Pipe network prediction and intelligent decision system based on physical spectrum sensing and dynamic scanning

Abstract

The application discloses a pipe network prediction and intelligent decision system based on physical spectrum sensing and dynamic scanning, which deeply mines the correlation characteristics among multi-source data through dynamic physical spectrum decomposition and frequency domain adaptive topology evolution technology, effectively overcomes the problems of sensor noise interference and data deviation, simultaneously introduces a HydroScan-SSM model, effectively captures the long-term dependence relationship in the time sequence, accurately captures the coupling evolution law of the water quality parameters and the pipe network structure, finally, in view of the pipe network intercepting valve control scene, combines the pipe network data characteristics and intelligent decision demand, uses an intelligent decision framework based on dynamic data driving, realizes the dynamic optimization of the intercepting valve control strategy through real-time data fusion and machine learning algorithm, and effectively avoids the problems of misdischarge, leakage and untimely discharge.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of urban pipe network management and artificial intelligence, and more particularly to a pipe network prediction and intelligent decision-making system based on physical spectrum sensing and dynamic scanning. BACKGROUND

[0002] In recent years, with the acceleration of urbanization, urban pipe network systems have become increasingly complex, and the monitoring and prediction of their operating states are crucial for the stable operation of urban infrastructure. Underground pipe network systems have become increasingly important in urban infrastructure. By enhancing pipe network monitoring and prediction technologies, early warning of waterlogging risks, rapid positioning and repair of pipe network problems can be achieved, thereby effectively preventing urban waterlogging, ensuring the stable operation of drainage systems, improving the quality of life of residents, and providing strong support for urban renewal and resilience construction.

[0003] Currently, the monitoring and prediction process of pipe networks on the market mainly includes the following steps: real-time collection of pipe network operating data (such as liquid level, flow rate, water quality, etc.) through sensors, data cleaning and standardization processing, establishment of prediction models using statistical analysis, machine learning or deep learning methods, training and optimization of models combined with historical data, and finally real-time monitoring and prediction of future operating states, providing decision support for pipe network maintenance and management. However, existing pipe network prediction systems still have many shortcomings and cannot meet the needs of modern urban management. The following are the main problems in existing technologies:

[0004] ①Sensor data is easily affected by faults, environmental interference (such as water accumulation or silt blockage), or transmission errors, resulting in frequent occurrence of abnormal values and missing values. Traditional processing methods do not fully consider the physical characteristics of pipe network multivariate and the topology of pipe network, and cannot accurately distinguish between real anomalies and noise. The supplementation of missing values also often fails to effectively utilize the spatial correlation of adjacent node data under the constraint of pipe network physical properties, resulting in insufficient data reliability and affecting the accuracy of subsequent analysis.

[0005] ②Current monitoring devices mainly focus on real-time data collection and simple analysis, lacking the ability to predict future operating states or having low prediction accuracy. This lag makes it difficult for management personnel to identify potential problems in advance and take preventive measures. For example, when a pipe leaks or becomes blocked, existing devices can only issue an alarm after the problem has occurred, rather than predicting and warning in advance, leading to further deterioration of the problem.

[0006] ③The existing technology has obvious limitations in handling the spatial correlation of pipe networks. Model training often relies on single monitoring point data, and fails to effectively integrate and utilize data information from multiple adjacent or related monitoring nodes. This makes it difficult for the model to learn and capture the overall dynamic behavior of the pipe network system and the mutual influence between nodes, resulting in poor generalization performance of the trained model and difficulty in adapting to the operation rules of different regions (such as commercial and industrial areas) or different pipe network states.

[0007] ④The existing method has inherent bottlenecks in processing multi-scale time series data of pipe networks. Noise signals and signals representing key events (such as low-frequency leakage) are easily mixed in the frequency spectrum, and traditional filtering methods are difficult to effectively separate, often resulting in loss or suppression of key features. There is a conflict between modeling long-term dependencies (such as seasonal and periodic trends) and capturing short-term dynamic details (such as instantaneous pressure mutations), and global normalization methods often compress or blur important peak-valley fluctuation features. Redundant or outdated information in historical data lacks effective dynamic screening mechanisms, and its interference can easily accumulate and eventually affect the accuracy of the prediction model. SUMMARY

[0008] Therefore, the present application provides a pipe network prediction and intelligent decision-making system based on physical spectrum perception and dynamic scanning, which can realize accurate prediction and intelligent decision-making of urban pipe networks through the cooperative mechanism of PhySync physical-spectrum dual-drive perception model and HydroScan-SSM model (Hydro-Scanning State Space Model), help to identify risks such as waterlogging, overflow, water hammer, and backflow in advance, facilitate to provide real-time decision support and resource optimization scheduling capabilities for city managers, prevent problems from occurring, shorten response and maintenance time, and reduce operation and maintenance costs.

[0009] To achieve the above purpose, the present application adopts the following technical solutions:

[0010] The pipe network prediction and intelligent decision-making system based on physical spectrum perception and dynamic scanning provided by the embodiment of the present application comprises a data processing module, a PhySync physical-spectrum dual-drive perception model, a HydroScan-SSM model, a data assimilation module, and an intelligent decision-making module. Wherein:

[0011] The data processing module performs outlier identification processing on the multivariate time series data obtained by the sensor, and supplements missing values and corrects outliers.

[0012] The PhySync physical-spectrum dual-drive perception model extracts time characteristics from the processed data through a dynamic physical spectrum resolver, constructs a channel mask matrix using a frequency domain adaptive topology evolution device, and fuses the time characteristics and the channel mask matrix using a dynamic mask focusing fusion device.

[0013] The HydroScan-SSM model performs linear transformation on the fused characteristics, captures long-term dependencies in time series through structured state space calculation using a state transition matrix and an input coupling matrix, and adaptively determines the state update direction according to the real-time flow direction of the pipe network to perform forward scanning or reverse scanning, and predicts risks or abnormal conditions in the pipe network.

[0014] The data assimilation module dynamically adjusts and optimizes the model parameters by comparing the model prediction results with the actual observation data.

[0015] The intelligent decision-making module generates decision execution instructions according to the prediction results.

[0016] In one specific embodiment, in the data processing module, the dynamic pulsatile threshold method is used to identify and process abnormal values in the multivariate time series data obtained by the sensor, and the missing values are supplemented and the abnormal values are corrected by combining field strength weighted interpolation.

[0017] In one specific embodiment, in the data processing module, after supplementing the missing values and correcting the abnormal values, the time series data is subjected to data standardization processing, and the multi-scale periodic mode of the pipe network operation is identified and divided by combining the segmented normalization of the pipe network operation cycle characteristics.

[0018] In one specific embodiment, in the PhySync physical-spectrum dual-drive perception model, the dynamic physical spectrum resolver extracts time characteristics from the processed data, including a dynamic mode initialization stage, a time series mode decomposition stage, a feature recombination and output stage, wherein:

[0019] In the dynamic mode initialization stage, sliding window feature extraction and triple parallel convolution kernels are used to synchronize the scanning of input signals through convolution kernels of different scales, respectively extracting short, medium and long time feature fragments, and then performing multi-scale feature splicing and fusion.

[0020] In the time series mode decomposition stage, a learnable wavelet transform layer is used for multi-scale frequency domain decomposition, and a three-level adaptive filter bank is used to dynamically generate a three-level wavelet basis function, each level corresponding to a typical frequency band feature of the pipe network. Then, a deep separable convolution is used in the time dimension for deep convolution. Finally, a channel attention optimization feature selection is introduced to strengthen the key frequency bands and suppress the noise dominant frequency bands according to the spectral characteristics of the input signal.

[0021] In the feature reorganization and output stage, the frequency domain features are integrated with the original time sequence features, and a standardized space-time feature matrix is output.

[0022] In a specific embodiment, in the PhySync physical-spectrum dual-driven perception model, the frequency domain adaptive topology evolutioner constructs a channel mask matrix through fast Fourier transform, evolvable relationship learning and reparameterization.

[0023] In a specific embodiment, in the PhySync physical-spectrum dual-driven perception model, the dynamic mask focus fusioner fuses the time features extracted by the dynamic physical spectrum decomposer and the channel mask matrix generated by the frequency domain adaptive topology evolutioner based on a mask attention mechanism, to realize collaborative expression of time-space two-dimensional information.

