A highway infrastructure monitoring data anomaly identification method
By constructing an anomaly pattern generation space and cross-modal embedding training model, the problem of anomaly identification in highway infrastructure monitoring data is solved, and the identification stability and accuracy are improved. It is applicable to various engineering scenarios such as bridges, pavements and slopes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUIZHOU PROVINCIAL QUALITY & SAFETY TRAFFIC ENG MONITORING & TESTING CENT CO LTD
- Filing Date
- 2026-02-14
- Publication Date
- 2026-05-29
AI Technical Summary
Highway infrastructure monitoring data is susceptible to sensor failures, unstable power and communication, and environmental interference during collection and transmission, resulting in anomalies such as outliers, missing segments, trend drift, and square wave interference. Traditional methods are difficult to adapt to complex anomaly patterns, resulting in unstable detection performance. Furthermore, the scarcity of samples in engineering scenarios leads to insufficient anomaly identification accuracy and generalization ability.
An anomaly pattern generation space based on real monitoring data is constructed to generate a synthetic monitoring data set. An anomaly recognition embedding model is generated by training through cross-modal embedding space alignment. The visual feature vectors of the monitoring data to be identified are extracted to identify its anomaly category.
While maintaining the consistency of real data distribution, it generates diverse abnormal samples, improves the model's recognition stability and accuracy under small sample conditions, enhances the clarity of the discrimination boundary between abnormal categories, and is applicable to a variety of engineering scenarios.
Smart Images
Figure CN122116138A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of highway infrastructure health monitoring, and in particular to a method, apparatus, and storage medium for identifying anomalies in highway infrastructure monitoring data. Background Technology
[0002] Health monitoring data for highway infrastructure is susceptible to sensor malfunctions, unstable power and communication, and environmental interference during actual collection and transmission, leading to anomalies such as outliers, missing segments, trend drift, and square wave interference. Due to the diverse types of anomalies and the non-stationary and noisy characteristics of the data, traditional anomaly detection methods relying on statistical features or threshold rules struggle to adapt to complex anomaly patterns, resulting in unstable detection performance. Meanwhile, learning-based methods relying on a large number of labeled samples fail to meet the constraints of sample scarcity in engineering scenarios, leading to insufficient anomaly identification accuracy and generalization ability, thus affecting the accuracy of structural damage identification and condition assessment. Summary of the Invention
[0003] The first aspect of this invention provides a method for identifying anomalies in highway infrastructure monitoring data, comprising the following steps: Based on the statistical prior of real monitoring data, an anomaly pattern generation space is constructed. The anomaly pattern generation space represents the distribution structure of generation parameters of different anomaly patterns and their perturbation mapping relationship under the statistical distribution constraints of the real monitoring data. Within the anomaly pattern generation space, a synthetic monitoring data set is generated based on the generation parameter distribution structure and its perturbation mapping relationship. Each synthetic monitoring data sample includes a sequence of monitoring values arranged according to a predetermined sampling time order and a corresponding anomaly category identifier. Based on the synthetic monitoring data set, cross-modal embedding space alignment training is performed on the visual features of the monitoring data and the semantic features of the corresponding anomaly categories, so that samples of the same category form a clustered distribution structure in the embedding space, and an anomaly recognition embedding model is generated. Using the aforementioned anomaly identification embedding model, visual feature vectors of the monitoring data samples to be identified are extracted, and based on the visual feature vectors, the anomaly category corresponding to the monitoring data samples to be identified is determined.
[0004] In some embodiments, the abnormal mode generation space is generated through the following steps: Statistical feature parameters characterizing the distribution characteristics of the data are extracted from real monitoring data, and these statistical feature parameters are used to limit the distribution boundary of the synthetic monitoring data. For different anomaly patterns, a corresponding set of generation parameters is constructed, which is used to characterize the morphological feature change parameters of the anomaly pattern; The value range of the generated parameter set is constrained based on the statistical feature parameters to form a generated parameter distribution structure; The original monitoring value sequence that is operating normally is selected from the real monitoring data as the reference sequence for anomaly generation, and a perturbation mapping relationship between the generation parameter and the original monitoring value sequence is constructed so that the generation parameter acts on the original monitoring value sequence under the statistical distribution constraint.
[0005] In some embodiments, the statistical characteristic parameters are obtained by performing sliding window segmentation on the real monitoring data and statistical analysis on each segment of the data. They include one or more of the following: amplitude distribution range, mean range, variance range, extreme value range, frequency domain energy distribution characteristics, dominant frequency range, or trend change range.
[0006] In some embodiments, under any anomaly mode, the original monitoring value sequence is used to generate a synthetic monitoring data sample through the following steps: Based on the distribution structure of the generated parameters corresponding to the abnormal mode, the abnormal generated parameters are selected within its limited value range as parameter instances corresponding to this abnormal mode. Based on the disturbance mapping relationship corresponding to the abnormal mode, the original monitoring value sequence is structurally transformed based on the parameter instance to generate an abnormal monitoring value sequence. By associating the abnormal monitoring numerical sequence with the abnormal category identifier corresponding to the abnormal pattern, a synthetic monitoring data sample corresponding to the abnormal pattern is generated.
