A central heating network pressure monitoring method and system based on artificial intelligence

Abstract

The application discloses a kind of based on artificial intelligence's central heating network pressure monitoring method and system, it is related to heating system monitoring technical field, including, first in pipe network layout optical fiber and point pressure sensor, acoustic strain and pressure data are collected;By the same frequency alignment and Delaunay triangulation, construct digital twin grid, reconstruct continuous pressure field;Residual error of reconstructed pressure and measured value at monitoring node is calculated;Based on pipe network topology, construct graph wavelet operator, extract multi-scale residual feature to form feature matrix;Characteristic is input into pre-trained spectral graph wavelet neural network, and output normalized anomaly score;Finally, according to the score, locate fault pipe section and hierarchical alarm.The whole process combines physical modeling and deep learning, realizes closed-loop monitoring from data acquisition to fault diagnosis, significantly improves the precision and timeliness of pipe network anomaly detection.The application does not need artificial detection and monitoring process is more flexible.

Description

Technical Field

[0001] This invention relates to the field of heating system monitoring technology, and in particular to a method and system for monitoring the pressure of centralized heating network pipes based on artificial intelligence. Background Technology

[0002] Currently, the heating and power generation pipeline network of the heating company covers a wide area and has many nodes. During operation, it is necessary to consider the coordinated operation of heating load and power generation equipment. Traditional pressure control relies on manual periodic inspection of pressure gauges and manual adjustment of valve openings. However, due to limitations in the frequency of inspections and differences in personnel experience, it is difficult to respond to dynamic changes in pipeline pressure in real time. Especially during the peak heating season in winter or when the power generation load fluctuates, the complex layout of the pipeline network with many branches and large differences in pipe diameter can easily cause local pressure over-limits or under-limits, resulting in insufficient heating for distant users, cavitation of heat exchange station equipment, and even safety hazards such as leakage at pipeline connections.

[0003] Existing control systems generally employ single-point pressure monitoring and independent equipment control modes, lacking a linkage analysis and coordinated adjustment mechanism for network-wide pressure data. Pressure sensors are sparsely deployed and communication is lagging, meaning actuators (such as regulating valves and circulating pumps) can only start and stop based on local setpoints, failing to perform global optimization based on the overall pressure distribution of the pipeline network. When the power generation load at the heat source changes or the flow rate at the heat user changes abruptly, pressure imbalance problems often require several hours for manual troubleshooting and repair, accompanied by multiple drawbacks such as ineffective energy loss, increased equipment wear and tear, and rising user complaint rates. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based method for monitoring the pressure of centralized heating networks to solve the problems of sparse deployment of pressure sensors and communication lag in existing point pressure monitoring technologies, as well as the problems of relying more on manual inspection and long identification time.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for monitoring the pressure of centralized heating network pipes based on artificial intelligence, which includes collecting acoustic echo, micro-strain curve and measured pressure value data to construct an acoustic strain-pressure dataset.

[0008] The acoustic strain-pressure dataset is aligned to the same frequency and mapped onto a digital twin geometric mesh constructed using triangulation to reconstruct a continuous spatial pressure field.

[0009] In a continuous spatial pressure field, the corresponding tetrahedral element is located for each predefined monitoring node. The centroid coordinates of the tetrahedral element are calculated, and the reconstructed pressure value of the node is obtained by linear interpolation of the reconstructed pressure value of the tetrahedral vertex. The pressure residual of the node is obtained by combining the measured pressure value.

[0010] An approximate graph wavelet operator is constructed, and the pressure residual is convolved and projected at each scale to extract multi-scale residual signal features. The projection results at each scale are then stitched together in node order to form a graph wavelet multi-scale feature matrix.

[0011] The spectral wavelet neural network is trained based on historical data. The original score is obtained by inputting the spectral wavelet multi-scale feature matrix. The original score is then normalized by the Sigmoid function to obtain the normalized anomaly score.

[0012] Based on standardized anomaly scoring screening and faulty pipe segments located through topology, risk classification and alarm are performed.

[0013] As a preferred embodiment of the artificial intelligence-based centralized heating network pressure monitoring method of the present invention, the method of collecting acoustic echo, micro-strain curve and measured pressure value data to construct an acoustic strain-pressure dataset includes laying single-mode optical fibers along the outer wall of the main pipeline and branch pipe sections, and arranging detection units and point pressure sensors at both ends and nodes of the single-mode optical fibers.

[0014] Periodic scanning was initiated to obtain acoustic echoes, micro-strain curves, and measured pressure values ​​with uniform time-series labels, and an acoustic strain-pressure dataset was constructed.

