A multi-modal sensor fusion pipeline micro-leak precise positioning system and method thereof

Abstract

The application discloses a kind of multi-modal sensor fusion's pipeline micro-leak precise positioning system and method, it is related to pipeline safety monitoring field.The system includes multi-modal sensor array, edge positioning host and cloud diagnosis platform.Multi-modal sensor array acquires acoustic and negative pressure wave mode sensing signal, after edge positioning host pre-processes signal, its multi-modal anomaly probability discrimination module fuses depth time-frequency feature and statistical waveform feature, adaptive space-time filtering and leak point positioning module adopt particle filtering algorithm based on attention mechanism, fuse the above information and carry out iterative convergence, obtain the optimal leak point position estimation, leakage event comprehensive research and judge module combines real-time working condition with historical model output positioning result, cloud diagnosis platform converges data and is optimized, forms intelligent closed loop.The application realizes the high sensitivity, high reliability and high-precision positioning of pipeline micro-leak.

Description

Technical Field

[0001] This invention relates to the field of pipeline safety monitoring, and in particular to a multimodal sensor fusion-based system and method for precise location of pipeline micro-leaks. Background Technology

[0002] Existing pipeline leak monitoring technologies primarily rely on single-sensor modes, which have inherent limitations. Methods based on negative pressure waves are insensitive to the weak pressure gradients generated by minute leaks and are easily interfered with in complex pipe networks or under fluctuating flow rates, resulting in low location accuracy and high false alarm rates. While methods based on acoustic waves are sensitive to micro-leaks, the signals attenuate rapidly in pipelines, propagate complexly, and are easily submerged by environmental noise, making it difficult to achieve high-precision long-distance location. Furthermore, single-modal systems lack the ability to verify leak events in multiple dimensions, failing to reliably distinguish between actual leaks and operational disturbances. Currently, many multi-sensor fusion studies remain at the level of simple weighted fusion at the data or feature layers, failing to establish a unified spatiotemporal benchmark and lacking in-depth modeling of the inherent coupling mechanisms and propagation differences of multi-modal signals during leaks. This results in limited fusion effects, especially when dealing with low signal-to-noise ratio micro-leak events, where the improvement in location accuracy is not significant, and the error remains large.

[0003] Therefore, there is an urgent need for a system and method that can deeply integrate multimodal sensing signals, establish a unified spatiotemporal reference, and possess adaptive signal enhancement and collaborative reasoning capabilities, in order to achieve highly sensitive, reliable, and accurate localization of pipeline micro-leakage events. Summary of the Invention

[0004] This invention proposes a multimodal sensor fusion-based system and method for precise location of micro-leaks in pipelines. Addressing the issues of low sensitivity, poor anti-interference capability, and unsatisfactory shallow multimodal fusion effects of existing single-sensor modal location technologies, this invention constructs a three-level collaborative architecture consisting of a sensor array, an edge positioning host, and a cloud-based diagnostic platform. This architecture enables intelligent processing throughout the entire process, from micro-leak event perception and deep fusion of multimodal signals to precise location of the leak point.

[0005] This invention first provides a multimodal sensor fusion-based precise positioning system for pipeline micro-leaks, which includes: a multimodal sensor array, an edge positioning host, and a cloud-based diagnostic platform;

[0006] A multimodal sensor array consists of multiple sensor nodes arranged at equal intervals along the monitored pipeline or at key nodes. Each sensor node synchronously acquires multimodal sensor signals of the pipeline's operating status. The multimodal sensor signals include acoustic mode sensor signals and negative pressure wave mode sensor signals.

[0007] The edge positioning host connects to the pipeline's real-time operating parameters and historical leak model library, and communicates with the multimodal sensor array to process the multimodal sensor signals and generate a micro-leak event location report containing leak confidence and location coordinates. The host integrates a signal acquisition and adaptive preprocessing module, a multimodal anomaly probability discrimination module, an adaptive spatiotemporal filtering and leak point location module, a leak event comprehensive analysis module, and a location result output and diagnosis module.

