Multi-source data fusion underground pipe network data intelligent detection method, module, system and medium

Abstract

The invention belongs to the technical field of data processing, and provides a multi-source data fusion underground pipe network data intelligent detection method, which comprises the following steps of: firstly, acquiring multi-source detection data of an underground pipe network; then, according to the multi-source detection data and an isolated forest algorithm based on Bayesian optimization, abnormal point information is determined; and finally inputting the abnormal point information into a pre-established pipeline state model, and determining a fault prediction value of the abnormal point of the underground pipe network. Wherein parameters of the pipeline state model are updated in real time according to an online variational inference algorithm; the online variational inference algorithm is provided with a first learning rate of a global parameter and a second learning rate of a local parameter; the first learning rate is used for reflecting common characteristics of the whole monitoring system or the pipeline network; the second learning rate is used for reflecting the change of the sensor along with the change of the underground environment. The differential learning rate is adopted for the global parameters and the local parameters, so that the pipeline state model is continuously updated in real time, and the safety risk of the underground pipe network in the future is accurately predicted.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and particularly relates to an intelligent detection method for underground pipeline network data with multi-source data fusion. Background Art

[0002] With the acceleration of the urbanization process, the underground pipeline network system has become increasingly complex, covering various types of pipelines such as water supply, drainage, gas, electricity, and communication. Due to the high concealment of the underground pipeline network, the long-term influence of corrosive media, and the susceptibility to electromagnetic interference, traditional detection methods are difficult to meet the requirements of high-precision and high-efficiency monitoring. There is an urgent need for an underground pipeline network detection technology that integrates multi-source data and improves the level of intelligence.

[0003] Traditional underground pipeline network detection mainly relies on single technical means, including: using ground penetrating radar to collect reflected wave characteristics for pipeline positioning and buried depth estimation; using equipment such as high-precision magnetometers to trace ferromagnetic pipelines based on electromagnetic induction; using piezoelectric sensors, distributed optical fibers, etc. to detect pipeline leakage through acoustic wave detection; relying on manual inspections or single-point monitoring of fixed sensors; obtaining prior attribute data such as pipeline type and material through geographic information systems, but lacking real-time dynamic analysis capabilities.

[0004] However, single technology is easily interfered by the complex underground environment and there is no cross-verification of multi-source data, resulting in large deviations in detection results. In the existing technology, there is also a lack of intelligent analysis and real-time feedback capabilities for multi-source data, and it is impossible to perform dynamic modeling and fault warning. Therefore, the accuracy of the current intelligent detection of underground pipeline network data is relatively low. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent detection method for underground pipeline network data with multi-source data fusion, aiming to solve the problem of relatively low accuracy in the intelligent detection of underground pipeline network data in the existing technology.

[0006] The first aspect of the embodiment of the present invention provides an intelligent detection method for underground pipeline network data with multi-source data fusion, including: Obtaining multi-source detection data of the underground pipeline network; Determining abnormal point information according to the multi-source detection data and the isolation forest algorithm based on Bayesian optimization; Inputting the abnormal point information into a pre-established pipeline state model to determine the fault prediction value of the abnormal points of the underground pipeline network; Among them, the parameters of the pipeline state model are updated in real time according to the online variational inference algorithm; the online variational inference algorithm is set with a first learning rate for global parameters and a second learning rate for local parameters; the first learning rate is used to reflect the common characteristics of the entire monitoring system or pipeline network; the second learning rate is used to reflect the changes caused by the sensors with the change of the underground environment.

[0007] In a possible implementation, the initial value of the second learning rate is greater than the initial value of the first learning rate. The method further includes: When the number of optimization rounds is greater than the first preset number of rounds and not greater than the second preset number of rounds, keep the first learning rate unchanged and control the second learning rate to decay exponentially; wherein, the second preset number of rounds is greater than the first preset number of rounds; When the number of optimization rounds is greater than the second preset number of rounds, freeze the local parameters and set the first learning rate to the minimum value.

[0008] In a possible implementation, after obtaining the multi-source detection data of the underground pipe network, the method further includes: Perform denoising processing on the multi-source detection data according to the extended Kalman filter algorithm; The method further includes: Determine the initial value of the second learning rate according to the covariance in the denoising process of the extended Kalman filter algorithm.

[0009] In a possible implementation, the online variational inference algorithm is also provided with a forgetting factor; the method further includes: Calculate the data change rate of the outlier according to the outlier information; Determine the forgetting factor according to the data change rate.

[0010] In a possible implementation, the online variational inference algorithm is also provided with a forgetting factor; the stages of intelligent detection of pipe network data include the corrosion detection stage and the leakage response stage; the method further includes: Calculate the data change rate of the outlier and determine the current stage according to the outlier information; Determine the forgetting factor according to the data change rate and the current stage.

