Method for identifying blocked pipe section of drainage pipe network system based on dynamic characteristics

Abstract

The invention discloses a blocked pipe section identification method of a drainage pipe network system based on dynamic characteristics, and relates to the technical field of drainage pipe network monitoring. Real-time hydraulic association between pipe network nodes is quantified through a dynamic adjacency matrix to generate a spatial topology matrix, and a time sequence is divided based on a sliding window; a space-time fusion model based on GCN and Transform is constructed to carry out multi-scale dynamic coding, and then a decoder is utilized to reconstruct normal working condition data of a pipe network. By calculating the deviation degree of the pipe section level reconstruction error and the threshold value, the blocked pipe section is accurately recognized, end-to-end modeling from'pipe network topology-drainage time sequence data-external rainfall 'multi-source data to blocked pipe section recognition is achieved, and the technical difficulties of a traditional method in the aspects of dynamic topology modeling, long time sequence dependence capture and multi-modal feature fusion are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sewer network monitoring, and in particular to a method for identifying a blocked pipe section of a sewer network system based on dynamic characteristics. BACKGROUND

[0002] As an important part of the city lifeline, the safe operation of the urban sewer network system is directly related to urban flood control, environmental protection and public safety. Through intelligent blocked pipe section identification technology, real-time monitoring and accurate early warning of the operation state of the pipe network can be achieved, effectively preventing secondary disasters such as urban waterlogging and road collapse, and significantly improving the city's disaster prevention and mitigation capabilities. This technology combines Internet of Things sensing and artificial intelligence analysis, not only greatly reducing the cost of manual inspection and improving operational efficiency, but also extending the service life of the pipe network and avoiding sewage overflow pollution, providing important data support and decision-making basis for smart city construction. It is a key technical means to achieve fine management of the city and ensure sustainable development of the city.

[0003] The prior art has the following several kinds:

[0004] Traditional detection methods: relying on manual inspection, sensor deployment or water model simulation, with high cost, limited coverage, poor real-time performance and other problems.

[0005] Image / video-based detection: identifying pipe defects through CCTV or sonar images, but limited by data quality, environmental interference and generalization ability under complex working conditions.

[0006] Single machine learning model: such as support vector machine or random forest, only using local features, and difficult to model the pipe network topology structure and time series dynamic characteristics.

[0007] The defects of the prior art include:

[0008] Ignoring the pipe network graph structure features (such as node connection relationship, flow direction dependence);

[0009] Insufficient long process modeling capability for time series data (such as flow, liquid level fluctuation);

[0010] Unable to achieve end-to-end "topology + time series" multi-modal feature fusion, resulting in high missed detection rate and low positioning accuracy.

[0011] Summary of the application

[0012] The purpose of the present application is to solve the problems in the prior art and provide a method for identifying a blocked pipe section of a sewer network system based on dynamic characteristics.

[0013] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0014] S1. Construct a pipe network directed graph model, and dynamically update the adjacency matrix based on real-time flow data:

[0015]

[0016] wherein Q i (t) is the real-time flow of the node, and a, b are weight coefficients;

[0017] S2. Divide the multi-source time series data by a sliding window to generate a space-time feature matrix:

[0018]

[0019] The multi-dimensional time series data includes flow, liquid level, flow rate, and dynamic adjacency matrix;

[0020] S3. Construct a space-time fusion encoder:

[0021] Spatial encoding layer: adopt an L-layer graph convolution network to extract node spatial features;

[0022] Time encoding layer: adopt an L-layer Transformer encoder to extract time series features;

[0023] Multi-modal fusion: perform feature-level fusion of rainfall time series data and node time series features;

[0024] S4. Construct a space-time decoder:

[0025] Reverse Transformer decoder to restore time series features;

[0026] Reverse graph convolution network to restore spatial features;

[0027] Calculate reconstruction error

[0028] S5. Abnormality detection:

[0029] Establish a threshold value τ = μ recon +k·σ recon based on the reconstruction error of the verification set, and determine abnormal nodes by node-level error

[0030] Determine pipe segment blockage in combination with topological relationship

[0031] As a further improvement of the present application, the dynamic adjacency matrix satisfies:

[0032] Downstream nodes only receive upstream node information;

[0033] Weight coefficients a ∈ [0.1, 0.3], b ∈ [0.5, 0.8];

[0034] Real-time synchronization of traffic data through edge computing devices ensures that the adjacency matrix update delay is ≤5 minutes.

