Centralized heating network management monitoring method and system based on artificial intelligence

Abstract

The application provides a kind of centralized heating network management monitoring method and system based on artificial intelligence, and the time series data of pressure, temperature, flow and pipe wall acoustic vibration of each node are collected.By calculating the hydraulic unbalance degree, enthalpy difference change rate and extracting the MFCC feature of acoustic vibration signal, a multi-modal fusion feature vector is generated.The pipe network topology is initialized as a graph structure, and the correlation weight between nodes is dynamically calculated using the relationship attention mechanism, and a time-varying heat propagation graph is constructed accordingly.The current features are used as node attributes, and the spatial features are extracted by graph convolution on the graph structure.The spatio-temporal features are extracted by a double-branch time convolution module, which are input into a fully connected network to output the fault type and location.The topology-aware weighted loss function is used to optimize the model during training.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of heat supply network monitoring, and particularly relates to a centralized heat supply network monitoring method and system based on artificial intelligence. BACKGROUND

[0002] The pipe network structure of centralized heat supply is complex, covers a wide range, and has a variable operating environment. Factors such as corrosion, aging, and external damage of the pipeline often cause leakage, blockage, and other failures. Artificial inspection not only consumes a lot of manpower and resources, but also is not timely in response, and it is difficult to find early hidden failures. Although the threshold-based alarm system can achieve real-time monitoring, it is easily disturbed by normal operating fluctuations, resulting in high false and missed alarm rates, and cannot provide accurate fault type and location information, which brings great inconvenience and challenges to subsequent maintenance work, and cannot meet the fine management needs of modern intelligent heat supply.

[0003] Abnormal patterns are identified by analyzing historical data such as pressure, temperature, and flow. However, most models only use single modal operating parameters, ignoring the complementarity of different physical information sources. For example, for a small leak, the acoustic vibration signal of the pipe wall is more sensitive than conventional thermal parameters, but existing methods rarely effectively integrate acoustic features with hydraulic and thermal features. Many models treat each monitoring node as an independent unit or use a fixed spatial proximity relationship when dealing with spatial relationships, failing to fully utilize the dynamic thermal propagation laws inherent in the topology of the heat supply pipe network, and cannot accurately capture the mutual influence between nodes that adjusts with operating conditions. Moreover, in terms of time series feature extraction, it is difficult to simultaneously consider both long-term dependence reflecting gradual failures and short-term fluctuations representing sudden failures. In terms of fault location, traditional loss functions usually treat the prediction deviation of all positions equally when optimizing the model, without considering the connectivity of the pipe network. SUMMARY

[0004] In order to improve the monitoring accuracy of centralized heat supply pipe networks, in a first aspect of the application, a centralized heat supply pipe network monitoring method based on artificial intelligence is provided, comprising the following steps:

[0005] Obtaining time series data of a plurality of monitoring nodes in the heat supply pipe network, the time series data including node pressure, temperature, flow, and pipe wall acoustic vibration signal;

[0006] Based on the node pressure, temperature, and flow data, the hydraulic imbalance and enthalpy difference change rate between each adjacent node are calculated as physical model features; the pipe wall acoustic vibration signal is subjected to mel-frequency cepstral coefficient transformation to extract acoustic frequency domain features; and the physical model features and the acoustic frequency domain features are spliced to generate a multi-modal fusion feature vector;

[0007] The heat supply pipe network topology structure is initialized as a basic graph structure, the correlation weight between nodes is calculated based on the multi-modal fusion feature vector of each node at the historical time by using a relational attention mechanism, and the adjacency matrix of the basic graph structure is adjusted according to the correlation weight, so as to construct a time-varying heat propagation graph;

[0008] The multi-modal fusion feature vector of each node at the current time is taken as a node attribute, and a graph convolution operation is performed on the time-varying heat propagation graph to extract a spatial coupling feature; the spatial coupling feature is input into a double-branch time convolution module, wherein a first branch uses a group of large-inflation-rate hollow convolution kernels to extract long-term time sequence dependence, a second branch uses a group of small-inflation-rate hollow convolution kernels to extract short-term time sequence fluctuation, and the outputs of the two branches are fused to obtain a space-time representation.

[0009] The space-time representation is input into a fully connected network to output the fault type and position of each pipe section, and in the model training stage, a topology-aware weighted loss function is used for parameter optimization, wherein the topology-aware weighted loss function performs weighted punishment on the position prediction deviation according to the shortest path distance between the predicted fault position and the real fault position on the heat supply pipe network topology structure.