[0024] In a specific embodiment, in the HydroScan-SSM model, a transient event detection module is also used to capture and respond to transient events in the pipe network, adopt a three-level response mechanism, predict transient events, and perform fault tracing and wave source positioning.

[0025] In the HydroScan-SSM model, in combination with transient event detection, the prediction process is divided into two modes: regular prediction and transient response prediction.

[0026] In a specific embodiment, the data assimilation module uses a hybrid model to mix observation data and model prediction results, and captures the internal structure and pattern of the data by decomposing and recombining the data.

[0027] In a specific embodiment, after the decision is executed, the system collects execution result data, compares the execution result data with the decision target, evaluates the execution effect of the decision, and uses the collected feedback data to update and optimize the PhySync physical-spectrum dual-driven perception model, the HydroScan-SSM model, and the intelligent decision module.

[0028] According to the above technical solution, the present application provides a pipe network prediction and intelligent decision system based on physical spectrum perception and dynamic scanning, which has the following technical advantages compared with the prior art:

[0029] 1. The system can realize accurate prediction and intelligent decision of urban pipe networks through the cooperative mechanism of the PhySync physical-spectrum dual-driven perception model and the HydroScan-SSM model, which helps to identify risks such as waterlogging, overflow, water hammer, and backflow in advance, and provides real-time decision support and resource optimization scheduling capabilities for urban managers.

[0030] 2.The application overcomes the problems of sensor noise interference and data bias by deeply mining the correlation characteristics between multi-source data through dynamic physical spectrum decomposition and frequency domain adaptive topology evolution technology; simultaneously, the HydroScan-SSM model is introduced to effectively capture the long-term dependence relationship in time series and accurately capture the coupling evolution law of water quality parameters and pipe network structure. Finally, for the pipe network intercepting valve control scene, combined with the pipe network data characteristics and intelligent decision-making requirements, the dynamic data driven intelligent decision-making module (DDF) is used to realize the dynamic optimization of intercepting valve control strategy through real-time data fusion and machine learning algorithm, effectively avoiding the problems of mis-discharge, leakage and delayed discharge.

[0031] 3.The application establishes a feedback and optimization mechanism. Through the feedback and optimization mechanism, the application can form a virtuous cycle of continuous improvement, continuously improve the prediction accuracy of the model and the intelligent level of decision-making, and make the model better adapt to the changing environment and needs.

[0032] The technical solutions of the application will be further described in detail below through the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0033] The accompanying drawings are used to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application.

[0034] Figure 1 The system architecture and workflow schematic diagram provided for the embodiments of the application.

[0035] Figure 2 The system working principle schematic diagram provided for the embodiments of the application.

[0036] Figure 3 The working process schematic diagram of the PhySync physical-spectrum dual-drive perception model provided for the embodiments of the application.

[0037] Figure 4 The working process schematic diagram of the HydroScan-SSM model provided for the embodiments of the application.

[0038] Figure 5 The working process schematic diagram of the DAM data assimilation module provided for the embodiments of the application.

[0039] Figure 6 The working process schematic diagram of the DDF intelligent decision-making module provided for the embodiments of the application. DETAILED DESCRIPTION

[0040] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations.

[0041] Referring to Figure 1 and Figure 2 The present application aims to provide a pipe network prediction and intelligent decision-making system based on physical spectrum sensing and dynamic scanning. Through the collaborative mechanism of the PhySync physical-spectrum dual-drive sensor (PhySync physical-spectrum dual-drive sensing model) and the HydroScan-SSM model, accurate prediction and intelligent decision-making of urban pipe networks are realized, and risks such as waterlogging, overflow, water hammer, and backflow are identified in advance, providing real-time decision support and resource optimization scheduling capabilities for city managers. The system can assist managers in dynamically adjusting operation parameters such as pump station start-stop, valve opening, pressure relief strategy, etc. based on prediction results to prevent problems from occurring. At the same time, the system has real-time monitoring and fault tracing capabilities, and can locate the impact source within seconds after a transient event (such as water hammer) occurs, significantly reducing response and repair time and reducing operation and maintenance costs.

[0042] Traditional prediction methods often struggle to cope with the complexity of multivariate time series data in urban pipe network systems, especially in terms of time dimension heterogeneity patterns (such as morning and evening peak, weekend differences) and spatial dimension complex interactions (such as pipe segment topology, hydraulic coupling). The present application realizes multi-scale period division (four segments per day / two segments per week) through a dynamic physical spectrum decomposer (Dyna Spectrum module), and extracts short-term transient and long-term evolution characteristics by combining dynamic spectrum decomposition technology (DSDT) to accurately model time heterogeneity. Through a frequency domain adaptive topology evolution module (Evo Topo module), a dynamic channel correlation matrix is constructed based on frequency domain characteristics to realize adaptive modeling of nonlinear relationships between variables, mask noise channels, and strengthen key variable collaborative modeling. Through a dynamic mask focus fusion module (MaskFuse module), time characteristics extracted by the DynaSpectrum module and channel mask matrix generated by the Evo Topo module are fused based on the mask attention mechanism to realize collaborative expression of time-spatial two-dimensional information, and finally generate prediction results.

[0043] Traditional pipe network monitoring systems often face problems such as sensor noise interference, transient event response lag, and difficulty in balancing long-term dependence and short-term fluctuations. The Hydro-SSM core (Hydro-Physical State-Space Module) in the HydroScan-SSM model of the system contains the following specific structures: ① State transition matrix X, which embeds the negative feedback of downstream to upstream, self attenuation and one-way propagation of hydraulic characteristics, used to describe the dynamic change relationship of state variables (such as pressure, flow) over time; ② Input coupling matrix Y, which represents the coupling strength between input signals and state variables, reflecting how input signals (such as monitoring data) affect system state; ③ Discretization time step ΔT, which represents the time resolution of the model, determines the sampling frequency of the model in time, ensures that the model can capture the dynamic changes of fluid propagation and maintain real-time. Hydro-SSM receives the fused feature vector from MaskFuse, constructs the state equation combined with the physical law, and dynamically adjusts through the discrete time step to update the system state in real time. Hydro-SSM can directly embed pipe diameter, material, flow rate and other hydraulic properties into the state transition matrix and input coupling matrix, so that the model has pipe network level fluid dynamics constraints, which can not only suppress the negative feedback of downstream pressure to upstream, but also retain the long-term memory of seasonal flow attenuation, thus forming a natural barrier to transient noise caused by electromagnetic interference or equipment failure; then according to the reliability of eight real-time indexes such as liquid level, temperature, flow rate and ammonia nitrogen, the flow direction is dynamically determined to complete forward or reverse scanning switching, realizing zero delay response of transient events; when the pressure acceleration is abnormal, the three-level early warning (level I record, level II pressure reduction, level III pump stop) is triggered, and the gradient positioning impact source is used to start the transient mode synchronously, output the shock wave propagation path and valve and pump station control effect prediction, which can publish risk warning in advance for several hours to several days.

[0044] In order to further verify and adjust the accuracy of the model, the system introduces a data assimilation module (DAM, Data Assimilation Module). This module compares the model prediction results with the actual observation data, and uses assimilation algorithm to dynamically adjust and optimize the model parameters.

[0045] Further, the system uses an intelligent decision module (DDF decision engine, Dynamic Data Fusion) to directly convert the prediction results into executable instructions such as valve opening degree, pump station start-stop, etc., forming a self-evolution closed loop of "perception-modeling-prediction-decision-execution-feedback", which significantly reduces the impact of failure on city operation and improves the resilience and intelligent level of infrastructure.

[0046] The main functions of the system are as follows:

[0047] ① Deploy multiple sensors at key nodes of the urban pipe network to collect data such as liquid level, flow rate, and water quality (e.g., ammonia nitrogen, conductivity) in real time. The sensor data is transmitted to the monitoring system via wireless or wired networks. Integrate algorithms into the equipment and use the feature extraction module of the algorithm to perform real-time cleaning and noise reduction on the sensor data. For outlier handling, combine dynamic pulsation thresholding method with field strength weighted interpolation technology, and use data from adjacent nodes to supplement missing values ​​to ensure data quality.