[0008] Signal acquisition and adaptive preprocessing module: Receives multimodal sensing signals, adds a unified time stamp to them based on a high-precision synchronous clock, and performs adaptive noise reduction and signal enhancement processing to generate aligned high-quality preprocessed signals;

[0009] Multimodal anomaly probability discrimination module: Taking high-quality preprocessed signal as input, it extracts the depth time-frequency features of acoustic modal sensing signals and the statistical waveform features of negative pressure wave modal sensing signals in parallel; using an improved DS evidence theory framework, it integrates the two types of features to independently discriminate the preliminary probability of micro-leakage events occurring at each sensing node, and outputs the multimodal local anomaly probability of each node and its uncertainty measure.

[0010] The adaptive spatiotemporal filtering and leak point location module, along with its uncertainty metric, is used as input to construct a leak probability field in the pipeline space. A particle filtering algorithm based on an attention mechanism is employed to perform spatiotemporal filtering along the pipeline axis, suppressing false probability peaks. The module also integrates the arrival time difference of acoustic modal sensing signals with the propagation velocity estimation of negative pressure wave modal sensing signals, and iteratively converges to obtain the optimal leak point location estimate and its probability distribution.

[0011] Leakage event comprehensive analysis module: Receives the optimal leak point location estimate and its probability distribution, and combines it with real-time pipeline operating parameters and historical leak model library to conduct a comprehensive confidence assessment and conflict resolution, and outputs the final leak event analysis conclusion, including the exact leak point coordinates, leak confidence level and possible leak aperture range.

[0012] Location result output and diagnosis module: The module formats the leakage event assessment conclusions to generate a micro-leakage event location report, and uploads it to the cloud diagnosis platform through the communication network, while driving the local audible and visual alarm device.

[0013] The cloud-based diagnostic platform, deployed on a remote server, communicates with the edge positioning host via a cellular network. This platform continuously receives and aggregates micro-leakage event location reports and their associated multimodal sensor signal fragments, forming a pipeline operation database. Based on this database, the platform optimizes the historical leak model library and performs centralized training and parameter iteration on the improved DS evidence theory framework used in the multimodal anomaly probability discrimination module, the particle filtering algorithm used in the adaptive spatiotemporal filtering and leak point location module, and the evaluation rules of the leak event comprehensive judgment module. The optimized model parameters and algorithm rules are then distributed to the corresponding edge positioning host.

[0014] Furthermore, the multimodal anomaly probability discrimination module calculates the multimodal local anomaly probability, specifically including the following steps:

[0015] Step M1: For the sensing node, using a high-quality preprocessed signal as input, the acoustic modal sensing signal is generated by short-time Fourier transform and then input into a lightweight convolutional neural network to extract the deep time-frequency feature vector; for the negative pressure wave modal sensing signal, the waveform kurtosis, approximate entropy and wavelet energy entropy within the sliding time window are calculated to form a statistical waveform feature vector.

[0016] Step M2: Input the depth time-frequency feature vector and the statistical waveform feature vector into two independent evidence generation networks respectively; each evidence generation network outputs a basic probability allocation function that reflects the modality of the evidence it belongs to;

[0017] Step M3: Using the improved DS evidence theory framework, the two basic probability allocation functions obtained in step M2 are fused to obtain the fused basic probability allocation function;

[0018] Step M4: Calculate the confidence level and similarity level of the micro-leakage event at the sensing node based on the fusion basic probability allocation function; take the midpoint value of the confidence interval determined by the confidence level and similarity level as the multimodal local anomaly probability output of the node, and output the width of the confidence interval as the uncertainty measure of the anomaly probability.

[0019] Furthermore, the adaptive spatiotemporal filtering and leak point localization module employs an attention-based particle filter algorithm for leak point localization, specifically including the following steps:

[0020] Step L1: Using the multimodal local anomaly probability and its uncertainty measure of each sensing node as input, define the state vector as the location of the leak point, and define the observation vector as the multimodal local anomaly probability sequence and its uncertainty measure of all nodes; establish a state transition model describing the evolution of the state vector, and an observation likelihood model describing the probability of the observation vector appearing under a given state vector.

