An abnormality identification and repair method and system for geological disaster monitoring data

Abstract

The application discloses an abnormality identification and repair method and system for geological disaster monitoring data, and belongs to the technical field of geological disaster monitoring and early warning, time series data processing and machine learning application. The method compares the original historical segment and the prediction error after the disturbance along the prototype direction to obtain the noise suspiciousness score and the prediction improvement degree by constructing the historical shape prototype of each monitoring variable. Further, the clean time and the noise time are divided by the adaptive threshold, and the multi-variable results are fused to form the trusted time set and the to-be-repaired time set. Then, the self-repair network is trained on the trusted time to generate the conditional vector constraint segment by the conditional historical segment coding, and the repair anchor point is selected according to the improvement degree to perform the write-back repair on the abnormal interval. The application can improve the abnormality identification accuracy, the repair segment reliability and the early warning analysis support capability without changing the basic evolution law of the geological disaster monitoring data.

Description

Technical Field

[0001] This invention relates to the fields of geological disaster monitoring and early warning, data processing, and machine learning applications. Specifically, it relates to a method and system for anomaly identification and repair of geological disaster monitoring data, addressing scenarios such as landslides, collapses, debris flows, slope instability, ground subsidence, and reservoir bank deformation. More specifically, this invention is applicable to anomaly identification, reliable sample screening, fragment repair, and early warning analysis support for multi-source monitoring data deployed on slopes, landslide bodies, unstable rock masses, gullies, and subsidence-sensitive areas. Background Technology

[0002] Geological hazard monitoring typically relies on rain gauges, GNSS displacement monitoring equipment, crack gauges, inclinometers, deep displacement gauges, soil moisture sensors, pore water pressure gauges, groundwater level gauges, and acceleration or microseismic monitoring equipment to continuously observe the evolution of the hazard body at different time scales. These multi-source observations constitute multivariate time-series data for geological hazard monitoring, serving as a crucial data foundation for slope stability analysis, landslide deformation trend assessment, hazard threshold early warning, risk level evaluation, and retrospective analysis of remediation effectiveness.

[0003] However, geological disaster monitoring is characterized by long-term field deployment, strong environmental disturbances, complex maintenance conditions, and non-stationary disaster processes. Monitoring terminals may be affected by factors such as rainfall erosion, siltation, lightning surges, unstable power supply, communication link jitter, foundation loosening, and equipment aging, leading to local spikes, continuous shifts, random noise, missing observations, fragment distortion, timestamp misalignment, and manual recording errors in the monitoring sequence. If such raw data is directly used as the basis for early warning analysis, trend modeling, or risk assessment, abnormal observations may be misjudged as real geological responses, or real changes in hazards may be drowned out by noise, thus affecting the reliability of monitoring conclusions.

[0004] Existing technologies can generally be divided into three categories. The first category is statistical thresholding or rule-based methods, such as identifying outliers based on mean, variance, sliding window thresholds, quantile upper and lower bounds, or empirical rules, and performing repair through mean backfilling, median backfilling, forward filling, or linear interpolation. These methods are simple to implement and have low deployment costs, but they are less adaptable to nonlinear accelerated deformation of landslide bodies after rainfall triggering, hysteretic responses caused by reservoir water level fluctuations, and coupling relationships between multiple source variables. They are prone to misclassifying normal changes as anomalies or missing anomalous segments. The second category is predictive methods, which use time-series prediction models to predict future segments and then identify anomalies based on the prediction residuals. These methods are better able to utilize time-dependent structures than rule-based methods, but when the historical input itself is contaminated, large prediction residuals cannot distinguish whether the future actual response is abnormal, the historical context is abnormal, or both. The third category is reconstruction or generative methods, such as autoencoders, sequence-to-sequence networks, diffusion models, and generative adversarial networks, which repair anomalous segments through reconstruction errors or generation results. This type of method can handle complex correlations at the distribution level, but it usually requires relatively pure training samples. If the training data has been mixed with a lot of outliers or misaligned segments, the model will often learn the contamination patterns as well, leading to incorrect reinforcement.

[0005] Related research also indicates that pre-trained time series base models and generative inpainting models have become important development directions in this field. Ansari et al. proposed "Chronos: Learning the Language of TimeSeries" in 2024, demonstrating that large-scale pre-trained time series models have strong zero-shot prediction capabilities and can transfer and utilize prior patterns across different monitoring variables and sites. Chen et al. proposed "ImDiffusion: Imputed Diffusion Models for Multivariate Time Series Anomaly Detection" in 2023, showing that generative models can reconstruct smoother, more reasonable, and structurally consistent results at the local segment level through distribution fitting. Yang et al.'s 2024 review, "A Survey on Diffusion Models for Time Series and Spatio-Temporal Data," further summarized the application trends of diffusion-type models in time series generation, imputation, and anomaly handling.

[0006] However, existing solutions still have the following common problems in geological disaster monitoring applications: First, there is insufficient coupling between anomaly identification indicators and early warning and prediction tasks, making it difficult to directly characterize whether a certain period of historical monitoring data has damaged the predictability of future changes; Second, in multivariate scenarios, there is a lack of a unified reliable moment screening mechanism, making it difficult to ensure that the data used for training the generative model simultaneously meets the multi-channel reliability conditions; Third, although the repair results may be numerically smooth, if there is a lack of typical historical patterns and predictive effect linkage constraints, the repaired segments may still deviate from the actual disaster evolution law; Fourth, when a certain variable has been faulty for a long time or is too disturbed, continuing to force repair may introduce greater risks.

[0007] To better illustrate the above issues, typical predictive anomaly detection can be abstracted as follows: given a historical segment of length (L) Future predictions are obtained from the predictor. And through residuals

[0008]

[0009] Assess the degree of anomaly, where P represents the prediction length. Represents the true future value. The key problem with this type of method is that when... When local distortions caused by communication anomalies, sensor drift, or manual recording errors are already present, an increase in residuals does not directly indicate anomalies in future observations themselves. It may simply mean that historical inputs have disrupted the predictive model's understanding of future monitoring processes. Similarly, typical interpolation-based repair can be expressed as...

[0010]

[0011] in, This represents the interpolation coefficients. While this method can maintain numerical continuity, it cannot guarantee that the repaired fragments, once re-introduced into early warning models, stability analysis models, or trend judgment models, will retain a predictable structure consistent with the actual geological disaster evolution process. Especially in complex dynamic scenarios such as rainfall-displacement linkage, pore pressure-deep displacement hysteresis coupling, and crack propagation-dip angle change, simple interpolation often only guarantees "numerical continuity" but not "mechanistic consistency."