[0048] ② In the PhySync physical-spectrum dual-drive perception model, temporal feature extraction is performed using a dynamic physical spectrum decomposer (DynaSpectrum module). First, a sliding window feature extraction and triple parallel convolution kernel design are employed to extract feature segments for short time (3 points / 15 minutes), medium time (5 points / 25 minutes), and long time (7 points / 35 minutes). Simultaneously, the physical parameter vector of the current pipe segment is obtained from the pipeline network digital twin system, including key attributes such as pipe diameter, material elastic modulus, and service life. After normalization, this vector is fused with the features output from the triple convolution kernel across modalities to generate joint features.

[0049] ③ In the PhySync physical-spectrum dual-drive sensing model, a frequency-domain adaptive topology evolver (Evo Topo) is used to construct the spatial correlation matrix of the pipeline network. First, a fast Fourier transform (rFFT) is performed on the time series data of each channel, projecting it onto the frequency space to obtain the frequency representation of each channel. Then, an initial relation matrix is ​​constructed based on the pipeline network GIS topology. Through evolutionary relation learning and reparameterization steps, the nonlinear correlation between channels is dynamically captured, adapting to changes in pipeline network operating conditions, and generating a binary channel mask matrix.

[0050] ④ In the feature fusion stage, a dynamic mask focusing fusion unit (MaskFuse) is used to fuse temporal features and channel mask matrices. Through the mask attention mechanism, it adaptively focuses on channels and time points that are more important to the prediction task, while shielding the influence of irrelevant channels, thereby achieving effective feature fusion and forming a more expressive feature representation.

[0051] ⑤ In the HydroScan-SSM model, the fused features are first subjected to a linear transformation to adapt to the processing requirements of subsequent modules. Then, through structured state-space computation, the long-term dependencies in the time series are captured using the state transition matrix and input coupling matrix to simulate the temporal evolution of the pipeline network's operating state. The state update direction is adaptively determined based on the real-time flow direction of the pipeline network, performing either forward or reverse scanning.

[0052] ⑥ A transient event detection module is set up in the HydroScan-SSM model, specifically for capturing and responding to transient events in the pipeline network, especially water hammer. By calculating the acceleration of pressure changes and using the second-order time derivative to capture the acceleration changes of pressure, a three-level response mechanism is adopted to provide early warning of transient events and to perform fault tracing and wave source localization.

[0053] ⑦ Use the test dataset to make predictions on the model, monitor the operation status of the pipeline network based on the prediction results set according to the scheduling requirements at different time scales, issue early warnings for abnormal events in the pipeline network, generate corresponding emergency plans and dynamic scheduling optimizations, and improve the resilience of the urban pipeline network.

[0054] The following is combined Figures 1-6 As shown, the specific implementation methods and working principles of the system of the present invention will be described in detail below:

[0055] (1) Data processing:

[0056] The data in this invention originates from multivariate time-series data of liquid level, temperature, flow rate, ammonia nitrogen value, current value, conductivity, resistivity, and salinity detected by urban pipeline monitoring equipment systems, with time intervals of 5 minutes. Data from any one month at a specific monitoring point is taken as the sample data for this invention. During the data collection process, outliers or missing values ​​may occur due to sensor malfunctions, environmental interference (water accumulation or silt blockage), or data transmission errors.

[0057] ①Outliers are assessed using a dynamic pulsating threshold method:

[0058] First, calculate the mean (μ) and standard deviation (σ) of the sample data, and then set the dynamic field strength threshold based on the pipeline physical parameters:

[0059] [μ-η·σ·ι,μ+η·σ·ι]

[0060] Where ι is the pipe material attenuation coefficient (e.g., 1.2 for cast iron, 0.8 for PVC, and 1.0 for steel pipe), and η represents the hydraulic sensitivity factor. (v is the flow velocity, and D is the diameter of the pipe's inner wall).

[0061] Newly collected data points are monitored in real time to determine if they exceed threshold ranges. If they do, they are identified as anomalies, triggering a pipeline health status alert.

[0062] ② Field strength weighted interpolation to supplement missing values ​​and correct outliers:

[0063] First, data from adjacent monitoring nodes at the same time point are collected. Different weights are assigned to the data from adjacent nodes based on the physical properties of the pipeline network. The formula for calculating the field strength attenuation coefficient is as follows:

[0064]

[0065] Among them, l ij Ξ represents the network topology distance between nodes i and j. ij λ represents the hydraulic resistance coefficient of the pipe section (determined by the roughness of the pipe material), and λ represents the medium attenuation factor (determined according to the water quality type).

[0066] The obtained field strength attenuation coefficient is used to calculate the weight value, as shown in the following formula:

[0067]

[0068] Finally, the weighted average is calculated as the correction value, using the following formula:

[0069]

[0070] in, Indicates the corrected outlier or missing value, x j ω represents the value of adjacent monitoring node j, N represents the number of nodes, and ω j This indicates the corresponding weight.

[0071] To ensure that the model is not affected by scale differences in the input data during training, this paper adopts data standardization to unify the distribution of training and testing data, thereby accelerating the convergence speed of the model and improving prediction performance.

[0072] In the previous steps, outliers and missing values ​​have been corrected, and the data distribution is relatively concentrated. Therefore, this invention uses the periodic adaptive pulsation normalization method to standardize the input time series data, and performs segmented normalization in combination with the characteristics of pipeline operation cycle, deeply coupling the inherent periodic characteristics of pipeline operation.

[0073] First, the multi-scale cyclical patterns of pipeline network operation are identified and classified: In the daily cycle dimension, four characteristic time periods are clearly defined: morning peak (6:00-10:00), midday stable period (10:00-17:00), evening peak (17:00-21:00), and nighttime trough (21:00-6:00); in the weekly cycle dimension, two operation modes are distinguished: weekday (Monday to Friday) and weekend (Saturday and Sunday). Based on this spatiotemporal framework, dynamic statistical parameters are independently calculated for each specific cycle combination (e.g., "weekday evening peak"), and the cycle mean and cycle standard deviation are calculated.

[0074]

[0075] Where, N current T represents the number of data points within the current period. t R represents the characteristic period of the daily cycle. tIndicates a week-cycle date type, T t ∩R t This represents a spatiotemporal periodic intersection unit, where t represents the data timestamp, and x represents the data time stamp. t This represents the original monitoring value at timestamp t.

[0076] During normalization, the spatiotemporal periodic unit to which the data point belongs is automatically matched based on the timestamp, and the following formula is applied for conversion:

[0077]

[0078] By using independent parameters for different time periods, the true fluctuation characteristics of the pipeline network are preserved, avoiding the compression of peak and valley features caused by global normalization.

[0079] (2) PhySync physical-spectrum dual-drive sensing feature extraction method:

[0080] See Figure 3 As shown, the normalized data is processed by a dynamic physical spectrum decomposer (Dyna Spectrum) to extract time-dimensional features, and a frequency domain adaptive topology evolver (Evo Topo) is used to construct the network spatial correlation matrix. The two modules are processed in parallel to form a physical-spectrum dual-drive perception architecture.

[0081] ① The Dynamic Physical Spectrum Decomposer (Dyna Spectrum) employs Dynamic Spectrum Decomposition Technology (DSDT), which mainly includes three core steps to achieve high-precision decoupling of time-series features.

[0082] A. Dynamic mode initialization phase:

[0083] Based on the multi-scale characteristics and real-time processing requirements of pipeline network monitoring data, a sliding window feature extraction and triple parallel convolution kernel design are adopted in the dynamic mode initialization stage. By synchronously scanning the input signal through convolution kernels of different scales, feature segments of short time (3 points / 15 minutes), medium time (5 points / 25 minutes) and long time (7 points / 35 minutes) are extracted respectively. At the same time, rapid transient events (such as pump valve switching) and long-term evolution patterns (such as changes in daily water consumption) in the pipeline network are perceived.

[0084] First, the normalized data is input into a processing layer consisting of three parallel convolutional branches, each containing a trainable weight matrix W. [a] and bias term b [a] The ReLU activation function is used to ensure feature nonlinearity.

[0085]

[0086] in Convolution kernel weight W [a] Independent training, bias term b [a]It is also trainable.