[0007] In some embodiments, generating the anomaly detection embedding model includes the following steps: A visual feature encoding module is constructed, and the monitoring numerical sequence of the synthetic monitoring data sample is feature encoded to obtain a visual feature vector representing the abnormal morphological features. A semantic feature encoding module is constructed, and the anomaly category identifiers of the synthetic monitoring data samples are semantically vectorized to obtain the corresponding semantic feature vectors; The visual feature vector and the semantic feature vector are mapped to a unified embedding space. The unified embedding space is optimized and trained based on the objective function of minimizing the distance between samples of the same class and maximizing the distance between samples of different classes. This minimizes the embedding distance between samples of the same class and maximizes the embedding distance between samples of different classes, thereby forming multiple feature clustering structures centered on the semantic representation vector of the abnormal class in the unified embedding space.
[0008] In some embodiments, identifying the anomaly category corresponding to the monitoring data sample to be identified based on the visual feature vector includes the following steps: Calculate the similarity value between the visual feature vector and the semantic representation vector of each anomaly category to obtain the similarity distribution corresponding to the monitoring data sample to be identified; Based on the similarity distribution, the anomaly category is determined according to the following rules: when the similarity between the visual feature vector and one or more anomaly category semantic representation vectors meets the preset determination conditions, the corresponding anomaly category is determined as one or more anomaly categories corresponding to the monitoring data sample to be identified.
[0009] In some embodiments, the semantic representation vector is: In the feature clustering structure, the semantic center vector is obtained by aggregating the semantic feature vectors corresponding to the same anomaly category.
[0010] In some embodiments, the preset determination condition is that the similarity value between the visual feature vector and the corresponding anomaly category semantic center vector is greater than a preset threshold.
[0011] Secondly, based on the method for identifying anomalies in highway infrastructure monitoring data provided in the first aspect, the present invention also provides a device for identifying anomalies in highway infrastructure monitoring data, comprising: The statistical modeling module is used to construct an anomaly pattern generation space based on the statistical prior of real monitoring data. The anomaly pattern generation space represents the distribution structure of generation parameters of different anomaly patterns and their perturbation mapping relationship under the statistical distribution constraints of the real monitoring data. The sample generation module is used to generate a synthetic monitoring data set based on the generation parameter distribution structure and its perturbation mapping relationship within the abnormal pattern generation space. Any synthetic monitoring data sample includes a sequence of monitoring values arranged according to a predetermined sampling time order and a corresponding abnormal category identifier. The embedding training module is used to perform cross-modal embedding space alignment training on the visual features of the monitoring data and the semantic features of the corresponding anomaly category based on the synthetic monitoring data set, so that samples of the same category form a clustered distribution structure in the embedding space, and generate an anomaly recognition embedding model. An anomaly identification module is used to extract the visual feature vector of the monitoring data sample to be identified using the anomaly identification embedding model, and to identify the anomaly category corresponding to the monitoring data sample to be identified based on the visual feature vector.
[0012] Thirdly, based on the method for identifying anomalies in highway infrastructure monitoring data provided in the first aspect, the present invention also provides a computer storage medium on which a computer program is stored. When the computer program is executed by a processor, the processor performs the steps of the method for identifying anomalies in highway infrastructure monitoring data described in any of the above embodiments.
[0013] The method, apparatus, and storage medium for identifying anomalies in highway infrastructure monitoring data provided by this invention have gains including at least: This invention constructs an anomaly pattern generation space based on statistical prior constraints of real monitoring data. This allows for the generation of diverse anomaly samples while maintaining the consistency of real data distribution, addressing the problem of insufficient anomaly labeling samples in engineering scenarios. Compared to models trained solely on real samples, this invention significantly improves the coverage and distribution completeness of anomaly categories, thereby enhancing the model's recognition stability under small sample conditions. Furthermore, this invention parameterizes the anomaly pattern into a generation parameter distribution structure and establishes a perturbation mapping relationship between the generation parameters and the original monitoring value sequence, so that the anomaly generation process has clear statistical constraints and morphological control logic, avoiding the distribution offset problem caused by random perturbation, thereby improving the engineering credibility and physical rationality of the synthetic data. Furthermore, this invention constructs a unified embedding space for visual and semantic features, and trains it based on a comparative optimization objective function that minimizes the distance between the same category and maximizes the distance between different categories. This enables different anomaly categories to form separable clustering structures in the embedding space, thereby enhancing the clarity of the discrimination boundary between anomaly categories and reducing the misclassification rate. Furthermore, this invention generates abnormal samples under the constraints of real statistical distribution and combines them with an embedded semantic alignment mechanism, enabling the recognition model to adapt to data distribution under different sampling frequencies and sensor conditions, while maintaining the consistency of category expression. This makes it suitable for various engineering scenarios such as highway bridges, pavement structures, and slope monitoring. Attached Figure Description
[0014] From the following description of embodiments in conjunction with the accompanying drawings, aspects, features, and advantages of the present invention will become clearer and more readily understood, in which: Figure 1 This is a schematic diagram of the method for identifying anomalies in highway infrastructure monitoring data provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the functional modules of the highway infrastructure monitoring data anomaly identification device provided in an embodiment of the present invention. Detailed Implementation
[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.
