test environment around 0.06 Hz, 0.09 Hz, and 16.5 Hz. 0.06 Hz and 0.09 Hz correspond to DC drift components, while 16.5 Hz is located in the region of concentrated low-frequency environmental noise. To address this noise characteristic, a fourth-order Butterworth high-pass filter is used, with a cutoff frequency set to 20 Hz. Zero-phase processing is employed during filtering to avoid peak time shifts caused by ordinary unidirectional filtering. For a waveform of the internal sensor of the rebound hammer with a length of 4097 points, the baseline before filtering may drift slowly around 0.018 V, while the baseline after filtering returns to a range close to 0 V, and the impact peak and damped oscillations maintain their original timing relationship.

[0050] S220 performs pre-emphasis and peak alignment on the filtered signal.

[0051] The high-pass filtered signal undergoes further pre-emphasis processing to weaken low-frequency components and highlight the rapid changes caused by the impulse excitation. The pre-emphasis uses a first-order differential form:

[0052] in, The original discrete signal, The discrete signal after pre-emphasis. This is the pre-emphasis coefficient. For example, Taking 0.97, the time-domain comparison of the internal sensor signals of the rebound hammer before and after pre-weighting is as follows: Figure 5As shown. The time-domain comparison of the signals from the built-in sensor of the rebound hammer before and after pre-weighting is as follows: Figure 5 As shown. Under the parameters of this embodiment, the pre-emphasis processing did not cause a significant shift in the position of the main impact peak, and the overall attenuation trend remained basically unchanged. At the same time, it enhanced the rapid oscillation component in the impact attenuation stage, which facilitates the subsequent extraction of time-domain features and cepstrum features.

[0053] Peak alignment uses the peak value of the sound pressure envelope as the alignment benchmark. For each synchronized sample, the peak position of the impact envelope is first obtained in the external sound pressure channel, and then the cochlear channel and the built-in channel of the rebound device are synchronously cropped according to the same impact event. For example, a 4097-point segment is formed at a sampling frequency of 40 kHz, with the truncation intervals being 0.03 s before and 0.05 s after the peak. This interval covers the stable background before the impact, the main impact peak, and the main attenuation segment, which can balance trigger robustness and physical response integrity. If the peak position of a certain channel deviates from the set tolerance in time correlation with the same impact event, the sample enters quality review and is not directly involved in subsequent training.

[0054] S230, abnormal shocks are removed according to the sample quality rules, and a fixed-length shock response segment is extracted.

[0055] Sample quality control is based on peak amplitude, root mean square (RMS) value, and multi-channel consistency. For multiple acquisitions of signals under the same tensile condition, the peak amplitude and RMS value of each channel are first obtained, and then the quality boundary is formed using the mean and standard deviation. If a record exceeds the quality boundary in any channel, or if the impact time relationship of the three channels is inconsistent, the record is discarded. The data distribution after quality control is as follows: Figure 6-8 As shown, Figure 6 For the cochlear sensor data quality control results, Figure 7 Schematic diagram of data quality control results from external sound pressure sensor. Figure 8 This is the distribution result of sensor data inside the rebound hammer; retaining samples can reduce the influence of accidental impact force, sensor contact state and environmental noise on subsequent features.

[0056] For example, if the mean peak amplitude of the cochlear channel at a certain Epsilon level is 0.82 V and the standard deviation is 0.06 V, then the peak amplitude quality boundary can be set to 0.70 V to 0.94 V. If the peak amplitude of the cochlear channel of a sample is 1.10 V, and the external sound pressure channel simultaneously exhibits amplitude saturation, then this sample is marked as an abnormal impact. If the peak values ​​of all three channels of another sample are within the boundary, and the main peak position falls within the same fixed window range after peak alignment, then this sample is considered a valid impact response sample and enters feature extraction. The frequency domain characteristics of different sensors (sensor 1, sensor 2, sensor 3) are as follows: Figure 9As shown, the energy of the cochlear sensor and the external sound pressure sensor is mainly concentrated in the 2000 Hz to 4000 Hz frequency band, while the energy of the internal sensor of the rebound device can cover the 1000 Hz to 5000 Hz range. This difference provides complementary information for subsequent multi-sensor fusion.

[0057] S300, extract multi-scale acoustic features from the effective impact response samples, and filter key features according to cross-channel consistency to form a fused feature vector.

[0058] The valid impact response samples consist of three sensor channels. Time-domain features, frequency-domain features, and cepstral features are extracted for each channel to form a candidate feature set. Subsequently, the maximum mutual information coefficient is used to evaluate the nonlinear correlation between the candidate features and the tension label, and the feature intersection of multiple channels that satisfy the correlation condition is extracted. This processing does not rely on highly correlated features of a single channel, but rather retains features with consistent responses across channels to reduce the impact of sensor installation location differences and local noise on the model input.

[0059] S310 extracts time-domain features, frequency-domain features, and cepstral features from each valid impulse response sample.

[0060] For each channel of each valid sample, 705 candidate features are extracted. These candidate features consist of 80 time-domain features, 85 frequency-domain features, and 540 cepstral features. Time-domain features include short-time average amplitude, short-time average amplitude difference, short-time autocorrelation coefficient, short-time energy, and short-time zero-crossing rate and their statistics. Frequency-domain features include spectral centroid, spectral stretch, spectral entropy, spectral flux, 12th-order linear prediction coefficients, frequency-domain energy, and the energy ratios of multiple subbands and their statistics. Cepstral features include Mel-frequency cepstral coefficients, gamma-pass filter cepstral coefficients, and power-normalized cepstral coefficients; each type of cepstral feature is 180-dimensional.

[0061] For example, Mel-frequency cepstral coefficients use a Mel-frequency filter bank to describe the nonlinear characteristics of human ear frequency perception; gamma-ton filter cepstral coefficients use a gamma-ton filter bank to describe the basilar membrane frequency decomposition mechanism; and power-normalized cepstral coefficients introduce power normalization and nonlinear compression, making them suitable for robust feature representation in complex noise environments. For instance, for an effective sample with a tension of 120 kN in channel 1, the cochlear channel, external sound pressure channel, and rebound thermometer built-in channel each form a 705-dimensional candidate feature vector. This stage retains a relatively wide candidate space, enabling subsequent screening to select features strongly correlated with prestress changes from different acoustic description methods.

[0062] S320 uses the maximum mutual information coefficient to evaluate the correlation between candidate features and prestressed labels for each channel.

[0063] The maximum mutual information coefficient is used to evaluate the nonlinear dependency between candidate features and prestressed labels. Mutual information can be expressed as:

[0064] in, Represents candidate feature variables. Indicates the prestressed label variable. Denotes the joint probability. and This represents the marginal probability. In practical implementation, the maximum mutual information coefficients are obtained for the 705-dimensional candidate features of the cochlear sensor, external sound pressure sensor, and internal sensor of the rebound device. For example... Figure 10 As shown, the screening results from different channels are used to determine key features consistent across channels, where, Figure 10 (a) is a comparison diagram of the micro-microphones (MICs) of the cochlear sensor. Figure 10 (b) is a comparison chart of external sensor MICs. Figure 10 (c) is a comparison diagram of the internal sensors (MIC) of the rebound hammer.

