Pipe network leakage self-learning diagnosis system based on voiceprint recognition

Abstract

The invention discloses a pipe network leakage self-learning diagnosis system based on voiceprint recognition, which comprises a management platform, the management platform comprises a central processing unit, and one end of the central processing unit is connected with an indication module, a touch panel and a cloud server. The other end of the central processor is connected with a voiceprint acquisition array module, a signal preprocessing module, a multi-modal feature extraction module, a self-learning diagnosis core module, a visual interaction platform module, an edge calculation integration module and a multi-source data fusion positioning module. According to the method, the diagnosis accuracy can be greatly improved, the anti-interference capability is enhanced, the environmental adaptability is greatly improved, the service life of the system is prolonged, the operation and maintenance cost is optimized, the positioning accuracy is enhanced, manual calibration is not needed, the maintenance efficiency is improved, and the expansibility can be improved; the system has an advanced framework, an efficient working process and a remarkable application effect, and can provide powerful guarantee for safe operation of an urban pipe network.

Description

Technical Field

[0001] The present invention relates to the technical field of self-learning diagnosis of pipe network leakage, and in particular to a self-learning diagnosis system for pipe network leakage based on voiceprint recognition. Background Art

[0002] In the operation and maintenance of underground pipelines such as urban water supply, gas supply, and oil pipelines, leakage diagnosis of underground pipelines is required. However, the existing technology has the following defects: 1. Environmental noise interference: Traditional voiceprint sensors are easily interfered by environmental noise such as turbulent water flow in pipes and vibration from ground traffic, which affects the accurate collection of leakage signals; 2. Model iteration lag: The fixed threshold diagnosis system is difficult to adapt to the voiceprint feature deviation caused by pipeline aging and corrosion, resulting in reduced diagnostic accuracy; 3. Ambiguous leak location: A single sensor system cannot achieve meter-level accurate positioning, often requiring multiple rounds of manual investigation, which is inefficient. 4. High false alarm rate: Conventional frequency domain analysis methods cannot effectively distinguish normal water flow fluctuations from real leakage signals, increasing the workload of manual inspections; 5. Data island problem: There is a lack of dynamic correlation analysis between voiceprint data and operating parameters such as pressure and flow, making it difficult to comprehensively assess the health status of the pipeline network.

[0003] To this end, we proposed a pipe network leakage self-learning diagnosis system based on voiceprint recognition to solve the above problems. Summary of the Invention

[0004] The purpose of the present invention is to solve the shortcomings of the existing technology and propose a pipe network leakage self-learning diagnosis system based on voiceprint recognition.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: The self-learning diagnosis system for pipe network leakage based on voiceprint recognition includes a management platform, which includes a central processing unit. One end of the central processing unit is connected to an indication module, a touch panel and a cloud server, and the other end of the central processing unit is connected to a voiceprint acquisition array module, a signal preprocessing module, a multimodal feature extraction module, a self-learning diagnosis core module, a visual interaction platform module, an edge computing integration module and a multi-source data fusion positioning module.

[0006] Preferably, the soundprint acquisition array module is composed of a distributed network of multiple groups of high-sensitivity acoustic wave composite sensors, covering key nodes of the pipeline network. The sensor array adopts a spiral topology layout and reduces environmental electromagnetic interference through adaptive power regulation technology. Each node contains a dual-modal data acquisition unit of a piezoelectric sensor and a fiber optic acoustic wave sensor, or contains one of the combinations of an embedded fiber optic acoustic vibration sensor, a surface mobile quantum dot laser interferometer, and a pipe wall contact piezoelectric film sensor, which is used to collect pipeline vibration soundprint signals in real time.

[0007] Preferably, the signal preprocessing module performs noise reduction, filtering and multi-channel synchronous calibration on the original voiceprint signal to generate a standardized time-frequency domain mixed signal, and adopts magnetohydrodynamic noise suppression technology to offset the pipeline medium flow noise through the Lorentz force.