[0021] Step L2: Generate an initial set of particles uniformly along the pipe axis. Each particle contains a state vector and an initial weight. All particles have the same weight.

[0022] Step L3: Taking the particle set as input, randomly perturb the state vector of each particle according to the state transition model to complete the state prediction and output the predicted particle set.

[0023] Step L4: Using the predicted particle set and the current observation vector as input, an attention mechanism is introduced to calculate the spatial distance between each particle and each sensing node, and an attention weight that is inversely proportional to the distance and the node uncertainty metric is constructed. Using this attention weight, the information of each node is fused to calculate the observation likelihood value corresponding to each particle, and the particle weight is updated accordingly. The particle set with updated weights is then output.

[0024] Step L5: Using the particle set with updated weights as input, calculate the weighted average of all particle state vectors as the current leak point location estimate, and resample to duplicate high-weight particles and eliminate low-weight particles to generate a new set of equally weighted particles.

[0025] Step L6: Using the new equally weighted particle set as the starting point for the next iteration, repeat steps L3 to L5. When the change in the estimated location of the leak point is less than the threshold or the maximum number of iterations is reached, the iteration stops. Output the weighted average location of the final particle set as the optimal leak point location estimate, and calculate its probability distribution.

[0026] This invention also provides a method for accurate localization of micro-leaks in pipelines using multimodal sensor fusion applied to the above-mentioned system, comprising the following steps:

[0027] Step A: Through a multi-modal sensor array deployed along the monitored pipeline, each sensor node synchronously collects acoustic modal sensing signals and negative pressure wave modal sensing signals of the pipeline operation.

[0028] Step B: The edge positioning host receives multimodal sensor signals, performs time synchronization and adaptive preprocessing, and generates aligned high-quality preprocessed signals;

[0029] Step C: Extract acoustic depth time-frequency features and negative pressure wave statistical waveform features in parallel from the high-quality preprocessed signal, and fuse them using the improved DS evidence theory to calculate the multimodal local anomaly probability and its uncertainty measure at each sensing node;

[0030] Step D: Using the multimodal local anomaly probability and its uncertainty metric of all nodes as input, a particle filter algorithm based on attention mechanism is used to perform spatiotemporal filtering along the pipeline axis to suppress false probability peaks; at the same time, the arrival time difference of acoustic modal sensing signals and the propagation velocity estimation of negative pressure wave modal sensing signals are fused, and the optimal leak point location estimation and its probability distribution are obtained through multiple rounds of iteration convergence.

[0031] Step E: Combine the real-time operating parameters of the accessed pipeline with the historical leakage model library to conduct credibility assessment and conflict resolution, and output the final leakage event judgment conclusion.

[0032] Step F: Format the leakage event assessment conclusion into a micro-leakage event location report, activate the local audible and visual alarm, and upload the report to the cloud diagnostic platform via the communication network;

[0033] Step G: The cloud-based diagnostic platform gathers micro-leakage event location reports and their associated raw signal fragments from various edge positioning hosts, performs centralized training and iterative optimization of the historical leakage model library and various algorithm parameters, and then distributes the optimized models and parameters to the edge positioning hosts.

[0034] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0035] To address the shortcomings of single-mode signals, such as susceptibility to interference and insignificant features, this invention deploys a multi-mode sensing array and designs a signal acquisition and adaptive preprocessing module. This enables high-precision synchronous acquisition of acoustic and negative pressure wave mode signals. Simultaneously, adaptive noise reduction and enhancement algorithms are used to specifically improve the signal-to-noise ratio of weak leakage signals. Furthermore, a unified high-precision time stamp is added to all signals to generate aligned, high-quality preprocessed signals. This effectively overcomes the inherent defects of single-mode signals and lays a solid data foundation for subsequent micro-leakage identification and localization.