[0065] For example, a dimensionless screening threshold of 0.5 is set. For the cochlear channel, if the maximum mutual information coefficient of the short-time average amplitude skewness is 0.63, the feature is retained in the cochlear channel; if the maximum mutual information coefficient of a certain frequency domain statistic is 0.41, the feature is not retained in the cochlear channel. The same feature also needs to be evaluated separately in the external sound pressure channel and the rebounder's built-in channel. This method can capture the nonlinear relationship between the prestressed label and the acoustic feature, without requiring them to be linearly related.

[0066] S330, Select candidate features that satisfy the correlation condition in multiple channels as key features, and concatenate them to form the fused feature vector.

[0067] Cross-channel key features are obtained by the intersection of the screening results from the three channels. Unlike selecting only highly correlated features from a single channel, intersection screening requires that the same candidate feature exhibits correlation with the prestressed label in multiple propagation paths. This condition can exclude highly correlated features caused solely by a sensor installation location, local contact state, or occasional noise, allowing the input features to more stably reflect the acoustic differences dominated by tension changes.

[0068] For example, after screening, 11 cross-channel consistent key features are obtained, including 5-dimensional time-domain features, 1-dimensional frequency-domain features, 1-dimensional Mel frequency cepstral features, 2-dimensional gamma-pass filter cepstral features, and 2-dimensional power-normalized cepstral features. Time-domain features may include short-time average amplitude skewness, peak index, impulse index, margin index, and average short-time autocorrelation impulse index. Frequency-domain features may include spectral stretching and kurtosis indices. Mel frequency cepstral features may include the standard deviation of the 6th-order coefficients of the first-order difference. Gamma-pass filter cepstral features may include the mean and median of the static 2nd-order coefficients. Power-normalized cepstral features may include the mean of the static 3rd-order coefficients and the median of the 1st-order coefficients of the first-order difference. The 11-dimensional key features from the three sensors are concatenated at the feature level to form a 33-dimensional fused feature vector.

[0069] In one verification embodiment, 720 valid real samples from the three channels were divided into 504 training samples and 216 test samples in a 7:3 ratio. Support vector regression models and Gaussian process regression models were trained using single-sensor features and three-sensor fusion features, respectively. The coefficient of determination (COP) of the single-channel cochlear sensor on the test set was approximately 0.89, with a root mean square error (RMSE) of approximately 18.97 kN; the COP of the single-channel external sound pressure sensor on the test set was approximately 0.88, with an RMS error of approximately 18.80 kN; and the COP of the single-channel built-in sensor of the rebound device on the test set was approximately 0.77, with an RMS error of approximately 30.76 kN. After three-sensor fusion, the COP of the Gaussian process regression model on the test set was approximately 0.93, with an RMS error of approximately 13.62 kN. It is easy to understand that cross-channel consistent features and three-sensor fusion can provide a more stable input for prestress recognition.

[0070] In another embodiment, if the sensor arrangement at a certain engineering site results in fewer than a preset number of intersection features across the three channels, the sample processing unit first maintains the principle of cross-channel consistency and then introduces candidate supplementary features based on channel reliability. Channel reliability can be evaluated by the number of saturation events during acquisition, peak alignment deviation, environmental noise level, and historical rejection rate. For example, if the cochlear sensor does not saturate within a detection cycle, has a peak alignment deviation of 0.4 ms, and a historical rejection rate of 1.5%, its channel reliability is higher than that of the external sound pressure sensor. If the intersection of the three channels yields only 9 key features, the sample processing unit can select 2-dimensional supplementary features from the shared candidate set of the cochlear sensor and the external sound pressure sensor to ensure that the fused feature vector still maintains sufficient input dimensions. The supplementary feature needs to meet two conditions: first, the maximum mutual information coefficient between the feature and the prestressed label is not lower than the cross-channel screening threshold; second, the feature was not dominated by abnormal impact samples in the most recent quality control. Through the above processing, feature selection can adapt to differences in the contact state of sensors at the site while maintaining physical consistency.

[0071] The sample processing unit can also record the source channel and statistical meaning of each key feature. For example, the first 11 dimensions of the 33-dimensional fused feature vector come from the cochlear sensor, the middle 11 dimensions from the external sound pressure sensor, and the last 11 dimensions from the internal sensor of the rebound sensor. If the 3rd dimension is the peak index, the 12th dimension is the peak index corresponding to the external sound pressure sensor, and the 23rd dimension is the peak index corresponding to the internal sensor of the rebound sensor, then the three dimensions maintain a fixed position in the model input. This fixed positional relationship allows the subsequent generation enhancement unit and recognition evaluation unit to track the channel source of each feature dimension. If a channel is replaced during on-site detection, the system can re-perform channel calibration and feature screening, and generate a new feature version number in the sample storage unit, avoiding the mixing of different versions of fused features.

[0072] S400, using the effective impact response samples as real samples, train a generative data augmentation network with physical constraints of steel strand dynamics to obtain virtual acoustic samples.

[0073] When the size of the real sample is limited, relying solely on real samples to train the regression model can easily lead to insufficient generalization. While ordinary generative adversarial networks (GANs) can generate additional waveforms, if the generated samples only satisfy statistical appearance similarity and do not satisfy the spectral peak, attenuation rate, and frequency band energy relationships of the steel strand anchoring system, it may introduce augmented data inconsistent with the actual dynamic process. This embodiment constructs a one-dimensional physically constrained generative data augmentation network, so that the generated samples are simultaneously guided by the distribution of real samples and dynamic physical constraints. The training process of the generative network is executed according to S410 to S430, and the comparison between the generated samples and real samples is illustrated through time-domain waveforms, peak-to-mean error, and frequency-domain distribution data. Specifically: S410 performs batch standardization on the extracted real samples.

[0074] During network training, a batch is first drawn from the real samples of the current duct or the current tension level. For example, 8 real samples are randomly drawn in each round to form a real sample matrix. The average value was obtained for this batch. and standard deviation And the samples are standardized as follows:

[0075] The generator outputs standardized virtual samples. Then, perform de-standardization:

[0076] This round-by-round standardization strategy allows the generator to encounter real samples with different distributions in different rounds, reducing the risk that the generator only memorizes waveforms from fixed batches. For example, if the mean of a sample batch in channel 1 of a certain round is 0.004 V and the standard deviation is 0.126 V, and the standardized peak value of the generator output is 1.85, then the destandardized peak value is approximately 0.237 V. If the mean and standard deviation of the sample batch change in the next round, the generator output will be mapped back to the corresponding waveform amplitude range according to the statistical characteristics of that batch.

[0077] S420 introduces a physical constraint loss consisting of the spectral peak, decay rate, and band energy into the adversarial training loss.

[0078] The generator employs a one-dimensional transposed convolutional structure, and the discriminator employs a one-dimensional convolutional structure. In one embodiment, the generator input is a 100×1 random vector, which passes through three transposed convolutional layers sequentially. The kernel size of each of the three transposed convolutional layers is 25×1, the stride is 5, and the number of output channels is 64, 32, and 1 respectively. The activation functions of the first two layers are leaky rectified linear units (RCUs), and the activation function of the last layer is a hyperbolic tangent function. The generator outputs a waveform of length 12500×1, and the first 4097 points are truncated as virtual acoustic samples. The discriminator input is a 4097×1 waveform, which passes through three convolutional layers sequentially. The kernel size of each layer is 25×1, the stride is 5, and the number of output channels is 32, 64, and 1 respectively. The activation functions of the first two layers are leaky rectified linear units (RCUs), and the final output is a real score after passing through a fully connected layer.

