A system and method for assessing the current status of urban reclaimed water utilization based on multi-source data.

Abstract

This invention relates to the field of urban reclaimed water pipeline network monitoring and assessment technology, and discloses an urban reclaimed water utilization status assessment system and method based on multi-source data. The method includes: dividing sensor groups by batch and fitting group degradation curve models, calculating degradation bias estimates; performing preliminary compensation on multi-source monitoring time-series data and extracting three-scale spatiotemporal map data sequences; training a long-scale spatiotemporal map autoencoder using an anchored benchmark training strategy, extracting preliminary slow-varying signals and calculating the upper bound of compensated residual noise; classifying and labeling nodes and reversing the degradation curve model parameters; generating decoupled pipeline network physical slow-varying anomaly detection results; performing multi-source data fusion with group degradation correction; and outputting a comprehensive assessment result of the urban reclaimed water utilization status.

Description

Technical Field

[0001] This invention relates to the field of urban reclaimed water pipeline network monitoring and assessment technology, and more specifically, to an urban reclaimed water utilization status assessment system and method based on multi-source data. Background Technology

[0002] In the scenario of monitoring and assessing the current status of urban reclaimed water pipeline network operation, a large-scale sensor cluster is deployed in the network to collect multi-source monitoring data such as water quality, water pressure, and flow rate. The reliability of each sensor is calculated independently based on data quality labels, calibration time intervals, and cross-source consistency metrics. Then, the multi-source data is fused using the Dempster-Shafer evidence theory framework, combined with the pipeline network topology and pipe segment physical properties, to generate a comprehensive assessment result of the current status of reclaimed water utilization.

[0003] However, during long-term operation of the pipeline network, two types of slowly varying signals are superimposed in the monitoring data. The first type involves the collective, gradual degradation of sensors from the same batch due to shared environmental exposure conditions. Each sensor degrades in a similar manner and rate, resulting in a high level of cross-source consistency measurement. Consistency-based confidence assessments cannot detect this collective bias. Degraded sensors mutually confirm erroneous readings, creating a blind spot in the assessment. The Dempster-Shafer framework further amplifies this bias when multiple similar bias evidence bodies are fused, generating false high confidence levels. The second type involves the pipeline network itself experiencing physical, slowly varying degradation, such as scaling on the pipe walls leading to a reduction in effective pipe diameter or a monthly increase in corrosion leakage. The diurnal variation of this degradation is far lower than normal fluctuation noise, and its manifestation in the monitoring data is highly overlapping with sensor degradation drift. If sensor degradation is compensated for before physical degradation is detected, the compensation error is amplified into a spurious signal in long-scale detection; if detection is performed directly, the source of data offset cannot be distinguished.

[0004] The superposition of the two types of slowly varying signals leads to a systematic shift in the multi-source data fusion results, making the accuracy and reliability of the assessment of the current status of reclaimed water utilization unreliable. Summary of the Invention

[0005] This invention provides a system and method for assessing the current status of urban reclaimed water utilization based on multi-source data, which solves the technical problems in related technologies, such as the difficulty in effectively separating sensor degradation noise from slowly changing abnormal signals in the pipeline network, the distortion of assessment results due to group system bias in the fusion of multi-source monitoring data, and the difficulty in quantifying and labeling monitoring blind areas.

[0006] This invention discloses a method for assessing the current status of urban reclaimed water utilization based on multi-source data, including: acquiring sensor device metadata and historical calibration record data, dividing the data into groups by batch and fitting a group degradation curve model, and calculating the degradation deviation estimate and estimation uncertainty interval of each sensor. Long-term historical multi-source monitoring time series data are acquired, and preliminary compensation is performed using the degradation bias estimate. Feature extraction is performed in three time windows: short-scale, medium-scale, and long-scale. The spatiotemporal graph data sequence at three scales is generated by combining the weighted directed graph structure of the pipeline network. The long-scale spatiotemporal map data sequence is input into a long-scale spatiotemporal map autoencoder that adopts an anchored benchmark training strategy. Preliminary slowly varying signals are extracted from the reconstruction error time series, and the upper bound of sensor compensation residual noise is calculated based on the estimated uncertainty interval. The initial slow-varying signal amplitude of each node is compared with the upper bound of the corresponding sensor compensation residual noise. The nodes are classified and marked as high-confidence slow-varying candidate nodes and indistinguishable nodes. When the correlation coefficient of the group to which the indistinguishable node belongs is higher than the preset threshold, the degradation curve model parameters are reversed. The original monitoring data is recompensated using the corrected degradation bias estimate and the updated slow-varying signal is extracted. The decoupled pipeline physical slow-varying anomaly detection results are generated by spatial consistency filtering. Group degradation perception correction is performed on sensor credibility scores, and Dempster-Shafer evidence combination operation is performed on evidence bodies in the same group after common deviation component stripping to generate multi-source data fusion evaluation index. Based on the decoupled network physical slow-change anomaly detection results and the multi-source data fusion evaluation index, the physical degradation quantification results and monitoring blind zone range are calculated, and the comprehensive evaluation results of the current status of urban reclaimed water utilization are output.

[0007] Furthermore, the step of dividing the data into batches and fitting a group degradation curve model to calculate the degradation bias estimate and the estimation uncertainty interval for each sensor includes: Sensors are grouped into the same batch based on batch number and installation date; Historical environmental parameter data of all sensor installation locations within each group are obtained. After Z-score standardization of each environmental parameter, the mean value is taken at all sensor locations within the group to generate a common environmental exposure feature vector for the group. Using the cumulative runtime and the common environmental exposure feature vector of the group as input, the historical calibration deviation data of all sensors in the group are jointly fitted to generate a group degradation curve model. The group degradation curve model is a parametric degradation model, which includes the product of the group's basic degradation rate coefficient and the time degradation basis function, the product of the inner product of the environmental influence coefficient vector and the group's common environmental exposure feature vector with the environmental and time coupling functions, and the group's initial deviation constant term. Each parameter is obtained by least squares fitting. Substitute the cumulative runtime of each sensor into the group degradation curve model of its respective group to obtain the degradation bias estimate; the radius of the uncertainty interval is determined by the square root of the sum of the squares of the standard deviation of the fitting residuals of the group degradation curve model and the standard deviation of the historical deviation of the individual sensors.

[0008] Furthermore, the feature extraction performed in three time windows—short-scale, medium-scale, and long-scale—includes: Within a short-scale time window, the statistical fluctuation of the time series data after preliminary compensation of each node is calculated within the window, and the instantaneous fluctuation characteristics are extracted. Within the mesoscale time window, the dominant frequency component and its corresponding amplitude are obtained by performing Fourier decomposition on intraday data segments as intraday periodic features, and the slope and residual are obtained by performing linear fitting on the daily average series of consecutive days as intraday trend features. Within a long-scale time window, seasonal components are obtained as seasonal periodic features by seasonally decomposing annual data, and long-term drift features are obtained by trend fitting of the residual sequence after removing seasonal components. In the weighted directed graph of the pipeline network, the edge weights are determined by weighted synthesis of pipe segment length, pipe diameter and pipe material attributes after mean normalization based on range, and pipe segment flow direction information is introduced so that downstream nodes receive feature propagation from upstream nodes.

[0009] Furthermore, the execution process of the anchoring benchmark training strategy includes: Acquire long-scale spatiotemporal map data of the pipeline network during the period when it is confirmed to be in good condition, and save it as an anchoring reference dataset. The anchoring reference dataset remains unchanged during the system operation cycle. In periodic retraining, the anchored benchmark dataset is mixed with recently collected long-scale spatiotemporal graph data in a fixed ratio as training data, wherein the proportion of the anchored benchmark dataset is not less than the preset lower limit of the total training data. The period in which the pipeline network is confirmed to be in good condition refers to the operational phase after the pipeline network has undergone comprehensive manual inspection and sensor calibration. The deviations of each sensor reading from the standard reference value are all within the factory accuracy range, and the pipeline network has no known physical defects.

[0010] Furthermore, the long-scale spatiotemporal graph autoencoder includes an encoder and a decoder. The encoder extracts the spatial features of each node in the graph structure through a graph convolutional layer, and then extracts the temporal evolution pattern of each node's features along the time dimension through a temporal coding layer, outputting a low-dimensional spatiotemporal implicit representation. The decoder takes the low-dimensional spatiotemporal implicit representation as input, and recovers the time dimension information and spatial dimension information sequentially through a temporal decoding layer and a graph deconvolutional layer, outputting a reconstructed spatiotemporal graph data sequence. The long-scale spatiotemporal graph autoencoder uses the mean squared error covering all nodes, time steps, and feature dimensions between the reconstructed output and input as the training loss function. The step of extracting the initial slow-varying signal from the reconstruction error time series includes: performing median filtering on the long-scale reconstruction error time series of each node and then performing linear regression to extract the monotonic trend component as the initial slow-varying signal. The calculation method for the upper bound of the sensor compensation residual noise is as follows: sum the squares of the estimated uncertainty interval radii of each sensor at the node, take the square root, and then multiply it by the square root of the number of sampling steps included in the long-scale time window. The estimated uncertainty interval radius is converted to a standardized dimensionless space consistent with the initial slowly varying signal before participating in the calculation.