[0035] As a further improvement of the application, the spatio-temporal encoder comprises:

[0036] 3-layer graph convolutional network, with the number of channels in each layer being 16-32-64;

[0037] 4-layer Transformer encoder, with the hidden layer dimension being 512 and the number of heads being 8;

[0038] The position encoding adopts the sine-cosine function (5.16-5.17).

[0039] As a further improvement of the application, the sliding window parameter is set to:

[0040] The window length ω ∈ [6, 24] hours;

[0041] The step size S = ω / 2;

[0042] The input data dimension where N ≥ 10 nodes.

[0043] As a further improvement of the application, the multi-source data fusion comprises:

[0044] The rainfall data is linearly mapped to the node feature dimension;

[0045] Feature-level weighted and fused:

[0046] Spatio-temporal feature joint attention mechanism.

[0047] As a further improvement of the application, the decoder comprises:

[0048] 3-layer deconvolutional graph network to restore spatial dimensions;

[0049] 4-layer Transformer decoder to restore temporal dimensions;

[0050] The reconstruction loss function adopts mean square error.

[0051] As a further improvement of the application, the anomaly determination comprises:

[0052] The anomaly determination of the last 3 time windows is regarded as suspected congestion;

[0053] The determination at the pipe segment level requires that the upstream and downstream nodes are simultaneously abnormal;

[0054] Triggering the early warning requires 5 consecutive abnormal determinations.

[0055] As a further improvement of the application, the model training adopts:

[0056] Normal operating condition data unsupervised pre-training

[0057] Adam optimizer, learning rate 1e-4

[0058] Batch normalization and Dropout regularization

[0059] As a further improvement of the application, the system comprises:

[0060] The sliding window generation module, the space-time coding and decoding module, the anomaly detection module and the early warning output module.

[0061] As a further improvement of the application, the system further comprises:

[0062] The real-time data acquisition module and the dynamic graph construction module.

[0063] 1. Dynamic graph construction method

[0064] Traditional pipe network graph structure modeling relies on static topology (such as fixed pipe segment connection relationship), which cannot reflect the influence of real-time water flow state on the correlation strength between nodes, resulting in deviation in spatial feature extraction. The application dynamically adjusts the weight of the adjacency matrix based on real-time sensor data, captures the real-time state of the pipe network, can flexibly cope with changes in the pipe network, and has stronger generalization ability.

[0065] 2. Spatio-temporal feature joint modeling and fusion technology

[0066] A single model cannot capture both the spatial topology dependence of the pipe network and the long-term dynamic changes, and traditional feature fusion methods (such as splicing and weighted averaging) cannot adapt to the contribution differences of spatio-temporal features. The application can quickly and accurately locate the position of the blocked pipe segment by extracting information of the entire drainage pipe network through GCN and Transformer.

[0067] 3. Unsupervised training framework

[0068] In real scenarios, there is a lack of labeled data for anomalies. The application trains an autoencoder structure model on normal historical data, sets an anomaly threshold through reconstruction error to identify blocked pipe segments in actual applications, and does not require a large amount of labeled data.

[0069] The application effectively captures the topology structure of the pipe network using GCN and effectively captures the long-term data features using Transformer, fully excavating the potential information of data at different nodes and time points. The application can accurately locate the position of the blocked pipe segment at an early time and give an early warning, ensuring the healthy operation of the urban drainage pipe network. The algorithm has strong generalization performance and is suitable for identifying blocked pipe segments in drainage pipe network systems of different cities. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 Model training pre-warning flowchart. DETAILED DESCRIPTION

[0071] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application.

[0072] 1. Data acquisition and preprocessing

[0073] 1.1 Drainage pipe network topology data acquisition

[0074] Let the pipe network be a directed graph G=(V, E), where:

[0075] V={v1, v2,..., v N} represents the nodes such as inspection wells, outlets, etc.

[0076] E={e ij =(v i , v j )} represents the pipe sections, and the direction is from v i to v j .

[0077] Then its adjacency matrix is:

[0078]

[0079] wherein the element a ij =1 represents that there is a pipe section e ij , otherwise a ij =0.

[0080] The traditional pipe network graph structure modeling relies on static topology (such as fixed pipe section connection relationship), and cannot reflect the influence of real-time water flow state on the correlation strength between nodes, resulting in deviation in spatial feature extraction. In the present scheme, a dynamic adjacency matrix is used, and the update formula is as follows:

[0081]

[0082] In the formula:

[0083] Q i (t) is the real-time flow of node i, and a, b are weight coefficients. In order to avoid the interference of counterflow, the downstream node only receives upstream information.