[0079] The adversarial loss adopts the Wasserstein form and is combined with gradient penalty to stabilize training. The total loss can be expressed as:

[0080] in, and These are the scores output by the discriminator for generated samples and real samples, respectively; during batch training, the average of the discriminator outputs for the corresponding batch can be used. The term is Wasserstein's adversarial loss. The gradient penalty term is determined by the deviation of the discriminator gradient norm on the difference samples. For gradient penalty weights, here . Used to constrain the main peak frequency of the spectrum between generated samples and real samples. Constrained impact attenuation rate, It is used to constrain the energy distribution in different frequency bands. In addition, the physical constraint term cannot replace adversarial training, but rather ensures that the generated waveforms approximate the real distribution while maintaining the dynamic characteristics of the steel strand anchoring system.

[0081] The spectral peak constraint can be obtained by the deviation of the peak frequencies between real samples and generated samples. For example, after performing a Fast Fourier Transform on a real sample, its main energy peak is located at 3120 Hz; the main peak of a generated sample at the same tensile level is located at 3185 Hz, then the spectral peak deviation between the two is 65 Hz. If the average main peak of 8 real samples in the same batch is 3100 Hz, and the average main peak of the generated batch is 3132 Hz, then the spectral peak deviation of this batch is 32 Hz. This deviation value is entered into... This is used to limit the main vibration frequency band of the generated sample. The decay rate constraint can be obtained based on the envelope curve after the impact peak. For example, if the envelope of the real sample decays from 0.82 V to 0.31 V within 20 ms after the peak, and the envelope of the generated sample decays from 0.80 V to 0.42 V in the same time period, then the decay of the generated sample is slower. It will increase. Bandwidth energy constraints can divide the response frequency band into multiple sub-bands and compare the energy proportion of each sub-band. If the energy proportion of the real sample in the 2000 Hz to 4000 Hz sub-band is 0.64, and the corresponding proportion of the generated sample is 0.58, then this difference will be considered. And so on.

[0082] The three physical constraints mentioned above correspond to the frequency position, energy attenuation, and frequency band distribution in the impact response of the steel strand anchorage system, respectively. The frequency position is related to the tension state of the steel strand, the anchor clamping conditions, and the local boundary stiffness; the energy attenuation is related to damping, contact state, and sensor coupling state; and the frequency band distribution is related to the impact propagation path and the structure of the anchorage system. If only adversarial loss is used, the generator may learn waveforms that are visually similar but deviate from the real samples in the aforementioned physical quantities. By incorporating all three constraints into the total loss, the generator needs to achieve consistency between the waveform distribution and the physical response, making the virtual samples more suitable as training data for the regression model.

[0083] S430 employs a weighting strategy that adjusts based on training phases and generated quality feedback to fuse various losses.

[0084] The physical constraint weights are adaptively adjusted in stages. The initial training phase consists of rounds 1 to 500. , and All values ​​are set to 0.01, allowing the adversarial loss to dominate training, and the generator first learns the basic waveform envelope of the real signal. During the mid-training period (rounds 501 to 1500), the weights of the three physical constraints are gradually increased to the target value according to a cosine annealing curve. For example, Increase to 0.1, Increase to 0.05, Increase to 0.05. In the later stages of training, starting from round 1501, introduce generation quality feedback, evaluating the physical parameter deviation of the current batch of generated samples every 100 rounds. If the deviation of the main peak of the spectrum exceeds 5%, temporarily increase... If the attenuation rate deviation exceeds 10%, then temporarily increase the value. If the weight is below the corresponding deviation threshold for three consecutive evaluation periods, the corresponding weight will gradually fall back to the target value.

[0085] For example, under the fixed-weight strategy, the generated sample has a spectral bias of 45.2 Hz, a decay rate bias of 12.3%, a bandgap energy ratio bias of 0.082, and requires 2100 convergence rounds. Under the adaptive-weight strategy, the generated sample has a spectral bias of 28.7 Hz, a decay rate bias of 7.8%, a bandgap energy ratio bias of 0.051, and requires 1580 convergence rounds. This validation result shows that after the physical constraint weights change with the training phase and the generation quality, the generated samples more closely approximate the real impact response in terms of spectrum, decay, and energy distribution.

[0086] A separate generative network is trained for each channel to avoid contamination of the same generated distribution by differences in boundary conditions and sensor placement between different channels. For example, 1000 virtual samples are generated for each channel, each containing synchronized virtual signals from the cochlear sensor, external sound pressure sensor, and internal sensor of the rebound sensor. A total of 3000 virtual samples are generated for the three channels. Through independent channel training and physically constrained generation, virtual samples can be used to augment the training set rather than replace real samples.

[0087] In one alternative implementation, a virtual-real domain consistency process can be introduced before virtual samples are incorporated into the regression model training. This process uses a domain adversarial learning network to extract domain-invariant features, resulting in more similar distributions of real and virtual samples in the feature space. The input to the domain adversarial learning network is a 33-dimensional fused feature vector, and the output is the domain-invariant features used for regression training. During training, the feature extraction branch serves pre-stress recognition, while the domain discrimination branch distinguishes between the real and virtual domains, forming adversarial training through gradient backpropagation. This alternative process reduces the distributional differences between virtual and real samples without altering the process by which the physically constrained generative data augmentation network generates samples.

[0088] S500, using the real samples and the virtual acoustic samples to train a Gaussian process regression recognition model, and obtaining the prestress recognition result and prediction variance based on the fused feature vector to be tested.

[0089] After generating virtual samples, the real samples and virtual acoustic samples are jointly organized into an enhanced training set. The Gaussian process regression identification model takes a 33-dimensional fused feature vector as input and outputs tension force or effective prestress value. To verify the effectiveness of generative data augmentation, the Gaussian process regression prediction results for condition III are as follows: Figures 11 to 13 As shown; where, Figure 11 The horizontal axis represents the "sample number", and the vertical axis represents the "prestress (kN)". Figure 12 The horizontal axis represents "actual prestress (kN)" and the vertical axis represents "predicted prestress (kN)". Figure 13 The x-axis represents the "sample number," and the y-axis represents the "prediction residual (kN)." To verify the effectiveness of multi-sensor feature fusion, the prediction results of the three-sensor fusion Gaussian process regression model are as follows: Figures 14 to 16 As shown. Among them, Figure 14 The horizontal axis represents the sample number, and the vertical axis represents the prestress (kN). Figure 15 The horizontal axis represents "actual prestress (kN)" and the vertical axis represents "predicted prestress (kN)". Figure 16 The horizontal axis represents the "sample number", and the vertical axis represents the "prediction residual (kN)".

[0090] S510, input the enhanced training set into the Gaussian process regression model.

[0091] In one embodiment, a total of 720 valid real samples are generated from the three channels, and 1000 virtual samples are generated for each channel, for a total of 3000 virtual samples. For the three-sensor fusion recognition model, the 11-dimensional key features corresponding to the cochlear sensor, external sound pressure sensor, and rebound sensor are concatenated at the feature level to form a 33-dimensional fusion feature vector. The model output label is tension force or effective prestress value. The Gaussian process regression model uses a radial basis function kernel function, and the hyperparameters are obtained through five-fold cross-validation. This model describes the mapping relationship between the fusion feature vector and the prestress value in a nonparametric Bayesian manner, and can provide both the predicted mean and predicted variance under small sample and nonlinear response conditions.