[0036] This invention achieves highly robust probabilistic identification of node-level micro-leakage events by designing a multimodal anomaly probability discrimination module and constructing an improved DS evidence theory fusion framework. The module deeply integrates the deep time-frequency features of acoustic signals and the statistical waveform features of negative pressure wave signals, and introduces intermodal conflict factors to dynamically adjust the fusion weights, significantly improving the accuracy and robustness of single-point anomaly discrimination and effectively reducing the false alarm rate caused by operating condition disturbances. At the same time, the output multimodal local anomaly probability and its uncertainty measure provide a reliable input with confidence information for subsequent spatial positioning.

[0037] This invention achieves high-precision spatial positioning of leak points and a self-evolving intelligent closed loop by combining an adaptive spatiotemporal filtering and a leak point location module based on an attention mechanism-based particle filtering algorithm, along with a leak event comprehensive analysis module and a cloud-based diagnostic platform. Specifically, the attention particle filtering algorithm effectively focuses on observation information from high-confidence, nearby nodes, suppressing false probability peaks caused by noise, and quickly and stably converging to the true leak point location, significantly improving positioning accuracy and reliability. Meanwhile, the edge-side comprehensive analysis module integrates physical models and real-time operating conditions to output reliable leak event analysis conclusions. The cloud-based diagnostic platform utilizes multi-node historical data for big data analysis, continuously optimizing the leak identification model and location algorithm parameters, and iteratively improving the performance of edge devices through downlink updates, enabling the system to possess self-evolving capabilities and maintain high monitoring performance over the long term. Attached Figure Description

[0038] Figure 1 This is a structural block diagram of a multimodal sensing fusion-based precise positioning system for pipeline micro-leakage proposed in this invention.

[0039] Figure 2 This is a flowchart of the multimodal anomaly probability discrimination module proposed in Embodiment 2 of the present invention. Detailed Implementation

[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] Example 1:

[0042] according to Figure 1 This invention provides a multimodal sensor fusion-based precise location system for pipeline micro-leakage, applicable to the monitoring of micro-leakage events in oil, gas, and water pipelines. The system includes:

[0043] Multimodal sensor array: A sensor node is arranged every 200 m along the pipeline to be monitored. Each node integrates a high-sensitivity piezoelectric acoustic wave sensor (model PCB102B06, frequency response 0.5-15 kHz, sensitivity 50 mV / g) and a high-frequency response dynamic pressure sensor (model KellerPR-25Y, range 0-2 MPa, resonant frequency >50 kHz). The sensors are encapsulated in a waterproof and explosion-proof housing and connected to the edge positioning host via armored cables. Data acquisition of all sensor nodes is triggered uniformly by the host, with a synchronization accuracy better than 10 μs. The acoustic wave sensor and the dynamic pressure sensor respectively acquire acoustic modal sensing signals and negative pressure wave modal sensing signals.

[0044] Edge positioning host: Adopting an industrial-grade rugged design with an IP67 protection rating, the main control unit is based on the NVIDIA Jetson Xavier NX core, equipped with a 24-bit synchronous acquisition card (sampling rate up to 100 kHz / channel), 16 analog input channels, 4 digital I / O channels, as well as dual gigabit Ethernet ports and 4G / 5G communication modules; The host has built-in signal acquisition and adaptive preprocessing modules, multimodal anomaly probability discrimination modules, adaptive spatiotemporal filtering and leak point location modules, leak event comprehensive analysis modules, and location result output and diagnosis modules, and is installed in valve chambers along the pipeline;

[0045] Cloud-based diagnostic platform: Deployed on Huawei Cloud ECS (Elastic Cloud Server, configured with 8 cores, 16 GB of memory, and 1 TB of cloud disk); This platform adopts a microservice architecture, providing data access services, storage services (time series database InfluxDB), algorithm model management services (based on the PyTorch framework), and web visualization services. It maintains a long connection with each edge positioning host via the MQTT protocol, receives micro-leakage event location reports in real time, and periodically receives multimodal sensor signals for model training.