[0084] Then the dynamic adjacency matrix at time t is:

[0085]

[0086] For data with a time series length of T, the dynamic adjacency matrix is:

[0087]

[0088] 1.1.2 Obtain sewer network drainage data

[0089] Drainage-related time series data: assuming that the sewer network has N nodes, the time series length is T, and the monitoring indicators are flow (Q), liquid level (H), and flow rate (V), for each time node t (t = 1,...T), the matrix form of the network data is:

[0090]

[0091] wherein Q , H , and V

[0092] represent the flow, liquid level, and flow rate of node i at time t, respectively.

[0093]

[0094] Influencing factors: rainfall. For a single site rainfall time series, assuming the observation length is T time points, then it can be represented by a matrix as:

[0095] P = [p1, p2,..., p T ](1.7)

[0096] wherein p t t represents the rainfall at the tth time step.

[0097] 1.2 Data preprocessing

[0098] Missing value processing

[0099] Outlier processing

[0100] Data standardization

[0101] Data set division

[0102] 1.3 Sliding window sample generation

[0103] In this scheme, the model needs to reconstruct data for multiple time steps, so the time series needs to be divided into multiple windows. Assuming that the window size is w and the step size is s, then for the drainage-related time series data, the sub-matrix of the kth window is:

[0104]

[0105] For the dynamic adjacency matrix, the sub-matrix of the kth window is:

[0106]

[0107] For rainfall data, the sub-matrix of the kth window is:

[0108]

[0109] where t k = 1 + (k - 1) s, and K is the number of data generated after applying the sliding window.

[0110] The drainage-related data generated by the sliding window is represented by the matrix:

[0111]

[0112] The generated dynamic adjacency matrix data is represented by the matrix:

[0113]

[0114] The generated rainfall data is represented by the matrix:

[0115]

[0116] where x k , b k and r k correspond to the data of a time window.

[0117] 1.4 Model Introduction

[0118] 1.4.1 Graph Convolutional Network (GCN) Input: Node Feature Matrix Adjacency Matrix

[0119] GCN update formula:

[0120]

[0121] where, (add self-loop), is the degree matrix, W (l) is a learnable parameter, and σ is an activation function (such as ReLU).

[0122] Output: High-order node embedding vector representing the spatial dependency between nodes.

[0123] 1.4.2 Time Encoder (Transformer)

[0124] Input: Time-series feature matrix

[0125] Perform word embedding dimension encoding on the input to obtain the encoded data:

[0126]

[0127] Add Positional Encoding. Positional Encoding (PE) generates a unique vector representation for each position in the sequence using sine-cosine positional encoding, and the dimension of the encoded data is The formula is as follows:

[0128]

[0129] Where pos represents the position of the token in the sequence (starting from 0), i represents the dimension index (0≤i model / 2), d model represents the embedding dimension.

[0130] Attention weight formula:

[0131]

[0132] Where Q, K, V are input and three weight matrices W Q ,W K ,W V are multiplied respectively. The formula for calculating the Transformer encoder module is:

[0133] TansformerEncoder(X) (1.19)

[0134] =LayerNorm(FFN(LayerNorm(X+MultiHead(XW Q ,XW K ,XW V )))

[0135] Where,

[0136] MultiHead is multi-head attention: Concat(head1,...,head h )W o .

[0137] Each attention head: head i =Attention(Q,K,V).

[0138] FFN is a feedforward network: ReLU(xW1+b1)W2+b2.

[0139] The formula for calculating the Transformer decoder module is:

[0140] The decoder adds mask self-attention and cross-attention on the basis of the encoder:

[0141] TransformerDecoder(Y,Z) (1.20)

[0142] LayerNorm(FFN(LayerNorm(Y+MultiHead(YW Q ,YW K ,YW V )))+CrossAttn(Y,Z))

[0143] where:

[0144] Masked attention is implemented by a lower triangular matrix M:

[0145]

[0146] K, V of CrossAttn come from encoder output Z, Q comes from decoder input Y.

[0147] Output: time-series feature vector Characterize the dynamic behavior patterns of each node.

[0148] 1.5 Model construction and training

[0149] 1.5.1 Model structure

[0150] Original input data: defined by equations (1.11), (1.12), (1.13), respectively. Where K represents the number of samples, w represents the length of the time series, N represents the number of nodes, and 3 is the number of node features.