[0092] For example, to evaluate the effect of generative data augmentation separately, three training scenarios were compared using data from one channel, with the same 11-dimensional key features, Gaussian process regression model structure, and kernel function parameters used in each scenario. Scenario I used only virtual samples for training and real samples for independent testing; Scenario II used only real samples for training and testing; Scenario III combined real and virtual samples to form an augmented dataset, and divided the training and testing sets in a 3:1 ratio. In this channel validation, Scenario I used 1000 virtual samples as the training set and 248 real samples as the independent testing set; Scenario II divided the 248 real samples into 186 training samples and 62 testing samples; Scenario III combined the 248 real samples and 1000 virtual samples to form 1248 augmented samples, and divided them into 936 training samples and 312 testing samples. The coefficient of determination (COP) for the test set under test condition I was 0.78, and the root mean square error (RMSE) was 26.25 kN; for test condition II, the COP was 0.92, and the RMSE was 14.62 kN; and for test condition III, the COP was 0.96, and the RMSE was 10.59 kN. The results show that virtual samples cannot completely replace real samples, but by appropriately introducing virtual samples into the real sample set, they can supplement the sparse regions in the real sample distribution and improve the model's prediction accuracy and generalization ability.

[0093] S520 uses kernel functions and cross-validation to obtain model hyperparameters.

[0094] The kernel parameters of the Gaussian process regression model are obtained through cross-validation. During training, the training set is divided into multiple folds, with each fold serving as the validation set in turn, and the remaining folds used as the fitting data. For each set of candidate kernel parameters, the error and prediction variance distribution on the validation set are recorded. Finally, the parameters with smaller validation errors and stable prediction variance distributions across multiple folds are selected. This selection method reduces the randomness introduced by a single partition.

[0095] For example, the radial basis function kernel function can be written as:

[0096] in, and This is the fused feature vector of the two samples. For signal variance, For length measurement. To adapt to Word document conversion, the exponential function can be written as... The equivalent form is as follows. In implementation, the length scale determines the model's sensitivity to feature changes, while the signal variance determines the range of function output fluctuations. If the length scale is too small, the model is easily sensitive to local noise; if the length scale is too large, the model may not be able to distinguish feature differences under different tension levels.

[0097] S530, output the prestress identification result and the corresponding prediction variance.

[0098] For the steel strand to be tested, a 33-dimensional fused feature vector is obtained according to S100 to S330, and then input into the trained Gaussian process regression recognition model. The model outputs the predicted mean. As a result of prestress identification, the predicted variance is output. This serves as a basis for confidence assessment. If the prediction variance is small, it indicates that the sample to be tested is near the model training distribution; if the prediction variance is large, it indicates that the sample to be tested may be located in a region where the training samples are insufficiently covered or in an abnormal input region.

[0099] For example, the test set contains 216 samples, covering five tension levels: 40 kN, 80 kN, 120 kN, 160 kN, and 200 kN. For the 40 kN load case, the model prediction is 40.12 kN with an absolute error of 0.12 kN and a relative error of 0.30%; for the 80 kN load case, the model prediction is 79.85 kN with an absolute error of 0.15 kN and a relative error of 0.19%; for the 120 kN load case, the model prediction is 119.92 kN with an absolute error of 0.08 kN and a relative error of 0.07%; for the 160 kN load case, the model prediction is 160.35 kN with an absolute error of 0.35 kN and a relative error of 0.22%; and for the 200 kN load case, the model prediction is 199.62 kN with an absolute error of 0.38 kN and a relative error of 0.19%.

[0100] In a field application example, after receiving a sample to be tested, the identification and evaluation unit first verifies the sample index, number of channels, and feature version number. If the sample to be tested comes from the bridge anchorage end, and the anchorage end is only allowed to be detected for a short time, the system can first extract the fused feature vector using the feature screening table obtained from the most recent laboratory calibration, and then input the fused feature vector into the Gaussian process regression identification model. For example, in the 33-dimensional fused feature vector of a sample to be tested, the peak index of the cochlear channel is 2.31, dimensionless; the kurtosis index of the spectral extension of the external sound pressure channel is 4.08, dimensionless; and the median of the power normalized cepstrum of the built-in sensor of the rebound hammer is 0.37, dimensionless. The predicted mean of the model output is 118.6 kN, the predicted standard deviation is 2.8 kN, and the corresponding predicted variance is 7.84 kN². The system displays the predicted mean and the corresponding predicted variance at the same time, so that the subsequent confidence assessment can determine whether the predicted result can be used as an engineering inspection record value.

[0101] When continuous testing is required at the same anchorage end, the identification and evaluation unit can retain the fused feature vector and model output of each test. For example, maintenance personnel perform three consecutive impact acquisitions at the same steel strand location. The three predicted mean values ​​are 118.6 kN, 119.1 kN, and 118.4 kN, with predicted standard deviations of 2.8 kN, 2.6 kN, and 3.0 kN, respectively. If all three results pass the rejection rule, the system can use the median of the three predicted mean values ​​as the recorded value for that inspection, and the maximum of the three predicted standard deviations as the variance description for that record. This processing avoids the direct impact of random errors from a single impact on the inspection record, while retaining the predicted variance information output by the Gaussian process regression model.

[0102] When outputting detection records, the identification and evaluation unit can also provide a sample coverage description. This description explains the relationship between the fused feature vector of the test sample and its neighboring samples in the training set. For example, if the average distance between the fused feature vector of a test sample and its five neighboring samples in the training set is 0.18 (dimensionless), and these five neighboring samples come from tension levels of 120 kN and 160 kN respectively, then the test sample is within the training sample coverage area. If the average distance between another test sample and its neighboring samples is 0.76 (dimensionless), and all neighboring samples come from a single channel, the system indicates in the detection record that the sample coverage is narrow. This description does not replace the prediction variance, but it helps maintenance personnel understand why the model gives a high or low confidence level for a given input.

[0103] S600, the confidence level of the prestress identification result is evaluated based on the predicted variance, and rejection is triggered if the evaluation fails.

[0104] In engineering testing, issues such as increased environmental noise, improper operation, loose sensors, waveform truncation, and abnormal amplitude scaling may occur. If the identification model provides usable results for all inputs, anomalous samples may lead to incorrect evaluations. The prediction variance output by the Gaussian process regression model can be used as a source of confidence assessment; in this embodiment, it is used for confidence evaluation and rejection handling.

[0105] S610, based on the predicted variance, form an absolute variance evaluation.

[0106] Absolute variance evaluation compares the predicted variance of the test sample with the average predicted variance of the training set. The evaluation rule can be expressed as:

[0107] in, The variance of the predicted sample is denoted as . The average prediction variance of the training set samples. This is the absolute variance threshold. For example, Take 2.0, dimensionless. If the prediction variance of a sample to be tested is 0.018 and the average prediction variance of the training set is 0.006, then the ratio is 3.0, which exceeds the threshold of 2.0, and the sample is rejected.

[0108] S620, relative interval evaluation is formed based on the relative width of the confidence interval.

[0109] Relative interval evaluation is used to handle differences in absolute variance caused by different magnitudes of predicted values. The evaluation rule can be expressed as:

[0110] in, To predict the mean, To predict the standard deviation, This is a relative interval threshold. For example, Let's take 5%. If a sample has a predicted mean of 100 kN and a predicted standard deviation of 4 kN, then the 95% confidence interval half-width is approximately 7.84 kN, and the relative width is 7.84%, which exceeds 5%, triggering rejection for that sample. If another sample has a predicted mean of 160 kN and a predicted standard deviation of 2 kN, then the relative width is 2.45%, and rejection is not triggered under this rule.