[0012] Therefore, there is an urgent need for a multivariate time series data repair technology that can directly correlate anomaly identification with future prediction consistency, form a reliable time set in multivariate geological disaster monitoring scenarios, perform distribution-level segment reconstruction under typical historical pattern constraints, and provide failure protection for high-noise variables. Summary of the Invention

[0013] The purpose of this invention is to overcome the problems in the existing technology, such as the disconnect between anomaly identification and early warning prediction tasks, the difficulty in unifying the reliable moments of multiple variables, the susceptibility of generative repair to the influence of contaminated training samples, and the high risk of blindly repairing high-noise variables. It provides an anomaly identification and repair method and system for geological disaster monitoring data, so as to achieve noise measurement that is closer to the early warning task, more robust monitoring time division, and more controllable reconstruction of anomaly segments.

[0014] To achieve the above objectives, the present invention adopts the following technical solution:

[0015] In a first aspect, embodiments of the present invention provide a method for anomaly identification and repair of geological disaster monitoring data, comprising the following steps:

[0016] S1. Obtain multivariate time series data for geological disaster monitoring;

[0017] S2. For each monitored variable, extract a historical observation segment of the first historical length from the preset sampling index set, and perform pattern induction on the historical observation segment to obtain at least one historical form prototype corresponding to the monitored variable.

[0018] S3. For each sampling time in the sampling index set, extract the historical observation segment and the future true segment of the second prediction length corresponding to the monitoring variable; call the pre-trained time series prediction model, use the historical observation segment as input to obtain the future prediction segment, and calculate the baseline prediction error; then, based on each historical morphological prototype, perturb the prototype direction of the historical observation segment and re-predict, and calculate the perturbation prediction error; based on the difference between the baseline prediction error and each perturbation prediction error, obtain the prediction improvement degree of the sampling time relative to each historical morphological prototype, and obtain the noise suspicion score based on the baseline prediction error;

[0019] S4. Based on the noise suspicion score of each monitored variable, determine the noise segmentation threshold of the sub-variables and divide the time position of each monitored variable into net time and noisy time; merge the net time and noisy time of each monitored variable to obtain a set of credible time for repair training and a set of time to be repaired.

[0020] S5. For variables that are not marked as skipped repair, construct training samples on the set of reliable times and train a conditional generative self-repairing network.

[0021] S6. For the variable to be repaired, extract the conditional historical fragments before the time to be repaired, obtain the conditional vector through the conditional encoder, and select the repair anchor point according to the predicted improvement degree of the time to be repaired relative to each historical prototype. Call the self-repairing network to generate the target fragment and write it back to the corresponding time interval to obtain the repaired geological disaster monitoring multivariate time series data.

[0022] In one embodiment, in step S1, the multivariate time series data includes at least multi-channel observations at different time steps from the same monitoring point, the same slope unit, the same landslide body, or the same disaster hazard area.

[0023] In one embodiment, the multi-channel observations include at least one of the following: rainfall, surface displacement, deep displacement, crack width, dip angle, surface subsidence, soil moisture content, pore water pressure, groundwater level, GNSS displacement, acceleration, vibration response, or microseismic response data; the multivariate time series data originates from multi-sensor joint observations at a single monitoring point or from joint monitoring records of multiple neighboring monitoring points after spatial alignment.

[0024] In one embodiment, in step S2, the historical morphology prototype is obtained by performing clustering operations on historical observation segments of each monitoring variable. The clustering operations employ K-means clustering, Gaussian mixture clustering, or spectral clustering. Each cluster center represents a historical change pattern that recurs during geological disaster monitoring and is used for subsequent disturbance assessment and remediation anchoring.

[0025] In one embodiment, in step S3, the prototype orientation perturbation is achieved by linearly combining the difference vectors of the current historical observation segment and the historical morphological prototype according to a preset perturbation coefficient, and a zero-mean random perturbation term can be superimposed; the random perturbation term is used to improve the robustness of the evaluation process to local small morphological differences.

[0026] In one embodiment, in step S4, the noise segmentation threshold for the discrete variables is determined using the elbow rule; specifically, it includes:

[0027] The noise suspicion scores of the monitored variables are sorted in ascending order and then normalized.

[0028] Calculate the degree of deviation of each normalized point on the sorting curve from the baseline line, and use it as the vertical distance;

[0029] The original score corresponding to the point that maximizes the vertical distance is determined as the noise segmentation threshold for the monitored variable.

[0030] In one embodiment, step S4, fusing the net time and noisy time of each monitored variable, includes:

[0031] The intersection of the net time sets of each monitored variable is used to obtain the set of reliable times;

[0032] The set of noisy time points of each monitored variable is combined to obtain the set of time points to be repaired.

[0033] In one embodiment, step S4 further includes:

[0034] When the proportion of noisy moments of a certain monitoring variable exceeds the preset limit, the monitoring variable is marked as a skip-repair variable to avoid continuing training on variables that are heavily contaminated overall, which would lead to error reinforcement.

[0035] In one embodiment, step S5, training a conditionally generative self-healing network, includes:

[0036] Based on the set of reliable moments, a training sample pair consisting of conditional history fragments and target fragments is constructed.

[0037] The conditional history segment is input into the conditional encoder to obtain the conditional vector;

[0038] Calculate the similarity between the conditional vector and the embedding vector corresponding to each historical morphological prototype, and select the historical morphological prototype with the highest similarity as the training anchor point.

[0039] Starting with noise superimposed on the training anchor points and ending with the target segment, an interpolated state is constructed on the time parameters.

[0040] A velocity network is trained, with the interpolation state, time parameters, and conditional vector as input, and the learning objective is to predict the update direction from the current state to the target segment.

[0041] Secondly, embodiments of the present invention also provide an anomaly identification and repair system for geological disaster monitoring data, employing an anomaly identification and repair method for geological disaster monitoring data as described in any of the first aspects, the system comprising:

[0042] The data acquisition unit is used to acquire and organize multivariate time series data for geological disaster monitoring;

[0043] An anomaly intelligent identification and evaluation unit is used to construct historical pattern prototypes, call pre-trained time series prediction models, and output noise suspicion scores and prediction improvement degrees.

[0044] The sample partitioning unit is used to determine the net time and the noisy time based on the noise suspicion score, and to form a set of credible times and a set of times to be repaired;

[0045] The self-repair unit is used to train the self-repair network on a set of reliable time points and to perform target segment generation and write-back on the interval corresponding to the time point to be repaired.

[0046] The data parsing and output unit is used to write the restored monitoring data back to the geological disaster monitoring database, slope monitoring platform, or early warning system.

[0047] As can be seen from the above technical solution, compared with the prior art, the present invention has the following technical effects:

[0048] First, the intelligent anomaly identification metric is more aligned with geological disaster early warning tasks. This invention establishes a direct link between "anomaly" and "prediction difficulty" by comparing the baseline prediction error given by the pre-trained prediction model with the error after perturbation of the prototype direction, reducing misjudgments caused by relying solely on amplitude thresholds or local residuals.