[0011] Furthermore, the reverse correction of the degradation curve model parameters for the group to which the indistinguishable nodes belong includes: Obtain the group number of the group to which each indistinguishable node belongs, and calculate the mean of the Pearson correlation coefficients between each pair of the preliminary slow-varying signal sequences of the corresponding nodes of each sensor within the same group, as the correlation coefficient within the group; When the correlation coefficient within the group is higher than a preset threshold, the mean direction and mean amplitude of the initial slow-changing signal of the indistinguishable nodes within the group are calculated. The equivalent rate increment is obtained by dividing the mean amplitude by the duration of the long-scale time window. This equivalent rate increment is then superimposed onto the group's basic degradation rate coefficient in the mean direction. The parameter increment for each correction does not exceed the preset upper limit of the current parameter value. If the upper limit is exceeded, the correction is made according to the upper limit value. The degradation bias estimates of each sensor were recalculated using the modified group degradation curve model.

[0012] Furthermore, the spatial consistency filtering includes: The initial slow-varying signals of high-confidence slow-varying candidate nodes are combined with the updated slow-varying signals of indistinguishable nodes to obtain the comprehensive result of the slow-varying signals of each node. Based on the topology of the weighted directed graph of the pipeline network, for nodes that are adjacent in the topology and share the same pipe segment, when the slow-changing signals of the two nodes are in the same direction and the amplitude difference is within the reasonable attenuation range determined by the physical properties of the pipe segment, the slow-changing signal is retained; when the slow-changing signal of a node has no spatial continuity support in its topological neighborhood, the confidence flag of the signal is reduced. The reasonable attenuation range is determined by the pipe segment length and pipe segment material, and a pipe segment age weighting factor is introduced to determine the criteria for relaxing the spatial consistency constraint on the nodes connected to pipe segments with longer pipe ages.

[0013] Furthermore, the group degradation perception correction for the sensor confidence score includes: Obtain the basic reliability score for each sensor. When the expected degradation deviation of the sensor group exceeds the accuracy tolerance threshold, multiply the basic reliability score by a reduction factor to obtain the reduced reliability score. The reduction factor is the larger of the ratio of the absolute value of the degradation deviation estimate to the accuracy tolerance threshold to the accuracy tolerance threshold, after being weighted by the reduction rate parameter and subtracted from 1 and 0. A confidence upper limit constraint is applied to the sensors in the same group, which is monotonically decreasing with the cumulative running time. The confidence upper limit constraint is a monotonically decreasing piecewise linear function with respect to the cumulative running time, and its inflection point position and slope are determined according to the dispersion parameter of the individual degradation rate within the group. The process of stripping the common deviation component is as follows: obtain the basic probability assignment function corresponding to each sensor in the same group, calculate the mean value of the projection component of each basic probability assignment function in the expected deviation direction as the common deviation component, and then perform normalization processing after subtracting the common deviation component from each basic probability assignment function. When the proportion of evidence from the same degenerate group in the combined evidence exceeds a preset group proportion threshold, the evidence from that group is downsampled.

[0014] Furthermore, the calculation of physical degradation quantification results and monitoring blind zone range outputs a comprehensive assessment result of the current status of urban reclaimed water utilization, including: After the amplitude of the slowly varying physical signal is restored to its original dimensional value through inverse normalization transformation, the cumulative deviation of physical degradation and degradation rate of each region are calculated in combination with the physical attribute data of the pipe section. Based on the characteristic patterns of the slow-changing physical signals, the causes of slow-changing anomalies are attributed. When the water pressure continues to drop and the flow rate gradually decreases, it is attributed to the shrinkage of the effective pipe diameter caused by scaling on the pipe wall. When the flow rate continues to increase abnormally and the water pressure at the downstream node decreases abnormally, it is attributed to the increase in corrosion leakage. Based on the degradation bias estimate and confidence score distribution of each sensor group, the pipeline area covered by sensors whose group confidence score is lower than the usable threshold is identified as the monitoring blind zone. Each assessment item is labeled with a confidence level. The area corresponding to the candidate node with high confidence and slow change is labeled with a high confidence level. The area corresponding to the node that cannot be distinguished after the degradation model parameters are reverse corrected is labeled with a medium confidence level. The area within the monitoring blind zone is labeled with a low confidence level.

[0015] This invention provides a system for assessing the current status of urban reclaimed water utilization based on multi-source data, comprising: The group degradation modeling module is used to acquire sensor device metadata and historical calibration record data, divide the data into groups by batch and fit the group degradation curve model, and calculate the degradation deviation estimate and estimation uncertainty interval for each sensor. The multi-scale spatiotemporal map construction module is used to acquire long-term historical multi-source monitoring time series data, perform preliminary compensation using degradation bias estimates, extract features at three time scales, and generate spatiotemporal map data sequences at three scales by combining the pipeline network weighted directed graph structure. The slow-varying signal extraction module is used to input the long-scale spatiotemporal map data sequence into the long-scale spatiotemporal map autoencoder which adopts the anchored benchmark training strategy, extract the preliminary slow-varying signal from the reconstruction error time series, and calculate the upper bound of the sensor compensation residual noise. The node classification and degradation correction module is used to compare the initial slow-varying signal amplitude of each node with the upper bound of the sensor compensation residual noise, classify and label the nodes, and reverse the degradation curve model parameters of the group to which the indistinguishable nodes belong. The physical slow-varying anomaly detection module is used to recompensate the original monitoring data with the corrected degradation bias estimate and extract the updated slow-varying signal. After spatial consistency filtering, it generates the decoupled pipeline physical slow-varying anomaly detection results. The multi-source data fusion module is used to correct the group degradation perception of sensor credibility scores, and to perform Dempster-Shafer evidence combination operation after stripping common deviation components from evidence in the same group to generate multi-source data fusion evaluation indicators. The comprehensive evaluation output module is used to calculate the physical degradation quantification results and monitoring blind zone range based on the decoupled pipeline network physical slow-change anomaly detection results and multi-source data fusion evaluation indicators, and output the comprehensive evaluation results of the current status of urban reclaimed water utilization.

[0016] This invention proactively predicts and initially compensates for the progressive degradation of sensors in the same batch by using a group degradation curve model. It also employs an anchored benchmark training strategy to train a long-scale spatiotemporal map autoencoder, ensuring that the reconstructed benchmark does not drift with the slow degradation of the pipeline network. This solves the technical problem of the high overlap between sensor group degradation and slow degradation of the pipeline network in monitoring data, which leads to a systematic shift in the multi-source data fusion results. The invention achieves the technical effect of being able to distinguish the sources of the two types of slow-changing signals, reducing false high confidence levels generated by sensors in the same group of degradation in evidence fusion, and outputting a comprehensive assessment result of the current status of urban reclaimed water utilization with confidence level annotations. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for assessing the current status of urban reclaimed water utilization based on multi-source data provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the degradation deviation estimate and uncertainty radius of the node sensor in batch A provided by an embodiment of the present invention; Figure 3This is a schematic diagram of the distribution of feature values ​​of node P01 at various time scales provided in an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the initial slowly varying signal amplitude at each node with the upper bound of the compensated residual noise, provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the signal correlation coefficient matrix within an indistinguishable node group provided in an embodiment of the present invention; Figure 6 This is a schematic diagram showing the comparison of sensor reliability score before and after correction according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the decoupled physical slow-varying signal amplitude (normalized space) of the pipeline network provided in an embodiment of the present invention; Figure 8 This is a schematic diagram comparing the degradation bias estimates before and after parameter correction of the group degradation model provided in this embodiment of the invention; Figure 9 This is a schematic diagram of the comprehensive integrated evaluation indicators and effective transmission and distribution capacity of each evaluation area provided in the embodiments of the present invention. Detailed Implementation

[0018] In the scenario of monitoring and assessing the current status of urban reclaimed water pipeline network operation, a large-scale sensor cluster is deployed in the network to collect multi-source monitoring data such as water quality, water pressure, and flow rate. The reliability of each sensor is calculated independently based on data quality labels, calibration time intervals, and cross-source consistency metrics. Then, the multi-source data is fused using the Dempster-Shafer evidence theory framework, combined with the pipeline network topology and pipe segment physical properties, to generate a comprehensive assessment result of the current status of reclaimed water utilization.

[0019] However, during long-term operation of the pipeline network, two types of slowly varying signals are superimposed in the monitoring data. The first type involves the gradual degradation of sensors from the same batch due to shared environmental exposure conditions. Each sensor degrades in a similar manner and rate, resulting in a high level of cross-source consistency measurement. Consistency-based reliability assessments cannot detect this group bias, and the degraded sensors mutually confirm erroneous readings, creating a blind spot in the assessment. The Dempster-Shafer framework further amplifies this bias when multiple similar bias evidence bodies are fused, generating false high confidence levels. The second type involves physical slowly varying degradation within the pipeline network itself, such as scaling on the pipe walls leading to a reduction in effective pipe diameter and a monthly increase in corrosion leakage. The diurnal variation of this degradation is far lower than normal fluctuation noise, and its manifestation in the monitoring data is highly overlapping with sensor degradation drift. If sensor degradation is compensated for before physical degradation is detected, the compensation error is amplified into a spurious signal in long-scale detection; if detection is performed directly, the source of data offset cannot be distinguished. The superposition of these two types of slowly varying signals causes a systematic shift in the multi-source data fusion results, making the accuracy and reliability of the reclaimed water utilization status assessment unreliable.

