Urban residential water use behavior classification method based on smart water meter

Abstract

The application discloses a city resident water use behavior classification method based on a smart water meter, and relates to the technical field of intelligent water affairs. The application is used for improving the accuracy of water use behavior identification and anomaly detection. First, the smart water meter generates a water flow change direction sequence in real time, captures based on stable and continuous triggering events, extracts flow gradient and timestamp density features; then, an adaptive filtering method is used to eliminate water hammer noise, key turning points are extracted by combining curvature analysis and compared with behavior pattern library topology, and a double-channel behavior feature vector is generated; further, the metering grid is divided according to the valve distribution in the building water supply system, and the pressure gradient is combined to perform distributed clustering, and a cross-grid abnormal feature conduction path is constructed; finally, the multi-layer graph convolution network is used to trace the abnormal diffusion path, locate the water pressure mutation inflection point and map it to the building component, output the leakage point identifier and behavior pattern label, and realize high-precision classification and anomaly identification of the resident water use behavior.

Description

Technical Field

[0001] The present invention relates to the field of smart water technology, and specifically to a method for classifying urban residents' water use behavior based on smart water meters. Background Art

[0002] With the continuous advancement of urbanization, the scale of urban water supply systems continues to expand, and residents' water demand is becoming increasingly diverse and fluctuating. This makes the refined identification and management of user water use behavior an important foundation for ensuring the efficiency of urban water resource scheduling, optimizing water supply structure layout, and improving leakage warning capabilities. The widespread deployment of smart water meters has enabled urban water supply companies to obtain high-frequency, fine-grained flow data, providing key support for data-driven water use behavior modeling and anomaly detection. However, relying solely on raw water meter readings to conduct water use statistical analysis is difficult to effectively reveal the dynamic changes in residents' water use behavior and its complex correlation with the status of the water supply system. There is an urgent need to develop a classification method with higher recognition accuracy and scenario adaptability.

[0003] Existing water use behavior classification methods mainly rely on fixed threshold trigger mechanisms or template matching recognition methods. Their core processes generally have two prominent problems: First, traditional methods generally use static thresholds or fixed rules to detect water use events, ignoring the seasonal changes and data volatility of individual users' historical behavior patterns, resulting in insufficient trigger sensitivity and prone to false detection or missed detection; Second, although some methods introduce event waveform comparison mechanisms, most of them match with a single time or flow sequence, failing to effectively characterize the spatial topological evolution of behavior patterns. In particular, in multi-user environments, there are problems of sample feature confusion and blurred boundaries, which cannot support high-precision abnormal behavior judgment and spatial positioning, limiting their application efficiency in complex urban water systems. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention provides a method for classifying urban residents' water usage behavior based on smart water meters, which solves the problems of the above-mentioned background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a classification method for urban residents' water use behavior based on smart water meters, comprising the following steps: S1. Calculating the flow difference between adjacent sampling points in real time through the smart water meter terminal to generate a water flow change direction sequence, triggering event capture when the continuous stability time of the direction sequence exceeds the user's historical extreme value range for the same period, and extracting the flow change gradient distribution and timestamp density characteristics within the event; S2. Adaptively configuring the noise elimination window size according to the timestamp distribution density characteristics, performing water hammer effect filtering on the water use event waveform, locating the waveform curvature mutation point in combination with the flow change gradient distribution characteristics, constructing a key turning point sequence, and connecting the nodes of the sequence The topological consistency comparison is performed on the typical waveform templates in the relationship and behavior pattern library to generate a dual-channel feature vector that describes the waveform similarity and topological connection integrity; S3. The independent metering grid units are divided according to the physical valve distribution of the building water supply network, and the dual-channel feature vectors of users in the same grid unit are subjected to pressure gradient-driven distribution clustering analysis. The real-time fluctuation data of the pipe network pressure sensor is combined to construct the cross-grid abnormal feature transmission path; S4. The diffusion direction of the abnormal feature in the transmission path is analyzed through a multi-layer graph convolutional network, and the water pressure attenuation mutation point is traced back along the physical direction of the water supply pipeline. The leakage point identification and abnormal water use behavior pattern mark specific to the building component location are output.

[0006] Furthermore, step S1 includes the following steps: using the smart water meter terminal to calculate the flow difference between adjacent sampling points in real time to generate a water flow change direction sequence, marking the flow increase as a positive state, the flow decrease as a negative state, and the flow stability as a zero state; triggering an event capture mechanism when the direction sequence continuously maintains the same state for a duration that exceeds the maximum duration threshold of the user's history during the same period; extracting the gradient distribution characteristics of the flow change in the water use event, calculating the distribution law of the flow change rate in each time period through sliding window analysis, extracting the timestamp density characteristics, and statistically analyzing the distribution characteristics of the number of sampling points in the unit time window.

[0007] Furthermore, the noise elimination window size is adaptively configured according to the timestamp distribution density characteristics, and the specific process of performing water hammer effect filtering on the water use event waveform is as follows: the noise elimination window size is dynamically matched based on the timestamp distribution density characteristics, a compact filtering window is used in the timestamp high-density area, and an extended filtering window is used in the timestamp low-density area; water hammer effect elimination processing is performed according to the hydraulic characteristics of water-using appliances, a high-frequency oscillation suppression mode is used for the faucet start-stop pulse waveform, and a low-frequency damping mode is used for the long-period waveform of bathing equipment, so as to eliminate the waveform distortion caused by the sudden change of pipe network pressure and retain the core morphological characteristics of water use behavior.

[0008] Furthermore, the waveform curvature mutation points are located in combination with the flow change gradient distribution characteristics, and the specific process of constructing the key turning point sequence is as follows: the curvature detection sensitivity is adaptively adjusted in combination with the flow change gradient distribution characteristics, the detection sensitivity is reduced in the low change intensity area to avoid accidentally touching small fluctuations, and the detection sensitivity is increased in the high change intensity area to accurately capture the state transition points; the key turning positions of the waveform are located through curvature extreme value analysis, and an ordered turning sequence including starting points, peak points, valley points, and ending points is constructed, and the time coordinates and type attributes of each turning point are recorded.

[0009] Furthermore, the node connection relationship of the sequence is compared with the typical waveform template in the behavior pattern library for topological consistency, and the specific process of generating a dual-channel feature vector describing the waveform similarity and topological connection integrity is as follows: an adjacency matrix of the node connection relationship of the turning sequence is established, and a three-level topological consistency verification is performed with the typical waveform template in the behavior pattern library. First, the sequence consistency of the key turning point type is matched, and then the continuity of the connection relationship between adjacent turning points is verified, and finally the integrity of the special topological structure is detected; a first-channel feature vector is generated to independently characterize the similarity of the waveform morphology, and a second-channel feature vector is simultaneously generated to quantify the integrity of the topological connection relationship, forming a dual-channel behavioral feature code.

[0010] Furthermore, independent metering grid units are divided according to the physical valve distribution of the building water supply network, and the specific process of pressure gradient-driven distribution clustering analysis of the dual-channel feature vectors of users in the same grid unit is as follows: independent metering grid units are divided according to the actual installation positions of the physical valves in the building water supply network, and each grid unit forms a closed hydraulic monitoring area; the dual-channel feature vectors of all users in the same grid unit are collected, and the real-time pressure difference vector direction between adjacent user nodes is calculated; the position of the user behavior clustering center is dynamically adjusted according to the pressure difference vector direction; the waveform similarity feature of the first channel and the topological connection integrity feature of the second channel are integrated to calculate the user behavior association weight coefficient; and a spatial distribution heat map of abnormal water use behavior in the grid unit is generated.

