A radar robot tunnel lining structure anomaly positioning method and system
By collecting radar echo sequences of tunnel lining structures using radar robots, constructing time-domain and frequency-domain feature channels, and generating anomaly confidence scores using lightweight networks, the problem of false alarms and missed alarms caused by electromagnetic interference in tunnel lining structure detection was solved, achieving high robustness and real-time anomaly localization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 中铁长江交通设计集团有限公司
- Filing Date
- 2026-04-16
- Publication Date
- 2026-06-05
AI Technical Summary
Existing radar detection methods are susceptible to electromagnetic interference in tunnel lining structure detection, resulting in low echo signal-to-noise ratio and unstable waveforms. Traditional methods are unable to fully capture multi-dimensional anomaly information, have poor adaptability, and produce false alarms and missed alarms.
A radar robot was used to collect radar echo sequences. Through moving average filtering and linear normalization preprocessing, a time-domain convolutional feature channel and a frequency-domain statistical threshold channel were constructed. A lightweight one-dimensional convolutional network was used to extract local morphological features, and anomaly confidence was generated through weighted fusion. Adjacent anomaly points were merged to form anomaly segments.
It significantly improves the robustness and interpretability of anomaly identification under complex interference, meets the real-time processing requirements in mobile inspection scenarios, and provides a reliable basis for safety assessment of tunnel lining structures.
Smart Images
Figure CN122151238A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of tunnel engineering inspection technology, and in particular to a radar robot method and system for locating anomalies in tunnel lining structures. Background Technology
[0002] Tunnel lining structure is crucial for ensuring safe tunnel operation, and early and accurate detection of hidden defects such as cavities and cracks is essential. Currently, tunnel lining detection methods based on vehicle-mounted or mobile ground-penetrating radar are gradually being applied. These methods transmit electromagnetic waves and receive echoes, using structural information carried in the echo signals to identify anomalies. However, in practical applications, tunnels are subject to multi-source electromagnetic interference from numerous electrical devices, vehicle operation, and surrounding rock scattering, resulting in low radar echo signal-to-noise ratios and unstable waveforms. Traditional detection methods based on single thresholds or single classification models are prone to false alarms and missed alarms. Existing methods often rely on limited features such as time-domain waveform amplitude and energy, or use fixed thresholds in the frequency domain, making it difficult to comprehensively capture multi-dimensional information such as local waveform distortion and spectral distribution changes caused by anomalies. They are particularly unsuitable for situations involving material differences and changes in coupling conditions. Therefore, how to construct a lining anomaly localization method that is both highly robust and interpretable, and adaptable to real-time end-side calculations in complex tunnel detection environments, has become a pressing technical challenge in this field. Summary of the Invention
[0003] The purpose of this invention is to provide a radar robot method and system for locating anomalies in tunnel lining structures, which significantly improves the robustness and interpretability of anomaly identification under complex interference, while meeting the real-time processing requirements in mobile inspection scenarios, and providing a reliable basis for safety assessment and maintenance decisions of tunnel lining structures.
[0004] To achieve the above objectives, in a first aspect, the present invention provides a radar robot method for locating anomalies in tunnel lining structures, comprising the following steps:
[0005] Along the pre-set detection points on the tunnel lining, the radar robot is controlled to collect radar echo sequences point by point;
[0006] The radar echo sequence is preprocessed, including denoising and normalization;
[0007] Based on the preprocessed echo sequence, anomalous evidence is generated for the temporal convolutional feature channel and the frequency domain statistical threshold channel, respectively.
[0008] By fusing anomalous evidence from both channels, the anomalous confidence level of the detection point is obtained;
[0009] Anomaly detection points are located based on anomaly confidence levels, and adjacent anomaly points are merged into anomaly segments.
[0010] The denoising process employs a moving average filter, and the normalization process uses linear normalization to map the echo amplitude to the [0,1] interval.
[0011] The temporal convolutional feature channel extracts local morphological features of the echo sequence through a lightweight one-dimensional convolutional network and outputs the initial anomaly probability.
