Civil structure deformation anomaly detection method based on time series data

Abstract

The application provides a civil structure deformation anomaly detection method based on time series data, and relates to the field of civil structure deformation anomaly detection; a disturbance intensity response value, a disturbance curvature and a local disturbance folding feature are constructed, a range adjustment enhancement value is formed by combining a symbol jump flag and a neighborhood disturbance difference value, a latent guide feature is generated based on the range adjustment enhancement value, a disturbance reconstruction feature is constructed by superimposing a diffusion residual and an asymmetric difference item, normalized features are obtained through normalization mapping, a disturbance spinor modulation factor is constructed by combining a nonlinear inhibition correlation coefficient, a disturbance spinor tensor is formed by applying orthogonal phase coding to a logarithmic compression channel and a square amplification channel, and a tensor potential mapping atlas is generated under a path coupling and self-coupling mechanism; on the basis of the above, a disturbance energy offset and an extreme value deflection are constructed, a probability potential index is formed, a civil structure deformation anomaly detection model is trained based on the probability potential index, and the detection of civil structure deformation anomalies is realized.

Description

Technical Field

[0001] This invention belongs to the field of civil structure deformation anomaly detection, specifically relating to a method for civil structure deformation anomaly detection based on time-series data. Background Technology

[0002] As the scale of large-scale civil engineering structures such as bridges, tunnels, and buildings continues to increase, even minute deformations during their long-term service have a significant impact on structural safety. Existing structural health monitoring systems typically rely on multi-source sensors such as displacement gauges, strain gauges, and accelerometers to collect time-series response data of structures to assess the presence of potential damage. However, traditional deformation anomaly identification methods are mostly based on threshold discrimination, statistical fluctuation detection, or empirical model construction, which are sensitive to complex environmental disturbances and struggle to accurately characterize the dynamic evolution of structures under different working conditions. Furthermore, traditional methods still suffer from high false positive rates, response lag, and insufficient model generalization when facing factors such as sensor noise, seasonal trends, and load cycles in the early identification of abnormal deformations. Therefore, it is necessary to propose a deformation anomaly detection method that can automatically extract structural deformation patterns from time-series data and is adaptable to non-stationary disturbances, in order to improve the safety assessment capabilities of civil structures in long-term monitoring scenarios.

[0003] Publication No. CN115615386A calculates whether the structure has deformed by comprehensively analyzing the trace of the monitoring matrix, settlement overshoot parameters, and vibration overshoot parameters. When the model result exceeds a set threshold, it is determined that there is abnormal deformation. This scheme mainly realizes the comprehensive monitoring of structural settlement and vibration through matrix modeling, and belongs to the structural deformation identification method based on the fusion of static and dynamic parameters.

[0004] In existing technologies, civil structure deformation monitoring mainly relies on traditional monitoring equipment such as total stations, levels, and strain gauges to obtain settlement, displacement, or vibration response characteristics. However, these instruments often suffer from long monitoring cycles, low acquisition frequency, and difficulty in achieving automation and continuous operation. Furthermore, the measurement process is easily affected by external environmental factors such as climate, lighting, and foundation stability, leading to insufficient long-term monitoring accuracy. At the same time, traditional methods often focus on a single parameter and lack the ability to model the multidimensional temporal characteristics of civil structures over time, making it difficult to capture early, subtle, and hidden abnormal deformations in a timely manner. Therefore, there is an urgent need for a new method for detecting abnormal deformations in civil structures that can integrate multi-source time-series data, possess real-time performance, and high sensitivity, in order to improve monitoring reliability and response speed and provide more accurate early warning capabilities for structural safety. Summary of the Invention

[0005] This invention proposes a method for detecting deformation anomalies in civil structures based on time-series data. By performing time-slicing processing on multi-source deformation data, the temporal correlation of each monitoring channel is maintained at a unified time scale. Based on this, disturbance intensity response values, disturbance curvature, and local disturbance folding features are constructed sequentially. These are combined with sign jump indicators and neighborhood disturbance differences to form range-adjusted enhancement values, achieving a joint characterization of disturbance direction changes and local abrupt changes. Furthermore, latent guidance features are generated based on the temporal variation relationship of the range-adjusted enhancement values. Disturbance reconstruction features are constructed by superimposing diffusion residuals and asymmetric difference terms, completing a normalized mapping to obtain normalized features. These normalized features are then used as input. A perturbation spinor modulation factor is constructed by combining nonlinear suppression correlation coefficients. A perturbation spinor tensor is formed by applying orthogonal phase encoding through logarithmic compression and square amplification channels. A multipath structure coupling tensor is then constructed under path coupling and self-coupling mechanisms to further generate a tensor potential mapping spectrum. Based on the tensor potential mapping spectrum, the theoretical symmetry center response is calculated in the symmetric time neighborhood, and a probabilistic potential index is formed by constructing perturbation energy offset and extreme value deflection. Finally, a probabilistic potential index matrix is ​​constructed using the probabilistic potential index, and the deformation state error loss function is combined to complete the training of the civil structure deformation anomaly detection model, realizing the temporal discrimination and modeling of civil structure deformation anomalies.

[0006] A method for detecting deformation anomalies in civil structures based on time-series data, the specific method is as follows:

[0007] S1. Collect civil structure state data and preprocess the collected civil structure state data to obtain civil structure deformation training set and civil structure deformation test set.

[0008] S2. Slice the civil structure deformation training set and the civil structure deformation test set to construct the civil structure deformation time series training set and the civil structure deformation time series test set.

[0009] S3. The change amplitude of the feature vector between the current window time step and the previous time step is used to construct the disturbance intensity response value. The disturbance curvature is calculated based on the disturbance intensity response values ​​at adjacent time steps and combined with the folding adjustment factor to generate local disturbance folding features. The sign difference between the local disturbance folding features of the current and previous time steps is compared to construct the sign jump flag. The difference between the local disturbance folding features of adjacent time steps is calculated and combined with the control coefficient to obtain the disturbance adjustment term. The sign jump flag and the neighborhood disturbance difference are fused to generate the range adjustment enhancement value.

[0010] S4. Calculate and normalize the difference ratio between the range adjustment enhancement value and the previous time step, and introduce the perturbation guidance adjustment coefficient to generate the latent guidance feature. Construct the diffusion residual through the temporal difference of the latent guidance feature, and generate the perturbation reconstruction feature by superimposing the asymmetric difference term of the range adjustment enhancement value. Normalize the perturbation reconstruction feature and combine it with the range adjustment enhancement value mapping to obtain the normalized feature.

[0011] S5. Using the normalized feature as input, the perturbation spinor modulation factor is constructed by combining the nonlinear suppression correlation coefficient and the amplitude is modulated. The perturbation spinor tensor is formed by applying orthogonal phase encoding through the logarithmic compression channel and the square amplification channel. The multipath structure coupling tensor is constructed by the perturbation spinor tensor and its transpose, combined with the normalized feature, and the tensor potential mapping spectrum is formed under the nonlinear suppression coefficient and scale compression processing.

[0012] S6. Using the tensor potential mapping spectrum as input, construct the channel-level theoretical symmetry center response in the symmetric time neighborhood, and obtain the perturbation energy offset by comparing each channel. Based on the directional distinction relationship of the perturbation energy offset in the previous and next time steps, construct the extreme value deflection corresponding to the current time point. Perform normalization mapping on the extreme value deflection in the local time neighborhood to form the probability potential index corresponding to the current time point.

[0013] S7. Construct a civil structure deformation anomaly detection model. Input the civil structure deformation time series training set, and use the joint deformation state error loss function to pass through steps S2 to S6 in sequence. Train the civil structure deformation anomaly detection model until it converges to realize the anomaly detection of civil structure deformation.

[0014] Preferably, in step S1, the civil structure deformation dataset is obtained through a combination of scaled-down model tests, long-term operational monitoring data of existing projects, finite element numerical simulation calculations, and indoor controlled loading tests. A complete time-series sample system is constructed by utilizing the complementarity of measured and simulated data in the time dimension. Deformation monitoring points are set up at beam ends, column bases, support nodes, mid-span locations, foundation top surfaces, and near expansion joints. Displacement sensors, tilt sensors, strain sensors, acceleration sensors, temperature sensors, humidity sensors, and crack width gauges are installed at each monitoring point. These sensors are connected to a unified monitoring platform via wired and wireless acquisition units, with a sampling frequency set to 1 time per second. Vertical displacement, horizontal displacement, component rotation, and reinforcement deformation are measured. Strain, concrete surface strain, vertical acceleration, horizontal acceleration, ambient temperature, ambient humidity, crack width, and expansion joint opening and closing volume are continuously collected. The monitoring point number, installation height, floor, component type, load condition identifier, construction stage identifier, and corresponding timestamp are recorded simultaneously to form multi-source deformation time series raw data. After all the collected multi-source deformation time series raw data are transmitted to the central server, time alignment, invalid segment removal, outlier cleaning based on physical thresholds and statistical criteria, missing data imputation, dimension unification, and numerical normalization are performed in sequence. The processed data are used to construct a civil structure deformation dataset, which is then divided into a civil structure deformation training set and a civil structure deformation test set.