[0020] It should be understood that the operating hardware environment of this embodiment includes: a sensor cluster deployed at each node of the urban reclaimed water pipeline network, and each sensor uploads monitoring data to the data processing server through a data acquisition gateway; the data processing server is equipped with a computing unit for performing spatiotemporal graph autoencoder training and inference, and stores pipeline network topology data, pipe segment physical attribute data and sensor device metadata.

[0021] At least one embodiment of the present invention discloses a method for assessing the current status of urban reclaimed water utilization based on multi-source data, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain sensor device metadata and historical calibration record data, divide them into groups by batch and fit the group degradation curve model, and calculate the degradation deviation estimate of each sensor; Acquire device metadata and historical calibration records for all sensors in the pipeline network. Device metadata includes batch number, installation date, installation location, and sensor model. Historical calibration records include calibration deviation values ​​and calibration timestamps for each calibration. Group sensors into batch groups based on batch number and installation date. Each group contains a set of sensors deployed within the same production batch and around the same installation time.

[0022] For each group, historical environmental parameter data for all sensor installation locations within that group, including temperature, humidity, and media corrosivity, are acquired. Before fitting the group degradation curve model, the raw environmental parameter data are Z-score standardized to eliminate the influence of dimensional differences between different environmental parameters on the joint fitting. The average of each standardized environmental parameter across all sensor locations within the group is then calculated to generate a common environmental exposure feature vector for the group. ,in This is the group number.

[0023] Obtain the historical calibration deviation sequence and cumulative runtime of all sensors within each group, using the cumulative runtime as the basis. and group shared environment exposure feature vector Using the expected deviation as the output, the historical calibration deviation data of all sensors in the group are jointly fitted to generate a group degradation curve model. Based on the group degradation curve model Calculate each sensor At the present moment Degradation bias estimator Estimating the uncertainty interval and the dispersion parameter of individual degradation rate within the group .

[0024] Among them, degradation bias estimator By using sensors The cumulative runtime and the group's shared environment exposure feature vector are substituted into the group degradation curve model. Obtain; estimate the radius of the uncertainty interval Degeneration curve model of groups The standard deviation of the fitting residuals and the sensor Determined in conjunction with the historical deviation of the group's degradation mean curve, specifically, ,in Group degradation curve model The standard deviation of the fit residuals over all historical calibration data within the group. For sensors The standard deviation of the historical calibration deviation sequence relative to the predicted values ​​of the group degradation mean curve; the dispersion parameter of individual degradation rate within the group. is the standard deviation of the degradation rate of each sensor within the group, where the degradation rate of each sensor is determined by the slope obtained by linear regression of its historical calibration deviation sequence against the cumulative runtime.

[0025] It should be noted that the group degradation curve model The parameterized degradation model takes the following form:

[0026] in, The group's basic degradation rate coefficient. For time-degrading basis functions, This is a vector of environmental impact coefficients. Represents the vector of environmental impact coefficients transpose, This is an environment-time coupling function. The initial deviation constant of the group. This represents the cumulative runtime. These parameters were obtained by performing a least-squares fit on the historical calibration deviation data of all sensors within the group. and A monotonically increasing function, such as a polynomial or logarithmic function, can be selected, with the specific form determined based on the trend characteristics of the group's historical calibration deviation. This is because the group shares a common environmental exposure feature vector. The group degradation curve model has been standardized by Z-score. All items maintain consistency on the numerical scale. The environmental impact component in the item and Items under the same dimension can be added directly.

[0027] Step 2: Obtain long-term historical multi-source monitoring time series data, perform preliminary compensation using degradation bias estimates, and extract features at three time scales to generate spatiotemporal map data sequences at three scales; Long-term historical multi-source monitoring time-series data for each node in the pipeline network were acquired, including time-series data for water quality parameters, water pressure, and flow rate. Since the dimensions of water quality parameters, water pressure, and flow rate are different, Z-score standardization was applied to each type of monitoring time-series data before initial compensation. This unified the different physical quantities into a dimensionless standardized numerical space, eliminating computational biases caused by dimension inconsistencies in subsequent multidimensional feature extraction and spatiotemporal graph autoencoder processing. Degradation bias estimation was then used. The standardized readings of each sensor at each sampling time are subtracted and compensated point by point to generate preliminary compensated time series data.

[0028] Feature extraction was performed on the pre-compensated time-series data in three time windows: short-scale, meso-scale, and long-scale. Within the short-scale time window, statistical fluctuations were calculated for each node's pre-compensated time-series data, and instantaneous fluctuation features were extracted. Within the meso-scale time window, intraday periodic features and inter-day trend features were extracted from each node's pre-compensated time-series data. Intraday periodic features were obtained by performing Fourier decomposition on intraday data segments to obtain the dominant frequency component and its corresponding amplitude. Inter-day trend features were obtained by linearly fitting the daily average sequence over multiple consecutive days to obtain the slope and residuals. Within the long-scale time window, seasonal periodic features and long-term drift features were extracted from each node's pre-compensated time-series data. Seasonal periodic features were obtained by seasonal decomposition of annual data to obtain the seasonal component. Long-term drift features were obtained by trend fitting on the residual sequence after removing the seasonal component.

[0029] Obtain the pipeline network topology data and represent the pipeline network as a weighted directed graph. ,in For each monitoring point in the node set, The edge set corresponds to the pipe segment connection relationship. This is the edge weight matrix, where the weights are determined by the pipe segment length, pipe diameter, and pipe material properties. Since the dimensions of pipe segment length, pipe diameter, and pipe material properties are different, before calculating the edge weights, each of these physical properties is normalized using a mean-based method based on its range, unifying all properties to the same standard. The data is then weighted and synthesized within intervals to eliminate the influence of differences in the dimensions of different physical attributes on the weight calculation. The node features extracted at each scale are combined with the weighted directed graph structure of the pipeline network to generate short-scale, medium-scale, and long-scale spatiotemporal graph data sequences.

[0030] It should be noted that the short-scale time window ranges in length from minutes to hours, the medium-scale time window ranges in length from days to weeks, and the long-scale time window ranges in length from months to quarters. The length of the time windows at each scale is determined based on the sampling frequency of the pipeline monitoring data and the operating characteristics of the pipeline network. In the spatiotemporal map data sequence at each scale, each time step corresponds to a snapshot of the map, and each node in the snapshot carries the feature vector extracted at that scale.

[0031] In this embodiment of the application, in order to preserve the spatial dependency information between nodes in the spatiotemporal graph data sequence at various scales, the flow direction information of the pipe segment is also introduced in addition to the physical attributes of the pipe segment when calculating the edge weight. For directed edges... ,in and These are the starting and ending nodes of the edge, respectively. When the pipe segment starts from node... Flow to Node At that time, the weight of the edge is multiplied by the flow direction coefficient on the basis of the physical attribute weight, so that the downstream node can receive the feature propagation of the upstream node, thereby capturing the signal propagation law in the direction of water flow in the subsequent spatiotemporal graph autoencoder processing.

[0032] Step 3: Input the long-scale spatiotemporal map data sequence into the long-scale spatiotemporal map autoencoder using the anchored benchmark training strategy, extract the initial slowly varying signal, and calculate the upper bound of the sensor-compensated residual noise. The long-scale spatiotemporal graph data sequence is input into the long-scale spatiotemporal graph autoencoder for reconstruction, and the long-scale reconstruction error time series of each node is calculated.

[0033] It should be understood that a long-scale spatiotemporal graph autoencoder consists of two components: an encoder and a decoder. The encoder takes a long-scale spatiotemporal graph data sequence as input, where the input contains a graph structure. and the feature sequences of each node within a long-scale time window. ,in For the number of nodes, For time steps, The encoder extracts the spatial features of each node in the graph structure through graph convolutional layers, and then extracts the temporal evolution pattern of each node's features along the time dimension through temporal coding layers, outputting a low-dimensional spatiotemporal latent representation. ,in This represents the hidden representation dimension. The decoder uses low-dimensional spatiotemporal hidden representation. As input, the temporal dimension information is recovered through a temporal decoding layer, and the spatial dimension information is recovered through a graph deconvolution layer, outputting a reconstructed long-scale spatiotemporal graph data sequence. The low-dimensional spatiotemporal implicit representation of the encoder output. Simultaneously serving as input to the decoder, the two constituent units are implicitly represented in a low-dimensional spatiotemporal spacetime. To transmit data.