[0087] Based on the multi-scale features of the parallel convolutional branches, constraint injection is achieved through a learnable physical parameter projection matrix. For each time point t, the physical parameter vector P of the current pipe segment is first obtained from the pipeline network digital twin system. t This includes key attributes such as pipe diameter (in mm), material elastic modulus (MPa), and service life (in years). These parameters, after normalization, are then fused across modalities with the features output from the triple convolution. The specific formula is as follows:

[0088] Pipe diameter properties are logarithmically scaled and normalized.

[0089] Where D represents the diameter distance of the inner wall of the pipe, D min D represents the smallest branch. max This indicates the upper limit of the main pipeline.

[0090] The elastic modulus is piecewise linearly normalized to differentiate material properties:

[0091]

[0092] The service life is normalized using exponential decay to reflect the nonlinearity of material aging.

[0093]

[0094] Where y represents the service life of the pipeline.

[0095] Concatenate the normalized parameters into a vector:

[0096] Among them, M one-hot This represents a one-hot encoding matrix, used to represent the category information of different nodes or areas in a pipeline network system.

[0097] Cross-modal fusion of features from triple convolution output:

[0098]

[0099] in, Ω represents the output of the triple convolution branch; t Represents the normalized vector of the physical properties of the pipeline network; The vector concatenation method generates joint features; the projection matrix U compresses the dimension of the concatenated features to the target dimension; the tanh activation function ensures that the values ​​of the fused features are in the range [-1, 1].

[0100] B. Timing Pattern Decomposition Stage:

[0101] In pipeline monitoring scenarios, the construction of the frequency domain feature extraction branch is based on the physical characteristics of the pipeline system and the essential patterns of the monitoring data. Multi-scale frequency domain decomposition is performed through a learnable wavelet transform layer. A three-level adaptive filter bank is designed for the spectral characteristics of pipeline monitoring indicators (liquid level, flow velocity, temperature, etc.). A parameterized neural network dynamically generates three levels of wavelet basis functions, each level corresponding to typical frequency band characteristics of the pipeline network.

[0102]

[0103] in, and These are trainable parameters used to generate the coefficients of the k-th level small filter, adapting to the signal characteristics of different frequency bands; φ t The output is for the dynamic mode initialization phase; the ELU activation function ensures that the generated filter coefficients are smooth and continuous and allow negative values, which conforms to the mathematical characteristics of wavelet functions.

[0104] Then, depthwise separable convolution is used to achieve efficient wavelet transform. This structure first performs depthwise convolution in the time dimension (each input channel is convolved independently), and then incorporates channel information through 1×1 point convolution, making it suitable for parallel processing of multiple indicators in pipeline networks compared to standard convolution. The convolution stride is fixed at 1 to maintain temporal resolution, and symmetrical padding is used at the boundaries to avoid feature truncation.

[0105]

[0106] in, Denotes the frequency domain features extracted by the k-th level wavelet basis function; DepthwiseConv1D represents the depthwise separable convolution operation; A t-Δt:t This represents the multi-sensor time-series data matrix within the sliding time window; Δt represents the length of the time window, corresponding to the sampling time span; ψ (k) f represents the k-th order wavelet basis function; k The filter length is indicated, which determines the time scale for feature extraction.

[0107] Finally, channel attention is introduced to optimize feature selection, which strengthens key frequency bands based on the spectral characteristics of the input signal (e.g., automatically increasing the weight of low-frequency components when leakage occurs), suppresses noise-dominated frequency bands, and significantly improves feature discriminative power.

[0108]

[0109] Where, ω k represents the attention weight of the k-th frequency band; sig represents the sigmoid activation function, which restricts the weights to the range [0,1] to ensure balanced fusion of multiple frequency bands; MLP represents the multilayer perceptron used to learn the nonlinear relationship between frequency bands and output the attention weights of each frequency band; Represents the characteristics of each frequency band Perform global average pooling to extract global feature information; F t freq This represents the frequency domain feature vector after fusion at time point t.

[0110] C. Feature Recombination and Output Stage:

[0111] This stage completes the final feature representation construction of Dyna Spectrum, forming a standardized spatiotemporal feature matrix through a systematic feature integration process. First, the frequency domain features extracted by dynamic spectral decomposition are fused with the original time-series features:

[0112]

[0113] in, This represents the frequency domain characteristics of the output of B. time-series pattern decomposition; This represents the temporal characteristics of the initial output of A. dynamic mode; d is the feature dimension.

[0114] Then, the feature vectors from all time steps are stacked along the time dimension to form a structured feature matrix:

[0115]

[0116] Wherein, the row dimension r represents the length of the time series; the first d columns of the column dimension are the frequency domain feature subspace, and the last d columns are the time domain feature subspace.

[0117] Independently standardize the two feature subspaces to eliminate dimensional differences:

[0118]

[0119] Where, μ 3d / σ 3d This represents a statistic based on a 3-day sliding window, adapting to short-term fluctuations; μ 7d / σ 7d This is a statistical measure expressed over a 7-day sliding window, capturing long-term trends.

[0120] Construct the final output matrix:

[0121]

[0122] ② The frequency domain adaptive topology evolver (Evo Topo) mainly includes the Fast Fourier Transform (rFFT), evolutionary relation learning, and reparameterization steps.

[0123] First, a Fast Fourier Transform (rFFT) is performed on the time-series data of each channel, projecting it onto the frequency space to obtain the frequency representation of each channel. The Fourier Transform transforms the time-series data (such as liquid level, temperature, and flow rate) detected by the urban pipeline monitoring system from the time domain to the frequency domain, eliminating time-domain fluctuation interference, highlighting the periodic characteristics of pipeline operation, and thus making it easier to capture the frequency characteristics between different channels. The formula is as follows:

[0124] F i =norm(rFFT(c i ))

[0125] Among them, c i This represents the data at all time points of the i-th channel. norm represents the normalization process applied to the Fourier transform result to eliminate dimensional differences between different channels or different frequency components, making them comparable.

[0126] In the Fast Fourier Transform, only the first m dominant frequency components are retained, compressing the original sequence length to m. In the evolutionary relation learning step, the frequency domain feature vector used is the reduced-dimensional one (length m). First, an initial relation matrix M0 is constructed based on the pipeline GIS topology:

[0127]

[0128] in, Representing spatial topology, σ space The standard deviation of spatial correlation is represented by O; O represents the distance matrix between network nodes; GG T The matrix represents hydraulic coupling; G represents the design flow matrix of the pipe section; T represents the transpose of the matrix; α and β represent the regional adaptive coefficients (commercial area: 0.6 / 0.3; industrial area: 0.4 / 0.5); I represents the identity matrix.

[0129] Constructing a time-varying positive semi-definite matrix M t Dynamically capture the evolution of pipeline network channel associations:

[0130] M t =M t-1 +ΔM t

[0131]

[0132] Among them, M t-1 ΔM represents the positive semi-definite matrix of the previous time step; t Indicates the incremental learning term; κ represents the learnable evolution rate (default 0.05); This indicates the pipeline network operating condition memory window (default 6 hours); This represents the feature sequence within the time window, used to calculate incremental updates.

[0133] Using the evolutionary matrix M t Measuring nonlinear relationships between channels, dynamically capturing nonlinear correlations between channels, and adapting to changes in pipeline network operating conditions:

[0134]

[0135] in, F represents the dynamic distance between i and j at time t; i ,F j This represents the channel frequency vector.

[0136] Adaptive topology construction is performed, the distance matrix is ​​converted into a probability distribution, and a pipeline channel correlation matrix is ​​constructed. This enables dynamic binarization without the need for manual threshold setting, and automatically filters strongly correlated channels.

[0137]

[0138] in, This represents the probability of association between channels i and j at time t; This represents the weight and distance between node i and node j. The smaller the value, the larger the exponent, indicating that node j has a greater influence on node i; This indicates the learnable temperature coefficient, which controls the degree of correlation between the pipeline network and the temperature.

[0139] Finally, reparameterization is performed by converting the probability matrix into a binary channel mask matrix (containing only 0s and 1s) through Bernoulli resampling, while preserving gradient propagation. The formula is as follows:

[0140] B ij ≈Bernoulli(C ij )

[0141] Higher probability C ij This will lead to B ij A value closer to 1 indicates a relationship between channels i and j.

