Urban drainage pipeline anomaly identification method based on shallow radar wave feedback model

Abstract

The application relates to the technical field of urban underground pipeline detection, and discloses a city drainage pipeline anomaly identification method based on a shallow radar wave feedback model, which comprises the following steps: based on real-time radar echo, the time-space joint entropy value of radar wave transmission parameters is dynamically adjusted, so that signal energy is adaptively focused on a high information entropy area; a prior propagation path is constructed by using pipeline material attenuation characteristics, and missing data is repaired through a priority guided hollow convolution; a cross-scale attention pyramid network is adopted to fuse multi-scale defect features, and a lightweight discrimination network is combined to generate online adversarial sample fine-tuning main models. Through the collaborative optimization of dynamic radar wave regulation and a physical propagation model, the compatibility detection capability of metal and non-metal pipelines is significantly improved; through a noise sensitivity feedback mechanism, environmental noise is converted into model optimization parameters, so that stable identification under complex working conditions is realized.

Description

Technical Field

[0001] This invention relates to a method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model, belonging to the field of non-destructive testing technology for urban underground pipelines. Background Technology

[0002] As a crucial component of underground pipe networks, the real-time monitoring of the structural health status of urban drainage pipelines is vital for urban safe operation and maintenance. Currently, the industry commonly employs shallow radar wave detection technology, which analyzes internal pipeline defects such as cracks and corrosion by emitting electromagnetic waves with specific parameters and receiving reflected signals. Typical methods usually employ fixed waveform parameter configurations combined with time-frequency analysis algorithms for static feature extraction. However, the following technical bottlenecks exist in practical applications:

[0003] 1. Insufficient adaptability to dynamic environments: Existing methods use preset radar pulse width and frequency modulation parameters, making it difficult to adjust signal strength in real time according to dynamic changes such as internal pipe humidity and sediment distribution. For example, signals near the metal pipe are susceptible to interference from multiple reflections, while signals at the far end of non-metallic pipes experience a sharp drop in signal-to-noise ratio due to material attenuation. The industry typically uses manual re-inspection or multiple scans for compensation, but this significantly reduces detection efficiency.

[0004] 2. Lack of physical correlation in data missing repair: For the problem of missing radar echo signals, existing technologies mostly use interpolation or neighborhood mean filling algorithms, without considering the propagation attenuation characteristics of radar waves in the pipe medium. For example, if conventional interpolation is used for missing signals at the near end of a metal pipe, false reflection peaks will be introduced, leading to misjudgment; while if the material attenuation coefficient is ignored when repairing missing signals at the far end of a non-metallic pipe, it is difficult to restore the true defect contour.

[0005] 3. The contradiction between multi-scale defect identification and noise suppression: Traditional methods rely on single-scale convolutional networks to extract features, which cannot simultaneously capture the high-frequency details of millimeter-scale cracks and the low-frequency structural features of centimeter-scale corrosion. The industry has attempted to process defects of different scales by cascading multiple detection modules, but the feature coupling between modules is insufficient and they are easily affected by water flow noise in the pipeline, leading to an increased false detection rate.

[0006] To address the aforementioned issues, some improvement schemes have attempted to introduce adaptive filtering or static noise library matching. However, these methods have limitations. For example, the noise suppression module and the defect identification model operate independently, making it difficult to achieve collaborative optimization under dynamic environments. Furthermore, the separation of the physical propagation model and the data-driven algorithm results in deviations between the repaired signal and the actual propagation path. Therefore, how to achieve dynamic control of radar wave parameters, data repair under physical propagation constraints, and collaborative optimization of multi-scale feature decoupling and noise suppression has become the technical problem to be solved by this invention. Summary of the Invention

[0007] This invention provides a method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model. Its main purpose is to solve the problems of insufficient adaptability to dynamic environments, physical inaccuracies in data repair, and incompatibility between multi-scale detection and noise suppression.

[0008] To achieve the above objectives, this invention provides a method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model, comprising the following steps:

[0009] Step S1: Based on the received real-time radar echo signal, calculate the joint entropy value in the time and spatial domains using a sliding window. The size of the sliding window is negatively correlated with the diameter of the pipe being detected, and the radar wave transmission parameters are dynamically adjusted according to the joint entropy value to enhance the signal strength in the high information entropy region and generate an adaptive radar wave signal.

[0010] Step S2: Construct a priori propagation path for missing radar echo data using a radar wave propagation physical model, and repair the missing data based on the priori propagation path using priority-guided dilated convolution. The repair priority is related to the attenuation coefficient of the radar wave in the pipe medium. Inversely proportional, among which Indicates the pipe material type;

[0011] Step S3: Extract multi-scale defect features from the repaired radar echo data using a cross-scale attention pyramid network, and dynamically adjust the weight distribution of features at different scales in the cross-scale attention pyramid network based on the prior probability of the defect type to achieve the fusion of multi-scale defect features.