[0034] Furthermore, the spatial feature extraction process of the graph convolutional layer is as follows: using the node feature matrix and the graph adjacency weight matrix... As input, following the message passing mechanism of Graph Convolutional Networks (GCN), the features of each node are weighted and aggregated with the features of its neighboring nodes according to edge weights. After linear transformation and nonlinear activation, the spatial augmentation features of each node are output. The temporal evolution pattern extraction process of the temporal coding layer is as follows: taking the features of each node as input, the features of each node are weighted and aggregated with the features of its neighboring nodes according to edge weights. Taking the spatially enhanced feature sequence at each time step as input, a Long Short-Term Memory (LSTM) network is used to process it stepwise along the time dimension. The final hidden state is used as the temporal evolution code of that node, and concatenated with the temporal evolution codes of all nodes to output a low-dimensional spatiotemporal hidden representation. The temporal decoding layer and the graph deconvolution layer are the symmetrical inverse processes of the temporal coding layer and the graph convolution layer, respectively, which sequentially recover the temporal and spatial dimension information and output the reconstructed long-scale spatiotemporal graph data sequence. .

[0035] Long-scale spatiotemporal graph autoencoders use reconstruction loss as the training objective, and the loss function is the reconstruction output. With input Mean square error between:

[0036] in, For nodes At time step The 3D reconstruction of eigenvalues, For the corresponding input feature values, For the number of nodes, For time steps, For node feature dimensions, the summation range covers all. Each node Each time step and The loss function value reflects the average reconstruction error across all nodes and feature dimensions within the entire long-scale time window. This is due to the input feature sequence... The data has already undergone Z-score standardization in step 2, ensuring that all feature dimensions reside in a unified dimensionless numerical space. Therefore, all terms in the mean squared error loss function have consistent dimensions and can be directly summed. The Adam optimization algorithm is used during training to minimize this loss function.

[0037] The long-scale spatiotemporal graph autoencoder adopts an anchored benchmark training strategy. The execution process of the anchored benchmark training strategy is as follows: First, acquire long-scale spatiotemporal map data of the pipeline network when it is in a confirmed good condition, and save this part of the data as the anchoring benchmark dataset. The anchoring benchmark dataset remains unchanged throughout the entire system operation cycle and is not updated with retraining.

[0038] Second, in subsequent periodic retraining, the anchored benchmark dataset and recently collected long-scale spatiotemporal graph data are mixed in a fixed ratio as training data, with the proportion of the anchored benchmark dataset not less than the preset lower limit of the total training data volume, and the long-scale spatiotemporal graph autoencoder is retrained.

[0039] It should be noted that the period of confirmed good condition refers to the initial operation stage or the stage after maintenance of the pipeline network after comprehensive manual inspection and full calibration of sensors. During the period of confirmed good condition, the deviations of the readings of each sensor from the standard reference values ​​are all within the factory accuracy range, and there are no known physical defects in the pipeline network.

[0040] After obtaining the long-scale reconstruction error time series of each node, median filtering is applied to the long-scale reconstruction error time series of each node to remove the interference of isolated outliers on trend fitting. Then, linear regression is performed on the filtered long-scale reconstruction error time series to extract the monotonic trend component, which is used as the initial slow-varying signal. ,in Node numbering. Since the long-scale reconstruction error is calculated based on the standardized feature sequence, the initial slowly varying signal... It also exists in the standardized dimensionless numerical space.

[0041] Meanwhile, based on the radius of the uncertainty interval of the degradation bias estimator Calculate the upper bound of sensor-compensated residual noise for each node over a long-scale time window. The calculation method is as follows:

[0042] in, For nodes A collection of sensors deployed in the area. For sensor set The sensor number in the middle, For sensors The radius of the estimated uncertainty interval, This represents the number of sampling steps contained within a long-scale time window. Before participating in the above calculations, it needs to be converted from the original physical dimension space to a space consistent with the initial slowly varying signal. A consistent, standardized dimensionless space, the transformation method is to... Divide by the standard deviation of the corresponding historical sensor data to set the upper bound of the sensor's residual noise compensation. With initial slow-changing signal Being in the same numerical space, they support the amplitude comparison operation in step 4. The physical meaning of the above equation is: the compensation uncertainties of each sensor are independently superimposed at the nodes, and then... The accumulation of sampling steps, in the worst case, represents the maximum offset that can be generated for the long-scale trend component, and is used to determine whether the initial slowly varying signal exceeds the range that the sensor compensation error can explain.

[0043] Furthermore, The factor is derived from the worst-case estimation of the independent superposition of the compensation residual errors at each time step: when the compensation residual errors of each sampling step are independent and all are expressed as... When it is the upper boundary, it passes through After accumulating through each sampling step, the upper bound of the composite residual error of each sensor at the node is: The upper bound provides a worst-case offset estimate under the assumption that the errors are independent and of the same magnitude, which is used to ensure that the classification judgment in step 4 is sufficiently conservative and avoids misjudging the cumulative effect of sensor compensation error as a slowly changing signal of the pipeline network.

[0044] Step 4: Compare the initial slow-varying signal amplitude with the upper bound of the sensor-compensated residual noise, classify and label each node, and reverse-correct the degradation curve model parameters of the group to which the indistinguishable nodes belong. The initial slowly varying signal amplitude at each node The corresponding upper bound of sensor compensation residual noise Compare. To The nodes are marked as high-confidence slow-changing candidate nodes. The nodes are marked as indistinguishable nodes.

[0045] For indistinguishable nodes, obtain the group number of the group to which each indistinguishable node belongs, extract the preliminary slowly varying signals of the corresponding nodes of each sensor within the same group, and calculate the correlation coefficient within the group. Correlation coefficient within the group It is the mean of the Pearson correlation coefficients between each pair of the initial slow-varying signal sequences of corresponding nodes of each sensor within the same group.

[0046] When the correlation coefficient within the group Higher than the preset threshold At that time, it was determined that the slowly varying signals of indistinguishable nodes within the group mainly originated from the group degradation curve model. The systemic compensation is insufficient. Preliminary slowly varying signals of each indistinguishable node within the group are obtained, and the mean direction and mean amplitude of the slowly varying signals within the group are calculated. Based on the mean direction and mean amplitude, the group degradation curve model is... The parameters are reversed: the group's basic degradation rate coefficient is adjusted. Adjustments are made by increasing or decreasing the parameter value according to the equivalent rate increment of the mean amplitude over a long-scale time window. The direction of adjustment is consistent with the direction of the mean, and the parameter increment for each adjustment does not exceed a preset upper limit for the current parameter value. When the calculated adjustment exceeds this upper limit, the adjustment is made according to the upper limit value, and the remaining adjustment is carried over to the next iteration. The modified group degradation curve model is then used. Recalculate the degradation bias estimates for each sensor.

[0047] It should be noted that the preset threshold The range of values ​​is The specific value is determined based on the number of sensors within the group and the distribution of the pipeline topology. When the sensors within the group are distributed in different topological regions of the pipeline, a high intra-group correlation coefficient indicates that these sensors have collectively experienced similar non-physical offsets, because physical degradation in different regions of the pipeline typically does not have such a high correlation.

[0048] Furthermore, the equivalent rate increment of the mean amplitude over a long-scale time window is calculated as follows: the mean amplitude of the initial slow-varying signal of the indistinguishable nodes within the group is divided by the duration of the long-scale time window to obtain the degradation rate increment per unit time. This degradation rate increment is then added to the current group's baseline degradation rate coefficient. Above, the modified group degradation curve model It can bridge the systematic bias between preliminary compensation and actual observations over a long-scale time window.

[0049] Step 5: Recompensate the original monitoring data using the corrected degradation bias estimate, extract the updated slow-varying signal, and combine it with spatial consistency filtering to generate decoupled pipeline physical slow-varying anomaly detection results; After re-standardizing the original monitoring data of each node in the pipeline network using the corrected degradation bias estimate, point-by-point subtraction compensation is performed to generate recompensated time-series data. Long-scale node features are then extracted from the recompensated time-series data using the long-scale time window feature extraction method described in step 2, and combined with the weighted directed graph structure of the pipeline network to generate an updated long-scale spatiotemporal graph data sequence. This updated long-scale spatiotemporal graph data sequence is input into a long-scale spatiotemporal graph autoencoder to calculate the updated reconstruction error time series of each node. Linear regression is then performed on the updated reconstruction error time series to extract the monotonic trend component, which is used as the updated slowly varying signal.

[0050] The initial slow-changing signals of high-confidence slow-changing candidate nodes are merged with the updated slow-changing signals of indistinguishable nodes to obtain the comprehensive result of the slow-changing signals of each node. Pipeline physical attribute data, including pipe material, age, diameter, and historical maintenance records, are acquired. Based on the topology of the weighted directed graph of the pipeline network, spatial consistency constraints are applied to the comprehensive result of the slow-changing signals of each node for filtering: for node pairs that are topologically adjacent and share the same pipe segment, the slow-changing signals of the two nodes are retained when their directions are consistent and the amplitude difference is within a reasonable attenuation range determined by the physical attributes of the pipe segment; when a node's slow-changing signal lacks spatial continuity support within its topological neighborhood, the confidence flag of that signal is reduced. After spatial consistency filtering, decoupled pipeline physical slow-changing anomaly detection results are generated, which include the amplitude, direction, and spatial confidence flag of the physical slow-changing signals of each node. In this process, the amplitude and direction of the slow-changing physical signal are both in the standardized dimensionless numerical space established in step 2. When outputting the comprehensive evaluation results, the inverse standardization transformation is performed based on the historical mean and standard deviation of various monitoring quantities to restore the amplitude of the slow-changing physical signal to the original dimensional value of the corresponding physical quantity, so as to conduct an interpretable comparative analysis with the physical attribute data of the pipe section.

