Pipe section anomaly detection method based on pipe network topology
By constructing a spatiotemporal matrix and an error matrix combined with graph convolution and LSTM models, and utilizing dynamic thresholds and linkage mechanisms, multi-level anomaly detection of water supply networks is achieved. This solves the problem of insufficient sensitivity in detecting minute leaks in existing technologies, and realizes accurate and timely anomaly detection of water supply networks.
Patent Information
- Application Number
- CN202511485133.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-17
- Publication Date
- 2026-02-13
AI Technical Summary
Existing technologies for detecting anomalies in water supply networks are not sensitive enough to detect minor leaks, resulting in frequent false alarms and underreporting, making it difficult to achieve accurate and timely anomaly detection in water supply networks.
High-frequency sampling technology is used in combination with graph convolution and long short-term memory (LSTM) network models to construct spatiotemporal matrix and error matrix. Anomaly detection is performed by spatiotemporal data fusion and pipeline topology map. Multi-level anomaly detection is performed by using dynamic threshold and linkage mechanism.
It enables real-time monitoring of flow and pressure in the water supply network, improves the accuracy and timeliness of anomaly detection, captures instantaneous and long-term trend anomalies, reduces false alarms and missed alarms, and assists in fault location and decision-making.
Smart Images
Figure CN121520541A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of water supply network detection technology, and particularly relates to a method for detecting pipe section anomalies based on network topology. Background Technology
[0002] Existing anomaly detection methods for water supply networks primarily focus on leak detection, relying mainly on traditional and simple machine learning models such as clustering, trend decomposition, production and sales analysis, and threshold adjustment to identify anomalies. However, water supply networks are complex, highly coupled, and lagging systems. Current technologies are insufficiently sensitive to even minor leaks, leading to frequent false alarms and underreporting. Therefore, further exploration of multi-source data fusion (such as flow and pressure) detection is needed to achieve more sensitive and accurate anomaly management. Summary of the Invention
[0003] The purpose of this invention is to address the potential and complexity of abnormal operating conditions in pipeline networks by incorporating a layered detection (long-term and short-term) approach when constructing the model to analyze and manage abnormal situations, thereby improving the timeliness and accuracy of anomaly identification. Furthermore, this invention employs high-frequency sampling technology. For distant pipe sections within the network, where the impact of anomalies is minimal, high-frequency sampling increases the probability of detecting anomalies, thus improving the timeliness and accuracy of the abnormal emission early warning algorithm.
[0004] The present invention adopts the following technical solution:
[0005] A method for detecting pipe segment anomalies based on pipeline network topology includes the following steps:
[0006] S1. Data Acquisition: Obtain the set of water supply network segments ε = {e1, e2, ..., e...} M Each pipe segment e k Corresponding starting coordinates End point coordinates Obtain the set of monitoring points on the water supply network section
[0007] N = {1, 2, ..., n}, where each monitoring point n ∈ N collects two physical indicators in real time at each time t ∈ {1, 2, ..., T}: flow rate, denoted as Q. n (t)∈R; pressure (unit: MPa), denoted as P n (t)∈R; obtain the coordinates (x) of each monitoring point n∈N. (n) ,y (n) );
[0008] S2. Construct a spatiotemporal matrix and analyze the obtained traffic Q. n (t) and pressure P n(t) Data cleaning work is performed, including outlier handling and missing value handling; the flow rate Q at each time step t∈{1,2,...,T} is obtained using S1.2. n (t) Construct the traffic feature matrix Q: The pressure P at each time t∈{1,2,...,T} obtained using S1.2 n (t) Construct the pressure feature matrix P: For the obtained pipeline segments ε={e1,e2,...,e M Each pipe segment e in} k The geometric length is: The spatial geometric path is: Based on the coordinates (x, y, y) of all monitoring points n∈N on the pipe segment obtained in S1.2 (n) ,y (n) ), calculate the measuring point n using equation (1-3) i and n j On Line (k) All distances d(n) on i ,n j The cumulative distance of the shortest path along the pipeline segment is... Among them, P ij Represents all slave nodes n i to n j All paths, with all monitoring points n∈N as nodes, and the distance d between the monitoring points. ij As edge weights, construct the distance adjacency matrix. Where, if node n i to n j If not connected, then D ij =d ij =∞, using the adjacency matrix of equation (1-5) Calculate the weight matrix between measurement points for: Where ε>0 is a small constant used to prevent [the problem] in D. ij When n = 0, a division by zero occurs; if node n i to n j If not connected, then W' ij =0, for the weight matrix W' ij Normalization process yields
[0009] S3. Construct a spatiotemporal model to perform spatial structure modeling and time modeling.