[0015] As a preferred embodiment of the artificial intelligence-based centralized heating network pressure monitoring method of the present invention, the following steps are included: The acoustic strain-pressure dataset is aligned to the same frequency, and the acoustic strain-pressure dataset is mapped to a digital twin geometric grid constructed using triangulation to reconstruct a continuous spatial pressure field. This includes: acquiring the acoustic strain-pressure dataset and performing denoising and frequency alignment processing on the acoustic echo and micro-strain curves; constructing a digital twin geometric grid using Delaunay triangulation; mapping the denoised and corrected acoustic echo and micro-strain curves to the measured pressure values ​​on the grid nodes to obtain discrete pressure and strain field quantities; reconstructing based on a weighted joint least squares regularization method for discrete pressure and strain field quantities; obtaining the continuous pressure value of each node on the grid through numerical iteration; and reconstructing a continuous spatial pressure field on the digital twin geometric grid.

[0016] As a preferred embodiment of the artificial intelligence-based centralized heating network pressure monitoring method of the present invention, the step of calculating the centroid coordinates of the tetrahedral element and performing linear interpolation on the reconstructed pressure values ​​of the tetrahedral vertices to obtain the reconstructed pressure values ​​of the nodes includes: locating the tetrahedral element to which each monitoring node belongs through an octree index and Delaunay adjacency query; then calculating the centroid coordinates of the tetrahedral element and performing weighted linear interpolation on the pressure values ​​of the tetrahedral vertices in combination with the micro-strain differences between the monitoring point and the vertex to obtain the weighted reconstructed pressure value of each monitoring node; finally, subtracting the weighted reconstructed pressure value from the corresponding measured pressure point by point and summarizing them in node order to generate the pressure residual of the monitoring node.

[0017] As a preferred embodiment of the artificial intelligence-based centralized heating network pressure monitoring method of the present invention, the step of extracting multi-scale residual signal features by convolutional projection of the pressure residual at each scale and stitching the projection results of each scale into a graph wavelet multi-scale feature matrix in node order includes: constructing a weighted undirected graph based on the inverse ratio of pipe segment resistance between nodes; constructing a graph wavelet operator by using Gaussian spectral wavelet function and Chebyshev polynomial approximation at multiple preset scales; performing convolutional projection of the residual vector at each scale to extract multi-scale residual signal features; and stitching the projection results of each scale into a graph wavelet multi-scale feature matrix in node order.

[0018] As a preferred embodiment of the artificial intelligence-based centralized heating network pressure monitoring method of the present invention, the method of obtaining the original score by inputting the graph wavelet multi-scale feature matrix and normalizing the original score by the Sigmoid function to obtain the normalized anomaly score includes inputting the graph wavelet multi-scale feature matrix into a spectral graph wavelet neural network trained with historical data. When the spectral graph wavelet neural network is running online, it performs two layers of graph convolution and nonlinear mapping on the input features and outputs the original score, which is then normalized by the Sigmoid function to obtain the normalized anomaly score.

[0019] As a preferred embodiment of the artificial intelligence-based centralized heating network pressure monitoring method described in this invention, wherein:

[0020] Secondly, the present invention provides a centralized heating network pressure monitoring system based on artificial intelligence, including a data acquisition module, a grid reconstruction module, a residual calculation module, a feature extraction module, an anomaly scoring module, and an alarm module.

[0021] The data acquisition module is used to collect acoustic echo, micro-strain curve and measured pressure value data to construct an acoustic strain-pressure dataset.

[0022] The mesh reconstruction module is used to perform frequency-aligned processing on the acoustic strain-pressure dataset, mapping the acoustic strain-pressure dataset to a digital twin geometric mesh constructed using triangulation, and reconstructing a continuous spatial pressure field.

[0023] The residual calculation module is used to locate the corresponding tetrahedral element for each predefined monitoring node in a continuous spatial pressure field, solve the centroid coordinates of the tetrahedral element, and perform linear interpolation on the reconstructed pressure value of the tetrahedral vertex to obtain the reconstructed pressure value of the node. Combined with the measured pressure value, the pressure residual of the node is obtained.

[0024] The feature extraction module is used to approximate the construction of the graph wavelet operator, perform convolution projection on the residual vector at each scale to extract multi-scale residual signal features, and concatenate the projection results at each scale in node order to form a graph wavelet multi-scale feature matrix.

[0025] The anomaly scoring module is used to train a spectrogram wavelet neural network based on historical data. The input spectrogram wavelet multi-scale feature matrix is ​​used to obtain the original score, and the original score is normalized by the Sigmoid function to obtain the normalized anomaly score.

[0026] The alarm module is used to filter based on standardized anomaly scores and locate faulty pipe segments through topology, and to perform risk-level alarms.

[0027] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the artificial intelligence-based centralized heating network pressure monitoring method described in the first aspect of the present invention.

[0028] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the artificial intelligence-based centralized heating network pressure monitoring method described in the first aspect of the present invention.