[0016] It should be noted that the described embodiments are only a portion of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0017] In the description of this application, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary details.
[0018] In addition, the term "connection" should be interpreted broadly, where "A connects to B" includes both the case where A is directly connected to B and the case where A is indirectly connected to B via one or more middleware. "A connects between B and C" includes both the case where A is directly connected to B and C respectively and the case where A is indirectly connected between B and C via one or more middleware.
[0019] In one embodiment, see Figure 1 , Figure 1 This is a schematic diagram of the process for identifying anomalies in highway infrastructure monitoring data provided in an embodiment of the present invention.
[0020] It should be noted that the highway infrastructure described in this invention includes, but is not limited to, bridges, tunnels, pavements, roadbeds, piers, guardrails, drainage facilities, curtain walls, and other engineering structures that require structural stress or operational status monitoring.
[0021] In any of the above-mentioned highway infrastructure health monitoring scenarios, the highway infrastructure monitoring data refers to the raw time-series numerical signals collected by one sensor or multiple sensors of the same type from the monitored structure at a predetermined sampling time sequence node. These data may include, but are not limited to, sensor identification information, timestamp information, and corresponding measurement values.
[0022] Among them, highway infrastructure monitoring data can come from different types of monitoring sensors, including but not limited to data collected by structural health monitoring equipment such as acceleration sensors, vibration sensors, temperature sensors, humidity sensors, atmospheric pressure sensors, and ultrasonic detectors.
[0023] like Figure 1 As shown in this embodiment, the method for identifying anomalies in highway infrastructure monitoring data specifically includes the following steps: S01. Based on the statistical prior of real monitoring data, an anomaly pattern generation space is constructed. The anomaly pattern generation space represents the distribution structure of generation parameters of different anomaly patterns and their disturbance mapping relationship under the statistical distribution constraints of the real monitoring data.
[0024] In this embodiment, the statistical prior of real monitoring data refers to the statistical feature parameters extracted from real monitoring data samples that can characterize the overall distribution characteristics of the data, and are used to depict the distribution boundary and change pattern of monitoring data in real engineering scenarios.
[0025] Furthermore, the anomaly pattern generation space is a constrained parameterized anomaly generation domain constructed based on the aforementioned statistical prior, used to uniformly model different anomaly patterns under the constraints of the statistical distribution of real monitoring data.
[0026] In this embodiment, the construction of the anomaly pattern generation space based on statistical priors of real monitoring data includes the following steps: S011. Extract statistical feature parameters that characterize the distribution characteristics of the data from the real monitoring data. The statistical feature parameters are used to limit the distribution boundary of the synthetic monitoring data.
[0027] Furthermore, the statistical characteristic parameters include, but are not limited to, one or more of the following: amplitude distribution range, mean range, variance range, extreme value range, frequency domain energy distribution characteristics, dominant frequency range, or trend change range of the monitored values.
[0028] The aforementioned statistical characteristic parameters are used to characterize the distribution boundaries and dynamic changes of real monitoring data in the time and frequency domains, thereby forming statistical constraints on the abnormal pattern generation process.
[0029] Specifically, one or more of the above-mentioned statistical characteristic parameters can be obtained by segmenting the real monitoring data into sliding windows and performing statistical analysis on the data within each window.
[0030] It should be noted that the above statistical characteristic parameters are merely illustrative examples and do not constitute a limitation of the present invention. In practical applications, the statistical characteristic parameters can be extended or adjusted according to factors such as the structural characteristics of different types of highway infrastructure, the type of monitoring sensor, and the sampling frequency, so as to achieve reasonable constraints on the abnormal generation parameters.
[0031] S012. Construct a corresponding set of generation parameters for different abnormal modes. The set of generation parameters is used to characterize the morphological feature change parameters of the abnormal mode.
[0032] Furthermore, the generated parameters can be continuous variables, discrete variables, or multidimensional parameter vectors, and their value range is limited by the statistical prior of the real monitoring data to ensure that the generated abnormal samples are consistent with the real monitoring data in terms of overall statistical characteristics.
[0033] S013. Constrain the value range of the generated parameter set according to the statistical characteristic parameters to form a generated parameter distribution structure.
[0034] S014. Select the original monitoring value sequence of normal operation from the real monitoring data as the anomaly generation reference sequence, and construct the perturbation mapping relationship between the generation parameter and the original monitoring value sequence, so that the generation parameter acts on the original monitoring value sequence under the statistical distribution constraint.
[0035] Furthermore, the perturbation mapping relationship is used to establish a function mapping mechanism between the generated parameters and the original monitoring numerical sequence, including one or more combinations of amplitude scaling, trend superposition, periodic disturbance injection, abrupt pulse embedding, missing segment insertion, or frequency domain energy redistribution.
[0036] Through this perturbation mapping relationship, the generation parameters are applied to the original monitoring numerical sequence within the statistical distribution boundary, thereby generating an anomaly monitoring numerical sequence with the morphological characteristics of the target anomaly.
[0037] S02. Within the abnormal mode generation space, a synthetic monitoring data set is generated based on the generation parameter distribution structure and its disturbance mapping relationship. Any synthetic monitoring data sample includes a sequence of monitoring values arranged according to a predetermined sampling time order and a corresponding abnormal category identifier.