[0111] S630: When any evaluation exceeds the corresponding threshold, a rejection flag is output and a prompt is given for re-collection or manual review.

[0112] When either the absolute variance evaluation or the relative interval evaluation meets any rejection condition, the system outputs a rejection flag and prompts "Insufficient reliability of the detection result; re-collection or manual verification is recommended." Rejected samples do not enter the regular prestressed output process, but their channel waveforms, fused feature vectors, prediction variance, and rejection reasons are saved to the sample storage unit for subsequent sensor maintenance, anomaly source analysis, and model iteration. After being saved, rejected samples only appear in the anomaly statistics field of the drift observation window; they do not enter the incremental candidate window or participate in subsequent incremental updates as pseudo-label samples. If the proportion of rejected samples within the same drift observation window exceeds a preset threshold, the system prioritizes identifying it as an abnormal acquisition quality or sensor status, prompting for re-collection or manual verification, rather than directly triggering model incremental updates.

[0113] For example, 10% of outlier samples were mixed into the test set. These outlier samples were obtained by adding high-frequency noise, truncating the waveform, and scaling the amplitude. Without the rejection mechanism, the outlier sample pass rate was 100%, the normal sample rejection rate was 0%, and the overall root mean square error was 14.23 kN. With the rejection mechanism enabled, the outlier sample pass rate was 8.7%, the normal sample rejection rate was 2.1%, and the overall root mean square error was 9.86 kN. This validation demonstrates that the rejection mechanism can utilize the prediction variance of Gaussian process regression to intercept outlier inputs and use rejected samples as the basis for subsequent maintenance and updates.

[0114] Post-processing of rejected samples does not directly discard the original data. The system first records the rejection trigger type, including absolute variance trigger, relative interval trigger, and both; then it records the channel quality information at the time of triggering, including the peak amplitude, root mean square value, peak alignment deviation, and low-frequency residue after filtering for the three channels. For example, if a sample's prediction variance increases due to waveform truncation in the external sound pressure channel, the system records this sample as a relative interval trigger and marks the effective window length of the external sound pressure channel as insufficient. After maintenance personnel re-acquire data, if the new sample passes quality control and rejection rules, it enters the normal detection record, while the original rejected sample is retained as abnormal data. If the same sensor triggers rejection consecutively in multiple samples, the system can prompt to check the sensor installation status or the acquisition cable connection status.

[0115] For rejected samples that still have analytical value, the sample storage unit can use them as candidate data for model iteration. If manual verification confirms that the prestressed label of the sample is reliable, but there are interpretable disturbances in the acquisition environment, the sample can be included in the subsequent retraining data pool. If manual verification determines that the sample was caused by misoperation, sensor detachment, or acquisition instrument saturation, the sample is only recorded as an anomaly type and does not participate in model updates. By distinguishing the destination of rejected data, the system avoids anomaly samples from contaminating the training set and can collect field anomaly types, providing a basis for subsequent sensor deployment and quality control rule adjustments.

[0116] S700 continuously monitors the temporal changes of the prediction variance, generates drift samples when feature drift is detected, and incrementally updates the Gaussian process regression identification model.

[0117] During long-term operation, factors such as strand slack, temperature changes, anchor slippage, and corrosion can cause slow changes in acoustic response characteristics. When a fixed-train model faces these changes, its prediction variance may gradually increase, and its recognition error may grow. This embodiment utilizes the prediction variance of a Gaussian process regression model to construct drift detection, drift range estimation, drift virtual sample generation, incremental fine-tuning, and lifetime loop mechanisms.

[0118] The S710 continuously records the prediction variance during the detection process and determines whether feature drift has occurred.

[0119] After system deployment, each time a detection record is generated via S100 to S500, the predicted mean, predicted variance, and rejection status are recorded. The sample storage unit maintains a drift observation window and an incremental candidate window. The drift observation window has a length of 20 and stores the 20 most recent detection records that have completed quality control, including samples that passed the rejection rule and samples that triggered rejection. Samples that triggered rejection are only included in the anomaly ratio statistics based on the rejection flag, predicted variance, channel quality information, and rejection reason; they are not used as subsequent incremental update samples. The incremental candidate window only stores high-confidence, unlabeled field samples that did not trigger rejection and whose predicted variance is lower than the median variance within the window.

[0120] When the drift update unit determines drift, it first calculates the rejection rate in the drift observation window. If the rejection rate exceeds a preset threshold, it determines that the current window is more likely to have abnormal acquisition quality or sensor status, and the system prompts for re-acquisition or manual review, without directly executing S720 to S740. If the rejection rate does not exceed the preset threshold, the drift update unit calculates the monotonicity index of the prediction variance on the sample subsequences that have not triggered rejection. If the monotonicity index M is greater than 0.6, and the mean variance of the subsequence exceeds 1.5 times the mean variance of the initial training set, then feature drift is determined to have occurred.

[0121] For example, in 20 consecutive detections of a steel strand, 2 rejections were triggered, a rejection rate of 10%, which is lower than the 30% threshold. After the drift update unit removes these 2 rejected records, it calculates the Kendall rank correlation coefficient on the remaining 18 non-rejected records. If the prediction variance of these 18 records gradually increases from 0.004 to 0.011, the Kendall rank correlation coefficient is 0.72, and the mean variance is 1.8 times the mean variance of the initial training set, then the current detection sequence is determined to have time-varying feature drift. If 8 rejections are triggered in 20 consecutive detections, a rejection rate of 40%, the system does not directly interpret this window as feature drift, but instead prompts to check the impact operation, sensor coupling status, and acquisition channel quality.

[0122] S720, the drift range is estimated with the current prestress prediction value as the center.

[0123] After determining the drift, the drift update unit estimates the current prestress state based on the predicted mean sequence within the window. Let the average of the 20 predicted means within the window be... The standard deviation is Then the drift range Possible options are:

[0124] For example, if the average of 20 predicted values ​​within the window is 100 kN and the standard deviation is 5 kN, then the drift range is 90 kN to 110 kN. This range is not the final detection conclusion, but rather the target prestress range used to generate the drift virtual sample. By limiting the generated range to the vicinity of the current detection window, interference from irrelevant tension level samples on incremental updates can be reduced.

[0125] S730 uses a trained generative data augmentation network to generate drift samples.

[0126] The drift update unit invokes a trained physical constraint generative data augmentation network to generate multi-channel virtual acoustic samples within the drift range. For example, 200 prestress values ​​are uniformly selected within the range of 90 kN to 110 kN, and synchronous virtual signals for the cochlear sensor, external sound pressure sensor, and internal sensor of the rebound device are generated for each prestress value. Because the generative network includes spectral peaks, attenuation rates, and bandgap energy constraints, the drift virtual samples are still limited by the dynamic characteristics of the steel strand anchoring system.

[0127] To prevent drift samples from being entirely determined by the generator network, the system also selects recent high-confidence unlabeled field samples from the incremental candidate window. These samples are not real samples already labeled during training, nor do they have on-site measured prestress labels. The system saves the model's predicted mean as a pseudo-label only when the sample meets the following conditions: the rejection rule is not triggered, the prediction variance is lower than the median variance within the window, the dispersion of the predicted mean collected multiple times within the same inspection cycle is lower than a preset threshold, and the main physical features do not violate the spectral peak and attenuation constraints. The pseudo-label is recorded together with the prediction variance, rejection status, and physical consistency check results, and its training weight is lower than that of manually labeled samples. The virtual drift samples and recent high-confidence unlabeled field samples together constitute the incremental update data.