[0029] The beneficial effects of this invention are as follows: By aligning the acoustic strain-pressure dataset with the same frequency, this invention maps the data onto a triangulated digital twin geometric grid to reconstruct a continuous spatial pressure field, achieving accurate characterization of pipeline pressure distribution and providing a benchmark for subsequent residual analysis, thus realizing high-resolution dynamic monitoring. By approximating the construction of a graph wavelet operator, the residual vector is convolved and projected at multiple scales and concatenated into a graph wavelet multi-scale feature matrix, enabling multi-scale anomaly feature extraction and providing multi-dimensional input for anomaly scoring and pipe segment location, making the entire detection process more sensitive and robust. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of a centralized heating network pressure monitoring method based on artificial intelligence.

[0032] Figure 2 This is a schematic diagram of a centralized heating network pressure monitoring system based on artificial intelligence.

[0033] Figure 3 This is a flowchart of the residual calculation process for a centralized heating network pressure monitoring method based on artificial intelligence.

[0034] Figure 4 This is a flowchart of an anomaly detection process for a centralized heating network pressure monitoring method based on artificial intelligence. Detailed Implementation

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0036] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0037] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0038] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for monitoring the pressure of a centralized heating network based on artificial intelligence, including the following steps:

[0039] S1: Collect acoustic echo, micro-strain curve and measured pressure data to construct an acoustic strain-pressure dataset.

[0040] Specifically, the process of collecting acoustic echo, micro-strain curves and measured pressure data to construct an acoustic strain-pressure dataset includes laying single-mode optical fibers along the outer wall of the main pipeline and branch pipe sections, and arranging detection units and point pressure sensors at both ends and nodes of the single-mode optical fibers.

[0041] Periodic scanning was initiated to obtain acoustic echoes, micro-strain curves, and measured pressure values ​​with uniform time-series labels, and an acoustic strain-pressure dataset was constructed.

[0042] Furthermore, the detection unit is usually selected as C-OTDR (Coherent Optical Time Domain Reflectometry). C-OTDR can provide spatially resolved vibration and strain information along the pipeline, and can also reflect dynamic changes by combining time series analysis. It is a key means of real-time monitoring of the entire pipeline with "zero blind spots".

[0043] It should be noted that the C-OTDR periodically emits laser pulses to each measuring point on the optical fiber and receives the scattered echo signals along the fiber length. A differential operator is used to calculate the micro-strain curve, which reflects the minute deformation of the pipeline under thermal expansion or external impact. Simultaneously, the pressure sensor on site outputs the measured pressure values ​​in the same time sequence. The acoustic vibration waveform, micro-strain curve, and measured pressure values, after filtering, noise reduction, and time delay correction, are combined with a PTP timestamp and then assembled into an acoustic strain-pressure dataset.

[0044] S2: The acoustic strain-pressure dataset is aligned to the same frequency and mapped to a digital twin geometric mesh constructed using triangulation to reconstruct a continuous spatial pressure field.

[0045] Specifically, the acoustic strain-pressure dataset is frequency-aligned and mapped onto a digital twin geometric mesh constructed using triangulation to reconstruct a continuous spatial pressure field. This process involves obtaining the acoustic strain-pressure dataset, denoising and aligning the acoustic echoes and micro-strain curves, constructing a digital twin geometric mesh using Delaunay triangulation, mapping the denoised and corrected acoustic echoes and micro-strain curves with measured pressure values ​​onto the mesh nodes to obtain discrete pressure and strain field quantities, reconstructing the discrete pressure and strain field quantities using a weighted joint least squares regularization method, and obtaining the continuous pressure value of each node on the mesh through numerical iteration. This process reconstructs a continuous spatial pressure field on the digital twin geometric mesh.

[0046] Furthermore, the acoustic strain-pressure dataset was acquired, and the acoustic echoes and micro-strain curves underwent denoising and frequency synchronization processing. The denoising process consisted of two stages: the first stage used a fourth-order Butterworth bandpass filter, and the second stage used the Daubechies 4 wavelet basis for denoising. After denoising, frequency synchronization processing was performed. First, the data from different modes were time-aligned to eliminate clock deviations caused by the acquisition equipment and transmission links. For each pair of acoustic echoes and micro-strain curves, the cross-correlation function was calculated, and the maximum correlation delay was found. Then, the wavelet-denoised signals were shifted as a whole to ensure precise overlap on the time axis. Simultaneously, all point pressure sensor data (sampling point coordinates x...) were... j The measured pressure values ​​are also calibrated to the same time reference through the PTP clock synchronization subsystem, ultimately forming consistent multimodal timing data.

[0047] In the synchronously corrected micro-strain curve, the curve value at each sampling point reflects the cumulative strain at the pipe location. To characterize instantaneous dynamic changes, the local deformation rate ε is used. j The difference in microstrain between two adjacent moments is used as the deformation rate of the node to quantify the instantaneous rate of change of the micro deformation of the pipeline.