[0010] In a second aspect of the present application, a centralized heat supply pipe network monitoring system based on artificial intelligence is provided, comprising the following units:

[0011] A data acquisition unit is configured to acquire time series data of a plurality of monitoring nodes in a heat supply pipe network, wherein the time series data includes node pressure, temperature, flow rate and pipe wall acoustic vibration signal.

[0012] A feature fusion unit is configured to calculate the hydraulic imbalance and enthalpy difference change rate between adjacent nodes based on the node pressure, temperature and flow rate data as physical model features, perform mel frequency cepstral coefficient transformation on the pipe wall acoustic vibration signal to extract acoustic frequency domain features, and splice the physical model features and the acoustic frequency domain features to generate a multi-modal fusion feature vector.

[0013] A heat propagation graph construction unit is configured to initialize the heat supply pipe network topology structure as a basic graph structure, calculate the correlation weight between nodes based on the multi-modal fusion feature vector of each node at the historical time by using a relational attention mechanism, and adjust the adjacency matrix of the basic graph structure according to the correlation weight to construct a time-varying heat propagation graph.

[0014] The feature extraction unit takes the multi-modal fusion feature vector of each node at the current time as a node attribute, performs a graph convolution operation on the time-varying heat propagation graph, and extracts spatial coupling features; the spatial coupling features are input into a double-branch time convolution module, wherein a first branch uses a group of large-expansion-rate hollow convolution kernels to extract long-term time sequence dependence, a second branch uses a group of small-expansion-rate hollow convolution kernels to extract short-term time sequence fluctuations, and the outputs of the two branches are fused to obtain a space-time representation;

[0015] The monitoring unit inputs the space-time representation into a fully connected network to output the fault type and location of each pipe section, and in the model training stage, a topology-aware weighted loss function is used for parameter optimization, which weights and punishes the position prediction deviation according to the shortest path distance between the predicted fault location and the real fault location on the heat supply pipe network topology.

[0016] In addition, the present application also provides a computer program, which, when executed by a processor, implements the method of the first aspect.

[0017] The present application fuses heat parameters such as pipe network pressure, temperature, flow and pipe wall acoustic vibration signals to generate multi-modal features, improves the recognition sensitivity and accuracy of early hidden faults such as micro-leakage, further constructs a graph structure that can reflect the influence relationship between each monitoring node, and uses a graph convolution network to deeply mine the spatial coupling characteristics between nodes, which can grasp the propagation and evolution law of the fault in the pipe network from a global perspective, overcoming the limitations of traditional methods that regard each node as an isolated unit. In the time dimension, through the double-branch structure of parallel processing of long-term time sequence dependence and short-term time sequence fluctuations, the model can simultaneously consider the characteristics of gradual faults and sudden events, enhancing the robustness of fault diagnosis. Moreover, in the model training, a topology-aware weighted loss function is introduced, so that the prediction result of the fault location is more consistent with the actual path of the pipe network, improving the accuracy of the positioning result. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of the first embodiment;

[0019] Figure 2 is a schematic diagram of the multi-modal fusion feature generation process;

[0020] Figure 3 is a schematic diagram of the double-branch space-time feature extraction model;

[0021] Figure 4 is a schematic diagram of the topology-aware weighted loss function. DETAILED DESCRIPTION

[0022] For the purposes of the present application, the technical solutions and advantages will be clearer, below will be combined with the specific embodiments of the present application and the corresponding drawings to clearly and completely describe the technical solutions of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor belong to the scope of protection of the present application. It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards, and provide corresponding operation portal for user to choose authorization or refusal.

[0023] In the first embodiment, the present application proposes a centralized heating pipe network monitoring method based on artificial intelligence, as shown in the following steps: Figure 1

[0024] S1, obtaining time series data of a plurality of monitoring nodes in a heating pipe network, the time series data including node pressure, temperature, flow and pipe wall acoustic vibration signal;

[0025] At the key nodes of the heating pipe network, such as branch points, junction points, pressurizing pump stations and the like, pressure sensors, temperature sensors, ultrasonic flowmeters and piezoelectric acceleration sensors are installed on the outside of the pipe wall. Through the data acquisition unit DAU, the analog or digital signals of each sensor are collected at a preset frequency, for example, once per second, and the data is transmitted in real time to the center server through the industrial bus or wireless communication network, such as LoRa or NB-IoT, and stored in the time series database to form a raw data set indexed by time stamp.