[0008] Preferably, the multimodal feature extraction module extracts the frequency domain energy distribution, time domain pulse features and nonlinear resonance features of the voiceprint signal based on the improved wavelet packet decomposition algorithm and Mel-frequency cepstral coefficients, performs multi-scale hybrid feature extraction, and adopts adaptive weighted fusion of time and frequency domain features.

[0009] Preferably, the self-learning diagnosis core module adopts a dynamic weighted residual network to fuse multimodal features. The dynamic weighted residual network introduces a channel attention mechanism to dynamically adjust the weight coefficients of different frequency band features. The network training adopts a transfer learning framework, and the pre-training model is generated based on a public acoustic database and simulated leakage voiceprint data, or adopts one of the dual-channel deep confidence network architectures, including a main channel leakage pattern classifier and an auxiliary channel environmental noise discriminator, and a dynamic model update mechanism based on confidence evaluation.

[0010] Preferably, the dynamic weighted residual network retains deep features through residual connections, the channel attention mechanism strengthens the weights of leakage-related frequency bands, and combines incremental learning to achieve rapid adaptation of the model to new leakage patterns. The dual-channel deep confidence network classifies leakage patterns through the main channel and the auxiliary channel identifies environmental noise. The model is dynamically updated based on confidence evaluation. The knowledge distillation mechanism adopts teacher-student model interaction. The lightweight student model on the edge and the expert teacher model on the cloud perform knowledge distillation based on dynamic entropy values ​​to improve the efficiency of model iteration.

[0011] Preferably, the visual interactive platform module includes integrated GIS map positioning, leakage level heat map, maintenance decision suggestion output, diagnosis result output, leakage probability heat map, and pipe health index three-dimensional cloud map.

[0012] Preferably, the edge computing integration module deploys an embedded edge computing unit at the pipe network node to realize local feature extraction and primary diagnosis of the voiceprint signal, and only uploads key feature data to the cloud to reduce network bandwidth occupancy. A lightweight diagnostic model is deployed at the edge, and model retraining and parameter distribution are performed on the cloud.

[0013] Preferably, the multi-source data fusion positioning module adopts the fusion method of voiceprint time difference positioning and vibration wave attenuation model, the improved generalized cross-correlation delay estimation method and the back propagation algorithm based on the attenuation model of material stress wave, combined with the real-time correction of the pipeline network GIS topology data to achieve accurate positioning of the leakage point.

[0014] The workflow of the present invention is multi-dimensional synchronous acquisition of voiceprints - field strength coupling noise reduction - spatiotemporal feature decoupling - dual diagnosis decision-making - multi-mode positioning compensation - knowledge distillation update. The specific steps include voiceprint signal acquisition, environmental noise stripping, multi-dimensional feature fusion, dynamic model diagnosis, collaborative positioning of leakage points, and incremental update of the knowledge base. The leakage probability calculation adopts the improved DS evidence theory, which integrates the acoustic energy distribution confidence, material vibration mode matching, and historical leakage case similarity. The system includes a self-test mode, and regularly verifies the sensor sensitivity and frequency response characteristics through white noise excitation and Chirp signal. Virtual leakage scenario simulation tests are performed every month. The workflow and self-test mode can ensure the accuracy and reliability of the system and timely discover leakage problems in the pipeline network.

[0015] The present invention has the following advantages: 1. Multimodal feature fusion significantly improves the accuracy of leakage identification to 98.7%-98.9%, a 22%-23% improvement over traditional methods, and reduces the false alarm rate to below 0.3 times per thousand hours, greatly improving diagnostic accuracy. 2. The dynamic noise reduction algorithm enables the system to effectively extract leakage features even under high background noise (75dB-85dB), improving the anti-interference ability by 300% compared with traditional methods and greatly improving environmental adaptability; 3. The system's false alarm rate is less than 0.5%, reducing the frequency of manual inspections by more than 80% and lowering operation and maintenance costs. The edge-cloud collaborative architecture reduces backbone network load by 75%, extending system life and optimizing operation and maintenance costs. 4. Fusion positioning technology achieves ±0.2m-±0.5m positioning accuracy, which enhances positioning accuracy and eliminates the need for manual calibration, improving maintenance efficiency; 5. Supports access to multi-source sensor data such as pressure and flow to build a comprehensive evaluation system for pipeline network health, greatly improving scalability; 6. The knowledge distillation mechanism improves the system's model iteration efficiency by 60% per quarter, supports autonomous identification of 17 new types of leakage patterns, can quickly adapt to changes in the pipeline network, and enhances its adaptive evolution capabilities; To sum up, the present invention can not only greatly improve the accuracy of diagnosis, but also enhance the anti-interference ability, greatly improve the environmental adaptability, and at the same time extend the system life, realize the optimization of operation and maintenance costs, enhance the accuracy of positioning, and do not require manual calibration, thereby improving maintenance efficiency and improving scalability. It has an advanced architecture, efficient workflow and significant application effects, and can provide strong guarantees for the safe operation of urban pipeline networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a system structure diagram of the present invention; Figure 2 This is a structural diagram of the voiceprint collection array module of the present invention; Figure 3 This is a diagram showing the content of the visual interactive platform module of the present invention; Figure 4 This is a process flow chart of the system of the present invention; Figure 5 This is a structural diagram of the management platform of the present invention; Figure 6 It is the workflow diagram of the present invention. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0018] Reference Figure 1-6 The self-learning diagnosis system for pipe network leakage based on voiceprint recognition includes a management platform. The management platform includes a central processing unit. One end of the central processing unit is connected to an indication module, a touch panel and a cloud server. The other end of the central processing unit is connected to a voiceprint acquisition array module, a signal preprocessing module, a multimodal feature extraction module, a self-learning diagnosis core module, a visual interaction platform module, an edge computing integration module and a multi-source data fusion positioning module. The management platform serves as the core hub of the entire system. The central processing unit coordinates the work of each module. The indication module connected to one end can not only intuitively display the operating status of the system through indicator lights, such as the green light is always on during normal operation and the red light flashes when an abnormality occurs, but also inform the operator of the specific situation of the system through voice prompts. The touch panel provides a convenient interaction method for the operator. It adopts a high-definition capacitive screen with fast response speed and supports multi-touch. The operator can easily input instructions, query data, etc. The cloud server is used to store a large amount of historical data and model parameters, and can realize remote data synchronization and sharing, facilitating remote diagnosis and analysis by experts. The soundprint acquisition array module is composed of a distributed network of multiple groups of high-sensitivity acoustic wave composite sensors, covering key nodes of the pipeline network. The sensor array adopts a spiral topology layout and reduces environmental electromagnetic interference through adaptive power regulation technology. Each node contains a dual-modal data acquisition unit of a piezoelectric sensor and a fiber optic acoustic wave sensor, or contains one of the following combinations: an embedded fiber optic acoustic vibration sensor, a surface-moving quantum dot laser interferometer, and a pipe-wall-contact piezoelectric thin film sensor, for real-time collection of pipeline vibration soundprint signals. The sensor power can be automatically adjusted according to the intensity of environmental electromagnetic interference to ensure the accuracy of the collected soundprint signals. Each node contains a dual-modal data acquisition unit of a piezoelectric sensor and a fiber optic acoustic wave sensor, or contains one of the following combinations: an embedded fiber optic acoustic vibration sensor, a surface-moving quantum dot laser interferometer, and a pipe-wall-contact piezoelectric thin film sensor, for real-time collection of pipeline vibration soundprint signals. Different types of sensor combinations can adapt to different pipeline network environments and monitoring needs. For example, in some complex underground pipeline networks, a combination of embedded fiber optic acoustic vibration sensors and pipe-wall-contact piezoelectric thin film sensors can be used to more accurately collect pipeline soundprint signals. The signal preprocessing module performs noise reduction, filtering, and multi-channel synchronous calibration on the original voiceprint signal to generate a standardized time-frequency domain mixed signal. It uses magnetohydrodynamic noise suppression technology to offset the pipeline medium flow noise through the Lorentz force. The motion characteristics of the magnetic fluid in the magnetic field are used to generate a force opposite to the pipeline medium flow noise, thereby effectively canceling the noise. At the same time, the module also uses advanced filtering algorithms to automatically adjust the filtering parameters according to the frequency characteristics of the voiceprint signal, further improving the signal quality. The multimodal feature extraction module extracts the frequency domain energy distribution, time domain pulse characteristics, and nonlinear resonance characteristics of the voiceprint signal based on an improved wavelet packet decomposition algorithm and Mel-frequency cepstral coefficients, performs multi-scale hybrid feature extraction, and adopts adaptive weighted fusion of time-frequency domain features. The improved wavelet packet decomposition algorithm can more finely decompose the frequency components of the voiceprint signal and extract more feature information. The Mel-frequency cepstral coefficients can simulate the characteristics of the human auditory system and extract more representative voiceprint features. Through adaptive weighted fusion of time-frequency domain features, the weights of different features can be automatically adjusted according to their importance, thereby improving the accuracy of feature extraction. The core module of self-learning diagnosis uses a dynamic weighted residual network to fuse multimodal features. The dynamic weighted residual network introduces a channel attention mechanism to dynamically adjust the weight coefficients of features in different frequency bands. Network training adopts a transfer learning framework. The pre-trained model is generated based on a public acoustic database and simulated leakage voiceprint data, or adopts one of the dual-channel deep belief network architectures, which includes a main channel leakage pattern classifier and an auxiliary channel environmental noise discriminator. The dynamic model update mechanism is based on confidence assessment. The dynamic weighted residual network retains deep features through residual connections. The channel attention mechanism strengthens the weights of leakage-related frequency bands. Combined with incremental learning, the model can quickly adapt to new leakage patterns. The dual-channel deep belief network classifies leakage patterns through the main channel and identifies environmental noise through the auxiliary channel. The model is dynamically updated based on confidence assessment. The knowledge distillation mechanism uses a teacher-student model interaction. The lightweight student model on the edge and the expert teacher model on the cloud perform knowledge distillation based on dynamic entropy value to improve model iteration efficiency. The self-learning mechanism enables the system to continuously adapt to new leakage patterns and environmental changes, improving the accuracy and reliability of diagnosis. The visualization interactive platform module includes integrated GIS map positioning, leakage level heat map, maintenance decision suggestion output, diagnosis result output, leakage probability heat map, and pipe health index three-dimensional cloud map. The integrated GIS map positioning can intuitively display the geographical location of the pipeline network and the specific location of the leakage point, making it easy for maintenance personnel to quickly reach the site. The leakage level heat map uses different colors to represent different leakage levels, allowing operators to quickly understand the leakage situation of the pipeline network. The maintenance decision suggestion output provides maintenance personnel with the best maintenance plan based on the diagnosis results and historical data. The diagnosis result output displays the diagnosis results in intuitive charts and text forms, facilitating analysis and judgment by operators. The leakage probability heat map predicts the leakage probability of various parts of the pipeline network by analyzing the voiceprint signal. The pipe health index three-dimensional cloud map can intuitively display the health status of the pipe, providing a reference for the maintenance and management of the pipeline network. The edge computing integration module deploys embedded edge computing units at the network nodes to achieve local feature extraction and primary diagnosis of voiceprint signals. Only key feature data is uploaded to the cloud, reducing network bandwidth usage. A lightweight diagnostic model is deployed at the edge, and model retraining and parameter distribution are performed in the cloud. Distributed computing can improve the system's response speed and reliability while reducing network transmission pressure. The embedded edge computing unit uses low-power, high-performance chips and can operate stably in harsh environments. The multi-source data fusion positioning module adopts the fusion method of voiceprint time difference positioning and vibration wave attenuation model, the fusion method of improved generalized cross-correlation time delay estimation method and the back propagation algorithm based on the attenuation model of material stress wave, combined with the real-time correction of pipeline network GIS topology data to achieve accurate positioning of leaks. Voiceprint time difference positioning determines the location of the leak by measuring the propagation time difference of voiceprint signals between different sensors. The vibration wave attenuation model calculates the location of the leak according to the propagation characteristics of vibration waves in the pipeline. The improved generalized cross-correlation time delay estimation method and the back propagation algorithm based on the attenuation model of material stress wave can improve the accuracy of positioning. Combined with the real-time correction of pipeline network GIS topology data, the positioning accuracy can be further improved. The workflow of the present invention is multi-dimensional synchronous acquisition of voiceprints - field strength coupling noise reduction - spatiotemporal feature decoupling - dual diagnosis decision-making - multi-mode positioning compensation - knowledge distillation update. The specific steps include voiceprint signal acquisition, environmental noise stripping, multi-dimensional feature fusion, dynamic model diagnosis, collaborative positioning of leakage points, and incremental update of the knowledge base. The leakage probability calculation adopts the improved DS evidence theory, which integrates the acoustic energy distribution confidence, material vibration mode matching, and historical leakage case similarity. The system includes a self-test mode, and regularly verifies the sensor sensitivity and frequency response characteristics through white noise excitation and Chirp signal. Virtual leakage scenario simulation tests are performed every month. The workflow and self-test mode can ensure the accuracy and reliability of the system and timely discover leakage problems in the pipeline network.