[0011] In a possible implementation, after obtaining the multi-source detection data of the underground pipe network, the method further includes: Perform denoising processing on the multi-source detection data according to the extended Kalman filter algorithm; The method further includes: Determine the forgetting factor adjustment value for the corrosion detection stage according to the covariance in the denoising process of the extended Kalman filter algorithm; Calculate the consistency index according to the time series variance of each sensor data; Determine the forgetting factor adjustment value for the leakage response stage according to the consistency index.

[0012] In a possible implementation, the method further includes: Generate an alarm message when the fault prediction value is greater than the preset threshold.

[0013] In a second aspect of the embodiments of the present invention, a data processing module is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the intelligent detection method for underground pipeline network data with multi-source data fusion in the first aspect above are implemented.

[0014] In a third aspect of the embodiments of the present invention, an intelligent detection system for underground pipeline network is provided, which includes a data acquisition module and the data processing module in the second aspect above.

[0015] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the intelligent detection method for underground pipeline network data with multi-source data fusion in the first aspect above are implemented.

[0016] The intelligent detection method for underground pipeline network data with multi-source data fusion provided by the embodiments of the present invention first obtains multi-source detection data of the underground pipeline network; then determines abnormal point information according to the multi-source detection data and the isolation forest algorithm based on Bayesian optimization; finally, inputs the abnormal point information into a pre-established pipeline state model to determine the fault prediction value of the abnormal points of the underground pipeline network; wherein, the parameters of the pipeline state model are updated in real time according to the online variational inference algorithm; the online variational inference algorithm is provided with a first learning rate for global parameters and a second learning rate for local parameters; the first learning rate is used to reflect the common characteristics of the entire monitoring system or pipeline network; the second learning rate is used to reflect the changes caused by the sensors with the change of the underground environment. By adopting different learning rates for global parameters and local parameters, the pipeline state model is continuously updated in real time, so as to accurately predict the safety risks of the future underground pipeline network. Description of the Drawings

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 is a schematic structural diagram of the intelligent detection system for underground pipeline network provided by the embodiments of the present invention; Figure 2 is an implementation flowchart of the intelligent detection method for underground pipeline network data with multi-source data fusion provided by the embodiments of the present invention. Detailed Embodiments

[0019] In the following description, specific details such as specific system structures and technologies are presented for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present invention. However, those skilled in the art should understand that the present invention can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from interfering with the description of the present invention.

[0020] Figure 1 is a schematic structural diagram of an underground pipe network intelligent detection system provided by an embodiment of the present invention. As Figure 1 shown, in some embodiments, the underground pipe network intelligent detection system includes: a data acquisition module 11 and a data processing module 12.

[0021] The data acquisition module 11 integrates multiple types of sensors to cover multi-dimensional physical characteristics and prior attributes of the underground pipe network, realizes the acquisition of all-element data, and then sends it to the data processing module 12 to realize the intelligent detection of the underground pipe network.

[0022] Among them, the data acquisition module 11 includes various acquisition devices such as ground penetrating radar, terrestrial lidar, high-precision magnetometer, piezoelectric sensor, distributed optical fiber, acoustic wave detection device, infrared thermal imager (carried by unmanned aerial vehicle), and geographic information system.

[0023] The data processing module 12 specifically includes: Data preprocessing module: performs denoising, calibration, and format unification processing on the collected raw data.

[0024] Data analysis: performs fusion processing on data from different sources, and then identifies the type, location, depth, and status of the pipe network through intelligent analysis, and detects abnormal situations (such as breakage, leakage, etc.).

[0025] Data management module: stores, queries, and shares the detection data, and supports integration with other geographic information systems.

[0026] Figure 2 is an implementation flowchart of an intelligent detection method for underground pipe network data with multi-source data fusion provided by an embodiment of the present invention. As Figure 2 shown, in some embodiments, the intelligent detection method for underground pipe network data with multi-source data fusion includes: S210, obtaining multi-source detection data of the underground pipe network; S220, determining abnormal point information according to the multi-source detection data and the isolated forest algorithm based on Bayesian optimization; S230, inputting the abnormal point information into a pre-established pipeline state model to determine the fault prediction value of the abnormal points of the underground pipe network; Among them, the parameters of the pipeline state model are updated in real time according to the online variational inference algorithm; the online variational inference algorithm is set with a first learning rate for global parameters and a second learning rate for local parameters; the first learning rate is used to reflect the common characteristics of the entire monitoring system or pipeline network; the second learning rate is used to reflect the changes caused by the sensors with the change of the underground environment.

[0027] In the embodiments of the present invention, the ground penetrating radar collects the reflection wave time, amplitude and frequency characteristics for pipeline positioning and buried depth estimation. The terrestrial lidar extracts the surface deformation and the coordinates of the pipeline network entrance markers. The high-precision magnetometer collects the magnetic field gradient changes of ferromagnetic pipelines to track deeply buried metal pipelines. The piezoelectric sensors and distributed optical fibers can collect the acoustic fingerprint characteristics and vibration spectrum information of pipeline leaks for leakage location of water / gas pipelines. The infrared thermal imager monitors the abnormally hot areas on the ground surface for hot spot detection of heat supply pipeline leakage points. The geographic information system is used to provide key information such as pipeline type, pipe diameter, material, laying age, etc., and can also be used for data verification and prior knowledge supplementation.