[0026] S2, based on the node pressure, temperature and flow data, calculating the hydraulic imbalance and enthalpy difference change rate between each adjacent node as the physical model feature; performing mel frequency cepstral coefficient transformation on the pipe wall acoustic vibration signal to extract acoustic frequency domain features; concatenating the physical model features and the acoustic frequency domain features to generate a multi-modal fusion feature vector;

[0027] ​The hydraulic unbalance degree is calculated by the law of conservation of mass, i.e. for any pipe segment, the instantaneous flow rate at the inlet node is subtracted from the instantaneous flow rate at the outlet node. Enthalpy is a function of temperature and pressure, and is calculated according to the temperature and pressure of each node by consulting a water vapor table. The enthalpy difference between the two nodes of the pipe segment is calculated, and the first-order difference of the time series of the enthalpy difference is taken to obtain the enthalpy difference change rate or the heat power loss change rate. For the acoustic vibration signal, pre-emphasis, framing and windowing processing are performed, for example, a Hamming window is used, and the frequency spectrum is obtained by fast Fourier transform. The frequency spectrum is passed through a set of mel triangular filters to obtain the mel spectrum, and the mel frequency cepstral coefficient MFCC is obtained by discrete cosine transform. The calculated hydraulic unbalance degree, enthalpy difference change rate and extracted MFCC coefficient vector are directly spliced along the feature dimension to form a longer numerical vector, as shown in Figure 2 .

[0028] In an optional embodiment, the calculation of the hydraulic unbalance degree and the enthalpy difference change rate between each adjacent node comprises:

[0029] For a pipe segment composed of an upstream node i and a downstream node j, the hydraulic unbalance degree at time t is the difference between the upstream node flow rate and the downstream node flow rate .

[0030] The enthalpy difference change rate is calculated by the following steps:

[0031] The thermal power loss of the pipe segment at time t is calculated based on the node temperature T and the flow rate Q , where c is the specific heat capacity of water and p is the density of water.

[0032] The difference between the thermal power loss at the current time and the previous time is calculated to obtain the enthalpy difference change rate.

[0033] The hydraulic imbalance degree reflects the mass conservation in the pipe section. For example, an upstream monitoring node i has a flow reading of 100 cubic meters per hour at a certain time t, while its immediate downstream node j has a flow reading of 99.5 cubic meters per hour. The hydraulic imbalance degree at that time is 0.5 cubic meters per hour. In normal operating conditions, due to measurement errors or minor normal leaks, this value will fluctuate around zero within a very small range. If a leakage fault occurs, the flow at the downstream node will decrease significantly, resulting in a sharp increase in the imbalance degree, for example, to 10 cubic meters per hour, which is a strong fault signal. The enthalpy difference rate is used to capture the dynamic changes in the thermodynamic state. Assuming the specific heat capacity of water c is 4200 joules per kilogram per degree Celsius, and the density p is 970 kilograms per cubic meter. At time t, upstream node i has a flow of 100 cubic meters per hour and a temperature of 85 degrees Celsius, while downstream node j has a flow of 99.5 cubic meters per hour and a temperature of 84.8 degrees Celsius. Based on this, the thermal power loss AP(t) can be calculated. Combined with the data at the previous time t-1, AP(t-1) can be calculated. The enthalpy difference rate is the difference between the two power loss values. In stable operating conditions, this rate is close to zero. When a pipe section breaks and hot water is ejected, the heat loss of the pipe section will increase sharply, resulting in a much larger thermal power loss AP(t) at the current time than at the previous time, thus producing a significantly positive enthalpy difference rate, indicating the occurrence of a thermal abnormal event.

[0034] In an optional embodiment, the Mel-frequency cepstral coefficient transformation of the pipe wall acoustic vibration signal, extracting acoustic frequency domain features, includes:

[0035] Frame the pipe wall acoustic vibration signal and window it; perform a fast Fourier transform on each frame of signal to obtain a frequency spectrum;

[0036] Pass the frequency spectrum through a Mel filter bank and calculate the log energy;

[0037] Discrete cosine transform the log energy and extract low-order coefficients as acoustic frequency domain features.