[0012] Step S4: Construct a lightweight discriminant network to evaluate the credibility of the defect identification results in real time, and generate adversarial examples based on the evaluation results to fine-tune the parameters of the main defect identification model online, forming a dynamic optimization closed loop. The adversarial examples are generated for regions in the current detection with a confidence level lower than a preset threshold.

[0013] In a preferred embodiment, in step S1, the radar wave transmission parameters include pulse width. and frequency modulation slope And the dynamic adjustment is: when the joint entropy value At that time, the pulse width is increased or the frequency modulation slope is adjusted to enhance the signal energy; when the joint entropy value At this time, the pulse width is reduced or the frequency modulation slope is adjusted to reduce the energy output, where The first preset threshold, This is the second preset threshold.

[0014] In a preferred embodiment, in step S2, the attenuation coefficient of the radar wave in the pipe medium... According to the material type of the pipe The process is dynamically determined: for metal pipes, priority is given to repairing missing signal points near the end; for non-metal pipes, priority is given to repairing missing signal points far from the end.

[0015] In a preferred embodiment, in step S3, the defect types include cracks and corrosion, and a pipe-defect type mapping table is pre-established to record the defect characteristic frequency bands corresponding to different pipes. The prior probability of the defect type is statistically updated based on the occurrence frequency of different types of defects in historical detection data.

[0016] In a preferred embodiment, in step S4, the generation of the adversarial example also simultaneously introduces a dynamic noise feature mapping mechanism, including the following sub-steps:

[0017] Step S4.1: Extract the noise region signal not covered by defect features in the radar echo data in real time, and construct a noise baseline template based on its spectral distribution characteristics;

[0018] Step S4.2: By comparing the offset of the current noise zone signal with the noise baseline template, a noise sensitivity weight matrix is ​​generated. The noise sensitivity weight matrix is ​​used to quantify the interference intensity of noise in different frequency bands on defect identification.

[0019] Step S4.3: Correlate the noise sensitivity weight matrix with the gradient perturbation direction of the adversarial example in reverse, dynamically adjust the generation path of the adversarial example, and make the model fine-tuning prioritize suppressing the influence of high interference noise.

[0020] In a preferred embodiment, in step S4.1, defect feature mask filtering technology is used to automatically mask the identified defect areas in the time-frequency graph in order to extract pure noise area signals. The size of the defect feature mask is dynamically adjusted based on the sliding window size.

[0021] In a preferred embodiment, in step S4.2, the frequency band energy proportion threshold is used. The noise zone is determined by marking the corresponding area as a noise zone when the energy proportion of the non-defect frequency band is higher than a preset ratio. The preset ratio is automatically adjusted according to the frequency domain distribution of defect features in historical detection data.

[0022] In a preferred embodiment, a coupling attenuation factor is introduced during the generation of the noise sensitivity weight matrix. The coupling attenuation factor is calculated using the following formula:

[0023] ,

[0024] in, Indicates the energy of the noise frequency band. Indicates the energy of the defect characteristic frequency band. The time-domain correlation coefficient is represented, and the energy of the defect characteristic frequency band is determined based on the dynamic loading of the pipe-defect type mapping table.

[0025] In a preferred embodiment, after step S4, the method further includes storing historical detection data, and the parameters of the lightweight discriminant network and the defect identification master model are iteratively updated based on real-time generated adversarial examples and the historical detection data.

[0026] In a preferred embodiment, the method is applied to an urban drainage pipeline system comprising cast iron pipes and / or PVC pipes.

[0027] Compared to the problems described in the background art, the beneficial effects of the present invention are:

[0028] 1. By using spatiotemporal joint entropy values ​​to perceive the information density distribution of radar echo signals in real time, the energy allocation of radar wave transmission parameters is dynamically adjusted to prioritize signal energy focusing on highly complex areas (such as crack edges or structural deformation). Simultaneously, a priori propagation path model is constructed based on the attenuation characteristics of the pipe material. This synergistic mechanism of dynamic energy focusing and physical path constraints overcomes the signal mismatch problem of traditional fixed waveform parameters in complex pipe environments. Furthermore, the priority repair strategy effectively suppresses data loss and misjudgment caused by material differences, significantly improving the compatibility detection capability for both metallic and non-metallic pipes.

[0029] 2. Based on a cross-scale attention pyramid network, the fusion weights of features at different scales are dynamically allocated through the prior probability of defect type, enabling adaptive fusion of local high-frequency features of millimeter-scale cracks and global structural features of centimeter-scale corrosion. This mechanism overcomes the limitations of traditional single-scale detection, and combined with defect masking filtering technology for precise isolation of noise areas, it ensures the effectiveness and purity of the signal during the decoupling process of multi-scale features, thus maintaining high-precision defect localization and classification even under complex background noise.