[0012] The frequency domain statistical threshold channel includes the following steps:
[0013] Perform a discrete Fourier transform on the same echo sequence to obtain the spectral amplitude;
[0014] Frequency domain thresholds are set based on the statistical distribution of normal sample spectral amplitudes.
[0015] Calculate the proportion of frequency points exceeding the threshold and generate a frequency domain anomaly evidence score.
[0016] Specifically, the abnormal evidence from both channels is fused using a weighted fusion method:
[0017] .
[0018] Where λ is the weighting coefficient. This represents the probability of time-domain channel anomalies. The score is the evidence score for the frequency domain channel normalization.
[0019] Among them, the method of locating anomaly detection points based on anomaly confidence and merging adjacent anomaly points into anomaly segments includes:
[0020] Set an anomaly confidence threshold;
[0021] Detection points with a confidence level exceeding the threshold are identified as outliers;
[0022] Consecutive or adjacent outliers are merged into outlier segments, and a minimum number of points constraint is applied to suppress isolated noise.
[0023] The lightweight one-dimensional convolutional network consists of multiple basic modules connected in series. Each module includes an input layer, a feature extraction layer, and an output layer, and the modules satisfy the input-output dimension matching.
[0024] The lightweight network aims to minimize the number of parameters in its design of convolutional kernel size, number, and layer structure. The total number of parameters in the model is expressed as follows:
[0025] ;
[0026] in, For the first Layer convolution kernel length or channel-related scaling factor, This represents the number of convolutional kernels / channels in the corresponding layer.
[0027] In a second aspect, the present invention provides a radar robot tunnel lining structure anomaly localization system, which is applied to a radar robot tunnel lining structure anomaly localization method as provided in the first aspect. The radar robot tunnel lining structure anomaly localization system includes a data acquisition module, a preprocessing module, a dual-channel evidence generation module, a fusion module, and a localization output module.
[0028] The data acquisition module is used to control the radar robot to collect radar echo sequences along the tunnel lining;
[0029] The preprocessing module is used to denoise and normalize the echo sequence;
[0030] The dual-channel evidence generation module is used to extract temporal convolutional features and frequency domain statistical anomaly evidence, respectively.
[0031] The fusion module is used to fuse dual-channel evidence and calculate anomaly confidence.
[0032] The positioning output module is used to output the location of abnormal points and information about abnormal sections.
[0033] The radar robot tunnel lining structure anomaly localization system is embedded in the radar robot, which supports closed-loop real-time reasoning from data acquisition to anomaly localization and outputs anomaly segment information bound to mileage / circumferential position.
[0034] This invention discloses a radar robot method and system for anomaly localization in tunnel lining structures. The method involves a radar robot collecting radar echo sequences from various detection points on the tunnel lining, followed by moving average denoising and linear normalization preprocessing. A temporal convolutional feature channel and a frequency domain statistical threshold channel are constructed to extract local morphological features and generate anomaly evidence based on the statistical distribution of normal samples. The dual-channel evidence is weighted and fused to obtain the anomaly confidence level of each detection point. Anomalies are determined based on the confidence threshold, adjacent anomalies are merged to form anomaly segments, and the corresponding mileage or circumferential location information is output. The system employs a lightweight network structure, controlling the number of parameters and computational complexity to achieve real-time inference and localization output on the radar robot end. This invention significantly improves the robustness and interpretability of anomaly identification under complex interference through a dual-channel evidence complementarity mechanism, while meeting the real-time processing requirements of mobile inspection scenarios, providing a reliable basis for safety assessment and maintenance decisions of tunnel lining structures. Attached Figure Description
[0035] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.
[0036] Figure 1This is a schematic diagram of the steps of a radar robot method for locating anomalies in tunnel lining structures according to the first embodiment of the present invention.
[0037] Figure 2 This is a flowchart illustrating a radar robot method for locating anomalies in tunnel lining structures, as provided by this invention.
[0038] Figure 3 This is a schematic diagram of a radar robot tunnel lining structure anomaly positioning system according to the second embodiment of the present invention.
