All-weather pipeline anomaly perception method and system based on multi-modal acoustic-vibration fusion

Abstract

The application discloses a kind of all-weather pipeline abnormality perception method and system based on multimodal acoustic vibration fusion, it is related to pipeline abnormality perception technical field. Including: along the pipeline to be monitored acoustic vibration acquisition unit, edge computing unit, cloud processing unit, edge computing unit and acoustic vibration acquisition unit are communicatively connected, cloud processing unit and edge computing unit are communicatively connected, obtain arrangement unit item.The application introduces physical constraint acoustic vibration depth fusion module and small sample lifelong learning dynamic adaptation module, constructs multi-level interference filtering mechanism, significantly improves the accuracy and reliability of pipeline anomaly detection, based on pre-stored acoustic vibration propagation physical parameter library, the acoustic vibration data collected in real time is strictly checked, by setting check threshold, the system automatically eliminates environmental noise that does not conform to the law of physical propagation, only real abnormal data is retained, so that false positive rate is greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of pipeline anomaly sensing technology, specifically to an all-weather pipeline anomaly sensing method and system based on multimodal acoustic-vibration fusion. Background Technology

[0002] Pipeline anomaly detection refers to a technical system that uses sensor monitoring and algorithm analysis to identify pipeline operational anomalies in real time or periodically for physical pipelines such as long-distance oil and gas pipelines, municipal water supply and drainage, heating pipe networks, and material transportation pipelines in chemical industrial parks. The monitoring dimensions typically cover characteristics such as pipeline pressure, flow fluctuations, leakage sound patterns, pipe wall vibration, temperature changes, corrosion rates, and pipe wall deformation. Commonly used technologies include distributed fiber optic sensing, IoT node acquisition, sonar detection, and AI anomaly recognition models. Since risks such as pipeline leaks, third-party construction damage, geological deformation damage, and freezing cracks occur randomly without clear time or scenario patterns, all-weather detection is required. This can significantly reduce the cost of manual inspections, especially for pipeline sections that are difficult to reach manually, such as those crossing rivers, buried underground, or in uninhabited areas, and can replace high-frequency manual inspections.

[0003] A method and system for intelligent sensing and anomaly analysis of the operating status of natural gas pipelines, with patent publication number CN121388913A, achieves multimodal feature compatibility and fusion while controlling noise robustness by processing energy entropy features, acoustic spatiotemporal features, gradient change rate features, and mutation coefficient features through a feature splicing module. It utilizes an LSTM attention mechanism to process multi-dimensional spliced ​​features, enabling multi-granularity sensing of complex pipeline anomaly patterns. This improves the detection accuracy of the dynamic fusion network for composite anomalies while maintaining computational efficiency. The system determines the level of anomaly risk based on anomaly probability values ​​and adopts corresponding solution strategies, achieving a classification of anomaly risks and solutions, significantly improving the efficiency of monitoring and subsequent operation and maintenance. Finally, a dual-engine mechanism combining location topology constraints and Bayesian inference is used to trace the source of anomalies at monitoring points, achieving efficient location of anomaly equipment in complex pipelines.

[0004] The aforementioned and similar technical solutions generally adopt a shallow fusion mode of feature splicing or result voting driven by pure data. They do not take advantage of the physical characteristics of the natural co-coupled sound and vibration signals in pipeline scenarios. For example, the sound and vibration signals of real pipeline anomalies such as leakage and deformation have strict physical correspondence in terms of time delay, energy and spectrum. They cannot separate weak abnormal signals such as micro-leakage and micro-deformation of pipe walls from strong environmental noise such as wind blowing vegetation, passing vehicles and rainstorm erosion, resulting in a high false alarm rate in strong interference scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide an all-weather pipeline anomaly sensing method and system based on multimodal acoustic-vibration fusion, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an all-weather pipeline anomaly sensing method based on multimodal acoustic-vibration fusion, comprising:

[0007] Along the pipeline to be monitored, sound and vibration acquisition units, edge computing units, and cloud processing units are deployed. The edge computing units are connected to the sound and vibration acquisition units, and the cloud processing units are connected to the edge computing units to obtain the deployment unit items.

[0008] A physical constraint acoustic-vibration deep fusion module is embedded in the edge computing unit and the cloud processing unit. The acoustic-vibration propagation physical parameter library of the pipeline to be monitored is pre-stored in the physical constraint acoustic-vibration deep fusion module to obtain the stored dataset.

[0009] The design parameter information of the pipeline to be monitored is obtained. Based on the matching results of the design parameter information in the stored dataset, the reference acoustic and vibration signal parameters are obtained, and then the benchmark data items are obtained.

[0010] The acoustic and vibration data of the monitored pipeline are acquired in real time by the acoustic and vibration acquisition unit to obtain the acquisition dataset. The pre-embedded physical verification items are called in the physical constraint acoustic and vibration deep fusion module to perform data verification and data removal based on the acquisition dataset, and to filter out interference signals that do not conform to the physical laws of acoustic and vibration propagation in the pipeline to obtain the filtered dataset.

[0011] Based on the small-sample lifelong learning dynamic adaptation module preset by the cloud processing unit, the cover feature set is obtained by calling the pre-built acoustic-vibration coupling feature base library of the module. Based on the cover feature set, the filter dataset is subjected to secondary feature matching and filtering to further filter out known scene interference signals and obtain the secondary filter dataset.