[0030] 3. The baseline spectrum of the noise region is dynamically mapped to the control parameters generated by adversarial examples. By inversely correlating the noise sensitivity weight matrix with the model gradient perturbation direction, the real-time fine-tuning process prioritizes suppressing the impact of high-frequency interference bands on defect identification. This mechanism creatively transforms traditional discarded noise signals into an active driving force for model optimization. Combined with the online evaluation of a lightweight discriminant network, it forms a dynamic closed loop of noise perception, feature repair, and model iteration, effectively solving the problem of model performance degradation caused by the time-varying characteristics of environmental noise.

[0031] 4. Based on the differences in attenuation coefficients of pipe materials, the repair priorities for near-end and far-end signals are set differently. Simultaneously, path parameters from the radar wave propagation physical model are reused to construct a material-adaptive hole convolution repair strategy. This mechanism, through deep coupling of physical constraints and data-driven approaches, significantly improves the repair efficiency of rapidly reflected signals from the near end of metal pipes and attenuated signals from the far end of non-metallic pipes without requiring additional detection modules, achieving dual optimization of computational resources and detection accuracy.

[0032] 5. Actively masking the identified region using a time-frequency domain defect mask, extracting clean noise signals to construct a dynamic baseline template, and converting its offset into a constraint condition for generating adversarial examples. This mechanism forces the model to continuously enhance its sensitivity to effective defect frequency bands during fine-tuning through a spectral game between noise features and defect features, while suppressing interference from irrelevant noise, thereby maintaining stable detection robustness in dynamic environments (such as water flow disturbances during the rainy season). Attached Figure Description

[0033] Figure 1 This is a framework diagram of the urban drainage pipeline anomaly identification system based on a shallow radar wave feedback model according to the present invention.

[0034] Figure 2 This is a flowchart illustrating the anomaly identification of urban drainage pipelines based on a shallow radar wave feedback model, as presented in this invention.

[0035] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0036] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0037] This application provides a method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model, including the following steps:

[0038] Step S1: Based on the received real-time radar echo signal, calculate the joint entropy value in the time and spatial domains using a sliding window. The size of the sliding window is negatively correlated with the diameter of the pipe being detected, and the radar wave transmission parameters are dynamically adjusted according to the joint entropy value to enhance the signal strength in the high information entropy region and generate an adaptive radar wave signal.

[0039] Step S2: Construct a priori propagation path for missing radar echo data using a radar wave propagation physical model, and repair the missing data based on the priori propagation path using priority-guided dilated convolution. The repair priority is related to the attenuation coefficient of the radar wave in the pipe medium. Inversely proportional, among which Indicates the pipe material type;

[0040] Step S3: Extract multi-scale defect features from the repaired radar echo data using a cross-scale attention pyramid network, and dynamically adjust the weight distribution of features at different scales in the cross-scale attention pyramid network based on the prior probability of the defect type to achieve the fusion of multi-scale defect features.

[0041] Step S4: Construct a lightweight discriminant network to evaluate the credibility of the defect identification results in real time, and generate adversarial examples based on the evaluation results to fine-tune the parameters of the main defect identification model online, forming a dynamic optimization closed loop. The adversarial examples are generated for regions in the current detection with a confidence level lower than a preset threshold.

[0042] In a preferred embodiment, in step S1, the radar wave transmission parameters include pulse width. and frequency modulation slope And the dynamic adjustment is: when the joint entropy value At that time, the pulse width is increased or the frequency modulation slope is adjusted to enhance the signal energy; when the joint entropy value At this time, the pulse width is reduced or the frequency modulation slope is adjusted to reduce the energy output, where The first preset threshold, This is the second preset threshold.

[0043] In a preferred embodiment, in step S2, the attenuation coefficient of the radar wave in the pipe medium... According to the material type of the pipe The process is dynamically determined: for metal pipes, priority is given to repairing missing signal points near the end; for non-metal pipes, priority is given to repairing missing signal points far from the end.

[0044] In a preferred embodiment, in step S3, the defect types include cracks and corrosion, and a pipe-defect type mapping table is pre-established to record the defect characteristic frequency bands corresponding to different pipes. The prior probability of the defect type is statistically updated based on the occurrence frequency of different types of defects in historical detection data.

[0045] In a preferred embodiment, in step S4, the generation of the adversarial example also simultaneously introduces a dynamic noise feature mapping mechanism, including the following sub-steps:

[0046] Step S4.1: Extract the noise region signal not covered by defect features in the radar echo data in real time, and construct a noise baseline template based on its spectral distribution characteristics;

[0047] Step S4.2: By comparing the offset of the current noise zone signal with the noise baseline template, a noise sensitivity weight matrix is ​​generated. The noise sensitivity weight matrix is ​​used to quantify the interference intensity of noise in different frequency bands on defect identification.

[0048] Step S4.3: Correlate the noise sensitivity weight matrix with the gradient perturbation direction of the adversarial example in reverse, dynamically adjust the generation path of the adversarial example, and make the model fine-tuning prioritize suppressing the influence of high interference noise.