[0019] Deployment test cases of the present invention: 1. Urban water supply network A set of composite sensors (including piezoelectric sensors, fiber optic acoustic sensors, etc.) is deployed every 200 meters along the DN600 cast iron main pipeline, and inspection drones equipped with laser vibrometers conduct monthly mobile scanning.

[0020] Edge computing nodes perform real-time diagnostics and activate deep cloud-based analysis when suspected leaks are detected.

[0021] The system automatically generates a leakage probability heat map and pushes it to the pipeline network GIS platform.

[0022] After maintenance, the excavation data is automatically transmitted back to update the pipe corrosion soundprint feature library.

[0023] It proactively identified 12 hidden leakage points within six months, saving an estimated 380,000 tons of water. The system's self-learning module shortened the detection time for micro-leaks with a diameter of less than 5mm from 72 hours to 4.2 hours through three incremental updates. The GIS visualization platform achieved a 5-meter positioning accuracy for leakage points, reducing maintenance response time by 65%.

[0024] 2. Urban gas pipeline network A set of fiber optic Bragg sensors is deployed every 50 meters along the PE200 gas pipeline, and drones equipped with quantum dot lasers perform daytime inspections.

[0025] The edge computing node runs the Mobile-Transformer model in real time, and activates cloud-based 3D-ResNeXt deep analysis when the diagnosis confidence level is lower than 85%.

[0026] The system automatically generates a three-dimensional leakage situation map and simulates maintenance plans through the digital twin platform.

[0027] The pipe wall vibration data during the maintenance process is automatically transmitted back to update the material aging feature library.

[0028] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A self-learning diagnosis system for pipe network leakage based on voiceprint recognition, including a management platform, characterized in that: The management platform includes a central processing unit, one end of which is connected to an indication module, a touch panel and a cloud server, and the other end of which is connected to a voiceprint acquisition array module, a signal preprocessing module, a multimodal feature extraction module, a self-learning diagnosis core module, a visual interaction platform module, an edge computing integration module and a multi-source data fusion positioning module.

