Edge side water supply system scheduling and control method and device

Abstract

The application discloses an edge side water supply system scheduling and control method and device, acquires pipe network pressure data and water flow data, generates load deviation characteristics through cloud edge load deviation identification, generates linkage fault groups according to the load deviation characteristics, generates pump group operation deviation amount by combining pump group cut-in frequency determination and light load reference working condition verification, implements cross-working condition reverse tracing and sudden drop blind area correction on the linkage fault groups to generate fine selection fault groups and lock the leading fault items, completes cloud edge source identification and cross-shift responsibility deflection verification based on the leading fault items and the pump group operation deviation amount, generates a health diagnosis report through root cause tracing, and finally combines the cross-working condition recurrence rate to perform inverse proportion weight superposition and toughness margin evaluation, and generates scheduling control instructions according to the toughness margin attenuation rate, so that the accuracy and stability of cloud edge collaborative scheduling of the water supply system are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of water supply system automation control technology, and in particular to a method and apparatus for scheduling and controlling a water supply system on the edge side. Background Technology

[0002] The operation of urban water supply networks is affected by multiple factors, including fluctuations in water demand, aging of the network, and degradation of pump performance. The coordination deviation between the cloud-based dispatch system and the edge-side execution equipment has been difficult to distinguish effectively for a long time. Existing monitoring methods usually only record the overall operational deviation but cannot identify whether the deviation is due to improper cloud instructions or deterioration of edge equipment, resulting in unclear fault determination and reliance on manual experience for maintenance resource allocation.

[0003] Meanwhile, switching behaviors vary significantly under different operating conditions, and a systematic reference for light-load conditions has long been lacking. The transfer of responsibility between shifts due to the division of dispatch authority is difficult for existing systems to capture. There is a lack of linkage between pipeline health assessment and pump maintenance scheduling, and the dynamic decay of resilience margins is not reflected in the dispatch control strategy in a timely manner. The overall management and control system has significant shortcomings in proactive early warning and precise dispatch. Summary of the Invention

[0004] This invention discloses a scheduling and control method and device for an edge-side water supply system. It aims to incorporate pipeline pressure and water flow data into a unified cloud-edge load deviation identification framework, locate the root cause of linkage faults through degradation separation and switching behavior analysis, and achieve accurate fault attribution by combining cross-condition recurrence rate correction and root cause tracing mechanism. Furthermore, it links health diagnosis results with resilience margin assessment to output scheduling and control commands, forming a complete link from deviation perception to closed-loop regulation.

[0005] The first aspect of this invention proposes a method for scheduling and controlling an edge-side water supply system, comprising the following steps: Collect pipeline pressure data and water flow data, and perform cloud-edge load deviation identification on the pipeline pressure data and water flow data to generate load deviation characteristics; Based on the load deviation characteristics, cloud-edge degradation separation is performed to generate a linkage fault group. The pump switching frequency of the linkage fault group is measured to generate a switching frequency spectrum. The switching frequency spectrum is then verified under light load reference conditions to generate the pump operation deviation. Cross-condition fault reverse tracing is performed on the linked fault group to generate cross-condition reproducibility rate. Based on the cross-condition reproducibility rate, the linked fault group is subjected to sudden drop blind zone pre-correction to generate selected fault group. Low reproducibility isolated items of the selected fault group are reverse locked in reverse order to determine the dominant fault item. Based on the dominant fault item and the pump set operation deviation, cloud-edge source identification is performed to determine the source of control responsibility. Based on the source of control responsibility, cross-shift responsibility deviation verification is performed to determine the cross-shift responsibility inversion item. Based on the cross-shift responsibility inversion item, root cause tracing is performed to generate a health diagnosis report. Based on the health diagnosis report and the cross-operating condition recurrence rate, an inverse weighted superposition is performed to generate a pump group maintenance scheduling queue. The pump group maintenance scheduling queue and the health diagnosis report are used to evaluate the resilience margin and generate a resilience margin attenuation rate. Based on the resilience margin attenuation rate, a high attenuation rate reverse frequency reduction linkage regulation is performed to generate a scheduling control command.

[0006] A second aspect of the present invention provides a scheduling and control device for an edge-side water supply system, comprising: The deviation identification module is used to collect pipeline pressure data and water flow data, and perform cloud-edge load deviation identification on the pipeline pressure data and the water flow data to generate load deviation characteristics. The decoupling verification module is used to separate cloud-edge degradation based on the load deviation characteristics to generate a linkage fault group, measure the pump switching frequency of the linkage fault group to generate a switching frequency spectrum, and verify the switching frequency spectrum under light load reference conditions to generate the pump operation deviation. The correction and screening module is used to perform cross-operating condition fault reverse tracing to generate cross-operating condition reproducibility rate for the linkage fault group, perform sudden drop blind zone pre-correction on the linkage fault group based on the cross-operating condition reproducibility rate to generate selected fault group, and implement low reproducibility rate working condition isolated item reverse locking to determine the dominant fault item in the selected fault group. The responsibility diagnosis module is used to identify the source of control responsibility based on the dominant fault item and the pump set operation deviation, to conduct cross-shift responsibility deviation verification based on the source of control responsibility to determine the cross-shift responsibility inversion item, and to generate a health diagnosis report based on the root cause tracing based on the cross-shift responsibility inversion item. The instruction generation module is used to generate a pump group maintenance scheduling queue by inversely weighting the health diagnosis report and the cross-operating condition recurrence rate, evaluate the resilience margin of the pump group maintenance scheduling queue and the health diagnosis report to generate a resilience margin attenuation rate, and generate scheduling control instructions by performing high attenuation rate reverse frequency reduction linkage regulation based on the resilience margin attenuation rate.

[0007] The beneficial effects of this invention are reflected in the following points: First, attributing the source of load deviation is a core challenge that cloud-edge collaborative systems have long failed to effectively address. By using the convergence depth of the load stagnation interval and the direction of the source as joint judgment criteria, the overall deviation is decomposed into two independent causes: cloud command malfunction and edge device degradation. Based on this, the degradation separation of the linkage fault group and the light load reference verification of the pump group switching behavior are completed, enabling fault identification to have a distinguishing ability at the source attribution level that traditional overall deviation monitoring lacks. Second, the combination of pre-correction of sudden drop blind zone and reverse locking of isolated items in low recurrence rate conditions is the key difference between this invention and existing methods at the fault location level. The continuous suppression of some fault items by high-severity conditions results in a systematically low triggering frequency in historical records. The discontinuous sudden increase in recurrence rate at the condition switching node is a direct signal of the fading suppression effect. Based on this, the omission error caused by the masking effect is included in the correction range, enabling the fault location results to actively identify scenarios of alternating conditions. Finally, traditional root cause analysis stops at the equipment level, and the phenomenon of responsibility transfer at the work team jurisdiction boundary has long been outside the assessment system. This case extends responsibility attribution to the work team jurisdiction level by checking cross-work team responsibility deviation. It locates the root cause by combining recurrence trend scanning and reverse anomaly verification of reported lead time quantities, and includes low responsibility deviation units in the health diagnosis report through reverse weighting. The health diagnosis results and cross-operating condition recurrence rate are inversely weighted to form a maintenance scheduling queue. Based on the resilience margin decay rate, the high decay rate pump group frequency reduction protection and load redistribution are driven, so that the health assessment results are directly transformed into an executable scheduling control strategy. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a method for scheduling and controlling an edge-side water supply system according to the present invention.

[0009] Figure 2 This is a structural block diagram of a scheduling and control device for an edge-side water supply system according to the present invention. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0012] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0013] The technical solutions of the embodiments of this application will be described below.

[0014] like Figure 1 As shown, this embodiment of the invention provides a method for scheduling and controlling an edge-side water supply system, including the following steps S101-S105: Step S101: Collect pipeline pressure data and water flow data, and perform cloud-edge load deviation identification on the pipeline pressure data and water flow data to generate load deviation characteristics.

[0015] Specifically, pipeline pressure data and water flow data are collected. Pipeline pressure data is continuously reported by pressure sensing units deployed at various monitoring nodes in the water supply network. Data acquisition is triggered in two ways: fixed-period acquisition and differential pressure over-limit triggering. Fixed-period acquisition maintains the integrity of the time-series records, while differential pressure over-limit triggering automatically increases the acquisition frequency to capture rapid fluctuations when the pressure change at a node exceeds a set range. Both methods operate in parallel, with over-limit triggering prioritizing coverage of the fixed-period data to ensure that the pipeline pressure data density during critical periods meets the requirements for deviation identification. Water flow data is synchronously collected by flowmeter units at the main pipe and branch pipe nodes. The acquisition time axis is aligned with the pipeline pressure data. Alignment is achieved through an edge-side timestamp synchronization mechanism. When the synchronization deviation exceeds half a acquisition cycle, a time-series offset label is added to the corresponding water flow data record to prevent pressure-flow time-series misalignment from interfering with subsequent load analysis. After the pipeline pressure data is collected, the zero drift deviation is corrected point by point according to the calibration parameters of each pressure sensing unit. The zero drift correction brings the pipeline pressure data value back to the comparable benchmark. When the zero drift correction exceeds the allowable range of the calibration parameters, the data of the corresponding node time period is automatically marked with a low confidence label. After the correction is completed, the zero drift offset of each sensing unit and the corrected reading are archived together for use as a reference for historical offset trends when the edge side parameter library is periodically recalibrated. Instantaneous pulse readings caused by pipeline valve switching or temporary equipment access in the water flow data are judged as abnormal acquisition points according to the flow continuity constraint. Abnormal acquisition points are replaced by linear interpolation of adjacent valid times. The interpolation operation is only enabled when the consecutive abnormalities do not exceed two acquisition cycles. When the limit is exceeded, the water flow data of the corresponding node for that time period is marked with a low confidence label. This label, along with the time series offset label, belongs to the quality label and is used as the basis for weight reduction.

[0016] In some embodiments, the step of identifying cloud-edge load deviations and generating load deviation features by comparing the pipeline pressure data and the water flow data includes: performing measured load analysis on the pipeline pressure data and the water flow data to generate an edge load sequence; performing cloud-edge double-sided load deviation location on the edge load sequence according to a scheduling window to generate a cloud-edge deviation sequence; identifying residual abnormal convergence periods for the cloud-edge deviation sequence to determine load stagnation intervals; and analyzing the convergence depth and following the source direction based on the load stagnation intervals to generate load deviation features.

[0017] The measured load analysis of pipeline pressure data and water flow data generates an edge load sequence. The measured load analysis is performed independently at the node level. For each node, pipeline pressure data and water flow data are paired one-to-one according to the acquisition time on the aligned time axis. After pairing, they are grouped by node identifier to form pressure-flow pairing sets. Each node's pressure-flow pairing set independently enters the subsequent hydraulic equation solution process. After solution, the data is summarized according to node time sequence. The node readings of pipeline pressure data and water flow data are paired moment-by-moment on the aligned time axis. After pairing, they are grouped by node identifier to form pressure-flow pairing sets. Each pair of readings in the pressure-flow pairing set is converted into the measured load value of that node at that moment using the pipeline hydraulic equation. The core form of the hydraulic equation is... Where L is the measured load value, Q is the flow meter reading, ΔP is the node pressure difference value, and C is the comprehensive load factor. C is determined by the pipe section material, pipe diameter, and local resistance coefficient. The C value for different pipe sections is entered into the parameter library through field calibration and plays a role in dimension harmonization. The measured load values ​​are arranged according to the time sequence of each node to form the load time sequence of each node. The load time sequences of each node are summarized to form the edge load sequence. This sequence is organized with the node identifier as the row and the alignment time as the column. The spatial distribution of the measured load values ​​of each node at the same time describes the overall load pattern of the current pipeline network. The pressure-flow pairing sets carry pairing items with time sequence offset labels or low confidence labels. Their solution results are synchronously labeled with corresponding quality labels in the edge load sequence. The quality labels serve as the confidence reference for the load value at that time in the subsequent cloud edge deviation positioning. The edge load sequence time with low confidence labels is reduced in weight in the deviation positioning comparison.