[0015] Preferably, in step S2, under the condition of a sampling frequency of 1 time per second, the civil structure deformation dataset is processed by sliding slices in chronological order to construct a time window as the model input. Each time window retains the temporal change features of the high-dimensional vector of all physical and environmental features. The structural deformation probability corresponding to the continuous sampling points after the end of each input window is used as the prediction target. By combining the time window and the prediction target window, a civil structure deformation time series training set and a civil structure deformation time series test set are constructed.

[0016] Furthermore, under the condition of a sampling frequency of 1 time per second, the civil structure deformation dataset is processed by sliding slices in chronological order to construct time windows as model inputs. Each time window retains the temporal change features of the high-dimensional vector of all physical and environmental characteristics. The structural deformation probability corresponding to the continuous sampling points after the end of each input window is used as the prediction target. By combining the time windows and the prediction target windows, a civil structure deformation time series training set and a civil structure deformation time series test set are constructed. The above method, combined with the needs of modeling structural deformation trends and identifying anomalies, not only ensures the complete linkage between the original physical features and environmental disturbances, but also enhances the ability to characterize the evolution path of structural behavior within continuous time series through sliding slices. This helps the model capture potential deformation patterns that accumulate slowly and change gradually, and enhances the fitting effect and early warning capability of deformation probability response. Thus, it meets the comprehensive requirements of data integrity, prediction sensitivity and engineering practicality in the scenario of long-term monitoring and risk identification of civil structures.

[0017] Preferably, in step S3, the change amplitude between the feature vector of the current window time step and the feature vector of the previous time step is calculated to obtain the disturbance intensity metric value. The disturbance intensity metric value of the current time step and the previous time step is normalized while a minimal constant stable denominator is introduced to construct the disturbance activation factor. Based on this, a nonlinear suppression coefficient is combined and the disturbance intensity response value is calculated through an exponential function.

[0018] Based on the disturbance intensity response values ​​at the current time point and the two adjacent times before and after it, the disturbance curvature at the current time point is calculated. Based on the disturbance curvature, a folding adjustment factor that includes the disturbance intensity response value and the folding adjustment coefficient is introduced. Combining the disturbance curvature and the folding adjustment factor, the local disturbance folding characteristics corresponding to the current time window are output.

[0019] By comparing the sign of the current local perturbation folding feature with that of the previous time step, and combining the adjustment factor to suppress the response intensity of the abrupt change region, a sign jump sign is formed; the neighborhood perturbation difference of the local perturbation folding feature before and after the current time step is introduced, and the neighborhood perturbation difference is scaled by the control coefficient to form a perturbation adjustment term; the sign jump sign and the perturbation adjustment term are multiplied to construct the range adjustment enhancement value.

[0020] Furthermore, by constructing disturbance intensity response values, local disturbance folding features, and range adjustment enhancement values, the fine-grained perception capability of the civil structure deformation anomaly detection model for time-varying structures is effectively enhanced. Specifically, the disturbance intensity response value measures the change amplitude of the feature vector between the current time step and the previous time step, and introduces a nonlinear suppression mechanism to regulate amplitude fluctuations. On this basis, the disturbance curvature is constructed by the difference between adjacent disturbance intensity response values, and further combined with the folding adjustment factor to generate local disturbance folding features, thereby highlighting the response trend in continuous changes. Subsequently, a sign jump flag is generated by judging the sign difference between the local disturbance folding features of the current and previous time steps to capture significant abrupt changes in the disturbance direction. Then, the difference between the local disturbance folding features between adjacent time steps is combined to construct the neighborhood disturbance difference value, and its amplitude is adjusted by combining the control coefficient. Finally, the sign jump flag and the disturbance adjustment term are fused to generate the range adjustment enhancement value, realizing the dual regulation of dynamic enhancement and abrupt suppression of local disturbance behavior at the time step, which helps to improve the feature sensitivity and robustness of the civil structure deformation anomaly detection model under complex time series.

[0021] Preferably, in step S4, the difference between the current range adjustment enhancement value and the previous time step range adjustment enhancement value is introduced, the change ratio of the two is calculated and the ratio is normalized, and a stability constant is added to ensure the numerical stability of the calculation process; on this basis, the normalized change ratio is adjusted by the perturbation guidance adjustment coefficient, and the adjustment result is combined with the current range adjustment enhancement value to obtain the current latent guidance characteristics.

[0022] Using the latent guidance features at the current time as input, the latent guidance features of the previous time step are introduced to construct the diffusion residual information between adjacent time steps. The diffusion residual information is smoothed and restricted using a nonlinear function. The asymmetric difference term formed by the range adjustment enhancement value between adjacent time steps is combined and the asymmetric difference term is constrained. The diffusion residual term and the asymmetric suppression term are superimposed on the current latent guidance features in a weighted form, thereby constructing the perturbation reconstruction features at the current time step.

[0023] The perturbation reconstruction features are normalized, and the range adjustment enhancement value is transformed into a dynamic weight term through a mapping function. The perturbation reconstruction features and perturbation enhancement weights are then operated on element by element. The result of the element-by-element operation is then fused with the normalization operation of the perturbation reconstruction features to obtain the normalized features after perturbation reconstruction.

[0024] Furthermore, the variation ratio of the range adjustment enhancement value to its previous moment is introduced, and through normalization and stabilization processing, the disturbance change participates in the modeling in a relative form, which can meticulously characterize the small fluctuations and gradual evolution trends in the disturbance sequence. On this basis, the disturbance-oriented adjustment coefficient is integrated to generate latent guidance features, so that the direction and magnitude of the variation of the range adjustment enhancement value in the time dimension are uniformly expressed, enhancing the ability to continuously characterize the disturbance evolution process. Furthermore, the temporal differences of the latent guidance features are used to construct diffusion residuals, and combined with the asymmetric difference terms formed by the range adjustment enhancement values ​​for superposition processing, so that the civil structure deformation anomaly detection model has strong structural adaptability when facing disturbance changes of different directions and magnitudes. Finally, by normalizing the disturbance reconstruction features and introducing dynamic weights obtained by mapping the range adjustment enhancement values ​​for fusion, the obtained normalized features can fully reflect the change characteristics of the disturbance in the time series while maintaining numerical stability, thus providing a fine and consistent disturbance response basis for subsequent civil structure deformation anomaly detection.

[0025] Preferably, in step S5, a perturbation screw modulation factor that adapts to the strength of the perturbation is constructed using a nonlinear suppression correlation coefficient and a normalized feature. Based on this, the normalized feature is uniformly modulated using the perturbation screw modulation factor. The modulated feature is then sent to the logarithmic compression channel and the square amplification channel, respectively, and orthogonal phase encoding is applied to them. Finally, the results of the two channels are coupled by an element-wise tensor outer product to form a perturbation screw tensor.

[0026] The perturbation screw tensor is transposed as input. The perturbation screw tensor and its transpose are combined element-wise to form a path coupling term that contains bidirectional perturbation correlation information. Based on this, the normalized feature after perturbation reconstruction is introduced and nonlinearly suppressed and mapped to construct the perturbation response vector. A self-coupling tensor is generated by vector outer product. The coupling factor and diffusion adjustment factor are adaptively generated according to the amplitude distribution of the normalized feature. The path coupling term and the self-coupling tensor are weighted and superimposed to obtain the multipath structure coupling tensor.

[0027] The multipath structure coupling tensor and normalized feature are mapped dimension-by-dimensionally at the tensor-vector level. A nonlinear suppression coefficient is introduced to adjust the amplitude of the mapping result. The result is then compressed to a uniform scale through continuous nonlinear mapping, ultimately forming a tensor potential mapping spectrum.

[0028] Furthermore, the normalized features, through layer-by-layer recombination and mapping, organize the originally dispersed multi-source deformation information into a representation with a clear structural hierarchy. The perturbation spinor modulation factor, guided by the nonlinear suppression correlation coefficient, can adaptively adjust the feature scale according to the strength of the perturbation, reducing the impact of environmental changes and differences in working conditions on feature consistency. The perturbation spinor tensor obtained by combining logarithmic compression channels and square amplification channels with orthogonal phase encoding enables perturbations at different amplitude levels to obtain complementary expressions in the same feature space. The multipath structural coupling tensor constructed by combining the normalized features after perturbation reconstruction characterizes the forward and reverse perturbation transmission paths between multiple monitoring quantities, and synchronously reflects the change characteristics of the perturbation of each monitoring quantity within the current time window. The resulting tensor potential mapping spectrum remains numerically stable under the constraint of the nonlinear suppression coefficient and highlights the perturbation differences, which is beneficial for subsequent detection of anomalies in civil structures and the discrimination of complex deformation states.

[0029] Preferably, in step S6, the tensor potential mapping spectrum is used as input, and several adjacent time steps before and after the current time point are selected to form a symmetrical time neighborhood. The maximum and minimum tensor potential mapping spectrum values ​​in the channel are calculated within the neighborhood, and the maximum and minimum values ​​are fused to obtain the theoretical symmetry center response. The tensor potential mapping spectrum value at the current time is compared with the theoretical symmetry center response channel by channel to obtain the deviation degree of each channel. The deviation results of all channels are aggregated to generate the perturbation energy offset corresponding to the current time point.

[0030] Select the perturbation energy offset corresponding to several time steps before and after, and distinguish the forward and backward time positions to form a local trend center value with time orientation. Calculate the perturbation energy offset and the local trend center value after direction distinction to construct the extreme value deflection corresponding to the current time point.

[0031] Extract all extreme value deflections within the local time neighborhood and sum them to obtain a normalized reference value; perform intensity mapping on the extreme value deflection corresponding to the current time point, and standardize the normalized reference value and the extreme value deflection to obtain the probability potential index of the current time point.

[0032] Furthermore, by using the tensor potential mapping spectrum as input and constructing a symmetric time neighborhood, a theoretical symmetry center response is introduced. This allows the perturbation energy offset to characterize the degree of symmetry disruption of the tensor potential in the time dimension from the perspective of the entire channel, thereby effectively reducing the interference of local random fluctuations on the judgment results and improving the stable representation ability of the real perturbation response. On this basis, by distinguishing the direction of the perturbation energy offset in the forward and backward time positions, an extreme value deflection is constructed, which allows the abrupt change characteristics caused by trend amplification or trend reversal during the anomaly evolution process to be explicitly expressed, enhancing the ability to characterize the critical moment of the anomaly. By normalizing the extreme value deflection in the local time neighborhood and constructing a probabilistic potential index, the anomaly responses at different time points have unified dimensions and comparability, thus providing an input index with probabilistic semantics and good engineering stability for the subsequent anomaly judgment model.

[0033] Preferably, in step S7, the probability potential indexes are constructed into a probability potential index matrix in chronological order and input into the multilayer perceptron structure to output the detection results of the future civil structure deformation anomaly detection model; by weighting and averaging the error between the detected value and the true value of the civil structure deformation anomaly detection model with the corresponding probability potential index matrix, a deformation state error loss function is constructed.

[0034] Input a time series training set of civil structure deformation, set hyperparameters and train a civil structure deformation anomaly detection model by combining deformation state error loss function, and realize the anomaly detection of civil structure deformation anomaly detection model.

[0035] Furthermore, the deformation state error loss function guides the civil structure deformation anomaly detection model to focus on the key time point errors at high anomaly probability locations during training, thereby achieving refined response optimization to disturbance trends. Specifically, the deformation state error loss function is based on the absolute error between the model's detected value and the true value. It is weighted and fused with the probability potential index matrix time-by-time, and the errors at each future time step are weighted and accumulated to form a unified deformation state error evaluation mechanism. This mechanism can dynamically adjust the importance distribution of the learning objective during the model training phase. The introduction of the probability potential index matrix significantly increases the weight of time periods with high anomaly probability in the loss function, thereby improving the model's ability to distinguish important disturbance regions and avoiding the problem of weak learning of anomaly regions caused by conventional error balancing mechanisms. At the same time, through weighted aggregation operations in the time dimension, the deformation state error loss function achieves holistic modeling of future time-series detection results, enhancing the adaptability of the civil structure deformation anomaly detection model to nonlinear disturbance evolution trends.

[0036] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0037] This invention processes multi-source deformation time-series data of civil structures into time slices and sequentially constructs disturbance intensity response values, disturbance curvature, and local disturbance folding features, enabling synchronous expression of disturbance change relationships across monitoring channels at a unified time scale. Furthermore, it introduces sign jump indicators and neighborhood disturbance differences to construct range-adjusted enhancement values, incorporating disturbance direction changes, local abrupt changes, and continuous change characteristics into the same disturbance modeling process, thus achieving a structured description of deformation disturbances in the time dimension.

[0038] This invention uses the range adjustment enhancement value as the core input to construct latent guidance features and combine diffusion residuals and asymmetric difference terms to generate perturbation reconstruction features. Normalized features are obtained through normalization mapping, so that the evolution direction and intensity of perturbation across different time steps can be uniformly described. On this basis, a perturbation spinor modulation factor is constructed to form a perturbation spinor tensor. Multipath structure coupling tensor is generated by combining path coupling and self-coupling mechanisms, and further mapped to obtain a tensor potential mapping spectrum. This allows the bidirectional perturbation correlation between multiple monitoring channels and the perturbation features of a single channel to be structurally expressed in a unified tensor space.

[0039] This invention uses a tensor potential mapping spectrum as input to construct a theoretical symmetry center response within a symmetric time neighborhood. It generates a probabilistic potential index through joint modeling of perturbation energy offset and extreme value deflection, enabling the deformation perturbation state at each time point to have a unified scale and temporal comparability. Furthermore, the probabilistic potential index is constructed into a probabilistic potential index matrix and trained with a deformation state error loss function. This allows the civil structure deformation anomaly detection model to update parameters around key locations of perturbation evolution, achieving continuous modeling and stability discrimination of civil structure deformation anomalies. Attached Figure Description

[0040] Figure 1 This is a flowchart of a method for detecting deformation anomalies in civil structures based on time-series data, provided by the present invention.

[0041] Figure 2 This is a structural diagram of the range adjustment enhancement value provided by the present invention.

[0042] Figure 3 This is a structural diagram of the normalization feature provided by the present invention.

[0043] Figure 4 This is a structural diagram of the tensor potential mapping spectrum provided by the present invention.

[0044] Figure 5 This is a structural diagram of the probability potential index provided by the present invention.

[0045] Figure 6 This is a graph showing the loss variation of the deformation state error loss function provided by the present invention.

[0046] Figure 7 This is a visualization diagram of the tensor potential mapping spectrum provided by the present invention.

[0047] Figure 8 This is a comparison chart of deformation anomaly detection results provided by the present invention. Detailed Implementation

[0048] This invention proposes a method for detecting deformation anomalies in civil structures based on time-series data. By performing time-slicing processing on multi-source deformation data, the temporal correlation of each monitoring channel is maintained at a unified time scale. Based on this, disturbance intensity response values, disturbance curvature, and local disturbance folding features are constructed sequentially. These are combined with sign jump indicators and neighborhood disturbance differences to form range-adjusted enhancement values, achieving a joint characterization of disturbance direction changes and local abrupt changes. Furthermore, latent guidance features are generated based on the temporal variation relationship of the range-adjusted enhancement values. Disturbance reconstruction features are constructed by superimposing diffusion residuals and asymmetric difference terms, completing a normalized mapping to obtain normalized features. These normalized features are then used as input. A perturbation spinor modulation factor is constructed by combining nonlinear suppression correlation coefficients. A perturbation spinor tensor is formed by applying orthogonal phase encoding through logarithmic compression and square amplification channels. A multipath structure coupling tensor is then constructed under path coupling and self-coupling mechanisms to further generate a tensor potential mapping spectrum. Based on the tensor potential mapping spectrum, the theoretical symmetry center response is calculated in the symmetric time neighborhood, and a probabilistic potential index is formed by constructing perturbation energy offset and extreme value deflection. Finally, a probabilistic potential index matrix is ​​constructed using the probabilistic potential index, and the deformation state error loss function is combined to complete the training of the civil structure deformation anomaly detection model, realizing the temporal discrimination and modeling of civil structure deformation anomalies.

[0049] Please see Figure 1 As shown in the figure, a method for detecting abnormal deformation of civil structures based on time-series data in an embodiment of this application has the following specific steps.

[0050] S1. Collect civil structure state data and preprocess the collected civil structure state data to obtain civil structure deformation training set and civil structure deformation test set.

[0051] In this embodiment, the deformation monitoring data of civil structures is not directly derived from actual bridge and extreme condition tests. Instead, it is obtained through a combination of scaled-down model experiments, long-term monitoring data of existing projects, numerical simulation calculations, and artificially controlled loading tests. A complete time-series sample is constructed by complementing simulation data and measured data, thereby meeting the research and verification needs of civil structure deformation anomaly detection models. Deformation monitoring points are set up at beam ends, column bases, support nodes, mid-span locations, foundation top surfaces, and near expansion joints. Displacement sensors, tilt sensors, strain sensors, and accelerometers are installed at each deformation monitoring point. Temperature sensors, humidity sensors, crack width gauges, and structural deformation probability values ​​are connected to a unified monitoring platform via wired and wireless acquisition units. The sampling frequency is set to once per second, continuously recording vertical displacement, horizontal displacement, component rotation angle, steel reinforcement strain, concrete surface strain, vertical acceleration, horizontal acceleration, ambient temperature, ambient humidity, crack width, expansion joint opening / closing amount, and simultaneously recording monitoring point number, installation height, floor level, component type, load condition identifier, construction condition identifier, and timestamp, forming multi-source deformation time-series raw data. After all monitoring signals are transmitted to the central server via industrial Ethernet or fieldbus, data is first filtered and time-aligned based on timestamps to remove duplicate records, equipment maintenance sections, and obviously offline sections. Then, an outlier removal step is performed, using a combination of the allowable displacement range of civil structures, strain limit values, acceleration safety thresholds, and the statistical three-standard-deviation rule to determine abnormal measurement points. Records that do not meet the physical and statistical constraints are marked as outliers and deleted from the dataset. Missing data segments in the remaining records are imputed, with continuous gaps not exceeding 300. The missing time segments were filled using linear interpolation. By calculating the time ratio between two adjacent sampling times for the missing time, the difference between the measurement values ​​of the later time and the previous time was multiplied by the time ratio and added to the measurement value of the previous time to obtain the estimated value of the missing time. Segments with continuous gaps exceeding 300 seconds were removed from the dataset. After interpolation, unit conversion and dimension unification were performed: displacement was unified to millimeters, strain to microstrain, acceleration to meters per second squared, temperature to degrees Celsius, humidity to relative humidity percentage, and crack width and expansion joint opening and closing amount to millimeters.Based on unified dimensions, the Min-Max normalization method is adopted. The minimum value of each dimension's sample feature value is subtracted, and then divided by the difference between the maximum and minimum values. Vertical displacement, horizontal displacement, component rotation angle, steel reinforcement strain, concrete surface strain, vertical acceleration, horizontal acceleration, ambient temperature, ambient humidity, crack width, expansion joint opening / closing, monitoring point height, floor number, component type coding, load condition coding, and construction condition coding are mapped to the [0,1] interval, forming a civil structure deformation dataset with 50,000 samples. The continuous monitoring data is divided into a civil structure deformation training set and a civil structure deformation test set at an 8:2 ratio. The training set contains 40,000 samples, and the test set contains 10,000 samples, providing standardized input data for subsequent time-series feature extraction, deformation pattern construction, and deformation anomaly detection.

[0052] S2. Slice the civil structure deformation training set and the civil structure deformation test set to construct the civil structure deformation time series training set and the civil structure deformation time series test set.

[0053] In this embodiment, using a civil structure deformation training set and a civil structure deformation test set with a sampling frequency of 1 time per second, the continuously monitored time series data is processed into sliding slices of fixed length. These slices are then arranged in chronological order to construct time windows of length 200 units. Each time window covers 200 consecutive seconds of structural operation. Within each time window, vertical displacement, horizontal displacement, component rotation angle, steel reinforcement strain, concrete surface strain, vertical acceleration, horizontal acceleration, ambient temperature, ambient humidity, crack width, expansion joint opening and closing, and synchronously recorded monitoring point numbers are fully preserved. The sequence of changes in the time-series characteristics of the component number, installation height, floor, component type, load condition identifier, construction condition identifier, and timestamp over 200 sampling times is combined to form a time slice. The data of the 30 consecutive sampling points after the end of the current slice are used as the prediction target window to characterize the deformation evolution trend and abnormal behavior of the structure in the next 30 seconds. The sliding step size is set to 5. Finally, the civil structure deformation time-series training set has a total of 7961 time slices and prediction target windows, and the civil structure deformation time-series test set has a total of 1961 time slices and prediction target windows.

[0054] Let the k-th time slice be represented as Each of them The prediction target corresponding to the k-th time slice is defined as a high-dimensional vector containing all physical and environmental features except for the deformation probability value. This is used to describe the probability of structural deformation within the next 30 seconds. By performing the above-mentioned sliding slice mapping on the civil structure deformation training set and the civil structure deformation test set, the civil structure deformation training set and the civil structure deformation test set are transformed into a large number of overlapping short time-series samples. This enables the model to learn the mapping relationship between the structural deformation pattern of the past 200 seconds and the deformation state of the next 30 seconds, thereby realizing near real-time structural deformation anomaly judgment and early warning, and providing a fast and stable decision basis for engineering safety monitoring.

[0055] S3. The change amplitude of the feature vector between the current window time step and the previous time step is used to construct the disturbance intensity response value. The disturbance curvature is calculated based on the disturbance intensity response values ​​at adjacent time steps, and the local disturbance folding feature is generated by combining the folding adjustment factor. The sign difference between the local disturbance folding features of the current and previous time steps is compared to construct the sign jump flag. The difference between the local disturbance folding features of adjacent time steps is calculated and the disturbance adjustment term is obtained by combining the control coefficient. The sign jump flag and the disturbance adjustment term are fused to generate the range adjustment enhancement value.

[0056] Furthermore, in step S3, the range adjustment enhancement value is established, and the process is as follows: Figure 2 As shown, the specific steps for establishing the range adjustment enhancement value are as follows.

[0057] S31. Calculate the change amplitude between the feature vector of the current window time step and the feature vector of the previous time step to obtain the disturbance intensity metric value. Normalize the disturbance intensity metric value between the current and previous time steps and introduce a minimal constant stable denominator to construct the disturbance activation factor. Based on this, combine the nonlinear suppression coefficient and complete the calculation of the disturbance intensity response value through an exponential function.

[0058] In this embodiment, a perturbation intensity suppression metric is constructed based on the local rate of change of the time slice sequence. By introducing the perturbation intensity metric value and the perturbation nonlinearity suppression function, weak perturbations in the sequence are suppressed, while regions of drastic change are preserved. The perturbation intensity suppression metric is calculated in the time slice. The Euclidean distance between each position and the previous time step yields a perturbation intensity metric. This perturbation intensity metric characterizes the magnitude of characteristic changes in the time series between adjacent time steps, reflecting whether a significant abrupt change has occurred at the current time point. The mathematical model for the perturbation intensity metric is as follows:

[0059] ;

[0060] in, For the first time slice Feature vectors at each time step; For the first time slice Feature vectors at each time step; Operations on Euclidean norms; For time window Mid-moment The disturbance intensity metric is used to capture the magnitude of local deformation changes; the larger the value, the greater the deformation at the current moment.

[0061] The perturbation nonlinear suppression function introduces a normalized ratio between perturbation intensity metrics as a perturbation activation factor, reflecting the significance of the current perturbation in the preceding and following perturbation sequences. An exponential decay function is used to nonlinearly suppress the perturbation activation factor, making the output of small perturbations close to 1 and the output of strong perturbations close to 0, thus forming a perturbation response with both stability and sensitivity. This avoids the noise sensitivity problem of traditional difference methods, and further improves the performance by adjusting the exponential coefficient. To achieve compressive strength control; the mathematical model of the disturbance nonlinear suppression function is:

[0062] ;

[0063] in, This is a nonlinear suppression coefficient, initially set to 4, used to control the steepness of the suppression curve; For perturbation activators; This is a very small constant used to prevent the denominator from being zero; its value is set to... ; It is a natural exponential function; The nonlinear suppression function for the perturbation within the time window Mid-moment The output disturbance intensity response value indicates that the smaller the value, the stronger the anomaly is likely.

[0064] S32. Based on the disturbance intensity response values ​​at the current time point and the two adjacent times before and after it, calculate the disturbance curvature at the current time point. Based on the disturbance curvature, introduce a folding adjustment factor that includes the disturbance intensity response value and the folding adjustment coefficient. Combine the disturbance curvature and the folding adjustment factor to output the local disturbance folding characteristics corresponding to the current time window.

[0065] In this embodiment, to address the issues of periodic disturbances, abrupt disturbances, and short-term noise interference in the disturbance intensity sequence, a disturbance curvature based on second-order difference is designed. While preserving key abrupt signals, a folding adjustment factor is introduced to stably suppress high-amplitude disturbances, thereby outputting a disturbance curvature sequence. The disturbance curvature is constructed by calculating the disturbance intensity response values ​​between the current time point and its two adjacent points, and is used to identify sudden trend reversals or regions of abrupt slope changes in disturbances. The mathematical model of the disturbance curvature is as follows:

[0066] ;

[0067] in, The nonlinear suppression function for the perturbation within the time window Mid-moment Output disturbance intensity response value; The nonlinear suppression function for the perturbation within the time window Mid-moment Output disturbance intensity response value; For time window Mid-moment The larger the curvature of the disturbance, the more prominent the trend of change;

[0068] A folding adjustment factor is introduced based on the calculation of disturbance curvature to form a disturbance folding adjustment function. When the original disturbance amplitude is large, the curvature value output is reduced to prevent excessive amplification of outliers. In the normal region, the disturbance folding adjustment function does not produce a suppression effect, thereby achieving dynamic smooth control of the high disturbance region. The mathematical model of the disturbance folding adjustment function is as follows:

[0069] ;

[0070] in, This is the folding adjustment coefficient, initially set to 0.5, which controls the suppression force and smooths the curvature. This is a folding adjustment factor; The disturbance folding adjustment function in the time window Mid-moment The output local perturbation folding features.

[0071] S33. By comparing the sign of the current local perturbation folding feature with that of the previous time step, and combining the adjustment factor to suppress the response intensity of the abrupt change region, a sign jump flag is formed; the neighborhood perturbation difference of the local perturbation folding feature before and after the current time step is introduced, and the neighborhood perturbation difference is scaled by the control coefficient to form a perturbation adjustment term; the sign jump flag and the perturbation adjustment term are multiplied to construct the range adjustment enhancement value.

[0072] In this embodiment, a range adjustment enhancement mechanism is constructed based on local perturbation folding features to solve the problem of information attenuation of the original perturbation features in the boundary ambiguity region and abnormal peak region. By adjusting the joint response of perturbation polarity, neighborhood range and continuity change, the enhancement of effective edge signals and the suppression of pseudo signals in the feature region are achieved.

[0073] By comparing the local disturbance folding characteristics at the current moment Compared to the previous moment The sign difference was calculated. If the differences are of the same polarity, the difference is 0; if the differences are of opposite polarity, the difference is 2. Multiply the differences by the adjustment factor. As the suppression strength of the perturbation jump region; a neighborhood perturbation difference term is introduced. This is used to estimate the probability that the current point is in a local peak or a rapidly changing region. The larger the difference in the neighborhood disturbance term, the more likely the current point is a local jump center, and its response amplitude needs to be appropriately reduced. The neighborhood disturbance term is multiplied by a control coefficient. The disturbance adjustment term is then subtracted from the current local disturbance folding feature. Finally, the mathematical model of the range regulation enhancement mechanism is as follows:

[0074] ;

[0075] in, As a regulatory factor, the initial value is set to 0.3 based on experience, which is used to control the inhibition intensity of polar mutation regions; As a control coefficient, the initial value is set to 0.5 based on experience, which is used to control the suppression amplitude of the edge response; It is a symbolic function; For the range adjustment enhancement mechanism in the time window Mid-moment Output range adjustment enhancement value.

[0076] S4. Calculate and normalize the difference ratio between the range adjustment enhancement value and the previous time step, and introduce the perturbation-guided adjustment coefficient to generate the latent guidance feature. Construct the diffusion residual through the temporal difference of the latent guidance feature, and generate the perturbation reconstruction feature by superimposing the asymmetric difference term of the range adjustment enhancement value. Normalize the perturbation reconstruction feature and combine it with the range adjustment enhancement value mapping to obtain the normalized feature.

[0077] Furthermore, in step S4, normalized features are established, and the process is as follows: Figure 3 As shown, the specific steps for establishing normalized features are as follows.

[0078] S41. Introduce the difference between the current range adjustment enhancement value and the previous time step range adjustment enhancement value, calculate the change ratio of the two and normalize the change ratio, and add a stability constant to ensure the numerical stability of the calculation process; on this basis, adjust the normalized change ratio through the perturbation guidance adjustment coefficient, and combine the adjustment result with the current range adjustment enhancement value to obtain the current latent guidance characteristics.

[0079] In this embodiment, in the long-term deformation monitoring of civil structures, deformation evolution often exhibits characteristics of both slow accumulation and local abrupt changes; to avoid directly using the range adjustment enhancement value. In cases where the response to sudden changes is insufficient or the steady-state region is over-amplified, this step constructs a perturbation-guided mapping to explicitly introduce the relationship between perturbation changes at adjacent time steps into the feature expression, giving the potential deformation trend a directional guiding ability in time series. First, the ratio of the change between the current range adjustment enhancement value and the previous time step is calculated to characterize the relative change trend of the perturbation. This ratio is then normalized to prevent it from being affected by the absolute amplitude scale. Subsequently, a perturbation-guided adjustment coefficient is introduced to linearly amplify and shrink the ratio, and multiplied by the range adjustment enhancement value, thereby introducing a directional guiding effect while maintaining the original perturbation sign and scale characteristics. When the perturbation change is significant, the latent guiding feature is enhanced; when the perturbation change is gradual, the latent guiding feature is close to the original input, achieving adaptive differentiation of different evolutionary states. The mathematical model of the perturbation-guided mapping is as follows:

[0080] ;

[0081] in, This is the disturbance guidance adjustment coefficient, used to control the intensity of the influence of adjacent disturbance differences on the latent missile mapping. The initial value is set to 0.3. This is a stability constant used to avoid zero denominators and suppress numerical instability under minimal perturbations; its initial value is set to... ; To guide the mapping of disturbances within a time window Mid-moment Output latent characteristics; For the range adjustment enhancement mechanism in the time window Mid-moment Output range adjustment enhancement value.

[0082] S42. Using the latent guidance features at the current time as input, the latent guidance features of the previous time step are introduced to construct the diffusion residual information between adjacent time steps. The diffusion residual information is smoothed and restricted using a nonlinear function. The asymmetric difference term formed between adjacent time steps is combined with the range adjustment enhancement value, and the asymmetric difference term is constrained. The diffusion residual term and the asymmetric suppression term are superimposed on the current latent guidance features in a weighted form, thereby constructing the perturbation reconstruction features at the current time step.

[0083] In this embodiment, the deformation of civil structures is often affected by the combined effects of load changes, environmental disturbances, and the evolution of the internal state of the structure. A single time step feature is insufficient to accurately reflect the true trend. This step, based on latent guidance features, introduces diffusion residual information from consecutive time steps to construct a disturbance reconstruction mechanism with temporal continuity, thereby characterizing the propagation and attenuation characteristics of deformation disturbances in the time dimension. The difference between latent guidance features of adjacent time steps is used to characterize the diffusion trend of the disturbance in time, and the hyperbolic tangent function is used to limit this trend within a stable range to avoid large fluctuations. Subsequently, an asymmetric suppression term based on range adjustment enhancement value is introduced to constrain the disturbance fallback stage. By combining the diffusion trend term and the suppression term and superimposing them on the latent guidance feature in a weighted form, the residual reconstruction of the disturbance evolution path is finally achieved, so that the output feature contains both historical trend information and maintains the ability to respond to abnormal changes. The mathematical model of the disturbance reconstruction mechanism is as follows:

[0084] ;

[0085] in, This is a diffusion adjustment factor used to balance the influence ratio between latent conduction characteristics and diffusion residual terms; its initial value is set to 0.2. It is a hyperbolic tangent function used to limit the amplitude of the diffusion difference and prevent abnormal amplification; This is the activation function used to suppress decreasing residuals that are contrary to the current perturbation trend; It is an asymmetric suppression term; For the perturbation reconstruction mechanism within the time window Mid-moment The output perturbation reconstruction features are used to comprehensively characterize the perturbation trend.

[0086] S43. Normalize the perturbation reconstruction features, and transform the range adjustment enhancement value into a dynamic weight term through a mapping function. Perform element-wise operations with the perturbation reconstruction features and perturbation enhancement weights, and fuse the element-wise operation results with the perturbation reconstruction features normalization operation to obtain the normalized features after perturbation reconstruction.

[0087] In this embodiment, in actual engineering monitoring data, the distribution of disturbance energy varies greatly across different sensing channels and time periods. Directly using this data for subsequent modeling can easily lead to bias in the civil structure deformation anomaly detection model. This step uses a dynamic redistribution mechanism of disturbance energy to perform unified scale calibration on the diffusion reconstruction features, while simultaneously using range adjustment enhancement values ​​to guide adaptive enhancement of the disturbance region. Specifically, the disturbance reconstruction features are processed for overall energy normalization to ensure that the feature distributions of different time steps and channels maintain a consistent scale. Subsequently, the range adjustment enhancement values ​​are smoothly mapped to construct dynamic weight terms, which are then combined with the reconstruction features element-wise. The disturbance enhancement weights control the overall influence amplitude, thereby achieving adaptive enhancement of high-disturbance regions. This process ensures that significantly disturbed regions receive high expression weights, while weakly disturbed regions remain in a normalized state, thus achieving a prominent expression of deformation disturbances while maintaining overall stability. The mathematical model of the dynamic redistribution mechanism of disturbance energy is as follows:

[0088] ;

[0089] in, This is a normalization stability factor used to prevent the denominator from becoming zero when energy is normalized; its initial value is set to... ; The perturbation enhancement weight is used to control the modulation intensity of the range adjustment enhancement value on the reconstructed features, and the initial value is set to 0.5. This is a dynamic weighting term used to map the range adjustment enhancement value to a smoothing weight; for Second normal form operation; This is an element-wise multiplication operation; For the dynamic redistribution mechanism of perturbation energy in the time window Mid-moment The normalized feature of the output after perturbation reconstruction.

[0090] S5. Using the normalized feature as input, a perturbation spinor modulation factor is constructed by combining the nonlinear suppression correlation coefficient and the amplitude is modulated. The perturbation spinor tensor is formed by applying orthogonal phase encoding through the logarithmic compression channel and the square amplification channel. A multipath structure coupling tensor is constructed by the perturbation spinor tensor and its transpose, combined with the normalized feature. The tensor potential mapping spectrum is formed under the nonlinear suppression coefficient and scale compression processing.

[0091] Furthermore, in step S5, a tensor potential mapping spectrum is established, the process of which is as follows: Figure 4 As shown, the specific steps for establishing the tensor potential mapping spectrum are as follows.

[0092] S51. Using nonlinear suppression correlation coefficient and normalized feature, construct a perturbation screw modulation factor that adapts to the strength of perturbation. Based on this, use the perturbation screw modulation factor to perform uniform amplitude modulation on the normalized feature. Send the modulated feature into the logarithmic compression channel and the square amplification channel respectively, and apply orthogonal phase encoding respectively. Finally, couple the results of the two channels through element-wise tensor outer product to form a perturbation screw tensor.

[0093] In this embodiment, addressing the issue that civil structures are simultaneously subjected to multiple disturbance factors during actual service, including temperature changes, load fluctuations, and support constraint adjustments, with significant coupling between various monitored quantities, the disturbance information, originally unfolded according to feature dimensions, is mapped into a disturbance spinor tensor with dual angle-amplitude characteristics. This is achieved through... Multi-channel nonlinear transformation and tensor outer product reconstruction are performed to couple the disturbance responses at different monitoring locations (beam ends, column bases, and mid-span) within a unified screw space, providing a basic data structure for subsequent construction of elastic anomaly features at the structural level. Specifically, the method involves first utilizing the normalized features after disturbance reconstruction. Construct two complementary channels: one channel passes through the pair Each dimension is incremented by 1, then the natural logarithm is taken, and multiplied by the perturbation spinor modulation factor. This forms a scale compression channel that is relatively sensitive to small disturbances; the other channel is... After squaring each element, multiply by the same factor. This process creates a nonlinear amplification channel that is more sensitive to large perturbations. Subsequently, the compressed channel is fed with a cosine function, and the amplified channel is fed with a sine function, resulting in two sets of phase-type response vectors. By performing a tensor outer product operation on the phase-type response vectors, the joint response between the two perturbation dimensions under different phase channels is simultaneously encoded at each element position of the matrix, thus forming a perturbation spinor tensor. The mathematical model for the perturbation spinor tensor is:

[0094] ;

[0095] in, The perturbation spinor modulation factor is adaptively adjusted based on the overall perturbation level of the current window. N is the current window size, with a value of 200. This is the nonlinear suppression coefficient in step S31, with an initial value of 4; For perturbation spinor tensors.

[0096] S52. The perturbation screw tensor is transposed as input. The perturbation screw tensor and its transpose are combined in an element-wise correspondence manner to form a path coupling term that contains bidirectional perturbation correlation information. On this basis, the normalized feature after perturbation reconstruction is introduced and nonlinearly suppressed and mapped to it to construct the perturbation response vector. A self-coupling tensor is generated by the vector outer product. The coupling factor and diffusion adjustment factor are adaptively generated according to the amplitude distribution of the normalized feature. The path coupling term and the self-coupling tensor are weighted and superimposed to obtain the multipath structure coupling tensor.

[0097] In this embodiment, the perturbation spinor tensor is used. As input, considering the characteristics of civil structures with multiple components, multiple monitoring points, and multiple load channels in spatial layout, a multipath structural coupling tensor is constructed. ; By element-wise coupling of the spinor tensor and its symmetric mapping, the normalized features reconstructed from the perturbation are superimposed. The generated nonlinear self-coupling tensor employs coupling and diffusion coefficients that adaptively vary with the feature distribution, ensuring that each position in the matrix contains both inter-channel interference path information and the perturbation and squeezing effect of a single channel. Specifically, the perturbation spinor tensor is first... Transpose to obtain , representing the spinor responses of the forward and reverse disturbance paths between the monitored quantities, respectively; then, for and Element-wise multiplication is performed, allowing the value at each position in the matrix to simultaneously incorporate path information from both directions, and an adaptive coupling factor is used. Adjusting the coupling strength between different monitoring channels to form path coupling terms, such as assigning higher weights to the path between mid-span deflection and support displacement, and relatively lower weights to channels with weaker interrelationships; simultaneously, utilizing the normalized features reconstructed from the perturbation. and nonlinear suppression coefficient Construct the disturbance response vector Self-coupled tensors are formed through outer product. Finally, with an adaptive diffusion regulation factor To balance this, the path coupling term is superimposed with the self-coupling tensor to obtain the multipath structure coupling tensor. The mathematical model for the multipath structure coupling tensor is as follows:

[0098] ;

[0099] in, This is the perturbation response vector after nonlinear suppression, used to construct the self-coupling tensor. This is the non-linear suppression weight, and its value is set to 0.25; This is an adaptive coupling factor, whose value combines... The normalized feature sizes of each dimension are adaptively scaled, and its mathematical model is expressed as follows: D is Feature dimension, For the dynamic redistribution mechanism of perturbation energy in the time window Mid-moment No. The normalized features after perturbation reconstruction of each feature The normalization stability factor takes a value of , For the first The perturbation importance weights of each feature; This is a diffusion regulation factor, whose value is related to the current window. The values ​​are inversely proportional to prevent the diffusion term from being too strong or too weak; It is a multipath structure coupling tensor.

[0100] S53. Perform dimension-by-dimensional mapping calculations on the multipath structure coupling tensor and normalized features at the tensor-vector level, introduce nonlinear suppression coefficients, adjust the amplitude of the mapping results, and perform uniform scale compression on the results through continuous nonlinear mapping to finally form a tensor potential mapping spectrum.

[0101] In this embodiment, the tensor is coupled using a multipath structure. and normalization characteristics As input, construct a tensor potential mapping spectrum. Tensor potential mapping is used to compress complex multipath structural coupling relationships into elastic anomaly detection feature vectors that can be directly used in anomaly detection, so that each feature component corresponds to a perturbation potential level of the civil structure at the current moment; the specific method is: to couple the multipath structure tensor With normalization characteristics A matrix-vector product is performed so that each dimension of the output simultaneously incorporates the coupling relationships between the current feature and all other monitored features, as well as the self-feedback pattern. Then, the nonlinear suppression coefficient from S31 is introduced. The amplification intensity is adjusted by combining the coupling relationship, and then the amplified result is fed into the hyperbolic tangent function for boundary compression, so that the output will not become numerically uncontrollable due to local extreme perturbations, while maintaining sufficient resolution for potential differences; the mathematical model of the tensor potential mapping spectrum is:

[0102] ;

[0103] in, For tensor potential mapping spectrum; This is the nonlinear suppression coefficient, with an initial value of 4; It is a non-linear activation function.

[0104] S6. Using the tensor potential mapping spectrum as input, construct the channel-level theoretical symmetry center response in the symmetric time neighborhood, and obtain the perturbation energy offset by comparing each channel. Based on the directional distinction relationship of the perturbation energy offset in the previous and next time steps, construct the extreme value deflection corresponding to the current time point. Perform normalization mapping on the extreme value deflection in the local time neighborhood to form the probability potential index corresponding to the current time point.

[0105] Furthermore, in step S6, a probabilistic potential index is established, the process of which is as follows: Figure 5 As shown, the specific steps for establishing the probability potential index are as follows.

[0106] S61. Using the tensor potential mapping spectrum as input, and taking the current time point as the center, select several adjacent time steps to form a symmetric time neighborhood. Calculate the maximum and minimum tensor potential mapping spectrum values ​​within the channel in the neighborhood. Merge the maximum and minimum values ​​to obtain the theoretical symmetric center response. Compare the tensor potential mapping spectrum value at the current time with the theoretical symmetric center response channel by channel to obtain the deviation degree of each channel. Then, aggregate the deviation results of all channels to generate the perturbation energy offset corresponding to the current time point.

[0107] In this embodiment, a tensor potential mapping spectrum is used. As input, the aim is to identify perturbation-sensitive regions in the tensor time series response. By constructing the offset between the current response value and the center of symmetry extrema within a local time window, the symmetry disruption phenomenon caused by the perturbation is characterized, thereby extracting the energy shift response of the tensor potential under the perturbation, providing a basic driving force for subsequent trend deflection modeling; the specific method is: using each time point Centered on, extract the parts before and after it. A symmetrical neighborhood interval consisting of time steps For each channel Tensor potential mapping spectrum values ​​on Extract the maximum and minimum values, and take their average as the current channel. At the present moment The theoretical symmetric central response, at the current time Tensor potential mapping spectrum value The absolute difference between the response at the theoretical center of symmetry and the response at the theoretical center of symmetry represents the degree of shift after the perturbation breaks the symmetry. Summing these values ​​over all channels generates the overall perturbation energy shift. The larger the disturbance energy offset, the closer the current moment is to the region of abnormal disturbance response; the mathematical model for the disturbance energy offset is:

[0108] ;

[0109] in, This represents the disturbance energy offset. The initial value for the local symmetry analysis window length is 5. The characteristic dimension of the tensor potential mapping spectrum is 64; For the first The first window The channel in the symmetrical interval to which each time step belongs The maximum value; For the first The first window The channel in the symmetrical interval to which each time step belongs The minimum value.