[0038] The acoustic sensor installed on the pipe wall will collect continuous vibration signals. Suppose an audio data of 2 seconds long and a sampling rate of 44100hz is collected. This long signal is divided into a series of short, partially overlapping frames, for example, each frame is 25 milliseconds long, and there is a 10 millisecond overlap between frames. Apply a Hamming window or Hanning window function to each frame to smooth the edges of the frame and reduce the spectral leakage effect produced by the subsequent Fourier transform, which helps to more accurately analyze the frequency components.

[0039] After framing and windowing, a fast Fourier transform is performed on each frame to convert it from a time-domain signal to a frequency-domain spectrogram, which shows the energy distribution at different frequencies. The linear frequency spectrum is filtered through a mel filter bank, which contains, for example, 40 triangular filters. The mel scale is a non-linear scale that mimics the human ear's perception of frequency, with higher resolution at lower frequencies. The log values of the energy of each filter output are calculated, resulting in a 40-dimensional log energy vector. A discrete cosine transform is performed on this vector, and only the first 13 coefficients are kept. These 13 values constitute the mel-frequency cepstral coefficients (MFCCs) of the frame, which are compact and efficient acoustic features that can effectively distinguish background noise, water flow, and unique sounds produced by different types of leaks.

[0040] S3, initializing the heat supply pipe network topology as a base graph structure, calculating the correlation weight between nodes based on the multi-modal fusion feature vectors of the nodes at historical time points using a relational attention mechanism, and adjusting the adjacency matrix of the base graph structure according to the correlation weight to construct a time-varying heat propagation graph;

[0041] A static adjacency matrix is constructed as a base graph structure according to the physical connection relationship of the pipe network, where the elements between directly connected nodes are 1 and the elements between unconnected nodes are 0. A graph attention network model is introduced, and the multi-modal fusion feature vector sequence of each node within a historical time window is taken as input. Through a learnable linear transformation and an attention function, such as a scaled dot-product attention calculation, the mutual influence strength between any two nodes is quantified, and a dynamic attention score matrix is generated. The score matrix is the correlation weight between nodes, which is used to replace or weight the original static adjacency matrix to obtain a time-varying heat propagation graph reflecting the heat propagation relationship under the current working condition, as shown in Figure 2 .

[0042] In an optional embodiment, the correlation weight between nodes is calculated based on the multi-modal fusion feature vectors of the nodes at historical time points using a relational attention mechanism, comprising:

[0043] The multi-modal fusion feature vector sequence of each node within a historical time window is transformed through a learnable linear transformation to generate a query vector, a key vector, and a value vector.

[0044] The scaled dot-product attention score of the query vector and the key vector is calculated, and normalized by a Softmax function to obtain the correlation weight between nodes.

[0045] Exemplarily, assume that data in the past 30 minutes are concerned, and a multi-modal feature vector that integrates hydraulic, thermal and acoustic information is generated every minute. For each node in the pipe network, a sequence of 30 feature vectors is obtained. When the relevance between node A and all other nodes such as B, C, D, etc. needs to be calculated, the feature sequence of node A is input into a linear layer to generate its query vector . The feature sequences of all nodes including A, B, C, D are input into two other different linear layers to generate their key vectors 、 、 、 and value vectors 、 、 、 .

[0046] To determine the degree of influence of other nodes on node A, the dot products of the query vector of node A and the key vectors of all nodes, i.e. , , , etc. are calculated. These dot product scores reflect the original similarity or relevance strength between nodes. To ensure the stability of the training process, these scores are divided by a scaling factor, usually the square root of the dimension of the key vector. Then the scaled scores are normalized by the Softmax function to obtain a set of weights with a sum of 1. For example, for node A, the attention weights it gets for nodes B, C, D may be 0.7, 0.2 and 0.1 respectively. This indicates that at the current time, the state of node A is most relevant to the state of node B. These weights are the relevance weights, which quantitatively reflect the information propagation strength and mutual influence degree between any two nodes in the pipe network in real time.

[0047] In an optional embodiment, the adjusting the adjacency matrix of the base graph structure according to the relevance weights comprises:

[0048] element-wise multiplying the adjacency matrix of the base graph structure with a matrix composed of the relevance weights to obtain the adjacency matrix of the time-varying thermal propagation graph, which modulates the strength of the base topology connection by the relevance weights.