[0018] The edge load sequence is used to perform cloud-edge double-sided load deviation positioning according to the scheduling time window to generate a cloud-edge deviation sequence. The length of the scheduling time window is read according to the configuration parameters of the cloud scheduling system. The start and end times of each time window are aligned and divided on the edge load sequence. The division boundary corresponds to the instruction switching time recorded by the cloud scheduling. Within the same time window, the measured load value on the edge side and the preset target value on the cloud constitute the paired calculation object for deviation positioning, and the calculation is carried out step by step. The edge load sequence is divided into several time window segments according to the scheduling time window. Each time window segment contains the measured load values ​​of the edge side at all times within that time window. These values ​​are compared with the preset targets in the scheduling records issued by the cloud on a time-by-time basis. The deviation value δ(t) = L_edge(t) - L_cloud(t) is calculated, where L_edge(t) is the measured load value of the edge load sequence at time t, and L_cloud(t) is the preset load target of the cloud at time t. A positive δ(t) indicates that the actual load on the edge side is higher than the cloud target, and a negative δ(t) indicates that it is lower than the target. The time-by-time sequence of δ(t) depicts the deviation trajectory of the perceived load on both sides of the cloud and the edge within that time window. The δ(t) values ​​at each time within a time window are arranged in time sequence to form the deviation time sequence of that time window. The deviation time sequences of all time windows are summarized to form the cloud-edge deviation sequence. The difference in the deviation amplitude between different time windows corresponds to the change in cloud-edge coordination quality within different scheduling cycles. A time window with a consistently high amplitude indicates a systematic mismatch between the cloud target and the actual water supply demand on the edge side within that scheduling cycle. The δ(t) corresponding to the time carrying the quality label in the edge load sequence is synchronously inherited in the cloud edge deviation sequence. The label transfer ensures that the data confidence information is not lost in the cloud edge deviation sequence.

[0019] To identify load stagnation intervals, residual abnormal convergence periods are identified for cloud edge deviation sequences. The cloud edge deviation sequence is scanned point-by-point at the time-to-time level, following a three-layer judgment: the first layer calculates the sliding residual for each time-to-time deviation value; the second layer compares the absolute value of the residual with a convergence threshold; and the third layer counts whether the number of consecutive times meeting the conditions exceeds the minimum duration threshold. When all three conditions are met, the continuous period is identified as a load stagnation interval. The identification of residual abnormal convergence periods uses the deviation value at each time point in the cloud edge deviation sequence as the detection object. The sliding residual r(t) = δ(t) - δ_w is calculated for each time point of the cloud edge deviation sequence, where δ_w is the average deviation within the sliding window centered at time t. The sliding window width is set to half the length of the scheduling window. The absolute value of r(t) characterizes the degree of deviation of the deviation from the local mean at that time point. |r(t)| is considered stable if it is lower than the threshold ε calibrated by the low quantile of the residual in the non-stagnation segment. High stability and a sustained δ(t) above the upper quantile of the mean in the non-stagnation segment indicate a loss of tracking ability during that period. When the number of consecutive moments satisfying |r(t)|<ε exceeds the minimum duration threshold, this consecutive period is determined to be a load stagnation interval. The minimum duration threshold is set to 2 to 4 times the length of the scheduling window. Excessively short consecutive low residual segments are considered occasional low fluctuations rather than genuine load stagnation phenomena. The r(t) corresponding to moments carrying low-confidence labels in the cloud-edge deviation sequence is weighted 0.5 times for participation in the convergence threshold comparison to prevent data quality issues from affecting the accurate location of the load stagnation interval boundary. The load stagnation interval is described by three elements: the start time, the end time, and the mean interval deviation δ_stall. δ_stall quantifies the stable height of the deviation between the cloud edges when stagnation occurs, serving as a direct input for subsequent convergence depth analysis.

[0020] The analysis focuses on the convergence depth and load deviation characteristics generated based on the load stagnation intervals. The analysis is performed independently on each load stagnation interval. For each interval, the convergence depth D is first calculated. Then, the edge load sequence within the interval is linearly fitted to obtain the local slope, which is compared with the direction of change of the target in the cloud to determine the type of direction of the target. These two indicators, along with the start and end times of the interval and the node identifier, are packaged together as the load deviation characteristic entries corresponding to that interval. After all intervals are processed, the results are summarized and written into the load deviation characteristic set. The convergence depth D is determined by the formula D=|δ_stall-δ_ref| / σ_δ, where δ_stall is the mean deviation within the load stagnation interval, δ_ref is the mean of the δ(t) sequence during the non-stagnation period, and σ_δ is the standard deviation of the cloud edge deviation sequence during the non-stagnation period. D measures the distance between the deviation level during the stagnation period and the normal convergence reference as a multiple of the standard deviation. The deeper the D value, the more severe the deviation during load stagnation. The determination of the source direction is based on the slope of the linear fit of the edge load sequence with respect to time within the load stagnation interval. When the local slope direction is consistent with the direction of change of the cloud target, it is determined as a limited following direction, indicating that although the edge side does not completely follow, it still retains a certain directional response capability. When the local slope direction is opposite to the direction of change of the cloud target, it is determined as reverse decoupling. When the absolute value of the slope approaches zero, it is determined as passivated decoupling, indicating that the measured load on the edge side within this interval is mainly driven by local factors rather than cloud scheduling. The load deviation characteristics consist of four elements: the time range of the load stagnation interval, the convergence depth D value, the result of the source direction determination, and the node identification. The load deviation characteristics of multiple nodes are summarized to cover the deviation distribution pattern of the entire pipeline network edge side. Nodes with high convergence depth and decoupling are assigned to the edge source group, nodes with limited following and high D value are assigned to the cloud source group, and other nodes with limited following are assigned to the mixed source group.

[0021] Step S102: Based on the load deviation characteristics, cloud edge degradation separation is performed to generate a linkage fault group. The pump switching frequency of the linkage fault group is measured to generate a switching frequency spectrum. The switching frequency spectrum is verified by light load reference working condition to generate the pump operation deviation.

[0022] In some embodiments, the step of generating a linked fault group by separating cloud-edge degradation based on the load deviation characteristics includes: generating a source grouping table by grouping the load deviation characteristics into cloud-edge dual-time window sources; calculating the edge anomaly rate during the normal cloud period and the cloud anomaly rate during the normal edge period based on the source grouping table to form a degradation direction analysis table; identifying fault pairs with no edge anomalies but continuous cloud-edge collaborative failures based on the degradation direction analysis table to form a degradation isolation group; and selecting the pre-triggered items for the last stable deterioration based on the degradation isolation group to generate a linked fault group.

[0023] Load deviation characteristics are grouped into a source grouping table using a cloud-edge dual-time window approach. Decoupled records in the load deviation characteristics are assigned to the edge source group, records with limited following and high convergence depth D are assigned to the cloud source group, and other records with limited following are labeled as mixed sources. Mixed sources are listed separately in a third group with a weight of 0.5 and participate independently in anomaly rate statistics to avoid feature distortion and misalignment caused by forced unilateral classification. The three groups correspond to three typical causes: edge-independent degradation, cloud-cooperative failure, and mixed-driven factors. The grouping operation is performed on all load deviation characteristic records one by one, using the scheduling time window as the smallest granularity. The source attribution of the same node within different scheduling time windows is determined independently. The source grouping table is indexed by both node identifier and scheduling time window, and each entry records the source attribution category and the corresponding convergence depth D value. For entries where the source attribution of the same node remains on the same side across multiple consecutive scheduling windows, the weight is normally included in the degradation direction analysis. Entries from nodes with frequently switching source attribution are weighted down to 0.5 times to avoid distorting the anomaly rate of unstable attribution samples. This weighting ensures that subsequent anomaly rate indicators are primarily driven by stable attribution nodes, while the disturbance effect of transitional nodes is appropriately suppressed. During scheduling shift changes, the cloud command issuance rhythm is briefly adjusted, and some nodes experience a single source attribution switch within the shift change window. The source grouping table adds a time-isolated annotation to these single-switch entries. Time-isolated annotated entries are weighted down to 0.5 times in the anomaly rate statistics to exclude shift change disturbances that artificially inflate the anomaly rate reading of specific nodes. Shift change periods typically last no more than two scheduling windows. Nodes that continue to switch source attribution outside this range are not subject to time-isolated annotation and are treated as normal attribution records.

[0024] A degradation direction analysis table is generated by calculating the edge anomaly rate during the normal cloud period and the cloud anomaly rate during the normal edge period from the source grouping table. The edge anomaly rate during the normal cloud period is calculated using the subset of entries from the source grouping table that are classified as normal for that time window as having an edge source as an anomaly. The cloud anomaly rate during the normal edge period is calculated symmetrically according to the source grouping table. The difference between the two rates, calculated using the same criteria when conditions are normal on the opposite side, is recorded in the degradation direction analysis table as ΔR = R_edge - R_cloud. A positive ΔR indicates a stronger tendency for independent degradation on the edge side, while a negative ΔR indicates a stronger tendency for collaborative failure in the cloud. The larger the absolute value of ΔR, the more singular and definite the source of degradation. Entries with time-series isolated labels are weighted down to 0.5 times when included in the anomaly rate statistics to avoid planned short-term fluctuations that could inflate the anomaly rate readings of individual nodes. Nodes with both R_edge and R_cloud being high but with a low absolute value of ΔR are marked as having simultaneous degradation on both sides. The sources of degradation for these nodes are mixed, and they are treated differently from single-sided dominant nodes in the subsequent degradation isolation group screening. Nodes with simultaneous degradation on both sides are separately included in the pending verification list and processed uniformly in the S104 responsibility deflection verification stage. This step does not involve fault pair pairing to avoid distorting the time window overlap calculation of single-sided dominant fault pairs by nodes with mixed sources. Trunk master nodes simultaneously receive cloud commands and edge local voltage regulation actions, and the situation of both rates being high is more common than for end nodes. The interpretation of the rate difference for such nodes in the degradation direction analysis table needs to be differentiated in conjunction with the node topology location. The ΔR threshold for trunk master nodes is relaxed by 1.2 times the threshold for end nodes to exclude interference from topology factors.

[0025] Based on the degradation direction analysis table, fault pairs with no edge anomalies but persistent cloud-based collaborative failures are identified and isolated. Nodes with low edge anomaly rates and negative absolute values ​​exceeding the 75th percentile of ΔR in the degradation direction analysis table are included in the verification. Candidate nodes with consecutive anomalies exceeding three scheduling time windows in the degradation direction analysis table are confirmed as nodes with persistent cloud-based collaborative failures. Single isolated cloud-based anomalies do not meet the persistence condition. The time window overlap O(i,j) = |W_i∩W_j| / |W_i∪W_j| is calculated for each pair of nodes with persistent cloud-based collaborative failures, where W_i and W_j are the time window sets for node i and node j, respectively, where W_i and W_j are the time windows for persistent cloud-based failures. Node pairs with O(i,j) exceeding the pairing threshold are confirmed as degradation-isolated fault pairs. Pairing thresholds are set according to the pipeline topology hierarchy. The pairing threshold for adjacent nodes within the same pressure zone is higher than that for nodes across pressure zones, to prevent nodes with large topological distances from being mistakenly identified as strongly correlated fault pairs due to coincidental scheduling timing. The pairing threshold for cross-pressure zones is usually set to 1.3 times that of the pairing threshold within the same pressure zone, giving cross-zone pairing a higher requirement for time window synchronization. Each fault pair in the degradation isolation group is accompanied by three attributes: time window overlap O(i,j), their respective cloud-based anomaly rate, and the length of the persistent failure interval. Fault pairs with a persistent failure interval length exceeding five scheduling time windows are marked with a high persistence label, which serves as the basis for weight upgrade in the P4 final degradation identification. The degradation isolation group is sorted from high to low according to the time window overlap O(i,j) value. Fault pairs ranked higher are given priority to enter the final stable degradation pre-trigger item screening stage, while fault pairs ranked lower are easily covered by the sudden drop blind zone pre-correction mechanism during the selected fault group screening stage.