[0012] By using the secondary filtered dataset as a benchmark, pipeline anomaly detection is performed, and anomaly determination is completed, thereby realizing multi-level interference filtering and acoustic-vibration fusion-based pipeline anomaly detection.

[0013] Furthermore, the method for obtaining the arrangement unit item includes:

[0014] The acoustic vibration acquisition unit adopts a hybrid deployment of distributed optical fiber acoustic vibration sensing and point piezoelectric acoustic vibration sensor. The distributed optical fiber acoustic vibration sensing cable is laid in the same trench as the pipeline to be monitored, which is the first arrangement item.

[0015] Set scene trigger conditions, including construction frequency, river crossing, road crossing, and key buried sections, to obtain scene trigger items. Set a placement threshold based on the scene trigger items. The placement threshold is a fixed distance value. Place point piezoelectric acoustic vibration sensors based on the placement threshold to obtain the second placement item.

[0016] The first arrangement item and the second arrangement item are combined to obtain the arrangement unit item.

[0017] Furthermore, the method for obtaining the stored dataset includes:

[0018] The stored dataset contains thresholds for propagation delay difference, energy attenuation coefficient, and spectrum correspondence of acoustic and vibration signals for different anomaly types, including micro-leakage, third-party excavation, pipe wall deformation, and external force collision, for different pipe materials, pipe diameters, and transported media.

[0019] Based on the pipe material, pipe diameter, and transport medium, create a category storage table, and combine the category storage tables to obtain a category combination set;

[0020] Based on the data acquisition method, using the category combination set as a benchmark, the thresholds for propagation delay difference, energy attenuation coefficient, and spectrum correspondence of acoustic and vibration signals under the corresponding combination are obtained, thereby obtaining the stored dataset.

[0021] Furthermore, the method for obtaining the benchmark data item includes:

[0022] The design parameter information of the pipeline to be monitored, including pipeline material, pipe diameter and conveying medium, is used to obtain the design parameter items;

[0023] The design parameters are matched one by one with the parameters in the stored dataset, and the initial thresholds of the parameters in the corresponding scenarios are extracted to obtain general benchmarking items.

[0024] Set a collection threshold, which is a fixed time value. During the trial operation of the pipeline to be monitored, collect normal operating sound and vibration data to obtain pre-collected items. Set a fluctuation value, which is a fixed range value. Based on the fluctuation value, fine-tune the threshold of the general benchmarking items to adapt to the specific characteristics of the current pipeline and obtain the calibration benchmarking items.

[0025] The average combination of calibration benchmarks and general benchmarks is used as the benchmark data item.

[0026] Furthermore, the physical verification items include:

[0027] A verification threshold is set as a fixed percentage value. The verification threshold is combined with the benchmark data items to obtain a combined judgment item. When the judgment deviation exceeds the combined judgment item, it is judged as a non-pipeline-related interference signal, and then the physical verification item is obtained.

[0028] Furthermore, the method for obtaining the filtered dataset includes:

[0029] The acoustic and vibration acquisition unit continuously acquires the two raw acoustic and vibration signals of the pipeline under test, sets a time window value, and slices and encapsulates the data to obtain the acquisition dataset.

[0030] The acoustic and vibration signals in the acquired dataset are preprocessed by timestamp alignment and power frequency denoising to obtain the preprocessed dataset.

[0031] The physical verification item in the physical constraint acoustic-vibration deep fusion module is called to calculate the parameter values ​​of the actual propagation delay difference, energy ratio, and spectral correlation degree of the acoustic-vibration signal in each time window, and the parameter calculation item is obtained.

[0032] The calculated parameters are compared with the corresponding thresholds of the benchmark data items. If the deviation of any dimension parameter exceeds the physical verification item, it is determined to be environmental interference that does not conform to the physical propagation law. The data in that time window is then removed, and the remaining valid data is integrated to obtain the filtered dataset.

[0033] Furthermore, the method for obtaining the secondary filtered dataset includes:

[0034] The acoustic-vibration coupling feature base library, which is pre-built by the small-sample lifelong learning dynamic adaptation module, is called. The base library covers known scene interference features in different regions, working conditions and seasons, and a covered feature set is obtained.

[0035] Multi-dimensional coupled features in the time domain, frequency domain, and time-frequency domain are extracted from the acoustic and vibration signals of the filtered dataset to obtain the feature terms to be matched. The cosine similarity between the feature terms to be matched and the covered feature set is calculated. A similarity judgment threshold is set. When the calculation result exceeds the similarity judgment threshold, it is judged as known scene interference, and the corresponding data is removed to obtain the secondary filtered dataset.

[0036] Furthermore, the anomaly determination includes:

[0037] The edge computing unit performs a preliminary classification of the secondary filtered dataset. If it determines that the dataset is suspected of being abnormal, it uploads the corresponding original acoustic and vibration data to the cloud. Otherwise, it only uploads the compressed normal operation feature archive to obtain the initial screening judgment items.

[0038] The cloud processing unit calls the high-precision fusion model to make a secondary judgment on the suspected abnormal data, clarify the abnormality type, and calculate the location of the abnormal point by combining the sound and vibration propagation time delay difference to obtain the precise judgment result item;

[0039] The precise judgment results are compared with the network GIS system, and the AI ​​video surveillance around the anomaly point is automatically retrieved for verification. At the same time, the anomaly type, location and handling suggestions are pushed to the mobile terminal of the operation and maintenance personnel to form a closed loop of handling.