[0048] A digital twin geometric mesh is constructed using Delaunay triangulation based on the 3D mesh of the BIM pipeline network. The coordinates x of each sampling point are... j Measured pressure value P j With the deformation rate ε at the monitoring point j Let j be the monitoring point node. First, calculate the coordinates of the monitoring point node and all nodes in the digital twin geometric mesh. Let M be the total number of grid nodes and k be the number of grid nodes. Find the grid node with the smallest distance. Then, construct the set of pressure values ​​for each grid node by using the measured pressure values ​​of the sampling points with the smallest distance from the grid node. Similarly, construct the set of deformation rates of the monitoring points for each grid node. When the same grid node corresponds to multiple sampling points, average the pressure values ​​of all sampling points to obtain the grid pressure field vector b of the grid node, expressed as:

[0049] b = [b1, ..., b M ] T ;

[0050] Where T is the transpose.

[0051] Correspondingly, the strain values ​​at all sampling points are also averaged to obtain the micro-strain field vector f of the grid nodes, which is expressed as:

[0052] f = [f1, ..., f M ] T ;

[0053] Subsequently, a joint least squares objective function J(P) with weights for microstrain and pressure deviation is constructed, expressed as:

[0054]

[0055] Wherein, P(x k (x) represents the continuous pressure field to be determined at the grid node coordinates. k The value at b k Let w be the grid pressure field vector of the k-th grid node. b (k) represents the weights of the fitted terms, and the weight function is defined as w. b (k)=1+γ|f k |, γ is the strain weighting coefficient, γ>0, and δ is the pressure deviation weighting coefficient, δ≤0.5. For x k The pressure gradient vector at a given location is used to calculate the nodal pressure gradient using the central difference method. The regularization coefficient λ is determined by the L-curve method, balancing fitting accuracy and gradient consistency. k Let J(P) be the micro-strain field vector at the k-th node. For J(P) with respect to the pressure vector p = [P(x1), ..., P(x...]... M )] T Taking the derivative and setting it to zero, we obtain the weighted normal equation, expressed as:

[0056] (A T W b A+λD T W f D)p=A T W b b+λD T W f f;

[0057] in This is the measurement point mapping matrix, which represents the correspondence between sampling points and grid nodes. Given an L-column, L-row real matrix, This is a gradient difference matrix used to approximate the nodal gradient. It constructs a three-row equation (L rows total) by differencing the pressure along the three axes of each node in the three directions. The diagonal weight matrix is ​​W. b =diag(w b (1),…,w b (L)) and W f =diag(w f (1),…,w f (L)) correspond to the weighting of the fitting term and the gradient term, respectively. Finally, the conjugate gradient iteration method is used to solve the problem, and the iteration stops when the residual norm is lower than the standard deviation of the field pressure and the micro-strain measurement noise by the same order of magnitude. The solution vector is then output to reconstruct the pressure value. With grid node coordinates {x k Pairing completes the generation of the reconstructed pressure field.

[0058] S3: In a continuous spatial pressure field, locate the corresponding tetrahedral element for each predefined monitoring node, solve the centroid coordinates of the tetrahedral element, and perform linear interpolation on the reconstructed pressure value of the tetrahedral vertex to obtain the reconstructed pressure value of the node. Combine the measured pressure value to obtain the pressure residual of the node.

[0059] Specifically, the acoustic strain-pressure dataset is aligned to the same frequency, and the acoustic strain-pressure dataset is mapped to a digital twin geometric mesh constructed using triangulation to reconstruct a continuous spatial pressure field. This includes obtaining the acoustic strain-pressure dataset, denoising and aligning the acoustic echo and micro-strain curves, constructing a digital twin geometric mesh using Delaunay triangulation, and mapping the denoised and corrected acoustic echo and micro-strain curves with the measured pressure values ​​onto the mesh nodes to obtain discrete pressure and strain field quantities.

[0060] The pressure field is reconstructed using a weighted joint least squares regularization method based on discrete pressure and strain field quantities. The continuous pressure value of each node on the grid is obtained through numerical iteration, and a continuous spatial pressure field is reconstructed on the digital twin geometric grid.

[0061] Furthermore, based on a right-handed Cartesian coordinate system (X-axis pointing due east, Y-axis pointing due north, Z-axis representing elevation), the reconstructed pressure field and measured pressure values ​​are used as inputs to calculate the pressure residuals at the nodes.

[0062] First, the coordinates of all grid nodes {x k} and the coordinates of all monitoring point nodes {x j In this process, an octree is used to accelerate indexing, combined with Delaunay adjacency lookup to locate each monitoring point node x. j The tetrahedral unit to which it belongs.