[0026] Based on the degradation isolation group, the preceding triggers for the last stable degradation are selected to generate linked fault groups. The starting time window of the cloud-based persistent failure interval for each fault pair in the degradation isolation group is used as the backtracking anchor point. The backtracking proceeds chronologically to the end of the continuous normal time window sequence. The node record in the degradation direction analysis table within the last normal time window where the ΔR deviates most from the mean of the normal time window segment is selected as a candidate preceding trigger. The candidate is the single record with the largest deviation, quantified by the absolute value of the difference between the current time window's ΔR and the mean ΔR of the normal segment. When the same fault pair in the degradation isolation group deteriorates multiple times, only the preceding trigger for the last degradation interval is used as the formal linked fault group entry. The preceding triggers for earlier degradation intervals are retained as historical triggers in an auxiliary field, only used as supplementary candidates for backtracking when the number of dominant fault items is insufficient. Preceding triggers with a convergence depth D value exceeding a high-depth threshold are annotated with a high-induction intensity in the linked fault group. This annotation, along with the span of the corresponding fault pair's cloud-based persistent failure interval, describes the overall severity of the linked event. Pre-triggered items triggered by external water use mode switching usually see their deviations quickly disappear after the switching is completed, and the ΔR within the last normal time window shows a brief symmetrical fluctuation before returning to the baseline. Pre-triggered items triggered by pump unit deterioration, on the other hand, show a monotonically high ΔR within the last normal time window, and the deviation continues to accumulate without returning when the external operating conditions do not change. The difference in the shape of ΔR within the last normal time window between the two types of triggering root causes is the physical basis for identifying the triggering mechanism by the source category field. This field is used as the basis for weighting in the ranking of dominant fault items, with the weight of the unit deterioration item being higher than that of the externally induced item.

[0027] For the linked fault group, the pump group switching frequency is measured to generate a switching frequency spectrum. Pump group operating status signals are collected synchronously; a change from 0 to 1 is recorded as one activation action, and a change from 1 to 0 is recorded as one deactivation action. The switching frequency spectrum is organized as "pump group × time window," with each cell filled with the total number of switching actions (in and out) within that time window. The absolute value of the difference between the number of in and out actions indicates whether the pump group's activation and deactivation actions are balanced within that time window; a large difference indicates an abnormal pattern of unidirectional continuous switching. The switching frequency parameter values ​​of each fault-related pump group in the linked fault group are arranged according to all pump groups and all time windows, forming a switching frequency spectrum with pump group identifiers as rows and scheduling time windows as columns. Under the same operating condition type, the mean distribution of each column provides the normal driving benchmark for pump group switching behavior under that condition. Column values ​​exceeding twice the standard deviation of the mean are marked as abnormal switching time windows. The frequency distribution of each pump group in the switching frequency spectrum also records the time distribution of switching actions. A small variance in the switching time within a time window corresponds to a concentrated burst of switching actions, while a large variance corresponds to a uniform dispersion. The mixed occurrence of these two distribution patterns indicates that the pump group is driven by the superposition of multiple trigger signals within that time window. The mixed pattern is recorded in the switching frequency spectrum by variance interval labeling and multi-source drive labeling. These two labels are used to distinguish the judgment window in P9 occlusion switching identification. Abnormal switching time windows in the switching frequency spectrum are classified and stored in the database according to the concentration coefficient. The concentration coefficient is the ratio of the switching time variance to the time window length. When the ratio is less than 0.2, it is stored as a high concentration abnormal time window; when it is between 0.2 and 0.5, it is a medium concentration abnormal time window; and when it is greater than 0.5, it is a low concentration abnormal time window. High concentration abnormal time windows trigger a single time window high-sensitivity judgment window, low concentration abnormal time windows trigger a multi-time window tolerance judgment window, and medium concentration abnormal time windows use the default judgment window. The three types of time windows correspond to different occlusion switching identification sensitivity configurations.

[0028] In some embodiments, the step of generating pump unit operating deviation by performing light load reference condition verification on the switching frequency spectrum includes: screening light load switching parameters from the switching frequency spectrum to construct a light load reference set; calculating the switching difference between the light load reference set and the switching frequency spectrum to form a switching deviation sequence; identifying switching items maintained by conventional noise filtering threshold micro-switching reverse amplification on the switching deviation sequence to generate a deviation locking item group; and determining the pump unit operating deviation by performing light load blanking condition reverse weighted verification based on the deviation locking item group.

[0029] Light load reference benchmarks are constructed by screening light load switching parameters from the switching frequency spectrum. Time windows in the switching frequency spectrum where the ratio of actual network load to rated water supply capacity is lower than the light load threshold are extracted as light load subsets. The light load threshold is set according to the network capacity parameters of each water supply zone; daytime off-peak hours and nighttime maintenance periods typically meet the light load conditions. The switching frequency parameter values ​​of each pump group in the light load subset of the switching frequency spectrum are summarized, and the switching mean f_light(i) and standard deviation are calculated. The mean and standard deviation together constitute the light load reference benchmark entries for that pump group. Pump groups with larger standard deviations have their conventional noise filtering thresholds relaxed in subsequent difference calculations to prevent large fluctuations under normal operating conditions from being over-identified as deviation lock-in items. Pump group entries with fewer than three light load time windows are considered insufficient samples and are supplemented by data from the next light load time window, weighted inversely proportional to the distance to the light load threshold. The supplemented results are annotated with low-precision labels, which are treated as 0.8 times the conservative threshold in subsequent difference calculations. Insufficient number of light-load time windows is common in water supply zones during the network expansion and renovation phase. During this phase, the actual operating load of the pipeline network is consistently high, and there are few time windows that meet the light-load conditions in the historical records. The weighted distance parameter for the supplementary calculation of the secondary light-load is appropriately relaxed in such zones to ensure sample coverage. The sample size field is carried along with the light-load reference benchmark set entries. Entries with high sample sizes have a greater reference weight in the subsequent switching difference calculation. The sample size differences of different pump groups are uniformly incorporated into the low-precision labeling system, and entries with low sample sizes are downweighted in the summary statistics to ensure that the overall accuracy of the light-load reference benchmark set is dominated by entries with high sample sizes.

[0030] For example, the step of calculating the switching difference between the light-load reference set and the switching frequency spectrum to form a switching deviation sequence includes: reading the switching frequency parameter values ​​for each operating condition from the switching frequency spectrum; performing item-by-item difference calculations between the switching frequency parameter values ​​and the light-load reference set to form a difference time series; identifying switching items that abruptly return to zero after the difference time series is subjected to difference periodic decay to generate a masked switching record group; and determining the associated deviation type of the difference time series based on the masked switching record group to generate a switching deviation sequence.

[0031] The switching frequency spectrum is used to read the switching frequency parameter values ​​for each operating condition. Operating conditions are categorized into three types: light load, normal load, and peak load. The categorization criteria are consistent with the operating condition definition used when constructing the light load reference set. Each column in the switching frequency spectrum includes an operating condition type label. For time windows with missing operating condition type labels, automatic labeling is performed based on the measured average load and the operating condition classification threshold. Time windows that continuously fluctuate around the operating condition type threshold are prone to repeated jumps between the two categories during automatic labeling. For such time windows, the average load of the three nearest time windows is smoothed before determining the operating condition type, ensuring stable labeling and avoiding repeated triggering of labeling due to single-point fluctuations. Time window columns that cannot be automatically labeled are left empty to maintain temporal continuity; these empty spaces are skipped in subsequent difference calculation stages. During periods of frequent operating condition switching, the switching frequency parameter values ​​of adjacent time windows may show a step change. Boundary time windows additionally record the operating condition type labels before and after the switch. During the difference calculation stage, this is used to distinguish between parameter value abrupt changes caused by operating condition switching and abnormal switching behavior of the pump unit itself. After the switching frequency parameter values ​​of each pump group in the switching frequency spectrum are read, they are grouped and archived according to the operating condition type. The distribution range of parameter values ​​within each operating condition group serves as the reference interval for switching behavior under the corresponding operating condition in the switching frequency spectrum. Readings outside the distribution range are marked as abnormal readings. The time windows marked with abnormal readings are directly excluded from the statistics in the subsequent masked switching identification to avoid outliers from entering the difference calculation and distorting the true switching baseline of the pump group. When the operating condition switches repeatedly within a short time window, the number of boundary time windows is relatively high. For boundary time windows that lack double-sided markings, the operating condition type of the adjacent time window is used to supplement the markings forward. The supplementary marking results are marked as inferred supplementary markings in the read archive record to distinguish them from direct reading markings. Inferred supplementary marking entries are counted separately from directly marked entries in the subsequent difference calculation to avoid the cumulative marking errors interfering with the deviation type determination results.

[0032] The switching frequency parameter value is compared with the light load reference set item by item to form the difference time series. The switching frequency parameter value f(i,t) of each time window is subtracted from the average switching value f_light(i) of the pump group in the light load reference set to obtain Δf(i,t)=f(i,t)-f_light(i). The difference is then compared with the light load reference set to make the difference time series comparable longitudinally. The Δf(i,t) corresponding to the peak load time window is usually a large positive value, the average load time window is a medium positive value, and the light load time window is close to zero. The time window position corresponding to the missing item in the difference time series is reserved as an empty space. The difference time series segments on both sides of the empty space are processed independently to avoid the introduction of time series misjudgment due to continuous identification across empty spaces. When the distance between adjacent empty spaces is too close and the interval between the two empty spaces does not exceed two scheduling time windows, the entire segment is downweighted as a low confidence segment to avoid the decrease in difference calculation accuracy caused by excessively dense local gaps being ignored in the subsequent identification of locked items. For pump groups with insufficient samples, the switching frequency parameter value is added to the difference time series after difference calculation. The difference value obtained by inferring the supplementary labeling entry in the switching frequency parameter value is inherited synchronously. The low-precision difference value is processed with a conservative threshold of 0.8 times when identifying occlusion switching. When the working condition gradually switches from light load to level load, the difference time series shows a trend of slowly rising from near zero in this transition period. The slope is usually maintained between 0.1 and 0.3 standard deviations per time window, which is significantly different from the single time window step amplitude value of 2 to 3 standard deviations caused by sudden change in working condition. The two types of modes trigger different judgment window settings in the occlusion switching recording stage. The judgment window of the gradual rise scenario is appropriately extended to accommodate the slight amplitude accumulation in the slow offset process, while the step scenario uses a single time window high-sensitivity judgment window to quickly locate the sudden change boundary.

[0033] After the difference time series undergoes periodic decay, a sudden zeroing event that remains continuously zero is identified and a masked switching record group is generated. The decay trend is determined by the sign of the difference between the absolute values ​​of the amplitudes of three consecutive time windows Δf(i,t) in the difference time series. A decay trend is confirmed when all three time window differences are negative. A single time window difference in the opposite direction does not interrupt the confirmed decay trend, thus accommodating minor fluctuations during the decay process. The sudden zeroing event requires that the absolute value of the amplitude of the time window Δf(i,t) immediately following the confirmation of the decay trend drops sharply to below 0.3 times the standard deviation of the light-load switching of the pump group. A continuous zeroing event requires that the absolute values ​​of the amplitudes of two or more consecutive time windows following the sudden zeroing event are below this threshold. Occasional low values ​​in a single time window are filtered out by the continuity condition. After confirmation, the starting time window of the sudden zeroing event, the length of the continuous zeroing interval, and the corresponding pump group identifier are summarized into a masked switching record. All confirmed records constitute a masked switching record group. When the final stage of the attenuation trend is close to the boundary window of the operating condition switch, the authenticity of the sudden zeroing needs to be determined by combining the boundary window operating condition annotations. The step zeroing of the difference caused by the operating condition switch itself does not constitute a shielded switch. The record is only confirmed when the starting window of the sudden zeroing is outside the operating condition switch buffer and the continuous zero segment is not divided by the operating condition switch. When switching from level load to light load, the difference time series shows rapid attenuation. The difference of some pump groups drops to zero suddenly after attenuating to near the light load threshold and lasts for multiple time windows. The continuous segment of sudden zeroing is characterized by the attenuation slope changing abruptly from a moderately negative value to close to zero within a single time window. This morphological feature is the core criterion for distinguishing between shielded switch and normal asymptotic convergence. The slope of normal asymptotic convergence continues to decrease continuously at the end of the convergence stage rather than abruptly returning to zero. The slope abrupt change of the two types of forms is retained as an auxiliary quantification field of shielding intensity in the shielded switch record group.

[0034] Based on the masked switching record group, the difference time series is correlated with the deviation type to generate a switching deviation sequence. The sudden zeroing start time window of each record in the masked switching record group divides the corresponding difference time series into an active segment before zeroing and a masked segment after zeroing. The deviation type of the active segment is determined by the combination of Δf(i,t) amplitude and sign. The masked segment is uniformly determined as masked type, and the true amplitude is estimated by filling the linear extrapolation value of Δf(i,t) at the end of the active segment. The extrapolation slope is the average of the difference slopes of the three consecutive time windows at the end of the active segment. The deviation type determination rules for the active segment are as follows: if the Δf(i,t) amplitude exceeds twice the standard deviation and is positive, it is determined as overload trigger type; if it is positive and does not exceed twice the standard deviation, it is determined as transition drive type; if the amplitude is negative, it is determined as underload suppression type. The three types of results are recorded in the switching deviation sequence entries with deviation type labels. In the differential timing sequence, time windows not covered by the masked switching record group are directly used to determine the deviation type according to the amplitude and sign rules. All time window entries are arranged in time sequence and together with the original value of Δf(i,t) and the deviation type label, they form the switching deviation sequence. After the sudden zeroing segment in the differential timing sequence is filled by masking, the continuity of the deviation amplitude is restored. Underload suppression entries usually appear in the fallback segment after the overload triggering peak in the timing sequence. The connection between the two forms a complete overload-blanking cycle in the switching deviation sequence. The complete appearance of this cycle within a single scheduling time window indicates that the corresponding pump group has a typical degradation mode of first overload triggering and then entering the masking state under the current operating conditions. This mode identification provides a complete timing context for the light load blanking compensation of the subsequent pump group operating deviation.

[0035] For the switching deviation sequence, a deviation locking term group is generated by identifying switching terms that are maintained by micro-switching and reverse amplification under the conventional noise filtering threshold. The conventional noise filtering threshold is set according to the multiple relationship of the standard deviation of each pump group in the light load reference set. The pump group with the smaller standard deviation corresponds to a smaller conventional noise filtering threshold, so that minor deviations are not filtered out. Switching terms with an absolute amplitude value lower than the conventional noise filtering threshold in the switching deviation sequence are normally removed as noise. Micro-switching terms are exceptions. The initial absolute amplitude value of such switching terms is lower than the conventional noise filtering threshold, but the amplitude does not converge to zero in subsequent continuous time windows but is maintained or reverse amplified. The term is confirmed when the number of consecutive time windows that meet the condition |Δf(i,t+k)|≥|Δf(i,t)| and k exceeds the duration threshold. The duration threshold is twice the length of the scheduling time window. Periods that meet the condition too shortly are considered occasional jitter and are not confirmed. All micro-amplitude reverse amplification and maintenance items in the switching deviation sequence are summarized to generate a deviation locking item group. Each item records the starting time window, amplitude evolution trajectory, and corresponding pump group identifier. The amplitude evolution trajectory is saved as a Δf(i,t) sequence from the starting time window to the confirmation time window. When the sequence length is greater than twice the duration length threshold, the item priority is increased by one level. When the length is less than 1.2 times the duration length threshold, the item is only retained as a low priority. During the water supply trough period, when the equipment gradually deteriorates, the switching deviation of a certain zone pump group remains near the normal noise threshold for a long time under light load conditions. However, the reduction of switching activity does not cause the deviation to return to zero. The deviation accumulates continuously in the same direction without offsetting in the opposite direction. In contrast, the low amplitude deviation when the equipment is normal usually fluctuates randomly back to zero after one or two time windows. The two types of patterns form a distinguishable difference in the monotonicity of the trajectory slope and the variance of each window. Based on this, it is identified as a micro-amplitude reverse amplification and maintenance item and included in the deviation locking item group. This accumulation pattern is most typical when aging equipment enters the early stage of substantial deterioration.

[0036] The pump set operating deviation is determined by reverse weighted verification under light load blanking conditions based on the deviation locking item group. The weighting coefficient of the reverse weighted verification is w_lock(i,t)=1+α_hide·I_hide(i,t), where I_hide(i,t) is the light load blanking indicator function. When the switching deviation sequence is judged to be underload suppression type in the light load time window and the Δf amplitude is continuously lower than the conventional noise filtering threshold, and the switching in the non-light load time window is continuously high, I_hide is 1; otherwise, it is 0. The blanking compensation coefficient α_hide is calibrated according to the historical blanking depth statistics of each partition. The blanking depth statistics are obtained by taking the average deviation amplitude of the historical non-light load time window and the average deviation amplitude of the light load time window of the partition. The partition with a larger drop has a higher α_hide. The weighted total deviation of each pump group is obtained by summing the products of Δf(i,t) and w_lock(i,t) of each item in the deviation locking item group and then normalizing them. This amount is measured as a standardized deviation multiple, making different pump groups directly comparable under the same dimension. The pump group operating deviations are arranged in descending order of value to form a deviation sequence. The pump group at the top of the sequence has the highest deviation and the most significant deterioration, and serves as the main object for subsequent cloud-edge responsibility attribution. During the low water supply period, the overall switching activity weakens. Abnormal switching signals caused by pump group deterioration are difficult to distinguish from normal operating conditions in the low-frequency operating background. The switching deviation during the light load window is consistently below the noise threshold, while the switching deviation is significantly higher during the high load period. The strong difference between the two periods confirms the existence of the blanking mechanism. The reverse weighting verification assigns higher weight to the blanking window, restoring the deterioration contribution masked by the low load operating environment to the total weighted deviation, so that the final deviation accurately reflects the true degree of deterioration of the pump group rather than just the apparent level during the high load period. The w_lock product corresponding to entries with low-precision annotations in the deviation locking item group is included in the summary with a weight of 0.5.

[0037] Step S103: Perform cross-condition fault reverse tracing on the linkage fault group to generate cross-condition reproducibility rate; perform sudden drop blind zone pre-correction on the linkage fault group based on the cross-condition reproducibility rate to generate selected fault group; and implement reverse locking of low reproducibility isolated items in the selected fault group to determine the dominant fault item.

[0038] Specifically, a cross-condition fault reverse tracing is performed on the linked fault group to generate the cross-condition recurrence rate. Starting from the current triggering time window of the linked fault group, the process traces back to the historical time sequence, searching for the occurrence time windows of the same fault item in the historical scheduling records for each operating condition type. The tracing depth covers all completed scheduling cycles in the historical record, with no fixed upper limit on the tracing time, until the beginning of the historical record or the first occurrence time of the fault item. Each fault item in the linked fault group has a unique identifier based on a combination of node identifier and deviation characteristics. During cross-condition tracing, this unique identifier is used to search for faults in the historical scheduling records for each operating condition type. The operating condition types are divided into three categories: light load, average load, and peak load. The number of historical occurrences of the fault item under each operating condition type is divided by the total number of historical time windows for the corresponding operating condition type to obtain the fault recurrence rate for that operating condition type. The three types of operating conditions are calculated independently and recorded side by side, together forming the cross-condition recurrence rate vector of the fault item. The relative magnitudes of the recurrence rates of the three types of operating conditions in the cross-condition recurrence rate vector reveal the sensitivity of the fault item to the severity of the operating condition. Fault items with a peak load recurrence rate significantly higher than the light load recurrence rate exhibit a positive dependence characteristic of accelerated triggering as the operating condition worsens. Fault items with similar recurrence rates among the three types have weak correlations and are marked as uniformly distributed items across operating conditions in the P3 operating condition trigger category group. When the number of time windows for a certain operating condition type in the historical scheduling records is less than five, the recurrence rate is incorporated into the historical insufficiency labeling system and propagated along with the cross-condition recurrence rate vector. The recurrence rates of operating condition types marked with historical insufficiency are weighted down in subsequent hierarchical scans to exclude the distortion of hierarchical distribution caused by the scarcity of samples leading to recurrence rate estimation bias. In the early stages of system operation, when the historical time window accumulation for each operating condition type is insufficient, the cross-condition recurrence rate vector is marked as historical insufficiency as a whole. Fault items marked with historical insufficiency are subject to a conservative threshold during the selection and removal phase of the fault group to avoid over-removal of truly high-risk fault items due to insufficient historical samples.

[0039] In some embodiments, the step of performing a sudden drop blind zone pre-correction on the linked fault group based on the cross-operating condition recurrence rate to generate a selected fault group includes: grouping the linked fault group according to the fault triggering condition type to form an operating condition triggering category group; performing a reverse hierarchical scan of the operating condition severity of the operating condition triggering category group based on the cross-operating condition recurrence rate to form a negative correlation hierarchical distribution; determining a sudden drop blind zone pre-correction critical point based on the negative correlation hierarchical distribution to form a critical correction threshold; and performing threshold coverage screening on the linked fault group based on the critical correction threshold to generate a selected fault group.

[0040] The linked fault groups are grouped into trigger condition category groups based on the fault trigger condition type. Grouping is based on the trigger time window condition type label recorded for each fault item in the linked fault group. This label has been archived during the switching frequency spectrum reading phase; the grouping operation directly references the archived label without re-determining the condition type. When the trigger time window of the linked fault group is not strictly aligned with the switching frequency spectrum time window, the condition label of the scheduling time window to which the center point of the trigger time window belongs is used as the grouping basis, ensuring that the condition attribution remains unique and avoids ambiguity in dual condition attribution during cross-time window mapping. For fault items whose trigger time window condition type label is an inferred supplementary label, an inferred source label is added to the corresponding condition trigger category group. Inferred source label entries are counted separately from directly labeled entries in subsequent severity level scanning to avoid inference errors affecting the reliability of the level scanning results. When the same fault item in a linked fault group is triggered multiple times and the triggering window spans multiple operating condition types, the fault item is assigned to the corresponding operating condition category according to the triggering window in the operating condition triggering category group. The same fault item can exist in multiple operating condition triggering category groups simultaneously, and the number of triggers and the triggering window are recorded together in the entries of each category group. Sparse categories with fewer than three fault items in a category group are merged into adjacent severity categories during the hierarchical scanning stage and participate in the mean calculation to avoid the distortion of the negative correlation hierarchical distribution caused by the high variance of the mean recurrence rate due to independent statistics of small sample categories. After the operating condition triggering category group is completed, the distribution of the number of fault items in each operating condition category characterizes the operating condition triggering bias of the linked fault group. The number of fault items in the peak load category group is much greater than that in the light load category group, which indicates that the linked faults are concentrated in a condition-sensitive manner. A uniform distribution indicates that the linkage faults are weakly correlated with the operating condition type. The P3 severity hierarchical scanning uses uniform sampling of all operating conditions for this type of fault item instead of reverse hierarchical scanning.

[0041] Based on the cross-condition recurrence rate, a negative correlation hierarchical distribution was formed by performing a reverse hierarchical scan of the condition severity for each condition triggering category group. The condition severity increases sequentially from light load, average load, to peak load, with the direction of increase being positive. The reverse hierarchical scan proceeds layer by layer from the peak load condition to the light load condition, recording the mean cross-condition recurrence rate of fault items within the current condition triggering category group layer by layer. The mean increases as the severity decreases, indicating a negative correlation. That is, the more lenient the condition, the higher the recurrence rate of the fault item, suggesting that the fault item is more likely to be triggered under low-severity conditions. The reverse hierarchical scan is performed independently for each fault item in the condition trigger category group. The recurrence rate of conditions with insufficient historical data is weighted at 0.5. The mean recurrence rate of each fault item is sorted in descending order of condition severity to form the hierarchical recurrence rate sequence for that fault item. A positive slope indicates that the recurrence rate increases progressively as the condition severity decreases, indicating a negative correlation. A negative slope indicates that the recurrence rate decreases synchronously as the severity decreases, indicating a positive correlation. The slopes of the hierarchical recurrence rate sequences of all fault items are summed to form a negative correlation hierarchical distribution. This distribution is organized with the slope value on the horizontal axis and the number of fault items on the vertical axis. Fault items with a slope greater than zero are clustered on the right side of the distribution, forming a set of negatively correlated fault items, while fault items with a slope close to zero or negative are clustered on the left side. Differential pressure deviation faults in water supply branch pipe areas show a typical negative correlation with the severity of operating conditions: during peak load, the main pipe pressure is sufficient, and the terminal pressure control margin is adequate, so the triggering conditions for this type of fault are continuously suppressed; during light load, the contraction of water supply causes frequent local differential pressure regulation, and the suppressed faults are released in a concentrated manner, with the hierarchical recurrence rate showing a continuously increasing slope in the descending order; the triggering of pump set mechanical wear faults is unrelated to operating conditions, and the recurrence rate of each level is evenly distributed with a slope close to zero. The two types of faults cluster on the left and right sides of the negatively correlated hierarchical distribution, and the difference in their distribution patterns directly defines the applicable scope of the pre-correction threshold for the sudden drop blind zone.