[0010] S4. Anomaly Detection and Linkage Mechanism: After completing the flow and pressure prediction for the next τ time steps, the system constructs an error matrix to identify short-term and long-term anomalies and combines it with the pipeline topology map to construct a spatial linkage mechanism for qualitative and quantitative anomaly identification and fault source location.
[0011] Furthermore, the spatial structure utilizes W to perform graph convolution operations on the flow feature matrix Q and the pressure feature matrix P at each time step t, extracting an F-dimensional feature vector Q. t 'and P t ':
[0012] Q t '=σ(WQ t ),P t '=σ(WP t )
[0013] Where σ is the activation function ReLU, and F is an adjustable parameter. t 'and P t 'Feature concatenation yields spatially encoded features:'
[0014]
[0015] Based on spatial coding characteristics, the time steps are merged after spatial coding:
[0016]
[0017] Furthermore, temporal modeling involves converting the spatially encoded feature matrix after graph convolution. The input is fed into the Long Short-Term Memory (LSTM) network of the time module, and the hidden state is updated recursively:
[0018] h t ,c t =LSTM(F t ',h t-1 ,c t-1 )
[0019] in, The hidden state vector at step t; The unit memory vector at step t; h0 and c0 are H-dimensional all-zero vectors, and the hidden state dimension H is an adjustable parameter.
[0020] Furthermore, the spatial structure also uses a fully connected layer (FC) to connect the hidden state vector h of (1-11) t Mapped to predicted flow and pressure values for future time steps:
[0021]
[0022] Among them, mapping layer parameters and It is learned automatically during training, and τ is the number of prediction steps.
[0023] Furthermore, the anomaly detection and linkage mechanism also includes constructing an error matrix and collecting actual flow and pressure data to form an observation matrix:
[0024]
[0025] Construct the absolute error matrix using equations (1-12) and (1-13):
[0026]
[0027] Where E(t,i) represents the prediction error at detection point i in the t-th prediction step.
[0028] Furthermore, the anomaly detection and linkage mechanism also includes capturing sudden changes at the hourly level. The short-term anomaly detection mechanism constructs a sliding window of length w based on the flow and pressure data of the past hour. For each time point t≥w, the error subsequence of the past w steps is extracted based on the overall error matrix.
[0029]
[0030] Statistical sliding window error subsequence Historical mean μ i (t), calculate the standard deviation σ i (t)
[0031]
[0032] Then, based on the rule of thumb (3σ criterion), the dynamic threshold is set as follows:
[0033]
[0034] According to the formula for determining short-term anomalies in pipeline sections:
[0035]
[0036] Furthermore, the anomaly detection and linkage mechanism also includes a long-term anomaly detection mechanism based on trend changes over the past 7 days, calculating the cumulative error deviation using the CUSUM method, and employing a fixed-length sliding window W. L The mean error of each monitoring point is dynamically updated, and the mean error of monitoring point i is calculated. and standard deviation σ i :
[0037]
[0038] Define the error deviation as:
[0039]
[0040] The cumulative sum in recursive form is constructed from the error deviation:
[0041] S i (t)=max(0,S i (t-1)+d i (t))
[0042] Wherein, the initial value S i (0)=0
[0043] Based on the rule of thumb (3σ criterion), an alarm threshold based on equation (1-21) is set. The formula for determining long-term anomalies in the pipeline segment (defining 0 as normal and 1 as abnormal) is as follows:
[0044]
[0045] Furthermore, the anomaly detection and linkage mechanism also includes anomaly determination, utilizing the topology graph weight matrix. Characterizes the physical connectivity between detection points; for any detection point i, defines its abnormal linkage coefficient L. i (t), representing the "intensity of anomaly" perceived by measuring point i at the current time t:
[0046]
[0047] Where 1{·} is the indicator function, representing whether the surrounding measurement point j is abnormal. These indicate whether the surrounding measuring point j exhibits short-term or long-term anomalies at the current moment;
[0048] Pipeline segment anomaly identification function Define 0 as normal, 1 as isolated anomaly, and 2 as linked anomaly:
[0049]
[0050] Where, η i It is the system linkage threshold:
[0051] Compared with existing technologies, this invention achieves comprehensive monitoring of flow and pressure anomalies at multiple measuring points in a water supply network through data preprocessing, pipeline topology modeling, real-time and long-term multi-level anomaly detection, and linkage analysis. Its main advantages are:
[0052] 1. Combined short-term and long-term detection: It can capture instantaneous anomalies in real time and detect long-term trend anomalies, and is suitable for various abnormal situations such as water leakage and water theft.