[0039] In the diagram: 101 - Data acquisition module, 102 - Preprocessing module, 103 - Dual-channel evidence generation module, 104 - Fusion module, 105 - Positioning output module. Detailed Implementation
[0040] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.
[0041] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms "a," "the," and "the" as used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.
[0042] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."
[0043] The first embodiment of this application is as follows:
[0044] Please see Figures 1-2 This invention provides a method for locating anomalies in tunnel lining structures using a radar robot, comprising the following steps:
[0045] S1. Along the pre-set detection points of the tunnel lining, control the radar robot to collect radar echo sequences point by point.
[0046] Specifically, the radar robot, equipped with ground-penetrating radar, moves within the tunnel along a pre-set inspection path, performing point-by-point inspections on the tunnel lining at various predetermined detection points. At each detection point, the radar emits electromagnetic wave signals. These signals penetrate the tunnel lining and are reflected by different media. The radar receives the reflected signals, forming raw radar echo data. These different media include normal lining structures as well as abnormal areas containing voids or cracks.
[0047] During a tunnel inspection, assuming a total of [number] samples were collected... Data from the first detection point, the first The original radar echoes from each detection point can be represented as a one-dimensional discrete sequence:
[0048] (1)
[0049] in, This refers to the number of sampling points for a single 1-point echo signal. For the first The detection point at the first The echo amplitude values corresponding to each sampling time. Therefore, the raw radar data for the entire lining section can be represented as a set. It is used for subsequent preprocessing and anomaly localization.
[0050] S2. Preprocess the radar echo sequence, including denoising and normalization.
[0051] Specifically, raw radar echoes are often noisy due to factors such as tunnel electrical equipment, vehicle electromagnetic interference, and geological scattering. To improve the robustness of subsequent feature extraction and anomaly localization, this invention performs denoising and normalization processing on the echo sequence of each detection point sequentially.
[0052] The noise reduction process employs a moving average filter, specifically:
[0053] For a certain detection point raw sequence Perform moving average filtering to obtain the denoised sequence. :
[0054] (2)
[0055] in, For The set of indices for the central sliding window, with the window length denoted as . (Generally, an odd number between 3 and 7 is chosen, which can be adjusted according to the actual noise level.) This represents the number of samples within the window. To ensure valid calculations at the boundaries, In Always satisfied This means that the window is truncated at the beginning and end of the sequence.
[0056] Moving average filtering suppresses random noise through local smoothing while preserving the main morphological features of radar echoes as much as possible, providing input for subsequent convolutional feature extraction.
[0057] The normalization process uses linear normalization to map the echo amplitude to the [0,1] interval. Specifically, to eliminate the differences in the dimensions and amplitude range of the echo amplitude at different detection points, the denoised sequence is... Perform linear normalization to obtain :
[0058] (3)
[0059] in, , After normalization .when At that time, it can be made Or add a very small amount Avoid having a denominator of zero.
[0060] Normalization unifies the data scale, improves the comparability of data from different detection points, and facilitates the stable learning of lining anomaly-related features by lightweight networks.
[0061] S3. Based on the preprocessed echo sequence, generate anomalous evidence for the temporal convolutional feature channel and the frequency domain statistical threshold channel, respectively.
[0062] Specifically, the temporal convolutional feature channel extracts local morphological features of the echo sequence through a lightweight one-dimensional convolutional network and outputs an initial anomaly probability, as follows:
[0063] Will Input the feature extraction layer of the lightweight network and perform one-dimensional convolution to obtain local features:
[0064] (4)
[0065] in, The local morphological features extracted at the r-th position of the n-th detection point. The value of the normalized echo sequence at the nth detection point at the (r+s-1)th position; denoted by , where is the weight parameter corresponding to the s-th position of the convolution kernel; c is the kernel length; s is the kernel internal index; r is the convolution output position index; m is the number of sampling points for the echo signal; and m-c+1 is the length of the output feature sequence after convolution. Local morphological features refer to the waveform characterization information of the radar echo sequence reflecting lining anomalies within a local time window, including at least local amplitude abrupt change features, abnormal reflection enhancement features, waveform width variation features, local undulation / slope variation features, and local energy distribution variation features. These features are used to characterize the differences in echo morphology between abnormal areas such as voids and cracks and normal lining.