[0042] Based on the negative correlation hierarchical distribution, a critical correction threshold is formed by determining the pre-correction critical point for the sudden drop blind zone. The hierarchical recurrence rate sequence of each fault item in the negative correlation hierarchical distribution is processed level by level. The difference in recurrence rates between adjacent levels is calculated sequentially for all adjacent gear pairs. The adjacent gear pair with the largest difference is determined as the gear boundary where the sudden drop blind zone of that fault item is located. A sudden drop blind zone is confirmed when this difference exceeds twice the overall standard deviation of the sequence; otherwise, the fault item has no significant sudden drop blind zone and is not included in the threshold calculation. Fault items with confirmed sudden drop blind zones in the negative correlation hierarchical distribution are grouped according to the gear boundaries of the sudden drop condition. The average of the largest differences among all fault items under the same gear boundary is taken as the critical correction threshold benchmark for that gear boundary. The threshold benchmark plus one standard deviation of the group's differences is used as the critical correction threshold. This one standard deviation margin prevents fault items near the threshold from being incorrectly included or excluded due to minor estimation errors, thus providing a certain degree of fault tolerance in the boundary determination of fault items near the threshold. Critical correction thresholds are set separately for each operating condition level boundary within the set of negatively correlated fault items. Thresholds are calculated separately for the light load to flat load and flat load to peak load boundaries. The critical correction thresholds for each level boundary are dynamically updated as historical data accumulates. After each complete scheduling cycle, the thresholds are recalculated and overwritten. During seasonal changes, the water demand pattern shifts as a whole, and the average recurrence rate under each operating condition type moves synchronously. After the seasonal data is updated, the thresholds are tightened or loosened to better reflect the current operating characteristics. When the system is cold-started or during the initial stage of network expansion and renovation, the thresholds remain in a loose mode when historical samples are scarce. They are gradually tightened to a precise state as scheduling cycles accumulate.

[0043] Based on the critical correction threshold, a threshold coverage screening process is implemented for linked fault groups to generate a selected fault group. The threshold coverage screening process evaluates each fault item in the linked fault group individually. The judgment criterion is whether the difference in the hierarchical recurrence rate of the fault item at the corresponding gear boundary exceeds the critical correction threshold. Fault items exceeding the critical correction threshold are judged to have a significant sudden drop blind zone effect, and their cross-operating condition recurrence rate is systematically underestimated due to the sudden drop blind zone; these are directly retained and added to the selected fault group. Fault items that do not exceed the critical correction threshold at any gear boundary are judged to have a stable recurrence rate distribution and do not require sudden drop blind zone correction; they enter a further screening stage. This screening stage verifies whether the total historical trigger count of the fault item reaches the minimum retention threshold defined by the median of historical trigger counts. Items below this threshold are screened out, while those above are retained. When a fault item has difference data at multiple gear boundaries simultaneously, the threshold corresponding to the gear boundary with the largest difference is used as the screening criterion. This ensures that the gear with the most significant sudden drop blind zone effect dominates the screening decision, avoiding the low difference of the second most significant gears masking the need for correction of the primary significant gear. Fault items in the linked fault group with insufficient historical data are skipped from the threshold coverage screening and directly forcibly retained into the selected fault group according to the conservative threshold. This is to avoid the threshold calculation being distorted due to insufficient historical samples, which could lead to the wrong screening of truly high-risk items. The selected fault group consists of all fault items that have passed the retention judgment. Each item inherits the original fields of the linked fault group and is marked with a sudden drop blind zone confirmation status. The marking distinguishes between three sources: retention of significant sudden drop blind zone, retention of stable recurrence rate, and forced retention due to insufficient historical data. The three types of markings are distinguished by weights of 1.2, 1.0, and 0.7 times, respectively, in the screening of isolated candidate groups for P7 operating conditions.

[0044] In some embodiments, the step of reversing the order of isolated low-recurrence operating conditions to determine the dominant fault item in the selected fault group includes: performing cross-operating condition frequency scanning on each fault item in the selected fault group to generate an operating condition coverage distribution; filtering fault items that only occur during peak-valley switching moments based on the operating condition coverage distribution to form an isolated operating condition candidate group; performing isolated operating condition continuous triggering intensity assessment on the fault items in the isolated operating condition candidate group to generate a triggering intensity index; and reversing the order of the isolated operating condition candidate group to determine the dominant fault item based on the triggering intensity index.

[0045] A cross-condition frequency scan is performed on each fault item in the selected fault group to generate a condition coverage distribution. The cross-condition frequency scan counts the number of historical trigger windows for each fault item in the selected fault group under each condition type. The statistical range is consistent with the historical record range during cross-condition recurrence rate tracing, ensuring alignment of the statistical benchmarks between the two stages. Condition types with a non-zero number of trigger windows under each condition type are counted as one covered condition. The number of covered conditions is divided by the total number of all condition types (currently 3, expandable) to obtain the condition coverage rate. A coverage rate of 1 indicates that the fault item has historical trigger records under all condition types, while 1 / 3 indicates that it has only been triggered under one condition type. The condition coverage rate monotonically increases with the historical scheduling cycle. Adding a new trigger window increases the number of covered conditions, and existing covered conditions are not canceled due to subsequent non-triggering. This monotonically cumulative characteristic ensures that the condition coverage rate always reflects the maximum known condition distribution range of the fault item, rather than an instantaneous snapshot. The operating condition coverage distribution is organized using the numerical sequence of operating condition coverage rates for all fault items in the selected fault group. Clusters of fault items with low operating condition coverage rates indicate a significant tendency for isolated operating conditions within the selected fault group, while clusters with high operating condition coverage rates indicate that most fault items are weakly correlated with the operating condition type. Fault items with the same operating condition coverage rate but different trigger window distribution positions need further differentiation. Fault items triggered by peak-valley switching boundary windows and those triggered by windows during stable operation may have the same operating condition coverage rate, but the former's triggering is concentrated at the moment of operating condition switching, belonging to switching-induced isolated faults, while the latter belongs to stable-triggered faults within the operating condition. The two are distinguished at the same numerical position in the operating condition coverage distribution by the trigger window distribution concentration field, which records the proportion of the fault item's trigger window near the operating condition switching boundary window.

[0046] Based on the operating condition coverage distribution, fault items that only occur during peak-to-valley switching moments are selected to form an isolated operating condition candidate group. The peak-to-valley switching moment is defined by taking the operating condition type switching boundary window as the center, extending forward and backward by one scheduling window to form a switching buffer. Trigger windows within the switching buffer are determined to be triggered during the switching moment, while trigger windows outside the buffer are determined to be triggered in a steady state. When the operating condition type continues to switch twice within the switching buffer, the buffer of the second switching and the buffer of the first switching are merged into a joint buffer for unified determination, avoiding the division of the same continuous trigger segment into multiple independent buffer segments by rapid and repeated switching, which would artificially dilute the trigger concentration estimation. The selection criteria for the isolated operating condition candidate group are constrained simultaneously from two dimensions: the operating condition coverage dimension requires that the operating condition coverage of the fault item does not exceed the 25th percentile of the global operating condition coverage distribution, making the candidate items significantly limited to the operating condition; the trigger concentration dimension requires that the proportion of triggers within the switching buffer in all historical trigger windows of the fault item exceeds 60%, ensuring that the triggering behavior is indeed concentrated at the switching moment rather than randomly distributed throughout the low-severity operating condition. Fault items that meet both conditions and are weighted and normalized according to the three categories of P5 with weights of 1.2, 1.0, and 0.7 are included in the isolated candidate group of operating conditions. The short-term concentrated action of pressure regulating equipment during peak-valley load switching causes the pressure at the end of the pipeline to experience instantaneous fluctuations within the switching buffer. The triggering sequence of the differential pressure deviation fault associated with this is highly concentrated at the switching boundary and completely disappears throughout the steady-state light load. The differential pressure deviation caused by the progressive wear of pump valves is triggered throughout the light load, not dependent on the switching boundary, and the triggering sequence is evenly distributed. The former meets the screening conditions in both operating condition coverage and triggering concentration dimensions, while the latter only meets a single dimension.

[0047] For fault items in the isolated operating condition candidate group, an isolated operating condition continuous triggering intensity assessment is performed to generate a triggering intensity index. The assessment of continuous triggering intensity is based on the triggering timing of each fault item in the isolated operating condition candidate group within the switching buffer. The assessment objects are the number of consecutive triggering time windows and the triggering amplitude of the fault item within the switching buffer in a single peak-valley switching event. The more consecutive triggering time windows there are, the more likely the fault item is to be continuously activated rather than a single pulse trigger at the moment of switching. Continuous activation means that the fault item has a deep coupling relationship with the switching process rather than an occasional excitation. If the switching buffer is divided into multiple triggering segments by a brief return to normal operating conditions within a single switching event, N_cont(i) takes the number of time windows of the longest consecutive triggering segment instead of the sum of the lengths of each segment, to avoid incorrectly counting re-triggering after a brief steady-state interval as continuous activation and thus artificially inflating the triggering intensity index. The triggering amplitude is taken as the average deviation amplitude of each triggering time window within the switching buffer. The higher the average value, the more severe the deviation of the fault item at the moment of switching. The trigger intensity index S_trig(i) = N_cont(i) × V_amp(i) / σ_amp, where N_cont(i) is the number of consecutive trigger windows for fault item i within the switching buffer, V_amp(i) is the average amplitude of the trigger window deviation, and σ_amp is the summative standard deviation of the deviation amplitudes of all fault items in the isolated candidate group of the operating condition. This dimensionless processing makes the trigger intensity index a comparable pure value, allowing for direct comparison of trigger intensity indices for fault items of different node types under the same dimension. When a fault item has been triggered in multiple historical peak-valley switching events, S_trig(i) is taken as the weighted average of the evaluation results of each switching event. The weight of recent switching events is higher than that of earlier switching events. The weight decay coefficient is set according to the time span of the historical records; the longer the time span, the more significant the decay, indicating that the recent operating status has a higher reference value for fault evaluation than earlier historical records.

[0048] Based on the trigger strength index, the candidate groups of isolated faults are sorted in reverse order to determine the dominant fault items. The reverse order priority sorting adopts a compromise mechanism of first ascending order and then increasing the weight of high-ranking faults. First, the faults are sorted in ascending order according to the value of S_trig(i). In ascending order, fault items with low trigger strength index are temporarily placed in the first position. Although the isolated faults with low trigger strength index have significant limitations in the operating conditions, their continuous activation at the moment of switching is limited, and their threat to the overall health of the pipeline network is appropriately downgraded after the second weighting. Fault items with trigger strength index exceeding the 75th percentile of the global index are second-weighted after the reverse order sorting. The weighting operation extracts these high-intensity isolated faults from the end of the ascending order and re-inserts them into the first position in the order. The insertion position is mapped to the first interval of the queue in reverse order according to the ratio of the magnitude of the high-intensity item S_trig(i) exceeding the 75th percentile to the percentile. The larger the magnitude, the earlier the insertion position. The sorting result after the second weighting balances the downgrading effect of the reverse order sorting on low and medium intensity isolated faults and the need to prioritize the retention of high intensity isolated faults. After sorting, the top few items are selected as dominant fault items. The number of items selected is set according to a certain proportion of the scale of the linked fault group. The proportion parameter is written into the configuration item during the edge-side initialization phase (default is 20%). Large-scale linked fault groups allow more dominant fault items to be selected to cover multiple independent fault triggering links that may coexist in large-scale linked events. Dominant fault items are the set of faults in the isolated operating condition candidate group that are most deeply coupled with the operating condition switching and are repeatedly activated during each load alternation. They represent the weak links of the water supply system at the operating condition switching node. Operating condition isolated faults with high trigger intensity indicate that the corresponding node is subjected to action pressure exceeding the normal regulation capacity at the moment of operating condition switching and systematically generates deviations. They are actively exposed structural hidden dangers. Operating condition isolated faults with medium trigger intensity are occasionally activated at the moment of switching, and their linkage impact range is limited.