[0046] Example 2:

[0047] This embodiment is based on Embodiment 1. In this embodiment, the modules within the edge positioning host work together through the following process to detect and locate micro-leakage events:

[0048] Signal acquisition and adaptive preprocessing module: Based on the PPS signal generated by the GPS-disciplined oven-controlled crystal oscillator (OCXO) inside the host, all acquisition channels are controlled to start synchronously at a sampling rate of 51.2 kHz; for acoustic modal sensing signals, bandpass filtering of 0.5-12 kHz is performed, and then an adaptive joint noise reduction algorithm based on wavelet threshold and empirical mode decomposition is used to suppress broadband noise from pumps, valves, etc.; for negative pressure wave modal sensing signals, low-pass filtering of 0.01-10 Hz is performed, and a moving average filter is used to smooth the trend term caused by flow fluctuations; finally, a uniform timestamp (accuracy 1 μs) is added to all channel data to generate aligned high-quality preprocessed signals, which are then buffered in a circular buffer.

[0049] Multimodal anomaly probability discrimination module, such as Figure 2 As shown, this module extracts the latest 1-second high-quality preprocessed signal from the buffer for processing in a 1-second cycle, specifically including the following steps:

[0050] Step M1: Using a high-quality preprocessed signal as input, calculate the short-time Fourier transform (window length 1024, overlap 512) for the acoustic modal sensing signal of each sensing node to generate a 128×128 time-frequency spectrum; input the time-frequency spectrum into a pre-trained lightweight convolutional neural network to output a deep time-frequency feature vector; for the negative pressure wave modal sensing signal of the same node, calculate its waveform kurtosis, approximate entropy, and wavelet energy entropy within a 1-second window to form a statistical waveform feature vector;

[0051] Step M2: Input the depth time-frequency feature vector into the acoustic evidence generation network to obtain the basic probability allocation function corresponding to the acoustic modal sensing signal; input the statistical waveform feature vector into the negative pressure wave evidence generation network to obtain the basic probability allocation function corresponding to the negative pressure wave modal sensing signal;

[0052] Step M3: Calculate the cosine similarity of the two feature vectors and define the intermodal conflict factor accordingly; adopt the improved DS evidence theory framework, which introduces the intermodal conflict factor to dynamically adjust the evidence weights in the fusion process of the two basic probability allocation functions, and obtain the fusion basic probability allocation function of the node.

[0053] Step M4: Calculate the confidence level and similarity level of the micro-leakage event at the sensor node based on the fusion basic probability allocation function; take the midpoint value of the confidence interval determined by the confidence level and similarity level as the multimodal local anomaly probability output of the sensor node, and output the width of the confidence interval as its uncertainty measure. Finally, output the multimodal local anomaly probability sequence and its uncertainty measure sequence of all sensor nodes.

[0054] The adaptive spatiotemporal filtering and leak point localization module receives a multimodal local anomaly probability sequence and an uncertainty metric sequence. Given the length of the monitored pipe segment and the number of sensor nodes, it employs an attention-based particle filtering algorithm. The process includes initialization, prediction, updating, and resampling steps. During particle weight updates, an attention mechanism is innovatively introduced, allowing the localization algorithm to focus more on observation information from high-confidence, nearby nodes. Specifically, for each particle in the particle set (representing a candidate leak location), the weighted likelihood function upon which its weight update depends is expressed as: ;in, Representing the The weight of each particle; This is the chain multiplication operator; The probability density function representing the normal distribution; For nodes Observed multimodal local anomaly probability; This is calculated based on the physical attenuation model of the leakage signal when the leakage point is located... At the node The theoretical probability of an anomaly at the location is expected. It is a node The measurement of observational uncertainty; It is a dynamically calculated attention weight.