[0053] 2. Dynamic threshold: Dynamic thresholds are used to trigger alarms in real-time anomaly detection to avoid false alarms / missed alarms caused by static thresholds, especially in systems such as water supply networks that are highly periodic, noisy, and have rapidly changing states.
[0054] 3. Spatiotemporal data fusion: Spatiotemporal graph convolution and LSTM model are used to jointly model pipeline topology and time series data to improve the accuracy and robustness of anomaly detection.
[0055] 4. Quantitative analysis: Mathematical formulas are used to describe each key link to ensure that system parameters are quantifiable and easy to optimize.
[0056] 5. System linkage: Construct a distance matrix between measurement points based on a two-way graph to realize cross-regional anomaly linkage analysis and assist in fault location and decision-making. Attached Figure Description
[0057] Figure 1 This is a schematic diagram of the detection process of a pipe segment anomaly detection method based on pipeline network topology proposed in this invention. Detailed Implementation
[0058] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0059] A pipe segment anomaly detection method based on pipeline network topology, comprising:
[0060] S1 Data Acquisition
[0061] S1.1 Obtain the set of water supply network segments ε={e1,e2,...,e M Each pipe segment e k Corresponding starting coordinates End point coordinates
[0062] S1.2 Obtain the set of monitoring points N = {1, 2, ..., n} on the water supply network section, where each monitoring point...
[0063] Two physical parameters are collected in real time at each time t∈{1,2,…,T} for n∈N:
[0064] Flow rate (unit: m) 3 / h), denoted as Q n (t)∈R;
[0065] Pressure (unit: MPa), denoted as P n (t)∈R.
[0066] Obtain the coordinates (x, y) of each monitoring point n∈N (n) ,y (n) ).
[0067] S2 constructs the spacetime matrix
[0068] S2.1 Data Preprocessing
[0069] a) The traffic Q obtained in S1.2 n (t) and pressure P n (t) Data cleaning work such as outlier handling and missing value handling is performed;
[0070] S2.2 Feature Engineering
[0071] a) Utilize the flow rate Q at each time t∈{1,2,...,T} obtained from S1.2 n (t) Construct the traffic feature matrix Q:
[0072]
[0073] The pressure P at each time t∈{1,2,...,T} obtained using S1.2 n (t) Construct the pressure feature matrix P:
[0074]
[0075] S2.3 Construction of Measurement Point Weight Matrix
[0076] a) For the pipeline segments ε={e1,e2,...,e2,e3} obtained in S1.1 M Each pipe segment e in} k
[0077] The geometric length is:
[0078]
[0079] The spatial geometric path is:
[0080]
[0081] b) Based on the coordinates (x, y, y) of all monitoring points n∈N on the pipe segment obtained in S1.2 (n) ,y (n) ), calculate the measuring point n using equation (1-3) i and n j On Line (k) All distances d(n) on i ,n j The cumulative distance of the shortest path along the pipeline segment is...