[0049] Second, the historical morphology prototype library provides an interpretable set of typical monitoring patterns. As historical morphology prototypes, the cluster centers provide a unified semantics for subsequent disturbance verification, anchor point selection, and repair generation, enabling the repair process to have traceability "starting from typical geological disaster monitoring patterns".

[0050] Third, the multivariate fusion strategy enhances the credibility of training samples. By obtaining a global credible time set through the intersection of the net times of each monitoring variable, the training samples are more likely to simultaneously satisfy the consistency of multiple channels such as rainfall, displacement, pore pressure, and cracks, thereby reducing the interference of univariate anomalies on joint modeling.

[0051] Fourth, the elbow rule enables threshold adaptation. Compared to manually setting fixed thresholds or fixed quantiles, this invention utilizes the geometric inflection point characteristics of the score ranking curve to determine the segmentation threshold, making it more adaptable to score distributions under different monitoring points, different variables, and different seasonal conditions.

[0052] Fifth, self-repairing reconstructs monitoring segments at the distribution level. Compared to simple interpolation or point-level backfilling, this invention learns the mapping from anchor point distribution to target distribution under conditional vector constraints, which is more conducive to maintaining the rationality of local dynamic structure, numerical range, and cross-variable correlation.

[0053] Sixth, skipping the repair strategy provides failure protection. When the proportion of noisy moments in a certain monitoring variable is too high, continuing training and repair may introduce error reinforcement. This invention reduces systemic risk and retains the space for manual verification by skipping the repair process for that variable.

[0054] Seventh, a differentiated prototype selection strategy is adopted for training and repair. During the training phase, anchor points are selected based on the similarity between the conditional vector and the prototype embedding, while during the repair phase, anchor points are selected based on the predicted improvement. This helps to form a closed loop among encoding representation, intelligent anomaly identification, and fragment repair. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0056] Figure 1This is a flowchart of the method for anomaly identification and repair of geological disaster monitoring data provided in this embodiment of the invention;

[0057] Figure 2 This is a structural diagram of the anomaly identification and repair system for geological disaster monitoring data provided in this embodiment of the invention;

[0058] Figure 3 This is a schematic diagram of the electronic device structure provided in an embodiment of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Example 1:

[0061] This invention addresses the problem of data contamination in scenarios such as landslides, collapses, debris flows, ground subsidence, slope instability, and reservoir bank deformation. Multi-source monitoring data, including displacement, cracks, water levels, and rainfall pressure, are susceptible to contamination by sensor drift, power supply fluctuations, communication interruptions, equipment aging, and manual data entry errors. It proposes an intelligent anomaly identification mechanism based on a pre-trained time series prediction model. This mechanism constructs historical prototypes of each monitoring variable, compares the prediction errors of the original historical segments with those after perturbation along the prototype direction, and obtains noise suspicion scores and prediction improvement degrees. Furthermore, it uses adaptive thresholds to divide clean and noisy moments, and fuses multi-variable results to form a set of reliable moments and a set of moments to be repaired. A self-repairing network is then trained on the reliable moments, using conditional vectors obtained from conditional historical segment encoding to constrain segment generation. Repair anchor points are selected based on the improvement degree, and write-back repair is performed on the anomaly intervals. This invention improves the accuracy of anomaly identification, the reliability of repaired segments, and the support capability for early warning analysis without altering the fundamental evolutionary laws of geological disaster monitoring data.

[0062] Reference Figure 1 This embodiment provides a method for anomaly identification and repair of geological disaster monitoring data. The multivariate time series data for geological disaster monitoring is organized in matrix form, where rows represent time step numbers, columns represent monitoring variable numbers, and matrix elements are the observed values ​​at the corresponding time steps and variables. In practical applications, observed values ​​can be rainfall, surface displacement, deep displacement, crack width, dip angle, surface subsidence, soil moisture content, pore water pressure, groundwater level, GNSS displacement, acceleration, vibration response, microseismic counts, or pre-processed statistical characteristics. This invention does not impose a unique limitation on the specific types of physical quantities.

[0063] Specifically, this includes steps S1 to S6:

[0064] S1. Obtain multivariate time series data for geological disaster monitoring; the multivariate time series data shall include at least the multi-channel observations of the same monitoring point, the same slope unit, the same landslide body, or the same disaster hazard area at different time steps; the multivariate time series data shall be derived from multi-sensor joint observations of a single monitoring point or from joint monitoring records of multiple neighboring monitoring points after spatial alignment.

[0065] S2. For each monitored variable, extract historical observation segments of the first historical length from the preset sampling index set, and perform pattern induction on the historical observation segments to obtain at least one historical morphological prototype corresponding to the monitored variable; wherein, the historical morphological prototype is obtained by performing clustering operations on the historical observation segments of each monitored variable, and the clustering operation adopts K-means clustering, Gaussian mixture clustering, or spectral clustering, or other clustering methods that can output typical historical pattern centers are also acceptable; each cluster center represents a historical change pattern that recurs during the geological disaster monitoring process, which is used for subsequent disturbance assessment and repair anchoring.

[0066] S3. For each sampling time in the sampling index set, extract the historical observation segment and the future true segment of the second prediction length corresponding to the monitoring variable; call the pre-trained time series prediction model, use the historical observation segment as input to obtain the future prediction segment, and calculate the baseline prediction error; then, based on each historical morphological prototype, perturb the prototype direction of the historical observation segment and re-predict, and calculate the perturbation prediction error; based on the difference between the baseline prediction error and each perturbation prediction error, obtain the prediction improvement degree of the sampling time relative to each historical morphological prototype, and obtain the noise suspicion score based on the baseline prediction error;

[0067] Among them, the prototype orientation perturbation is achieved by linearly combining the difference vectors of the current historical observation segment and the historical morphological prototype according to the preset perturbation coefficient, and can be superimposed with a zero-mean random perturbation term; the random perturbation term is used to improve the robustness of the evaluation process to local small morphological differences.

[0068] S4. Based on the noise suspicion score of each monitored variable, determine the noise segmentation threshold of the sub-variables and divide the time position of each monitored variable into net time and noisy time; merge the net time and noisy time of each monitored variable to obtain a set of credible time for repair training and a set of time to be repaired.

[0069] In this step, the noise segmentation threshold for the discrete variables is determined using the elbow rule; specifically, it includes:

[0070] 1) Sort the noise suspicion scores of the monitored variables in ascending order and normalize them;

[0071] 2) Calculate the degree of deviation of each normalized point on the sorting curve from the baseline line, and use it as the vertical distance;

[0072] 3) The original score corresponding to the point that maximizes the vertical distance is determined as the noise segmentation threshold for the monitored variable.

[0073] S5. For variables that are not marked as skipped repair, construct training samples on the set of reliable times and train a conditional generative self-repairing network.

[0074] The training samples consist of conditional history segments and target segments; the length of the conditional history segment is the third conditional length, and the length of the target segment is the fourth target length; the conditional history segment is taken from the observations of the third consecutive conditional length preceding the target segment, and the target segment is taken from the observations of the variable to be repaired within the target interval of the fourth consecutive target length.