[0040] The all-weather pipeline anomaly detection system based on multimodal acoustic-vibration fusion utilizes the aforementioned all-weather pipeline anomaly detection method based on multimodal acoustic-vibration fusion, including:

[0041] Unit Layout Module: Along the pipeline to be monitored, sound and vibration acquisition units, edge computing units, and cloud processing units are laid out. The edge computing units are connected to the sound and vibration acquisition units, and the cloud processing units are connected to the edge computing units to obtain the layout unit items.

[0042] Data pre-storage module: Embed a physical constraint acoustic vibration deep fusion module in the edge computing unit and the cloud processing unit. The physical constraint acoustic vibration deep fusion module pre-stores the acoustic vibration propagation physical parameter library of the pipeline to be monitored, and obtains the stored dataset.

[0043] The data comparison module acquires the design parameter information of the pipeline to be monitored. Based on the matching results of the design parameter information in the stored dataset, it acquires the reference acoustic and vibration signal parameters, and then obtains the benchmark data items. Based on the acoustic and vibration acquisition unit, it acquires the acoustic and vibration data of the monitored pipeline in real time, and obtains the acquisition dataset. In the physical constraint acoustic and vibration deep fusion module, it calls the pre-embedded physical verification items, performs data verification and data removal based on the acquisition dataset, filters out interference signals that do not conform to the physical laws of acoustic and vibration propagation in the pipeline, and obtains the filtered dataset. Based on the small sample lifelong learning dynamic adaptation module preset by the cloud processing unit, it calls the acoustic and vibration coupling feature base library pre-built by the module to obtain the coverage feature set. Based on the coverage feature set, it performs secondary feature matching and filtering on the filtered dataset to further filter out interference signals of known scenarios, and obtains the secondary filtered dataset.

[0044] Anomaly detection module: Based on the secondary filtered dataset, pipeline anomaly detection is performed and anomaly detection is completed, thereby realizing multi-level interference filtering and sound and vibration fusion pipeline anomaly detection.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] This all-weather pipeline anomaly detection method and system based on multimodal acoustic-vibration fusion constructs a multi-level interference filtering mechanism by introducing a physical constraint acoustic-vibration deep fusion module and a small-sample lifelong learning dynamic adaptation module. This significantly improves the accuracy and reliability of pipeline anomaly detection. Based on a pre-stored acoustic-vibration propagation physical parameter library, the system rigorously verifies the real-time acquired acoustic-vibration data. By setting a verification threshold, the system automatically eliminates environmental noise that does not conform to the physical propagation laws. At the same time, using the acoustic-vibration coupling feature base library pre-built by the small-sample lifelong learning module, the system performs secondary feature matching on the filtered data. By extracting multi-dimensional coupling features in the time domain, frequency domain, and time-frequency domain, and based on cosine similarity calculation, the system accurately identifies and eliminates known interference signals, retaining only true anomaly data, thus significantly reducing the false alarm rate. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0048] Figure 2 This is a schematic diagram of the parameter library construction process of the present invention;

[0049] Figure 3 This is a schematic diagram of the parameter calibration process of the present invention;

[0050] Figure 4 This is a schematic diagram of the physical verification and filtering process of the present invention;

[0051] Figure 5 This is a schematic diagram of the feature matching and filtering process of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] With the increasing importance of pipeline transportation in energy, chemical and other fields, pipeline safety has received widespread attention. Acoustic-vibration fusion technology, as an emerging method for pipeline anomaly detection, combines the advantages of acoustic and vibration sensors, showing great potential in improving detection accuracy and reliability. However, existing acoustic-vibration fusion schemes generally adopt shallow fusion modes such as pure data-driven feature splicing or result voting, failing to fully exploit the natural homogeneous coupling of acoustic and vibration signals in pipeline scenarios. This leads to poor performance and a high false alarm rate in environments with strong interference. First, in abnormal situations such as pipeline leakage and deformation, acoustic waves and vibration signals are not independent but have a strict physical correspondence. For example, the acoustic waves generated by a leak will cause vibration of the pipeline wall, and the intensity, frequency, and other characteristics of the vibration signal are closely related to parameters such as the degree and location of the leak. Furthermore, acoustic waves and vibration signals also have a specific relationship in terms of time delay; the propagation speed of acoustic waves is different from that of vibration waves, causing them to reach the sensor at different times. While there are time differences between the detectors, existing fusion schemes often ignore these physical relationships, simply splicing the features of acoustic and vibration signals or voting on the detection results of the two, failing to fully utilize this physical information to improve detection accuracy. The technical solution provided in this application, by introducing a physical constraint acoustic-vibration deep fusion module and a small-sample lifelong learning dynamic adaptation module, constructs a multi-level interference filtering mechanism, significantly improving the accuracy and reliability of pipeline anomaly detection. Based on a pre-stored acoustic-vibration propagation physical parameter library, the real-time acquired acoustic-vibration data is rigorously verified. By setting a verification threshold, the system automatically removes environmental noise that does not conform to the physical propagation laws. Simultaneously, using the acoustic-vibration coupling feature base library pre-built by the small-sample lifelong learning module, secondary feature matching is performed on the filtered data. By extracting multi-dimensional coupling features in the time domain, frequency domain, and time-frequency domain, and based on cosine similarity calculation, the system accurately identifies and removes known interference signals, retaining only true anomaly data, thus significantly reducing the false alarm rate. Figure 1 As shown, it includes steps S100-S600.