[0038] In this embodiment, step S02 involves instantiating and selecting generation parameters within the parameterized anomaly generation domain constrained by statistical priors constructed in step S01, and applying them to the original monitoring value sequence through the corresponding perturbation mapping relationship, thereby generating a synthetic monitoring data sample.
[0039] It should be noted that the original monitoring value sequence is a normal operation time sequence signal derived from real monitoring data, used as a reference sequence for anomaly generation, such as the acceleration time sequence signal collected by the bridge structure under normal load conditions, the stable stage sequence of road settlement monitoring data, and the steady change sequence in slope displacement monitoring, etc.
[0040] In this embodiment, the basic statistical distribution characteristics of the original monitoring numerical sequence are not changed during the generation process; only within the range allowed by statistical prior constraints are structured perturbations superimposed or applied.
[0041] In this embodiment, under any abnormal mode, the original monitoring data sequence is used to generate a synthetic monitoring data sample through the following steps: S021. Based on the distribution structure of the generation parameters corresponding to the abnormal mode, select the abnormal generation parameters within its limited value range as the parameter instances corresponding to this abnormal mode.
[0042] In some specific embodiments, the selection of anomaly generation parameters can be carried out by random sampling within the continuous value range of the generation parameters to obtain corresponding parameter instances; or by selecting from the discrete candidate set of generation parameters according to a preset interval or combination rule to form parameter instances of different intensity levels; or by combining and selecting in a multi-dimensional parameter space based on a preset coverage strategy to fully cover the value combinations of different parameter dimensions.
[0043] In these embodiments, the parameter instances obtained in the above manner can reflect abnormal characteristics of different magnitudes, different rates of change, or different locations of occurrence, thereby improving the diversity and coverage of abnormal sample generation.
[0044] S022. Based on the disturbance mapping relationship corresponding to the abnormal mode, perform a structured transformation on the original monitoring value sequence based on the parameter instance to generate an abnormal monitoring value sequence.
[0045] In some specific embodiments, the structured transformation may include linear superposition of the original monitoring numerical sequence to make the sequence as a whole show a continuous changing trend; or cumulative offset processing of the original monitoring numerical sequence to make the sequence drift gradually; or replacement or mutation processing of local time segments of the original monitoring numerical sequence to introduce outliers.
[0046] In these embodiments, the above-described structured transformation method enables the anomaly generation parameters to be directly mapped to the characteristic changes of monitoring data in the time or frequency domain, thereby generating an anomaly monitoring numerical sequence with clear anomaly characteristics.
[0047] S023. Associate the abnormal monitoring numerical sequence with the abnormal category identifier corresponding to the abnormal pattern to generate a synthetic monitoring data sample corresponding to the abnormal pattern.
[0048] In some specific embodiments, the association includes assigning a category identifier consistent with the corresponding abnormal pattern to the anomaly monitoring numerical sequence, so that the anomaly monitoring numerical sequence and its anomaly category form a structured sample unit, and each synthetic monitoring data sample simultaneously contains anomaly feature expression and corresponding semantic category information, thereby providing a sample basis with supervised attributes for subsequent cross-modal embedding training.
[0049] Furthermore, in order to generate the synthetic monitoring data set, step S02 also includes repeatedly executing steps S021 to S023 under different anomaly modes and different parameter combinations, and integrating and distributing the generated samples to form a structured sample system with sufficient coverage and balanced distribution in the anomaly category dimension and parameter intensity dimension of the synthetic monitoring data set.
[0050] For example, when generating the synthetic monitoring data set, a predetermined number of samples can be set for different anomaly patterns. For instance, multiple sets of samples can be generated for trend anomalies, drift anomalies, and outlier anomalies to ensure the complete coverage of anomaly categories.
[0051] Furthermore, under the same anomalous pattern, the anomalous intensity level can be divided according to the different value ranges of the generated parameters. For example, the trend slope parameter can be divided into low, medium and high amplitude ranges, and anomalous samples of corresponding intensity levels can be generated respectively; or the mutation amplitude parameter can be divided into different amplitude levels to generate outlier anomalous samples of different significance, so that the anomalous samples form a hierarchical distribution in terms of anomalous significance.
[0052] As an example, the number of samples of different anomaly categories or different intensity levels can be adjusted according to a preset distribution ratio. For example, the number of samples of each anomaly category can be kept basically balanced, or a sample distribution structure of mild and severe anomalies can be constructed according to a preset ratio to improve the representativeness of the sample set and the training stability.
[0053] The synthetic monitoring dataset constructed in the above manner not only achieves complete coverage in terms of anomaly categories, but also forms a continuous distribution in terms of anomaly significance, thereby enhancing the generalization ability and recognition stability of subsequent model training.
[0054] S03. Based on the synthetic monitoring data set, perform cross-modal embedding space alignment training on the visual features of the monitoring data and the semantic features of the corresponding anomaly categories, so that samples of the same category form a clustered distribution structure in the embedding space, and generate an anomaly recognition embedding model.