[0051] It should be noted that the reasonable attenuation range in spatial consistency filtering is determined by both the pipe segment length and the pipe material. For metal pipe segments, the upper limit of signal attenuation between adjacent nodes is proportional to the pipe segment length; for non-metallic pipe segments, the upper limit of attenuation must also consider the impact of the number of pipe joints on signal propagation. A pipe segment age weighting factor is also introduced in the spatial consistency filtering process. For node pairs connected to pipe segments with longer ages, the judgment conditions for spatial consistency constraints are appropriately relaxed, allowing for larger signal amplitude differences and weaker spatial continuity support, in order to avoid missing isolated, slowly changing signals caused by concentrated degradation due to localized aging.

[0052] Step 6: Based on the modified group degradation curve model, perform group degradation perception correction on the sensor credibility score, and perform Dempster-Shafer evidence combination operation after preprocessing the evidence body by stripping common bias components to generate a multi-source data fusion evaluation index corrected for group degradation. Based on the modified group degradation curve model Calculate the expected degradation bias of each sensor group at the current time. Obtain the baseline reliability score for each sensor, which is determined by the data quality label and calibration time interval.

[0053] Group degradation perception correction is applied to the basic reliability scores of each sensor. When the expected degradation deviation of the sensor's group exceeds the accuracy tolerance threshold... At that time, based on the deviation amount and the accuracy tolerance threshold The baseline credibility score is reduced by a factor of 1, resulting in a reduced credibility score. The calculation method is as follows:

[0054] in, For sensors Basic credibility score, For sensors The degradation bias estimator, For the accuracy tolerance threshold, The deceleration rate parameter is used to control the confidence score from exceeding the accuracy tolerance threshold due to degradation deviation. The subsequent descent rate, The larger the value, the higher the precision tolerance threshold. The faster the credibility score decreases, the better. In the above formula... With precision tolerance threshold All sensors are located in the standardized dimensionless numerical space established in step 2, have consistent dimensions, and can be directly used for ratio calculations. Simultaneously, a confidence upper limit constraint, which monotonically decreases with cumulative runtime, is applied to the sensors in the same group. , Regarding cumulative runtime A monotonically decreasing piecewise linear function. The initial value is 1, and then... Increase the linearity to reduce to a preset minimum confidence level; the specific inflection point location and slope are determined based on the dispersion parameter of the individual degradation rate within the group. Determine: the dispersion parameter of individual degradation rate within the group The larger the slope, the earlier the inflection point and the greater the absolute value of the slope, indicating that the greater the individual differences in the rate of degradation, the faster the upper limit of credibility decreases. When the reduced credibility score exceeds the upper limit of credibility constraint... When applying the upper limit constraint on credibility. The value is used as the final credibility score.

[0055] Furthermore, the accuracy tolerance threshold The physical meaning is: when the absolute value of the sensor degradation bias estimate is... Exceeding the accuracy tolerance threshold At this point, it is considered that the degradation of the sensor has had a substantial impact on measurement accuracy, and its reliability score needs to be reduced. Accuracy tolerance threshold. The value is determined based on the sensor's factory accuracy specifications. Specifically, it is the corresponding value in the standardized dimensionless space established in step 2 for the upper limit of the sensor's factory allowable deviation, which is the result obtained by dividing the upper limit of the sensor's factory allowable deviation by the standard deviation of the corresponding historical data of the sensor.

[0056] Before multi-source data fusion, for the evidence bodies corresponding to each sensor belonging to the same degradation group, the modified group degradation curve model is used. The common deviation component is preprocessed by stripping the expected deviation direction. The specific process is as follows: obtain the set of basic probability assignment functions corresponding to each sensor in the same group, calculate the mean of the projection components of each basic probability assignment function in the set of basic probability assignment functions in the expected deviation direction, take the mean of the projection components as the common deviation component, subtract the common deviation component from each basic probability assignment function and then perform normalization processing to obtain the stripped independent information components.

[0057] Furthermore, the projection component of the basic probability assignment function in the expected deviation direction is calculated as follows: the sensor readings are compared with the corrected group degradation curve model. The expected deviations at the current moment are compared, and the components in each sensor reading that are in the same direction as the expected deviation are extracted as projection values. The average of the projection values ​​of all sensors in the same group is taken to obtain the common deviation component of the group. The common deviation component is subtracted from the basic probability assignment function corresponding to each sensor. That is, the basic probability assignment of each hypothesis is reduced proportionally according to the proportion of the common deviation component. After reduction, the basic probability assignment of all propositions is renormalized to satisfy the normalization constraint of the basic probability assignment function, and the stripped independent information components are obtained.

[0058] After stripping away the individual information components of each piece of evidence, a Dempster-Shafer evidence combination operation is performed, weighted according to the corrected credibility score. When the proportion of evidence from the same degraded group among the combined evidence exceeds a preset group proportion threshold, evidence from that group is downsampled, and some evidence is randomly selected to participate in the combination operation, reducing the concentration of evidence from the same group in the fusion result. Finally, a multi-source data fusion evaluation index corrected for group degradation is generated.

[0059] Step 7: Based on the decoupled network physical slow-change anomaly detection results and the multi-source data fusion evaluation index after population degradation correction, calculate the physical degradation quantification results and monitoring blind zone range, and output the comprehensive evaluation results of the current status of urban reclaimed water utilization; Based on the decoupled detection results of slow-varying anomalies in the pipeline network, a quantitative analysis of physical degradation is performed on each pipeline area. The amplitude and direction of the slow-varying physical signals at each node are acquired, and combined with the physical attribute data of the pipe segments, the cumulative deviation and degradation rate of physical degradation in each area are calculated. Before calculation, the amplitude of the slow-varying physical signals is denormalized from the standardized dimensionless space to its original dimensional value, and then compared with the physical attribute data of the pipe segments to ensure that the quantitative results of physical degradation have interpretable physical units. Simultaneously, the type of slow-varying anomaly is attributed based on the characteristic patterns of the slow-varying physical signals at each node: when the slow-varying physical signal manifests as a continuous decrease in water pressure and a gradual decrease in flow rate, it is attributed to scaling on the pipe wall leading to a reduction in the effective pipe diameter; when the slow-varying physical signal manifests as a continuous abnormal increase in flow rate and an abnormal decrease in water pressure at downstream nodes, it is attributed to an increase in corrosion and leakage.

[0060] The quantification results of physical degradation are integrated with multi-source data fusion evaluation indicators corrected for group degradation and degradation status information of each sensor group. Based on the cumulative deviation and degradation rate of physical degradation, the reduction in effective transmission and distribution capacity caused by pipeline physical degradation in each region is calculated. For scaling-related degradation, the effective pipe diameter reduction is estimated based on the ratio of the amplitude of the slowly varying physical signal to the original pipe diameter of the pipe section; for leakage-related degradation, the leakage flow loss is estimated based on the ratio of the amplitude of the slowly varying physical signal to the rated flow rate of the pipe section. Based on the degradation deviation estimates and reliability score distribution of each sensor group, the monitoring blind zone range caused by sensor degradation is calculated. The monitoring blind zone is defined as the pipeline area covered by sensors whose group reliability score is lower than the usable threshold.

[0061] The assessment results of the current status of urban reclaimed water utilization are summarized based on multi-source data, including the decline in effective transport and distribution capacity, the scope of monitoring blind zones, multi-source data fusion assessment indicators corrected for population degradation, and the attribution results of physical degradation types. When outputting the comprehensive assessment results of the current status of urban reclaimed water utilization, the confidence levels of each assessment item are also marked: the physical degradation assessment of the area corresponding to the high-confidence slow-changing candidate node is marked as high confidence level; the physical degradation assessment of the area corresponding to the indistinguishable node after inverse correction of degradation model parameters is marked as medium confidence level; and the area within the monitoring blind zone is marked as low confidence level with a note indicating that manual verification is required.

[0062] This implementation uses a group degradation curve model. Active prediction and preliminary compensation for the group-based progressive degradation of sensors in the same batch are performed, allowing sensor degradation bias to be separated from the mixed monitoring data in advance, reducing the degree of direct superposition of two types of slowly varying signals. This is due to the group degradation curve model. Exposure feature vectors of shared environment in groups By jointly fitting the data with the cumulative runtime as input, it can capture the similar degradation trends of sensors in the same batch due to common environmental conditions, thus overcoming the deficiency of being unable to detect group bias when relying solely on cross-source consistency metrics for credibility assessment.