[0142] (3) Feature fusion using the dynamic mask focusing fusion tool (MaskFuse):

[0143] See Figure 3 As shown, the system first receives temporal features extracted from the Dyna Spectrum module. These features capture the changing patterns and periodicities of urban pipeline monitoring data (such as liquid level, temperature, and flow rate) over time. It also receives a channel mask matrix generated by the EvoTopo module. This matrix indicates the correlation between different channels (monitoring indicators), providing channel-dimensional information for feature fusion.

[0144] To perform the attention mechanism calculation, three projection matrices U are first used. Q U K U V For time feature F output Perform a linear transformation. These three projection matrices correspond to the generation of the query, key, and value matrices, respectively.

[0145] The masked attention score is calculated by multiplying the dot product between the query matrix Q and the transpose of the key matrix K, and then scaling the result using a scaling factor to obtain the attention score matrix. This matrix is ​​then multiplied by the channel mask matrix B to perform a masking operation. The goal is to set the attention scores of irrelevant channel pairs to negative infinity, thus making the attention weights of these pairs close to zero in the subsequent Softmax operation, effectively masking the influence of irrelevant channels. The formula is as follows:

[0146]

[0147] in, ⊙ represents the scaled attention score. By scaling the attention score, the magnitude of the gradient can be controlled to avoid gradient vanishing or exploding. ⊙ represents element-wise multiplication, which is used to combine the attention score matrix with the channel mask matrix to achieve masking operations. 1-B represents the attention score after masking, which restricts the attention score to related channel pairs, focusing only on channels that are beneficial to the prediction task; 1-B represents the inverse matrix of the channel mask matrix, used to indicate which channels are not beneficial to the downstream prediction task; -∞ represents setting the attention score of unrelated channel pairs to negative infinity; (1-B)⊙(-∞) represents masking unrelated channel pairs, ensuring that attention is focused only on related channels, and the attention score of unrelated channel pairs is negative infinity, with the weights close to 0 after Softmax normalization.

[0148] The attention score matrix is ​​normalized using the Softmax function to obtain the attention weights. Then, the attention weights are multiplied by the value matrix to obtain the fused features. The formula is as follows:

[0149] F final =Softmax(MKscores)·(F output ·U V )

[0150] Among them, F final This represents the features after Softmax normalization. This process ensures that the sum of each row is 1, allowing the model to adaptively focus on channels and time points that are more important to the prediction task.

[0151] Finally, through feature fusion, the model can adaptively focus on channels and time points that are more important to the prediction task. In urban pipeline network monitoring, this can improve the accuracy of the model's prediction of the pipeline network's operating status. The formula is as follows:

[0152] F fused =F final ·(F output ·U V )

[0153] Among them, F fused Indicates the features after fusion;

[0154] The MaskFuse module adaptively focuses on channels and time points that are more important to the prediction task through an attention mechanism. Simultaneously, it uses a channel mask matrix to shield the influence of irrelevant channels, thereby achieving effective feature fusion. This fusion method captures key information in both time and channel dimensions, providing strong support for subsequent data assimilation and intelligent decision-making.

[0155] (4) HydroScan-SSM model:

[0156] ① The fused features undergo preliminary feature transformation through a linear transformation layer to meet the processing requirements of subsequent modules. The formula is as follows:

[0157] F in =U in ·F fused +b in

[0158] Among them, F in It is the feature representation after linear transformation, F fused It is a feature after fusion, U in Let b represent the projection weight matrix. in This represents the bias vector. In urban pipeline network monitoring, the fused features contain information from different monitoring points and time points. Through a linear transformation layer, this information is initially transformed and mapped to a unified feature space.

[0159] ②Structured state-space computation:

[0160] The transformed feature F in The computation is performed using a hydrophysically constrained SSM (Hydro-SSM). The Hydro-SSM module effectively captures long-term dependencies in time series data through its internal parameterized matrices and hardware-aware parallel computing algorithms.

[0161] F ssm =SSM(F′) in (X,Y,ΔT)

[0162] Among them, F' in These are the features processed by the one-dimensional convolutional module. X and Y are the parameterized matrices of the SSM module, ΔT represents the discretization time step, and F... ssm This is the characteristic output of the Hydro-SSM module. The Hydro-SSM module captures the dynamic characteristics of time series through a state-space model. In urban pipeline network monitoring, this helps the model understand the temporal evolution of the pipeline network's operating status (such as pressure propagation delay and periodic flow changes).

[0163] In the above formula, X is the state transition matrix, which describes the dynamic relationship between system states. In pipeline monitoring scenarios, it reflects the transmission and change patterns of state variables such as pressure and flow rate between time steps.

[0164]

[0165] Where n and h represent state indices, and represent different state dimensions.

[0166] When n > h This represents the negative feedback effect of state n on state h. In a pipeline network, this corresponds to the inhibitory effect of the downstream state on the upstream state, which helps stabilize fluctuations in system pressure or flow.

[0167] When n = h This represents the self-inhibition effect of the state. It reflects the natural decay of the state over time in the absence of external input, which helps the model capture the self-regulating ability of the pipeline system.

[0168] When n = h, X nh =0 indicates that state n has no direct effect on state h. This is consistent with the unidirectional propagation characteristic of fluid dynamics, meaning that the upstream state is not directly affected by the downstream state.

[0169] Y represents the input coupling matrix, indicating the coupling strength between the input signal and the state variable. In pipeline monitoring, the input signal can be monitoring data such as pressure and flow rate. Y reflects how these inputs affect the system state. The calculation formula is as follows:

[0170]

[0171] Among them, Y i L represents the coupling strength of the i-th input channel; i This represents the length of the i-th pipe segment. The input coupling strength is inversely proportional to the pipe segment length; shorter segments respond more strongly to the input signal and therefore have higher coupling strength. This design reflects the influence of the input signal on different pipe segments, ensuring that the model can accurately capture the dynamic changes of the input signal.

[0172] ΔT represents the discretization time step, indicating the model's temporal resolution and determining its sampling frequency over time. In pipeline monitoring, it ensures the model can capture the dynamic changes in fluid propagation while avoiding excessive computational burden caused by excessively high temporal resolution. The calculation formula is as follows:

[0173]

[0174] Where L represents the pipe length, v represents the fluid velocity, and the time step is the smaller of 10% of the ratio of pipe length to flow velocity and 5 seconds, to ensure that the model can capture the dynamic changes in fluid propagation while maintaining real-time performance.

[0175] ③ Dynamic scanning direction processing with multi-indicator fusion:

[0176] In the HydroScan-SSM model, a dynamic scanning direction processing mechanism replaces the traditional bidirectional processing, adaptively determining the state update direction based on the real-time flow direction of the pipeline network.

[0177] From H ssm Eight key pipeline monitoring variables were extracted from the state matrix:

[0178]

[0179] Where Lvl represents the liquid level, Tmp represents the water temperature, Spd represents the flow rate, Nit represents the ammonia nitrogen value, Crr represents the current value, Cnd represents the conductivity, Res represents the resistivity, Sal represents the salinity, and S represents the feature extraction matrix. A linear transformation S is performed from the state space H... ssm Eight physical quantities were decoupled from the middle.

[0180] 1. Construct a multivariate decision model that comprehensively considers the complex interactions of multiple monitoring variables, and determine the water flow direction by comprehensively considering multiple pipeline monitoring variables through a flow direction determination formula:

[0181]

[0182] Where, ω i This represents the dynamically calculated weights; g i (X i ) represents the characteristic function of the i-th monitored variable; θ represents a preset threshold used to control the sensitivity of the decision; FORWARD represents water flow from the source to the end, which is suitable for normal water supply conditions; BACKWARD represents water flow from the end to the source, which is suitable for detecting backflow or water hammer events.

[0183] The formula for calculating the dynamic weights mentioned above is:

[0184]

[0185] in, The confidence level of the i-th feature at time t is represented by the following formula:

[0186]

[0187] Where Δt represents the size of the time window, used to calculate the confidence level; μ represents the value of the i-th feature at time ζ; i and σ i The mean and standard deviation of the i-th feature are used for standardization.

[0188] Each feature function g i (X) is designed to capture dynamic changes in network monitoring variables that may indicate the direction of water flow. The characteristic function is defined as follows:

[0189] Liquid level Lvl: Indicates the rate of change of liquid level. Positive contribution increases;

[0190] Temperature Tmp: ΔTmp = |Tmp - 25℃| represents abnormal temperature changes. When ΔTmp > 2℃, the reverse contribution increases.