[0055] In this embodiment, the generated anomaly identification embedding model is a dual-encoding mapping model that maps the visual features of monitoring data and the semantic features of anomaly categories to the same embedding space. It includes a visual feature encoding module and a semantic feature encoding module.
[0056] Furthermore, the visual feature encoding module is used to perform visual vectorization representation of the monitored numerical sequence, and the semantic feature encoding module is used to perform semantic vectorization representation of the anomaly category identifier.
[0057] Furthermore, the anomaly detection embedding model is generated through the following steps: S031. Construct a visual feature encoding module and perform feature encoding on the monitoring numerical sequence of the synthetic monitoring data sample to obtain a visual feature vector representing the abnormal morphological features.
[0058] For example, the visual feature encoding module may include an encoding network for extracting features from time-series data, such as a time-series feature extraction network based on a convolutional structure, a sequence encoding network based on a recurrent structure, or a feature encoding network based on an attention mechanism, for converting the monitored numerical sequence into a fixed-dimensional visual feature vector representation.
[0059] S032. Construct a semantic feature encoding module and perform semantic vectorization representation on the anomaly category identifier of the synthetic monitoring data sample to obtain the corresponding semantic feature vector.
[0060] For example, the semantic feature encoding module may include a semantic encoding network for embedding representations of anomaly category identifiers, such as a category encoding module based on an embedding matrix, or a semantic vector generation module based on a text semantic representation model, for mapping anomaly categories to semantic feature vectors of a uniform dimension.
[0061] S033. Map the visual feature vector and the semantic feature vector to a unified embedding space, and optimize the unified embedding space by comparing the objective function of minimizing the distance between the same category and maximizing the distance between different categories, so that the embedding distance between samples of the same category tends to be minimized and the embedding distance between samples of different categories tends to be maximized, thereby forming multiple feature clustering structures centered on the semantic representation vector of the abnormal category in the unified embedding space.
[0062] In this embodiment, the unified embedding space refers to a vector space in which visual feature vectors and semantic feature vectors are mapped to the same dimension through an embedding mapping network.
[0063] Furthermore, in the unified embedding space, the similarity between different samples can be calculated and compared using methods such as Euclidean distance, cosine similarity, or vector inner product.
[0064] To ensure the comparability of the two types of modal features in the embedding space, the embedding mapping network may include structures such as fully connected mapping layers, nonlinear activation function layers, and normalization layers to achieve feature scale uniformity and distribution alignment.
[0065] In this implementation, after completing the unified embedding mapping, the embedding space structure is constrained and optimized by introducing a comparative optimization objective function that minimizes the distance between the same category and maximizes the distance between different categories, so that the different anomaly categories in the embedding space exhibit a separable distribution pattern.
[0066] In some specific implementations, positive sample pairs and negative sample pairs can be constructed for training.
[0067] Specifically, positive sample pairs are visual feature vectors and their corresponding semantic feature vectors under the same anomaly category; negative sample pairs are combinations of visual feature vectors and semantic feature vectors between different anomaly categories.
[0068] In these embodiments, by optimizing the training, the embedding distance between positive sample pairs is gradually reduced, while the embedding distance between negative sample pairs is gradually increased, thereby forming a class distribution structure with interval boundaries in the embedding space.
[0069] Through the above mechanism, different anomaly categories form structured and separable clustered regions in the embedding space, providing a stable basis for subsequent anomaly identification based on similarity calculation.
[0070] For example, for a trend anomaly sample, the visual encoding module extracts features from its monitored numerical sequence to obtain the corresponding visual feature vector; further, the visual feature vector is mapped to a unified embedding space; at the same time, the semantic feature vector corresponding to the trend anomaly category is also mapped to the unified embedding space.
[0071] Under the constraint of the comparative optimization objective function, the visual feature vectors of trend anomaly samples gradually converge toward their corresponding trend anomaly semantic center vectors, thereby forming a compact and concentrated cluster structure of trend anomaly samples in the embedding space.
[0072] Under the above training mechanism, the distance between the visual feature vector and its corresponding category semantic center vector continues to decrease, while the distance between it and the semantic vectors of other categories remains relatively large, thereby enhancing the structural stability and distinguishability of the category in the embedding space.
[0073] Similarly, the same embedding mapping and alignment training process is performed for both drift anomaly samples and outlier anomaly samples. Through the same-class distance convergence mechanism, the visual feature vectors of drift anomaly samples are clustered towards the drift anomaly semantic center vector; and the visual feature vectors of outlier anomaly samples are clustered towards the outlier anomaly semantic center vector.
[0074] When multiple categories participate in training, due to the distance constraint mechanism of negative sample pairs, a large embedding interval is maintained between the visual feature vectors of different anomaly categories and the non-corresponding semantic vectors. Therefore, each category forms a separate anomaly category region in a unified embedding space, avoiding feature overlap between categories.
[0075] For example, in a unified embedding space, the following distribution structure can be formed: trend anomaly samples cluster around the first semantic center region; drift anomaly samples cluster around the second semantic center region; and outlier anomaly samples cluster around the third semantic center region. A predetermined embedding interval boundary is maintained between different semantic center regions.