[0028] In the embodiments of the present invention, after the original data is collected, the original data is preprocessed to eliminate noise and errors. The noise is eliminated and the errors are corrected through a multi-level processing process to ensure that the data quality meets the requirements of subsequent fusion analysis.

[0029] Underground pipelines require more complex algorithms to process noise, perform real-time analysis, or predictive maintenance due to complex environments (such as high concealment (buried depth, non-visible), long-term exposure to corrosive media (soil, groundwater), electromagnetic interference, signal transmission problems, etc.). At the same time, it is necessary to fuse multi-modal data such as pressure, vibration, and acoustics for cross-verification. In other fields, such as agricultural monitoring (temperature, humidity + light), the data correlation is simple, and the algorithms are mainly linear regression; in electricity, it is mostly multi-physical field coupling, such as the interweaving of electricity (current, voltage), heat (wire temperature), mechanics (transformer vibration), and chemistry (gases in oil) data, etc.

[0030] In underground pipeline monitoring, there are many sources of noise, such as vehicle vibration, soil displacement, construction activities, soil temperature, humidity, climate temperature, light, etc., while the monitoring environments in other fields are relatively stable. Pipeline monitoring requires fusing data of different physical quantities, such as pressure and strain, etc., while other fields such as electricity may pay more attention to multi-sensor data of the same physical quantity.

[0031] Based on this, the present invention adopts a multi-sensor joint denoising method for denoising processing. This method is a technical method that eliminates noise and improves data quality by integrating data from multiple sensors. Compared with single-sensor denoising, it has the following significant advantages: (1) Complementary noise characteristics: Multi-sensor data joint denoising can utilize the complementary nature of the noise characteristics of different sensors to comprehensively suppress various interferences.

[0032] (2) Improve the signal-to-noise ratio: The signal-to-noise ratio (SNR) of a single sensor is limited by hardware performance and environmental interference. Multi-sensor data fusion can significantly improve the SNR through weighted averaging or filtering algorithms.

[0033] (3) Improve the anti-interference ability: For example, in a strong electromagnetic interference environment, the magnetometer data may fail, but GPR and LiDAR can still provide reliable data.

[0034] (4) Improve fault tolerance: Even if some sensors fail, the target information can still be recovered through the data of other sensors. For example, if the DAS optical fiber breaks, the pipeline can still be located through the data of GPR and magnetometer.

[0035] (5) Improve data accuracy: Multi-sensor joint denoising can improve data accuracy through cross-validation and error compensation algorithms.

[0036] (6) Multi-sensor joint denoising can dynamically adjust weights according to environmental conditions and data quality.

[0037] (7) Multi-sensor joint denoising can dynamically adjust weights according to environmental conditions and data quality. For example, GPR + LiDAR + magnetometer: jointly extract electromagnetic reflection, geometric shape, and magnetic field gradient features to improve the accuracy of pipeline material identification; DAS + infrared thermal imager: jointly extract vibration spectrum and temperature distribution features to accurately locate leakage points.

[0038] The specific content of denoising is as follows: 1. Multi-sensor joint denoising 1.1 Extended Kalman Filter (EKF). The extended Kalman filter algorithm extends the Kalman filter to nonlinear systems and deals with nonlinear problems by linearizing the nonlinear system model. It uses Taylor expansion to approximate the nonlinear function and updates the state estimate and covariance matrix at each time step. Its state equation is: =

[0039] where, is the system state vector (such as GPR reflection wave amplitude, LiDAR point cloud coordinates), is the process noise (assumed to be zero-mean Gaussian noise with covariance matrix ), is the nonlinear state transition function.

[0040] Its measurement equation is:

[0041] where, is the sensor observation value, is the measurement noise (covariance matrix ). is the non - linear measurement function.

[0042] 1.2 Denoising Filtering Steps In a multi - sensor system, assume there are N sensors, and the measurement value of each sensor is .

[0043] First, initialize the state estimate and the covariance matrix , and use the state equation to predict the state and covariance at the next moment:

[0044]

[0045] where, is the Jacobian matrix of the state transition function , is the process noise covariance matrix.

[0046] For each sensor i , calculate the measurement residual:

[0047] At the same time, calculate the covariance of the measurement residual:

[0048] where, is the Jacobian matrix of the measurement function , is the sensor measurement noise covariance matrix.

[0049] Subsequently, calculate the Kalman gain:

[0050] Update the state estimate and covariance:

[0051]

[0052] Finally, perform a weighted average of the state estimates according to the measurement accuracy (covariance matrix of the measurement noise) of each sensor:

[0053]

[0054] After denoising, it is necessary to align multi-source data to make them consistent and comparable in terms of time, space, semantics, etc. And perform normalization processing on the multi-source data to eliminate differences in dimension, value range, etc. of the data, so that different types of data are comparable.