[0054] Step S100: Deploy acoustic and vibration acquisition units, edge computing units, and cloud processing units along the pipeline to be monitored. Connect the edge computing units to the acoustic and vibration acquisition units and the cloud processing units to the edge computing units to obtain the deployment unit items.

[0055] It should be noted that the method for obtaining the arrangement unit item includes: the acoustic vibration acquisition unit adopts a hybrid deployment of distributed fiber optic acoustic vibration sensors and point piezoelectric acoustic vibration sensors, and the distributed fiber optic acoustic vibration sensor cable is laid in the same trench as the pipeline to be monitored to obtain the first arrangement item; setting scene trigger conditions, including construction frequency, crossing rivers, crossing roads, and key buried sections, to obtain scene trigger items; setting an arrangement threshold based on the scene trigger items, the arrangement threshold being a fixed distance value; and arranging point piezoelectric acoustic vibration sensors based on the arrangement threshold to obtain the second arrangement item; the first arrangement item and the second arrangement item are combined to obtain the arrangement unit item.

[0056] Specifically, the distributed fiber optic acoustic vibration sensing cable is laid in the same trench as the pipeline to be monitored, 30cm away from the outer wall of the pipeline, with the same burial depth as the pipeline, covering all pipeline sections throughout the entire process, serving as the basis for full-area perception. At the same time, scenario triggering conditions are set. When there are high-construction frequency sections with construction sites under construction or planned construction nearby, river crossing sections, road crossing sections, or high-pressure buried key sections near residential areas, a threshold of 200m is set for all marked key sections, that is, one point-type piezoelectric acoustic vibration sensor is deployed every 200m. The signal is calibrated with the fiber optic acquisition points at the same location to supplement the perception accuracy of key sections. All acquisition points are connected to the nearest edge computing box to conduct full-point signal testing to ensure no perception blind spots, thus completing the deployment.

[0057] Step S200: Embed a physical constraint acoustic-vibration deep fusion module in the edge computing unit and the cloud processing unit, and pre-store the acoustic-vibration propagation physical parameter library of the pipeline to be monitored in the physical constraint acoustic-vibration deep fusion module to obtain the stored dataset.

[0058] It is important to note that, such as Figure 2 As shown, the method for obtaining the stored dataset includes: the stored dataset contains threshold values ​​for propagation delay difference, energy attenuation coefficient, and spectrum correspondence of acoustic and vibration signals for different anomaly types with different pipe materials, pipe diameters, and transport media. Different anomaly types include micro-leakage, third-party discovery, pipe wall deformation, and external force collision; a category storage table is created based on the pipe material, pipe diameter, and transport media, and the category storage tables are combined to obtain a category combination set; based on the data acquisition method, the threshold values ​​for propagation delay difference, energy attenuation coefficient, and spectrum correspondence of acoustic and vibration signals under the corresponding combination are obtained using the category combination set as a benchmark, thereby obtaining the stored dataset.

[0059] Specifically, three core parameter dimensions are first defined: the first is pipeline attributes, including material, pipe diameter, and transported medium; the second is anomaly type, including micro-leakage, third-party excavation, pipe wall deformation, and external force collision; and the third is physical verification parameters, namely, the threshold for sound and vibration propagation time delay difference, the threshold for energy attenuation coefficient, and the threshold for spectrum correspondence. A category storage table is first created, with three categories: the first category is pipeline material, the second is pipeline attributes, and the third is the transported medium. The category storage table is then combined to obtain a category combination set. A unique ID is generated for all possible combinations; for example, L360M steel pipe + DN600 + natural gas is combination 1, PE pipe + DN300 + tap water is combination 2, and so on, to obtain the category combination set. For each combination, the thresholds of the three physical verification parameters under different anomaly types are collected through laboratory calibration, historical testing of operational pipelines, and big data acquisition, and stored under the corresponding combination entry to form a complete storage dataset.

[0060] Step S300: Obtain the design parameter information of the pipeline to be monitored, and based on the matching results of the design parameter information in the stored dataset, obtain the reference acoustic and vibration signal parameters, and then obtain the benchmark data items.

[0061] It is important to note that, such as Figure 3 As shown, the method for obtaining benchmark data items includes: obtaining design parameter information of the pipeline to be monitored, including pipeline material, pipe diameter, and transport medium, to obtain design parameter items; matching the design parameter items with the parameters in the stored dataset one by one to extract the initial threshold of the parameters under the corresponding scenario, to obtain general benchmark items; setting a collection threshold, which is a fixed time value, and collecting normal operating sound and vibration data during the trial operation phase of the pipeline to be monitored to obtain pre-collected items; setting a fluctuation value, which is a fixed range value, and fine-tuning the threshold of the general benchmark items based on the fluctuation value to adapt to the specific characteristics of the current pipeline, to obtain calibration benchmark items; and using the average combination result of calibration benchmark items and general benchmark items as benchmark data items.