[0049] The physical connection of the heating pipe network is fixed and unchangeable, which can be represented by a base adjacency matrix A. In this matrix, if there is a physical pipe directly connecting node i and node j, the element at the corresponding position in the matrix is 1, otherwise 0. The matrix only describes the static, physical connection relationship and cannot reflect the actual working condition of the pipe network in operation. For example, even if there is a pipeline between two nodes, the thermal and hydraulic interaction between them will be interrupted if the valve is closed. In order to enable the graph model to perceive such dynamic changes, the correlation weight matrix W calculated in the previous link is Hadamard multiplied with the basic adjacency matrix A, that is, element-by-element multiplication. The element of the correlation weight matrix W represents the actual correlation strength between node i and node j at the current time. For example, although node i and j are physically connected, i.e. equal to 1, but due to slow water flow, the correlation weight between them may only be 0.3. After multiplication, the element value of the corresponding position in the time-varying adjacency matrix becomes 0.3. Conversely, if the water flow between another pair of physically connected nodes k and l is turbulent, the correlation weight is as high as 0.9, and the corresponding element value in the new matrix is 0.9. In this way, the binary adjacency matrix originally representing the connection or not is adjusted to a weighted adjacency matrix that can reflect the real-time information flow strength, so that the graph neural network can more accurately learn the propagation pattern of the fault on this basis.

[0050] S4, taking the multi-modal fusion feature vector of each node at the current time as the node attribute, performing graph convolution operation on the time-varying heat propagation graph to extract spatial coupling features; inputting the spatial coupling features into a double-branch time convolution module, wherein a first branch uses a group of large dilation rate hollow convolution kernels to extract long-term time sequence dependence, a second branch uses a group of small dilation rate hollow convolution kernels to extract short-term time sequence fluctuations, and the outputs of the two branches are fused to obtain a space-time representation;

[0051] Taking the multi-modal fusion feature vector of each node at the current time as the initial feature of the graph node, using the graph convolution network GCN layer to aggregate the feature information of the neighbor nodes according to the adjacency relationship of the time-varying heat propagation graph, and generating a node embedding representation containing global spatial correlation, i.e. spatial coupling features, as shown in Figure 3 , inputting the time sequence of the features into a double-branch time convolution network TCN. The dilation rate sequence of the hollow convolution layer of the first branch is set to an exponential growth type, such as 1, 2, 4, 8, to obtain a large receptive field and capture long-term trends such as leakage and aging. The dilation rate sequence of the second branch is set to a smaller fixed value or a slow growth type, such as 1, 1, 2, 2, to finely capture short-term patterns of sudden events such as pipe wall impact. The feature maps output by the two branches are element-wise added or spliced and fused to obtain a space-time representation that takes into account long-term and short-term dynamics, as shown in Figure 3 .

[0052] More specifically, the double-branch time convolution module comprises:

[0053] The first branch adopts a set of dilated convolution kernels with exponentially increasing dilation rates to capture feature dependencies at large time scales; the second branch adopts a set of dilated convolution kernels with a fixed dilation rate of 1 to extract local feature fluctuations at small time scales.

[0054] The output feature maps of the two branches are fused to integrate long and short term temporal information.

[0055] The first branch is mainly used to capture long-term trends and slow-changing patterns. Preferably, the first branch is stacked with multiple dilated convolution layers, whose dilation rates are exponentially increasing, for example, set to 1, 2, 4, 8. The first layer of convolution kernels observes data at adjacent time points; the second layer has a dilation rate of 2, and its convolution kernels have a receptive field that skips one time step, observing data at time points with an interval of 2; the third layer skips three time steps and observes data with an interval of 4, and so on. When the dilation rate is large, the top layer of convolution kernels can cover a very long time span; when the dilation rate is small, it can effectively learn long-term dependencies such as slow pressure decline caused by pipeline corrosion or seasonal load changes.

[0056] The second branch mainly captures short-term and sudden events. Preferably, the second branch uses standard convolution layers with a fixed dilation rate of 1. The convolution kernels always observe consecutive adjacent time point data, and are extremely sensitive to local details and high-frequency changes in the signal, for example, they can capture sudden pressure drops and flow surges caused by sudden pipe ruptures within a time scale of seconds or tens of seconds. The feature maps obtained after processing the two branches are combined, for example, by element-wise addition or splicing operations. The resulting feature representation contains both macroscopic trend understanding and microscopic mutation sensitivity.