[0191] Flow rate Spd: sgn(Spd)·|Spd| 0.7 Combine information on the direction and intensity of the flow velocity;

[0192] Ammonia nitrogen value (Nit): Indicates the pollutant diffusion rate. The reverse contribution increases over time;

[0193] Current value Crr: Indicates cathodic protection status. The reverse contribution increases over time;

[0194] Conductivity Cnd: This indicates the change in ion concentration; the larger the value, the greater the reverse contribution.

[0195] Res: Res indicates abnormal water quality. norm When the value is greater than 0.8, the reverse contribution increases;

[0196] Salinity: Represents the salinity gradient. The reverse contribution increases over time.

[0197] 2. Steering drive state update:

[0198] Forward scan (water source → tip):

[0199] When the decision engine outputs FORWARD, the system executes state updates according to the pipeline topology. First, nodes are sorted in ascending order of distance from the water source: node 1 (water source), node 2, ..., node p (endpoint), ensuring that state updates start from the water source and propagate towards the endpoint. For each node o, the state update equation is:

[0200]

[0201] Where X is the state transition matrix and Y is the input coupling matrix, XY is calculated in the structured state space in step ②. This represents the input to node o at time t. The state is transmitted from the water source to the end point, simulating the propagation direction of pressure waves during normal water supply.

[0202] Reverse scan (terminus → water source):

[0203] When a BACKWARD is detected, the system performs a state update in reverse pipeline topology order. First, nodes are sorted in descending order of distance from the water source: node p (endpoint), node p-1, ..., node 1 (water source), ensuring that state updates start from the endpoint and propagate towards the water source. For each node o, the state update equation is:

[0204]

[0205] Among them, X T It is the transpose of the state matrix, encoding the characteristics of reverse wave propagation. The state is transmitted from the end to the water source, simulating the reverse propagation of backflow / water hammer waves.

[0206] Arrange the state features of all nodes in order and integrate them into a vector H. scan :

[0207]

[0208] ④ Transient event detection module:

[0209] The transient event detection module is a key component of the dynamic scanning direction processing flow. It is specifically designed to capture and respond to transient events in pipeline networks, especially water hammer. Water hammer is one of the most dangerous transient events in pipeline networks, capable of generating enormous pressure surges in a very short time, posing a serious threat to the safety of the pipeline system.

[0210] In transient event detection heads, the core formula mainly involves the pressure component, because water hammer events primarily manifest as instantaneous pressure changes. The formula is as follows:

[0211]

[0212] in, The acceleration due to pressure change is calculated using the five-point central difference method. The adaptive threshold is calculated using the following formula: Where σ Pr Pr represents the standard deviation of pressure. max This represents the maximum pressure value. The second time derivative is used to capture the acceleration changes of pressure, and the ReLU activation function is used to filter out background noise, focusing on sudden pressure events.

[0213] A three-level response mechanism is adopted: when g transient ∈(0,5), a Level I warning is issued, and the event is recorded; when g transient ∈[5,10], Level II early warning is adopted, and the pressure reducing valve opening is 50%; when g transient If the value is greater than 10, a Level III warning will be issued, and the pump will be stopped immediately with the pressure relief valve fully opened.

[0214] Perform fault tracing and wave source location: Among them, u i This indicates the location of the i-th monitoring point in the pipeline network. G represents the probability of a transient event. transient The location u of monitoring point i i The gradient of the transient event probability reflects the rate at which the transient event probability changes with location at monitoring point i. By calculating the gradient of the transient event probability with respect to each monitoring point, the location of the impact source can be determined.

[0215] ⑤ Prediction module:

[0216] Combining transient event detection, the prediction process is divided into two modes: conventional prediction and transient response prediction. The specific steps are as follows:

[0217] 1. Standard prediction mode (without transient events):

[0218] The state feature H output by dynamic scanning scan Input into the prediction engine structure:

[0219]

[0220] in, χ represents the output vector for the predicted future time step, containing the predicted values ​​of all monitored variables; χ represents the output matrix, containing the mapping coefficients from each state variable to the output; e XΔt H represents the matrix exponent of the state transition matrix X, used to describe the evolution of the state over time; scan Represents the state feature vector output by dynamic scanning; Θ represents the input compensation matrix, which maps the input features to the state space; T i This represents the input feature vector at the current time step.

[0221] 2. Transient response prediction mode:

[0222] Triggering conditions:

[0223] Shock wave pressure core equation:

[0224] in, The predicted shock wave pressure value is represented by Pr0; the initial pressure value is represented by ρ; and the liquid density is represented by v. pre The pressure wave velocity is expressed by the following formula: Among them, E vol The bulk modulus of the fluid is represented by ρ, the density by D, and the pipe diameter by E. mate Here, Π represents the elastic modulus of the pipe material, v0 represents the initial flow velocity, ι represents the pipe material attenuation coefficient, and Φ represents the fundamental frequency of water hammer. The calculation formula is as follows: Where L represents the pipe length; l represents the distance along the pipe network; and t represents time.

[0225] Shock wave propagation prediction: Where Ψ represents the pipeline topology transfer matrix, describing the propagation characteristics of pressure waves in the pipeline; This represents the shock wave pressure prediction vector.

[0226] Prediction of control effect: Where Λ represents the control matrix, which includes the valve adjustment coefficient Λ valve and pump station adjustment coefficient Λ pump ;Θ con This represents the control input vector, which includes valve opening and pump station status.

[0227] 3. Predicted output:

[0228] Transient event probability: This represents the probability of a transient event occurring. Based on the probability of a transient event, three risk level indicators are established: when... When it is low risk, At the time, it was considered a medium-risk period. It is a high-risk situation.

[0229] ⑥ Output layer:

[0230] The final output includes:

[0231] 1. Conventional prediction results: Predicted values ​​for liquid level, temperature, flow rate, ammonia nitrogen value, current value, conductivity, resistivity, and salinity at each monitoring point.

[0232] 2. Transient event prediction results (when a water hammer event is triggered): predicted pressure values ​​at each monitoring point; predicted impact values ​​of control measures on each monitoring variable.

[0233] 3. Probability and risk level of transient events: The probability of a transient event occurring. Risk level assessment based on probability.

[0234] (5) DAM data assimilation:

[0235] Data assimilation effectively combines observational data with model predictions. By fusing information from multiple sources, it corrects model biases, supplements missing data, and captures the inherent structure and patterns of the data. This improves the system's monitoring accuracy and prediction reliability of the urban pipeline network's operational status, ensuring that the model can more accurately reflect the actual pipeline network's operational status and providing stronger support for the management and maintenance of the urban pipeline network.

[0236] This invention employs a hybrid model, which combines observed data with model predictions to capture the data's inherent structure and patterns through data decomposition and reconstruction. The hybrid model can be represented as:

[0237] g assimilated =ξ·g observed +(1-ξ)·g predicted

[0238] Among them, g assimilated It is the assimilated data, g observed It is observation data, g predicted ξ is the model prediction data, and ξ is the mixing coefficient, which is used to control the weights of the observed data and the model prediction data in the assimilation process.

[0239] (6) Method for updating model parameters and evaluating the model using the Adam optimizer and mean squared error loss function:

[0240] The Adam optimizer combines the advantages of AdaGrad and RMSProp, adaptively adjusting the learning rate. By combining first-moment and second-moment estimations of the gradient to adaptively adjust the learning rate of each parameter, the model converges more quickly during training. Adam also demonstrates good adaptability to the massive datasets used in this invention. Furthermore, the Adam optimizer is relatively robust to hyperparameter selection, simplifying the model training process and reducing the workload of parameter tuning. Since urban pipeline network operation status monitoring and prediction systems typically require processing large amounts of monitoring data and constructing complex models to capture the inherent patterns and regularities of the data, the Adam optimizer effectively meets these needs.