[0063] This implementation uses an anchored benchmark training strategy to train a long-scale spatiotemporal graph autoencoder. Since the anchored benchmark dataset stores data from a period when the pipeline network was in a confirmed good state, and it always participates in the mixed training at a fixed proportion in subsequent retraining, the reconstruction benchmark of the long-scale spatiotemporal graph autoencoder will not gradually adapt to the shift as the physical degradation of the pipeline network slowly changes. This allows the asymptotic deviation of the pipeline network state from the initial good benchmark to be continuously reflected in the long-scale reconstruction error, overcoming the factor that conventional autoencoders cannot detect the shift because all retraining data uses recent data and gradually adapts to the slow shift.

[0064] This implementation classifies nodes into high-confidence slow-varying candidate nodes and indistinguishable nodes by comparing the initial slow-varying signal amplitude with the upper bound of the sensor-compensated residual noise. It then analyzes the correlation of slow-varying signals within the groups to which indistinguishable nodes belong. When the correlation coefficient within the group... Significantly high reverse correction group degradation curve model The parameters. Since the physical degradation of different topological regions of the pipeline network usually does not have the same high correlation pattern as the degradation of sensors in the same batch, the high correlation coefficient within the group can effectively indicate that the slow-varying signal comes from insufficient compensation for sensor group degradation rather than physical degradation of the pipeline network, thus overcoming the factor of attribution uncertainty when the two types of slow-varying signals are mixed.

[0065] This implementation preprocesses the evidence from the same group by removing common bias components before multi-source data fusion, and applies group degradation perception reduction and a confidence cap constraint to the sensor confidence score. Since only independent information components participate in the Dempster-Shafer evidence combination calculation after the common bias components are stripped away, similar bias evidence provided by degraded sensors in the same group no longer reinforces each other during the fusion process. This overcomes the factor that causes false high confidence levels in the fusion results due to mutual confirmation of erroneous readings by degraded sensors. In summary, this implementation method enables the assessment results of the current status of reclaimed water utilization after multi-source data fusion to simultaneously reflect the impact of both pipeline physical degradation and sensor degradation, thereby improving the reliability of the assessment results.

[0066] The following is an example of an application of the present invention, such as... Figure 2-9 As shown, the implementation process is as follows: A city's reclaimed water pipeline system was completed and put into operation in 20XX. The network covers three water use areas: industrial parks, municipal green spaces, and road washing areas. A total of 58 monitoring nodes are deployed throughout the network, each equipped with sensors for water quality (turbidity), water pressure, and flow rate. The network sensors were procured and installed in two batches: Batch A (number BA) was installed in early 20XX, comprising 34 nodes; Batch B (number BB) was installed in early 20XX+1, comprising 24 nodes. After approximately three years of operation, system monitoring personnel discovered a systematic shift in the comprehensive assessment results for some areas, suspecting a combination of sensor degradation and network physical degradation. Therefore, this method was initiated to conduct a dual-source slow-varying signal collaborative decoupling analysis. The following selects six representative nodes (nodes P01 to P06, all belonging to Batch A) as core examples, where P01 to P04 are located in the industrial water supply main pipe, and P05 to P06 are located in the green space branch pipe.

[0067] For step 1, the system extracts the basic information and historical calibration records of batch A sensors from the device metadata database, calculates the common environmental exposure feature vector of the group after Z-score standardization of the environmental parameters of the installation location, and then performs joint fitting on the historical calibration deviation data to obtain the group degradation curve model parameters and the degradation deviation estimate of each sensor.

[0068] The installation environment of batch A (BA) primarily consists of high humidity and moderately corrosive media in industrial parks. After Z-score standardization of temperature, humidity, and media corrosivity indicators, the average values ​​were taken at 34 nodes within the group to obtain the common environmental exposure feature vector for the group. These correspond to the standardized average temperature, average humidity, and average corrosivity, respectively.

[0069] Choosing time degradation basis functions , environment-time coupling function By performing least-squares fitting on all historical calibration deviation data of sensors in batch A, the parameters of the group degradation curve model were obtained: , , , standard deviation of fitting residuals .

[0070] Taking the water pressure sensor (numbered S-P01-WP) at node P01 as an example, its cumulative running time Months, substituted into the group degradation curve model:

[0071]

[0072]

[0073] Uncertainty radius according to Calculation, taking S-P01-WP as an example:

[0074] Table 1 shows the estimation results of degradation bias for node sensors in batch A.

[0075] The dispersion parameter of the individual degradation rate within the group is determined by the standard deviation of the slope obtained by linear regression of the historical calibration deviation of each sensor on the runtime. The dispersion parameter of the degradation rate of batch A group is 0.00083 (unit: standardized deviation / month).

[0076] For step 2, the system extracts historical multi-source monitoring time-series data from the past 36 months, performs Z-score standardization on the three types of data (water quality, water pressure, and flow rate), and then subtracts them point by point using the degradation bias estimate in Table 1 to generate preliminary compensated time-series data. Subsequently, features are extracted at the three time scales and combined with the weighted directed graph structure of the pipeline network to generate spatiotemporal graph data sequences at the three scales.

[0077] Taking pipe segment P01-P02 as an example, the edge weight calculation for the weighted directed graph of the pipeline network is as follows: the pipe segment is 180m long, 200mm in diameter, and made of ductile iron. After normalizing the mean of the length, diameter, and material score (1 for metal pipes, 0 for non-metal pipes) of all pipe segments in batch A, the normalized length of pipe segment P01-P02 is 0.62, the normalized diameter is 0.71, the normalized material score is 1.00, the weighted composite edge weight is 0.74, and a flow direction coefficient of 1.15 is assigned according to the water flow direction (P01 to P02). The final edge weight is:

[0078] Table 2 shows the feature extraction results at three time scales (taking node P01 as an example).

[0079] The long-term drift slope in the table above is positive, indicating that node P01 still has a slow upward trend after the initial compensation, and its source needs to be further analyzed in subsequent steps.

[0080] For step 3, the long-scale spatiotemporal graph data sequence is input into the long-scale spatiotemporal graph autoencoder using the anchored benchmark training strategy. The long-scale reconstruction error time series of each node is calculated. After median filtering and linear regression, the preliminary slowly varying signal is extracted. At the same time, the upper bound of the sensor compensation residual noise of each node is calculated.

[0081] The anchoring benchmark dataset is taken from the pipeline network's good condition data for the first three months after its completion in 20XX (confirmed by comprehensive manual inspection and sensor calibration). In each subsequent retraining iteration, the anchoring benchmark dataset accounts for no less than 30% of the dataset. The long-scale time window is three months, and the sampling steps are... (One sampling step per hour, for a total of 90 days).

[0082] Taking node P03 (flow sensor) as an example, this node only deploys one sensor, S-P03-FL, with an uncertainty radius of 0.0073 (converted to normalized dimensionless space). Therefore, the upper bound for sensor-compensated residual noise is:

[0083] Table 3. Preliminary Slowly Varying Signals and Upper Bounds of Compensated Residual Noise for Each Node

[0084] For step 4, based on the comparison results in Table 3, nodes P01 and P02 are marked as high-confidence slow-changing candidate nodes, and nodes P03, P04, P05, and P06 are marked as indistinguishable nodes.

[0085] For the indistinguishable nodes P03, P04, P05, and P06, all belonging to batch A, the preliminary slowly varying signal sequences of these four nodes are extracted, and the correlation coefficients within the group are calculated. Taking the Pearson correlation coefficients of P03 and P04 (0.921), P03 and P05 (0.887), P03 and P06 (0.894), P04 and P05 (0.903), P04 and P06 (0.896), and P05 and P06 (0.918) as examples, the mean correlation coefficients within the group are as follows:

[0086] This value is higher than the preset threshold of 0.80 (set according to the distribution of batch A sensors in different topological areas of the pipeline network), indicating that the slow-changing signals of indistinguishable nodes are mainly due to insufficient systematic compensation of the group degradation curve model.

[0087] Calculate the mean amplitude of the slowly varying signals P03, P04, P05, and P06:

[0088] The mean is positive (slightly high). The long-scale time window is 3 months long, and the equivalent rate increment is... (Standardized Deviation / Month). Current The preset maximum adjustment percentage is 20%, meaning the maximum adjustment amount is [missing value]. Since the calculated required correction amount of 0.0903 far exceeds the upper limit, this round will be corrected according to the upper limit:

[0089] The remaining corrections are reserved for the next iteration. The degradation bias estimates of each sensor are recalculated using the corrected parameters. Taking S-P03-FL as an example, the corrected degradation bias estimate is approximately 0.0436 (an improvement over the original value of 0.0362, with the direction consistent with the mean).

[0090] For step 5, the original monitoring data is re-standardized and compensated using the corrected degradation bias estimate to generate re-compensated time series data. Long-scale node features are extracted again and input into the long-scale spatiotemporal graph autoencoder to calculate the updated slow-varying signal.

[0091] After the update, the amplitude of the slow-varying signals of indistinguishable nodes P03, P04, P05, and P06 decreased significantly (the residual trend decreased after more thorough compensation), while the slow-varying signals of high-confidence slow-varying candidate nodes P01 and P02 remained basically unchanged (the portion of their amplitude exceeding the upper limit of noise originates from actual physical degradation). The initial slow-varying signals of P01 and P02 were merged with the updated slow-varying signals of P03 to P06, and a spatial consistency constraint was applied: P01 and P02 are topologically adjacent (both located in the industrial water supply main), and their slow-varying signals are in the same direction (both positive, corresponding to a slow increase in water pressure), with amplitude differences... Within the allowable attenuation range of the P01-P02 metal pipe section (3 years old), the signals of both are retained and marked with high confidence. P03 is located downstream of the main pipe and has a topological connection with P01 and P02, but after the update, the amplitude of the slowly varying signal of P03 drops to 0.0412, which is not supported by significant spatial continuity, thus reducing the confidence mark.

[0092] Table 4. Results of slow-varying physical anomalies in the pipeline network after decoupling (restored to original dimensions by denormalization)

[0093] The water pressure at nodes P01 and P02 continued to rise while the flow rate did not increase significantly. Considering the pipe section characteristics (3-year-old pipe, ductile iron material), the cause was attributed to scaling on the pipe wall, which reduced the effective pipe diameter.

[0094] For step 6, the expected degradation deviation of each sensor in batch A is calculated based on the modified group degradation curve model. The basic confidence score is corrected for group degradation perception. After stripping the common deviation components of the evidence in the same group, the Dempster-Shafer evidence combination operation is performed.

[0095] The accuracy tolerance threshold is determined based on the upper limit of the factory-permitted deviation of the water pressure sensor (1% of the measuring range, corresponding to the standardized space value). In this example, we take... Deceleration rate parameter Taking S-P01-WP as an example, the corrected degradation bias estimator If the basic credibility score is 0.88, then:

[0096] Simultaneously check the confidence limit constraint: the degradation rate dispersion parameter for batch A group is 0.00083 (relatively small), the inflection point is set at the 30th month of operation, and the current operation has been running for 26 months, with a confidence limit of [missing information]. .because The final credibility score was 0.335.

[0097] Table 5 shows the revised results of the reliability score for node sensors in batch A.

[0098] For evidence from the same group as batch A, the mean of the projection of each sensor reading into the expected deviation of the corrected group degradation curve model is calculated as the common deviation component (approximately 0.0412 in this example). This common deviation component is then proportionally subtracted from the basic probability assignment function of each sensor and renormalized to obtain the isolated independent information components. Since the proportion of sensors from batch A in the total number of evidences participating in this round of fusion is 68%, exceeding the preset group proportion threshold of 60%, the evidence from batch A is downsampled, and 65% of them are randomly selected to participate in the Dempster-Shafer combination operation. Finally, a multi-source data fusion evaluation index corrected for group degradation is generated (confidence level of 0.71 for comprehensive water quality compliance of reclaimed water in industrial trunk area, and confidence level of 0.58 for sufficient water pressure supply).

[0099] For step 7, the quantitative results of physical degradation are integrated with the fusion evaluation indicators to output the comprehensive evaluation results of the current status of urban reclaimed water utilization.

[0100] The amplitude of the slowly varying water pressure signals at nodes P01 and P02, after denormalization, is approximately +0.031 MPa. Combined with the original pipe diameter (200 mm), the estimated effective pipe diameter reduction is: the scale thickness is approximately...

[0101] The equivalent effective pipe diameter is approximately 178 mm, resulting in a decrease in effective distribution capacity of approximately 20.4%. In Table 5, the sensors with a final confidence score below the usable threshold of 0.30 are S-P03-FL (0.190) and S-P04-FL (0.226), and the corresponding middle section of the industrial trunk line is marked as a monitoring blind zone.

[0102] Table 6 Summary of Comprehensive Assessment Results of Urban Reclaimed Water Utilization Status

[0103] Throughout the implementation process, the data undergoes a clear logical flow from its initial state through each step: Step 1 starts with equipment metadata and historical calibration records, calculates the degradation bias estimate using a group degradation curve model, and provides a benchmark for subsequent compensation; Step 2 uses this estimate to perform preliminary compensation on the original monitoring data and extracts spatiotemporal map features at three time scales; Step 3 reconstructs the long-scale spatiotemporal map using an autoencoder trained on the benchmark and extracts the initial slowly varying signal, while simultaneously calculating the upper bound of the compensation residual noise; Step 4 compares the signal amplitude with the noise upper bound and finds that P03 to P06 are in the indistinguishable range. The high correlation coefficient (0.903) within the group was used to determine insufficient compensation and the degradation model parameters were corrected in reverse. Step 5: the corrected parameters were used to recompensate, and P01 and P02 were identified as the real physical degradation nodes by combining spatial consistency filtering. Step 6: the sensor confidence was reduced based on the corrected model and common deviation components were removed before evidence fusion was performed. Step 7: the physical degradation quantification (pipe diameter reduction of about 11.2 mm), monitoring blind zone (P03-P04 segment) and fusion evaluation index were finally summarized and output. Each item was labeled with a confidence level. The data maintained the uniformity of dimensions and logical consistency throughout the process.

[0104] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for assessing the current status of urban reclaimed water utilization based on multi-source data, characterized in that, Includes the following steps: Obtain sensor device metadata and historical calibration record data. The device metadata includes batch number, installation date, installation location and sensor model. The historical calibration record data includes calibration deviation value and calibration timestamp at each calibration time. Divide the data into groups by batch and fit the group degradation curve model. Calculate the degradation deviation estimate and estimation uncertainty interval for each sensor. The process of grouping sensors by batch and fitting a group degradation curve model to calculate the degradation deviation estimate and uncertainty interval for each sensor includes: dividing sensors into groups based on batch number and installation date; acquiring historical environmental parameter data for all sensor installation locations within each group, including temperature, humidity, and media corrosivity indices; performing Z-score standardization on each environmental parameter and averaging the values ​​across all sensor locations within the group to generate a common environmental exposure feature vector for the group; and using the cumulative runtime and the common environmental exposure feature vector as inputs to jointly fit the historical calibration deviation data of all sensors within the group to generate a group degradation curve model. This group degradation curve model is a parametric degradation model, and its form is as follows: ,in The group's basic degradation rate coefficient. For time-degrading basis functions, This is a vector of environmental impact coefficients. Represents the vector of environmental impact coefficients transpose, This is an environment-time coupling function. The initial deviation constant of the group. The cumulative runtime is obtained by least-squares fitting of each parameter; the cumulative runtime of each sensor is substituted into the group degradation curve model of its respective group to obtain the degradation bias estimate; the radius of the uncertainty interval is estimated. ,in The standard deviation of the fitting residuals of the group degradation curve model over all historical calibration data within the group. is the standard deviation of the historical calibration deviation sequence of sensor i relative to the predicted value of the group degradation mean curve; Long-term historical multi-source monitoring time series data are acquired, and preliminary compensation is performed using the degradation bias estimate. Feature extraction is performed in three time windows: short-scale, medium-scale, and long-scale. The spatiotemporal graph data sequence at three scales is generated by combining the weighted directed graph structure of the pipeline network. The long-scale spatiotemporal map data sequence is input into a long-scale spatiotemporal map autoencoder that adopts an anchored benchmark training strategy. Preliminary slowly varying signals are extracted from the reconstruction error time series, and the upper bound of sensor compensation residual noise is calculated based on the estimated uncertainty interval. The initial slow-varying signal amplitude of each node is compared with the upper bound of the corresponding sensor compensation residual noise. The nodes are classified and marked as high-confidence slow-varying candidate nodes and indistinguishable nodes. When the correlation coefficient of the group to which the indistinguishable node belongs is higher than the preset threshold, the degradation curve model parameters are reversed. The original monitoring data is recompensated using the corrected degradation bias estimate and the updated slow-varying signal is extracted. The decoupled pipeline physical slow-varying anomaly detection results are generated by spatial consistency filtering. Group degradation perception correction is performed on sensor credibility scores, and Dempster-Shafer evidence combination operation is performed on evidence bodies in the same group after common deviation component stripping to generate multi-source data fusion evaluation index. Based on the decoupled network physical slow-change anomaly detection results and the multi-source data fusion evaluation index, the physical degradation quantification results and monitoring blind zone range are calculated, and the comprehensive evaluation results of the current status of urban reclaimed water utilization are output.

2. The method for assessing the current status of urban reclaimed water utilization based on multi-source data according to claim 1, characterized in that, Feature extraction is performed in three time windows: short-scale, medium-scale, and long-scale, including: Within a short-scale time window, the statistical fluctuation of the time series data after preliminary compensation of each node is calculated within the window, and the instantaneous fluctuation characteristics are extracted. Within the mesoscale time window, the dominant frequency component and its corresponding amplitude are obtained by performing Fourier decomposition on intraday data segments as intraday periodic features, and the slope and residual are obtained by performing linear fitting on the daily average series of consecutive days as intraday trend features. Within a long-scale time window, seasonal components are obtained as seasonal periodic features by seasonally decomposing annual data, and long-term drift features are obtained by trend fitting of the residual sequence after removing seasonal components. In the weighted directed graph of the pipeline network, the edge weights are determined by weighted synthesis of pipe segment length, pipe diameter and pipe material attributes after mean normalization based on range, and pipe segment flow direction information is introduced so that downstream nodes receive feature propagation from upstream nodes.

3. The method for assessing the current status of urban reclaimed water utilization based on multi-source data according to claim 1, characterized in that, The execution process of the anchored benchmark training strategy includes: Acquire long-scale spatiotemporal map data of the pipeline network during the period when it is confirmed to be in good condition, and save it as an anchoring reference dataset. The anchoring reference dataset remains unchanged during the system operation cycle. In periodic retraining, the anchored benchmark dataset is mixed with recently collected long-scale spatiotemporal graph data in a fixed ratio as training data, wherein the proportion of the anchored benchmark dataset is not less than the preset lower limit of the total training data. The period in which the pipeline network is confirmed to be in good condition refers to the operational phase after the pipeline network has undergone comprehensive manual inspection and sensor calibration. The deviations of each sensor reading from the standard reference value are all within the factory accuracy range, and the pipeline network has no known physical defects.

4. The method for assessing the current status of urban reclaimed water utilization based on multi-source data according to claim 1, characterized in that, The long-scale spatiotemporal graph autoencoder includes an encoder and a decoder. The encoder extracts the spatial features of each node in the graph structure through a graph convolutional layer, and then extracts the temporal evolution pattern of each node's features along the time dimension through a temporal coding layer, outputting a low-dimensional spatiotemporal latent representation. The decoder takes the low-dimensional spatiotemporal latent representation as input, and recovers the temporal and spatial dimension information sequentially through a temporal decoding layer and a graph deconvolutional layer, outputting a reconstructed spatiotemporal graph data sequence. The long-scale spatiotemporal graph autoencoder uses the mean squared error between the reconstructed output and input, covering all nodes, time steps, and feature dimensions, as the training loss function. The step of extracting the initial slow-varying signal from the reconstruction error time series includes: performing median filtering on the long-scale reconstruction error time series of each node and then performing linear regression to extract the monotonic trend component as the initial slow-varying signal. The calculation method for the upper bound of the sensor compensation residual noise is as follows: sum the squares of the estimated uncertainty interval radii of each sensor at the node, take the square root, and then multiply it by the square root of the number of sampling steps included in the long-scale time window. The estimated uncertainty interval radius is converted to a standardized dimensionless space consistent with the initial slowly varying signal before participating in the calculation.

5. The method for assessing the current status of urban reclaimed water utilization based on multi-source data according to claim 1, characterized in that, The reverse correction of the degradation curve model parameters for the group to which the indistinguishable node belongs includes: Obtain the group number of the group to which each indistinguishable node belongs, and calculate the mean of the Pearson correlation coefficients between each pair of the preliminary slow-varying signal sequences of the corresponding nodes of each sensor within the same group, as the correlation coefficient within the group; When the correlation coefficient within the group is higher than a preset threshold, the mean direction and mean amplitude of the initial slow-changing signal of the indistinguishable nodes within the group are calculated. The equivalent rate increment is obtained by dividing the mean amplitude by the duration of the long-scale time window. This equivalent rate increment is then superimposed onto the group's basic degradation rate coefficient in the mean direction. The parameter increment for each correction does not exceed the preset upper limit of the current parameter value. If the upper limit is exceeded, the correction is made according to the upper limit value. The degradation bias estimates of each sensor were recalculated using the modified group degradation curve model.

6. The method for assessing the current status of urban reclaimed water utilization based on multi-source data according to claim 1, characterized in that, The spatial consistency filtering includes: The initial slow-varying signals of high-confidence slow-varying candidate nodes are combined with the updated slow-varying signals of indistinguishable nodes to obtain the comprehensive result of the slow-varying signals of each node. Based on the topology of the weighted directed graph of the pipeline network, for nodes that are adjacent in the topology and share the same pipe segment, when the slow-changing signals of the two nodes are in the same direction and the amplitude difference is within the reasonable attenuation range determined by the physical properties of the pipe segment, the slow-changing signal is retained; when the slow-changing signal of a node has no spatial continuity support in its topological neighborhood, the confidence flag of the signal is reduced. The reasonable attenuation range is determined by the pipe segment length and pipe segment material, and a pipe segment age weighting factor is introduced to determine the criteria for relaxing the spatial consistency constraint on the nodes connected to pipe segments with longer pipe ages.

7. The method for assessing the current status of urban reclaimed water utilization based on multi-source data according to claim 1, characterized in that, The group degradation perception correction for sensor credibility scores includes: Obtain the basic reliability score for each sensor. When the expected degradation deviation of the sensor group exceeds the accuracy tolerance threshold, multiply the basic reliability score by a reduction factor to obtain the reduced reliability score. The reduction factor is the larger of the ratio of the absolute value of the degradation deviation estimate to the accuracy tolerance threshold to the accuracy tolerance threshold, after being weighted by the reduction rate parameter and subtracted from 1 and 0. A confidence upper limit constraint is applied to the sensors in the same group, which is monotonically decreasing with the cumulative running time. The confidence upper limit constraint is a monotonically decreasing piecewise linear function with respect to the cumulative running time, and its inflection point position and slope are determined according to the dispersion parameter of the individual degradation rate within the group. The process of stripping the common deviation component is as follows: obtain the basic probability assignment function corresponding to each sensor in the same group, calculate the mean value of the projection component of each basic probability assignment function in the expected deviation direction as the common deviation component, and then perform normalization processing after subtracting the common deviation component from each basic probability assignment function. When the proportion of evidence from the same degenerate group in the combined evidence exceeds a preset group proportion threshold, the evidence from that group is downsampled.

8. The method for assessing the current status of urban reclaimed water utilization based on multi-source data according to claim 1, characterized in that, The calculation of physical degradation quantification results and monitoring blind zone range outputs a comprehensive assessment result of the current status of urban reclaimed water utilization, including: After the amplitude of the slowly varying physical signal is restored to its original dimensional value through inverse normalization transformation, the cumulative deviation of physical degradation and degradation rate of each region are calculated in combination with the physical attribute data of the pipe section. Based on the characteristic patterns of the slow-changing physical signals, the causes of slow-changing anomalies are attributed. When the water pressure continues to drop and the flow rate gradually decreases, it is attributed to the shrinkage of the effective pipe diameter caused by scaling on the pipe wall. When the flow rate continues to increase abnormally and the water pressure at the downstream node decreases abnormally, it is attributed to the increase in corrosion leakage. Based on the degradation bias estimate and confidence score distribution of each sensor group, the pipeline area covered by sensors whose group confidence score is lower than the usable threshold is identified as the monitoring blind zone. Each assessment item is labeled with a confidence level. The area corresponding to the candidate node with high confidence and slow change is labeled with a high confidence level. The area corresponding to the node that cannot be distinguished after the degradation model parameters are reverse corrected is labeled with a medium confidence level. The area within the monitoring blind zone is labeled with a low confidence level.

9. A system for assessing the current status of urban reclaimed water utilization based on multi-source data, used to execute the method for assessing the current status of urban reclaimed water utilization based on multi-source data as described in any one of claims 1 to 8, characterized in that, include: The group degradation modeling module is used to acquire sensor device metadata and historical calibration record data, divide the data into groups by batch and fit the group degradation curve model, and calculate the degradation deviation estimate and estimation uncertainty interval for each sensor. The multi-scale spatiotemporal map construction module is used to acquire long-term historical multi-source monitoring time series data, perform preliminary compensation using degradation bias estimates, extract features at three time scales, and generate spatiotemporal map data sequences at three scales by combining the pipeline network weighted directed graph structure. The slow-varying signal extraction module is used to input the long-scale spatiotemporal map data sequence into the long-scale spatiotemporal map autoencoder which adopts the anchored benchmark training strategy, extract the preliminary slow-varying signal from the reconstruction error time series, and calculate the upper bound of the sensor compensation residual noise. The node classification and degradation correction module is used to compare the initial slow-varying signal amplitude of each node with the upper bound of the sensor compensation residual noise, classify and label the nodes, and reverse the degradation curve model parameters of the group to which the indistinguishable nodes belong. The physical slow-varying anomaly detection module is used to recompensate the original monitoring data with the corrected degradation bias estimate and extract the updated slow-varying signal. After spatial consistency filtering, it generates the decoupled pipeline physical slow-varying anomaly detection results. The multi-source data fusion module is used to correct the group degradation perception of sensor credibility scores, and to perform Dempster-Shafer evidence combination operation after stripping common deviation components from evidence in the same group to generate multi-source data fusion evaluation indicators. The comprehensive evaluation output module is used to calculate the physical degradation quantification results and monitoring blind zone range based on the decoupled pipeline network physical slow-change anomaly detection results and multi-source data fusion evaluation indicators, and output the comprehensive evaluation results of the current status of urban reclaimed water utilization.

Images