[0049] In a preferred embodiment, in step S4.1, defect feature mask filtering technology is used to automatically mask the identified defect areas in the time-frequency graph in order to extract pure noise area signals. The size of the defect feature mask is dynamically adjusted based on the sliding window size.

[0050] In a preferred embodiment, in step S4.2, the frequency band energy proportion threshold is used. The noise zone is determined by marking the corresponding area as a noise zone when the energy proportion of the non-defect frequency band is higher than a preset ratio. The preset ratio is automatically adjusted according to the frequency domain distribution of defect features in historical detection data.

[0051] In a preferred embodiment, a coupling attenuation factor is introduced during the generation of the noise sensitivity weight matrix. The coupling attenuation factor is calculated using the following formula:

[0052] ,

[0053] in, Indicates the energy of the noise frequency band. Indicates the energy of the defect characteristic frequency band. The time-domain correlation coefficient is represented, and the energy of the defect characteristic frequency band is determined based on the dynamic loading of the pipe-defect type mapping table.

[0054] In a preferred embodiment, after step S4, the method further includes storing historical detection data, and the parameters of the lightweight discriminant network and the defect identification master model are iteratively updated based on real-time generated adversarial examples and the historical detection data.

[0055] In a preferred embodiment, the method is applied to an urban drainage pipeline system comprising cast iron pipes and / or PVC pipes.

[0056] Example 1: In this example, the radar wave transmission parameters include pulse width. and frequency modulation slope According to the description of step S1, we calculate the spatiotemporal joint entropy value. This is used to sense the complexity distribution of radar signals, thereby adjusting these transmission parameters to ensure that radar wave energy is focused on areas of higher complexity (such as cracks and deformed regions). In practice, the joint entropy value is calculated as follows:

[0057] ,

[0058] in, It is the probability distribution of the observed values. This represents the characteristics of radar signals in the time and spatial domains. According to... The value can be dynamically adjusted during signal transmission. and The specific adjustment rules are as follows: when Greater than the preset threshold At that time, increase the pulse width Or adjust the frequency modulation slope To enhance signal energy; conversely, when Less than At that time, reduce the pulse width Or adjust the frequency modulation slope This reduces the signal output power. This adjustment mechanism ensures that the radar waves can adaptively focus on the area to be detected, thereby improving signal quality.

[0059] In step S2, the propagation physics model of radar waves can be used as a basis, combined with the attenuation coefficient of the pipe material. The missing radar echo data is repaired. Specifically, for metal pipes, priority is given to repairing signal loss at the near end; while for non-metal pipes, priority is given to repairing signal loss at the far end. During data repair, the attenuation coefficient is considered. The calculation method is as follows:

[0060] ,

[0061] in, It is the attenuation coefficient for metal or non-metal pipes. This refers to the signal loss coefficient caused by environmental factors (such as humidity and temperature) within the pipeline. By combining physical models with data-driven approaches, the propagation path of missing data can be accurately predicted and repaired, thus avoiding the spurious reflection peak problem commonly found in traditional methods.

[0062] Step S3 proposes a cross-scale attention pyramid network for extracting defect features at multiple scales. In this embodiment, defect types include, for example, cracks and corrosion, and the prior probability of each defect type is dynamically updated based on historical detection data. During this process, we optimize the feature fusion weights at each scale. For example, for millimeter-level cracks, high-frequency features are extracted first, while for centimeter-level corrosion, low-frequency features are extracted first. The weights after feature fusion are dynamically adjusted based on the frequency of different defect types in historical data, thereby improving detection accuracy.

[0063] To avoid noise interference, defect feature masking techniques from the time-frequency image can be used to shield identified areas. Signal extraction from noisy regions is performed as follows: First, noise regions not covered by defect features are extracted from the time-frequency image, and a noise baseline template is constructed based on the spectral characteristics of these regions. Then, by comparing the offset of the current noise region signal with the noise baseline template, a noise sensitivity weight matrix is ​​generated to quantify the interference intensity of different frequency band noise on defect identification. By inversely correlating the noise sensitivity weight matrix with the gradient perturbation direction of the generated adversarial examples, the model can dynamically adjust to suppress the impact of high-frequency interference noise on defect identification.

[0064] In step S4, the process of generating adversarial examples involves fine-tuning based on feedback from historical detection data. Specifically, the main defect identification model can be iteratively optimized based on the real-time generated adversarial examples, forming a dynamic optimization loop. When a low-confidence region is detected, the system generates adversarial examples and incorporates them into the model fine-tuning. The generation of adversarial examples not only incorporates a noise sensitivity feedback mechanism but also optimizes parameters based on historical detection data, thereby improving the robustness of the model and ensuring that it maintains efficient and stable identification performance even under complex operating conditions (such as water flow disturbances or weather changes).