[0128] S740, the Gaussian process regression identification model is incrementally updated using drift samples and recent high-confidence unlabeled field samples.

[0129] Incremental updates employ a Bayesian update method with confidence weights. Original manually labeled training samples are retained as baseline data and are not overwritten by high-confidence unlabeled field samples. For sparse Gaussian process regression, only the induced point parameters and local kernel parameters can be updated without retraining the entire model. High-confidence unlabeled field samples participate in the update as pseudo-labeled samples; their pseudo-labels are derived from the predicted mean output by the model under low prediction variance conditions and are not recorded as true prestressed labels. The system assigns update weights to pseudo-labeled samples that are higher than the observation noise of manually labeled samples or lower than those of manually labeled samples, ensuring they are only used for local adaptation of the current drift interval. This process reduces update time and decreases reliance on manually added labeled samples.

[0130] For example, a frequency shift from 0 Hz to 50 Hz is applied to the test set to simulate acoustic feature drift after 5 years of operation. Three strategies are compared: the first strategy uses the original model directly without compensation; the second strategy re-collects labeled samples after drift and performs full retraining; the third strategy employs the lifelong learning closed loop of this embodiment. Before drift, the root mean square error (RMSE) of the model was 9.86 kN, and after drift, the RMSE of the uncompensated strategy increased to 18.73 kN. With the full retraining strategy, after collecting more than 50 new samples, the compensated RMSE was 10.15 kN. In this embodiment, under the condition of zero newly added manually labeled samples, only pseudo-labels of unlabeled field detection samples, prediction variance weights, and physical consistency screening results are used for local incremental updates, resulting in a compensated RMSE of 10.21 kN. This comparison shows that drift sample generation and incremental updates based on prediction variance feedback can restore the model's adaptability to the current state with less human intervention. The aforementioned zero newly added manually labeled samples do not indicate that the field-tested samples have genuine prestress labels, but rather that the system has not obtained additional prestress label values ​​that have been manually loaded or manually verified. High-confidence unlabeled field samples only participate in the update as pseudo-labeled samples, and their impact is limited by prediction variance weighting, physical consistency checks, consistency checks before and after updates, and model version rollback mechanisms, thereby avoiding the unconstrained use of model prediction values ​​as genuine labels for repeated bootstrapping.

[0131] During incremental updates, the drift update unit also performs a model consistency check before and after the update. Before the update, the system records the predicted mean and predicted variance of the old model on the most recent high-confidence unlabeled field samples; after the update, the system re-inputs the new model with the same batch of samples and compares the changes in predicted mean and variance. If the predicted mean changes abnormally after the update, or the predicted variance of samples outside the drift interval increases, the update is marked as pending review and the online model is not immediately replaced.

[0132] For example, if the median mean predicted by the old model for the most recent 10 high-confidence samples is 101.2 kN, and the median mean predicted by the new model is 101.6 kN, with a difference of 0.4 kN, and the median prediction variance decreases from 0.007 to 0.005, then the update is acceptable. If the median mean predicted by the new model becomes 108.4 kN, this change exceeds the drift range estimate, and the system retains the old model and prompts for manual review.

[0133] To mitigate the risk of model forgetting during continuous incremental updates, the sample storage unit retains summaries of the initial training set and historical high-confidence samples. The initial training set summary may include the feature mean, feature variance, and representative sample index for each tension level; the historical high-confidence sample summary may include samples that passed the rejection rule within the most recent detection periods. During incremental fine-tuning, drifting virtual samples, recent high-confidence unlabeled field samples, and summary samples are all updated together. This process ensures that the model retains its ability to recognize existing tension levels and existing duct conditions while adapting to the current drift interval.

[0134] The incremental update record also includes the trigger time, drift window, drift range, number of generated samples, number of unlabeled field samples participating in the update, and model versions before and after the update.

[0135] For example, an update record shows the trigger time as the 8th inspection, the drift window as the last 20 inspections, the drift range as 90 kN to 110 kN, the generation of 200 virtual drift samples, and the inclusion of 12 recent high-confidence unlabeled field samples in the update. The model version is updated from version 3 to version 4. If the version 4 model consistently passes the rejection rules and the prediction variance decreases in subsequent inspections, the version 4 model is retained. If the version 4 model leads to an increase in the rejection rate, the system can roll back to the version 3 model and mark the updated data as pending review. Through version recording and rollback mechanisms, the system can maintain traceability between unattended inspections and manual review.

[0136] In another implementation, the drift update unit can set different update cycles based on the service environment of different steel strands. For exposed anchorage areas of bridges with large temperature variations, the system can use a shorter monitoring cycle and increase the sampling frequency after drastic temperature changes; for indoor components or components less affected by the environment, the system can use a longer monitoring cycle. Regardless of the cycle used, drift determination is still triggered by the temporal change of the prediction variance, rather than simply by the service time. This avoids unnecessary updates when the model is still adapted to the current state and allows for timely initiation of the compensation process when abnormal drift occurs in the short term.

[0137] When multiple steel strands are managed by the same system simultaneously, the drift update unit maintains a sliding window and model version for each strand. For example, strand A is located in an exposed area, while strand B is located indoors, and their temperature disturbances and corrosion risks differ. The system does not directly use drift samples from A for incremental updates of B; instead, it only shares sensor calibration records and quality control rules in the sample storage unit. If A drifts and completes its fourth model update, B can still maintain its second model until its own prediction variance sequence meets the drift trigger condition. This isolation mechanism prevents cross-contamination of the service conditions of different components while preserving the acquisition specifications and data tracking methods under a unified system.

[0138] S750, after the update is complete, returns to drift detection and forms a permanent loop.

[0139] After the incremental update is completed, the system continues to perform routine detection and writes the new prediction variance, prediction mean, and rejection status to the sample storage unit. If the prediction variance continues to rise again within a subsequent window, the drift update unit executes steps S720 to S740 again. If the prediction variance recovers to near the initial training set within the window, the system maintains the current model and continues monitoring. This loop allows the model to gradually adjust with changes in service life, rather than remaining static after initial training.

[0140] For example, in a bridge operation and maintenance scenario, the impact acoustic response of a single steel strand can be collected monthly or quarterly. Each time, the system outputs the prestressing identification result, prediction variance, and rejection status. For rejected samples, maintenance personnel can re-collect the data or perform manual verification; for samples showing a drift trend but not rejected, the system can enter an incremental update process. This approach extends one-time model training into a closed loop of continuous detection, continuous evaluation, and continuous updating.

[0141] like Figure 2 As shown, this embodiment provides a prestress identification system for a single steel strand based on physical constraint generative data augmentation, comprising a signal acquisition unit, a sample processing unit, a generative augmentation unit, an identification evaluation unit, a drift update unit, and a sample storage unit. Each unit can be implemented by a sensor, a data acquisition instrument, a processor, a memory, and a program module running in the processor; among them, the signal acquisition unit is more inclined towards physical hardware, while the sample processing unit, generative augmentation unit, identification evaluation unit, and drift update unit can be implemented by the processor executing instructions in the memory.