[0241] Mean Squared Error (MSE) represents the average of the squared differences between predicted and true values. The Adam optimizer algorithm updates model parameters based on the calculated MSE loss to minimize the loss function. By continuously adjusting the model parameters, the model's predicted values ​​gradually approach the true values, thereby improving the model's predictive performance. The formula is as follows:

[0242]

[0243] Where N represents the sample size, i.e., the total number of data points used to evaluate model performance. In the scenario of urban pipeline network operation status monitoring and prediction, N can represent the number of monitoring data points collected within a specific time period (e.g., one month), where each data point may contain values ​​for multiple monitoring indicators such as liquid level, temperature, and flow rate; x i This represents the actual observed value, that is, the data value actually detected by the urban pipeline network monitoring equipment system. This represents the model's predicted value. During training, the MSE loss is calculated and monitored periodically to understand the model's performance on the training data. The smaller the MSE loss, the smaller the difference between the model's predicted value and the true value, and the better the model's predictive performance.

[0244] (7) DDF Intelligent Decision Making:

[0245] Based on the future operating status of the pipeline network and sensor information, a DDF (Digital Decision Building) module can be constructed to provide early warnings of potential anomalies and enable timely preventative measures.

[0246] The future sequence predictions are input into the DDF intelligent decision-making module. These predictions include forecast values ​​for key indicators (such as liquid level and flow rate) of the urban pipe network over a future period. The module first interprets the predictions, identifying potential risks or anomalies. For example, it might predict that the liquid level in a certain area's pipe network will continue to rise and exceed the warning line. Based on the actual situation of the urban pipe network, a series of decision rules are set. These rules can be threshold comparisons, pattern matching, or more complex logical judgments. For example:

[0247] If the predicted liquid level is higher than the warning liquid level, an early warning will be triggered.

[0248] If the predicted flow velocity is less than the minimum flow velocity, consider increasing the pump station output.

[0249] ① Data processing:

[0250] The predicted flow velocity values ​​are organized in chronological order and associated with information such as the location of the corresponding monitoring points to form a structured data table. The predicted values ​​of other relevant variables (such as pressure and liquid level) are integrated to form multivariate time series data for comprehensive analysis.

[0251] ② Characteristic structure (taking flow velocity as an example):

[0252] A) Rate of change of flow velocity: By calculating the rate of change of flow velocity during morning and evening peak hours, the magnitude of change in flow velocity between adjacent time points is observed. The formula is as follows:

[0253]

[0254] Among them, v' pre Let x' be the predicted flow velocity value at the current moment. pre-1 v is the predicted flow velocity value at the previous moment. pre This represents the rate of change of flow velocity.

[0255] B) Velocity difference between adjacent monitoring points: Calculating the velocity difference between adjacent monitoring points can be used to determine abnormal velocity conditions in a local pipe network. The formula is as follows:

[0256] v i,j =v' i -v' j

[0257] Among them, v' i and v' j The predicted flow velocity values ​​for monitoring points i and j are respectively, v i,j This represents the velocity difference between the two.

[0258] C) Moving Average Velocity: Calculated as a moving average velocity over the past ΔT′ time steps. This smooths the velocity data and reflects the overall trend of the velocity. The formula is as follows:

[0259]

[0260] Among them, v ΔT′ The moving average velocity over the past ΔT′ time steps is denoted as .

[0261] ③ Decision-making rule setting:

[0262] Threshold comparison rule: Set a minimum threshold v for the flow rate. velocity_thres If the predicted flow rate v' velocity If the output falls below this threshold, it is considered that there may be an abnormal situation, triggering a decision to increase the output of the pumping station.

[0263] Pattern matching rules:

[0264] A) Daily Cyclic Pattern: Analyze and predict the daily cyclical variation of flow velocity. For example, flow velocity is typically higher during morning and evening peak hours on weekdays, and lower at night. If the predicted flow velocity remains below normal levels during peak hours, it may indicate problems such as localized blockages in the pipeline network.

[0265] B) Weekly Pattern: Analyze the periodic variation of predicted flow velocity. For example, the flow velocity pattern on weekends may differ from that on weekdays. If the predicted flow velocity is unusually low during the weekend peak period, further investigation may be necessary.

[0266] C) Continuous downward trend: Determine whether the predicted flow rate has been continuously decreasing for ΔT′ time steps. If so, issue an early warning, indicating possible pipeline blockage or insufficient water supply.

[0267] D) Continuous upward trend: Determine whether the predicted flow rate continues to rise for a continuous period of ΔT′. If so, it indicates a possible risk of excessively high pipeline pressure or leakage.

[0268] ④ Machine learning decision-making models:

[0269] The decision regarding flow velocity is a binary classification problem (whether to increase pump station output), and a logistic regression model is used for prediction. Historical flow velocity predictions and their constructed features (flow velocity change rate, flow velocity difference between adjacent monitoring points, moving average flow velocity, etc.) are used as the input feature vector. The output label 'l' indicates whether an increase in pump station output is needed, where l=1 indicates an increase is needed, and l=0 indicates no increase is needed. The model formula is:

[0270]

[0271] in, To predict flow velocity and related feature vectors, q is the model parameter (obtained through training), defined as:

[0272]

[0273] in, It is a linear combination. If... The prediction is that an increase in pump station output is needed.

[0274] (8) Decision execution and feedback optimization:

[0275] The prediction results are input into the rule engine, which matches and judges them according to preset rules, generating preliminary decision suggestions. The decision suggestions output by the DDF module are then translated into specific operational instructions, providing early warnings of potential anomalies and sending them to the relevant actuators, such as valve controllers and pump stations. During the decision execution process, the execution status needs to be monitored in real time to ensure that the decision is executed as expected.

[0276] If any abnormal situation occurs during execution, such as malfunction of the actuator or error of the instruction, it is necessary to handle the abnormality in a timely manner, such as resending the instruction or switching to a backup actuator.

[0277] After a decision is implemented, it is necessary to collect execution result data, such as valve on / off status and pump station output flow. The execution result data is then compared with the decision objectives to evaluate the effectiveness of the decision. For example, assessing whether the issuance of a warning signal effectively prevented pipeline overflow events.

[0278] The collected feedback data is used to update and optimize the PhySync physical-spectrum dual-drive sensing model, the HydroScan-SSM model, and the rule engine or optimization algorithm in the DDF intelligent decision-making module. Through continuous feedback and optimization, the predictive accuracy and decision-making intelligence of the models are improved, forming a virtuous cycle of continuous improvement.

[0279] From the description of the above embodiments, those skilled in the art will understand that the present invention proposes a pipeline network prediction and intelligent decision-making system based on physical spectrum sensing and dynamic scanning, which has the following technical advantages:

[0280] ① In the process of processing urban pipeline network monitoring data, improving data quality is a key challenge. Factors such as sensor failure, environmental interference, or data transmission errors often lead to outliers and missing values ​​in the data, which can seriously affect model training performance and prediction accuracy. For example, outliers may distort the data distribution, causing the model to learn incorrect patterns; missing values ​​will destroy the integrity of the data, preventing the model from effectively learning the continuity and correlation of time series. To solve this problem, this invention adopts a joint outlier and missing value processing mechanism. Outliers are identified and processed using a dynamic pulsation threshold method, and missing values ​​are supplemented using data from adjacent monitoring nodes using field strength weighted interpolation technology, while simultaneously correcting outliers. This method not only effectively improves data quality but also fully considers the spatiotemporal characteristics of pipeline network monitoring data, ensuring data integrity and accuracy, and providing a reliable data foundation for subsequent model training and prediction.

[0281] ②Urban pipeline network operation status data is typical multivariate time series data, containing various indicators such as liquid level, temperature, flow rate, and ammonia nitrogen value, and these indicators have complex spatiotemporal correlations. Traditional methods struggle to effectively extract representative and discriminative features from this multivariate time series data, leading to insufficient model prediction accuracy. For example, changes in liquid level may be affected by multiple factors such as flow rate, temperature, and water quality parameters, and these factors have complex interactions. Traditional methods often only focus on the features of a single variable or a simple combination of variables, making it difficult to comprehensively capture the complex patterns in multivariate time series data. This invention proposes the PhySync physical-spectrum dual-drive sensing technology. The Dyna Spectrum module is responsible for extracting the latent distribution features and long-term trends of the time series data. Through sliding window feature extraction and the design of triple parallel convolution kernels, it extracts short-, medium-, and long-term feature segments, and combines them with a learnable physical parameter projection matrix to achieve constraint injection, sensing rapid transient events and long-term evolution patterns in the pipeline network. The EvoTopo module projects time-series data onto the frequency space using Fast Fourier Transform (FFT) to construct an initial relational matrix. Through evolutionary relational learning and reparameterization steps, it dynamically captures nonlinear correlations between channels, adapts to changes in pipeline operating conditions, and generates a binary channel mask matrix to indicate the correlations between different channels (monitoring indicators). PhySync's physical-spectrum dual-drive sensing technology can extract features in parallel, capturing both temporal features and inter-channel correlations separately, improving the comprehensiveness and accuracy of feature extraction and providing richer feature representations for subsequent model predictions.