[0066] After local morphological features are extracted through multiple layers / modules of a lightweight one-dimensional convolutional network, the output layer integrates the extracted features to form a scalar result representing the degree of anomaly at the detection point. This result is then mapped to the [0,1] interval as the temporal channel anomaly probability of the detection point. The larger this probability value, the greater the degree to which the local morphology of the echo at the detection point deviates from the normal sample.
[0067] (5)
[0068] in, This represents the probability of time-domain channel anomalies.
[0069] This quantity reflects the degree of abnormality in the local morphology of the echo (abrupt changes, enhanced reflection, attenuation changes, etc.).
[0070] The processing procedure for frequency domain statistical threshold channels is as follows:
[0071] For the same Perform a discrete Fourier transform:
[0072] (6)
[0073] in, The signal obtained by the discrete Fourier transform of the echo sequence corresponding to the nth detection point is represented in the frequency domain at frequency f; where f is the frequency variable. Fourier transform operator; The input sequence for the nth detection point; The spectral amplitude of the nth detection point at frequency f; Represents the spectrum of complex numbers The modulus represents the magnitude of the amplitude.
[0074] Percentage-based frequency domain anomaly score:
[0075] (7)
[0076] in, The percentage-based frequency domain anomaly score for the nth detection point; represents the set of frequency points participating in the statistics; II(.) is the indicator function, which takes the value 1 when the condition in parentheses is true, and 0 otherwise. Let T(f) be the spectral amplitude of the nth detection point at frequency f; T(f) is the anomaly detection threshold at frequency f. Then, normalize it to a frequency domain channel normalized evidence score. It is used for integration with the initial network assessment.
[0077] S4. Fusion of abnormal evidence from the two channels to obtain the anomaly confidence level of the detection point.
[0078] Specifically, the initial judgment from the convolutional network is fused with frequency domain evidence to obtain the final anomaly confidence level. .
[0079] Weighted fusion:
[0080] (8)
[0081] Where λ is the weighting coefficient. This represents the probability of time-domain channel anomalies. The frequency domain channel normalized evidence score refers to the score obtained by performing frequency domain analysis on the echo sequence at the same detection point, establishing anomaly judgment thresholds based on the statistical distribution of normal samples at each frequency point, calculating the proportion of frequency points exceeding the threshold, and then normalizing this score to a numerical range consistent with the anomaly probability of the time domain channel for fusion. The weighting coefficient λ represents the relative contribution of the time domain channel and the frequency domain channel to the final anomaly confidence score. It can be predetermined based on the recognition performance on the training / validation sample set, or preset according to the on-site false alarm rate and false negative rate control requirements.
[0082] S5 locates anomaly detection points based on anomaly confidence and merges adjacent anomaly points into anomaly segments.
[0083] Specifically, set a judgment threshold. ,when:
[0084] (9)
[0085] Then determine the first Each detection point corresponds to an abnormal lining location, and its location index (detection point number / mileage / circumferential direction) is output. The abnormality judgment threshold can be preset based on the statistical results of normal samples and labeled abnormal samples to take into account both false alarm rate and false negative rate. The frequency domain abnormality judgment threshold can be set according to the statistical distribution of normal samples at each frequency point. In the implementation method, the mean and standard deviation can be used to construct the frequency point threshold.
[0086] Merge adjacent or consecutive outliers into anomaly segments:
[0087] (10)
[0088] in, Let q be the set of detection points corresponding to the qth abnormal segment, where q is the abnormal segment number; This is the starting detection point number for the q-th abnormal segment; This is the termination detection point number for the q-th abnormal segment.
[0089] It can also incorporate a minimum number of points constraint to suppress isolated noise. This minimum number of points constraint is used to suppress isolated noise points and can be set based on the spacing between detection points, the minimum engineering scale of abnormal sections, and the distribution of continuous points in the verification data.