[0049] Step S104: Based on the dominant fault item and the pump set operation deviation, cloud-edge source identification is performed to determine the source of control responsibility. Based on the source of control responsibility, cross-shift responsibility deviation verification is performed to determine the cross-shift responsibility inversion item. Based on the cross-shift responsibility inversion item, root cause tracing is performed to generate a health diagnosis report.

[0050] Specifically, the source of control responsibility is determined by cloud-edge source identification based on the dominant fault item and the pump group operation deviation. The dominant fault item and the pump group operation deviation are matched by the overlap of the pump group scheduling actions within the trigger time window. Dominant fault item items with an overlap exceeding the threshold are paired records. The cloud responsibility score is quantified by the magnitude of the deviation of the cloud scheduling instruction from the normal instruction range within the trigger time window, and the edge responsibility score is quantified by the weighted total deviation of the pump group operation deviation within the corresponding time window. Both scores are normalized to the range of 0 to 1 based on the historical global extreme value. The difference between the two scores is ΔS = S_cloud - S_edge. The larger the absolute value of ΔS, the more singular and concentrated the source of responsibility. When the absolute value of ΔS is lower than the judgment threshold set according to the 25th percentile of the historical distribution, it is classified as equally responsible on both sides. The responsibility attribution category is divided into three categories: cloud-dominant, edge-dominant, and equally responsible on both sides. The source of control responsibility is recorded on each paired record with the ΔS value, the responsibility attribution category, and the corresponding work group identifier. The paired records of the equally responsible category are processed separately in the subsequent cross-work group responsibility deviation verification and are not involved in the comparison calculation of opposing responsibility directions. When the water supply peak shifts to a flat period, the cloud-based dispatching system centrally issues a target to reduce water supply. Within a short time window, the deviation of the cloud-based command from the normal range increases significantly, while the edge-side pump units are still in the inertial operation range, resulting in a significantly positive ΔS value. The control responsibility mainly belongs to the cloud-based dispatching decision. When the pump units mechanically degrade, the edge-side deviation slowly accumulates during the stable period of cloud-based dispatching, and ΔS continuously turns negative. The control responsibility mainly belongs to the edge equipment. The difference between the two sources in the sign and amplitude change rhythm of ΔS gives the source of control responsibility an identifiable temporal characteristic, which is used for subsequent cross-shift responsibility deviation verification to locate the shift pairs with opposing responsibility directions at the shift jurisdiction boundary.

[0051] Based on the source of control responsibility, cross-shift responsibility deviation verification is conducted to determine cross-shift responsibility inversion items. The responsibility attribution category of each paired record in the control responsibility source record is cross-expanded along two dimensions: trigger window and shift identification. When the responsibility attribution categories of adjacent shifts within the same trigger window in the control responsibility source are opposite, deviation identification is triggered. One side is determined to be cloud-dominated, while the other is determined to be edge-dominated. The opposition of responsibility directions forms a locatable deviation node at the shift jurisdiction boundary. Differences in responsibility direction spanning two or more shift jurisdiction areas do not trigger deviation due to excessively long transmission links; only the directional opposition between spatially adjacent shifts proceeds to subsequent stability confirmation. Stability confirmation requires that the same adjacent shift pair continuously exhibits opposing responsibility directions within three or more consecutive trigger windows; a single instance of opposing directions within a trigger window is considered an occasional fluctuation. In adjacent shift pairs with stable deviation, the side with the larger absolute value of ΔS is marked as the high-deflection shift pair, and the side with the smaller absolute value is marked as the low-deflection shift pair. The deviation amplitude is the difference between the absolute values ​​of ΔS of the two shifts; the larger the deviation amplitude, the deeper the disagreement between the two shifts in responsibility determination. Team pairs that meet both the amplitude and persistence conditions are identified as cross-team responsibility inversion items. The deflection amplitude and the number of stable persistence windows constitute the two core quantitative attributes of cross-team responsibility inversion items. The two attributes support two-dimensional ranking. When amplitude is prioritized, team pairs with larger deflection amplitudes are given priority for root cause tracing. When persistence is prioritized, team pairs with more persistence windows are given priority. The two ranking results are weighted and fused during the anchoring stage of the interference team set to determine the final processing priority. The combined magnitude of the two attributes describes the comprehensive severity of the inversion event in both the amplitude and persistence dimensions.

[0052] In some embodiments, the step of generating a health diagnosis report based on the cross-team responsibility inversion item includes: anchoring the cross-team responsibility inversion item to a set of interfering teams by reverse decoy anchoring of the responsible abnormal stable teams; collecting the operational offset and the reporting time sequence advance based on the interfering team set to form a root cause tracing record group; reverse-checking the cloud-edge control responsibility deviation of each record item in the root cause tracing record group to form a root cause tracing conclusion; and generating a health diagnosis report by reverse weighting and aggregation of low responsibility deviation units based on the root cause tracing conclusion.

[0053] For cross-shift responsibility inversion items, a set of interfering shifts is formed by using reverse decoy anchoring for shifts with abnormal responsibility stability. Each record in the cross-shift responsibility inversion items carries a trigger window and a corresponding shift identifier. Shifts with high frequency of occurrence in the cross-shift responsibility inversion items but low mean absolute value of ΔS in non-inversion time windows demonstrate stable control. The difference in control performance between these two time periods constitutes the abnormal responsibility stability characteristic; the larger the difference, the higher the priority of reverse anchoring for that shift. Anchoring operations are sorted by the difference between the frequency of occurrence of cross-shift responsibility inversion items and the stability score of regular scheduling records. The difference is obtained by subtracting the normalized value of the inversion frequency from the normalized value of the stability score. The two normalization benchmarks are taken as their respective historical maximum values. The shift identifiers with the highest difference ranking are aggregated to form the set of interfering shifts. Each shift entry in the interfering shift set carries three attributes: inversion frequency, stability score difference, and a trigger window list. The trigger window list records all time windows in which the shift appears in the cross-shift responsibility inversion items. When the jurisdiction boundary of a work group shift changes due to scheduling adjustments or changes in authority, the inversion frequency and stability score of the corresponding work group in the interfering work group set are recalculated according to the new boundary to avoid misjudgment of anchoring caused by sample confusion before and after the boundary shift. The historical trigger windows before the shift are mapped to the new work group identifier according to the attribution rules and retained. The upper limit of the number of entries in the interfering work group set is configured to one-third of the total number of pipeline work groups. When the upper limit is exceeded, the top few work groups are retained according to the difference. The trigger windows of the excluded work groups are merged into the corresponding records of the adjacent work groups according to the topological proximity rules. The number of merged trigger windows is updated synchronously to the inversion frequency statistics of the receiving work group to ensure that the total number of trigger windows remains complete after the merging operation. Work groups with large differences in both inversion frequency and stability score are marked with high anchoring strength in the interfering work group set. This mark is used as the basis for priority processing during the root cause tracing record group collection stage.

[0054] For example, the step of forming a root cause tracing record group based on the collected operational offset and reported time-series advance amount of the interference team set includes: tracing the operational offset and reported time-series advance amount of the upstream pipeline unit of the interference team set; performing pipeline unit deterioration trend scanning based on the operational offset to determine the deteriorated operational segment after intervention; performing team reverse-order anomaly verification on the deteriorated operational segment after intervention and the reported time-series advance amount to generate suspected root cause items; and tracing the suspected root cause items back to the upstream pipeline unit across teams to form a root cause tracing record group.

[0055] The operational offset and reporting time advance are obtained by tracing the upstream pipeline unit offset of the interfering work group set. The tracing operation is carried out upstream step by step along the pipeline topology within each trigger window, prioritizing high anchoring strength. It starts from the pipeline unit under the jurisdiction of each work group in the interfering work group set and extends step by step until it reaches the main node of the trunk line or encounters the cross-work group jurisdiction boundary. The difference between the measured operating parameters and the scheduling target parameters of each node on the tracing path is recorded as the operational offset of that node. The jurisdiction boundary of the interfering work group set is determined by the work group scheduling authority division table. When tracing to the boundary node, the boundary node identifier and the corresponding work group identifier are recorded simultaneously. The difference in operational offset between the nodes on both sides of the boundary is recorded as the cross-work group offset transmission amount. The larger the transmission amount, the stronger the transmission of the offset of the upstream work group jurisdiction area downstream. The advance time of the reported sequence is collected within the same trigger window. The difference between the submission time of each reported record and the detection time of the same type of anomaly by the scheduling system is stored using a dual-dimensional index of node identifier and trigger window. When there are multiple reported records for the same node and the same trigger window, the earliest submission time is used to participate in the advance time calculation, so that the advance time of the reported sequence corresponds to the earliest time when the shift team perceives the anomaly of the node. The reporting record time of the main control node is earlier than the time scheduling system's perception of the same type of anomaly. At the same time, the running offset of the node has been continuously exceeding the normal range. The offset propagation in the cross-shift direction increases step by step along the upstream direction of the pipeline. The superposition of the three indicators being synchronously high indicates that there is an upstream source of continuous output interference on the topology path where the node is located. The three indicators are recorded together in the offset tracing result and indexed by a dual-dimensional index of node identifier and trigger window. The complete topology sequence of the upstream tracing path is saved along with the offset tracing result. The running offsets of each node in the path are arranged in order from upstream to downstream. The offset gradient in the arrangement direction describes the propagation direction and attenuation law of the offset on the topology path.

[0056] Based on the operational offset, a relapse trend scan of the pipeline unit is performed to determine the relapsed operational segment after intervention. The scan target is the operational offset time sequence of each node on the upstream tracing path within a continuous trigger time window. The determination of relapse is based on the basic pattern of two consecutive time windows of offset decrease followed by two consecutive time windows of offset increase. When the average offset after the increase exceeds 80% of the average offset before the decrease, it is confirmed as a relapse event; if it is below 80%, it is determined as normal fluctuation convergence. During the scan, the slope of the decreasing segment and the slope of the increasing segment are recorded simultaneously. When the absolute value of the increasing slope is greater than that of the decreasing slope, it indicates that the relapse speed is faster than the intervention convergence speed. Such rapid relapse events are marked with a high-speed relapse label. This label, together with the value of the increasing slope, quantifies the urgency of intervention failure. The post-intervention relapse operation segment consists of time windows covering all confirmed relapse events. Each post-intervention relapse operation segment records three time nodes: the initial time window, the lowest drop time window, and the relapse confirmation time window. The time window interval between the lowest drop time window and the relapse confirmation time window corresponds to the duration of the intervention effect. The shorter the interval, the faster the intervention effect fades and the weaker the health recovery capability of the corresponding pipeline unit. Cloud-based intervention adjusts scheduling commands to temporarily converge the offset to the normal range. However, when pressure-bearing components in the pipeline continue to age, the convergence effect is difficult to maintain. The offset rapidly rebounds within the immediate time window after intervention, exceeding the pre-intervention level. The rebound speed is significantly faster than the convergence speed, forming an asymmetrical V-shaped time trajectory. This rapid relapse after intervention indicates that the intervention measures only temporarily suppress the symptoms while the root cause continues to produce deviations. In contrast to pipeline units where the intervention effect is maintained stably, this is an important time signal for locating the source of control failure. The operational offset of each time window is completely preserved with the relapse operation segment.