[0062] Specifically, all core design parameters, including pipe material, pipe diameter, and transport medium, are exported from the pipeline as-built data to form design parameter items. The design parameters are then matched with the stored dataset, and all thresholds under the corresponding combinations are extracted as general benchmark items. The set collection threshold is seven days, and a seven-day trial operation collection period is set to collect acoustic and vibration data of the pipeline under normal operation. The fluctuation range of parameters under normal conditions is statistically analyzed, and the set fluctuation value is ±10%. The threshold of the general benchmark items is fine-tuned according to the ±10% fluctuation range to adapt to the specific conditions such as the burial depth and soil characteristics of this pipeline. Finally, the benchmark parameters are integrated, and the average value of the calibrated parameters and the general parameters is taken as the final benchmark data item.

[0063] In the specific implementation process, the acoustic and vibration propagation time delay difference of the existing target pipeline under the third-party excavation scenario is 0.23 ms / km, thus obtaining the general benchmark. At the same time, the acoustic and vibration time delay difference under normal airflow fluctuations was collected during the 7-day trial operation of the pipeline, which was 0.24 ms / km. At this time, according to the set fluctuation threshold, the general benchmark was adjusted, and the adjusted acoustic and vibration propagation time delay difference was 0.207 ms / km, thus obtaining the calibration benchmark. The average combination result of the calibration benchmark and the general benchmark was used as the benchmark data item, and the result was 0.2235 ms / km, which is more suitable for the actual operation of the target pipeline than the general benchmark.

[0064] Step S400: Based on the acoustic and vibration acquisition unit, the acoustic and vibration data of the monitored pipeline are acquired in real time to obtain the acquisition dataset. The pre-embedded physical verification items are called in the physical constraint acoustic and vibration deep fusion module. Data verification and data removal are performed based on the acquisition dataset to filter out interference signals that do not conform to the physical laws of acoustic and vibration propagation in the pipeline, and a filtered dataset is obtained.

[0065] It is important to note that, such as Figure 4 As shown, the physical verification items include: setting a verification threshold, which is a fixed percentage value; combining the verification threshold with the benchmark data items to obtain a combined judgment item; when the judgment deviation exceeds the combined judgment item, it is judged as a non-pipeline-related interference signal, thus obtaining the physical verification item.

[0066] Specifically, the set verification threshold is 15%. Based on the verification threshold and the benchmark data item, when there is a deviation exceeding 15% of the benchmark data item, it is judged as non-pipeline-related interference information.

[0067] It is important to note that the method for obtaining the filtered dataset includes: the acoustic and vibration acquisition unit continuously acquires the original acoustic and vibration signals of the pipeline under test, sets a time window value, and slices and encapsulates the signals within the time window value to obtain the acquisition dataset; the acoustic and vibration signals in the acquisition dataset are preprocessed by timestamp alignment and power frequency denoising to obtain the preprocessed dataset; the physical verification item in the physical constraint acoustic and vibration deep fusion module is called to calculate the parameter values ​​of the actual propagation delay difference, energy ratio, and spectral correlation degree of the acoustic and vibration signals within each time window to obtain the parameter calculation item; the parameter calculation item is compared with the corresponding threshold of the benchmark data item, and if the deviation of any dimension parameter exceeds the physical verification item, it is determined to be environmental interference that does not conform to the physical propagation law, and the data of that time window is removed, and the remaining valid data is integrated to obtain the filtered dataset.

[0068] Specifically, the acquisition unit continuously acquires both acoustic and vibration signals 24 hours a day. A 1-second time window is set, and each 1-second signal is packaged into a data frame with a timestamp to form an acquisition dataset. Each data frame is subjected to timestamp alignment, 50Hz power frequency denoising, and wavelet denoising to obtain a preprocessed dataset. At the same time, for each 1-second data frame, the propagation delay difference, energy ratio, and spectral correlation of the acoustic and vibration signals are calculated. The calculated three parameters are compared with the interval of the physical verification items. If any dimension exceeds the interval, the data frame is discarded. The remaining valid data is integrated into a filtered dataset.

[0069] Step S500: Based on the small sample lifelong learning dynamic adaptation module preset by the cloud processing unit, the cover feature set is obtained by calling the pre-built acoustic-vibration coupling feature base library of the module. The filter dataset is then subjected to secondary feature matching and filtering based on the cover feature set to further filter out known scene interference signals and obtain the secondary filter dataset.

[0070] It is important to note that, such as Figure 5 As shown, the method for obtaining the secondary filtering dataset includes: calling the pre-built acoustic-vibration coupling feature base library of the small-sample lifelong learning dynamic adaptation module, which covers known scene interference features in different regions, working conditions, and seasons to obtain the covered feature set; extracting time-domain, frequency-domain, and time-frequency-domain multi-dimensional coupling features from the acoustic-vibration signals of the filtering dataset to obtain the feature items to be matched; performing cosine similarity calculation between the feature items to be matched and the covered feature set; setting a similarity judgment threshold; when the calculation result exceeds the similarity judgment threshold, it is judged as known scene interference, and the corresponding data is removed to obtain the secondary filtering dataset.