[0065] Example 2: This example combines Figure 1 and Figure 2 This paper details the overall architecture and specific process of an anomaly identification method for urban drainage pipelines based on a shallow radar wave feedback model. This technical solution achieves efficient identification and accurate location of anomalies in urban drainage pipelines through radar wave emission control, defect identification and repair, dynamic optimization, and cross-scale feature fusion. (See also...) Figure 1 The system mainly consists of four modules: radar wave emission control module, defect identification module, dynamic optimization module, and pipeline material database module. Radar wave emission control module: This module calculates the spatiotemporal joint entropy value in real time (…). The system analyzes the received radar echo signals and dynamically adjusts the radar wave transmission parameters based on the analysis results. Specifically, when the joint entropy value exceeds a preset threshold (i.e., ...), ... The system increases the pulse width to improve signal resolution; when the joint entropy value is less than a preset threshold (i.e., The system reduces pulse width to optimize signal energy distribution. The defect identification module uses a multi-scale convolutional neural network (CNN) to extract defect features and evaluates the reliability of the defect identification results in real time through a lightweight discriminant network. By generating adversarial examples, the system can fine-tune the identification model online, thereby improving the accuracy and robustness of defect identification. The dynamic optimization module is responsible for optimizing the fusion of multi-scale features by adjusting the weights of feature fusion. By combining the prior probabilities of different types of defects in historical data, the system can dynamically adjust the fusion strategy of features at different scales, further improving the ability to identify complex defects. The pipe material database module stores attenuation characteristic data of pipe materials, including different attenuation coefficients for metal and non-metal pipes. Based on this data, the system can accurately repair missing radar echo data and apply different repair strategies to metal and non-metal pipes, such as prioritizing the repair of near-end signals for metal pipes and prioritizing the repair of far-end signals for non-metal pipes. Through the collaborative work of these modules, the system can detect and repair pipe anomalies in real time, thereby improving the detection accuracy and stability of urban drainage pipes.

[0066] Figure 2 This paper demonstrates a process for anomaly identification in urban drainage pipelines based on a shallow radar wave feedback model. The process, from radar wave emission to final detection result output, encompasses multiple steps including signal transmission, joint entropy value determination, radar wave propagation path construction, pipeline material identification, and repair strategy selection. Radar wave emission: The system initiates the detection process through the radar wave emission module, emitting specific radar signals to collect reflected waves within the pipeline. Joint entropy value determination: The received radar echo signals are used to calculate the joint entropy value (…). If the joint entropy value is greater than a preset threshold, the system decides to increase the pulse width to enhance signal resolution; if the joint entropy value is less than the preset threshold, the system reduces the pulse width to optimize signal energy distribution. Constructing a transmission path model: Based on the constructed radar wave propagation path model, the system repairs missing radar echo data. Repair strategies are selected based on the pipe material: for metal pipes, near-end signals are repaired first; for non-metal pipes, far-end signals are repaired first. Pipe material judgment: The system selects different repair schemes based on the pipe material. When the pipe is determined to be metal, a near-end priority repair strategy is adopted; when the pipe is determined to be non-metallic, a far-end priority repair strategy is adopted. Span transmission extraction: After repair, the system further optimizes the repair results through span transmission extraction technology to ensure data integrity and accuracy. Final detection result output: After completing all repairs and optimizations, the system finally outputs the detection results, providing the location and type analysis of pipe anomalies. This process ensures that every step from radar wave transmission to defect identification is executed accurately, making full use of spatiotemporal joint entropy values, pipe material data, cross-scale feature fusion and other technical means to ultimately achieve efficient and stable anomaly identification of drainage pipes.

[0067] Example 3: This example aims to further illustrate the dynamic adjustment of radar wave parameters, the calculation of attenuation coefficient and its application in data repair, as well as the generation process of noise sensitivity weight matrix.

[0068] In step S1, the radar wave transmission parameters, i.e., pulse width... and frequency modulation slope It is based on the spatiotemporal joint entropy value Adjustments have been made. (Formula) middle, It is the probability distribution of the observed values. This represents the characteristics of radar signals in the time and spatial domains. More specifically, It can be a feature vector obtained through time-frequency analysis of radar echo signals, which contains frequency components and energy intensities at multiple time points and spatial locations. Calculation First, the radar echo signal is segmented into a series of overlapping spatiotemporal windows. For each window, features of the radar signal are extracted to form a feature vector. Then, by counting the frequency of each feature vector appearing in all windows, its probability distribution is estimated. .when Greater than the preset threshold When the signal is complex and contains a lot of information, it indicates that the signal is complex and the pulse width is high. In this case, increasing the pulse width is appropriate. It can improve the penetration capability of radar waves or adjust the frequency modulation slope. This expands the frequency range of radar waves, thereby enhancing signal energy and providing more comprehensive coverage of the target area. Conversely, when... Less than the preset threshold At that time, reduce the pulse width Or adjust the frequency modulation slope This reduces signal output power, energy consumption, and noise interference. This dynamic adjustment mechanism allows the radar wave transmission parameters to match the complex environment inside the pipeline, optimizing detection performance.

[0069] In step S2, the attenuation coefficient of the radar wave in the pipe medium It depends on the type of pipe material. It is dynamically determined. The calculation method is as follows: ,in, It is the attenuation coefficient of metal or non-metal pipes, which characterizes the energy loss rate of radar waves propagating in a specific material. It is the signal loss coefficient caused by environmental factors inside the pipeline, including humidity, temperature, and sediment. The attenuation of radar waves can be determined by experimentally measuring the degree of attenuation of different pipe materials. Determining the value requires comprehensive consideration of various environmental factors inside the pipeline. These factors can be estimated through sensor measurements or empirical models, and then substituted into the corresponding model for calculation. .

[0070] Attenuation coefficient during data repair This is used to guide the priority of repairing dilated convolutions. For metal pipes, since their attenuation coefficient is usually small, radar waves can penetrate well. Therefore, priority is given to repairing missing points of the near-end signal to reduce interference caused by multiple reflections. For non-metallic pipes, their attenuation coefficient is large, and radar wave energy attenuates faster. Therefore, priority is given to repairing missing points of the far-end signal to restore the true outline of the far-end defect.

[0071] In step S4, a noise sensitivity weight matrix is ​​generated to suppress noise interference. Specifically, firstly, noise region signals not covered by defect features are extracted from the radar echo data, and a noise baseline template is constructed based on their spectral distribution characteristics. Then, the noise sensitivity weight matrix is ​​calculated by comparing the offset of the current noise region signal with the noise baseline template in each frequency band. Each element of this weight matrix represents the intensity of the noise energy in the corresponding frequency band relative to the energy in the defect feature frequency band, used to quantify the degree of interference of noise in different frequency bands on defect identification.

[0072] Calculate the coupling attenuation factor In this method, the factor is used to generate the noise sensitivity weight matrix: ,in, Indicates the energy of the noise frequency band. Indicates the energy of the defect characteristic frequency band. This represents the time-domain correlation coefficient. and The calculation is performed by performing Fourier transforms on the noise region signal and the defect feature region signal, and then calculating the energy of each frequency band. This reflects the similarity in the time domain between the signal from the noise region and the signal from the defect feature region, which can be obtained by calculating their cross-correlation function. By... Energy proportion threshold in the noise frequency band By comparing the results, noise regions can be identified, and a noise sensitivity weight matrix can be generated to guide the generation of adversarial examples. This allows the model to be fine-tuned to prioritize the suppression of the effects of high-interference noise. These are all extended implementation methods that are known to those skilled in the art.

[0073] Example 4: In this example, to enhance the adaptability of radar waves in complex pipeline environments, the spatiotemporal joint entropy value is specified. How to perceive the information density distribution of radar echo signals in real time and adjust transmission parameters accordingly? Specifically, the transmission parameters of radar waves mainly include pulse width. and frequency modulation slope And it is dynamically adjusted through the following steps, such as calculating the spatiotemporal joint entropy value. Its formula is: ,in, Represents the probability distribution of the observed values. To define the time and spatial characteristics of radar signals, based on the calculated joint entropy value, when... At that time, increase the pulse width Or adjust the frequency modulation slope To enhance signal resolution; when At that time, reduce the pulse width Or adjust the frequency modulation slope To optimize signal energy distribution, this mechanism dynamically adjusts the transmission parameters of radar waves, enabling them to prioritize focusing on areas with high information complexity (such as cracks and deformation areas), thereby improving signal quality and ensuring more accurate pipeline anomaly detection.

[0074] To address the issue of missing radar echo signals, this embodiment, based on the physical model of radar wave propagation, adjusts the attenuation coefficient. The calculation method is explained in detail. The attenuation coefficient depends on the pipe material type. Dynamically determined: For metal pipes, the attenuation coefficient For smaller attenuation coefficients, priority should be given to repairing near-end signal loss to avoid multiple reflections and interference; for non-metallic pipes, the attenuation coefficient... The radar wave energy attenuates rapidly due to its large size; therefore, priority is given to repairing the missing signal at the far end to restore the true outline of the far-end defect. The formula for calculating the attenuation coefficient is as follows:

[0075] ,

[0076] in, It is the attenuation coefficient of the pipe material. This refers to the signal loss coefficient caused by environmental factors (such as humidity and temperature) within the pipeline. For metal pipelines, Smaller, non-metallic pipes The relatively large value reflects the difference in the impact of different materials on radar wave propagation. This data restoration mechanism can accurately repair missing radar echo data, avoiding the false reflection peak problem common in traditional methods and improving the reliability of detection results.

[0077] To address the challenges of multi-scale defect identification and noise interference, this embodiment details how to extract defect features from radar echo data using a cross-scale attention pyramid network. By dynamically adjusting the fusion weights of features at different scales, it can simultaneously capture high-frequency features of millimeter-scale cracks and low-frequency features of centimeter-scale corrosion. After data repair, the system extracts features using the cross-scale attention pyramid network and dynamically adjusts the weight allocation of features at each scale based on the prior probability of the defect type, thereby achieving the fusion of multi-scale defect features. Simultaneously, a defect feature masking filtering technique is used to shield identified areas, ensuring the extraction and isolation of noise regions. Specifically, the signal of the noise region not covered by defect features is extracted to construct a noise baseline template; the offset between the current noise region signal and the noise baseline template is compared to generate a noise sensitivity weight matrix, which is used to quantify the interference intensity of different frequency band noise on defect identification. This noise sensitivity feedback mechanism can dynamically adjust the path for generating adversarial examples in real time, allowing the model to prioritize suppressing the impact of high-frequency interference noise on defect identification during fine-tuning, thereby improving detection robustness.

[0078] In the defect identification process, this embodiment further optimizes the real-time evaluation function of the lightweight discriminant network. Through the generation of adversarial examples and real-time feedback, the robustness of the model can be effectively enhanced. When a low-confidence region is detected, adversarial examples are generated and introduced into the model for fine-tuning, optimizing the identification results. The generation of adversarial examples is based on the current detection confidence and historical detection data, enabling dynamic optimization of model parameters and forming a closed loop of noise perception-feature repair-model iteration. This solves the problem of the time-varying characteristics of environmental noise affecting model performance, ensuring that the system can still operate stably and provide accurate detection results even under complex working conditions. These are all extended implementation methods known to those skilled in the art.

[0079] Example 5: This technical solution is applied, for example, to the structural health assessment of a cast iron pipe under a main urban road that has been in service for many years. The pipe has silt inside, and the pipe wall is wet due to rainfall and accompanied by continuous water flow. This condition poses a challenge to the stable propagation of radar wave signals and the accurate identification of defects, especially the identification of millimeter-level circumferential cracks caused by corrosion.

[0080] Before the detection task begins, an offline parameter calibration procedure is performed to determine the key threshold required for judging the joint entropy value. A standard pipe section of the same material and diameter with known cracks and corrosion defects is selected. Under dry and non-silting conditions, a scan is performed with a fixed pulse width and frequency modulation slope to collect a reference radar echo dataset. For this dataset, the echo signals corresponding to the known defect structure and the echo signals corresponding to the smooth and intact pipe wall are clearly identified, and their spatiotemporal joint entropy values ​​are calculated respectively. The spatiotemporal feature vector here This involves extracting the energy value sequence of each frequency component after applying a short-time Fourier transform to the radar echo signal within each sliding window, and then constructing its probability distribution using a kernel density estimation method. The result is calculated by taking the echo signals corresponding to all known defects. The 10th percentile of the statistical distribution of the value is set as the first preset threshold. ; Calculated by taking the echo signals of all corresponding intact pipe walls The 90th percentile of the statistical distribution of the value is set as the second preset threshold. .

[0081] Upon entering the on-site testing area, the system uses real-time calculations... Value and the already calibrated , Dynamically adjust transmission parameters. Attenuation coefficient is used when constructing a priori propagation paths to repair missing data. The determination is made by retrieving the material attenuation coefficient from the pipe-defect type mapping table. Based on this, the real-time temperature and humidity data inside the pipeline collected by field sensors are used as input, and a predetermined signal loss coefficient value is output through a pre-set two-dimensional lookup table. and combined and Calculated Given that the current detection object is a cast iron pipe, the system determines it to be a metal pipe. Therefore, when performing priority-guided void convolution, priority is given to repairing signal loss points near the end caused by silt or water flow disturbance.

[0082] In the closed loop of defect identification and model fine-tuning, when the lightweight discriminant network gives a confidence level lower than a preset threshold for the crack identification result in a certain area, the adversarial example generation process is triggered. The confidence threshold is set at the 25th percentile of the confidence scores of all identified defects in historical detection data. At this point, the system extracts clean noise signals from the time-frequency map based on the defect feature mask, and constructs a noise baseline template by calculating the moving average of the energy of these signals in each frequency band. The adjustment procedure for the frequency band energy proportion threshold is as follows: the system periodically statistically analyzes the energy distribution of the characteristic frequency bands corresponding to all confirmed corrosion and crack defects in historical data, and sets 5% of the total energy of this distribution as the frequency band energy proportion threshold. When the energy proportion of non-defect frequency bands is higher than this dynamic threshold, the region is marked as a strong interference noise region. Finally, the noise sensitivity weight matrix generated by the coupling attenuation factor is used to back-correlate the gradient perturbation direction of the adversarial examples. When generating adversarial examples for fine-tuning the main model, a larger perturbation is applied to the noise frequency bands that are judged to be strong interference. This mechanism transforms environmental noise into a driving force for optimizing model parameters, improves the model's sensitivity to specific defect frequency bands, and ensures that the identification of corrosion cracks remains stable even in humid conditions with water flow.

[0083] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model, characterized in that, Includes the following steps: Step S1: Based on the received real-time radar echo signal, calculate the joint entropy value in the time and spatial domains using a sliding window. The size of the sliding window is negatively correlated with the diameter of the pipe being detected, and the radar wave transmission parameters are dynamically adjusted according to the joint entropy value to enhance the signal strength in the high information entropy region and generate an adaptive radar wave signal. Step S2: Construct a priori propagation path for missing radar echo data using a radar wave propagation physical model, and repair the missing data based on the priori propagation path using priority-guided dilated convolution. The repair priority is related to the attenuation coefficient of the radar wave in the pipe medium. Inversely proportional, among which Indicates the pipe material type; Step S3: Extract multi-scale defect features from the repaired radar echo data using a cross-scale attention pyramid network, and dynamically adjust the weight allocation of features at different scales in the cross-scale attention pyramid network based on the prior probability of the defect type to achieve the fusion of multi-scale defect features. Step S4: Construct a lightweight discriminant network to evaluate the credibility of the defect identification results in real time, and generate adversarial examples based on the evaluation results to fine-tune the parameters of the main defect identification model online, forming a dynamic optimization closed loop. The adversarial examples are generated for regions in the current detection with a confidence level lower than a preset threshold.

2. The method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model according to claim 1, characterized in that, In step S1, the radar wave transmission parameters include pulse width. and frequency modulation slope And the dynamic adjustment is: when the joint entropy value At that time, the pulse width is increased or the frequency modulation slope is adjusted to enhance the signal energy; when the joint entropy value At this time, the pulse width is reduced or the frequency modulation slope is adjusted to reduce the energy output, where The first preset threshold, This is the second preset threshold.

3. The method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model according to claim 1, characterized in that, In step S2, the attenuation coefficient of the radar wave in the pipe medium According to the material type of the pipe The process is dynamically determined: for metal pipes, priority is given to repairing missing signal points near the end; for non-metal pipes, priority is given to repairing missing signal points far from the end.

4. The method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model according to claim 1, characterized in that, In step S3, the defect types include cracks and corrosion, and a pipe-defect type mapping table is pre-established to record the defect feature frequency bands corresponding to different pipes. The prior probability of the defect type is statistically updated based on the occurrence frequency of different types of defects in historical detection data.

5. The method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model according to claim 4, characterized in that, In step S4, the generation of the adversarial example also simultaneously introduces a dynamic noise feature mapping mechanism, including the following sub-steps: Step S4.1: Extract the noise region signal not covered by defect features in the radar echo data in real time, and construct a noise baseline template based on its spectral distribution characteristics; Step S4.2: By comparing the offset of the current noise zone signal with the noise baseline template, a noise sensitivity weight matrix is ​​generated. The noise sensitivity weight matrix is ​​used to quantify the interference intensity of noise in different frequency bands on defect identification. Step S4.3: Correlate the noise sensitivity weight matrix with the gradient perturbation direction of the adversarial example in reverse, dynamically adjust the generation path of the adversarial example, and make the model fine-tuning prioritize suppressing the influence of high interference noise.

6. The method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model according to claim 5, characterized in that, In step S4.1, defect feature mask filtering technology is used to automatically mask the identified defect areas in the time-frequency graph in order to extract pure noise area signals. The size of the defect feature mask is dynamically adjusted based on the sliding window size.

7. The method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model according to claim 5, characterized in that, In step S4.2, the frequency band energy ratio threshold is used. The noise zone is determined by marking the corresponding area as a noise zone when the energy proportion of the non-defect frequency band is higher than a preset ratio. The preset ratio is automatically adjusted according to the frequency domain distribution of defect features in historical detection data.

8. The method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model according to claim 5, characterized in that, A coupling attenuation factor is introduced during the generation of the noise sensitivity weight matrix. The coupling attenuation factor is calculated using the following formula: ， in, Indicates the energy of the noise frequency band. Indicates the energy of the defect characteristic frequency band. The time-domain correlation coefficient is represented, and the energy of the defect characteristic frequency band is determined based on the dynamic loading of the pipe-defect type mapping table.

9. The method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model according to claim 1, characterized in that, After step S4, the method further includes storing historical detection data, and the parameters of the lightweight discriminant network and the defect identification master model are iteratively updated based on the real-time generated adversarial examples and the historical detection data.

10. The method for identifying anomalies in urban drainage pipelines based on a shallow radar wave feedback model according to claim 1, characterized in that, The method is applied to urban drainage pipe systems that include cast iron pipes and / or PVC pipes.

Images