[0057] The post-intervention deterioration operation segment and the reported time-series advance amount are used for reverse-order anomaly verification by work group to generate suspected root cause items. The reverse-order verification of the post-intervention deterioration operation segment scans backward from the resumption confirmation time window to the starting time window, checking the sign and amplitude of the reported time-series advance amount for each work group in each time window. Work groups with a continuously positive reported time-series advance amount and an increasing trend in amplitude in the reverse-order scanning direction indicate that their perception of upstream anomalies during the deterioration process is continuously advanced. Work groups with the largest advance amount amplitude near the deterioration confirmation time window and gradually decreasing towards the starting time window are recorded as positive anomaly verification work groups. Work groups with irregular fluctuations or continuously negative advance amounts throughout the entire deterioration operation segment are recorded as negative anomaly verification work groups. The positive judgment requires that the advance amount of three consecutive time windows in the reverse-order scanning direction are all positive and the amplitude is monotonically increasing. If the advance amount of a single time window is positive but the preceding and following time windows do not meet the monotonic condition, a positive judgment is not triggered. A positive verification result, along with the three attributes of the corresponding post-intervention relapse operation segment's initial time window, relapse confirmation time window, and relapse amplitude, constitutes a suspected root cause record. The relapse amplitude is the difference between the average offset after recovery and the average offset before decline. When multiple work groups report time series covering the same relapse operation segment, reverse verification independently assesses the positive results for each work group. If the number of positive work groups exceeds half of the covered work groups, it is determined to be a group root cause event for that relapse segment and is marked with a collaborative root cause label, distinguishing it from an isolated root cause event with a single positive work group. Suspected root cause items are arranged in descending order of relapse amplitude to form a root cause priority queue. The suspected root cause item at the top of the queue serves as the path starting point in the cross-work group upstream tracing stage. The remaining items in the queue are filtered according to the cross-relationship threshold and retained for additional reference when generating root cause tracing conclusions. All positive records are summarized to form suspected root cause items, and the relapse amplitude of each suspected root cause item is recorded alongside the positive work group identifier.

[0058] For suspected root cause items, the source of the upstream pipeline unit across work shifts is traced to form a root cause tracing record group. The upstream source tracing across work shifts starts with the positive work shift identifier in the suspected root cause item, and continues upstream along the pipeline topology, crossing the work shift jurisdiction boundary. The tracing depth is capped at two work shift jurisdiction levels; anything exceeding this is marked as depth pending tracing and manually reviewed. Boundary node identifiers and crossing work shift identifiers are recorded at each level boundary. The crossing work shift identifiers are retained as candidate identifiers for potential root cause source work shifts in the root cause tracing record group. The operational offset of each upstream node along the cross-work shift tracing path is read node by node within the time series covered by the suspected root cause item start window to the recurrence confirmation window. Cross-correlation analysis is performed on the operational offset time series of upstream cross-work shift nodes with persistently high operational offsets and positive work shift jurisdiction nodes. The cross-correlation function... x(t) represents the time series of upstream cross-shift node offsets, y(t) represents the time series of offsets of nodes under the jurisdiction of positive shifts, τ represents the time shift, and the peak value of R(τ) when τ is positive corresponds to the number of time windows in which the upstream node offset leads the downstream node offset. The larger the leading time shift, the earlier the change in the upstream node offset occurs. Upstream cross-shift nodes with a leading time shift exceeding one scheduling time window are identified as cross-shift source nodes. The four attributes of the cross-shift source node identifier, the class identifier, the leading time shift, and the average running offset are merged with the suspected root cause record of the corresponding positive shift. After merging, they are arranged according to the dual-dimensional index of the positive shift identifier and the triggering time window to form a root cause tracing record group. When multiple positive shifts within the same triggering time window are identified as cross-shift source nodes, they are each independently formed into merged records. Multiple records are retained side by side in the root cause tracing record group, and the cross-shift source node with the largest leading time shift corresponding to the cross-shift source node is marked as the dominant source.

[0059] The root cause analysis results are derived by retrospectively verifying the cloud-edge control responsibility deviation of each record item in the root cause analysis record group. The retrospective verification traces back from the trigger window to the historical time series, recalculating the cloud responsibility score and edge responsibility score of each record item in the root cause analysis record group corresponding to the network unit for each time window. The retrospective depth covers the three complete scheduling cycles before the trigger window; if the number of trigger windows is insufficient, it is automatically extended. The average of the cloud and edge responsibility scores for the three cycles is used to form the retrospective average cloud score and retrospective average edge score for that record item. The difference between the two averages is the cloud-edge control responsibility deviation of that record item, D_resp = S_cloud_retro - S_edge_retro. A positive D_resp indicates that the cloud responsibility is consistently high during the retrospective period, while a negative D_resp indicates that the edge responsibility is consistently high. The larger the absolute value of D_resp, the more stably the source of responsibility is concentrated on one side during the retrospective period. The weighted average of D_resp values ​​from multiple records of the same work group within the root cause tracing record group is calculated. The weight is the normalized value of the operational offset of the corresponding trigger window. Records corresponding to trigger windows with higher operational offsets have greater weight in the mean calculation, causing the D_resp mean to tilt towards the responsibility direction of the period with the most severe deviation. Work groups with a D_resp mean exceeding twice the global standard deviation are identified as single-sided responsibility-dominant work groups. Work groups that do not exceed the threshold are marked as responsibility-ambiguous work groups. When the D_resp mean of a responsibility-ambiguous work group is not clearly assigned to a single side, but its historical failure frequency is similar to that of a single-sided responsibility-dominant work group, it participates in weight compensation as a key unit for reverse weighting during the downstream health diagnosis report generation stage. This mechanism allows hidden high-risk units that are underestimated in routine responsibility assessments to emerge at the forefront of the health priority sequence. The root cause tracing conclusion consists of three elements: work group identification, D_resp mean, and responsibility attribution category. Single-sided responsibility-dominant work groups and responsibility-ambiguous work groups are assigned to different conclusion categories in the root cause tracing conclusion.

[0060] Health diagnostic reports are generated by reverse weighting aggregation of low-responsibility deviation units based on root cause analysis results. Low-responsibility deviation units are network units in the root cause analysis results where the absolute value of the mean D_resp is lower than D_threshold. Each screened unit is calculated with a reverse weighting coefficient according to w_up(j) = 1 + β·(1 - |D_resp(j)| / D_threshold). Units with higher weighting coefficients receive greater weight in the weighted health score aggregation. The weighting coefficient for reverse weighting is w_up(j) = 1 + β·(1 - |D_resp(j)| / D_threshold), where D_threshold is 0.5 times the standard deviation of the global distribution of D_resp. This threshold definition allows the judgment criterion for low-responsibility deviation units to adapt to the global distribution pattern. β is the weighting compensation coefficient, which is calibrated according to the frequency of actual failures of units with unclear responsibility in history; the higher the frequency, the larger β. The weighted health score of all pipeline units is obtained by summing the mean D_resp values ​​of each record item in the root cause analysis conclusion using w_up(j). The weighted health scores are arranged from high to low to form a health priority sequence. The health diagnosis report is based on the health priority sequence, and each pipeline unit entry records the weighted health score, responsibility category, mean D_resp value, and corresponding work group identifier. When the weighted health scores of multiple pipeline units in the same spatial segment are simultaneously high, a concentrated regional labeling is added. The top 10% of pipeline units in the health priority sequence are selected into a high-priority set. The spatial distribution of units within the set is calculated based on topological proximity using a clustering coefficient. When the clustering coefficient is greater than 0.6, the entire topological segment triggers concentrated regional labeling. When the clustering coefficient is less than 0.2, the high-scoring units arranged in the sequence are labeled independently. The health diagnosis report lists the concentrated regional labeling segment as a key focus segment.

[0061] Step S105: Based on the health diagnosis report and the cross-operating condition recurrence rate, an inverse weight superposition is performed to generate a pump group maintenance scheduling queue. The pump group maintenance scheduling queue and the health diagnosis report are used to evaluate the resilience margin and generate a resilience margin attenuation rate. Based on the resilience margin attenuation rate, a high attenuation rate reverse frequency reduction linkage control is performed to generate a scheduling control command.

[0062] Specifically, a pump group maintenance scheduling queue is generated by inversely weighting and superimposing health diagnosis reports and cross-condition recurrence rates. The weighted health score of each pipeline unit in the health diagnosis report is associated with the corresponding pump group identifier. When the same pump group manages multiple pipeline units, the maximum weighted health score of each unit is taken as the representative health risk value of that pump group. The average of the three types of recurrence rates for the corresponding fault items in the cross-condition recurrence rate vector is calculated using weights of 0.5 for peak load, 0.3 for average load, and 0.2 for light load to form the comprehensive recurrence rate. The formula for calculating the inverse weighted summation is W_maint(i) = α·H_risk(i) + (1-α)·(1-R_comp(i)), where H_risk(i) is the normalized result of the health risk representative value, R_comp(i) is the normalized result of the comprehensive recurrence rate, and α is dynamically adjusted according to the skewness of the overall health score distribution in the health diagnosis report. When the distribution is right-skewed, α is increased to increase the ranking contribution of the health risk representative value; when the distribution is left-skewed, α is decreased to increase the ranking contribution of the comprehensive recurrence rate. W_maint(i) is arranged in descending order of value to form a pump group maintenance scheduling queue. Each entry carries four attributes: pump group identifier, health risk representative value, comprehensive recurrence rate, and W_maint(i) value. The pump group maintenance scheduling queue is divided into three levels according to the value of W_maint(i): high priority is given to pump groups with W_maint(i) greater than the global 75th percentile, medium priority is given to pump groups with W_maint(i) between the 25th and 75th percentiles, and low priority is given to pump groups with W_maint(i) less than the 25th percentile. The three levels correspond to maintenance scheduling intervals of 7 days, 14 days, and 30 days, respectively. High priority pump groups are usually concentrated in key nodes of the main water supply area with high health risk representative values ​​and low overall recurrence rate. Although their deterioration has not yet been reproduced in large quantities, the risk has accumulated and priority intervention is required to prevent it from evolving into a systemic failure. Low priority pump groups have both indicators in the normal range and maintain a regular observation cycle. The three-level configuration allows maintenance resources to be concentrated on the pump groups with the highest current risk rather than being diluted evenly.

[0063] A resilience margin assessment is performed on the pump set maintenance scheduling queue and health diagnosis report to generate a resilience margin decay rate. The water supply capacity maintenance duration is calculated by taking the time sequence of the operating offset of each pump set's corresponding pipeline unit in the health diagnosis report as input, and calculating the linear fitting slope of the most recent three complete scheduling cycles. A positive slope in the health diagnosis report indicates a continuous increase in the operating offset. The time required for the operating offset to increase from its current value to the safety threshold according to the slope is the water supply capacity maintenance duration, with the safety threshold being the upper limit of the design pressure. The load pressure resilience is the reciprocal normalized result of the extreme values ​​of the switching deviation sequence of each pump set in the pump set maintenance scheduling queue within the historical peak load time window. The smaller the extreme values ​​of the switching deviation sequence in the pump set maintenance scheduling queue, the higher the load pressure resilience; larger extreme values ​​indicate that the peak load impact has caused significant switching anomalies. The comprehensive value of resilience margin is obtained by dividing the product of the maintenance duration and the pressure-bearing elasticity by the product of the reference average values ​​of the two indicators for the normally operating pump group. The normalization of the product form makes the deterioration contribution of the two indicators equally reflected in the comprehensive value. The difference between the comprehensive value of resilience margin and the comprehensive value of resilience margin in the previous scheduling cycle is divided by the comprehensive value of resilience margin in the previous scheduling cycle to obtain the resilience margin decay rate. A positive resilience margin decay rate indicates that the resilience margin in this cycle has decreased compared with the previous cycle, and a negative rate indicates that it has recovered to some extent. The resilience margin attenuation rate is arranged in descending order to form an attenuation priority queue. After the peak operation of the water supply system, the extreme values ​​of the switching deviation and the time-series slope of the operating offset of some pump sets continue to accumulate, the comprehensive value of resilience margin shrinks continuously, and the attenuation rate remains stable at a high level. The remaining resilience margin of these pump sets is consumed the fastest and they need to be included in the forced frequency reduction protection immediately to prevent the margin from being exhausted. The comprehensive resilience value of pump sets that are operating smoothly does not fluctuate much, and the attenuation rate hovers at a low level. The scheduling interval does not need to be adjusted. The attenuation rate of pump sets whose resilience has shown a recovery trend falls and they are removed from the frequency reduction protection list. The three states are clearly distributed in different percentile intervals of the attenuation priority queue.