[0071] Specifically, the cloud-based small-sample lifelong learning module calls a pre-stored acoustic-vibration coupling feature base library. This library covers known interference features under different regions, seasons, and working conditions, such as dump trucks, pile driving, renovations, and lightning. For each data frame in the filtered dataset, it extracts coupling features in the time domain (peak value, kurtosis), frequency domain (peak frequency, spectral entropy), and time-frequency domain (wavelet energy coefficient) to form a feature term to be matched. The cosine similarity between the feature term to be matched and the features in the base library is calculated, with a threshold of 0.85. If the similarity exceeds 0.85, it is judged as known interference, and the corresponding data frame is discarded. The remaining data is the secondary filtered dataset.

[0072] Step S600: Based on the secondary filtering dataset, pipeline anomaly detection is performed, and anomaly determination is completed, thereby realizing multi-level interference filtering and sound-vibration fusion pipeline anomaly detection.

[0073] It is important to note that the anomaly detection process includes: the edge computing unit performs a preliminary classification of the secondary filtered dataset; for those deemed potentially anomaly, the corresponding original acoustic and vibration data is uploaded to the cloud; otherwise, only the compressed normal operation feature archive is uploaded, resulting in an initial screening judgment item; the cloud processing unit calls a high-precision fusion model to perform a secondary judgment on the suspected anomaly data, clarifying the anomaly type, and calculating the location of the anomaly point by combining the acoustic and vibration propagation time delay difference, resulting in a refined judgment result item; the refined judgment result item is then connected to the pipeline GIS system, automatically retrieving AI video surveillance around the anomaly point for verification, and simultaneously pushing the anomaly type, location, and handling suggestions to the mobile terminal of the maintenance personnel, forming a closed-loop handling process.

[0074] Specifically, the edge computing unit uses a lightweight classification model to make an initial judgment on the secondary filtered dataset. Suspected anomalies are uploaded to the cloud with complete original acoustic and vibration data, while normal data is only uploaded with compressed feature values ​​for archiving, saving bandwidth. At the same time, the cloud calls a high-precision multimodal fusion model to perform secondary classification on the suspected anomaly data to determine the anomaly type. The anomaly location is calculated by the acoustic and vibration time delay difference, with a positioning accuracy of less than 1 meter. The anomaly information is synchronized to the pipeline GIS system, and AI monitoring within 50 meters of the anomaly point is automatically called for verification. At the same time, the anomaly type, location, and handling suggestions are pushed to the mobile terminal of the operation and maintenance personnel. After the operation and maintenance is completed, the results are uploaded to close the loop.

[0075] In the specific implementation process, at kilometer marker K12+350 of a certain pipeline, adjacent to the boundary of a construction site awaiting development, unauthorized construction workers were digging with shovels to lay temporary water pipes. The system detected them when they were only 0.4 meters from the outer wall of the pipeline. At this point, the acquisition unit collected both acoustic and vibration signals over 24 hours, slicing and encapsulating them into 1-second time windows. A total of 27 frames of valid data were collected during this period, containing three types of signals: vibration from pile driving at the construction site, vibration from passing dump trucks, and vibration from unknown excavation. All data were then timestamped and aligned to the power frequency. After denoising preprocessing, the signal-to-noise ratio reached 27dB. The delay difference, energy ratio, and spectral correlation of each frame of signal were calculated. The delay difference of the piling signal was 0.68ms / km, which exceeded the physical verification range of 0.197-0.302ms / km. All 19 frames of piling signals were removed, and the remaining 8 frames were added to the filtered dataset. At the same time, a pre-built acoustic-vibration coupling feature base library covering 126 known interference features was called. 17-dimensional coupling features in the time domain, frequency domain, and time-frequency domain were extracted from the 8 frames of the filtered dataset. Cosine similarity calculation was performed: 7 frames of dump truck signals had a similarity of 0.93 with the dump truck features in the base storage area, exceeding the 0.85 judgment threshold, and were therefore identified as known interference and removed. Only 1 frame of unknown signal remained and entered the secondary filtering dataset. The edge-end lightweight model initially screened and determined that the signal was a suspected anomaly. The original data was uploaded to the cloud, while normal data was only uploaded with compressed features, reducing bandwidth usage by 91%. The cloud-based high-precision fusion model made a secondary judgment, confirming the anomaly type as "third-party mining". The location was calculated using latency difference and located at K12+350 with a positioning error of 0.7m. The system automatically connected to the pipeline GIS system and called a nearby 50-meter AI PTZ camera to capture images, confirming that 2 people were digging on site. Simultaneously, a work order was pushed to maintenance personnel 2.3 kilometers away. The work order included the anomaly type, location, on-site screenshots, and handling plan. The push took 17 seconds, and the maintenance personnel arrived at the site within 16 minutes to stop the construction. After performing non-destructive testing on the pipeline to confirm no damage, the handling record was uploaded. The work order was closed in 42 minutes, and no losses were caused.