[0082]
[0083] Among them, P ij Represents all slave nodes n ito n j All paths
[0084] All monitoring points n∈N are taken as nodes, and the distance d between the monitoring points is... ij As edge weights, construct the distance adjacency matrix.
[0085]
[0086] Where, if node n i to n j If not connected, then D ij =d ij =∞.
[0087] c) Construction of the measurement point weight matrix: In anomaly detection and linkage analysis, we usually want the weight to be lower for farther distances, thus reflecting the physical connectivity between nodes in the graph model. Using the adjacency matrix of equation (1-5)... Calculate the weight matrix between measurement points for:
[0088]
[0089] Where ε>0 is a small constant used to prevent [the problem] in D. ij When D = 0, a division by zero occurs; when D ij When smaller, W' ij A larger value indicates a lower value; conversely, a smaller value indicates a lower value. ij When it is large, W' ij The value decreases, thus reflecting the principle that "the farther the distance, the lower the weight." If node n i to n j If not connected, then W' ij =0.
[0090] For the weight matrix W' (1-6) ij Normalization process yields
[0091] S3 Spatiotemporal Model Construction
[0092] S3.1 Spatial Structure Modeling
[0093] a) Using W from (1-7), perform graph convolution operations at each time step t on the flow feature matrix Q and pressure feature matrix P in S2.2 to extract the F-dimensional feature vector Q. t 'and P t ':
[0094] Q t '=σ(WQ t ),P t '=σ(WP t (1-8)
[0095] Where σ is the activation function ReLU, and F is an adjustable parameter (usually 16, 32, or 64).
[0096] Q t 'and P t 'Feature concatenation yields spatially encoded features:'
[0097]
[0098] According to (1-9), the time steps are merged after spatial encoding:
[0099]
[0100] S3.2 Time Modeling
[0101] a) Encode the feature matrix of the (1-10) spatial region after graph convolution. The input is fed into the Long Short-Term Memory (LSTM) network of the time module, and the hidden state is updated recursively:
[0102] h t ,c t =LSTM(F t ',h t-1 ,c t-1 (1-11)
[0103] in, The hidden state vector at step t; The unit memory vector at step t; h0 and c0 are H-dimensional all-zero vectors, and the hidden state dimension H is an adjustable parameter (usually taken as 32, 64, or 128). b) The hidden state vector h0 of step t is stored in a fully connected layer (FC). t Mapped to predicted flow and pressure values for future time steps:
[0104]
[0105] Among them, mapping layer parameters and It is learned automatically during training, and τ is the number of prediction steps.
[0106] S4 Anomaly Detection and Linkage Mechanism
[0107] After completing the flow and pressure prediction for the next τ time steps, the system constructs an error matrix to identify short-term and long-term anomalies, and combines it with the pipeline topology map to construct a spatial linkage mechanism for qualitative and quantitative anomaly identification and fault source location.
[0108] S4.1 Error Matrix Construction
[0109] Collect actual flow and pressure data to construct an observation matrix:
[0110]
[0111] Construct the absolute error matrix using equations (1-12) and (1-13):
[0112]
[0113] Where E(t,i) represents the prediction error at detection point i (flow rate or pressure) in the t-th prediction step.