[0110] S62. Select the perturbation energy offset corresponding to several time steps before and after, and distinguish the forward and backward time positions to form a local trend center value with time orientation. Calculate the perturbation energy offset and the local trend center value after direction distinction to construct the extreme value deflection corresponding to the current time point.

[0111] In this embodiment, the disturbance energy offset is obtained. Furthermore, by tracking its evolution trend over time, an extreme deflection value reflecting the amplification and reversal trends of the disturbance is constructed. By introducing a sign function, the difference in the evolution trends of forward and backward disturbances within the time window is incorporated into the modeling, thereby improving the sensitivity of anomaly trend detection. The specific method is as follows: taking the current moment... Extract from the center, both forward and backward. Perturbation energy offset at each time step To enhance the perception of trend direction, a sign function is introduced. Past time steps Assign a value of -1, future time step Increment the value by 1 to the current time step. Assigning a value of 0, and then averaging the products of the disturbance offset and the corresponding sign, yields a direction-weighted local trend center value. This value is then used to calculate the current time value. The difference between the value and the local trend center value, and the absolute value, is used to construct the extreme value deflection. The larger the extreme value deflection, the stronger the deviation from the overall trend at the current moment, which may be an anomalous abrupt change point; the mathematical model for the extreme value deflection is:

[0112] ;

[0113] in, Set the size of the forward and backward offset trend analysis window to 3; This is a sign function used to distinguish between forward (+1), backward (−1), and current (0); For the first The perturbation energy offset at any given moment; This represents the extreme value deflection.

[0114] S63. Extract all extreme value deflections within the local time neighborhood and sum them to obtain a normalized reference value; perform intensity mapping on the extreme value deflection corresponding to the current time point, and standardize the normalized reference value and the intensity-mapped extreme value deflection to obtain the probability potential index of the current time point.

[0115] In this embodiment, based on the extreme value deflection amount A normalization mechanism is introduced to construct a probability potential index. This is used to measure the relative probability level of disturbance anomalies occurring at each time point. By processing the squared extreme value deflection and then normalizing the mapping, the ability to distinguish high-risk points is effectively enhanced, and the output is made probabilistically meaningful, which is convenient for use in the fusion of subsequent anomaly detection models. The specific method is as follows: taking the current time... Extract its neighborhood intervals centered on the target. All extreme deflection quantities within The values ​​are squared and summed to obtain a normalized reference value, which is the extreme value deflection at the current moment. Square the value and take the local perturbation energy intensity; then compare the two to obtain the value. Its value is in the range (0,1), representing the probability of an anomaly occurring in the relative neighborhood at the current time point. The closer the probability potential index value is to 1, the higher the degree of anomaly. The construction method introducing a normalization mechanism takes into account both the sensitivity of time series trends and the comparability of normalization, and has high engineering applicability; the mathematical model of the probability potential index is:

[0116] ;

[0117] in, The width of the normalized probability mapping window is initially set to 4. To prevent division by zero, the value is set to a very small positive number. ; For the first The first window Extreme deflection at each time step; For the first Time steps within a time window The probability potential index.

[0118] S7. Construct a civil structure deformation anomaly detection model. Input the civil structure deformation time series training set, and use the joint deformation state error loss function to pass through steps S2 to S6 in sequence. Train the civil structure deformation anomaly detection model until it converges to realize the anomaly detection of civil structure deformation.

[0119] Furthermore, in step S7, the specific steps for training the civil structure deformation anomaly detection model are as follows.

[0120] S71. The probability potential index is constructed in chronological order into a probability potential index matrix and input into the multilayer perceptron structure to output the detection results of the future civil structure deformation anomaly detection model. By weighting and averaging the error between the detected value and the true value of the civil structure deformation anomaly detection model with the corresponding probability potential index matrix, a deformation state error loss function is constructed.

[0121] In this embodiment, the probability potential index constructed in step S6 is used as input, and the probability potential indices at all times are sequentially concatenated to form a probability potential index matrix. The probability potential index matrix is ​​input into a multilayer perceptron structure, which outputs deformation state data corresponding to each time step within the next 30 seconds. The multilayer perceptron structure consists of two nonlinear mapping layers and one linear regression layer, and has the ability to model across time steps and perform nonlinear fitting. The output result is a set of deformation state detection values ​​for consecutive time steps. , used to characterize abnormal evolutionary trends in future stages;

[0122] A deformation state error loss function is constructed. This function explicitly enhances the contribution weight of errors at key time points, guiding the model to focus on locations with high anomaly probability, thereby optimizing the response to abrupt changes. Specifically, the method involves: first, calculating the error loss function for each future time step... The absolute error between the detected value and the true value, i.e. The absolute error and the probability potential index matrix are combined. Multiply to obtain the weighted residuals. For all time steps arrive The weighted residuals are summed, and the sum is divided by the total number of future detection time steps. The final deformation state error loss value is obtained, and the mathematical model of the deformation state error loss function is as follows:

[0123] ;

[0124] in, To indicate the prediction time length, it is set to 30; The output of the civil structure deformation anomaly detection model in the k-th window is the first... The detected value at any given time; For the future in the k-th window The true value of a moment; This represents the loss value of the deformation state error loss function.

[0125] S72. Input the time series training set of civil structure deformation, set the hyperparameters and train the civil structure deformation anomaly detection model in combination with the deformation state error loss function to realize the anomaly detection of the civil structure deformation anomaly detection model.

[0126] In this embodiment, a civil structure deformation anomaly detection model is constructed. This model, driven by a tensor potential mapping spectrum, takes a civil structure deformation anomaly detection set as input and sequentially performs multi-source disturbance coupling analysis and spatiotemporal disturbance tensor embedding to generate a tensor potential mapping spectrum. A symmetric time neighborhood is constructed centered on the current time point. The maximum and minimum values ​​of the tensor potential mapping spectrum within the channel are extracted, fused to generate a theoretical symmetric center response, and the deviation from the tensor potential mapping spectrum at the current time is calculated to obtain the disturbance energy offset. The model constructs a local trend center value by combining time directionality, thereby generating extreme value deflection. It then aggregates the extreme value deflection within the local neighborhood, completes normalization and intensity mapping, and generates a probability potential index for the current time point. The civil structure deformation anomaly detection model further integrates the full-time-series probability potential index to construct a probability potential index matrix, which is input into a multilayer perceptron network to output deformation state data for the next 30 seconds. A deformation state error loss function is also constructed to improve the identification accuracy and prediction stability of deformation anomalies, achieving high-precision anomaly detection and dynamic response modeling during the deformation evolution process of civil structures.

[0127] Furthermore, the civil structure deformation anomaly detection model proposed in this invention is implemented using the Python programming language and built and trained based on the PyTorch deep learning framework. During the training phase, the Adam optimizer is used for parameter updates, with an initial learning rate set to 0.001. An exponential decay strategy is introduced to dynamically adjust the learning rate, with a learning rate decay coefficient set to 0.95 to improve the stability of parameter updates and the convergence speed of the training process. In terms of hyperparameters, the batch size is set to 64, the total number of training epochs is 1000, and a gradient clipping threshold of 2.0 is set to prevent gradient explosion during training.

[0128] Furthermore, the collected time-series training set of civil structure deformation is input into the constructed civil structure deformation anomaly detection model for training. The changing trend of the deformation state error loss function used during the training process is as follows: Figure 6As shown in the figure, in the early stage of training, the civil structure deformation anomaly detection model is in the stage of responding to and fitting the tensor potential mapping spectrum, perturbation energy offset, and extreme value deflection. The loss function value is relatively high and fluctuates to a certain extent. As the number of training rounds increases, the civil structure deformation anomaly detection model gradually establishes the mapping relationship between the perturbation structure and the deformation state. The loss function value shows a continuous downward trend and tends to stabilize after about 850 training rounds, eventually converging to about 0.12. The above results show that the civil structure deformation anomaly detection model proposed in this invention has good convergence performance and numerical stability in the optimization process of the deformation state error loss function. It can accurately model multi-channel probability potential indicators under the action of high-dimensional perturbation features and complete the stable prediction of future deformation states.