[0090] (12)
[0091] Finally, the start and end detection points (or corresponding mileage / circumferential range) of each anomaly area are output, completing the engineering representation of the anomaly area location results.
[0092] To adapt to application scenarios where radar robots have limited computing resources and require real-time output of detection results, this invention constructs and embeds a lightweight network model on the robot's end. This lightweight network model adopts a modular network architecture composed of multiple basic modules, each of which includes an input layer, a feature extraction layer, and an output layer.
[0093] Input layer: Receives radar echo sequences (which can be single-detection-point sequences or sliding window sequences) collected and preprocessed by the radar robot.
[0094] Feature extraction layer: Extracts local features related to lining anomalies through convolution operations.
[0095] Output layer: Integrates features and outputs the feature representation or preliminary judgment result of this module.
[0096] Multiple basic modules are connected according to the following rules to form an overall network:
[0097] (13)
[0098] Where t is the basic module number; Indicates the first The output of each basic module Indicates the first Mapping functions for each module; It is the output of the (t-1)th basic module, and also serves as the input of the tth basic module; This represents the total number of basic modules in the overall network.
[0099] It also satisfies the condition that "the output of the previous module is directly used as the input of the next module", and that the output dimension matches the input dimension of the next module.
[0100] In the selection of convolution kernel size, number, and neuron scale of each layer, the goal is to minimize the number of parameters. At the same time, experiments are conducted to verify that the recognition performance of lining anomalies does not significantly decrease when the number of parameters is reduced, thereby strictly controlling the parameter scale and reducing the computational burden.
[0101] The total number of parameters in the model can be expressed as:
[0102] (14)
[0103] in, Here, L represents the total number of parameters in the model, and L represents the total number of network layers involved in the statistics. For the first Layer convolution kernel length or channel-related scaling factor, By employing the aforementioned lightweight strategy to adjust the number of convolutional kernels / channels for the corresponding layer, the model is able to achieve closed-loop inference of "acquisition—preprocessing—feature extraction—anomaly localization" at the radar robot end.
[0104] This invention proposes a "dual-channel evidence" anomaly determination mechanism based on radar robot lining inspection data: For the radar echo sequence of the same detection point after denoising and normalization, a time-domain convolutional feature channel and a frequency-domain statistical threshold channel are constructed. The former uses a lightweight network to extract local morphological features of the echo through one-dimensional convolution and outputs an initial anomaly probability. The latter performs Fourier transform on the same source signal and establishes a frequency-domain threshold based on the 3σ principle of normal samples to generate anomaly evidence scores. Subsequently, the two complementary evidences are weighted and fused or gated to obtain the final anomaly confidence score, which is used to locate anomaly points and merge adjacent areas. Its significance lies in the fact that, under conditions of fluctuating radar coupling conditions, complex on-site noise, and differences in lining materials that make a single model or single threshold method prone to false alarms / missed alarms, the evidence complementarity of "learnable discrimination + statistical verification" significantly improves the robustness and interpretability of anomaly identification, while simultaneously meeting the real-time inference and online output requirements of the radar robot, allowing the location results to be directly mapped to mileage / segment for review and maintenance decisions.
[0105] The second embodiment of this application is as follows:
[0106] Please see Figure 3 The present invention provides a radar robot tunnel lining structure anomaly localization system, which is applied to a radar robot tunnel lining structure anomaly localization method as provided in the first embodiment. The radar robot tunnel lining structure anomaly localization system includes a data acquisition module 101, a preprocessing module 102, a dual-channel evidence generation module 103, a fusion module 104 and a localization output module 105.
[0107] The data acquisition module 101 is used to control the radar robot to collect radar echo sequences along the tunnel lining;
[0108] The preprocessing module 102 is used to denoise and normalize the echo sequence.
[0109] The dual-channel evidence generation module 103 is used to extract temporal convolutional features and frequency domain statistical anomaly evidence, respectively.