[0075] The specific training process is as follows:

[0076] 1) Based on the set of credible moments, construct training sample pairs consisting of conditional history segments and target segments;

[0077] 2) Input the conditional history segment into the conditional encoder to obtain the conditional vector;

[0078] 3) Calculate the similarity between the conditional vector and the embedding vector corresponding to each historical morphological prototype, and select the historical morphological prototype with the highest similarity as the training anchor point;

[0079] 4) Starting with the noise superimposed on the training anchor point and ending with the target segment, construct the interpolation state on the time parameter;

[0080] 5) Train a velocity network, taking the interpolation state, time parameters and condition vector as input, and taking the prediction of the update direction from the current state to the target segment as the learning objective.

[0081] S6. For the variable to be repaired, extract the conditional historical fragments before the time to be repaired, obtain the conditional vector through the conditional encoder, and select the repair anchor point according to the predicted improvement degree of the time to be repaired relative to each historical prototype. Call the self-repairing network to generate the target fragment and write it back to the corresponding time interval to obtain the repaired geological disaster monitoring multivariate time series data.

[0082] The aforementioned conditional encoder includes: normalizing the input sequence; dividing the normalized sequence into several segments and obtaining a token sequence through linear mapping; performing multi-level causal self-attention operation on the token sequence, and taking the output latent vector as the conditional vector to represent the context state of the current monitored variable in a given historical window.

[0083] In the training phase, the training anchor point is selected based on the similarity between the conditional vector and the corresponding embedding vector of each historical form prototype; in the repair phase, the repair anchor point is selected based on the predicted improvement degree of the time to be repaired relative to each historical form prototype, thus forming a consistent mechanism in which the recognition phase and the repair phase are mutually closed.

[0084] In addition, in this embodiment, when the set of credible moments is empty, a preset proportion of low-scoring moments are selected as the alternative set of credible moments based on the multivariate joint score; or, after parsing and reshaping the original tabular records, steps S1 to S6 are executed, and the repair results are output to the database, file system or early warning platform interface in the original monitoring platform format.

[0085] This invention employs the following technical concept: First, using a pre-trained time series prediction model as a reference, the "anomaly" operation is transformed into a calculable index of "an abnormally increased difficulty in predicting future monitoring segments under given historical monitoring conditions." Second, a historical prototype library is used to characterize the typical historical change patterns of each monitoring variable, and controlled perturbations along the prototype direction are used to test whether the prediction error can be significantly improved, thereby obtaining information on the correlation strength related to the prototype. Third, the elbow rule is used to adaptively determine the segmentation threshold on the score ranking curve, forming the intersection of the net time of multiple variables and the union of the noisy time, and a skip repair strategy is enabled for variables with excessively high noise proportions. Finally, a conditional flow matching network is trained on credible time points, using a causal attention encoder to compress longer historical data as conditional vectors, and using the prototype as an anchor to perform flow matching generation. During the repair stage, anchor points are selected based on the degree of prediction improvement, realizing the structured write-back of the anomaly monitoring interval.

[0086] This embodiment's method comprises four logical stages: historical morphology prototype construction and anomaly intelligent identification and evaluation stage, adaptive threshold division and multivariate fusion stage, self-healing network training stage, and noisy interval segment repair and write-back stage. The first two stages complete noise identification and reliable sample division, while the latter two stages complete generative repair for each variable. For variables marked as skipping repair, training and write-back are not performed; only the scores, thresholds, and anomaly ratios obtained in the identification stage are retained for further verification by monitoring personnel.

[0087] 1. Historical prototype construction and anomaly intelligent identification and evaluation stage;

[0088] In the first sub-phase, the computing device independently performs the following process for each monitored variable:

[0089] First, determine the set of sampling times used for clustering and evaluation. For example... Figure 1 As shown, let the length of the first history be... , representing the length of the historical observation window; the second prediction length is , indicating the length of the future evaluation window; the sampling step size is To ensure that complete historical and future segments can be extracted at each sampling time, the sampling time must be no less than [number missing]. And it is guaranteed to still exist from that moment on. Starting from the location of future observations, subtract from the end of the sequence. Sampling continues along the time axis at preset step sizes. If the total number of time steps is denoted as T, then the sample set can be denoted as:

[0090]

[0091] A larger sampling step size helps reduce computational overhead, while a smaller sampling step size helps improve temporal resolution. In a geological disaster monitoring platform, this sampling step size parameter can be dynamically adjusted based on the sampling frequency, the number of monitoring points, and computational resources.

[0092] Secondly, for each time i and each variable j in the sampling set, extract the continuous time sequence of that variable before time i. The historical observation vector at each time step is denoted as . Historical observation segments of the same variable at various sampling times are vectorized and clustered to obtain K cluster centers. The number of clusters can be set according to the scale of monitoring data, variable complexity, and pattern diversity, or can be adaptively selected based on the silhouette coefficient, elbow rule, or other clustering quality indicators. Each cluster center, after being expanded along the same dimension as the historical observation segment, is denoted as K. This represents a typical historical pattern prototype of the monitored variable, such as a slow rise in displacement after continuous rainfall, a delayed change in deep displacement after a sudden increase in pore pressure, or a small fluctuation in crack width under a long-term stable background. Therefore, the historical pattern prototype is not a simple replication of a single sample, but a compressed expression of the evolutionary patterns of similar monitoring, which can provide a unified pattern library for subsequent definition of disturbance direction and anchoring of repair.

[0093] Next, the pre-trained time series prediction model is invoked. This model receives historical tensors during the inference phase and outputs a length of... The future predicted sequence. Based on current historical observation fragments. The model outputs a baseline future prediction fragment. Compare it with the actual future fragment corresponding to time i. Point-by-point comparison, calculate baseline prediction error:

[0094]

[0095] in, This represents the baseline prediction error at the j-th variable at the i-th sampling time. This represents the corresponding predicted value. This represents the true future value. The larger the first prediction error, the more difficult it is for the model to fit the future segment under the given historical conditions. This invention uses this error as the basis for the noise suspicion score.

[0096] The aforementioned pre-trained time series prediction model can be a causal sequence model pre-trained based on large-scale time series data, generating predictions of future segments through autoregression or direct mapping. This model can be deployed offline on a monitoring center server or perform inference on edge computing nodes or lightweight devices.

[0097] Then, for each historical prototype of this variable, a perturbed historical fragment is constructed. If the current historical fragment is denoted as... The k-th prototype is denoted as The first disturbance coefficient is denoted as The random disturbance term is denoted as Then the perturbated historical fragment can be represented as:

[0098]

[0099] in, Zero-mean Gaussian noise can be used, with its standard deviation controlled by a preset noise standard deviation parameter. Then, the perturbed historical segment is input into the same pre-trained time series prediction model to obtain the perturbed predicted segment, and the second prediction error between it and the future true segment is calculated:

[0100]

[0101] Based on this, the improvement in prediction relative to prototype k at this moment is defined as:

[0102]

[0103] If a perturbation in a certain prototype direction can significantly reduce the prediction error, then A positive and large value indicates that the current historical segment, after converging towards the prototype, better conforms to the normal prediction structure learned by the model. This prototype can be regarded as a favorable reference pattern for explaining the current anomalous history. The baseline error is used to characterize the prediction difficulty itself, while the improvement is used to characterize the strength of the interpretable correlation between the current segment and different typical patterns. In subsequent steps, they respectively serve different functions: noise measurement and anchor point selection.

[0104] Finally, the noise suspicion score and the prediction improvement on each prototype are recorded at the sampling time, where the noise suspicion score can be directly obtained:

[0105]

[0106] For unsampled moments, linear interpolation is performed based on the scores of two adjacent sampled moments on the time axis. If a certain unsampled moment... At adjacent sampling times Between these two points, the score can be expressed as:

[0107]

[0108] Predicted improvement Interpolation is performed on each prototype dimension in the same way. The significance of interpolation is that subsequent threshold division and repair write-back are performed at the time step granularity, requiring a score definition at each time position to avoid being operable only at sparse sampling points.

[0109] 2. Adaptive thresholding and multivariate fusion stage;

[0110] In the second sub-stage, the computing device performs threshold segmentation independently for each variable, and then performs multivariate fusion.

[0111] For a single variable, the noise suspicion scores of that variable at all time steps are sorted in ascending order to obtain a ranking sequence. To reduce the influence of dimensions, the ranking numbers are mapped to the interval between 0 and 1. The score is subtracted from the minimum value, divided by the difference between the maximum and minimum values, and a very small constant is added to prevent division by zero, thus obtaining the normalized ranking position and normalized score. A baseline line is formed by connecting the first and last points, and the perpendicular distance from each normalized point to this baseline line is calculated. If the baseline line is written as... Then the deviation of the (m)th sorting point can be written as:

[0112]

[0113] If the sorted score sequence is denoted as Then the normalized position and the normalized score can be expressed as follows:

[0114]

[0115] Where i represents the sorting position, and N represents the total number of time positions. This represents a minimal constant. Then, the first and last points are connected to form a baseline line, and the deviation of each normalized point from this baseline line is calculated. Within a preset search interval, a window smoothing search is performed, and the original score corresponding to the center of the window with the largest deviation is used as the noise segmentation threshold for that variable. This method can more naturally separate "normal background fluctuations" from "abnormally high-scoring areas," avoiding the problem of poor transferability of fixed thresholds between different monitoring points and different variables.

[0116] Time steps with scores below a threshold are classified as net time steps for that variable, while time steps with scores above the threshold are classified as noisy time steps. The proportion of noisy time steps to the total number of time steps for that variable is calculated. If this proportion exceeds a preset noise percentage limit (e.g., 20%), the variable is marked as a variable to skip repair. This strategy is suitable for situations involving long-term sensor failure, long-term data loss, or severe interference in the monitoring environment, preventing the system from forcibly fitting unreliable patterns to variables with excessive overall distortion.

[0117] At the multivariate level, the intersection of the net time sets of each monitored variable is used to obtain the global net time set; the union of the noisy time sets of each monitored variable is used to obtain the global noisy time set. Simultaneously, a noisy variable mapping structure is established: for each global noisy time, a list of indices of variables judged as noisy at that time is recorded. The meaning of the intersection and union is that the net time requires that all multi-source monitored variables be simultaneously reliable, facilitating the construction of consistent training samples across multiple channels; while the noisy time allows any variable that is suspected to enter the candidate set for repair, thus ensuring that no single-variable faults or local anomalies are overlooked.

[0118] If the global net time set is empty, it indicates that there is no time within the current time range where all variables are simultaneously reliable. In this case, a degradation strategy can be adopted: calculate the joint average of the multivariate noise suspicion scores for each time step, and select the preset proportion time position with the lowest average as the alternative reliable time set to ensure that the subsequent repair model still has a source of training samples. The above alternative strategy is particularly suitable for high-noise monitoring phases such as during periods of heavy rainfall in the flood season or during construction disturbances.

[0119] 3. Self-healing network training phase;

[0120] In the third sub-stage, the computing device independently trains a self-healing network for each variable that was not marked as skipping repair. Independent training can adapt to differences in the dimensions, noise mechanisms, and dynamic morphology of different monitored variables; in other embodiments, a shared backbone network combined with variable encoding can also be used for joint training, which is not the only option in this invention.

[0121] First, training samples are constructed. Training samples consist of conditional history fragments and target fragments, where the length of the conditional history fragment is denoted as . The length of the target segment is denoted as To ensure the extraction of complete conditional history fragments and target fragments, a sufficient length of historical observations must be retained before the target fragment. Furthermore, to ensure the training samples are not significantly contaminated by noise, this embodiment requires that, with the end time of the target fragment as a reference, the corresponding variable within its preceding continuous target length interval should not be marked as anomaly by noisy variable mappings. Time intervals satisfying this condition are used to construct training samples, where the conditional history fragment is taken from the continuous period preceding the target fragment. Observations at each time step, the target segment is taken from the continuous observations of the variable within the target interval. The observations are taken at each time step. The purpose of the above design is to ensure that both the conditional fragments and the target fragments come from a reliable distribution, so that the mapping learned by the repair network is closer to the real monitoring dynamics.

[0122] Secondly, historical prototypes are encoded. Each prototype fragment is input into the prototype encoding path of the conditional encoder to obtain a prototype embedding vector; simultaneously, the prototype fragments are organized into a prototype tensor with the same scale as the training target, serving as candidate anchor points for flow matching. The prototype encoding process can use the same length alignment, normalization, and segmentation mapping rules as the input in the training phase to ensure consistency of the representation space.

[0123] During training iterations, for a batch of training samples, conditional history fragments are input into the conditional encoder to obtain conditional vectors. The similarity between the conditional vectors and each prototype embedding vector is calculated, and the index with the highest similarity is selected. The corresponding prototype tensor is then extracted as the stream matching anchor point for this batch of training samples. If the conditional vector is denoted as... The k-th prototype embedding vector is denoted as Then, cosine similarity can be used:

[0124]

[0125] The prototype corresponding to the maximum value is selected. Then, the flow matching loss is calculated: the time parameter is uniformly and randomly sampled in the interval between 0 and 1; the short target segment is convexly combined with the starting point formed by superimposing random noise on the prototype tensor according to the time parameter to obtain the interpolation state; the interpolation state, the scaled time parameter and the conditional vector are input into the velocity network to obtain the network output; the squared error between the network output and the target segment is used as the training loss, and excessively large loss terms can be masked to suppress outlier batches.

[0126] Flow matching generation can employ a linear interpolation path connecting the target segment and the prototype noise-adding starting point on a zero-to-one time parameter, and construct a training target by regressing the difference between the target and the current state using a velocity network. The sampling phase uses multi-step discrete iterations to approximate the continuous flow. Preferably, for the target segment y, the prototype tensor p, and random noise... and time parameters The interpolation state can be written as:

[0127]

[0128] Speed ​​network with The time parameter t and the condition vector z are the input and output predictions. And optimized through squared error; during sampling from Starting from this point, the final generated fragment is obtained through multiple discrete updates.

[0129] One type of form can be written as

[0130]

[0131] Where B represents the batch size. This indicates that the b-th training sample has the following time parameter. The interpolation state under the following conditions This represents the corresponding conditional vector. This represents the corresponding short target segment. The above expression corresponds to the direct regression stream matching training method in this embodiment.

[0132] One implementation of a conditional encoder includes: calculating and normalizing the mean and standard deviation of the input one-dimensional sequence batch by batch; if the input historical sequence is denoted as c, its normalized form can be written as...

[0133]

[0134] and The mean and standard deviation of the current condition historical segment are represented respectively; the sequence is divided into several sub-segments, and each sub-segment is linearly mapped to a high-dimensional latent space; the token sequence is processed through a multi-layer causal self-attention module, where the self-attention uses a mask to ensure causality, and the query and key vectors are encoded by rotation; finally, the latent vector of the last token is taken as the condition vector.

[0135] 4. Noisy region segment repair and write-back stage;

[0136] In the fourth sub-stage, the computing device iterates through the time points in the time set where the variable to be repaired is marked as noise. For each time point to be repaired, a conditional history fragment is extracted and encoded into a conditional vector. Unlike the training stage, the prototype selection in the repair stage is based on the predicted improvement degree of the variable at that time point across different prototypes, selecting the prototype tensor corresponding to the index with the highest improvement degree as the repair anchor point. The significance of this differentiated strategy is that the training stage emphasizes matching with the encoder representation, while the repair stage emphasizes consistency with the most interpretable prototype determined in the noise identification stage, thus forming an end-to-end closed loop.

[0137] The multi-step iterative generation process of the flow matching generation network is then invoked: The initial state is recorded using a prototype tensor superimposed with random noise; iterating from 0 on a grid that discretizes the time interval from 0 to 1 into several steps, each step inputting the current state, current time parameter, and condition vector into the velocity network, updating the state according to the discrete step length, until the termination time is reached, thus obtaining the generated segment. Let the discrete step length be denoted as... The initial state is denoted as Then the s-th step can be expressed as:

[0138]

[0139] The generated segments are destandardized using the same historical statistics as during training to obtain the repaired values ​​in the target interval, and then written back to the corresponding positions in the data matrix. The destandardization process can be written as follows:

[0140]

[0141] in, This represents the normalized segment of the generator network output. and These represent the mean and standard deviation of the historical data under the current conditions, respectively.

[0142] It should be noted that when the repair process traverses noisy moments in chronological order, the conditional history fragments on which subsequent repairs depend may contain values ​​that have been previously written back, thus forming a certain degree of time consistency propagation. This invention does not make a unique limitation on whether to force sorted traversal, but it is recommended to sort by time index to reduce random overwriting.

[0143] Example 2:

[0144] Reference Figure 2 This embodiment provides an anomaly identification and repair system for geological disaster monitoring data, used to execute the method described in Embodiment 1. The system includes: a data acquisition unit, an anomaly intelligent identification and evaluation unit, a sample division unit, a self-repair unit, and a data parsing and output unit.

[0145] The data acquisition unit is used to read raw records such as rainfall, displacement, cracks, tilt angle, pore pressure, groundwater level, and GNSS displacement from monitoring terminals, historical databases, message queues, or file media, and organize them into multivariable matrices or tensors aligned by time index.

[0146] The anomaly intelligent identification and evaluation unit is used to perform sampling time extraction, historical fragment clustering, pre-training prediction, prototype perturbation evaluation, score calculation and interpolation propagation, and output noise suspicion score tensor and prediction improvement tensor.

[0147] The sample partitioning unit is used to perform elbow threshold search, noise ratio judgment, multivariate net time intersection and noisy time union fusion, skip repair marker, and generation of alternative reliable time set.

[0148] Self-healing units are used for prototype encoding, conditional vector extraction, training sample construction, speed network training, noise interval repair generation, and fragment write-back.

[0149] The data parsing and output unit is used to restore the internal matrix results into the table structure, file structure or interface format required by the monitoring platform, and supports outputting fields such as "original value, repair value, score, threshold, and whether it is repaired" for display by the early warning platform, manual review and subsequent model calls.

[0150] The modules described above can be implemented by different program units on the same computing device, or they can be distributed among monitoring edge nodes, central servers, and business platforms. This invention does not limit the specific deployment form.

[0151] Example 3: Typical examples of parameters and data formats;

[0152] This embodiment provides a set of examples of workable parameter combinations, which are only used to illustrate the feasibility of the present invention.

[0153] In a typical configuration, the first history length A total of 96 time steps can be selected, and the second prediction length is [not specified]. One or more time steps can be selected; third condition length Can be with The same or different settings may be required depending on the repair context; fourth target length Can be with The number of clusters K can be set to be the same or based on the target interval length; the perturbation coefficient can be 10 to 15. The value can be taken as 0.4 to 0.6; the standard deviation of random disturbance can be taken as 0.05 to 0.1; the sampling step size The step size can be 1 to 10; the upper limit for the noise percentage can be 20%; when extracting training samples, the first 80% of the time steps can be used as the priority cleaning training interval, and the last 20% of the time steps can be used as the effect comparison interval. For high-frequency monitoring points, the sampling step size can be appropriately increased; for strong deformation stages, the step size can be appropriately decreased to enhance recognition sensitivity.

[0154] Data input can be in long table format, wide table format, database time series table format, or message stream format. For cases where the original records contain fields for station number, monitoring point number, variable name, and sampling time, a perspective reshaping can be performed first based on the monitoring point and variable dimensions, and then aligned according to a unified time grid to form the multivariate time series matrix required by the algorithm.

[0155] Furthermore, in practical deployment, the data pipeline can adopt a segmented strategy of training interval and evaluation interval. For example, the first 80% of the time series can be used as the training interval, in which the complete anomaly identification and repair process is executed, and the repaired data is used to train the self-repairing network; the last 20% of the time series can be used as the evaluation interval, in which the original observations are kept unchanged, but the repair results generated by the method of this invention are also run, so as to conduct independent and objective comparative evaluation of the repair effect.

[0156] The repair output can either form a complete repair sequence on the timeline, or it can only output the repair suggestion values ​​and corresponding scores of the intervals that are judged to be abnormal, for manual review to decide whether to revert to the previous version.

[0157] Example 4: Electronic device and storage medium;

[0158] Taking a landslide monitoring point as an example, the system continuously collects data on rainfall, surface displacement, deep displacement, crack width, groundwater level, and pore water pressure. During continuous rainfall, due to power supply fluctuations and sensor drift, local spikes and short-term breaks appear in the crack width and pore water pressure sequences. If a fixed threshold method is used, it is easy to confuse the real rapid response caused by heavy rainfall with instrument malfunction; if a simple interpolation method is used, it may weaken the real precursors of accelerated deformation.

[0159] When employing the scheme of this invention, the system first clusters historical segments of each variable to form prototypes, and then uses a pre-trained prediction model to calculate the baseline prediction error and the perturbation prediction error. When a historical segment of crack width is perturbed towards a prototype direction of "sustainable stabilization followed by slow rise," the future prediction error decreases significantly, indicating that the original historical segment is likely to contain local anomalies rather than true structural mutations. Subsequently, the time period is identified as a noisy moment by using noise suspicion score and elbow threshold, and generative repair is performed on the interval in a self-repairing network trained on the set of credible moments, based on the prototype with the greatest improvement. The repaired crack width sequence retains the overall rising trend during the rainfall process and eliminates sharp perturbations that do not conform to the coupling relationship of surrounding variables, thus making it more conducive to stability analysis by the early warning platform.

[0160] Example 5: Verification example based on real monitoring data;

[0161] To further illustrate the feasibility of the method of this invention in real geological disaster monitoring scenarios, this embodiment uses a set of anonymized monitoring data of real landslide hazard points as the verification object. The data comes from the historical database of the on-site automated monitoring terminal and monitoring platform, with a monitoring period covering approximately 10 consecutive months and a sampling granularity of 1 hour. To avoid exposing specific engineering information, this embodiment anonymizes the monitoring point names, coordinates, project units, and some absolute value ranges, but retains core information related to the technical effects of this invention, such as data structure, variable composition, anomaly types, and processing results.

[0162] In this embodiment, six types of monitoring variables were selected: hourly rainfall, surface displacement, deep displacement, crack width, groundwater level, and pore water pressure, forming a six-channel multivariate time series for geological disaster monitoring. The original time series consisted of approximately 7200 time steps. Combining on-site inspection records, equipment maintenance logs, and platform alarm records, several typical anomalies were identified in the data, including instantaneous spike noise, slow drift, short-term communication interruptions, duplicate reporting, and a small number of manual data entry errors. Preliminary analysis indicated that the aforementioned anomaly-related segments accounted for approximately 5% to 7% of the total time steps.

[0163] Table 1 lists the basic information of the experimental data used in this embodiment.

[0164] Table 1 Overview of Actual Monitoring Data

[0165]

[0166] As shown in Table 1, the data used in this embodiment has a long coverage period and includes rainfall, displacement, cracks and hydrological variables. It can better reflect the actual characteristics of multi-source variable coupling, complex anomaly types and high requirements for time series continuity in geological disaster monitoring scenarios. Therefore, it is suitable as a verification object of the method of this invention.

[0167] To generate experimental data suitable for verification, this embodiment manually reviewed a portion of the abnormal segments based on on-site inspection records, maintenance logs, and the coupling relationships between adjacent variables, resulting in a reference set containing over 80 abnormal segments. Unlabeled but generally reliable time ranges were then used as supplementary reliable sample sources. In the experiment, the first 80% of time steps were used as the main intervals for prototype construction, threshold determination, and network training, while the last 20% were used as the comparison observation interval to avoid relying entirely on the training interval results during the verification phase. Historical observation window length. Set to 96, future evaluation window length Set to 12, sampling step size Set to 4, cluster center number K to 10, perturbation coefficient Set the value to 0.4, the standard deviation of random disturbance to 0.05, and the upper limit of noise percentage to 20%.

[0168] Before the data enters the method of this invention, all variables are aligned according to a unified time grid, and the monitoring point number, variable name, sampling time, and observation value fields are reshaped to form a monitoring matrix organized according to "time step-variable". Then, processing is performed according to the four-stage process in the aforementioned embodiment: first, a historical prototype is constructed for each variable, and a noise suspicion score is calculated. and predicted improvement Secondly, the elbow rule is used to determine the threshold of each variable and form a global set of credible moments and a set of moments to be repaired; then, conditional history segments and target segments are constructed on the credible moments to train the self-repairing network; finally, anchor point selection, segment generation and de-standardization are performed on the identified abnormal segments.

[0169] In the comparative analysis of this embodiment, to illustrate the technical effects of the present invention, the method of the present invention is compared with linear interpolation repair, moving average smoothing, and anomaly identification methods based solely on predicted residual thresholds. The results show that, on the aforementioned anonymized monitoring data, the method of the present invention has a relatively stable identification effect on manually reviewed anomaly segments, with a recall rate of approximately 84% and an accuracy rate of approximately 80%. Compared with identification methods based solely on residual thresholds, both recall and accuracy show a certain improvement. For the three variables that have a significant impact on early warning judgment—crack width, deep displacement, and pore water pressure—compared with reference values ​​constructed from credible segments after manual review, the average mean square error of the method of the present invention in anomaly segments decreases by approximately 18% to 25% compared to before repair, and also shows further improvement compared to the linear interpolation scheme.

[0170] Table 2 presents exemplary statistics of the main comparison results in this embodiment.

[0171] Table 2 Comparison of Treatment Effects

[0172]

[0173] As shown in Table 2, under the condition of anonymized real monitoring data, the method of the present invention exhibits relatively stable comprehensive performance in both abnormal segment identification and abnormal section repair. Compared with identification methods based solely on prediction residual thresholds, the present invention achieves a more balanced result between recall and precision. Compared with linear interpolation repair methods, the present invention performs better in terms of the reduction of the mean square error of abnormal sections, trend continuity, and cross-variable coupling consistency, which better meets the actual needs of subsequent trend analysis and early warning judgment of geological disaster monitoring data.

[0174] Furthermore, from a monitoring perspective, the repaired time series exhibits a more continuous trend response during rainfall, and the phased coupling relationship between rainfall, groundwater level, and displacement is more consistent with manual inspection records. Within several short-term communication interruption intervals, the repaired segments generated by the method of this invention maintain a consistent slope and fluctuation amplitude with adjacent time periods, without exhibiting significant excessive flattening. Within slow offset intervals caused by sensor drift, the method of this invention can utilize historical morphological prototypes and intelligent anomaly identification and evaluation results to adjust unreasonable drifts back to numerical ranges that better conform to the overall evolutionary pattern. This demonstrates that the present invention not only improves the quality of abnormal segments in terms of statistical indicators but also enhances the stability of early warning judgment at the monitoring interpretation level, and provides a more reliable data foundation for subsequent manual review, threshold early warning, and risk trend analysis.

[0175] It should be noted that the above-mentioned experimental data sources, sample size, number of variables, parameter settings, and effect indicators are merely a set of exemplary verification results illustrating the feasibility of this invention under real monitoring conditions, and do not constitute a limitation on the applicable disaster types, monitoring variable combinations, sampling frequency, monitoring scale, or upper limit of the effect of this invention. In other embodiments, minute-level sampling, daily sampling, joint modeling of more monitoring points, or fewer variable combinations can also be used as data organization forms, and the method of this invention is equally applicable.

[0176] Example 6: Electronic device and storage medium;

[0177] Reference Figure 3 As shown, this embodiment of the invention also provides an electronic device for geological disaster early warning. The electronic device may include a processor, a memory, a communication bus and a communication interface, and may also include a computer program stored in the memory and capable of running on the processor, for executing the method for anomaly identification and repair of geological disaster monitoring data in Embodiment 1.

[0178] In some embodiments, the processor may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions. This includes combinations of one or more central processing units, microprocessors, digital processing chips, graphics processors, and various control chips. The processor is the control core of the electronic device, connecting various components of the device through various interfaces and lines. It executes programs or modules stored in memory and calls data stored in the memory to perform various functions and process data within the electronic device.

[0179] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, devices, or computer program products, etc. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer.

[0180] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0181] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for anomaly identification and repair of geological disaster monitoring data, characterized in that, Includes the following steps: S1. Obtain multivariate time series data for geological disaster monitoring; S2. For each monitored variable, extract a historical observation segment of the first historical length from the preset sampling index set, and perform pattern induction on the historical observation segment to obtain at least one historical form prototype corresponding to the monitored variable. S3. For each sampling time in the sampling index set, extract the historical observation segment and the future true segment of the second predicted length corresponding to the monitoring variable; A pre-trained time series prediction model is invoked, and the future prediction segment is obtained by taking the historical observation segment as input. The baseline prediction error is calculated. Then, the historical observation segment is perturbed by the prototype direction based on each historical morphological prototype and re-predicted, and the perturbation prediction error is calculated. The prediction improvement degree of the sampling time relative to each historical morphological prototype is obtained according to the difference between the baseline prediction error and each perturbation prediction error, and the noise suspicion score is obtained according to the baseline prediction error. S4. Based on the noise suspicion score of each monitored variable, determine the noise segmentation threshold of the sub-variables and divide the time position of each monitored variable into net time and noisy time; merge the net time and noisy time of each monitored variable to obtain a set of credible time for repair training and a set of time to be repaired. S5. For variables that are not marked as skipped repair, construct training samples on the set of reliable times and train a conditional generative self-repairing network. S6. For the variable to be repaired, extract the conditional historical fragments before the time to be repaired, obtain the conditional vector through the conditional encoder, and select the repair anchor point according to the predicted improvement degree of the time to be repaired relative to each historical prototype. Call the self-repairing network to generate the target fragment and write it back to the corresponding time interval to obtain the repaired geological disaster monitoring multivariate time series data.

2. The method as described in claim 1, characterized in that, In step S1, the multivariate time series data includes at least the multi-channel observations at different time steps from the same monitoring point, the same slope unit, the same landslide body, or the same disaster hazard area.

3. The method as described in claim 2, characterized in that, The multi-channel observations include at least one of the following: rainfall, surface displacement, deep displacement, crack width, dip angle, surface subsidence, soil moisture content, pore water pressure, groundwater level, GNSS displacement, acceleration, vibration response, or microseismic response data; the multivariate time series data are derived from multi-sensor joint observations at a single monitoring point or from joint monitoring records of multiple neighboring monitoring points after spatial alignment.

4. The method as described in claim 1, characterized in that, In step S2, the historical morphology prototype is obtained by performing clustering operations on the historical observation segments of each monitoring variable. The clustering operation adopts K-means clustering, Gaussian mixture clustering, or spectral clustering. Each cluster center represents a historical change pattern that recurs during geological disaster monitoring and is used for subsequent disturbance assessment and repair anchoring.

5. The method as described in claim 1, characterized in that, In step S3, the prototype orientation perturbation is achieved by linearly combining the difference vectors of the current historical observation segment and the historical morphological prototype according to a preset perturbation coefficient, and a zero-mean random perturbation term can be superimposed; the random perturbation term is used to improve the robustness of the evaluation process to local small morphological differences.

6. The method as described in claim 1, characterized in that, In step S4, the noise segmentation threshold for the discrete variables is determined using the elbow rule; specifically, it includes: The noise suspicion scores of the monitored variables are sorted in ascending order and then normalized. Calculate the degree of deviation of each normalized point on the sorting curve from the baseline line, and use it as the vertical distance; The original score corresponding to the point that maximizes the vertical distance is determined as the noise segmentation threshold for the monitored variable.

7. The method as described in claim 1, characterized in that, In step S4, the net time and noisy time of each monitored variable are merged, including: The intersection of the net time sets of each monitored variable is used to obtain the set of reliable times; The set of noisy time points of each monitored variable is combined to obtain the set of time points to be repaired.

8. The method as described in claim 7, characterized in that, Step S4 also includes: When the proportion of noisy moments of a certain monitoring variable exceeds the preset limit, the monitoring variable is marked as a skip-repair variable to avoid continuing training on variables that are heavily contaminated overall, which would lead to error reinforcement.

9. The method as described in claim 1, characterized in that, In step S5, a conditionally generative self-healing network is trained, including: Based on the set of reliable moments, a training sample pair consisting of conditional history fragments and target fragments is constructed. The conditional history segment is input into the conditional encoder to obtain the conditional vector; Calculate the similarity between the conditional vector and the embedding vector corresponding to each historical morphological prototype, and select the historical morphological prototype with the highest similarity as the training anchor point. Starting with noise superimposed on the training anchor points and ending with the target segment, an interpolated state is constructed on the time parameters. A velocity network is trained, with the interpolation state, time parameters, and conditional vector as input, and the learning objective is to predict the update direction from the current state to the target segment.

10. A system for anomaly identification and repair of geological disaster monitoring data, characterized in that, The system, comprising the method for anomaly identification and repair of geological disaster monitoring data as described in any one of claims 1-9, includes: The data acquisition unit is used to acquire and organize multivariate time series data for geological disaster monitoring; An anomaly intelligent identification and evaluation unit is used to construct historical pattern prototypes, call pre-trained time series prediction models, and output noise suspicion scores and prediction improvement degrees. The sample partitioning unit is used to determine the net time and the noisy time based on the noise suspicion score, and to form a set of credible times and a set of times to be repaired; The self-repair unit is used to train the self-repair network on a set of reliable time points and to perform target segment generation and write-back on the interval corresponding to the time point to be repaired. The data parsing and output unit is used to write the restored monitoring data back to the geological disaster monitoring database, slope monitoring platform, or early warning system.

Images