[0063] If the query fails or the cell geometry degenerates, switch to the nearest neighbor mapping to ensure that each x j All have mapping units. Next, construct the barycenter coordinate matrix, represented as:

[0064]

[0065] Among them, V j Let l be the centroid coordinate coefficient matrix. j Let be the target coordinate vector. and Represents network node x j The X, Y, Z coordinate components of the first vertex of the tetrahedral element in a three-dimensional Cartesian coordinate system, x j ,y j,z j For monitoring point node x j The X, Y, and Z coordinates in the same three-dimensional Cartesian coordinate system.

[0066] Solving M using singular value decomposition j λ j =l j Obtain the barycenter coordinates and the barycenter coordinate vector λ. j =[λ1,…,λ4] T Then, for each vertex of the tetrahedral element, a weighted coefficient w based on the difference in local deformation rate between the monitoring point and the vertex is applied. n,j , represented as:

[0067]

[0068] Where, ε j and These represent the deformation rate at the monitoring point and the local deformation rate at the vertex, respectively; β>0 is the micro-strain weighting coefficient; λ n (x j This represents the weight from the monitoring point to each vertex of the tetrahedral cell. The weighting coefficient is w. n,j Reconstruction pressure value of tetrahedron vertices Normalized interpolation is expressed as:

[0069]

[0070] Among them, P recon (x j ,t) is x j Weighted reconstruction pressure. This enables an adaptive amplified response to changes in the spatial gradient of the pressure field.

[0071] Finally, the difference between the weighted reconstructed pressure and the measured pressure is calculated point by point and expressed as:

[0072] r j (t)=P(x j ,t)-P recon (x j ,t);

[0073] Where, r j (t) represents the pressure residual at time t of the j-th monitoring node, P(x) j ,t) represents the measured pressure value at the monitoring point at time t.

[0074] All residuals are combined sequentially according to the monitoring node number to form the monitoring node pressure residual vector.

[0075] S4: Approximately construct a graph wavelet operator, perform convolution projection on the pressure residual at each scale to extract multi-scale residual signal features, and stitch the projection results of each scale into a graph wavelet multi-scale feature matrix in node order.

[0076] Specifically, the process of calculating the centroid coordinates of the tetrahedral elements and performing linear interpolation on the reconstruction pressure values ​​of the tetrahedral vertices to obtain the reconstruction pressure values ​​of the nodes includes locating the tetrahedral elements to which each monitoring node belongs through octree indexing and Delaunay adjacency query.

[0077] The centroid coordinates of the tetrahedral element are calculated, and the pressure value of the tetrahedral vertex is weighted linearly interpolated by combining the micro-strain difference between the monitoring point and the vertex to obtain the weighted reconstructed pressure value of each monitoring node.

[0078] The weighted reconstructed pressure value is subtracted from the corresponding measured pressure point by point and summarized in node order to generate the pressure residual of the monitoring node.

[0079] Furthermore, an undirected weighted graph G(V,E) is first constructed based on the inverse ratio of the pipe segment resistance coefficient, where the node set V represents the monitoring nodes and the edge weight w ij =1 / R ij , where R ij To monitor the pipe segment resistance between monitoring points i and j, where i represents a monitoring point different from j, the degree matrix D and weight matrix W are calculated, and a normalized graph Laplacian operator is constructed, expressed as:

[0080]

[0081] Where I is the identity matrix, W = [w ij Let I be the weight matrix, and D be the degree matrix. I is the identity matrix, a J×J diagonal matrix where all diagonal elements are 1 and off-diagonal elements are 0. J is the total number of monitoring nodes. W is a J×J symmetric matrix. D is a J×J diagonal matrix where the diagonal elements of a row are equal to the sum of all elements in the corresponding row of W.

[0082] Then, spectral scaling maps the spectral range to [-1, 1], represented as:

[0083]

[0084] Where, λ max L estimated by the power method norm The largest eigenvalue. Next, define the adaptive spectral wavelet kernel function:

[0085] g s (λ)=(1+η‖r(t)‖2)e -sλ ;

[0086] Where s is the wavelet scale, ||r(t)||² is the L2 norm of the residual vector, η>0 is the global residual gain coefficient, and g is fitted using the least squares method. s (λ) yields the expansion coefficient α s,K K is the order of the Chebyshev polynomial approximation expansion.

[0087] Using Chebyshev polynomial approximation expansion to order K (4 in this embodiment), the wavelet operator is constructed as follows:

[0088]

[0089] Among them, T n This forms a Chebyshev polynomial basis. Next, the residual vector is represented by a multi-scale projection as follows:

[0090] r s (t)=Ψ s r(t);

[0091] The local and global anomaly responses at each scale were obtained respectively.