[0076] The all-weather pipeline anomaly detection system based on multimodal acoustic-vibration fusion utilizes the aforementioned all-weather pipeline anomaly detection method based on multimodal acoustic-vibration fusion, including: a unit deployment module: deploying acoustic-vibration acquisition units, edge computing units, and cloud processing units along the pipeline to be monitored; connecting the edge computing units to the acoustic-vibration acquisition units and the cloud processing units to the edge computing units to obtain deployment unit items; a data pre-storage module: embedding a physically constrained acoustic-vibration deep fusion module in the edge computing units and the cloud processing units; pre-storing a physical parameter library of acoustic-vibration propagation for the pipeline to be monitored in the physically constrained acoustic-vibration deep fusion module to obtain a stored dataset; and a data comparison module: acquiring the design parameter information of the pipeline to be monitored; and based on the matching results of the design parameter information in the stored dataset, obtaining the comparison acoustic-vibration signal parameters, and then... The system obtains benchmark data items, acquires real-time acoustic and vibration data of the monitored pipeline based on the acoustic and vibration acquisition unit, and obtains the acquisition dataset. In the physical constraint acoustic and vibration deep fusion module, the pre-embedded physical verification items are called to perform data verification and data removal based on the acquisition dataset, filtering out interference signals that do not conform to the physical laws of acoustic and vibration propagation in the pipeline, and obtaining the filtered dataset. Based on the small sample lifelong learning dynamic adaptation module preset by the cloud processing unit, the system calls the acoustic and vibration coupling feature base library pre-built by the module to obtain the coverage feature set. Based on the coverage feature set, the filtered dataset is subjected to secondary feature matching and filtering to further filter out interference signals of known scenarios, and obtains the secondary filtered dataset. The anomaly detection module uses the secondary filtered dataset as a benchmark to perform pipeline anomaly detection and complete anomaly detection, thereby realizing multi-level interference filtering and acoustic and vibration fusion-based pipeline anomaly detection.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. An all-weather pipeline anomaly detection method based on multimodal acoustic-vibration fusion, including: Along the pipeline to be monitored, sound and vibration acquisition units, edge computing units, and cloud processing units are deployed. The edge computing units are connected to the sound and vibration acquisition units, and the cloud processing units are connected to the edge computing units to obtain the deployment unit items. Its characteristic is that it further includes: A physical constraint acoustic-vibration deep fusion module is embedded in the edge computing unit and the cloud processing unit. The acoustic-vibration propagation physical parameter library of the pipeline to be monitored is pre-stored in the physical constraint acoustic-vibration deep fusion module to obtain the stored dataset. The design parameter information of the pipeline to be monitored is obtained. Based on the matching results of the design parameter information in the stored dataset, the reference acoustic and vibration signal parameters are obtained, and then the benchmark data items are obtained. The acoustic and vibration data of the monitored pipeline are acquired in real time by the acoustic and vibration acquisition unit to obtain the acquisition dataset. The pre-embedded physical verification items are called in the physical constraint acoustic and vibration deep fusion module to perform data verification and data removal based on the acquisition dataset, and to filter out interference signals that do not conform to the physical laws of acoustic and vibration propagation in the pipeline to obtain the filtered dataset. Based on the small-sample lifelong learning dynamic adaptation module preset by the cloud processing unit, the cover feature set is obtained by calling the pre-built acoustic-vibration coupling feature base library of the module. Based on the cover feature set, the filter dataset is subjected to secondary feature matching and filtering to further filter out known scene interference signals and obtain the secondary filter dataset. By using the secondary filtered dataset as a benchmark, pipeline anomaly detection is performed, and anomaly determination is completed, thereby realizing multi-level interference filtering and acoustic-vibration fusion-based pipeline anomaly detection.

2. The all-weather pipeline anomaly detection method based on multimodal acoustic-vibration fusion according to claim 1, characterized in that: The method for obtaining the arrangement unit item includes: The acoustic vibration acquisition unit adopts a hybrid deployment of distributed optical fiber acoustic vibration sensing and point piezoelectric acoustic vibration sensor. The distributed optical fiber acoustic vibration sensing cable is laid in the same trench as the pipeline to be monitored, which is the first arrangement item. Set scene trigger conditions, including construction frequency, river crossing, road crossing, and key buried sections, to obtain scene trigger items. Set a placement threshold based on the scene trigger items. The placement threshold is a fixed distance value. Place point piezoelectric acoustic vibration sensors based on the placement threshold to obtain the second placement item. The first arrangement item and the second arrangement item are combined to obtain the arrangement unit item.

3. The all-weather pipeline anomaly detection method based on multimodal acoustic-vibration fusion according to claim 1, characterized in that: The method for obtaining the stored dataset includes: The stored dataset contains thresholds for propagation delay difference, energy attenuation coefficient, and spectrum correspondence of acoustic and vibration signals for different anomaly types, including micro-leakage, third-party excavation, pipe wall deformation, and external force collision, for different pipe materials, pipe diameters, and transported media. Based on the pipe material, pipe diameter, and transport medium, create a category storage table, and combine the category storage tables to obtain a category combination set; Based on the data acquisition method, using the category combination set as a benchmark, the thresholds for propagation delay difference, energy attenuation coefficient, and spectrum correspondence of acoustic and vibration signals under the corresponding combination are obtained, thereby obtaining the stored dataset.

4. The all-weather pipeline anomaly detection method based on multimodal acoustic-vibration fusion according to claim 1, characterized in that: The method for obtaining the benchmark data items includes: The design parameter information of the pipeline to be monitored, including pipeline material, pipe diameter and conveying medium, is used to obtain the design parameter items; The design parameters are matched one by one with the parameters in the stored dataset, and the initial thresholds of the parameters in the corresponding scenarios are extracted to obtain general benchmarking items. Set a collection threshold, which is a fixed time value. During the trial operation of the pipeline to be monitored, collect normal operating sound and vibration data to obtain pre-collected items. Set a fluctuation value, which is a fixed range value. Based on the fluctuation value, fine-tune the threshold of the general benchmarking items to adapt to the specific characteristics of the current pipeline and obtain the calibration benchmarking items. The average combination of calibration benchmarks and general benchmarks is used as the benchmark data item.