[0142] The signal acquisition unit includes a rebound hammer, a cochlear sensor, an external sound pressure sensor, internal sensors of the rebound hammer, and a synchronous data acquisition device. The rebound hammer applies impact excitation to the anchoring system, the cochlear sensor acquires solid-borne sound signals, the external sound pressure sensor acquires airborne sound signals, and the internal sensors of the rebound hammer acquire mechanical vibration signals from the excitation source. The synchronous data acquisition device provides a unified sampling time reference for multiple channels and sends the response signals of multiple channels under the same impact event to the sample processing unit.

[0143] After receiving the multi-source acoustic response signals, the sample processing unit performs downsampling, high-pass filtering, pre-emphasis, sample quality control, peak alignment, and fixed window truncation to form valid impulse response samples. Subsequently, the sample processing unit extracts 705-dimensional candidate features from the valid impulse response samples, uses the maximum mutual information coefficient to obtain 11-dimensional key features consistent across channels, and concatenates them to form a 33-dimensional fused feature vector. The sample processing unit writes the fused feature vector, sample labels, and sample indices into the sample storage unit.

[0144] The generation and enhancement unit receives valid impact response samples and sample grouping information, and trains a physically constrained generative data augmentation network for each channel or prestress level. During training, the generation and enhancement unit performs round-by-round normalization, integrates adversarial training loss, gradient penalty, and physical constraint loss, and employs a phased adaptive weighting strategy. After training, the generation and enhancement unit outputs multi-channel synchronous virtual acoustic samples and sends them to the sample storage unit. The recognition and evaluation unit reads real samples, virtual samples, and fused feature vectors from the sample storage unit, trains a Gaussian process regression recognition model, and outputs the prestress recognition result, prediction variance, and rejection flag during the detection phase.

[0145] The drift update unit reads the prediction variance continuously output by the recognition and evaluation unit, maintains the drift observation window and incremental candidate window, and determines whether feature drift has occurred based on the rejection rate, monotonicity index, and mean variance. When drift occurs, the drift update unit estimates the drift range, calls the generation and enhancement unit to generate drift virtual samples, and reads recent high-confidence unlabeled field samples and their pseudo-labels, prediction variance, and physical consistency check results from the sample storage unit. The drift update unit then sends these samples to the recognition and evaluation unit to perform incremental updates with confidence weights.

[0146] It should be noted that the sensor combination for the multi-source acoustic response signals can be adjusted without changing the technical approach of this application. The cochlear sensor can be replaced with other sensors capable of acquiring solid-borne sound or structural vibration responses, the external sound pressure sensor can be replaced with an airborne sound sensor with the same frequency band coverage capability, and the internal sensor of the rebound hammer can be replaced with an acceleration or force sensor mounted on the exciter. The replaced sensors must still maintain time synchronization under the same impact event and must be able to provide response information for multiple propagation paths for cross-channel consistent feature screening.

[0147] Preprocessing parameters can also be adjusted according to the sampling equipment and the engineering object. The fixed window length, sampling frequency, filter order, and cutoff frequency can be reset without disrupting the target frequency band response. If the main response frequency band of the structure under test differs from the 2000 Hz to 5000 Hz range in this embodiment, the target frequency band can be determined first through spectrum analysis, and then the downsampling and filtering parameters can be set accordingly. The pre-emphasis coefficient can also be adjusted within the range of 0.90 to 0.99, with the impact peak position, attenuation pattern, and subsequent identification error as selection criteria.

[0148] The internal structure of generative data augmentation networks can employ other one-dimensional sequence generation structures. The generator can use one-dimensional transposed convolution, upsampling convolution, or sequence decoding structures, while the discriminator can use one-dimensional convolution, spectral normalization convolution, or multi-scale discriminant structures. Regardless of the structure used, the generation process must retain the physical constraints of the steel strand anchoring system and constrain the spectral peak, attenuation rate, and bandgap energy. Gaussian process regression identification models can also be implemented using sparse Gaussian process regression, depending on the sample size, to reduce the incremental update overhead in long-term deployments.

[0149] Example 2 After executing S430 and before the identification and evaluation unit executes S500, the generation enhancement unit can also perform virtual sample generation based on the counterfactual path of adjacent tension levels and physical residual calibration, so as to supplement intermediate level samples located between adjacent tension levels when the real tension level samples are sparse, and write the supplemented samples into the sample storage unit in the form of training samples with observation noise.

[0150] Specifically, the sample processing unit selects two real samples of adjacent tension levels from the same channel, the same sensor installation location, and the same sampling conditions as anchor point samples.

[0151] Each anchor point sample includes a 33-dimensional fused feature vector and a physical parameter vector. The fused feature vector includes frequency domain peak value, bandwidth energy, attenuation envelope, time domain statistics, and time-frequency joint features. The physical parameter vector includes physical constraint parameters obtained from the steel strand length, linear density, boundary conditions, dominant frequency, and energy attenuation coefficient. To avoid bias in the interpolation path caused by different feature dimensions, the sample processing unit first standardizes each feature dimension based on the mean and standard deviation of the training set.

[0152]

[0153] In the formula, For the k-th dimension fusion feature, Let r be the physical parameter of dimension r; and These are the mean values ​​of the corresponding variables in the training set; and These are the standard deviations of the corresponding variables in the training set; and These represent the standardized fusion features and physical parameters, respectively; k = 1, 2, ..., 33, r = 1, 2, ..., R; R is the dimension of the physical parameter vector. The above standardization process ensures that data of different dimensions enter the same interpolation path, avoiding unreasonable dominance of the generation direction by features such as dominant frequency, energy, and attenuation due to their different numerical ranges.

[0154] When the intermediate tension level is located between two adjacent anchor point levels, the enhancement unit is generated to calculate the position coefficient of the intermediate tension level within the adjacent level interval, and based on this, the target fusion feature path point and target physical parameter path point of the intermediate level are obtained:

[0155]

[0156]

[0157] In the formula, and For the tension level of two adjacent anchor points, The intermediate tensioning level to be generated; η is In the interval [ , Position coefficient within ]; and This is the standardized fusion feature vector of the two anchor point samples. The target fusion feature path points for intermediate tension levels; and Given the standardized physical parameter vectors of the two anchor point samples, The target physical parameter path points for intermediate tensioning levels.

[0158] The generation enhancement unit inputs the vibration signal of the source sample into the first two one-dimensional convolutional layers of the discriminator and performs global average pooling on the convolution output to obtain the source sample encoding vector. Subsequently, the generation enhancement unit inputs the source sample encoding vector, tension level offset, target fusion feature path points, and target physical parameter path points into the generator to generate a counterfactual virtual vibration signal corresponding to the intermediate tension level. This counterfactual virtual vibration signal then goes through the same feature extraction process in S200 to obtain the generated fusion feature vector and the generated physical parameter vector.

[0159] To ensure that virtual samples not only closely approximate real samples in statistical distribution but also vary along the physical change path between adjacent tension levels, the generation enhancement unit incorporates a counterfactual path consistency loss in addition to the generator loss of S420:

[0160] In the formula, This is due to the loss of consistency in the counterfactual path. To determine the deviation between the generated fused feature vector and the target fused feature path points; To determine the deviation between the generated physical parameter vector and the target physical parameter path points; The monotonic constraint loss is the deviation of the filtered monotonic feature dimensions from the expected direction of change. , and These are the weighting coefficients for the three losses mentioned above. Monotonic constraints are applied only to the feature dimensions and physical parameter dimensions that have passed the Kendall rank correlation screening; dimensions that did not pass the screening only participate in the path deviation constraint.