[0092] Finally, all scale projection results are horizontally stitched together according to node order to form a graph wavelet multi-scale feature matrix R(t) with dimensions of node number × scale number, expressed as:

[0093]

[0094] S5: Train a spectrogram wavelet neural network based on historical data, input the spectrogram wavelet multi-scale feature matrix to obtain the original score, and normalize the original score through the Sigmoid function to obtain the normalized anomaly score.

[0095] Specifically, the input graph wavelet multi-scale feature matrix is ​​used to obtain the original score. The original score is then normalized using the Sigmoid function to obtain the normalized anomaly score. This involves inputting the graph wavelet multi-scale feature matrix into a spectral graph wavelet neural network trained on historical data. During online operation, the spectral graph wavelet neural network performs two layers of graph convolution and nonlinear mapping on the input features and outputs the original score. The normalized anomaly score is then obtained by normalizing the original score using the Sigmoid function.

[0096] Furthermore, the historical operation and monitoring logs of the centralized heating network are retrieved to extract manual inspection reports and alarm events for each monitoring node, including alarm time and fault type. These events are then precisely aligned with the time-series data collected by sensors using a PTP clock or timestamp. For each monitoring node j, the tag y is added within the time window of the corresponding alarm event. j Set as fault (i.e., y) j =1), for nodes that do not trigger any alarms within the time window, the tag will be set to normal (i.e., y = 1). j =0);

[0097] It should be noted that the existing spectral wavelet neural network (GW-GNN) model is used for real-time anomaly scoring calculation. The GW-GNN model is trained under supervised conditions using historical operational data and fault labels during the offline phase, with a total training set sample size of N. s 3000 samples (range 1000–5000) were selected and divided into 2400 training samples and 600 validation samples in an 8:2 ratio. Each sample consists of a graph wavelet multi-scale feature matrix and the corresponding node fault label vector y = [y1,…,y]. J ] T Composition, label y j =1 indicates a fault at node j, y j =0 indicates normal. Training uses the binary classification cross-entropy loss function, expressed as:

[0098]

[0099] Where, q j This represents the raw score output by the GW-GNN at node j. The optimization algorithm used is Adam, with a learning rate of 0.0005 (range 0.0001–0.001), 100 epochs (range 50–200), and a batch size of 32. After training, the GW-GNN model parameters are saved.

[0100] After the GW-GNN model was trained offline, the Spectral Graph Wavelet Neural Network (GW-GNN) was evaluated using an independent validation set. First, the true fault labels of the nodes in the validation set were mapped to the original scores output by the model. The true positive rate (recall) and true negative rate (specificity) at different score thresholds were calculated, and ROC curves were plotted accordingly. Then, the high-risk threshold was selected as the score point corresponding to a specificity of 95%. At this point, the GW-GNN model achieved a recall of 90% or higher while maintaining a false positive rate of only 5%, effectively identifying most faulty nodes. Next, the position on the same ROC curve where the recall and specificity curves intersect or the difference between them is minimized was found. This score was used as the medium-risk threshold to achieve a balanced distinction between moderately anomalous nodes.

[0101] During online inference, R(t) is input into the GW-GNN, and after two layers of spectral wavelet convolution and nonlinear activation operations, the original scores {u} of each node are output. j Subsequently, the original scores are normalized using the Sigmoid function to obtain the normalized anomaly score. j , represented as

[0102]

[0103] The rating range is limited to 0 to 1.

[0104] To facilitate tiered response for operations and maintenance, a high-risk threshold T is defined. high =0.8 and medium risk threshold T mid =0.5, the threshold is determined based on the ROC curve plotted on the validation set of GW-GNN, under the premise of ensuring 95% specificity and 90% recall for node faults. Normalized anomaly scoring o j If greater than or equal to T high If α j Between T mid With T high If α is between these values, it is considered medium risk. j Below T mid If so, it is judged as low risk.

[0105] S6: Based on standardized anomaly scoring, the faulty pipe segment is located through topology and risk classification alarm is performed.

[0106] Specifically, the risk classification and alarm process based on standardized anomaly scoring screening and faulty pipe segments located through topology includes screening high-risk nodes using node anomaly scoring vectors, performing node location in conjunction with the pipe network topology, classifying each segment into high, medium, and low risk categories according to network management type and scoring, and generating graded alarm work orders.

[0107] Furthermore, firstly, according to the preset high-risk threshold T high High-risk nodes with anomaly scores greater than or equal to the high-risk threshold are selected from all monitoring nodes to obtain a preliminary set of faulty nodes. Then, a depth-first search is performed on these nodes based on the pipeline topology to extract a subset of preliminary faulty nodes that are connected to each other. Each subset of preliminary faulty nodes corresponds to a candidate range of a faulty pipe segment.

[0108] Next, for each subset of faulty pipe segments, the maximum normalized anomaly score is taken as the anomaly score for that segment, and risk classification is performed in conjunction with the network type identifier of each node (main network, return trunk line, or branch network):

[0109] If a faulty segment subset contains a main network node and its anomaly score reaches the high-risk threshold, it is determined to be a high-risk main network.