[0151] 1.5.1.1 Spatial encoder

[0152] Use multi-layer GCN to extract spatial features, perform graph convolution on the node features of each time step, and then merge the time steps.

[0153] Input: node feature matrix of each time step and dynamic adjacency matrix (total w time steps).

[0154] Processing flow: time-step-by-time-step spatial encoding, for each time step t = 1, 2,... w, independently perform GCN encoding, defined by equation (1.14):

[0155]

[0156] where, (initial input).

[0157] After L-layer GCN encoding, merge the time steps:

[0158]

[0159] Output:

[0160] 1.5.1.2 Temporal Encoder

[0161] Use multi-layer Transformer to extract time series features, encode the features of each node in the time dimension.

[0162] Input: Spatially encoded features

[0163] Rainfall time series

[0164] Process flow:

[0165] (1) Convert the input matrix H GCN Transformed into the format required by Transformer through matrix transformation:

[0166]

[0167] (2) Encode the node time series features using equations (1.15), (1.16), (1.17) and add position encoding:

[0168]

[0169] (3) Rainfall time series encoding: mapped to the same dimension as node features through a linear layer:

[0170]

[0171] (4) Fuse node feature data and rainfall data as encoder input data:

[0172]

[0173] (5) Use L-layer Transformer encoder to extract time series features, the formula is given by (1.19):

[0174] H (l+1) = TansformerEncoder(H (l) ) (1.26)

[0175] Where, (initial input).

[0176] Output: Time series features with rainfall fusion

[0177] 1.5.1.3 Temporal Decoder

[0178] Data shape is recovered using a multi-layer Transformer decoder.

[0179] Input: latent representation H encode with encoder input

[0180] Process flow:

[0181] Data shape is recovered using a L-layer Transformer decoder, formula given by (1.20):

[0182] H (l+1) = TansformerDecoder(H (l) , H encode ) (1.27)

[0183] where, (initial input).

[0184] Decoder output is transformed into

[0185] Output: temporal feature decoding

[0186] 1.5.1.4 Spatial decoder

[0187] Data original dimension is recovered using a multi-layer GCN decoder.

[0188] Input: H decode with dynamic adjacency matrix B.

[0189] Output: Similar to GCN encoder process flow, output reconstructed node features:

[0190]

[0191] 1.5.1.5 Loss function

[0192]

[0193] 1.6 Unconventional running pipe segment identification

[0194] 1.6.1 Threshold setting

[0195] Reconstruction error of normal data follows Gaussian distribution, threshold is determined by calculating statistical characteristics of validation set error. Reconstruction error formula is:

[0196]

[0197] where, x (i) is the input data of the i-th pipe segment, W is the number of time steps, and F is the number of features.

[0198] The threshold formula is:

[0199] τ = μ recon + k · σ recon (1.30)

[0200] where μ recon is the mean of reconstruction errors of normal samples in the validation set, σ recon is the standard deviation of reconstruction errors of normal samples in the validation set, and k is a tunable parameter (usually 3-5, corresponding to 99.7%-99.99% confidence interval).

[0201] 1.6.2 Abnormality detection

[0202] (1) Data input and reconstruction:

[0203] Real-time or batch input window data

[0204] Self-encoder output reconstruction data

[0205] (2) Reconstruction error calculation:

[0206] Calculate the error of each time window by node:

[0207]

[0208] (3) Abnormality determination:

[0209] At time t, the clogging node determination formula (0 for normal, 1 for abnormal):

[0210]

[0211] Clogging pipe segment determination formula:

[0212]

[0213] where a ij is defined by formula (1.1).

[0214] In the next N time windows, count the number of abnormal occurrences and trigger an early warning:

[0215]

[0216] where N is the number of consecutive windows monitored, and m is the threshold for confirming abnormality, m≤N.

[0217] Improve timeliness and accuracy

[0218] Real-time performance and dynamic response capability: GCN directly models the connection relationship between pipeline nodes and pipe segments through graph structure, without relying on complex physical equations or manual rules, and can complete the dynamic update of topological features in milliseconds; Transformer's self-attention mechanism supports parallel processing of long-term sequences, avoiding the recursive delay problem of traditional RNN / LSTM.

[0219] Optimization of detection accuracy and false alarm rate: Spatial modeling based on GCN captures pipeline topology dependencies (such as upstream blockage causing downstream flow reduction), reducing false alarms caused by local sensor noise; Temporal modeling based on Transformer analyzes flow periodicity (such as morning and evening peak hours) and abrupt change patterns (such as rainstorm impact), distinguishing between normal fluctuations and actual blockages.