[0110] The fusion module 104 is used to fuse dual-channel evidence and calculate the anomaly confidence level;
[0111] The positioning output module 105 is used to output the location of abnormal points and information about abnormal sections.
[0112] The radar robot tunnel lining structure anomaly localization system is embedded in the radar robot end, supporting closed-loop real-time reasoning from data acquisition to anomaly localization, and outputting anomaly segment information bound to mileage / circumferential position.
[0113] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.
[0114] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.
[0115] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.
Claims
1. A method for locating anomalies in tunnel lining structures using a radar robot, characterized in that, Includes the following steps: Along the pre-set detection points on the tunnel lining, the radar robot is controlled to collect radar echo sequences point by point; The radar echo sequence is preprocessed, including denoising and normalization; Based on the preprocessed echo sequence, anomalous evidence is generated for the temporal convolutional feature channel and the frequency domain statistical threshold channel, respectively. By fusing anomalous evidence from both channels, the anomalous confidence level of the detection point is obtained; Anomaly detection points are located based on anomaly confidence levels, and adjacent anomaly points are merged into anomaly segments. The temporal convolutional feature channel extracts local morphological features of the echo sequence through a lightweight one-dimensional convolutional network and outputs the initial anomaly probability. The frequency domain statistical threshold channel includes the following steps: Perform a discrete Fourier transform on the same echo sequence to obtain the spectral amplitude; Frequency domain thresholds are set based on the statistical distribution of normal sample spectral amplitudes. Calculate the proportion of frequency points exceeding the threshold and generate a frequency domain anomaly evidence score; The denoising process uses a moving average filter, and the normalization process uses linear normalization to map the echo amplitude to the [0,1] interval; The anomalous evidence from both channels is fused using a weighted fusion method: ； Where λ is the weighting coefficient. This represents the probability of time-domain channel anomalies. The score is the evidence score for the frequency domain channel normalization.
2. The radar robot method for locating anomalies in tunnel lining structures as described in claim 1, characterized in that, Anomaly detection points are located based on anomaly confidence levels, and adjacent anomaly points are merged into anomaly segments, including: Set an anomaly confidence threshold; Detection points with confidence levels exceeding a threshold are identified as outliers; Consecutive or adjacent outliers are merged into outlier segments, and a minimum number of points constraint is applied to suppress isolated noise.
3. The radar robot method for locating anomalies in tunnel lining structures as described in claim 1, characterized in that, The lightweight one-dimensional convolutional network consists of multiple basic modules connected in series. Each module includes an input layer, a feature extraction layer, and an output layer, and the modules satisfy the input-output dimension matching.
4. The radar robot method for locating anomalies in tunnel lining structures as described in claim 3, characterized in that, The lightweight network aims to minimize the number of parameters in its design of convolutional kernel size, number of kernels, and layer structure. The total number of parameters in the model is expressed as follows: ； in, For the first Layer convolution kernel length or channel-related scaling factor, This represents the number of convolutional kernels / channels in the corresponding layer.
5. A radar robot tunnel lining structure anomaly localization system, applied to the radar robot tunnel lining structure anomaly localization method as described in claim 1, characterized in that, The radar robot tunnel lining structure anomaly localization system includes a data acquisition module, a preprocessing module, a dual-channel evidence generation module, a fusion module, and a localization output module. The data acquisition module is used to control the radar robot to collect radar echo sequences along the tunnel lining; The preprocessing module is used to denoise and normalize the echo sequence; The dual-channel evidence generation module is used to extract temporal convolutional features and frequency domain statistical anomaly evidence, respectively. The fusion module is used to fuse dual-channel evidence and calculate anomaly confidence. The positioning output module is used to output the location of abnormal points and information about abnormal sections.
6. The radar robot tunnel lining structure anomaly positioning system as described in claim 5, characterized in that, The radar robot tunnel lining structure anomaly localization system is embedded in the radar robot, supports closed-loop real-time reasoning from data acquisition to anomaly localization, and outputs anomaly segment information bound to mileage / circumferential position.