[0055] Among them, the alignment processing of multi-source data is specifically as follows: For spatial registration, it specifically includes rough registration and fine registration. The rough registration converts the coordinates of each sensor to a unified engineering coordinate system, and the fine registration uses the ICP (Iterative Closest Point, point cloud registration based on iterative closest point search) algorithm to optimize the registration parameters.

[0056] Exemplarily, when the spatial reference is unified, through coordinate transformation and feature anchor point matching, the error of multi-source data is controlled within the centimeter level, laying a foundation for subsequent analysis. If the data involves the time dimension (such as dynamic monitoring), the acquisition timestamps of different devices are aligned through clock synchronization technology.

[0057] The present invention can also perform feature dimension fusion, and specifically can use technologies such as Fourier transform and machine learning to achieve cross-modal association of time domain / frequency domain / spatial domain data, improving the accuracy of anomaly recognition.

[0058] Then perform normalization processing on the multi-source data to eliminate differences in dimension, value range, etc. of the data, so that different types of data are comparable. Specifically, normalization can be performed through the following formula: Min-Max:

[0059] This method is applicable to continuous variables such as GPR amplitude and LiDAR point cloud intensity.

[0060] Z-Score standardization:

[0061] This method is applicable to data such as magnetic field gradient and temperature that follow a Gaussian distribution.

[0062] Next, it is necessary to detect and remove outliers from the normalized data. For pipeline monitoring sensors (pressure / flow / vibration, etc.) with characteristics of high dimension (multi-parameter coupling monitoring, usually 3D or more than 4D), non-Gaussian distribution (multi-modal / skewed distribution affected by working conditions fluctuations), and local anomalies (sudden leakage is manifested as local outliers rather than global offset), the present invention adopts the Isolation Forest algorithm, and for the parameters in the Isolation Forest, the Bayesian optimization method is used. Compared with other methods, its advantages are as follows: (1) Because its time complexity is , so it is suitable for processing large-scale data.

[0063] (2) It does not require data annotation and can be directly applied to unlabeled data sets.

[0064] (3) It has good robustness to noise and outliers.

[0065] (4) The degree of abnormality can be intuitively explained by the path length.

[0066] Randomly select a feature from the data set j , assuming that the data set has d features, randomly select one of them j . Among the minimum value j and the maximum value of the feature , randomly select a splitting value p .

[0067] According to the splitting value p , divide the data set into two subsets: left subset: points where the value of the feature j is less than p . Right subset: points where the value of the feature j is greater than or equal to p .

[0068] Recursively execute the above steps for each subset until one of the following termination conditions is met: there is only one point in the subset; the tree reaches the maximum depth (usually set to , where is the size of the data set).

[0069] For the data point x , its path length is the number of edges passed from the root node to the leaf node. Construct multiple isolation trees (usually 100 - 200 trees) to calculate x the average path length E in all trees.

[0070] For a data set containing n points, the approximation of the average path length is:

[0071] where is the harmonic number.

[0072] The anomaly score x of the data point is defined as:

[0073] Among them, if , then x it is very likely to be an abnormal point. If , then x it is a normal point. If , then x it is very likely to be a normal point (longer path length).

[0074] In the leakage monitoring of water supply pipe networks, due to the multi-factor coupling characteristics of the failure mechanism, as shown in the above case, leakage events are often triggered by the coupling effects of multiple physical quantities such as pressure fluctuations, flow anomalies, temperature changes, and vibration frequency offsets. However, in existing monitoring systems, the potential correlations between various sensing parameters are still uncertain. Based on this, the technical solution of this patent innovatively introduces a sensor data independence analysis mechanism: When the sensor data set satisfies the independence assumption, a univariate isolation forest algorithm is used for parallel training, and abnormal point detection is achieved through the weighted average of abnormal scores; if significant data correlations (Pearson coefficient ρ > 0.6) are detected, a multi-dimensional feature vector space is constructed. Using the PCA principal component analysis method, the dimension is reduced to 90% variance retention through PCA. The present invention allows the model to be updated gradually without the need to load all data at once. That is, if the data distribution changes, it is necessary to refit the PCA and isolation forest models regularly. This not only avoids the curse of dimensionality caused by feature redundancy but also ensures the model generalization ability under different working conditions.

[0075] In some embodiments, the initial value of the second learning rate is greater than the initial value of the first learning rate, and the method further includes: when the number of optimization rounds is greater than the first preset number of rounds and not greater than the second preset number of rounds, keeping the first learning rate unchanged and controlling the second learning rate to decay exponentially; wherein, the second preset number of rounds is greater than the first preset number of rounds; when the number of optimization rounds is greater than the second preset number of rounds, freezing the local parameters and setting the first learning rate to the minimum value.