[0055] Leakage event comprehensive analysis module: Receives the optimal leak point location estimate and its probability distribution, queries the real-time operating parameters (pressure, flow rate) of the pipeline section, if the location point is located at a location prone to noise such as an elbow or tee, and the probability distribution is relatively large, the confidence level is appropriately lowered, the optimal leak point location estimate and its probability distribution are matched with the historical leak model library, the leak aperture range is initially estimated, and after comprehensive judgment, the leakage event analysis conclusion is output.

[0056] Location result output and diagnosis module: The leakage event assessment conclusion is formatted into a micro-leakage event location report in JSON format, which is uploaded to the cloud diagnosis platform via 4G network, and at the same time triggers the local audible and visual alarm (buzzer and red LED flashing) on ​​the host.

[0057] Example 3:

[0058] This embodiment is based on Embodiment 2. In this embodiment, the adaptive spatiotemporal filtering and leakage point localization module adopts a particle filtering algorithm based on an attention mechanism. The specific steps are as follows:

[0059] Step L1: Define the current state of the system as the coordinates of the leak point along the pipeline axis; the observation vector is composed of the multimodal local anomaly probability sequence of all sensor nodes at the current moment and their corresponding uncertainty metric sequence.

[0060] Step L2: Within the pipeline monitoring range, a certain number of particles are randomly generated according to a uniform distribution. Each particle represents a candidate leak point location and is assigned the same initial weight.

[0061] Step L3: Based on the state transition model, apply a Gaussian process noise with zero mean and a given variance to the position of each particle to simulate the small uncertainty change in the location of the leak point;

[0062] Step L4: Assign differentiated attention to the observation data of different sensor nodes, and calculate the likelihood value of each particle to update its weight. First, calculate the spatial distance from each particle's position to each sensor node, and then calculate the attention weight through a softmax function based on the distance and the uncertainty measure of each node's observation, so that nodes that are close and have low uncertainty receive higher attention.

[0063] Step L5: Based on the updated particle weight set, calculate the weighted average estimate of the leakage point location at the current moment, determine the degree of particle weight degradation, and if the number of effective particles is lower than the set threshold, perform a resampling operation to copy high-weight particles and eliminate low-weight particles, thereby obtaining a new set of equal-weight particles for the next iteration.

[0064] Step L6: Repeat steps L3 to L5 until the preset number of iterations is reached or the change in the state estimate converges to an allowable range. Output the weighted average position obtained from the last iteration as the optimal leak point location estimate, and calculate its covariance based on the distribution of the final particle set as a measure of positioning uncertainty.

[0065] Example 4:

[0066] This embodiment demonstrates a specific application scenario: a water pipeline with a diameter of DN800 and a total length of 2 kilometers has a tiny leak with a diameter of about 2 millimeters at a distance of 1000 meters from the starting point. The system has 11 sensing nodes (including the nodes at both ends) arranged along the line, with a node spacing of 200 meters.

[0067] The edge positioning host continuously monitors the system. When a leak occurs, the nodes closest to the leak point (located at 800 meters and 1200 meters) are the first to show abnormal responses. The acoustic modal sensing signal shows a significant increase in energy in the 2000-5000 Hz frequency band, while the negative pressure wave modal sensing signal shows weak negative spikes. The signal acquisition and adaptive preprocessing module synchronously acquires and denoises the original signal to generate a time-aligned high-quality preprocessed signal.

[0068] The multimodal anomaly probability discrimination module analyzes high-quality preprocessed signals. The multimodal local anomaly probabilities output by the nodes 800 meters and 1200 meters closest to the leak point reach 0.85 and 0.78, respectively, with low uncertainty (0.1). The anomaly probabilities of other more distant nodes are lower (0.1-0.3), while the uncertainty is relatively higher.

[0069] The adaptive spatiotemporal filtering and leak point localization module was then activated. After the particle swarm was initialized, during the prediction update iteration process, due to the introduction of the attention mechanism, the 800-meter and 1200-meter nodes with high probability and low uncertainty were given higher weights. At the same time, the false probability peak (probability 0.4, uncertainty 0.4) caused by random noise at 400 meters was effectively suppressed. After 8 iterations, the particle swarm quickly converged to the 980-meter to 1020-meter range. Finally, the optimal leak point location estimate was 1003 meters with a standard deviation of 8 meters.