[0110] If all nodes are return water mains and their scores fall between the medium and high risk thresholds, they are classified as medium-risk (medium-risk) in the return water system. All other pipe sections are classified as low-risk (branch-risk). Finally, corresponding alarm work orders are automatically generated based on different risk levels and sent to the dispatch and maintenance terminals through the operation and maintenance platform. At the same time, the location of the faulty pipe section is highlighted on the GIS map to guide on-site personnel to respond quickly and complete accurate handling.

[0111] This embodiment also provides an artificial intelligence-based centralized heating network pressure monitoring system, including:

[0112] The system includes a data acquisition module, a grid reconstruction module, a residual calculation module, a feature extraction module, an anomaly scoring module, and an alarm module.

[0113] The data acquisition module is used to collect acoustic echo, micro-strain curve and measured pressure value data to construct an acoustic strain-pressure dataset.

[0114] The mesh reconstruction module is used to perform frequency-aligned processing on the acoustic strain-pressure dataset, mapping the acoustic strain-pressure dataset to a digital twin geometric mesh constructed using triangulation, and reconstructing a continuous spatial pressure field.

[0115] The residual calculation module is used to locate the corresponding tetrahedral element for each predefined monitoring node in a continuous spatial pressure field, solve the centroid coordinates of the tetrahedral element, and perform linear interpolation on the reconstructed pressure value of the tetrahedral vertex to obtain the reconstructed pressure value of the node. Combined with the measured pressure value, the pressure residual of the node is obtained.

[0116] The feature extraction module is used to approximate the construction of the graph wavelet operator. It performs convolution projection on the residual vector at each scale to extract multi-scale residual signal features, and concatenates the projection results at each scale in node order to form a graph wavelet multi-scale feature matrix.

[0117] The anomaly scoring module is used to train a spectrogram wavelet neural network based on historical data. The input graph wavelet multi-scale feature matrix is ​​used to obtain the raw score, and the raw score is normalized by the Sigmoid function to obtain the normalized anomaly score.

[0118] The alarm module is used to filter and locate faulty pipe segments based on standardized anomaly scores and perform risk-level alarms.

[0119] This embodiment also provides a computer device applicable to the centralized heating network pressure monitoring method based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the centralized heating network pressure monitoring method based on artificial intelligence as proposed in the above embodiment.

[0120] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0121] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the artificial intelligence-based centralized heating network pressure monitoring method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0122] In summary, this invention achieves accurate characterization of pipeline pressure distribution and provides a benchmark for subsequent residual analysis by: aligning the acoustic strain-pressure dataset with the same frequency and mapping the data onto a triangulated digital twin geometric grid to reconstruct a continuous spatial pressure field; and realizing high-resolution dynamic monitoring by approximating the graph wavelet operator and convolving and projecting the residual vector at multiple scales to form a graph wavelet multi-scale feature matrix. This enables multi-scale anomaly feature extraction and provides multi-dimensional input for anomaly scoring and pipeline segment location, making the entire detection process more sensitive and robust.