2. The pipe network leakage self-learning diagnosis system based on voiceprint recognition according to claim 1 is characterized by: The soundprint acquisition array module is composed of a distributed network of multiple groups of high-sensitivity acoustic wave composite sensors, covering key nodes of the pipeline network. The sensor array adopts a spiral topology layout and reduces environmental electromagnetic interference through adaptive power regulation technology. Each node contains a dual-modal data acquisition unit of a piezoelectric sensor and a fiber optic acoustic wave sensor, or contains one of the combinations of an embedded fiber optic acoustic vibration sensor, a surface mobile quantum dot laser interferometer, and a pipe wall contact piezoelectric film sensor, which is used to collect pipeline vibration soundprint signals in real time.

3. The pipe network leakage self-learning diagnosis system based on voiceprint recognition according to claim 1 is characterized by: The signal preprocessing module performs noise reduction, filtering and multi-channel synchronous calibration on the original voiceprint signal to generate a standardized time-frequency domain mixed signal, and adopts magnetohydrodynamic noise suppression technology to offset the pipeline medium flow noise through the Lorentz force.

4. The pipe network leakage self-learning diagnosis system based on voiceprint recognition according to claim 1 is characterized by: The multimodal feature extraction module extracts the frequency domain energy distribution, time domain pulse characteristics and nonlinear resonance characteristics of the voiceprint signal based on the improved wavelet packet decomposition algorithm and Mel-frequency cepstral coefficients, performs multi-scale hybrid feature extraction, and adopts adaptive weighted fusion of time and frequency domain features.

5. The pipe network leakage self-learning diagnosis system based on voiceprint recognition according to claim 1 is characterized by: The self-learning diagnosis core module adopts a dynamic weighted residual network to fuse multimodal features. The dynamic weighted residual network introduces a channel attention mechanism to dynamically adjust the weight coefficients of features in different frequency bands. The network training adopts a transfer learning framework. The pre-training model is generated based on a public acoustic database and simulated leakage voiceprint data, or adopts one of the dual-channel deep confidence network architectures, including a main channel leakage pattern classifier and an auxiliary channel environmental noise discriminator, and a dynamic model update mechanism based on confidence evaluation.

6. The pipe network leakage self-learning diagnosis system based on voiceprint recognition according to claim 5 is characterized by: The dynamic weighted residual network retains deep features through residual connections, and the channel attention mechanism strengthens the weights of leakage-related frequency bands. Combined with incremental learning, the model can quickly adapt to new leakage patterns. The dual-channel deep belief network classifies leakage patterns through the main channel and identifies environmental noise through the auxiliary channel. The model is dynamically updated based on confidence evaluation. The knowledge distillation mechanism adopts teacher-student model interaction. The lightweight student model on the edge and the expert teacher model on the cloud perform knowledge distillation based on dynamic entropy values ​​to improve the efficiency of model iteration.

7. The pipe network leakage self-learning diagnosis system based on voiceprint recognition according to claim 1 is characterized by: The visualization interactive platform module includes integrated GIS map positioning, leakage level heat map, maintenance decision suggestion output, diagnosis result output, leakage probability heat map, and pipe health index three-dimensional cloud map.

8. The pipe network leakage self-learning diagnosis system based on voiceprint recognition according to claim 1 is characterized by: The edge computing integration module deploys embedded edge computing units at the network nodes to realize local feature extraction and primary diagnosis of voiceprint signals, and only uploads key feature data to the cloud, reducing network bandwidth occupancy. A lightweight diagnostic model is deployed at the edge, and model retraining and parameter distribution are performed on the cloud.

9. The pipe network leakage self-learning diagnosis system based on voiceprint recognition according to claim 1 is characterized by: The multi-source data fusion positioning module adopts the fusion method of voiceprint time difference positioning and vibration wave attenuation model, the improved generalized cross-correlation delay estimation method and the back propagation algorithm based on the attenuation model of material stress wave, combined with the real-time correction of pipeline network GIS topology data to achieve accurate positioning of the leakage point.

Images