[0076] In the embodiments of the present invention, data from different sources will be fused and processed, and then the type, location, depth, and status of the pipe network will be identified through intelligent analysis, and abnormal conditions (such as breakage, leakage, etc.) will be detected. After determining some abnormal points of the pipeline, repairs or further monitoring will be carried out for these abnormal points. Specifically, an online variational inference algorithm can be introduced to effectively reduce the influence of sensor noise and data conflicts through real-time probability modeling and dynamic parameter update of multi-source heterogeneous data, thereby generating a monitoring result with higher confidence. That is, by optimizing a simple variational distribution q(z) to approximate the complex true posterior distribution p(z∣x), where z represents the latent variable (such as pipeline-related parameters) and x is the observed data (sensor signal). The goal is to maximize the evidence lower bound (ELBO):

[0077] Underground pipelines (such as water pipes and oil pipes) may leak or suffer structural damage due to aging, corrosion, or external force damage. Traditional detection methods (such as manual inspections and fixed sensors) are inefficient and unable to perform real-time analysis. Online variational inference can be used to process sensor data in real time, dynamically update the pipeline state model, and quickly detect anomalies.

[0078] First, define the sensor data and the variables to be detected.

[0079] Input data (observed data ) Pressure sensor data: The pressure value inside the pipeline changes over time (the pressure drops suddenly during a leak).

[0080] Acoustic signal: Leaks or corrosion will emit acoustic waves at specific frequencies, and the vibration energy (the vibration increases due to the leak) Flow data: The difference between the outlet flow and the inlet flow may indicate a leak.

[0081] Temperature / chemical sensor: Corrosion may be accompanied by temperature changes or the release of specific chemical substances. Infrared region temperature

[0082] Latent variables (unknown quantities to be detected ) Leak location (1D coordinate or 3D coordinate x = (x, y, z)).

[0083] Leak intensity (the size of the leak hole, such as the leak hole area A). It directly affects the pressure change and flow difference inside the pipeline.

[0084] Degree of corrosion (the reduction in the wall thickness of the pipe).

[0085] Pipeline health status (normal, mild corrosion, severe corrosion).

[0086] Sometimes, the leak rate can also be referred to and the sensor noise level .

[0087] Then, establish a probability model.

[0088] Assume that the pipeline state conforms to physical laws (such as the fluid mechanics model), and construct a joint probability distribution: (x, z) = (x|z)p(z) Assume that the sensor data follows a Gaussian distribution, and the mean is determined by the leak intensity A. The variance is the sensor noise variance. The initial probabilities of leaks and corrosion (for example, older pipelines have a higher corrosion probability). The dynamic prior can also be a prior for the leak rate based on pipeline pressure and material parameters .

[0089] Next, select the variational distribution.

[0090] Approximate the true posterior distribution with a simple distribution. For example: Leak location: Use a Gaussian distribution (the mean represents the most likely location, and the variance represents the uncertainty). Corrosion degree: Use a gamma distribution (positive values, suitable for representing gradually accumulating corrosion). As , where the variational parameter . When updating, adjust the gamma distribution parameters . Pipeline health status: Use a categorical distribution (probabilities for the three categories of normal, minor, and severe).

[0091] Finally, execute the online update process (taking data per minute as an example) Assume the pipeline health status is "normal", the corrosion degree is 0, and the leakage probability is 0. Set the learning rate (such as ). Receive real-time sensor data. For example, the pressure sensor shows that the pressure suddenly drops from 5 MPa to 4.8 MPa.

[0092] Among them, for local parameter optimization (leakage / corrosion speculation at the current time point): Input: Data such as current pressure and acoustic waves.

[0093] Calculation: Based on the current global model, speculate whether there is a leak at the current moment and the corrosion degree.

[0094] For example: A sudden drop in pressure may correspond to an increase in the leakage probability from 5% to 60%.

[0095] If the corrosion degree increases slowly, it may be updated from 0.1 mm to 0.12 mm.

[0096] Due to the particularity of pipeline monitoring, global parameters reflect the common characteristics of the entire monitoring system or pipeline network. For example: The global elastic modulus of the pipeline material, the overall pressure-flow model parameters of fluid transmission; Local parameters are susceptible to short-term environmental changes (such as sudden temperature changes) or local environmental interferences (such as soil stress concentration in a certain pipeline), and need to be dynamically adjusted. It may also be the zero drift or sensitivity factor of a certain sensor (such as the mV / V deviation of a pressure sensor), the loss coefficients of each fiber optic sensor, etc., which need to be adjusted in a timely manner. Local parameters in pipeline monitoring are mostly aimed at the changes caused by multi-source sensors themselves due to surface or underground environmental changes. Monitoring in other fields, such as power monitoring, may target the status of a single device and is relatively stable at the same time.

[0097] For this invention, global parameters (global parameters are for the entire dataset) and local parameters (local parameters are for each data point or sensor features) are considered. The learning rate of global parameters may require a slower decay, while the learning rate of local parameters can be adjusted more quickly. This invention comprehensively considers global parameters and local parameters and adjusts them together to obtain a better convergence effect.

[0098] The following gives an implementation example to illustrate the learning rate, but it is not a limitation: (1)Initial stage (the first 10 rounds): A high local learning rate (such as sensor sensitivity, local correction factor, etc.) quickly fits the characteristics of each pipe segment.

[0099] A low global learning rate (such as material properties, basic model parameters, corrosion rate, etc.) maintains the stability of the model.

[0100] (2)Middle stage (10 - 50 rounds): Apply exponential decay to the high local learning rate (×0.95 per round) to avoid overfitting noise.

[0101] Keep the low global learning rate unchanged to strengthen global feature extraction.

[0102] (3)Final stage (after 50 rounds): Freeze the local parameters and fine-tune the global parameters (such as 0.0001) to improve generalization.

[0103] High learning rate local adjustment: Reduce the fitting error of pipe segment specificity.

[0104] Low learning rate global optimization: Improve the cross-region prediction consistency.

[0105] Initial stage (less data): The learning rate is relatively high (such as ), quickly absorbing new information.

[0106] Final stage (stable data): The learning rate is reduced (such as 0.01), preventing noise interference.

[0107] In some embodiments, after obtaining multi-source detection data of the underground pipe network, the method further includes: denoising the multi-source detection data according to the extended Kalman filter algorithm; the method further includes: determining the initial value of the second learning rate according to the covariance during the denoising process of the extended Kalman filter algorithm.

[0108] In the embodiments of the present invention, the multi-sensor joint denoising effect also has a certain impact on the learning rate, and this process is mainly the adjustment of the local parameter learning rate. Among them, independent learning rates can be assigned to different sensor channels, and the weights are inversely proportional to the covariance of each channel. The learning rates of multiple sensors are dynamically adjusted according to the Kalman filter covariance. When the covariance is large (high data noise), the learning rate is reduced to stabilize the update; when the covariance is small, the learning rate is increased to accelerate convergence. The deviation between the observation and the prediction is used to evaluate the degree of data abnormality and trigger the attenuation or reset of the learning rate. For example, in the case of pipeline leakage, the residual sequence of the Kalman filter may mutate, and in this case, the model parameters will be reset or the learning rate will increase suddenly.

[0109] In some embodiments, the online variational inference algorithm is also provided with a forgetting factor; the method further includes: calculating the data change rate of the outlier according to the outlier information; determining the forgetting factor according to the data change rate.

[0110] In the embodiments of the present invention, the traditional update formula for global model update (taking corrosion intensity as an example) is: Current corrosion intensity estimate The present invention introduces a forgetting factor on the basis of the learning rate for update, wherein the forgetting factor is adaptively and dynamically adjusted based on the data change rate: Current corrosion intensity estimate For example, if the corrosion degree follows a gamma distribution, the posterior distribution parameters of the corrosion degree ( , ) are updated to:

[0111]

[0112] The present invention considers the change situation of data points and introduces variance for measurement. The monitoring index is the variance of sensor data within a sliding window .

[0113] Adjustment rule:

[0114] Among them, when the data mutates ( is large), (such as 0.6), the old state is forgotten more quickly. When the data is stable ( is small), (such as 0.95), the long-term memory is retained.

[0115] Piecewise constant forgetting factor: Corrosion detection stage (slow change): For example = 0.95, emphasizing the long-term trend leakage response stage (rapid change): such as = 0.7, quickly tracking mutations. If leakage signals are detected continuously for multiple times, the global model will gradually increase the leakage probability.

[0116] In some embodiments, the online variational inference algorithm is further provided with a forgetting factor; the stages of intelligent detection of pipeline network data include a corrosion detection stage and a leakage response stage; the method further includes: calculating the data change rate of the abnormal point and determining the current stage according to the abnormal point information; determining the forgetting factor according to the data change rate and the current stage.

[0117] In some embodiments, after obtaining the multi-source detection data of the underground pipeline network, the method further includes: denoising the multi-source detection data according to the extended Kalman filter algorithm; the method further includes: determining the forgetting factor adjustment value in the corrosion detection stage according to the covariance in the denoising process of the extended Kalman filter algorithm; calculating a consistency index according to the time series variance of each sensor data; determining the forgetting factor adjustment value in the leakage response stage according to the consistency index.

[0118] In the embodiments of the present invention, the online variational inference algorithm realizes dynamic modeling of the pipeline network state by setting a forgetting factor (γ), combines the different requirements of the corrosion detection stage and the leakage response stage, and forms an adaptive adjustment mechanism based on data characteristics. The pipeline network state changes slowly (such as the wall thickness of the pipe wall decreasing year by year), the sensor data shows low variance and progressive trend, and the noise is mainly periodic interference (such as seasonal soil moisture changes). The pipeline network state suddenly mutates (such as a sudden drop in pressure caused by a pipeline rupture), and the sensor data shows high variance and multi-parameter coupling anomalies (pressure, vibration, temperature synchronously deviate from the mean value). Anomaly detection is performed on the preprocessed multi-source data (such as pressure, flow rate, vibration) through the isolation forest algorithm, and an anomaly score s ( x ) (the value range is 0-1, and a score > 0.8 is determined as an abnormal point). Data change rate calculation: Change_rate = | x t - x t−n | / x t−n × 100% Among them, x t is the data value at the current moment, x t−n is the average value of the previous n cycles ( n take 10 - 30, adjusted according to the pipeline network type).