[0114] S4.2 Short-term anomaly detection mechanism
[0115] a) To capture sudden changes at the hourly level, the short-term anomaly detection mechanism constructs a sliding window of length w (e.g., 60 minutes) based on the flow and pressure data of the past hour. For each time point t≥w, the error subsequence of the past w steps is extracted according to the overall error matrix (1-14):
[0116]
[0117] b) Statistical formula (1-15) Error subsequence within the sliding window Historical mean μ i (t), calculate the standard deviation σ i (t)
[0118]
[0119] Then, based on the rule of thumb (3σ criterion), the dynamic threshold is set as follows:
[0120]
[0121] Based on (1-15) and (1-17), the formula for judging short-term anomalies in pipe sections (defining 0 as normal and 1 as abnormal) is as follows:
[0122]
[0123] S4.3 Long-term anomaly detection mechanism
[0124] a) The long-term anomaly detection mechanism is based on trend changes over the past 7 days. It calculates the cumulative error deviation using the CUSUM method and employs a fixed-length sliding window W. L (For example, 2016 means that data is collected every 5 minutes.) The mean error of each monitoring point is dynamically updated, and the mean error of measuring point i is calculated according to formula (1-16). and standard deviation σ i :
[0125]
[0126] b) The error deviation is defined according to equations (1-16) and (1-19) as follows:
[0127]
[0128] The cumulative sum in recursive form is constructed from the error deviation:
[0129] S i (t)=max(0,S i (t-1)+d i (t)) (1-22)
[0130] Wherein, the initial value S i (0)=0
[0131] Based on the rule of thumb (3σ criterion), an alarm threshold based on equation (1-21) is set. The formula for determining long-term anomalies in the pipeline segment (defining 0 as normal and 1 as abnormal) is as follows:
[0132]
[0133] S4.3 Anomaly detection utilizes the topology graph weight matrix of equation (1-7). Characterizes the physical connectivity between detection points.
[0134] a) For any measuring point i, define its abnormal linkage coefficient L. i (t), representing the "intensity of anomaly" perceived by measuring point i at the current time t:
[0135]
[0136] Where 1{·} is the indicator function, representing whether the surrounding measurement point j is abnormal. These indicate whether the surrounding measuring point j exhibits short-term or long-term anomalies at the current moment.
[0137] b) Then, from (1-18)(1-23)(1-24), we obtain the anomaly identification function for the pipe segment anomaly level. (Define 0 as normal, 1 as isolated anomaly, and 2 as linked anomaly):
[0138]
[0139] Where, η i It is the system linkage threshold:
[0140] Finally, 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. A method for detecting pipe segment anomalies based on pipeline network topology, characterized in that, include: S1. Data Acquisition: Obtain the set of water supply network segments ε = {e1, e2, ..., e...} M Each pipe segment e k Corresponding starting coordinates End point coordinates Obtain the set of monitoring points N = {1, 2, ..., n} on the water supply network section. At each time t ∈ {1, 2, ..., T}, collect two physical indicators in real time: flow rate, denoted as Q. n (t)∈R; pressure (unit: MPa), denoted as P n (t)∈R; obtain the coordinates (x) of each monitoring point n∈N. (n) ,y (n) ); S2. Construct a spatiotemporal matrix and analyze the obtained traffic Q. n (t) and pressure P n (t) Data cleaning work is performed, including outlier handling and missing value handling; the flow rate Q at each time step t∈{1,2,...,T} is obtained using S1.
2. n (t) Construct the traffic feature matrix Q: The pressure P at each time t∈{1,2,...,T} obtained using S1.2 n (t) Construct the pressure feature matrix P: For the obtained pipeline segments ε={e1,e2,...,e M Each pipe segment e in} k The geometric length is: The spatial geometric path is: Based on the coordinates (x, y, y) of all monitoring points n∈N on the pipe segment obtained in S1.2 (n) ,y (n) ), calculate the measuring point n using equation (1-3) i and n j On Line (k) All distances d(n) on i ,n j The cumulative distance of the shortest path along the pipeline segment is... Among them, P ij Represents all slave nodes n i to n j All paths, with all monitoring points n∈N as nodes, and the distance d between the monitoring points. ij As edge weights, construct the distance adjacency matrix. Where, if node n i to n j If not connected, then D ij =d ij =∞, using the adjacency matrix of equation (1-5) Calculate the weight matrix between measurement points for: Where ε>0 is a small constant used to prevent [the problem] in D. ij When n = 0, a division by zero occurs; if node n i to n j If not connected, then W' ij =0, for the weight matrix W' ij Normalization process yields S3. Construct a spatiotemporal model to perform spatial structure modeling and time modeling. S4. Anomaly Detection and Linkage Mechanism: After completing the flow and pressure prediction for the next τ time steps, the system constructs an error matrix to identify short-term and long-term anomalies and combines it with the pipeline topology map to construct a spatial linkage mechanism for qualitative and quantitative anomaly identification and fault source location.