[0070] The comprehensive analysis module for leakage events, combined with the stable real-time operating parameters of the pipe section, determined that the event was a high-confidence leak, estimating the leak diameter to be approximately 1-3 mm. The location results output and the diagnostic module generated a structured micro-leakage event location report, which was then uploaded to the cloud diagnostic platform.

[0071] After receiving the report, the cloud-based diagnostic platform archives the event. The platform regularly collects typical signal fragments uploaded by each edge host (including confirmed leakage and interference event samples) and performs incremental training and fine-tuning on the convolutional neural network feature extractor and evidence generation network in the multimodal anomaly probability discrimination module to further improve the model's generalization ability. The optimized model parameters are then distributed to each edge positioning host along the line, thereby achieving continuous evolution and improvement of the system's identification and positioning performance.

[0072] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A multi-modal sensor fusion-based pipeline microleak precise positioning system, characterized in that, The system includes: a multimodal sensor array, an edge positioning host, and a cloud-based diagnostic platform; The multimodal sensor array consists of multiple sensor nodes deployed along the monitored pipeline. Each sensor node synchronously acquires multimodal sensor signals of the pipeline's operating status. The multimodal sensor signals include acoustic mode sensor signals and negative pressure wave mode sensor signals. The edge positioning host connects to the pipeline's real-time operating parameters and historical leak model library, and communicates with the multimodal sensor array to process the multimodal sensor signals and generate a micro-leak event location report containing leak confidence and location coordinates. The edge positioning host integrates a signal acquisition and adaptive preprocessing module, a multimodal anomaly probability discrimination module, an adaptive spatiotemporal filtering and leak point location module, a leak event comprehensive analysis module, and a location result output and diagnosis module. The cloud-based diagnostic platform, deployed on a remote server, communicates with the edge positioning host via a communication network. It aggregates micro-leakage event location reports and their associated multimodal sensor signal fragments. It also performs centralized training and parameter iteration on the algorithm models used in the historical leak model library, the multimodal anomaly probability discrimination module, the adaptive spatiotemporal filtering and leak point location module, and the leak event comprehensive analysis module. The optimized model parameters and algorithm rules are then distributed to the corresponding edge positioning host. Among them, the signal acquisition and adaptive preprocessing module receives multimodal sensing signals, adds a unified time stamp to them based on a high-precision synchronous clock, and performs adaptive noise reduction and signal enhancement processing to generate aligned high-quality preprocessed signals. The multimodal anomaly probability discrimination module takes high-quality preprocessed signals as input, extracts the depth time-frequency features of the acoustic modal sensing signals and the statistical waveform features of the negative pressure wave modal sensing signals in parallel, and fuses them using an improved DS evidence theory framework to calculate the multimodal local anomaly probability and its uncertainty measure at each sensing node. The improved DS evidence theory framework defines the intermodal conflict factor by calculating the similarity between the depth time-frequency feature vector and the statistical waveform feature vector, and introduces the intermodal conflict factor to dynamically adjust the evidence weights during the fusion of the two basic probability allocation functions. The adaptive spatiotemporal filtering and leak point location module takes the multimodal local anomaly probability and its uncertainty metric of all nodes as input, and uses a particle filtering algorithm based on attention mechanism to perform spatiotemporal filtering along the pipeline axis. It integrates the arrival time difference of acoustic modal sensing signals and the propagation speed estimate of negative pressure wave modal sensing signals, and iteratively converges to obtain the optimal leak point location estimate and its probability distribution. The comprehensive analysis module for leakage events receives the optimal leak point location estimate and its probability distribution, and combines it with real-time pipeline operating parameters and historical leak model library to conduct a comprehensive credibility assessment and conflict resolution, and outputs the final leakage event analysis conclusion. The location result output and diagnosis module formats the leakage event assessment conclusions into a micro-leakage event location report, uploads it to the cloud diagnosis platform, and drives local alarms.

2. The multimodal sensing fusion-based precise positioning system for pipeline micro-leakage according to claim 1, characterized in that, The multimodal anomaly probability discrimination module calculates the probability of multimodal local anomalies and its uncertainty measure, specifically including the following steps: Step M1: For the sensing node, using a high-quality preprocessed signal as input, generate a time-frequency spectrum for the acoustic modal sensing signal and input it into a convolutional neural network to extract a deep time-frequency feature vector; for the negative pressure wave modal sensing signal, calculate its waveform statistical features within the sliding time window to form a statistical waveform feature vector. Step M2: Input the depth time-frequency feature vector and the statistical waveform feature vector into two independent evidence generation networks respectively, and output the corresponding basic probability assignment function; Step M3: Using the improved DS evidence theory framework, the two basic probability allocation functions obtained in step M2 are fused to obtain the fused basic probability allocation function; Step M4: Calculate the confidence level and similarity of the micro-leakage event occurring at the sensing node based on the fusion basic probability allocation function; take the midpoint value of the confidence interval determined by the confidence level and similarity as the multimodal local anomaly probability output of the sensing node, and output the width of the confidence interval as the uncertainty measure of the anomaly probability.

3. The multimodal sensing fusion-based precise positioning system for pipeline micro-leakage according to claim 1, characterized in that, The adaptive spatiotemporal filtering and leak point localization module employs an attention-based particle filtering algorithm for leak point localization, specifically including the following steps: Step L1: Using the multimodal local anomaly probability and its uncertainty measure of each sensing node as input, define the state vector as the location of the leak point, define the observation vector as the multimodal local anomaly probability sequence and its uncertainty measure of all nodes, and establish the state transition model and the observation likelihood model. Step L2: Generate an initial set of particles uniformly along the pipe axis, with each particle containing a state vector and an initial weight; Step L3: Taking the particle set as input, randomly perturb the state vector of each particle according to the state transition model to complete the state prediction and output the predicted particle set. Step L4: Using the predicted particle set and the current observation vector as input, an attention mechanism is introduced to calculate the spatial distance between each particle and each sensor node, construct an attention weight that is inversely proportional to the distance and the uncertainty measure, use this attention weight to fuse the information of each sensor node, calculate the observation likelihood value corresponding to each particle, and update the particle weight accordingly. Step L5: Using the particle set with updated weights as input, calculate the weighted average of all particle state vectors as the current leak point location estimate, and perform resampling to generate a new equally weighted particle set; Step L6: Using the new equally weighted particle set as the starting point for the next iteration, repeat steps L3 to L5. When the change in the estimated location of the leak point is less than the threshold or the maximum number of iterations is reached, the iteration stops, and the weighted average location of the final particle set is output as the optimal leak point location estimate.

4. The multimodal sensing fusion-based precise positioning system for pipeline micro-leakage according to claim 3, characterized in that, In step L4, the particle weights are updated based on a weighted likelihood function, which is calculated by exponentiating the corrected observation matching degree of each sensor node with its corresponding dynamic attention weight and multiplying them together.

5. The multimodal sensing fusion-based precise positioning system for pipeline micro-leakage according to claim 4, characterized in that, The weighted likelihood function is expressed as: ； in, Representing the The weight of each particle; This is the chain multiplication operator; The probability density function representing the normal distribution; For nodes Observed multimodal local anomaly probability; This is calculated based on the physical attenuation model of the leakage signal when the leakage point is located... At the node The theoretical probability of an anomaly at the location is expected. It is a node The measurement of observational uncertainty; It is a dynamically calculated attention weight.

6. The multimodal sensing fusion-based precise positioning system for pipeline micro-leakage according to claim 1, characterized in that, The cloud-based diagnostic platform performs centralized training and parameter iteration on the parameters of the evidence generation network in the improved DS evidence theory framework used in the multimodal anomaly probability discrimination module, and the particle filter algorithm used in the adaptive spatiotemporal filtering and leakage point location module.

Images