[0119] In the corrosion detection stage, if continuousm Within one cycle (e.g., m = 20), if the density of abnormal points < 5% and the data change rate < threshold 1 (e.g., 3%), it is determined to be in the normal operation state and enters the corrosion detection stage.

[0120] In the leakage response stage, if the density of abnormal points > 10% or the data change rate > threshold 2 (e.g., 15%) within a single cycle, it is determined to be in the abnormal burst state and enters the leakage response stage.

[0121] During the denoising process of the extended Kalman filter, the trace of the covariance matrix P (trace( P )) reflects the data noise intensity. The larger the value of the trace, the higher the noise. The adjustment formula is: γ 腐蚀 = γ base - k 1 * trace( P max ) * trace( P ) Where: γ base = 0.95 (basic forgetting factor); k 1 = 0.2 (adjustment coefficient to ensure γ corrosion ≥ 0.8); trace( P max ) is the historical maximum covariance trace value. The higher the noise, γ the smaller, enhancing the trust in the current data and avoiding the corrosion trend being masked by noise.

[0122] Define the sensor data mutation consistency index: Consistency_index = N The number of synchronously mutated sensors (if data from ≥ 2 types of sensors among pressure, DAS vibration, and infrared temperature exceed their respective variance thresholds simultaneously, it is determined as "synchronous mutation").

[0123] Adjustment formula: γ leakage = γ min + k 2 * Consistency_index Where, γ min = 0.6 (minimum forgetting factor); k 2 = 0.4 (adjustment coefficient to ensure γ leakage ≤ 0.8). The higher the consistency index, γ the larger, strengthening the collaborative trust of multi-sensor data and accelerating the model's response to leakage.

[0124] When switching from the corrosion detection stage to the leakage response stage, if the data change rate > threshold 2 but the consistency index < 0.5 (single sensor mutation), start a buffer window (e.g., 5 cycles) to avoid misjudgment.

[0125] When recovering from the leakage response stage to the corrosion detection stage, it is necessary to continuously m for several cycles to satisfy that the abnormal point density < 5% and the consistency index < 0.3, to prevent frequent switching caused by short-term interference.

[0126] When entering the leakage response stage, reset γ the leakage to the initial value of 0.6, clear the weight accumulation of the historical corrosion trend, and focus on the current abnormality.

[0127] When returning to the corrosion detection stage, re-initialize γ corrosion based on prior data (such as pipeline material, laying age), for example, γ set the base of old pipelines to 0.9 and that of new pipelines to 0.95 to improve the utilization rate of prior knowledge.

[0128] In some embodiments, the method further includes: generating an alarm message when the fault prediction value is greater than a preset threshold.

[0129] In the embodiments of the present invention, when the leakage probability exceeds 90% or the corrosion degree exceeds the safety threshold, an alarm is triggered; that is, the decision output, the current leakage probability, if it exceeds the threshold, an alarm is triggered. Combining the time difference of the acoustic wave sensor, the most likely leakage position is determined through the mean value of the variational distribution. After the alarm, manual or robotic inspections are carried out to confirm whether there is leakage / corrosion. If there is a false alarm (such as a sudden drop in pressure due to valve operation), new information is added to the model to correct the prior distribution.

[0130] The present invention has the following advantages: (1) Real-time performance: The model is updated every minute, earlier than traditional regular inspections.

[0131] (2) Anti-noise: A single sensor failure will not cause a false alarm (through probability smoothing).

[0132] (3) Self-adaptability: After the pipeline ages, the model automatically increases the prior probability of corrosion.

[0133] In some embodiments, a scene classifier is further set in the pipeline state model. The scene classifier is specifically a neural network or a preset formula. After inputting the historical fault prediction value and the data of each sensor into the scene classifier, the scene classifier will classify the current scene. The specific scene types can be high-load operation, slow corrosion development, accelerating corrosion, sudden pipe burst, interface leakage, etc.

[0134] After identifying the scenario, the corresponding dynamic adjustment rules for the first learning rate, the second learning rate, and the forgetting factor can be determined according to the current scenario, aiming to improve the adaptability of the model to different working conditions through parameter optimization.

[0135] The following uses several examples to illustrate the adjustment process using the preset formula, but it is not restrictive.

[0136] In the high-load operation scenario, the sensor data fluctuates periodically. The adjustment goal is to quickly respond to transient changes and avoid misjudging as abnormal. Specifically, the following formula can be used to capture the transient overload characteristics: =min( ×1.5, 0.1) where p 0 is the initial value of the learning rate, p t is the adjusted learning rate. By setting a threshold, it is ensured that the learning rate does not exceed 0.1 to avoid overfitting. This learning rate can be the first learning rate or the second learning rate. In this scenario, it is better to adjust the second learning rate according to this formula.

[0137] In the corrosion acceleration scenario, the adjustment goal is to reduce the learning rate to prevent noise from interfering with the long-term trend. Specifically, the following formula can be used to stably track the slow corrosion process: =max (η0×0.7, 0.001) By setting a threshold, it is ensured that the learning rate is not lower than 0.001 to maintain the continuous update ability of the model. This learning rate can be the first learning rate or the second learning rate. In this scenario, it is better to adjust the first learning rate according to this formula.

[0138] In the sudden pipe burst scenario, the adjustment goal is to reset the learning rate to the maximum value, that is =1.0, full-weight learning in an emergency state. This learning rate can be the first learning rate or the second learning rate. In this scenario, it is better to adjust the second learning rate according to this formula.

[0139] In the interface leakage scenario, the adjustment goal is a medium learning rate to balance the steady state and micro-changes, that is = , to avoid missing small leakage signals.

[0140] In the slow corrosion development scenario, the adjustment goal is to strengthen the weight of historical data. The long-term corrosion law can be obtained through the following formula: =min( +0.2, 0.99) where is the adjusted value of the forgetting factor, is the initial value of the forgetting factor, with an upper limit of 0.99 to preserve long-term corrosion trends.

[0141] In the case of sudden pipe burst, the adjustment goal is to sharply reduce λ and quickly forget the historical state, that is, =0.1 to focus on current emergencies.

[0142] In the corrosion acceleration scenario, the adjustment goal is to moderately reduce λ and weaken the impact of outdated corrosion data. The following formula can be used to prevent old data from covering up the acceleration trend: =max( -0.3,0.5) It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0143] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. An intelligent detection method for underground pipe network data with multi-source data fusion, characterized in that, Including: Obtain multi-source detection data of the underground pipe network; Determine abnormal point information according to the multi-source detection data and the Isolation Forest algorithm based on Bayesian optimization; Input the abnormal point information into a pre-established pipeline status model to determine the fault prediction value of the abnormal points in the underground pipe network; Among them, the parameters of the pipeline status model are updated in real time according to the online variational inference algorithm; the online variational inference algorithm is set with a first learning rate for global parameters and a second learning rate for local parameters; the first learning rate is used to reflect the common characteristics of the entire monitoring system or pipeline network; The second learning rate is used to reflect the changes caused by the sensors due to changes in the underground environment.

2. The intelligent detection method for underground pipeline network data with multi-source data fusion according to claim 1, characterized in that, The initial value of the second learning rate is greater than the initial value of the first learning rate, and the method further includes: When the number of optimization rounds is greater than the first preset round and not greater than the second preset round, keep the first learning rate unchanged and control the second learning rate to decay exponentially; where the second preset round is greater than the first preset round; When the number of optimization rounds is greater than the second preset round, freeze the local parameters and set the first learning rate to the minimum value.

3. The intelligent detection method for underground pipe network data with multi-source data fusion according to claim 2, wherein After obtaining the multi-source detection data of the underground pipe network, the method further includes: Denoise the multi-source detection data according to the Extended Kalman Filter algorithm; The method further includes: Determine the initial value of the second learning rate according to the covariance during the denoising process of the Extended Kalman Filter algorithm.

4. The intelligent detection method for underground pipeline network data with multi-source data fusion according to claim 1, wherein, The online variational inference algorithm is also set with a forgetting factor; the method further includes: Calculate the data change rate of the abnormal points according to the abnormal point information; Determine the forgetting factor according to the data change rate.

5. The intelligent detection method for underground pipeline network data with multi-source data fusion according to claim 3, characterized in that, The online variational inference algorithm is also set with a forgetting factor; the stages of intelligent detection of pipe network data include the corrosion detection stage and the leakage response stage; the method further includes: Calculate the data change rate of the abnormal points according to the abnormal point information and determine the current stage; Determine the forgetting factor according to the data change rate and the current stage.

6. The intelligent detection method for underground pipe network data with multi-source data fusion according to claim 4, characterized in that, After obtaining the multi-source detection data of the underground pipe network, the method further includes: Denoise the multi-source detection data according to the Extended Kalman Filter algorithm; The method further includes: Determine the forgetting factor adjustment value for the corrosion detection stage according to the covariance during the denoising process of the Extended Kalman Filter algorithm; Calculate the consistency index according to the time series variance of each sensor data; Determine the forgetting factor adjustment value for the leakage response stage according to the consistency index.

7. The intelligent detection method for underground pipeline network data with multi-source data fusion according to any one of claims 1-6, characterized in that, The method further includes: Generate an alarm message when the fault prediction value is greater than the preset threshold.

8. A data processing module, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent detection method for underground pipe network data with multi-source data fusion described in any one of claims 1 to 7 above.

9. An intelligent detection system for underground pipe networks, characterized in that, Including a data acquisition module and the data processing module described in claim 8 above.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps of the intelligent detection method for underground pipe network data with multi-source data fusion described in any one of claims 1 to 7 above.

Images