[0123] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An artificial intelligence-based centralized heating network pipe pressure monitoring method, characterized in that: The method comprises the following steps: Collecting acoustic echo, micro-strain curve and measured pressure value data to construct an acoustic-strain-pressure data set; Aligning the acoustic-strain-pressure data set in the same frequency, mapping the acoustic-strain-pressure data set to a digital twin geometric grid constructed by using triangular subdivision, and reconstructing a continuous spatial pressure field; Locating the corresponding tetrahedral unit for each predefined monitoring node in the continuous spatial pressure field, calculating the barycentric coordinates of the tetrahedral unit, and performing linear interpolation on the reconstructed pressure values of the tetrahedral vertices to obtain the reconstructed pressure values of the nodes, and combining the measured pressure values to obtain the pressure residual of the nodes; Approximately constructing a graph wavelet operator, and performing convolution projection on the pressure residual at each scale to extract multi-scale residual signal features, and splicing the projection results of each scale into a graph wavelet multi-scale feature matrix according to the node order; Training a spectral graph wavelet neural network based on historical data, inputting the graph wavelet multi-scale feature matrix to obtain original scores, and normalizing the original scores by a Sigmoid function to obtain normalized anomaly scores; Based on the normalized anomaly scores, filtering and locating the fault pipe segment through topology, and performing risk classification and warning. 2.The method of claim 1, wherein the method comprises: The method comprises the following steps: Arranging a single-mode optical fiber along the outer wall of the trunk and branch pipe segments, and arranging detection units and point pressure sensors at both ends of the single-mode optical fiber and nodes; 3.The method of claim 2, wherein the method comprises: Starting periodic scanning to obtain acoustic echo, micro-strain curve and measured pressure value with unified time sequence marking, and constructing an acoustic-strain-pressure data set. The method comprises the following steps: 4.The method of claim 3, wherein the method comprises: Obtaining the acoustic-strain-pressure data set, denoising and aligning the acoustic echo and micro-strain curve in the same frequency, constructing a digital twin geometric grid by using Delaunay triangular subdivision, mapping the denoised and corrected acoustic echo and micro-strain curve and the measured pressure value to the grid nodes to obtain discrete pressure and strain field quantities, and reconstructing a continuous spatial pressure field on the digital twin geometric grid. Based on the weighted joint least squares regularization method of the discrete pressure and strain field quantities, the continuous pressure values of each node on the grid are obtained by numerical iteration, and a continuous spatial pressure field is reconstructed on the digital twin geometric grid. The method comprises the following steps: Locating the tetrahedral unit to which each monitoring node belongs by using octree indexing and Delaunay adjacency query; Calculating the barycentric coordinates of the tetrahedral unit, and combining the micro-strain difference between the monitoring point and the vertex to perform weighted linear interpolation on the pressure values of the tetrahedral vertices, to obtain the weighted reconstructed pressure values of each monitoring node; Subtracting the weighted reconstructed pressure values from the corresponding measured pressure point by point and summarizing them according to the node order to generate the pressure residual of the monitoring node. 5.The method of monitoring pressure of a central heating network based on artificial intelligence according to claim 4, characterized in that: The convolutional projection of the pressure residual on each scale extracts a multi-scale residual signal feature, and the projection results of each scale are spliced into a graph wavelet multi-scale feature matrix in node order, including constructing a weighted undirected graph based on the inverse proportion of pipe section resistance between nodes, constructing a graph wavelet operator using a Gaussian spectral wavelet function on multiple preset scales and through Chebyshev polynomial approximation, and performing convolutional projection on the residual vector on each scale to extract a multi-scale residual signal feature, and splicing the projection results of each scale to obtain a graph wavelet multi-scale feature matrix. 6.The method of monitoring pressure of a central heating network based on artificial intelligence according to claim 5, characterized in that: The input graph wavelet multi-scale feature matrix obtains an original score, and the original score is normalized to obtain a normalized anomaly score through a Sigmoid function, including inputting the graph wavelet multi-scale feature matrix through a spectral graph wavelet neural network trained by historical data, and the spectral graph wavelet neural network performs two-layer graph convolution and non-linear mapping on the input feature when running online and outputs an original score, and then the original score is normalized through a Sigmoid function to obtain a normalized anomaly score. 7.The centralized heating network pipe pressure monitoring method based on artificial intelligence according to claim 6, wherein: The normalized anomaly score is used for screening and topological positioning of a fault pipe section, and risk classification alarm, including screening high-risk nodes with a node anomaly score vector, performing node positioning combined with pipe network topology, classifying each section according to network type and high, medium and low risk levels, and generating a classified alarm work order.

8. An artificial intelligence-based central heating network pipe pressure monitoring system based on the artificial intelligence-based central heating network pipe pressure monitoring method according to any one of claims 1 to 7, characterized in that: It comprises a data acquisition module, a grid reconstruction module, a residual calculation module, a feature extraction module, an anomaly scoring module, and an alarm module. The data acquisition module is used to collect acoustic echo, micro-strain curve and measured pressure value data to construct an acoustic strain pressure dataset. The grid reconstruction module is used for same-frequency alignment processing of the acoustic strain pressure dataset, mapping the acoustic strain pressure dataset to a digital twin geometric grid constructed by triangular subdivision, and reconstructing a continuous spatial pressure field. The residual calculation module is used to locate a corresponding tetrahedral unit for each pre-defined monitoring node in the continuous spatial pressure field, calculate the barycentric coordinates of the tetrahedral unit, and perform linear interpolation on the reconstructed pressure values of the tetrahedral vertices to obtain the reconstructed pressure values of the nodes, and combine the measured pressure values to obtain the pressure residual of the nodes. The feature extraction module is used to approximate the construction of a graph wavelet operator, perform convolutional projection on the residual vector on each scale to extract a multi-scale residual signal feature, and splice the projection results of each scale into a graph wavelet multi-scale feature matrix in node order. The anomaly scoring module is used to train a spectral graph wavelet neural network based on historical data, input the graph wavelet multi-scale feature matrix to obtain an original score, and normalize the original score through a Sigmoid function to obtain a normalized anomaly score. The alarm module is used for screening and topological positioning of a fault pipe section based on the normalized anomaly score, and risk classification alarm. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the centralized heating network pressure monitoring method based on artificial intelligence in any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the centralized heating network pressure monitoring method based on artificial intelligence in any one of claims 1-7.

Images