[0057] S5, inputting the spatio-temporal representation into a fully connected network to output the fault type and location of each pipe segment, and in the model training stage, using a topology-aware weighted loss function to optimize the parameters, the topology-aware weighted loss function weights the position prediction deviation according to the shortest path distance between the predicted fault location and the real fault location on the heat supply pipe network topology.

[0058] The flattened spatio-temporal representation is then fed into a multi-layer perceptron, i.e. a fully connected network, containing several hidden layers. The output layer of the network is split into two branches. One branch outputs the probability of each predefined fault type, such as leak, blockage, normal, etc. via a Softmax activation function. The other branch also outputs the probability of the fault occurring in each pipe segment via a Softmax function. During backpropagation of the fully connected network model, the loss function is composed of two parts, namely the cross-entropy loss for fault classification and the weighted cross-entropy loss for fault localization. For the localization loss, the shortest topological distance between any two pipe segments in the pipe network is pre-computed using a breadth-first search or Dijkstra algorithm. When the model’s predicted fault location does not match the true location, its loss term is multiplied by a weight factor that is positively correlated with the topological distance, e.g. the weight is 1 plus the topological distance multiplied by a hyperparameter, so that the farther apart the wrongly predicted locations are in the topology, the more heavily they are penalized.

[0059] In an optional embodiment, the formula for calculating the topologically-aware weighted loss function L is: wherein:

[0060] is the cross-entropy loss for fault type classification; a is a balancing hyperparameter;

[0061] is the weighted fault location localization loss, calculated as:

[0062] Calculate the shortest path distance d between the predicted fault location and the true fault location on the heat supply pipe network topology graph;

[0063] Calculate the penalty weight w based on the distance d, which increases with d;

[0064] Multiply the penalty weight w with the cross-entropy loss for location localization to obtain .

[0065] The loss function is composed of two parts, balanced by the hyperparameter a. For example, a can be set to 0.8, indicating that the model training places more emphasis on the accuracy of location localization. The first part is the standard classification loss, used to penalize the model’s errors in fault type judgment. For example, if the model incorrectly classifies a “leak” fault as a “blockage” fault, this part of the loss value will increase.

[0066] The second part The loss function is enabled to perceive the topology of the pipe network. Suppose a real fault in the pipe network occurs at node numbered 50. If the model predicts the fault location at node 51, and node 51 is directly connected to node 50 on the pipe network topology map, the shortest path distance d between them is 1. If the model incorrectly predicts the fault at node 80, and at least 6 intermediate nodes and pipe segments are needed to go from node 50 to node 80, the shortest path distance d is 6. The penalty weight w can be designed to linearly or exponentially increase with d, for example, set w equal to 1 plus d. Therefore, the penalty weight for the former prediction is 2, and the penalty weight for the latter prediction is 7. The weight w is multiplied to the base loss of location positioning. That is, although both predictions are wrong, the model will be punished much more for predicting a much more topologically remote location, as shown in Figure 4 The model is enabled to not only guess the node number during the learning process, but also to understand the physical connection relationship of the pipe network, to generate a positioning result that is more reasonable and closer to the real situation in the physical space.

[0067] In a second embodiment, an artificial intelligence-based central heating pipe network monitoring system is provided, comprising the following units:

[0068] A data acquisition unit is configured to acquire time series data of a plurality of monitoring nodes in the heating pipe network, the time series data including node pressure, temperature, flow rate, and pipe wall acoustic vibration signal;

[0069] A feature fusion unit is configured to calculate the hydraulic imbalance and enthalpy difference change rate between each adjacent node based on the node pressure, temperature, and flow rate data as physical model features, perform Mel frequency cepstral coefficient transformation on the pipe wall acoustic vibration signal to extract acoustic frequency domain features, and splice the physical model features and the acoustic frequency domain features to generate a multi-modal fusion feature vector;

[0070] A heat propagation graph construction unit is configured to initialize the topology of the heating pipe network as a base graph structure, calculate the correlation weight between nodes based on the multi-modal fusion feature vector of each node at a historical time using a relational attention mechanism, and adjust the adjacency matrix of the base graph structure according to the correlation weight to construct a time-varying heat propagation graph;

[0071] A feature extraction unit is configured to use the multi-modal fusion feature vector of each node at the current time as node attributes, perform graph convolution operation on the time-varying heat propagation graph to extract spatial coupling features, input the spatial coupling features into a double-branch time convolution module, wherein the first branch uses a group of large dilation rate hollow convolution kernels to extract long-term time series dependence, the second branch uses a group of small dilation rate hollow convolution kernels to extract short-term time series fluctuations, and the outputs of the two branches are fused to obtain a spatiotemporal representation;

[0072] The monitoring unit is configured to input the spatio-temporal representation into a full connection network, output a fault type and a location of each pipe section, and in a model training stage, adopt a topology-aware weighted loss function to optimize parameters, wherein the topology-aware weighted loss function is configured to weight and punish a location prediction deviation according to a shortest path distance of a predicted fault location and an actual fault location on the topology structure of the heat supply pipe network.

[0073] The second embodiment is implemented in the same way as the first embodiment, and thus details are not repeated here.

[0074] Those skilled in the art will understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, Compact Disc Read-Only Memory (CD-ROM), optical storage, etc.) containing computer-usable program code.

[0075] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices produce a device that implements the flowcharts and / or block diagrams. Figure 1 The device that implements the function specified in one flow or multiple flows and / or blocks. Figure 1 The device that implements the function specified in one flow or multiple flows and / or blocks.

[0076] The above only describes the embodiments of the present application and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for monitoring centralized heating pipe networks based on artificial intelligence, characterized in that, Includes the following steps: Acquire time-series data from multiple monitoring nodes in the heating network, including node pressure, temperature, flow rate, and acoustic vibration signals of the pipe wall; Based on the node pressure, temperature, and flow data, the hydraulic imbalance and enthalpy difference change rate between adjacent nodes are calculated as physical model features; the Mel frequency cepstral coefficient transform is performed on the acoustic vibration signal of the pipe wall to extract acoustic frequency domain features; the physical model features and the acoustic frequency domain features are concatenated to generate a multimodal fusion feature vector. The heating network topology is initialized into a basic graph structure. The relational attention mechanism is used to calculate the association weights between nodes based on the multimodal fusion feature vectors of each node at historical time. The adjacency matrix of the basic graph structure is adjusted according to the association weights to construct a time-varying heat propagation graph. The multimodal fusion feature vectors of each node at the current time are used as node attributes. Graph convolution is performed on the time-varying heat propagation map to extract spatial coupling features. The spatial coupling features are then input into a dual-branch temporal convolution module. The first branch uses a group of dilated convolution kernels with a large dilation rate to extract long-term temporal dependencies, and the second branch uses a group of dilated convolution kernels with a small dilation rate to extract short-term temporal fluctuations. The outputs of the two branches are then fused to obtain a spatiotemporal representation. The spatiotemporal representation is input into a fully connected network, and the fault type and location of each pipe segment are output. During the model training phase, a topology-aware weighted loss function is used for parameter optimization. The topology-aware weighted loss function applies a weighted penalty to the location prediction deviation based on the shortest path distance between the predicted fault location and the actual fault location on the heating pipe network topology.

2. The method according to claim 1, characterized in that, The calculation of the hydraulic imbalance and enthalpy difference change rate between adjacent nodes includes: For a pipe segment consisting of upstream node i and downstream node j, the hydraulic imbalance at time t is equal to the upstream node flow rate. Traffic with downstream nodes The difference; The rate of change of enthalpy difference is calculated through the following steps: The thermal power loss of the pipe section at time t is calculated based on the node temperature T and the flow rate Q. , where c is the specific heat capacity of water and ρ is the density of water; Calculate the difference in thermal power loss between the current moment and the previous moment. The rate of change of enthalpy difference is obtained.

3. The method according to claim 1, characterized in that, The step of performing Mel-frequency cepstral coefficient transform on the acoustic vibration signal of the pipe wall to extract acoustic frequency domain features includes: The acoustic vibration signal of the pipe wall is divided into frames and windowed; a fast Fourier transform is performed on each frame of the signal to obtain the spectrum; Pass the spectrum through a Mel filter bank and calculate the logarithmic energy; The logarithmic energy is subjected to a discrete cosine transform, and the low-order coefficients are truncated as acoustic frequency domain features.

4. The method according to claim 1, characterized in that, The method of calculating the association weights between nodes using the relational attention mechanism based on the multimodal fusion feature vectors of each node at historical moments includes: The multimodal fusion feature vector sequence of each node within the historical time window is transformed by a learnable linear transformation to generate query vector, key vector and value vector; The association weights between nodes are obtained by calculating the attention score of the scaled dot product of the query vector and the key vector and normalizing it using the Softmax function.

5. The method according to claim 1, characterized in that, The step of adjusting the adjacency matrix of the base graph structure according to the association weight includes: The adjacency matrix of the basic graph structure is multiplied element-wise with the matrix composed of the association weights to obtain the adjacency matrix of the time-varying heat propagation graph, and the strength of the basic topological connection is modulated by the association weights.

6. The method according to claim 1, characterized in that, The dual-branch temporal convolution module includes: The first branch uses a group of dilated convolutional kernels with an exponentially increasing dilation rate to capture feature dependencies at large time scales; the second branch uses a group of dilated convolutional kernels with a fixed dilation rate of 1 to extract local feature fluctuations at small time scales. The output feature maps of the two branches are fused to integrate long-term and short-term time series information.

7. The method according to claim 1, characterized in that, The formula for calculating the topology-aware weighted loss function L is as follows: ,in: Cross-entropy loss is used for fault type classification; α is the balancing hyperparameter. The weighted fault location loss is calculated as follows: Calculate the shortest path distance d between the predicted fault location and the actual fault location on the heating network topology map; The penalty weight w is calculated based on the distance d, and the weight increases as d increases; Multiplying the penalty weight w by the cross-entropy loss of the location positioning yields... .

8. A centralized heating network monitoring system based on artificial intelligence, characterized in that, Includes the following units: The data acquisition unit is used to acquire time-series data of multiple monitoring nodes in the heating pipeline network, including node pressure, temperature, flow rate and pipe wall acoustic vibration signals. The feature fusion unit is used to calculate the hydraulic imbalance and enthalpy difference change rate between adjacent nodes based on the node pressure, temperature and flow data, as physical model features; perform Mel frequency cepstral coefficient transformation on the pipe wall acoustic vibration signal to extract acoustic frequency domain features; and concatenate the physical model features with the acoustic frequency domain features to generate a multimodal fusion feature vector. The heat propagation graph construction unit is used to initialize the heating network topology into a basic graph structure, calculate the association weights between nodes based on the multimodal fusion feature vectors of each node at historical time using the relational attention mechanism, and adjust the adjacency matrix of the basic graph structure according to the association weights to construct a time-varying heat propagation graph. The feature extraction unit is used to take the multimodal fusion feature vector of each node at the current time as the node attribute, and perform graph convolution operation on the time-varying heat propagation map to extract spatial coupling features; The spatial coupling features are input into a dual-branch temporal convolution module. The first branch uses a group of dilated convolution kernels with a large dilation rate to extract long-term temporal dependencies, and the second branch uses a group of dilated convolution kernels with a small dilation rate to extract short-term temporal fluctuations. The outputs of the two branches are then fused to obtain a spatiotemporal representation. The monitoring unit is used to input the spatiotemporal representation into the fully connected network, output the fault type and location of each pipe segment, and optimize the parameters using a topology-aware weighted loss function during the model training phase. The topology-aware weighted loss function applies a weighted penalty to the location prediction deviation based on the shortest path distance between the predicted fault location and the actual fault location on the heating pipe network topology.

9. The system according to claim 8, characterized in that, The calculation of the hydraulic imbalance and enthalpy difference change rate between adjacent nodes includes: For a pipe segment consisting of upstream node i and downstream node j, the hydraulic imbalance at time t is equal to the upstream node flow rate. Traffic with downstream nodes The difference; The rate of change of enthalpy difference is calculated through the following steps: The thermal power loss of the pipe section at time t is calculated based on the node temperature T and the flow rate Q. , where c is the specific heat capacity of water and ρ is the density of water; Calculate the difference in thermal power loss between the current moment and the previous moment. The rate of change of enthalpy difference is obtained.

10. The system according to claim 8, characterized in that, The step of performing Mel-frequency cepstral coefficient transform on the acoustic vibration signal of the pipe wall to extract acoustic frequency domain features includes: The acoustic vibration signal of the pipe wall is divided into frames and windowed; a fast Fourier transform is performed on each frame of the signal to obtain the spectrum; Pass the spectrum through a Mel filter bank and calculate the logarithmic energy; The logarithmic energy is subjected to a discrete cosine transform, and the low-order coefficients are truncated as acoustic frequency domain features.

Images