2. The method for detecting pipe segment anomalies based on pipeline network topology according to claim 1, characterized in that, The spatial structure utilizes W to perform graph convolution operations on the flow feature matrix Q and the pressure feature matrix P at each time step t, extracting an F-dimensional feature vector Q. t 'and P t ': Q t '=σ(WQ t ),P t '=σ(WP t ) Where σ is the activation function ReLU, and F is an adjustable parameter. t 'and P t 'Feature concatenation yields spatially encoded features:' Based on spatial coding characteristics, the time steps are merged after spatial coding:
3. A method for detecting pipe segment anomalies based on pipeline network topology according to claim 1, characterized in that, Temporal modeling is the spatial encoding feature matrix after graph convolution. The input is fed into the Long Short-Term Memory (LSTM) network of the time module, and the hidden state is updated recursively: h t ,c t =LSTM(F t ',h t-1 ,c t-1 ) in, The hidden state vector at step t; The unit memory vector at step t; h0 and c0 are H-dimensional all-zero vectors, and the hidden state dimension H is an adjustable parameter.
4. A pipe segment anomaly detection method based on pipeline network topology according to claim 3, characterized in that, The spatial structure also uses a fully connected layer (FC) to connect the hidden state vector h of (1-11). t Mapped to predicted flow and pressure values for future time steps: Among them, mapping layer parameters and It is learned automatically during training, and τ is the number of prediction steps.
5. A pipe segment anomaly detection method based on pipe network topology according to claim 4, characterized in that, The anomaly detection and linkage mechanism also includes constructing an error matrix and collecting actual flow and pressure data to form an observation matrix. Construct the absolute error matrix using equations (1-12) and (1-13): Where E(t,i) represents the prediction error at detection point i in the t-th prediction step.
6. A method for detecting pipe segment anomalies based on pipeline network topology according to claim 5, characterized in that, The anomaly detection and linkage mechanism also includes capturing sudden changes at the hourly level. The short-term anomaly detection mechanism constructs a sliding window of length w based on the flow and pressure data of the past hour. For each time point t≥w, the error subsequence of the past w steps is extracted based on the overall error matrix. Statistical sliding window error subsequence Historical mean μ i (t), calculate the standard deviation σ i (t) Then, based on the rule of thumb (3σ criterion), the dynamic threshold is set as follows: According to the formula for determining short-term anomalies in pipeline sections:
7. A method for detecting pipe segment anomalies based on pipeline network topology according to claim 5, characterized in that, The anomaly detection and linkage mechanism also includes a long-term anomaly detection mechanism based on trend changes over the past 7 days. It calculates the cumulative error deviation using the CUSUM method and employs a fixed-length sliding window W. L The mean error of each monitoring point is dynamically updated, and the mean error of monitoring point i is calculated. and standard deviation σ i : Define the error deviation as: The cumulative sum in recursive form is constructed from the error deviation: S i (t)=max(0,S i (t-1)+d i (t)) Wherein, the initial value S i (0)=0 Based on the rule of thumb (3σ criterion), an alarm threshold based on equation (1-21) is set. The formula for determining long-term anomalies in the pipeline segment (defining 0 as normal and 1 as abnormal) is as follows:
8. A method for detecting pipe segment anomalies based on pipeline network topology according to claim 7, characterized in that, The anomaly detection and linkage mechanism also includes anomaly determination, which utilizes the topology graph weight matrix. Characterizes the physical connectivity between detection points; for any detection point i, defines its abnormal linkage coefficient L. i (t), representing the intensity of the surrounding "anomaly" perceived by measuring point i at the current time t: Where 1{·} is the indicator function, representing whether the surrounding measurement point j is abnormal. These indicate whether the surrounding measuring point j exhibits short-term or long-term anomalies at the current moment; Pipeline segment anomaly identification function Define 0 as normal, 1 as isolated anomaly, and 2 as linked anomaly: Where, η i It is the system linkage threshold: