Structural health monitoring method and system

Abstract

The application provides a structural health monitoring method and system, belongs to the field of building health safety, processes an input time sequence, generates a preliminary inference sequence, and forms a frequency spectrum feature after frequency domain processing of the preliminary inference sequence and a task visible context; a physical operator is adaptively selected and combined according to the frequency spectrum feature, an operator scheduling queue is generated, the operator scheduling queue is applied to the preliminary inference sequence, a physical correction sequence is obtained, a relationship reference is established with the input time sequence, a failure feature in the physical correction sequence is identified and a failure mode is constructed, an additional physical operator is selected according to the failure mode, the physical correction sequence is locally injected based on the additional physical operator, a candidate physical correction sequence is obtained, iterative verification is performed until a preset iteration number is reached or no failure feature exists, a final physical correction sequence is obtained, a task output result is obtained based on the final physical correction sequence, and a time sequence analysis model is trained; the accuracy of structural health monitoring is improved.

Description

Technical Field

[0001] This invention relates to the field of building health and safety technology, and in particular to structural health monitoring methods and systems. Background Technology

[0002] Structural health monitoring plays a crucial role in the safety status assessment and long-term operation and maintenance of critical facilities such as bridges and large buildings, and is an important technology for ensuring the stable service of engineering structures. This technology continuously collects multivariate sensor responses, including strain, displacement, vibration, and temperature, under long-term and unstable service conditions. It then performs time-series analysis tasks such as missing value completion, anomaly detection, and condition prediction, providing a basis for monitoring data recovery, abnormal state identification, and future condition assessment. With the development of deep learning, structural health monitoring time-series analysis technology has gradually evolved from traditional statistical methods to end-to-end models based on recurrent neural networks, convolutional neural networks, and Transformers. In recent years, the emergence of time-series foundational models has further promoted the development of unified time-series analysis technology, enabling the same model to support multiple structural health monitoring tasks through pre-training and downstream adaptation.

[0003] However, in actual structural monitoring data, dominant statistical correlations and fine-grained physical relationships exhibit significantly different characteristics: dominant statistical correlations have a strong impact on data fitting and are easily learned preferentially by deep models; while some fine-grained relationships, although possessing clear physical meaning, have relatively weak statistical significance and are easily masked by dominant statistical correlations. Existing time-series models primarily learn time-series representations based on observational data and statistical objectives, paying insufficient attention to the intrinsic physical relationships in structural responses. Taking bridge health monitoring as an example, during the representation learning process, frequency, phase, time delay, amplitude, and cross-sensor relationships may be weakened or destroyed, resulting in a lack of physical consistency and mechanistic interpretability in the learned representations. This makes it difficult for the model to accurately maintain the intrinsic relationships such as time delay and phase between different sensor responses, leading to missing physical relationships in the bridge. This results in distortion when completing missing bridge data, missed or false alarms when identifying abnormal bridge states, or deviations from the actual bridge response patterns when predicting future bridge states. Summary of the Invention

[0004] The technical problem to be solved by this invention is: how to improve the accuracy of structural health monitoring.

[0005] This invention solves the above-mentioned technical problems through the following technical solution: a structural health monitoring method, comprising: The collected structural health monitoring data is preprocessed to obtain the input time series sequence. The input time series sequence is then processed to generate the preliminary inference sequence. The preliminary inference sequence and the task visible context are processed in the frequency domain to form spectral features. Based on the spectral characteristics, physical operators are adaptively selected and combined to generate an operator scheduling queue. The operator scheduling queue is then applied to the preliminary inference sequence to obtain the physical correction sequence. A relational reference is established based on the input time series, failure features in the physical correction sequence are identified and failure modes are constructed, additional physical operators are selected based on the failure modes, and local injection is performed on the physical correction sequence based on the additional physical operators to obtain candidate physical correction sequences. The candidate physical correction sequences are iteratively verified until a preset number of iterations is reached or there are no failure features, and the final physical correction sequence is obtained. The task output result is obtained based on the final physical correction sequence, and the time series analysis model is obtained after training. Input the time series sequence to be analyzed into the time series analysis model to obtain the time series analysis results.

[0006] In this invention, step 2 involves constructing a parameterized frequency domain physical operator library, adaptively selecting and combining physical operators on task-related spectral components using long-range temporal context, and injecting the resulting operator scheduling queue into the preliminary inference sequence to obtain the initial physical correction timing sequence. Step 3 identifies inconsistencies in relationships that the current physical operator scheduling still cannot explain from complementary temporal and spectral views, and selects additional physical operators based on these inconsistencies. It then performs scheduling evolution and iterative verification on the corresponding local windows to reveal fine-grained physical relationships masked by dominant statistical correlation. This achieves the preservation of inherent physical relationships in structural responses and the revelation of hidden (masked by dominant statistical correlation) fine-grained physical relationships. This effectively alleviates the problem of physical relationship weakening or destruction caused by pure statistical learning, improves the physical consistency and mechanistic interpretability of timing representation and task output, and enhances the accuracy of structural health monitoring tasks such as missing value completion, anomaly detection, or state prediction.

[0007] Preferably, the process of processing the input time series sequence to generate a preliminary inference sequence includes: The input time series is divided into multiple time segments to obtain a time segment sequence. A learnable time summary tag is inserted at the beginning of the time segment sequence. The time summary tag and the time segment sequence are fed into a bidirectional long short-term memory encoder to obtain the encoded features. The encoded features are projected and the projection result is restored to the required time series length and sensor variable dimension to obtain the preliminary inference sequence.

[0008] Preferably, the process of initially inferring the sequence and task visible context to form spectral features after frequency domain processing includes: Based on the input time series, a task visibility context is constructed. Fast Fourier Transform is performed on the preliminary inference sequence and the task visibility context respectively to obtain the spectral representation of the input time series and the spectral representation of the task visibility context. The spectral representations of the time series and the task visibility context are aligned and concatenated to obtain the complete spectrum. The average spectral amplitude of the preliminary inferred sequence and the task visible context at each frequency position is calculated separately, and normalized using the average spectral amplitude of all frequency positions of the corresponding sequence; the scores of the normalized amplitude are summed according to preset weights to obtain the importance score of each frequency position; the top K frequency positions are selected according to the importance score, and the joint spectral features corresponding to the top K frequency positions are extracted to obtain the spectral features.

[0009] Preferably, the process of adaptively selecting and combining physical operators based on spectral characteristics to generate an operator scheduling queue includes: The spectral features are mapped to a unified operator selection space to obtain the feature projection result. The feature projection result is input into the recurrent decoder. At each decoding position, the recurrent decoder selects a physical operator from the parameterized physical operator library to form an ordered operator scheduling queue. The order of the physical operators in the operator scheduling queue determines the execution order of the physical operators. The parameterized physical operator library includes physical operators such as identity mapping, low-pass filtering, high-pass filtering, band-pass filtering, phase shifting, response delay adjustment, amplitude scaling, and spectral smoothing. Each physical operator has corresponding parameters.

[0010] Preferably, the process of applying the operator scheduling queue to the preliminary inference sequence to obtain the physical correction sequence includes: A Fast Fourier Transform is performed on the initial inferred sequence to obtain a spectral representation. The spectral coefficients of the first K frequency positions are extracted from the spectral representation. The physical operators in the operator scheduling queue are executed sequentially on the first K spectral coefficients according to their order, resulting in K modified spectral coefficients. The modified spectral coefficients are written back to the original frequency positions in the spectral representation to obtain a modified spectral representation. An Inverse Fast Fourier Transform is performed on the modified spectral representation to obtain a physical correction sequence.

[0011] Step 2 is used to preserve the inherent physical relationships that may be weakened or disrupted during the initial inference process. Based on spectral characteristics, Step 2 adaptively selects and combines physical operators such as band selection, phase shift, amplitude scaling, and time delay adjustment from a parametric physical operator library to generate an ordered operator scheduling queue. Subsequently, the operator scheduling queue is applied to the task-related spectral components of the initial inference sequence, allowing physical relationships such as frequency, phase, amplitude, and time delay to enter the timing inference process, resulting in a physical correction sequence.

[0012] Preferably, the process of establishing a relational reference based on the input time series, identifying failure features in the physical correction sequence, and constructing failure modes includes: The input time series and physical correction sequence are divided into position-aligned local windows. The relational representation of each local window is extracted through the same multi-view observation to obtain the relational representation feature vector of the input time series and physical correction sequence. Clustering is performed on the relational representation feature vectors of the input time series to obtain reference cluster centers, and the distance from the relational representation feature vectors of the input time series to the reference cluster centers is used as the reference neighborhood; The relation representation feature vector of the physical correction sequence is mapped to the reference neighborhood. Relation representations located outside the reference neighborhood are identified as failure features. The features are ranked according to the normalization deviation of the failure features relative to the nearest reference neighborhood. The top-ranked features are organized into failure modes.

[0013] Preferably, the process of locally injecting the physical correction sequence based on additional physical operators to obtain candidate physical correction sequences includes: Find the local window corresponding to the window of the additional physical operator in the physical correction sequence, transform the local window to the frequency domain to obtain the local spectrum, perform the additional physical operator on the local spectrum to obtain the corrected local window, restore the corrected local window to the time domain and merge it with the unmodified region in the physical correction sequence to obtain the candidate physical correction sequence.

[0014] Preferably, the process of iteratively verifying the candidate physics correction sequence includes: dividing the current candidate physics correction sequence into multiple local windows, extracting the relational representation of each local window through multi-view observation, and obtaining the relational representation feature vector of the candidate physics correction sequence; Map the relation representation feature vector of the current candidate physical correction sequence to the reference neighborhood, identify the relation representation outside the reference neighborhood as the current failure feature, obtain the current failure mode based on the current failure feature, calculate the scores of the current failure mode and the previous round failure mode respectively, and the score of the failure mode is the sum of the squared Euclidean distance from each failure feature in the failure mode to the nearest cluster center and the deviation of the squared Euclidean distance from the reference neighborhood threshold. If the difference between the score of the current failure mode and the score of the previous failure mode reaches the improvement amount, the current candidate physics correction sequence is accepted and proceeds to the next iteration; otherwise, the current candidate physics correction sequence is not accepted and the previous candidate physics correction sequence is returned.

[0015] Step 3 primarily addresses relation failures that the initial operator scheduling queue cannot interpret, revealing fine-grained physical relations masked by dominant statistical correlations. This step establishes relation references using the input time series and uses the physical correction sequence as the initial state for iterative correction. Failure features in the current physical correction sequence are identified and failure modes are constructed through multi-view observations. Subsequently, additional physical operators are selected based on the failure modes, and corresponding failure windows are locally injected. The correction results are verified through queue evolution and acceptance gates. This process is repeated, allowing the operator scheduling queue to gradually adapt to the current input time series, ultimately yielding a final physical correction sequence with more complete physical relations.

[0016] Preferably, the time series analysis model includes a trained bidirectional long short-term memory encoder, a task head, a physical operator parameter generator, a recurrent decoder, and a multilayer perceptron selector. The time series analysis result is any one of the following: a completed sequence, an anomaly score, or a predicted sequence.

[0017] The present invention also provides a structural health monitoring system, comprising: The feature extraction module is used to preprocess the collected structural health monitoring data to obtain the input time series sequence. The input time series sequence is processed to generate a preliminary inference sequence. The preliminary inference sequence and the task visible context are processed in the frequency domain to form spectral features. The physical operator injection module is used to adaptively select and combine physical operators based on spectral characteristics, generate an operator scheduling queue, and apply the operator scheduling queue to the preliminary inference sequence to obtain the physical correction sequence. The physical relationship revealing module is used to establish relationship references from the input time series, identify failure features in the physical correction sequence and construct failure modes, select additional physical operators based on the failure modes, perform local injection on the physical correction sequence based on the additional physical operators to obtain candidate physical correction sequences, iteratively verify the candidate physical correction sequences until a preset number of iterations is reached or no failure features are found, and obtain the final physical correction sequence. The task output result is obtained based on the final physical correction sequence, and the time series analysis model is obtained after training. The task output module is used to input the input time series sequence to be analyzed into the time series analysis model to obtain the time series analysis results. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] Figure 1 This is a schematic diagram of the structural health monitoring method provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the structural health monitoring method provided in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the physical operator injection module in the structural health monitoring system provided in Embodiment 2 of the present invention; Figure 4 This is a schematic diagram of the physical relationship disclosure module in the structural health monitoring system provided in Embodiment 2 of the present invention.

[0020] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] Structural health monitoring continuously collects multivariate sensor responses, such as strain, displacement, vibration, and temperature, under long-term and uncertain service conditions. Processing this multivariate sensor data yields time-series data, which can be analyzed to perform tasks such as missing value completion, anomaly detection, and condition prediction. Existing structural health time-series analysis methods generally suffer from insufficient physical consistency, weak mechanistic interpretability, and decreased generalization ability when handling high-dimensional, cross-sensor coupled, and non-stationary monitoring data, resulting in low accuracy of the output results. The main reason is that traditional deep time-series models primarily learn time-series representations based on observational data and statistical objectives, paying insufficient attention to the intrinsic physical relationships within the structural response. These relationships may be weakened or destroyed during the representation learning process, leading to distorted task outputs. Furthermore, while existing basic models possess strong temporal representation and multi-task adaptability, they still primarily rely on dominant statistical correlations that significantly impact data fitting. This makes it difficult to fully preserve fine-grained relationships with clear physical meaning but weak statistical significance. In actual structural health monitoring, this leads to over-reliance on dominant overall trends, neglecting subtle but physically meaningful time delay, phase, and frequency band energy relationships between sensors. This results in difficulty in timely identification of early local damage or weak anomalies, potentially causing missed anomalies, biased state predictions, and distortion in missing value completion. Simultaneously, existing methods lack effective mechanisms for adaptively maintaining intrinsic physical relationships and revealing hidden ones. This causes the model to gradually disrupt the intrinsic physical relationships of the bridge response during completion, anomaly detection, and state prediction, generating results inconsistent with the actual structural mechanisms. When environmental conditions, loads, sensor states, or bridge types change, false alarms, missed alarms, and decreased generalization performance across bridges are prone to occur, reducing the reliability of bridge condition assessment and maintenance early warning.

[0023] like Figure 1As shown, this invention provides a structural health monitoring method that can be used to perform time series analysis on structural monitoring data of bridges or other large buildings. If the target task of the time series analysis is missing value completion, the result of the time series analysis is a completed sequence, which is used to complete the missing values ​​of the original observed input time series. If the target task of the time series analysis is an anomaly detection task, the result of the time series analysis is an anomaly score. If the target task of the time series analysis is to predict the future state, the result of the time series analysis is a predicted sequence within a preset future time series range.

[0024] Example 1 This embodiment uses bridge structural health monitoring data and time series analysis as examples to illustrate the task of missing value completion. (See also...) Figure 2 The specific process of structural health monitoring methods includes: Step 1: Deploy multi-element sensors at key locations on the bridge. These sensors include strain sensors, displacement sensors, vibration sensors, and temperature sensors. The data collected by these sensors is structural health monitoring data. The collected structural health monitoring data is preprocessed, for example, by operating on the data from each sensor separately, reducing the data from the original second-level to the minute-level (by averaging), and then normalizing it to obtain the input time series. The input time series is then processed to generate a preliminary inference sequence. The preliminary inference sequence and the task visible context are then processed in the frequency domain to form spectral features.

[0025] Step 1 involves processing the input time series sequence to generate the preliminary inference sequence, which includes: See Figure 1 Step 1.1.1: Input time series sequence The time segment is divided into multiple time segments to obtain a time segment sequence, and the multiple time segments are time segments. , , ... , This represents the number of time segments. A learnable time summary marker is inserted at the beginning of the time segment sequence. .

[0026] Step 1.1.2: Tag the learnable time summary Together with the time segment sequence, it is fed into a bidirectional long short-term memory encoder. During the encoding process, learnable time summary tags are generated. By aggregating sequence-level temporal information, a bidirectional long short-term memory encoder simultaneously learns forward and backward temporal dependencies to obtain encoded features. .

[0027] Step 1.1.3: Encode the features Input task header, task header for encoding features The task-specific output projection is performed; in this embodiment, the task is missing value completion. Therefore, the task header encodes features. The output projection of the missing value completion task involves reconstructing the sequence in the task header and aligning it with the input window to obtain the preliminary inference sequence. For anomaly detection tasks, the task header also reconstructs a sequence aligned with the input window (a segment of continuous time-series data truncated from the input time-series sequence according to a preset window length) to obtain a preliminary inference sequence. For prediction tasks, the task head projects historical encoded features (encoded features obtained by encoding historical input windows using a bidirectional long short-term memory encoder) onto a preset future time range to obtain a preliminary inferred sequence. Preliminary inferences suggest that the sequence is a complete sequence generated by the bidirectional long short-term memory encoder and the task head based on the input encoding features.

[0028] The process of forming spectral features from the initial inference sequence and task visible context in step 1 after frequency domain processing includes: Step 1.2.1: Construct a task visibility context based on the input time series sequence. The task visibility context is an auxiliary information sequence extracted or constructed from the input time series sequence according to the current task, and is processed to align with the output time range and sensor variable dimensions of the preliminary inference sequence. For the missing value completion task, the task visibility context is constructed from the missing position mask of the input window, used to identify the observed position (not missing) and the missing position; for the anomaly detection task, the task visibility context is the current input window aligned with the preliminary inference sequence, used to provide the actual observation response; for the state prediction task, the task visibility context is obtained by mapping the historical boundary fragments at the end of the historical input window to a preset prediction range, used to provide historical boundary response information. The downstream task in this embodiment is the missing value completion task, specifically obtained by projecting the input time series sequence and its mask onto the reconstruction space to obtain the task visibility context. For anomaly detection tasks, the task visibility context is obtained by aligning the current observation window with the preliminary inference sequence. The current observation window is a continuous time segment extracted from the input time series and aligned with the preliminary inference sequence in terms of time range and sensor variable dimensions. For the prediction task, the historical boundary segment is projected onto the prediction range (the future time interval to be output) to obtain the task's visible context. Historical boundary segments are continuous time segments of the historical input window near the prediction start point in a prediction task.

[0029] Step 1.2.2: Perform preliminary inference on the sequence respectively. and task-visible context Perform a Fast Fourier Transform to obtain the spectral representation of the input time series and the spectral representation of the task visibility context. Align and concatenate the spectral representations of the time series and the task visibility context to obtain the complete spectrum. The specific process of aligning and concatenating the spectral representations of the time sequence and the task's visible context is as follows: For each frequency position, the initially inferred sequence... Spectral coefficients and task visibility context The spectral coefficients at the same frequency position are combined. The combined spectral coefficients include the real part, imaginary part, amplitude, relative phase information and frequency position information of both, to obtain the joint spectral characteristics corresponding to that frequency position. The joint spectral characteristics of all frequency positions constitute the complete spectrum.

[0030] Steps 1, 2, and 3: Calculate the preliminary inference sequence respectively. and task-visible context The average spectral amplitude at each frequency position is calculated and normalized using the average spectral amplitude of all frequency positions in the corresponding sequence. The scores of the normalized amplitudes are then summed according to preset weights to obtain the importance score for each frequency position. The top K frequency positions are selected based on their importance scores, and the joint spectral features corresponding to the top K frequency positions are extracted to obtain the spectral features. .

[0031] Step 2: Based on spectral characteristics Adaptively select and combine physical operators to generate an operator scheduling queue, and apply the operator scheduling queue to the preliminary inference sequence. The physical correction sequence is obtained.

[0032] See Figure 3 Step 2, which involves adaptively selecting and combining physical operators based on spectral characteristics to generate an operator scheduling queue, includes: Spectral characteristics First, through feature projection, the spectral features are... The feature projection results are obtained by mapping to a unified operator selection space. These results are then input into a recurrent decoder. At each decoding position, the recurrent decoder selects a physical operator from a parametric physical operator library, forming an ordered operator scheduling queue. :

[0033] in, Represents operator scheduling queue The Middle Physical operators for location selection Indicates the first Physical operator for location selection The corresponding parameters, physical operators are in the operator scheduling queue The order in which the physical operators are arranged determines the execution order of the physical operators.

[0034] The process of the loop decoder selecting a physical operator from the parameterized physical operator library at each decoding position is as follows: The feature projection result enters the loop decoder, and after decoding, the scores or probabilities of all physical operators are output. The physical operator with the highest score or probability is placed in the first decoding position of the feature projection result, and the hidden state of the first decoding position is recorded. Then, the feature projection result, the hidden state of the first decoding position, and the physical operator corresponding to the first decoding position are input into the loop decoder again. After decoding, the scores or probabilities of all physical operators are output, and the physical operator with the highest score or probability is placed in the second decoding position of the feature projection result. This process continues until the end marker is output or the preset length is reached. All decoding positions of the feature projection result are filled with the corresponding physical operators.

[0035] The parameterized physical operator library includes physical operators such as identity mapping, low-pass filtering, high-pass filtering, band-pass filtering, phase shifting, response delay adjustment, amplitude scaling, and spectral smoothing. Among them, identity mapping, low-pass filtering, high-pass filtering, and band-pass filtering physical operators are used to describe frequency selection in structural response; phase shifting physical operator is used to describe phase change in structural response; response delay adjustment is used to describe timing delay in structural response; amplitude scaling physical operator is used to describe response amplitude in structural response; and spectral smoothing physical operator is used to describe local spectral continuity in structural response.

[0036] The physical operator parameter generator sets corresponding parameters for different physical operators in the parameterized physical operator library. These parameters are driven by the training objective, learned through a gradient-based Adam optimization algorithm, and constrained within an effective parameter range by bounded transformations corresponding to the operator type. This yields the physical operator's action strength, cutoff frequency, passband boundary, phase or delay scale, and filtering characteristics. The physical operator parameters determine the specific execution method and spectral characteristics of that physical operator. Used to guide the loop decoder in selecting the type and order of physical operators.

[0037] During training, the physics operator parameter generator maintains globally learnable variables for various operators, such as cutoff frequency, passband boundaries, phase or delay scales, and amplitude adjustment parameters. These variables are learned through the training objective and the Adam optimizer and are shared across different samples and scheduling locations; they are not an independent set of parameters regenerated each time based on the input sequence. The input spectrum primarily guides the recurrent decoder in selecting which operators and their execution order. In practical applications, pre-trained parameters are loaded, and Adam training is not reused.

[0038] To ensure parameter validity, unconstrained learnable variables are transformed into actual operator parameters using bounded functions. For example, parameters that need to be restricted to a certain interval can be first processed by a Sigmoid function and then scaled to that interval; parameters that allow positive and negative variations can be processed by a Tanh function and then multiplied by the maximum scale of variation. Specific boundaries are defined by the corresponding operators. During execution, a Fourier transform is performed on the initial inference sequence, and the coefficients at the Top-K frequency positions are modified according to the scheduling order, while other frequency components remain unchanged. The modified coefficients are then written back to their original positions and inversely transformed to obtain the physical correction sequence. The training loss is fed back to the relevant parameters through differentiable computation and updated by Adam.

[0039] Step 2, which applies the operator scheduling queue to the preliminary inference sequence to obtain the physical correction sequence, includes: A Fast Fourier Transform (FFT) is performed on the initial inferred sequence to obtain a spectral representation. The spectral coefficients at the first K frequency positions are extracted from this representation. Following the order of the physical operators in the operator scheduling queue, the physical operators in the queue are executed sequentially on the first K spectral coefficients to obtain K modified spectral coefficients. These modified spectral coefficients are then written back to the original frequency positions in the spectral representation to obtain a modified spectral representation. An Inverse Fast Fourier Transform (IFFT) is then performed on this modified spectral representation to obtain the physical correction sequence. .

[0040] Step 2 is used to preserve the intrinsic physical relationships that may have been weakened or disrupted during the initial inference process. Step 2 is based on spectral characteristics. The system adaptively selects and combines physical operators such as frequency band selection, phase shift, amplitude scaling, and time delay adjustment from a parametric physical operator library to generate an ordered operator scheduling queue. Subsequently, the operator scheduling queue is applied to the task-related spectral components of the initial inference sequence, so that physical relationships such as frequency, phase, amplitude, and time delay are incorporated into the timing inference process, resulting in a physical correction sequence.

[0041] Step 3: Establish a relational reference based on the input time series and identify the physical correction sequence. The system identifies failure characteristics and constructs failure modes. Based on these failure modes, additional physical operators are selected, and the physical correction sequence is then adjusted accordingly. Local injection is performed to obtain candidate physical correction sequences. These sequences are then iteratively verified until a preset number of iterations is reached or no failure characteristics are found, at which point the final physical correction sequence is obtained. The task output is obtained based on the final physical correction sequence, and a time series analysis model is obtained after training.

[0042] See Figure 4 Step 3, which involves establishing a relational reference based on the input time series, identifying failure features in the physical correction sequence, and constructing failure modes, includes: Step 3.1.1: Input time series sequence and physical correction sequence The system is divided into position-aligned local windows, and the relational representation of each local window is extracted through the same multi-view observation to obtain the relational representation feature vectors of the input time series sequence and the physical correction sequence.

[0043] Multi-view observation includes four views: temporal continuity, response delay, band energy, and cross-sensor phase. First, the input sequence and the current physical correction sequence are divided into position-aligned local windows. Then, these four types of relationship features are extracted from windows at the same position. The temporal continuity view obtains the difference by subtracting adjacent sample values, i.e., calculating the change between adjacent time points, used to describe the direction and magnitude of the change. The response delay view first averages the standardized sequences of each sensor within the same window time-by-time to obtain the reference response. Then, it compares the window sequence of a sensor with the reference response at different time offsets, multiplying the corresponding values ​​at effective overlapping positions and averaging them as the similarity at that offset, used to indicate whether the sensor response is ahead or behind the reference response. The band energy view performs a Fourier transform on the window and statistically analyzes the energy proportions in the low-frequency, mid-frequency, and high-frequency ranges. The cross-sensor phase view compares the phase difference between different sensors and the reference response at the same frequency position, used to describe whether the sensor responses are synchronized.

[0044] For each local window and sensor channel, temporal continuity, response delay, bandgap energy, and cross-sensor phase characteristics are calculated separately, and these are combined into a relational representation for that local window. The input sequence and the physical correction sequence are extracted using the same method, and then their relationships are compared to see if they are consistent. The relational representation is used to describe structural response characteristics such as temporal continuity, response delay, bandgap energy, and cross-sensor phase relationships.

[0045] Step 3.1.2: Cluster the relational representation feature vectors of the input time series to obtain reference cluster centers. The distance from the relational representation feature vectors of the input time series to the reference cluster centers is used as the reference neighborhood.

[0046] Step 3.1.3: Map the relation representation feature vectors of the physical correction sequence to the reference neighborhood. Identify relation representations outside the reference neighborhood as failure features. Sort the features according to their normalized deviation from the nearest reference neighborhood. Organize the top-ranked features into failure modes. The top-ranked features are those with the strongest failure response. Failure Mode Used to summarize the main relation types, occurrence locations, and failure levels in the physical correction sequence that have not yet been interpreted by the physical operator scheduling queue.

[0047] Establishing the reference neighborhood involves two steps: reference feature clustering and distance threshold determination. First, normal relationship representation features extracted from the input time series are used as reference features, and clustering is performed according to four views: temporal continuity, response delay, frequency band energy, and cross-sensor phase, obtaining reference cluster centers for each view. During clustering, a small number of reference features with large distances are removed based on the distance distribution from the reference features to the nearest cluster center to reduce the impact of outliers on the reference cluster centers.

[0048] Subsequently, the squared distance from the retained reference features to their nearest reference cluster centers is calculated to form a reference deviation distance distribution. The 92nd percentile of this distribution is taken as the distance threshold for the reference neighborhood, hereinafter referred to as the reference neighborhood threshold. For any relation feature to be judged, the squared distance between it and each reference cluster center in the same view is first calculated, and the minimum value among them is taken as the deviation distance of the feature relative to the reference distribution. If the distance does not exceed the reference neighborhood threshold of the corresponding view, the feature is considered to be within the reference neighborhood and belongs to an interpretable relation; if it exceeds the reference neighborhood threshold, it is marked as a failed feature, and a consecutive failure score is calculated according to the relative degree to which it exceeds the reference neighborhood threshold.

[0049] Once the reference cluster center and reference neighborhood thresholds are established, they are kept fixed, and the same set of reference standards are used to judge the initial physical correction sequence and the candidate physical correction sequence. This avoids re-establishing the reference neighborhood after each candidate correction is generated, thereby ensuring that there is a consistent comparison standard between different correction results.

[0050] Step 3, the specific process of selecting additional physical operators based on failure modes, is as follows: The failure modes are input into the multilayer perceptron selector, which performs failure analysis on the failure modes and outputs the selection results of additional physical operators, window and sensor activation levels, and local interaction parameters. The queue evolution selects additional physical operators from the parameterized physical operator library based on the output of the multilayer perceptron selector. And it is applied to the corresponding local window of the physical correction sequence through failure window injection.

[0051] Step 3, which involves locally injecting the physical correction sequence based on additional physical operators to obtain candidate physical correction sequences, includes: Find the physical operator in the physical correction sequence. The local window corresponding to the given window is transformed to the frequency domain to obtain the local spectrum, and additional physical operators are performed on the local spectrum. The corrected local window is obtained, and after restoring the corrected local window to the time domain, it is merged with the unmodified region in the physical correction sequence to obtain the candidate physical correction sequence.

[0052] The physical correction sequence obtained in step 2 Used to initialize the iteration process, set at the... In the round of iterations, the current candidate physical correction sequence is See Figure 3 Step 3, the iterative verification of candidate physical correction sequences, includes: Step 3.2.1: The current candidate physical correction sequence The system is divided into multiple local windows, and the relational representation of each local window is extracted through multi-view observations to obtain candidate physical correction sequences. The relational representation of the feature vector; Step 3.2.2: The current candidate physical correction sequence The relation representation feature vector is mapped to the reference neighborhood in step 3.1.2. Relation representations located outside the reference neighborhood are identified as current failure features. The current failure mode is obtained based on the current failure features. The scores of the current failure mode and the previous failure mode are calculated respectively. The score of the failure mode is the sum of the squared Euclidean distance from each failure feature in the failure mode to the nearest cluster center and the deviation of the squared Euclidean distance from the reference neighborhood threshold.

[0053] For example, the current failure mode include Failure characteristics, previous round failure pattern include Each failure feature is calculated. The squared Euclidean distance from each of the failed features to the nearest cluster center can be obtained. The square Euclidean distance is then calculated. The average of squared Euclidean distances ,calculate The deviation of each squared Euclidean distance from a reference neighborhood threshold is calculated. The reference neighborhood threshold is set and the deviation is the absolute value of the difference between the squared Euclidean distance and the reference neighborhood threshold. One deviation, then calculate The average of the deviations Current failure mode fractions for Similarly, the failure mode of the previous round can be calculated. fractions .

[0054] Step 3.2.3: If the difference between the score of the current failure mode and the score of the previous failure mode is... When the improvement amount is reached, the improvement amount is obtained by setting the current candidate physical correction sequence. Accepted, proceed to the next round (number) (Round) Iteration, repeating steps 3.2.1 to 3.2.3; otherwise, the current candidate physical correction sequence... If rejected, return to the previous round of candidate physical correction sequences. Repeat steps 3.2.1 to 3.2.3.

[0055] Only the difference Candidate modifications must be at least equal to the minimum improvement amount (preset minimum improvement amount) to be accepted. The improvement amount is a decision parameter in the model configuration, determined as needed during training or validation. In practical applications, the determined settings are used, eliminating the need for retraining. The completion task primarily checks the continuity near observed locations and missing boundaries; the prediction task checks the level and trend connection with historical endpoints; the anomaly detection task limits the magnitude of modification of candidate corrections relative to the initial correction sequence. Upon acceptance, candidate results corresponding to the window (the current candidate physical correction sequence) are retained. When rejecting, retain the result before modification (the previous round of candidate physical correction sequence). ).

[0056] The multilayer perceptron selector receives failure modes and outputs selection scores for additional physical operators, window and sensor activation levels, and local action parameters. In a training process with candidate comparison enabled, the degree of relational improvement of candidate physical correction sequences is compared, and the better-performing candidate operators are used as automatically generated selection targets. Operator selection is trained using cross-entropy loss. (Cross-entropy loss) for:

[0057] in, The set of valid windows participating in the correction. For window The target operator, Assign a target operator to the multilayer perceptron selector The probability of action. The action location, activation level, and local action parameters are updated through backpropagation of the loss of the candidate or final physical correction sequence, without the need for manual labeling of which physical operator should be used for a particular window.

[0058] Step 3 primarily addresses relation failures that the initial operator scheduling queue cannot interpret, revealing fine-grained physical relationships masked by dominant statistical correlations. This step establishes a relation reference using the input time series and uses the physical correction sequence as the initial state for iterative correction. Failure features in the current physical correction sequence are identified and failure modes are constructed through multi-view observations. Subsequently, additional physical operators are selected based on the failure modes, and corresponding failure windows are locally injected. The correction results are verified through queue evolution and a receiving gate. This process is repeated, allowing the operator scheduling queue to gradually adapt to the current input time series, ultimately yielding a final physical correction sequence with more complete physical relationships. The receiving gate does not participate in gradient backpropagation.

[0059] The training process of the time series analysis model is divided into two main stages: the first stage trains the basic time series model, and the second stage trains the physical correction module. The second stage includes physical operator scheduling, relation correction, and candidate result verification; candidate verification is the screening step in this stage, used to determine whether to retain local corrections.

[0060] Data preparation: The bridge monitoring data was time-aligned, missing data flagged, normalized, and divided into sliding window sets, and then divided into training, validation, and test sets according to time order.

[0061] The first stage involves dividing the input time series into segments, adding learnable time summary labels, extracting encoded features using a bidirectional matrix long short-term memory encoder (Bi-mLSTM), and then generating an initial inference sequence from the task head. The completion and anomaly detection tasks generate reconstructed sequences aligned with the input window, while the prediction task generates predicted sequences within a preset future range. Task loss mechanisms drive the encoder and task head to learn temporal statistical patterns.

[0062] The first stage utilizes task loss to train preliminary inference capabilities: For missing value imputation, some known observations in the training sequence are artificially masked, and the true values ​​before masking are used as the supervision target. The mean square error of the masking position is calculated, and the calculation formula can be:

[0063] in, To compensate for missing values ​​and reduce task losses, A collection of locations where people are obstructed. For time In a sheltered location The original true observations, For time In a sheltered location The model output value.

[0064] For anomaly detection, reconstruction training is performed using normal observation sequences, and the mean square error between the reconstructed result and the normal input time series sequence is calculated. The calculation formula can be:

[0065] in, Losses due to anomaly detection tasks The input time series is a normal sequence. For the reconstruction result, For window length, This represents the number of sensor channels.

[0066] For future predictions, using the actual observation sequence after the historical window as the monitoring target, the weighted sum of the mean square error and the mean absolute error is calculated. The calculation formula can be:

[0067] in, To predict future mission losses, This refers to the predicted sequence output by the model in future forecasting. For the actual observation sequence corresponding to the predicted range, Indicates mean square error. Indicates the mean absolute error. These are the weighting coefficients.

[0068] Phase Two: Based on the preliminary inferred sequence and the task's visible context, spectral features are generated. A physical operator scheduler is trained to select and arrange physical operators, resulting in an initial physical correction sequence. Relational correction is then performed, identifying local failure features and forming failure modes. The selector determines additional operators and their application locations. Relational observation is re-performed on candidate sequences, retaining only those that provide sufficient improvement without worsening the task outcome. Relational Correction Training: The same multi-view feature extraction is performed on the input time-series sequence and the current physical correction sequence. Reference prototypes are established in each view using the features of the input sequence to identify deviation features in the correction sequence. High-response features are encoded and aggregated into failure modes based on similarity. The MLP selector outputs candidate operator scores, window and sensor channel activation levels, and local adjustment coefficients to determine "which operator to use, where to modify, and by how much." Candidate Validation: A Fourier transform is performed on the selected window, applying the candidate operators to the local spectrum, followed by an inverse transform back to the time domain. The corrected window is merged with the unmodified region, and overlapping parts are fused according to the composite weight to form a candidate sequence. Relational features are then re-extracted, the original reference center and threshold are reused, and failure scores are compared. Modifications that meet the acceptance criteria proceed to the next round, while those that do not are discarded. The process ends when the preset number of rounds is reached or there are no more activity failure windows to process. This forward candidate validation process can still be performed after the model is deployed, but the model parameters will no longer be updated via Adam.

[0069] The second stage incorporates a weighted average of task loss and failure improvement loss into the training objective, enabling the model to address inconsistencies in relationships while reducing task error. Therefore, the total loss function for the second stage of training is:

[0070] in, This represents the total loss during the second phase of training. For task loss, when the task involves filling in missing values, for When the task is anomaly detection, for When the task is to predict the future, for ; To mitigate losses from failures, Improve loss weighting for failures. Cross-entropy loss The weights of the factors, including the failure improvement loss. The calculation method is as follows:

[0071] in, The window before correction Failure score For candidate corrected windows Failure score The set of valid windows participating in the correction. To preset the relative improvement range, A constant to prevent the denominator from being too small.

[0072] The first stage uses training data to train the basic time series model. The optimizer in the first stage updates all parameters of the basic model, enabling the bidirectional matrix long short-term memory encoder and task head to learn the temporal statistical regularities and generate preliminary inference sequences. The second stage first loads the model parameters saved in the first stage and freezes all parameters of the basic time series model; then it only trains the physical correction-related parts, including physical operator scheduling, physical operator parameters, and modules related to local relation correction and candidate selection.

[0073] Phase 1: Load the pre-trained parameters of the basic time series model, learn the time series statistical patterns through the training data of the corresponding task, and generate preliminary inference sequences. All three tasks use the Adam optimizer and cosine learning rate scheduling, with the specific configuration as follows.

[0074] (1) Missing completion: The current strict training process with a missing rate of 25% freezes the pre-trained encoder and trains the unfrozen task adaptation part; the initial learning rate is 0.0002, the batch size is 32, and the maximum training time is 20 rounds.

[0075] (2) Future prediction: Freeze the pre-trained encoder and train the unfrozen task adaptation part; the initial learning rate is 0.00025, the batch size is 32, and the maximum training time is 16 rounds.

[0076] (3) Outlier detection: The pre-trained encoder is not frozen, and the encoder and task head are updated together; the initial learning rate is 0.00025, the batch size is 32, and the maximum training time is 15 rounds. After each round of training, the three tasks are evaluated on the validation set, and the model with better validation performance is saved and selected as the base model for the second stage.

[0077] The training data labels are constructed as follows: when the task is missing value completion, the known observations in the training sequence are artificially masked, and the true values ​​before masking are used as the supervision targets. The loss is calculated at the corresponding positions. Naturally missing positions that originally had no true values ​​cannot be directly used as supervised labels.

[0078] When the task is anomaly detection, the reconstruction target is mainly the normal observed sequence itself, and the reconstruction model is not trained with anomaly labels from the test set. If artificial perturbation is enabled to assist training, a program-generated perturbation is applied to the normal sequence, and the perturbation position is recorded by the program to construct auxiliary supervision.

[0079] When the task is future prediction, the real observations within a preset prediction range after the input historical window are used as the supervision target. The real future values ​​are only used for supervision during training and are not used as the task visibility context during actual prediction.

[0080] The second stage involves loading the model saved from the first stage, freezing the basic time-series model, and training parameters related to physical operator selection and ranking, operator parameters and local relation correction, and candidate selection. All three tasks employ the Adam optimizer and cosine learning rate scheduling. The training loss is backpropagated through frequency domain operations of the physical operators, inverse fast Fourier transform, and local correction processes. The Adam optimizer updates trainable parts such as the physical operator parameter generator, recurrent decoder, and multilayer perceptron selector. Operator selection in the recurrent decoder uses Gumbel-Softmax pass-through estimation during training, enabling the forward process to select specific operators while preserving approximate gradients during the backward process.

[0081] All three tasks involve validation set evaluation after each training round, saving and selecting models with better validation performance. Candidate correction validation is used to decide whether to retain a local correction and is a screening step in the physical correction process; validation set evaluation is used to select the post-trained model state. The two are not the same operation and do not constitute an additional training phase.

[0082] The training objectives include the error term for the specific task and the enabled relation correction auxiliary term. The completion task compares the completed value with the true value at the training occlusion location where the true value is available; the prediction task compares the predicted sequence with the future true sequence of the training samples; anomaly detection learns reconstruction patterns on normal training data, and during inference, obtains anomaly scores from reconstruction differences and relation deviations. The presence of anomaly scores does not necessarily indicate that anomaly label classification loss was used during training.

[0083] In this invention, step 2 involves constructing a parameterized frequency domain physical operator library, adaptively selecting and combining physical operators on task-related spectral components using long-range temporal context, and injecting the resulting operator scheduling queue into the preliminary inference sequence to obtain the initial physical correction timing sequence. Step 3 identifies inconsistencies in relationships that the current physical operator scheduling cannot explain from complementary temporal and spectral views, and selects additional physical operators based on these inconsistencies. It then performs scheduling evolution and iterative verification on the corresponding local windows to reveal fine-grained physical relationships masked by dominant statistical correlation. This achieves the preservation of inherent physical relationships in structural responses and the revelation of hidden (masked by dominant statistical correlation) fine-grained physical relationships. This effectively alleviates the problem of physical relationship weakening or destruction caused by pure statistical learning, improves the physical consistency and mechanistic interpretability of timing representation and task output, and enhances the accuracy of structural health monitoring tasks such as missing value completion, anomaly detection, or state prediction.

[0084] The time-series analysis model constructed in this invention can maintain the inherent physical relationships in the structural response under conditions such as missing mode changes, abnormal disturbances, and cross-bridge distribution shifts, reduce the masking of fine-grained physical relationships by dominant statistical correlations, and improve the physical consistency, task accuracy, and cross-bridge generalization ability of results from various tasks.

[0085] Step 4: Input the input time series sequence to be analyzed into the time series analysis model to obtain the time series analysis result. The time series analysis model includes a trained bidirectional long short-term memory encoder, a task head, a physical operator parameter generator, a recurrent decoder, and a multilayer perceptron selector. In this embodiment, the time series analysis result is the completed sequence. If the task is anomaly detection, the time series analysis result is the anomaly score. If the task is future time series prediction, the time series analysis result is the predicted sequence.

[0086] Step 4 involves inputting the input time series sequence to be analyzed into the time series analysis model to obtain the time series analysis results. The time segment sequence corresponding to the input time series sequence to be analyzed and the inserted time summary tag are fed together into the trained bidirectional long short-term memory encoder to obtain the encoded features; the encoded features are then input into the trained task head to obtain the preliminary inferred sequence. Preliminary inferences of the sequence and task visible context yield spectral features after frequency domain processing. The feature projection result after spectral feature mapping is input into the trained recurrent decoder. The trained physical operator parameter generator generates the parameters of physical operators in the parameterized physical operator library. The trained recurrent decoder selects a physical operator from the parameterized physical operator library at each decoding position, forming an ordered operator scheduling queue. The operator scheduling queue is applied to the initial inference sequence to obtain the physical correction sequence; The input time series and physical correction sequence are divided into position-aligned local windows. The relational representation of each local window is extracted through the same multi-view observation to obtain the relational representation feature vector of the input time series and physical correction sequence. Clustering is performed on the relational representation feature vectors of the input time series to obtain reference cluster centers, and the distance from the relational representation feature vectors of the input time series to the reference cluster centers is used as the reference neighborhood; The relation representation feature vector of the physical correction sequence is mapped to the reference neighborhood. Relation representations located outside the reference neighborhood are identified as failure features. The features are ranked according to the normalization deviation of the failure features relative to the nearest reference neighborhood. The top-ranked features are organized into failure modes.

[0087] The failure mode is input into the trained multilayer perceptron selector, which outputs the selection result of the additional physical operator, the window, the sensor activation degree, and the sensor local action parameters. Based on the output of the multilayer perceptron selector, the additional physical operator is selected from the parameterized physical operator library. The local window corresponding to the window of the additional physical operator is found in the physical correction sequence. The local window is transformed to the frequency domain to obtain the local spectrum. The additional physical operator is executed on the local spectrum to obtain the corrected local window. The corrected local window is restored to the time domain and merged with the unmodified region in the physical correction sequence to obtain the candidate physical correction sequence. The candidate physical correction sequence is iteratively verified until the preset number of iterations is reached or there are no failure features, and the final physical correction sequence is obtained. The final candidate physical correction sequence is subjected to task calibration transformation to obtain a specific task output. In this embodiment, the task is a missing value completion task. The task calibration uses the candidate physical correction sequence to restore the missing position and retains the original observations in the input time series sequence to generate a completed sequence. For the anomaly detection task, an anomaly score is generated based on the difference between the input time series sequence and the physical correction result and the degree of relationship failure. For the prediction task, the final physical correction sequence is converted into a prediction sequence within a preset future time series range.

[0088] Task calibration conversion is the process of converting the final physical correction sequence into a specific task result. All three types of tasks share the preceding time-series encoding, physical operator injection, and relational correction processes, but employ different output methods depending on the task type. Missing value completion: Existing observations in the input time-series are retained, and only the values ​​at corresponding positions in the final physical correction sequence are used to fill in missing positions, generating a completed sequence. Anomaly detection: The difference between the input time-series and the final physical correction sequence is calculated, and after normalization and sensor channel aggregation, anomaly scores are generated for each time step. State prediction: The task header has mapped the encoded features to a preset future time range, and subsequent physical corrections are also performed within this prediction range. The task calibration output is the final accepted future prediction sequence; this should not be interpreted as the historical sequence being converted into a future sequence only after all physical corrections are completed.

[0089] Example 2 The difference between this embodiment and Embodiment 1 is that this embodiment provides a structural health monitoring system, including: The feature extraction module is used to preprocess the collected structural health monitoring data to obtain the input time series sequence. The input time series sequence is then processed to generate a preliminary inference sequence. The preliminary inference sequence and the task visible context are processed in the frequency domain to form spectral features.

[0090] The process of processing the input time series sequence to generate a preliminary inference sequence includes: The input time series is divided into multiple time segments to obtain a time segment sequence. A learnable time summary tag is inserted at the beginning of the time segment sequence. The time summary tag and the time segment sequence are fed into a bidirectional long short-term memory encoder to obtain the encoded features. The encoded features are projected and the projection result is restored to the required time series length and sensor variable dimension to obtain the preliminary inference sequence.

[0091] The process of forming spectral features from the initial inference sequence and task visible context after frequency domain processing includes: Based on the input time series, a task visibility context is constructed. Fast Fourier Transform is performed on the preliminary inference sequence and the task visibility context respectively to obtain the spectral representation of the input time series and the spectral representation of the task visibility context. The spectral representations of the time series and the task visibility context are aligned and concatenated to obtain the complete spectrum. The average spectral amplitude of the preliminary inferred sequence and the task visible context at each frequency position is calculated separately, and normalized using the average spectral amplitude of all frequency positions of the corresponding sequence; the scores of the normalized amplitude are summed according to preset weights to obtain the importance score of each frequency position; the top K frequency positions are selected according to the importance score, and the joint spectral features corresponding to the top K frequency positions are extracted to obtain the spectral features.

[0092] The physical operator injection module is used to adaptively select and combine physical operators based on spectral characteristics, generate an operator scheduling queue, and apply the operator scheduling queue to the preliminary inference sequence to obtain the physical correction sequence.

[0093] See Figure 3 The process of adaptively selecting and combining physical operators based on spectral characteristics to generate an operator scheduling queue includes: The spectral features are mapped to a unified operator selection space to obtain the feature projection result. The feature projection result is input into the recurrent decoder. At each decoding position, the recurrent decoder selects a physical operator from the parameterized physical operator library to form an ordered operator scheduling queue. The order of the physical operators in the operator scheduling queue determines the execution order of the physical operators. The parameterized physical operator library includes physical operators such as identity mapping, low-pass filtering, high-pass filtering, band-pass filtering, phase shifting, response delay adjustment, amplitude scaling, and spectral smoothing. Each physical operator has corresponding parameters.

[0094] The process of applying the operator scheduling queue to the initial inference sequence to obtain the physical correction sequence includes: A Fast Fourier Transform is performed on the initial inferred sequence to obtain a spectral representation. The spectral coefficients of the first K frequency positions are extracted from the spectral representation. The physical operators in the operator scheduling queue are executed sequentially on the first K spectral coefficients according to their order, resulting in K modified spectral coefficients. The modified spectral coefficients are written back to the original frequency positions in the spectral representation to obtain a modified spectral representation. An Inverse Fast Fourier Transform is performed on the modified spectral representation to obtain a physical correction sequence.

[0095] The physical relationship revealing module is used to establish relationship references from the input time series sequence, identify failure features in the physical correction sequence and construct failure modes, select additional physical operators based on the failure modes, perform local injection on the physical correction sequence based on the additional physical operators to obtain candidate physical correction sequences, iteratively verify the candidate physical correction sequences until a preset number of iterations is reached or no failure features are found, and obtain the final physical correction sequence. The task output result is obtained based on the final physical correction sequence, and the time series analysis model is obtained after training.

[0096] See Figure 4 The process of establishing relational references based on input time series, identifying failure features in physical correction sequences, and constructing failure modes includes: The input time series and physical correction sequence are divided into position-aligned local windows. The relational representation of each local window is extracted through the same multi-view observation to obtain the relational representation feature vector of the input time series and physical correction sequence. Clustering is performed on the relational representation feature vectors of the input time series to obtain reference cluster centers, and the distance from the relational representation feature vectors of the input time series to the reference cluster centers is used as the reference neighborhood; The relation representation feature vector of the physical correction sequence is mapped to the reference neighborhood. Relation representations located outside the reference neighborhood are identified as failure features. The features are ranked according to the normalization deviation of the failure features relative to the nearest reference neighborhood. The top-ranked features are organized into failure modes.

[0097] The process of obtaining candidate physical correction sequences by locally injecting additional physical operators into the physical correction sequence includes: Find the local window corresponding to the window of the additional physical operator in the physical correction sequence, transform the local window to the frequency domain to obtain the local spectrum, perform the additional physical operator on the local spectrum to obtain the corrected local window, restore the corrected local window to the time domain and merge it with the unmodified region in the physical correction sequence to obtain the candidate physical correction sequence.

[0098] The process of iteratively verifying candidate physics correction sequences includes: dividing the current candidate physics correction sequence into multiple local windows, extracting the relational representation of each local window through multi-view observation, and obtaining the relational representation feature vector of the candidate physics correction sequence; Map the relation representation feature vector of the current candidate physical correction sequence to the reference neighborhood, identify the relation representation outside the reference neighborhood as the current failure feature, obtain the current failure mode based on the current failure feature, calculate the scores of the current failure mode and the previous round failure mode respectively, and the score of the failure mode is the sum of the squared Euclidean distance from each failure feature in the failure mode to the nearest cluster center and the deviation of the squared Euclidean distance from the reference neighborhood threshold. If the difference between the score of the current failure mode and the score of the previous failure mode reaches the improvement amount, the current candidate physics correction sequence is accepted and proceeds to the next iteration; otherwise, the current candidate physics correction sequence is not accepted and the previous candidate physics correction sequence is returned.

[0099] The task output module is used to input the input time series sequence to be analyzed into the time series analysis model to obtain the time series analysis results. The time series analysis model includes a trained bidirectional long short-term memory encoder, a task head, a physical operator parameter generator, a recurrent decoder, and a multilayer perceptron selector. The time series analysis results are any one of the following: a completed sequence, an anomaly score, or a predicted sequence.

[0100] This invention is based on a bidirectional long short-term memory encoder. By introducing a memory-guided physical operator injection module and a multi-view physical relationship revealing module, it achieves a step-by-step process from initial temporal inference, preservation of intrinsic physical relationships to the revelation of hidden physical relationships. Furthermore, it generates complete sequences, anomalous scores, or predicted sequences through task calibration.

[0101] Experimental Analysis 1. Dataset This experiment is based on a rich dataset of long-term bridge SHM records from China. This dataset contains monitoring data from multiple bridges, each typically providing one to two years of continuous multivariate sensor observation data under varying environmental and operational conditions. Two representative bridge datasets were selected for task evaluation: the "Cihuai Xinhe Bridge Dataset (CNRB)," covering 43 monitoring variables from April 10, 2024 to April 10, 2025, was used for in-data evaluation; the "Quanhe Bridge Dataset (QRB)," spanning from June 6, 2023 to June 2, 2025, contains 87 raw monitoring channels and was used for out-of-data bridge evaluation. All reported data were acquired at one-minute sampling intervals. To ensure comparability of task evaluations, the QRB dataset was further aligned with the CNRB using the same 43-variable structure. These two datasets were used to evaluate three representative SHM time series tasks: missing value imputation, anomaly detection, and predictive analysis.

[0102] 2. Evaluation Indicators For data imputation, we report MSE and MAE; for prediction, we report MSE, MAE, and RMSE. MSE focuses on reflecting significant bias, MAE measures the average pointwise error magnitude, and RMSE provides an error scale consistent with the original signal magnitude. For anomaly detection, we report accuracy, precision, recall, and F1 score.

[0103] 3. Implementation details All datasets were divided into training, validation, and test sets in chronological order, with a ratio of 6:2:2. During data imputation, the system generated random masks with missing rates of 0.250, 0.375, and 0.500. For anomaly detection, the training and validation sets contained only normal observations, while approximately 8% of the test sequences exhibited spike anomalies, horizontal shift anomalies, and drift anomalies [31-34]. In the prediction task, the historical window length, label length, and prediction duration were set to 96, 48, and 24, respectively. All models were implemented using PyTorch version 2.2.1+cu118. The input window length for interpolation and prediction tasks was 96, while it was 100 for anomaly detection tasks. The image patch length and stride for each task were 16 and 8, respectively. The Top-K frequency values ​​were 24, 16, and 12 for the three tasks, respectively. Training employed the Adam optimizer with a cosine learning rate decay function. The second-stage training lasted a maximum of 20 epochs.

[0104] 4. Experimental Results (1) Missing value completion Table 1 shows the imputation results on the CNRB and QRB datasets. Bold values ​​in Table 1 represent the best results, and underlined values ​​represent the second-best results. The values ​​0.500, 0.375, and 0.250 in the second row of Table 1 represent missing values. On the CNRB dataset, the proposed PhyRe model (a physical reification foundation model) achieves the best average MAE / MSE values ​​(0.033 / 0.011, respectively), reducing MAE by approximately 15.4% compared to the second-best results and reducing MSE by approximately 8.3% compared to the SymTime scheme. On the QRB dataset, PhyRe also achieves the best average MAE / MSE values ​​(0.030 / 0.014), improving MAE by approximately 3.4% compared to the MTSCI scheme and improving MSE by approximately 6.7% compared to the SymTime scheme. Compared to interpolators based on recurrent or attention mechanisms (such as brits and SAITS), PhyRe exhibits significantly lower reconstruction errors, demonstrating stronger modeling capabilities for time and intervariate dependencies. Furthermore, compared to interpolators based on diffusion mechanisms or low-rank Transformer architectures (such as MTSCI and ImputeFormer), PhyRe demonstrates more stable performance across different missing rates. These results suggest that MPOI recovers missing responses through physically guided spectral feature cues, while MPRR utilizes observational anchors and multi-perspective evidence to maintain the stability of the reconstruction results.

[0105] Table 1. Imputation results on CNRB and QRB datasets

[0106] (2) Anomaly detection Table 2 presents the anomaly detection results. Bold values ​​represent the best results, and underlined values ​​represent the second-best results. On the CNRB dataset, PhyRe achieved the best precision, accuracy, recall, and F1 score (0.978, 0.903, 0.823, and 0.861, respectively), exceeding the second-best results by 1.9, 14.7, 2.7, and 12.0 percentage points. The simultaneous improvement in precision and recall indicates that PhyRe enhances overall detection capabilities, rather than focusing solely on a single category of detections. On the QRB dataset, PhyRe achieved the best F1 score (0.713).

[0107] Table 2 Anomaly Detection Results

[0108] (3) Prediction Table 3 shows the prediction results of PhyRe on the CNRB and QRB datasets. Bold values ​​in Table 3 represent the best results, and underlined values ​​represent the second-best results. On the CNRB dataset, PhyRe achieved the best MAE value (0.028), a reduction of approximately 26.3% compared to the second-best iTransformer model and a reduction of 30.0% compared to the SymTime baseline model; its MSE and RMSE indices were comparable to the best results and both lower than SymTime. On the QRB dataset, PhyRe also achieved the best MAE value (0.020), an improvement of approximately 23.1% compared to the second-best iTransformer model and approximately 25.9% compared to the SymTime model, while its MSE and RMSE indices remained close to the optimal levels. Compared to decomposition-based methods and linear prediction models (such as Autoformer and LTSF-Linear), PhyRe produces smaller prediction errors, indicating its stronger ability to model nonlinear structural response evolution. Compared to fragment-based PatchTST and inverse channel iTransformer models, its MAE value is significantly reduced without significantly increasing the squared error or scale-sensitive bias. These results demonstrate that PhyRe can effectively reduce pointwise prediction errors while maintaining stable overall prediction performance.

[0109] Table 3. Prediction results of the present invention on CNRB and QRB datasets.

[0110] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A structural health monitoring method, characterized in that: include: The collected structural health monitoring data is preprocessed to obtain the input time series sequence. The input time series sequence is then processed to generate the preliminary inference sequence. The preliminary inference sequence and the task visible context are processed in the frequency domain to form spectral features. Based on the spectral characteristics, physical operators are adaptively selected and combined to generate an operator scheduling queue. The operator scheduling queue is then applied to the preliminary inference sequence to obtain the physical correction sequence. A relational reference is established based on the input time series, failure features in the physical correction sequence are identified and failure modes are constructed, additional physical operators are selected based on the failure modes, and local injection is performed on the physical correction sequence based on the additional physical operators to obtain candidate physical correction sequences. The candidate physical correction sequences are iteratively verified until a preset number of iterations is reached or there are no failure features, and the final physical correction sequence is obtained. The task output result is obtained based on the final physical correction sequence, and the time series analysis model is obtained after training. Input the time series sequence to be analyzed into the time series analysis model to obtain the time series analysis results.

2. The structural health monitoring method according to claim 1, characterized in that: The process of processing the input time series sequence to generate a preliminary inference sequence includes: The input time series is divided into multiple time segments to obtain a time segment sequence. A learnable time summary tag is inserted at the beginning of the time segment sequence. The time summary tag and the time segment sequence are fed into a bidirectional long short-term memory encoder to obtain the encoded features. The encoded features are projected and the projection result is restored to the required time series length and sensor variable dimension to obtain the preliminary inference sequence.

3. The structural health monitoring method according to claim 1, characterized in that: The process of forming spectral features from the initial inference sequence and task visible context after frequency domain processing includes: Based on the input time series, a task visibility context is constructed. Fast Fourier Transform is performed on the preliminary inference sequence and the task visibility context respectively to obtain the spectral representation of the input time series and the spectral representation of the task visibility context. The spectral representations of the time series and the task visibility context are aligned and concatenated to obtain the complete spectrum. The average spectral amplitude of the preliminary inferred sequence and the task visible context at each frequency position is calculated separately, and normalized using the average spectral amplitude of all frequency positions of the corresponding sequence; the scores of the normalized amplitude are summed according to preset weights to obtain the importance score of each frequency position; the top K frequency positions are selected according to the importance score, and the joint spectral features corresponding to the top K frequency positions are extracted to obtain the spectral features.

4. The structural health monitoring method according to claim 1, characterized in that: The process of adaptively selecting and combining physical operators based on spectral characteristics to generate an operator scheduling queue includes: The spectral features are mapped to a unified operator selection space to obtain the feature projection result. The feature projection result is input into the recurrent decoder. At each decoding position, the recurrent decoder selects a physical operator from the parameterized physical operator library to form an ordered operator scheduling queue. The order of the physical operators in the operator scheduling queue determines the execution order of the physical operators. The parameterized physical operator library includes physical operators such as identity mapping, low-pass filtering, high-pass filtering, band-pass filtering, phase shifting, response delay adjustment, amplitude scaling, and spectral smoothing. Each physical operator has corresponding parameters.

5. The structural health monitoring method according to claim 1, characterized in that: The process of applying the operator scheduling queue to the initial inference sequence to obtain the physical correction sequence includes: A Fast Fourier Transform is performed on the initial inferred sequence to obtain a spectral representation. The spectral coefficients of the first K frequency positions are extracted from the spectral representation. The physical operators in the operator scheduling queue are executed sequentially on the first K spectral coefficients according to their order, resulting in K modified spectral coefficients. The modified spectral coefficients are written back to the original frequency positions in the spectral representation to obtain a modified spectral representation. An Inverse Fast Fourier Transform is performed on the modified spectral representation to obtain a physical correction sequence.

6. The structural health monitoring method according to claim 1, characterized in that: The process of establishing relational references from input time series, identifying failure features in physical correction sequences, and constructing failure modes includes: The input time series and physical correction sequence are divided into position-aligned local windows. The relational representation of each local window is extracted through the same multi-view observation to obtain the relational representation feature vector of the input time series and physical correction sequence. Clustering is performed on the relational representation feature vectors of the input time series to obtain reference cluster centers, and the distance from the relational representation feature vectors of the input time series to the reference cluster centers is used as the reference neighborhood; The relation representation feature vector of the physical correction sequence is mapped to the reference neighborhood. Relation representations located outside the reference neighborhood are identified as failure features. The features are ranked according to the normalization deviation of the failure features relative to the nearest reference neighborhood. The top-ranked features are organized into failure modes.

7. The structural health monitoring method according to claim 1, characterized in that: The process of obtaining candidate physical correction sequences by locally injecting additional physical operators into the physical correction sequence includes: Find the local window corresponding to the window of the additional physical operator in the physical correction sequence, transform the local window to the frequency domain to obtain the local spectrum, perform the additional physical operator on the local spectrum to obtain the corrected local window, restore the corrected local window to the time domain and merge it with the unmodified region in the physical correction sequence to obtain the candidate physical correction sequence.

8. The structural health monitoring method according to claim 1, characterized in that: The process of iteratively verifying candidate physics correction sequences includes: dividing the current candidate physics correction sequence into multiple local windows, extracting the relational representation of each local window through multi-view observation, and obtaining the relational representation feature vector of the candidate physics correction sequence; Map the relation representation feature vector of the current candidate physical correction sequence to the reference neighborhood, identify the relation representation outside the reference neighborhood as the current failure feature, obtain the current failure mode based on the current failure feature, calculate the scores of the current failure mode and the previous round failure mode respectively, and the score of the failure mode is the sum of the squared Euclidean distance from each failure feature in the failure mode to the nearest cluster center and the deviation of the squared Euclidean distance from the reference neighborhood threshold. If the difference between the score of the current failure mode and the score of the previous failure mode reaches the improvement amount, the current candidate physics correction sequence is accepted and proceeds to the next iteration; otherwise, the current candidate physics correction sequence is not accepted and the previous candidate physics correction sequence is returned.

9. The structural health monitoring method according to claim 1, characterized in that: The time series analysis model includes a trained bidirectional long short-term memory encoder, a task head, a physical operator parameter generator, a recurrent decoder, and a multilayer perceptron selector. The time series analysis result is any one of the following: a completed sequence, an anomaly score, or a predicted sequence.

10. A structural health monitoring system, characterized in that: include: The feature extraction module is used to preprocess the collected structural health monitoring data to obtain the input time series sequence. The input time series sequence is processed to generate a preliminary inference sequence. The preliminary inference sequence and the task visible context are processed in the frequency domain to form spectral features. The physical operator injection module is used to adaptively select and combine physical operators based on spectral characteristics, generate an operator scheduling queue, and apply the operator scheduling queue to the preliminary inference sequence to obtain the physical correction sequence. The physical relationship revealing module is used to establish relationship references from the input time series, identify failure features in the physical correction sequence and construct failure modes, select additional physical operators based on the failure modes, perform local injection on the physical correction sequence based on the additional physical operators to obtain candidate physical correction sequences, iteratively verify the candidate physical correction sequences until a preset number of iterations is reached or no failure features are found, and obtain the final physical correction sequence. The task output result is obtained based on the final physical correction sequence, and the time series analysis model is obtained after training. The task output module is used to input the input time series sequence to be analyzed into the time series analysis model to obtain the time series analysis results.

Images