[0064] Based on the resilience margin attenuation rate, high attenuation rate reverse frequency reduction linkage regulation generates scheduling control commands. The high attenuation rate judgment threshold is set according to the 75th percentile of the global distribution of resilience margin attenuation rate. Pump groups exceeding the threshold are judged as high attenuation rate pump groups, and the scheduling trigger interval extension ratio is equal to the multiple of the attenuation rate exceeding the threshold (e.g., 1.4 times the threshold corresponds to 1.6 times the extension). The greater the exceedance, the greater the extension of the scheduling trigger interval. After the frequency of switching on high attenuation rate pump groups is reduced, the load pressure of the corresponding pipeline unit is proportionally taken over by pump groups with higher comprehensive resilience margin values ​​in the same pressure zone. The share of each pump group is obtained by dividing its comprehensive resilience margin value by the sum of the comprehensive resilience margin values ​​of all non-high attenuation rate pump groups in the same pressure zone. The higher the comprehensive resilience margin value, the greater the share of the pump group. After the share of ... The dispatch control instructions are based on the dispatch trigger interval and switching frequency target values ​​of each pump group. Dispatch control instructions for pump groups with high attenuation rates are marked with a frequency reduction protection label, and dispatch control instructions for pump groups receiving compensation are marked with a compensation acceptance label. The changes in the trigger interval and the changes in the switching frequency target value of each pump group are recorded side by side in the dispatch control instructions. The validity period of the dispatch control instructions is consistent with the current dispatch cycle. After the expiration, the instruction changes for each pump group are recalculated according to the resilience margin attenuation rate of the next cycle. Pump groups that have not had instruction changes for three consecutive cycles are removed from the frequency reduction linkage control checklist. The changes in the trigger interval and the changes in the switching frequency target value of each pump group in the current cycle are archived in two dimensions: pump group identification and dispatch cycle. The dispatch control instructions are refreshed synchronously with the update cycle of the resilience margin attenuation rate. Pump groups that are newly triggered with high attenuation rate judgment are automatically included in the frequency reduction protection. Pump groups whose resilience margin attenuation rate falls below the threshold are automatically restored to normal dispatch trigger intervals and removed from the frequency reduction protection label.

[0065] To implement the edge-side water supply system scheduling and control method corresponding to the above method embodiments, in order to achieve the corresponding functions and technical effects. See also Figure 2 , Figure 2 This paper illustrates a structural block diagram of an edge-side water supply system scheduling and control device 200 according to an embodiment of this application, including: The deviation identification module 201 is used to collect pipeline pressure data and water flow data, and perform cloud-edge load deviation identification on the pipeline pressure data and the water flow data to generate load deviation characteristics. The decoupling verification module 202 is used to separate cloud-edge degradation based on the load deviation characteristics to generate a linkage fault group, measure the pump switching frequency of the linkage fault group to generate a switching frequency spectrum, and perform light load reference condition verification on the switching frequency spectrum to generate the pump operation deviation. The correction and screening module 203 is used to perform cross-operating condition fault reverse tracing to generate cross-operating condition recurrence rate for the linkage fault group, perform sudden drop blind zone pre-correction on the linkage fault group based on the cross-operating condition recurrence rate to generate selected fault group, and implement low recurrence rate operating condition isolated item reverse locking to determine the dominant fault item in the selected fault group. The responsibility diagnosis module 204 is used to identify the source of control responsibility based on the dominant fault item and the pump group operation deviation, to conduct cross-shift responsibility deviation verification based on the source of control responsibility to determine the cross-shift responsibility inversion item, and to generate a health diagnosis report based on the root cause tracing based on the cross-shift responsibility inversion item. The instruction generation module 205 is used to generate a pump group maintenance scheduling queue based on the health diagnosis report and the cross-operating condition recurrence rate by inverse weight superposition, to evaluate the toughness margin of the pump group maintenance scheduling queue and the health diagnosis report to generate a toughness margin attenuation rate, and to generate a scheduling control instruction by performing high attenuation rate reverse frequency reduction linkage regulation based on the toughness margin attenuation rate.

[0066] The aforementioned edge-side water supply system scheduling and control device 200 can implement an edge-side water supply system scheduling and control method according to the above-described method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application embodiment can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0067] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

Claims

1. A method for scheduling and controlling an edge-side water supply system, characterized in that, include: Collect pipeline pressure data and water flow data, and perform cloud-edge load deviation identification on the pipeline pressure data and water flow data to generate load deviation characteristics; Based on the load deviation characteristics, cloud-edge degradation separation is performed to generate a linkage fault group. The pump switching frequency of the linkage fault group is measured to generate a switching frequency spectrum. The switching frequency spectrum is verified under light load reference conditions to generate the pump operation deviation. Cross-condition fault reverse tracing is performed on the linked fault group to generate cross-condition reproducibility rate. Based on the cross-condition reproducibility rate, the linked fault group is subjected to sudden drop blind zone pre-correction to generate selected fault group. Low reproducibility isolated items of the selected fault group are reverse locked in reverse order to determine the dominant fault item. Based on the dominant fault item and the pump set operation deviation, cloud-edge source identification is performed to determine the source of control responsibility. Based on the source of control responsibility, cross-shift responsibility deviation verification is performed to determine the cross-shift responsibility inversion item. Based on the cross-shift responsibility inversion item, root cause tracing is performed to generate a health diagnosis report. Based on the health diagnosis report and the cross-operating condition recurrence rate, an inverse weighted superposition is performed to generate a pump group maintenance scheduling queue. The pump group maintenance scheduling queue and the health diagnosis report are used to evaluate the resilience margin and generate a resilience margin attenuation rate. Based on the resilience margin attenuation rate, a high attenuation rate reverse frequency reduction linkage regulation is performed to generate a scheduling control command.

2. The method according to claim 1, characterized in that, The step of identifying load deviation characteristics by comparing the pipeline pressure data and the water flow data includes: The measured load analysis of the pipeline pressure data and the water flow data generates an edge load sequence. The edge load sequence is used to perform cloud-edge double-sided load deviation positioning according to the scheduling time window to generate a cloud-edge deviation sequence; For the cloud edge deviation sequence, residual abnormal convergence period identification is performed to determine the load stagnation interval; Based on the load stagnation interval analysis, the convergence depth and load deviation characteristics generated by following the source direction are analyzed.

3. The method according to claim 1, characterized in that, The process of separating cloud-edge degradation and generating linked fault groups based on the load deviation characteristics includes: The load deviation characteristics are grouped into cloud-edge dual-time-window sources to generate a source grouping table; The degradation direction analysis table is generated by calculating the edge anomaly rate during the normal period in the cloud and the cloud anomaly rate during the normal period in the edge based on the source grouping table. Based on the degradation direction analysis table, fault pairs with no edge anomalies but continuous cloud collaboration failures are identified and degradation isolation groups are formed. Based on the aforementioned deterioration isolation group, the pre-trigger items for the last stable deterioration are selected to generate a linked fault group.

4. The method according to claim 1, characterized in that, The step of generating pump set operating deviation by performing light load reference condition verification on the switching frequency spectrum includes: Light load reference benchmark set is constructed by screening the light load condition switching parameters from the switching frequency spectrum; The light load reference set and the switching frequency spectrum are used to calculate the switching difference to form a switching deviation sequence; The cut-and-cast deviation sequence is subjected to conventional noise filtering, and the cut-and-cast terms are identified and the deviation locking term group is generated by reverse amplification and preservation of the micro-amplitude cut-and-cast under the threshold. The pump set operating deviation is determined by reverse weighted verification under light load blanking conditions based on the aforementioned deviation locking item group.

5. The method according to claim 1, characterized in that, The step of generating a selected fault group by performing a sudden drop in blind zone pre-correction on the linked fault group based on the cross-operating condition recurrence rate includes: The linked fault groups are grouped according to the fault triggering condition type to form condition triggering category groups; Based on the cross-condition recurrence rate, the condition trigger category group is subjected to a reverse hierarchical scan of condition severity to form a negative correlation hierarchical distribution; Based on the aforementioned negative correlation hierarchical distribution, a critical correction threshold is formed by determining the pre-correction critical point for the sudden drop blind zone. Based on the critical correction threshold, the linked fault group is subjected to threshold coverage screening to generate a selected fault group.

6. The method according to claim 1, characterized in that, The step of performing low-recurrence-rate isolated item reverse locking on the selected fault group to determine the dominant fault item includes: A cross-condition frequency scan is performed on each fault item in the selected fault group to generate a condition coverage distribution. Based on the operating condition coverage distribution, fault items that only occur during peak-valley switching are selected to form an isolated operating condition candidate group. For the fault items in the isolated candidate group of the operating conditions, an isolated operating condition continuous triggering intensity assessment is performed to generate a triggering intensity index. Based on the trigger strength index, the isolated candidate groups of the operating conditions are sorted in reverse order to determine the dominant fault item.

7. The method according to claim 1, characterized in that, The process of generating a health diagnosis report based on the cross-team responsibility reversal includes: The cross-team responsibility inversion item is anchored by a reverse decoy for the responsibility anomaly stabilizing team to form an interference team set; Based on the interference team set, the operational offset and the reporting time sequence advance amount are used to form a root cause tracing record group; The root cause tracing record group is used to reverse-check the cloud-edge control responsibility deviation of each record item to form a root cause tracing conclusion. Based on the root cause analysis conclusions, a health diagnosis report is generated by reverse weighting and aggregation of low-responsibility-bias units.

8. The method according to claim 4, characterized in that, The step of calculating the cut-in difference between the light-load reference set and the cut-in frequency spectrum to form a cut-in deviation sequence includes: Read the switching frequency parameter values ​​for each operating condition based on the switching frequency spectrum; The difference time series is formed by performing a step-by-step difference operation between the switching frequency parameter value and the light load reference set; The difference time series is subjected to periodic decay of the difference and then suddenly returns to zero, generating a masked switching record group; Based on the masked switching record group, the time series difference is correlated with the deviation type to generate a switching deviation sequence.

9. The method according to claim 7, characterized in that, The root cause tracing record group, formed based on the operational offset and reporting time sequence advance of the interference team set, includes: The upstream pipeline unit offset tracing of the interference team set yields the operating offset and the reporting timing advance. Based on the aforementioned operational offset, a pipeline unit deterioration trend scan is performed to determine the deteriorating operational segment after intervention; The post-intervention deterioration operation segment and the reported time sequence advance amount are used to perform reverse-order anomaly verification at the work group level to generate suspected root cause items. The suspected root cause items are traced back to the upstream pipeline unit across different work groups to form a root cause tracing record group.

10. A scheduling and control device for an edge-side water supply system, characterized in that, include: The deviation identification module is used to collect pipeline pressure data and water flow data, and perform cloud-edge load deviation identification on the pipeline pressure data and the water flow data to generate load deviation characteristics. The decoupling verification module is used to separate cloud-edge degradation based on the load deviation characteristics to generate a linkage fault group, measure the pump switching frequency of the linkage fault group to generate a switching frequency spectrum, and verify the switching frequency spectrum under light load reference conditions to generate the pump operation deviation. The correction and screening module is used to perform cross-operating condition fault reverse tracing to generate cross-operating condition reproducibility rate for the linkage fault group, perform sudden drop blind zone pre-correction on the linkage fault group based on the cross-operating condition reproducibility rate to generate selected fault group, and implement low reproducibility rate working condition isolated item reverse locking to determine the dominant fault item in the selected fault group. The responsibility diagnosis module is used to identify the source of control responsibility based on the dominant fault item and the pump set operation deviation, to conduct cross-shift responsibility deviation verification based on the source of control responsibility to determine the cross-shift responsibility inversion item, and to generate a health diagnosis report based on the root cause tracing based on the cross-shift responsibility inversion item. The instruction generation module is used to generate a pump group maintenance scheduling queue by inversely weighting the health diagnosis report and the cross-operating condition recurrence rate, evaluate the resilience margin of the pump group maintenance scheduling queue and the health diagnosis report to generate a resilience margin attenuation rate, and generate scheduling control instructions by performing high attenuation rate reverse frequency reduction linkage regulation based on the resilience margin attenuation rate.

Images