[0129] Visual analysis of the tensor potential mapping spectrum was performed, and the results are as follows: Figure 7 As shown in the normal state tensor potential mapping spectrum, the tensor potential response of each channel is relatively uniformly distributed in the time dimension, with no obvious concentrated enhancement region. The overall energy distribution remains stable and consistent, reflecting that the structure is in a normal evolutionary state. In contrast, the tensor potential mapping spectrum of the abnormal state reveals that within a local time interval, the tensor potential response of multiple channels is significantly enhanced simultaneously, forming a continuously distributed high-response region. This indicates that the structural perturbation exhibits a cooperative activation characteristic in both the time and channel dimensions. These results demonstrate that the tensor potential mapping spectrum can effectively characterize the joint evolution of multi-channel perturbations in the temporal space, providing a stable and structurally distinguishable basic representation for the subsequent construction of perturbation energy shift, extreme value deflection, and probability potential indices.

[0130] To verify the effectiveness of the civil structure deformation anomaly detection model in the deformation anomaly identification stage, the anomaly detection results of the deformation state test set were compared and analyzed. The results are as follows: Figure 8 As shown in the figure, the horizontal axis represents the time evolution of a test sample within the next 30 seconds, and the vertical axis represents the deformation anomaly detection value at the corresponding time point. As can be seen from the figure, the deformation anomaly detection value output by the civil structure deformation anomaly detection model is highly consistent with the actual deformation anomaly value. The correspondence between the two on the time axis is clear and stable, without any obvious deviation. These results demonstrate that the civil structure deformation anomaly detection model can stably characterize the evolutionary features of deformation anomalies in the time dimension, accurately track the amplitude and trend of deformation anomalies in future periods, and verify that the civil structure deformation anomaly detection model has good discriminative consistency and temporal modeling reliability when identifying deformation anomalies under complex disturbance conditions.

[0131] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for detecting deformation anomaly of a civil structure based on time series data, characterized by, The method comprises the following steps: Collecting civil structure state data and preprocessing and slicing processing; Calculate the disturbance intensity response value between the current window time step feature vector and the previous time step feature vector, and introduce a minimum constant stable denominator while normalizing the disturbance intensity response value of the current and previous time, to construct the disturbance activation factor, and on this basis, combine the nonlinear suppression coefficient, and complete the calculation of the disturbance intensity response value through the exponential function; Based on the disturbance intensity response value of the current time point and the two adjacent time points before and after it, the disturbance curvature of the current time point is calculated, and based on the disturbance curvature, the folding adjustment factor containing the disturbance intensity response value and the folding adjustment coefficient is introduced, and the disturbance curvature and the folding adjustment factor are combined to output the local disturbance folding feature corresponding to the current time window; Training a civil structure deformation anomaly detection model to detect civil structure deformation anomalies. Through joint scale model test, long-term monitoring, numerical simulation and indoor test to obtain data, monitoring points are arranged at the structure detection position and multiple types of sensors are installed, deformation and environmental parameters are collected at a frequency of 1 per second, and related attribute information is recorded synchronously, forming multi-source deformation time sequence raw data; The multi-source deformation time sequence raw data is time-aligned, invalid data is removed, anomalies are cleaned, missing data is interpolated, dimensions are unified, and numerical values are normalized, and finally a civil structure deformation data set is constructed. Under the condition that the sampling frequency is 1 per second, the civil structure deformation data set is processed by sliding slicing according to time sequence, a time window is constructed as model input, each time window retains the time sequence change characteristics of the high-dimensional vector of all physical and environmental characteristics, the structure deformation probability corresponding to the continuous sampling points after the end of each input window is taken as the prediction target, and the time window and the prediction target window are combined to construct a civil structure deformation time sequence training set and a civil structure deformation time sequence test set.

2. The method according to claim 1, wherein, Calculate the change amplitude between the current window time step feature vector and the previous time step feature vector to obtain the disturbance intensity measure value, and introduce a minimum constant stable denominator while normalizing the disturbance intensity measure value of the current and previous time, to construct the disturbance activation factor, and on this basis, combine the nonlinear suppression coefficient, and complete the calculation of the disturbance intensity response value through the exponential function; Based on the disturbance intensity response value of the current time point and the two adjacent time points before and after it, the disturbance curvature of the current time point is calculated, and based on the disturbance curvature, the folding adjustment factor containing the disturbance intensity response value and the folding adjustment coefficient is introduced, and the disturbance curvature and the folding adjustment factor are combined to output the local disturbance folding feature corresponding to the current time window; 3. The method according to claim 1, wherein, ​ 4. The method according to claim 1, wherein, ​ ​ By comparing the current local disturbance folding feature with the local disturbance folding feature symbol of the previous time step, and combining the adjustment factor to suppress the response strength of the mutation region, a symbol jump mark is formed; The neighborhood disturbance difference value of the local disturbance folding feature before and after the current time is introduced, and the neighborhood disturbance difference value is scaled by combining the control coefficient to form a disturbance adjustment term, and the symbol jump mark and the disturbance adjustment term are multiplied to construct the range adjustment enhancement value.

5. The method according to claim 1, wherein, The difference relationship between the range adjustment enhancement value at the current time and the range adjustment enhancement value at the previous time step is introduced, the change ratio of the two is calculated and the ratio is normalized, and a stability constant is added to ensure the numerical stability of the calculation process; On this basis, the normalized change ratio is adjusted by the disturbance guide adjustment coefficient, and the adjustment result is combined with the current range adjustment enhancement value to obtain the latent guide feature at the current time; The latent guide feature at the current time is taken as the input, and the latent guide feature at the previous time step is introduced to construct the diffusion residual information between adjacent time steps, a nonlinear function is used to smooth and limit the diffusion residual information, an asymmetric difference term formed between adjacent time steps is combined with the range adjustment enhancement value, and the asymmetric difference term is constrained, and the diffusion residual term and the asymmetric suppression term are added to the current latent guide feature in a weighted form, thereby constructing the disturbance reconstruction feature at the current time; The disturbance reconstruction feature is normalized, and the range adjustment enhancement value is converted into a dynamic weight term through a mapping function, and the disturbance reconstruction feature and the disturbance enhancement weight are element-wise operated, and the element-wise operation result is fused with the disturbance reconstruction feature to obtain the normalized feature after disturbance reconstruction.

6. The method according to claim 1, wherein, A disturbance spinor modulation factor that changes adaptively with the disturbance strength is constructed using a nonlinear suppression correlation coefficient and a normalized feature. On this basis, the normalized feature is uniformly amplitude modulated using the disturbance spinor modulation factor. The modulated feature is sent into a logarithmic compression channel and a square amplification channel, respectively, and orthogonal phase encoding is applied. Finally, the two channel results are coupled through element-wise tensor outer product to form a disturbance spinor tensor. The disturbance spinor tensor is transposed, and the disturbance spinor tensor and its transpose are combined in an element-wise manner to form a path coupling term that contains bidirectional disturbance correlation information. On this basis, the disturbance reconstruction normalized feature is introduced, which is nonlinearly suppressed and mapped to construct a disturbance response vector. A self-coupling tensor is generated through vector outer product. According to the amplitude distribution of the normalized feature, a coupling factor and a diffusion adjustment factor are adaptively generated. The path coupling term and the self-coupling tensor are weighted and added to obtain a multi-path structure coupling tensor. The multi-path structure coupling tensor and the normalized feature are mapped and calculated in the tensor-vector layer, a nonlinear suppression coefficient is introduced to adjust the amplitude of the mapping result, and the result is uniformly scaled through continuous nonlinear mapping, finally forming a tensor potential mapping atlas.

7. The method according to claim 1, wherein, Take the tensor potential mapping atlas as input, take the current time point as the center, select several time steps before and after it to form a symmetric time neighborhood, calculate the maximum and minimum tensor potential mapping atlas values in the neighborhood, fuse the maximum and minimum values to obtain the theoretical symmetric center response, compare the current time step tensor potential mapping atlas value with the theoretical symmetric center response channel by channel to obtain the deviation degree of each channel, and aggregate the deviation results of all channels to generate the disturbance energy offset corresponding to the current time point; Select the corresponding disturbance energy offset in several time steps before and after it, and distinguish the direction of the forward and backward time positions to form a local trend center value with time direction, and operate the disturbance energy offset and the local trend center value after direction differentiation to construct the extreme value deflection corresponding to the current time point; Extract all extreme value deflections in the local time neighborhood and accumulate to obtain a normalized reference value; map the intensity of the extreme value deflection corresponding to the current time point, normalize the normalized reference value and the extreme value deflection to obtain the probability potential index of the current time point.

8. The method according to claim 1, wherein, The probability potential index constructs a probability potential index matrix in time sequence and inputs a multi-layer perception structure, and outputs a future civil structure deformation anomaly detection model detection result; by weighting and averaging the error between the civil structure deformation anomaly detection model detection value and the true value and the corresponding probability potential index matrix, a deformation state error loss function is constructed; Input the civil structure deformation time sequence training set, set the hyperparameters and train the civil structure deformation anomaly detection model combining the deformation state error loss function to realize the anomaly detection of the civil structure deformation anomaly detection model.

Images