[0161] The generation enhancement unit further calculates the residual score for each counterfactual virtual sample. This residual score is used to perform accept or discard processing on the virtual sample and to pass the uncertainty of the virtual sample to the subsequent Gaussian process regression model.

[0162]

[0163]

[0164] In the formula, Residual scores for counterfactual virtual samples; , and These are the residual weights for characteristic path deviation, physical path deviation, and monotonic constraint loss, respectively. The threshold for receiving virtual samples; Based on the variance of observational noise; The observation noise variance when the i-th counterfactual virtual sample is written into the Gaussian process regression model; Let be the residual score of the i-th counterfactual virtual sample. The larger, The larger the value, the smaller the impact of the virtual sample on the regression mean.

[0165] For example, at the same measuring point on the same steel strand, the sample storage unit already has real samples of two tension levels: 80 kN and 120 kN. The dominant frequency of the 80 kN sample is 2860 Hz, and the dominant frequency of the 120 kN sample is 3180 Hz. If a 100 kN sample is needed, the position coefficient is 0.5, and the target dominant frequency path point is 3020 Hz. The generation enhancement unit generates a 100 kN virtual vibration signal based on the standardized fusion characteristics, standardized physical parameters, and source sample encoding of the 80 kN and 120 kN anchor point samples. After processing by S200, the dominant frequency of this virtual vibration signal is 3035 Hz, and its attenuation coefficient, frequency band energy, and time-frequency ridge position are all included in the residual score calculation. When the receiving threshold is set to 1.0, and the residual score of the virtual sample is not greater than 1.0, the virtual sample is written to the sample storage unit; if the residual score is greater than 1.0, the virtual sample is discarded.

[0166] In another supplementary process, the generative augmentation unit can also generate virtual samples at three intermediate levels: 90 kN, 100 kN, and 110 kN, between 80 kN and 120 kN. For the 90 kN sample, the position coefficient is 0.25; for the 100 kN sample, it is 0.5; and for the 110 kN sample, it is 0.75. The target path points corresponding to each intermediate level are obtained by interpolation using the same anchor features and physical parameters. Based on this, the generative augmentation unit generates multiple candidate virtual samples and performs residual scoring, threshold filtering, and observation noise assignment on each. Through this processing, the generative augmentation unit can provide intermediate tensioning level training samples with physical path constraints for the S500 Gaussian process regression model without changing the original S420 discriminator loss, gradient penalty loss, frequency domain constraint loss, decay constraint loss, and energy constraint loss.

[0167] The above embodiments are used to illustrate the technical solution of this application, and are not intended to limit the scope of protection of this application. Those skilled in the art can make equivalent substitutions or combinations of sensor models, acquisition frequencies, window lengths, generator network structures, kernel function types, and drift determination parameters based on the disclosure of this application. As long as they still utilize multi-source acoustic response signals, physically constrained generative data augmentation, Gaussian process regression prediction variance assessment, and drift increment updates to identify the prestress of a single steel strand, they should all fall within the scope of protection of this application.

[0168] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit described above can be implemented in hardware.

[0169] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying the prestress of a single steel strand based on physical constraint generative data augmentation, characterized in that, include: Acoustic response signals from multiple sources were collected for the steel strand anchoring system under standardized impact excitation. The multi-source acoustic response signal is preprocessed to obtain effective impact response samples; Multi-scale acoustic features are extracted from the effective impact response samples, and key features are selected based on cross-channel consistency to form a fused feature vector; Using the effective impact response samples as real samples, a generative data augmentation network with physical constraints of steel strand dynamics is trained to obtain virtual acoustic samples; A Gaussian process regression recognition model is trained using the real samples and the virtual acoustic samples, and the prestress recognition result and prediction variance are obtained based on the fused feature vector to be tested. The confidence level of the prestress identification result is evaluated based on the predicted variance, and rejection is triggered if the evaluation fails. The temporal changes of the predicted variance are continuously monitored. When feature drift is detected, drift samples are generated and the Gaussian process regression identification model is incrementally updated.

2. The method for identifying the prestress of a single steel strand according to claim 1, characterized in that, The acquisition of the multi-source acoustic response signal includes: An impact excitation was applied to the anchoring system using a rebound hammer. Simultaneously acquire response signals from the cochlear sensor, external sound pressure sensor, and internal sensor of the rebound device; Multiple channel response signals under the same impact excitation are correlated as the same detection sample.

3. The method for identifying the prestress of a single steel strand according to claim 1, characterized in that, The preprocessing includes: Downsampling and high-pass filtering are performed on the original response signal; Pre-emphasis and peak alignment are performed on the filtered signal; Abnormal shocks were removed according to the sample quality rules, and a fixed-length shock response segment was extracted.

4. The method for identifying the prestress of a single steel strand according to claim 1, characterized in that, The extraction and screening of the multi-scale acoustic features include: Time-domain features, frequency-domain features, and cepstral features are extracted from each valid impact response sample; The maximum mutual information coefficient was used to evaluate the correlation between candidate features and prestressed labels for each channel; Candidate features that satisfy the correlation condition in multiple channels are selected as key features and concatenated to form the fused feature vector.

5. The method for identifying the prestress of a single steel strand according to claim 1, characterized in that, The training of the generative data augmentation network includes: Batch standardization was performed on the extracted real samples; Introduce a physical constraint loss consisting of the spectral peak, decay rate, and bandwidth energy into the adversarial training loss; A weighting strategy that adjusts based on training phase and generated quality feedback is used to fuse various losses.

6. The method for identifying the prestress of a single steel strand according to claim 5, characterized in that, The generation of the virtual acoustic samples includes: Organize real samples according to the steel strand duct or prestress level; Generative augmentation networks are used to generate multi-channel synchronous virtual signals using trained generative data. The virtual signal and the real sample are used together to form an enhanced training set.

7. The method for identifying the prestress of a single steel strand according to claim 1, characterized in that, The training of the Gaussian process regression identification model includes: Input the enhanced training set into the Gaussian process regression model; Model hyperparameters were obtained using kernel functions and cross-validation. Output the prestress identification result and the corresponding prediction variance.

8. The method for identifying prestress in a single steel strand according to claim 1, characterized in that, The confidence assessment includes: An absolute variance evaluation is generated based on the predicted variance; A relative interval evaluation is formed based on the relative width of the confidence interval; When any evaluation exceeds the corresponding threshold, a rejection flag is output and a prompt is made to re-collect data or conduct a manual review.

9. A prestress identification system for a single steel strand based on physical constraint generative data augmentation, characterized in that, include: The signal acquisition unit is used to acquire the multi-source acoustic response signals of the steel strand anchoring system under test under standardized impact excitation. The sample processing unit is used to preprocess the multi-source acoustic response signal and form a fused feature vector; Generative augmentation units are used to train generative data augmentation networks with physical constraints on steel strand dynamics and to obtain virtual acoustic samples. The identification and evaluation unit is used to train a Gaussian process regression identification model based on real samples and virtual acoustic samples, and output the prestress identification results, prediction variance and rejection labels. The drift update unit is used to monitor the temporal changes in the prediction variance and generate drift samples and update the Gaussian process regression identification model when feature drift is detected.

10. The prestressed identification system for a single steel strand according to claim 9, characterized in that, The system also includes a sample storage unit, which is connected to the sample processing unit, the generation and enhancement unit, the identification and evaluation unit, and the drift update unit, respectively. The sample storage unit is used to store real samples, virtual acoustic samples, rejected samples, and recent samples with high confidence, and to provide incremental update data to the drift update unit.

Images