[0076] With the above embedding structure, in the subsequent recognition stage, it is only necessary to calculate the similarity value between the visual feature vector of the sample to be identified and each semantic center vector. The corresponding anomaly category can be determined according to the maximum similarity or threshold judgment rule, thereby achieving a stable and separable anomaly classification result.
[0077] S04. Using the anomaly identification embedding model, extract the visual feature vector of the monitoring data sample to be identified, and based on the visual feature vector, identify the anomaly category corresponding to the monitoring data sample to be identified.
[0078] In this embodiment, the anomaly category semantic center vector refers to the semantic representation vector of each anomaly category in the unified embedding space after alignment training is completed, or the category semantic representation formed by training and convergence of the corresponding semantic feature vectors.
[0079] Further, step S04, which involves using the anomaly identification embedding model to extract the visual feature vector of the monitoring data sample to be identified, and identifying the anomaly category corresponding to the monitoring data sample to be identified based on the visual feature vector, specifically includes the following steps: S041. Input the monitoring data sample to be identified into the visual feature encoding module of the anomaly identification embedding model to obtain the visual feature vector corresponding to the monitoring data sample to be identified.
[0080] In this embodiment, the monitoring data sample to be identified is a sequence of monitoring values arranged according to a predetermined sampling time sequence, and it does not carry an anomaly category identifier.
[0081] Furthermore, the visual feature encoding module encodes the monitored numerical sequence into a fixed-dimensional vector representation to characterize the abnormal morphological features of the sample (such as overall trend, local mutation, slow drift, periodic perturbation, etc.) and ensures that the vector is adapted to a unified embedding space.
[0082] For example, when the sample to be identified is a time series of bridge vibration monitoring data, the sequence is input into the visual feature encoding module, which can output the corresponding visual feature vector V. This vector also contains a comprehensive representation of the signal morphology, such as visual information like "overall slow rise", "local spikes", and "frequency domain energy concentration changes".
[0083] S042. Calculate the similarity value between the visual feature vector and the semantic representation vector of each anomaly category to obtain the similarity distribution of the monitoring data sample to be identified.
[0084] In this embodiment, the anomaly category semantic center vector is used to represent the category semantic anchor point of each anomaly category in the unified embedding space. It can be obtained by semantically vectorizing the anomaly category identifier by the semantic feature encoding module, or by aggregating the semantic vectors or aligned semantic vectors of samples of the same category during the training phase.
[0085] Furthermore, the similarity value is used to quantify the degree of matching between the visual features of the monitoring data sample to be identified and the semantic representation of each category; specifically, the higher the similarity, the more the sample matches the semantic features of the corresponding anomaly category.
[0086] The similarity calculation method can be cosine similarity, vector inner product, or distance-based similarity mapping, etc. It should be noted that the similarity calculation method should be consistent with or compatible with the feature measurement method used in the embedding space alignment training to ensure that the category clustering structure formed during the training phase can be effectively maintained during the recognition phase.
[0087] For example, if the anomaly categories include trend anomalies, drift anomalies, and outlier anomalies, the corresponding anomaly category semantic center vectors are the trend anomaly semantic center vector, the drift anomaly semantic center vector, and the outlier anomaly semantic center vector, respectively.
[0088] For the monitoring data sample to be identified, after the visual feature vector is extracted by the visual feature encoding module, the similarity value between the visual feature vector and the semantic center vector of the trend anomaly, the similarity value between the visual feature vector and the semantic center vector of the drift anomaly, and the similarity value between the visual feature vector and the semantic center vector of the outlier anomaly are calculated respectively, thereby obtaining a similarity set containing multiple similarity values. Each similarity value in the set is used to characterize the degree of matching between the monitoring data sample to be identified and the corresponding anomaly category.
[0089] S043. Based on the similarity distribution, the anomaly category is determined according to the following rules: when the similarity between the visual feature vector and one or more anomaly category semantic representation vectors meets the preset determination conditions, the corresponding anomaly category is determined as one or more anomaly categories corresponding to the monitoring data sample to be identified.
[0090] In this embodiment, the preset judgment condition is used to limit whether the degree of matching between the visual feature vector and the semantic center vector of the anomaly category meets the anomaly judgment standard.
[0091] Furthermore, the determination criteria may be that the similarity value is greater than a preset threshold, or the similarity value ranks among the top few in all categories and exceeds the confidence threshold, or the difference between the maximum similarity and the second largest similarity satisfies a preset interval condition.
[0092] When the similarity between the visual feature vector and the semantic representation vectors of multiple anomaly categories all meet the preset judgment conditions, multiple corresponding anomaly categories can be output simultaneously.
[0093] This multi-category similarity matching mechanism enables the monitoring data samples to be identified to converge simultaneously to multiple anomaly category semantic directions in the embedding space, thereby supporting the parallel identification of complex anomalies and avoiding the loss of anomaly information caused by forced classification of a single category.
[0094] For example, when a monitoring data sample to be identified simultaneously exhibits both an overall slow upward trend and local abrupt peak characteristics, and the similarity between its visual feature vector and the semantic representation vectors of "trend anomaly" and "outlier anomaly" both exceed a preset threshold, then the sample is simultaneously determined to belong to both trend anomaly and outlier anomaly.