[0282] ③ Traditional time series forecasting models suffer from bottlenecks when processing urban pipeline network operation data, particularly in accurately capturing long-term dependencies and effectively resisting dynamic noise interference. The HydroScan-SSM model proposed in this invention, through a unique architecture and mechanism, utilizes state transition matrices and input coupling matrices to accurately capture long-term dependencies in time series data, simulating the temporal evolution of pipeline network operation. Simultaneously, the transient event detection module in the model accurately captures instantaneous pressure changes by calculating pressure change acceleration, effectively distinguishing noise data from real event signals. Furthermore, the model employs a bidirectional scanning mechanism, processing time series data from both forward and reverse directions. The forward scan captures evolutionary patterns from the past to the future, while the reverse scan captures inverse dependencies from the future to the past, integrating bidirectional information to improve prediction accuracy.

[0283] ④ In practical applications, there may be discrepancies between model predictions and actual observed data. These discrepancies may stem from limitations of the model itself, such as an unreasonable model structure or improper parameter settings, or from data quality issues, such as missing data or outliers. Model bias leads to inaccurate predictions, affecting the quality of decisions based on those predictions. Simultaneously, missing data is also a significant factor affecting the accuracy of model predictions. Missing data prevents the model from learning the complete data distribution, thus impacting the reliability of the prediction results. To correct model bias and supplement missing data, this invention employs the DAM data assimilation method. DAM data assimilation is a technique that fuses observed data with model predictions, capturing the inherent structure and patterns of the data through decomposition and recombination. This invention mixes observed data with model predictions, using the observed data to correct the model predictions and reduce model bias. Furthermore, the DAM data assimilation method can also reasonably estimate and supplement missing data based on existing data information, further improving the reliability of model predictions.

[0284] ⑤ To achieve intelligent decision-making and early warning of abnormal situations, this invention constructs a DDF intelligent decision-making module. Based on model prediction results, the DDF intelligent decision-making module builds a decision rule engine, sets decision rules, and generates preliminary decision suggestions. Simultaneously, the DDF intelligent decision-making module can also translate decision suggestions into specific operational instructions, enabling real-time monitoring and execution. Through the DDF intelligent decision-making module, this invention can make intelligent decisions and provide early warnings of abnormal situations based on prediction results, improving the management and maintenance efficiency of urban pipeline networks and promptly identifying and handling potential risks and abnormal situations.

[0285] ⑥ To achieve continuous improvement and optimization of the model, this invention establishes a feedback and optimization mechanism. This invention collects decision execution result data, evaluates the effectiveness of decision execution, and feeds the execution result data back to the model for updating and optimizing the PhySync physical-spectrum dual-drive sensing model, the HydroScan-SSM model, and the DDF intelligent decision module. Through this feedback and optimization mechanism, this invention can form a virtuous cycle of continuous improvement, constantly enhancing the model's predictive accuracy and the intelligence level of decision-making, enabling the model to better adapt to constantly changing environments and needs.

[0286] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A pipeline network prediction and intelligent decision-making system based on physical spectrum sensing and dynamic scanning, characterized in that, The system includes: a data processing module, a PhySync physical-spectrum dual-drive sensing model, a HydroScan-SSM hydraulic scanning state-space model, a data assimilation module, and an intelligent decision-making module; among which: The data processing module performs outlier identification processing on the multivariate time series data acquired by the sensor, and supplements missing values ​​and corrects outliers. The PhySync physical-spectrum dual-drive sensing model extracts time features from the processed data using a dynamic physical spectrum decomposer, constructs a channel mask matrix using a frequency domain adaptive topology evolver, and then fuses the time features and the channel mask matrix using a dynamic mask focusing fusion unit. In the PhySync physical-spectrum dual-drive sensing model, the dynamic physical spectrum decomposer extracts temporal features from the processed data, including a dynamic mode initialization stage, a temporal mode decomposition stage, and a feature recombination and output stage, wherein: In the dynamic mode initialization stage, sliding window feature extraction and triple parallel convolution kernels are used. The input signal is scanned synchronously by convolution kernels of different scales to extract different feature segments, and then multi-scale feature splicing and fusion are performed. In the temporal pattern decomposition stage, multi-scale frequency domain decomposition is performed through a learnable wavelet transform layer. A three-level adaptive filter bank is used to dynamically generate three-level wavelet basis functions, each level corresponding to the typical frequency band characteristics of the pipeline network. Then, depthwise separable convolution is used to perform depthwise convolution in the time dimension. Finally, channel attention is introduced to optimize feature selection, strengthen key frequency bands and suppress noise-dominated frequency bands based on the spectral characteristics of the input signal. In the feature recombination and output stage, the frequency domain features are integrated with the original time-series features to output a standardized spatiotemporal feature matrix; The HydroScan-SSM hydraulic scanning state space model performs linear transformation on the fused features, and captures long-term dependencies in the time series through structured state space calculation using state transition matrix and input coupling matrix; and adaptively determines the state update direction according to the real-time flow direction of the pipeline network, performing forward or reverse scanning to predict risks or abnormal situations in the pipeline network. In the HydroScan-SSM hydraulic scanning state space model, transient events in the pipeline network are also captured and responded to by a transient event detection module. A three-level response mechanism is adopted to predict transient events and perform fault tracing and wave source localization. In the HydroScan-SSM hydraulic scanning state-space model, combined with transient event detection, the prediction process is divided into two modes: conventional prediction and transient response prediction. The data assimilation module dynamically adjusts and optimizes the model parameters by comparing the model prediction results with the actual observation data. The intelligent decision-making module generates decision execution instructions based on the prediction results.

2. The pipeline network prediction and intelligent decision-making system based on physical spectrum sensing and dynamic scanning according to claim 1, characterized in that, In the data processing module, outlier identification processing is performed on the multivariate time series data acquired by the sensor using the dynamic pulsation threshold method, and missing values ​​are supplemented and outliers are corrected by field strength weighted interpolation.

3. The pipeline network prediction and intelligent decision-making system based on physical spectrum sensing and dynamic scanning according to claim 2, characterized in that, In the data processing module, after supplementing missing values ​​and correcting outliers, the time series data is standardized, and segmented normalization is performed in combination with the characteristics of pipeline operation cycle to identify and classify the multi-scale cycle pattern of pipeline operation.

4. The pipeline network prediction and intelligent decision-making system based on physical spectrum sensing and dynamic scanning according to claim 1, characterized in that, In the PhySync physical-spectrum dual-drive sensing model, the frequency domain adaptive topology evolver constructs the channel mask matrix through fast Fourier transform, evolutionary relation learning, and reparameterization.

5. The pipeline network prediction and intelligent decision-making system based on physical spectrum sensing and dynamic scanning according to claim 1, characterized in that, In the PhySync physical-spectrum dual-drive perception model, the dynamic mask focusing fusion unit, based on the mask attention mechanism, fuses the temporal features extracted by the dynamic physical spectrum decomposer with the channel mask matrix generated by the frequency domain adaptive topology evolver, thereby achieving collaborative expression of time-space dual-dimensional information.

6. The pipeline network prediction and intelligent decision-making system based on physical spectrum sensing and dynamic scanning according to claim 1, characterized in that, The data assimilation module employs a hybrid model, which mixes observed data with model predictions, and captures the inherent structure and patterns of the data by decomposing and recombining the data.

7. The pipeline network prediction and intelligent decision-making system based on physical spectrum sensing and dynamic scanning according to claim 1, characterized in that, After the decision is executed, the system collects execution result data, compares the execution result data with the decision objective, and evaluates the execution effect of the decision. The collected feedback data is used to update and optimize the PhySync physical-spectrum dual-drive sensing model, the HydroScan-SSM hydraulic scanning state space model, and the intelligent decision module.

Images