[0220] Comprehensive analysis of the entire drainage network

[0221] GCN's graph structure awareness capability: This invention abstracts the pipeline network into a directed graph, explicitly models the attributes of nodes (manholes) and edges (pipe segments), and identifies abnormal propagation paths.

[0222] System-level anomaly localization: Traditional methods can only identify local sensor anomalies, while the GCN-Transformer model, through global feature propagation, can accurately locate the blocked pipe segment, rather than falsely reporting all nodes.

[0223] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying blocked pipe sections in a drainage pipe network system based on dynamic features, characterized in that, The method includes the following steps: S1. Construct a directed graph model of the pipeline network and dynamically update the adjacency matrix based on real-time traffic data: Among them, Q i (t) represents the real-time traffic of the node, and α and β are the weighting coefficients; S2. Divide multi-source time-series data using a sliding window to generate a spatiotemporal feature matrix: Multidimensional time-series data including flow rate, liquid level, flow velocity, and dynamic adjacency matrix; S3. Construct a spatiotemporal fusion encoder: Spatial coding layer: An L-layer graph convolutional network is used to extract spatial features of nodes; Temporal coding layer: An L-layer Transformer encoder is used to extract temporal features; Multimodal fusion: merging rainfall time-series data with node time-series features at the feature level; S4. Constructing a spatiotemporal decoder: The inverse Transformer decoder recovers the timing features; Inverse graph convolutional networks recover spatial features; Calculate reconstruction error S5. Anomaly Detection: A threshold τ = μ is established based on the validation set reconstruction error. recon +k·σ recon , Node-level error determination Mark abnormal nodes Determining pipe blockage by combining topological relationships 2. The method for identifying blocked pipe sections in a drainage pipe network system based on dynamic features according to claim 1, characterized in that, The dynamic adjacency matrix satisfies: Downstream nodes only receive information from upstream nodes; The weighting coefficients α∈[0.1,0.3], β∈[0.5,0.8]; Traffic data is synchronized in real time through edge computing devices to ensure that the adjacency matrix update delay is ≤5 minutes.

3. The method for identifying blocked pipe sections in a drainage pipe network system based on dynamic features according to claim 1, characterized in that, The spatiotemporal encoder includes: A 3-layer graph convolutional network, with 16-32-64 channels per layer; A 4-layer Transformer encoder with 512 hidden layers and 8 encoder heads; The position coding uses a sine-cosine function (5.16-5.17).

4. The method for identifying blocked pipe sections in a drainage pipe network system based on dynamic features according to claim 1, characterized in that, The sliding window parameters are set as follows: Window length ω∈[6,24] hours; Step size S = ω / 2; Input data dimensions Where N ≥ 10 nodes.

5. The method for identifying blocked pipe sections in a drainage pipe network system based on dynamic features according to claim 1, characterized in that, The multi-source data fusion includes: Rainfall data is linearly mapped to node feature dimensions; Feature-level weighted summation and fusion: Spatiotemporal joint attention mechanism.

6. The method for identifying blocked pipe sections in a drainage pipe network system based on dynamic features according to claim 1, characterized in that, The decoder includes: A 3-layer deconvolutional graph network restores the spatial dimension; A 4-layer Transformer decoder restores the timing dimension; The reconstruction loss function uses mean squared error.

7. The method for identifying blocked pipe sections in a drainage pipe network system based on dynamic features according to claim 1, characterized in that, The anomaly determination includes: Three consecutive time windows of abnormality are identified as suspected congestion; Pipeline segment-level determination requires simultaneous anomalies at both upstream and downstream nodes; Five consecutive abnormalities must be detected to trigger an alert.

8. The method for identifying blocked pipe sections in a drainage pipe network system based on dynamic features according to claim 1, characterized in that, The model training uses: Unsupervised pre-training using normal operating condition data; Adam optimizer, learning rate 1e-4; Batch normalization and Dropout regularization.

9. The method for identifying blocked pipe sections in a drainage pipe network system based on dynamic features according to claim 1, characterized in that, The system includes: The module includes a sliding window generation module, a spatiotemporal encoding and decoding module, an anomaly detection module, and an early warning output module.

10. The method for identifying blocked pipe sections in a drainage pipe network system based on dynamic features according to claim 1, characterized in that, The system also includes: Real-time data acquisition module and dynamic graph construction module.

Images