[0095] In some embodiments, the semantic representation vector is a semantic center vector obtained by aggregating semantic feature vectors corresponding to the same anomaly category in the feature clustering structure.
[0096] Specifically, during training in the unified embedding space, multiple semantic feature vectors or multiple visual-semantic aligned samples may correspond to the same anomaly category. To enhance the stability and structural representativeness of the category representation, the semantic feature vectors under that category can be aggregated to obtain a semantic center vector that represents the overall semantic features of the anomaly category.
[0097] Furthermore, the aggregation operation may include mean operation, weighted mean operation, or weighted aggregation based on sample confidence. In some embodiments, the semantic feature vectors of the same category can be arithmetically averaged to obtain the semantic center vector of that category; in other embodiments, different weights can be assigned to different semantic feature vectors according to sample quality, training confidence, or embedding distance stability to form a weighted semantic center vector.
[0098] During training, the semantic center vector can be dynamically adjusted as batch data is updated. For example, in each training iteration, the corresponding semantic center vector is updated based on the distribution of semantic feature vectors and visual feature vectors belonging to the same anomaly category in the current batch, so that the semantic center vector is always located at the geometric center of the category embedding distribution, thereby improving the stability of the category clustering structure.
[0099] By constructing semantic center vectors, each anomaly category forms a clear category representative point in the embedding space. This can be used for contrast optimization constraints during the training phase, as well as as a reference benchmark for similarity calculation during the recognition phase, thereby enhancing the consistency of anomaly category representation.
[0100] In some embodiments, see Figure 2 , Figure 2 This is a schematic diagram of the functional modules of the highway infrastructure monitoring data anomaly identification device provided in an embodiment of the present invention.
[0101] Furthermore, the highway infrastructure monitoring data anomaly identification device proposed in this invention can be deployed in a highway infrastructure health monitoring server, an edge computing terminal, or a cloud analysis platform to implement the above-mentioned highway infrastructure monitoring data anomaly identification method.
[0102] like Figure 2 As shown, the highway infrastructure monitoring data anomaly identification device includes a statistical modeling module, a sample generation module, an embedding training module, and an anomaly identification module. The modules interact with each other through a data interface or bus.
[0103] Furthermore, the statistical modeling module is used to extract statistical feature parameters based on real monitoring data, construct an anomaly pattern generation space constrained by statistical distribution, and output the generation parameter distribution structure and disturbance mapping rules.
[0104] Furthermore, the sample generation module is connected to the statistical modeling module and is used to instantiate and select generation parameters according to the generation parameter distribution structure, and apply the perturbation mapping rule to the original monitoring value sequence to generate a synthetic monitoring data set.
[0105] Furthermore, the embedding training module is used to construct a visual feature encoding network and a semantic feature encoding network based on the synthetic monitoring data set, and to perform unified embedding space alignment training to output an anomaly recognition embedding model.
[0106] Furthermore, the anomaly identification module is used to receive monitoring data samples to be identified, extract visual feature vectors, and calculate the similarity between the vectors and semantic representation vectors of each anomaly category based on the anomaly identification embedding model, thereby determining the corresponding anomaly category.
[0107] The above modules work together to achieve a complete process of abnormal sample generation, embedding space construction, and abnormal category identification.
[0108] In one embodiment, a computer-readable storage medium is also provided, having a computer program stored thereon, which, when executed by a processor, causes the processor to perform the highway infrastructure monitoring data anomaly identification method as described in any of the foregoing embodiments.
[0109] It should be noted that the computer program may exist in the form of source code, object code, executable file or any other form, and may be loaded and executed by computing devices including general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices.
[0110] Furthermore, the computer-readable storage medium may be, but is not limited to, a read-only memory (ROM), a random access memory (RAM), a solid-state drive, flash memory, a disk, an optical disk, or other storage media capable of storing program code and readable by a computer device. It is used to store program instructions and is not limited to a specific hardware structure. As long as it can store and execute the program code of the method for identifying anomalies in highway infrastructure monitoring data, it falls within the protection scope of this invention.
[0111] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0112] It should be noted that the above embodiments can be freely combined as needed. The above are merely preferred embodiments of the present invention; it should be observed that those skilled in the art can make several improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for identifying anomalies in highway infrastructure monitoring data, characterized in that, Includes the following steps: Based on the statistical prior of real monitoring data, an anomaly pattern generation space is constructed. The anomaly pattern generation space represents the distribution structure of generation parameters of different anomaly patterns and their perturbation mapping relationship under the statistical distribution constraints of the real monitoring data. Within the anomaly pattern generation space, a synthetic monitoring data set is generated based on the generation parameter distribution structure and its perturbation mapping relationship. Each synthetic monitoring data sample includes a sequence of monitoring values arranged according to a predetermined sampling time order and a corresponding anomaly category identifier. Based on the synthetic monitoring data set, cross-modal embedding space alignment training is performed on the visual features of the monitoring data and the semantic features of the corresponding anomaly categories, so that samples of the same category form a clustered distribution structure in the embedding space, and an anomaly recognition embedding model is generated. Using the aforementioned anomaly identification embedding model, visual feature vectors of the monitoring data samples to be identified are extracted, and based on the visual feature vectors, the anomaly category corresponding to the monitoring data samples to be identified is determined.