5. The all-weather pipeline anomaly detection method based on multimodal acoustic-vibration fusion according to claim 1, characterized in that: The physical verification items include: A verification threshold is set as a fixed percentage value. The verification threshold is combined with the benchmark data items to obtain a combined judgment item. When the judgment deviation exceeds the combined judgment item, it is judged as a non-pipeline-related interference signal, and then the physical verification item is obtained.

6. The all-weather pipeline anomaly detection method based on multimodal acoustic-vibration fusion according to claim 1, characterized in that: The method for obtaining the filtered dataset includes: The acoustic and vibration acquisition unit continuously acquires the original acoustic and vibration signals of the pipeline under test, sets a time window value, and slices and encapsulates the data to obtain the acquisition dataset. The acoustic and vibration signals in the acquired dataset are preprocessed by timestamp alignment and power frequency denoising to obtain the preprocessed dataset. The physical verification item in the physical constraint acoustic-vibration deep fusion module is called to calculate the parameter values ​​of the actual propagation delay difference, energy ratio, and spectral correlation degree of the acoustic-vibration signal in each time window, and the parameter calculation item is obtained. The calculated parameters are compared with the corresponding thresholds of the benchmark data items. If the deviation of any dimension parameter exceeds the physical verification item, it is determined to be environmental interference that does not conform to the physical propagation law. The data in that time window is then removed, and the remaining valid data is integrated to obtain the filtered dataset.

7. The all-weather pipeline anomaly detection method based on multimodal acoustic-vibration fusion according to claim 1, characterized in that: The method for obtaining the secondary filtering dataset includes: The acoustic-vibration coupling feature base library, which is pre-built by the small-sample lifelong learning dynamic adaptation module, is called. The base library covers known scene interference features in different regions, working conditions and seasons, and a covered feature set is obtained. Multi-dimensional coupled features in the time domain, frequency domain, and time-frequency domain are extracted from the acoustic and vibration signals of the filtered dataset to obtain the feature terms to be matched. The cosine similarity between the feature terms to be matched and the covered feature set is calculated. A similarity judgment threshold is set. When the calculation result exceeds the similarity judgment threshold, it is judged as known scene interference, and the corresponding data is removed to obtain the secondary filtered dataset.

8. The all-weather pipeline anomaly detection method based on multimodal acoustic-vibration fusion according to claim 1, characterized in that: The anomaly determination includes: The edge computing unit performs a preliminary classification of the secondary filtered dataset. If it determines that the dataset is suspected of being abnormal, it uploads the corresponding original acoustic and vibration data to the cloud. Otherwise, it only uploads the compressed normal operation feature archive to obtain the initial screening judgment items. The cloud processing unit calls the high-precision fusion model to make a secondary judgment on the suspected abnormal data, clarify the abnormality type, and calculate the location of the abnormal point by combining the sound and vibration propagation time delay difference to obtain the precise judgment result item; The precise judgment results are compared with the network GIS system, and the AI ​​video surveillance around the anomaly point is automatically retrieved for verification. At the same time, the anomaly type, location and handling suggestions are pushed to the mobile terminal of the operation and maintenance personnel to form a closed loop of handling.

9. An all-weather pipeline anomaly detection system based on multimodal acoustic-vibration fusion, characterized in that: The all-weather pipeline anomaly detection method based on multimodal acoustic-vibration fusion, as described in any one of claims 1-8, includes: Unit Layout Module: Along the pipeline to be monitored, sound and vibration acquisition units, edge computing units, and cloud processing units are laid out. The edge computing units are connected to the sound and vibration acquisition units, and the cloud processing units are connected to the edge computing units to obtain the layout unit items. Data pre-storage module: Embed a physical constraint acoustic vibration deep fusion module in the edge computing unit and the cloud processing unit. The physical constraint acoustic vibration deep fusion module pre-stores the acoustic vibration propagation physical parameter library of the pipeline to be monitored, and obtains the stored dataset. The data comparison module acquires the design parameter information of the pipeline to be monitored. Based on the matching results of the design parameter information in the stored dataset, it acquires the reference acoustic and vibration signal parameters, and then obtains the benchmark data items. Based on the acoustic and vibration acquisition unit, it acquires the acoustic and vibration data of the monitored pipeline in real time, and obtains the acquisition dataset. In the physical constraint acoustic and vibration deep fusion module, it calls the pre-embedded physical verification items, performs data verification and data removal based on the acquisition dataset, filters out interference signals that do not conform to the physical laws of acoustic and vibration propagation in the pipeline, and obtains the filtered dataset. Based on the small sample lifelong learning dynamic adaptation module preset by the cloud processing unit, it calls the acoustic and vibration coupling feature base library pre-built by the module to obtain the coverage feature set. Based on the coverage feature set, it performs secondary feature matching and filtering on the filtered dataset to further filter out interference signals of known scenarios, and obtains the secondary filtered dataset. Anomaly detection module: Based on the secondary filtered dataset, pipeline anomaly detection is performed and anomaly detection is completed, thereby realizing multi-level interference filtering and sound and vibration fusion pipeline anomaly detection.

Images