2. The method for identifying anomalies in highway infrastructure monitoring data according to claim 1, characterized in that, The abnormal mode generation space is generated through the following steps: Statistical feature parameters characterizing the distribution characteristics of the data are extracted from real monitoring data, and these statistical feature parameters are used to limit the distribution boundary of the synthetic monitoring data. For different anomaly patterns, a corresponding set of generation parameters is constructed, which is used to characterize the morphological feature change parameters of the anomaly pattern; The value range of the generated parameter set is constrained based on the statistical feature parameters to form a generated parameter distribution structure; The original monitoring value sequence that is operating normally is selected from the real monitoring data as the reference sequence for anomaly generation, and a perturbation mapping relationship between the generation parameter and the original monitoring value sequence is constructed so that the generation parameter acts on the original monitoring value sequence under the statistical distribution constraint.
3. The method for identifying anomalies in highway infrastructure monitoring data according to claim 1, characterized in that, The statistical characteristic parameters are obtained by processing real monitoring data into sliding window segments and performing statistical analysis on each segment. They include one or more of the following: amplitude distribution range, mean range, variance range, extreme value range, frequency domain energy distribution characteristics, dominant frequency range, or trend change range.
4. The method for identifying anomalies in highway infrastructure monitoring data according to claim 1, characterized in that, Under any anomaly mode, the original monitoring numerical sequence is used to generate a synthetic monitoring data sample through the following steps: Based on the distribution structure of the generated parameters corresponding to the abnormal mode, the abnormal generated parameters are selected within its limited value range as parameter instances corresponding to this abnormal mode. Based on the disturbance mapping relationship corresponding to the abnormal mode, the original monitoring value sequence is structurally transformed based on the parameter instance to generate an abnormal monitoring value sequence. By associating the abnormal monitoring numerical sequence with the abnormal category identifier corresponding to the abnormal pattern, a synthetic monitoring data sample corresponding to the abnormal pattern is generated.
5. The method for identifying anomalies in highway infrastructure monitoring data according to claim 1, characterized in that, The generation of the anomaly detection embedding model includes the following steps: A visual feature encoding module is constructed, and the monitoring numerical sequence of the synthetic monitoring data sample is feature encoded to obtain a visual feature vector representing the abnormal morphological features. A semantic feature encoding module is constructed, and the anomaly category identifiers of the synthetic monitoring data samples are semantically vectorized to obtain the corresponding semantic feature vectors; The visual feature vector and the semantic feature vector are mapped to a unified embedding space. The unified embedding space is optimized and trained based on the objective function of minimizing the distance between samples of the same class and maximizing the distance between samples of different classes. This minimizes the embedding distance between samples of the same class and maximizes the embedding distance between samples of different classes, thereby forming multiple feature clustering structures centered on the semantic representation vector of the abnormal class in the unified embedding space.
6. The method for identifying anomalies in highway infrastructure monitoring data according to claim 5, characterized in that, Based on the visual feature vector, the anomaly category corresponding to the monitoring data sample to be identified is determined, including the following steps: Calculate the similarity value between the visual feature vector and the semantic representation vector of each anomaly category to obtain the similarity distribution corresponding to the monitoring data sample to be identified; Based on the similarity distribution, the anomaly category is determined according to the following rules: when the similarity between the visual feature vector and one or more anomaly category semantic representation vectors meets the preset determination conditions, the corresponding anomaly category is determined as one or more anomaly categories corresponding to the monitoring data sample to be identified.
7. The method for identifying anomalies in highway infrastructure monitoring data according to claim 6, characterized in that, The semantic representation vector is: In the feature clustering structure, the semantic center vector is obtained by aggregating the semantic feature vectors corresponding to the same anomaly category.
8. The method for identifying anomalies in highway infrastructure monitoring data according to claim 6, characterized in that, The preset judgment condition is that the similarity value between the visual feature vector and the corresponding anomaly category semantic center vector is greater than a preset threshold.
9. A device for identifying anomalies in highway infrastructure monitoring data, characterized in that, include: The statistical modeling module is used to construct an anomaly pattern generation space based on the statistical prior of real monitoring data. The anomaly pattern generation space represents the distribution structure of generation parameters of different anomaly patterns and their perturbation mapping relationship under the statistical distribution constraints of the real monitoring data. The sample generation module is used to generate a synthetic monitoring data set based on the generation parameter distribution structure and its perturbation mapping relationship within the abnormal pattern generation space. Any synthetic monitoring data sample includes a sequence of monitoring values arranged according to a predetermined sampling time order and a corresponding abnormal category identifier. The embedding training module is used to perform cross-modal embedding space alignment training on the visual features of the monitoring data and the semantic features of the corresponding anomaly category based on the synthetic monitoring data set, so that samples of the same category form a clustered distribution structure in the embedding space, and generate an anomaly recognition embedding model. An anomaly identification module is used to extract the visual feature vector of the monitoring data sample to be identified using the anomaly identification embedding model, and to identify the anomaly category corresponding to the monitoring data sample to be identified based on the visual feature vector.
10. A computer storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method for identifying anomalies in highway infrastructure monitoring data as described in any one of claims 1 to 8.