[0011] Furthermore, the specific process of constructing cross-grid abnormal characteristic conduction paths based on the real-time fluctuation data of pipeline network pressure sensors is as follows: real-time fluctuation data of pressure sensors at the boundaries of adjacent grid units are obtained, and the matching degree between the pressure fluctuation characteristics and the abnormal behavior cluster areas in the spatial distribution thermodynamic map is detected; when the matching degree exceeds the set threshold, a directional conduction link is established from the high-pressure area to the low-pressure area, the characteristic value of the topological connection integrity of the second channel is extracted, and the attenuation correction of the conduction link strength is performed according to the characteristic value to generate a cross-grid abnormal conduction path dataset.

[0012] Furthermore, a multi-layer graph convolutional network is used to analyze the diffusion direction of abnormal features in the conduction path, and the specific process of tracing the water pressure attenuation mutation point in reverse along the physical direction of the water supply pipeline is as follows: the cross-grid abnormal conduction path dataset is loaded into the graph convolution processing module, the bottom-level convolution operation is performed to extract the conduction direction characteristics within the grid, and the top-level convolution operation is performed to capture the cross-grid cascade conduction law; the water pressure loss gradient value of each pipe section is calculated in reverse along the physical connection direction of the water supply pipeline, the position of the sudden inflection point of the water pressure loss gradient curve is detected, and the inflection point position is mapped to the three-dimensional coordinates of the building in combination with the pipeline network spatial topology database.

[0013] Furthermore, the specific process of outputting leakage point identification and abnormal water use behavior pattern marking specific to the location of building components is as follows: according to the building's three-dimensional coordinates of the inflection point of the water pressure loss gradient mutation, the engineering component database is retrieved, the leakage point identification of the specific building component is output and associated with the flange sealing surface or valve interface number; the dual-channel feature change pattern in the conduction path dataset is simultaneously analyzed, when the waveform similarity distortion value of the first channel exceeds the threshold, the equipment sealing failure is marked, when the topological connection integrity fracture value of the second channel exceeds the threshold, the pipeline structure damage is marked, and when the dual-channel features are abnormal at the same time, the user operation function failure is marked.

[0014] The present invention has the following beneficial effects:

[0015] (1) A classification method for urban residents' water use behavior based on smart water meters introduces a dynamic trigger mechanism based on the user's historical fluctuation range during the same period during the data collection stage, which effectively avoids the misjudgment problem of fixed threshold detection in individual difference scenarios and improves the adaptability and accuracy of water use event detection. At the same time, the noise processing window is adaptively adjusted in combination with the timestamp density feature, and the water hammer effect filtering and flow gradient analysis are superimposed, so that the key curvature change points in the event waveform can be accurately extracted. The constructed key turning point sequence further generates a dual-channel feature vector through a topological consistency comparison mechanism, which fully integrates the waveform morphological information and structural connection features, enhances the expression ability of typical water use behavior patterns, and solves the defects of existing methods in terms of fuzzy behavior boundaries and low pattern recognition accuracy.

[0016] (2) A classification method for urban residents' water use behavior based on smart water meters improves the ability to locate abnormal behaviors and the interpretability of propagation paths through two levels: spatial clustering and graph neural network reasoning. Specifically, the metering grid is divided by the physical structure of the water supply network, and combined with the valve distribution and pressure gradient changes, grid-oriented multi-user feature clustering and abnormal path construction are achieved, avoiding the feature interference caused by regional coupling in the traditional global matching method; further, a multi-layer graph convolutional network is introduced to deeply analyze the diffusion direction of abnormal features, and the physical topology of the water supply pipeline is used to reversely trace the source of water pressure changes, so that leakage points and abnormal behavior users can be accurately identified, significantly improving the response efficiency and handling capabilities of risky behaviors in complex urban water supply networks.

[0017] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of the method for classifying urban residents’ water usage behavior based on smart water meters in the present invention. DETAILED DESCRIPTION

[0019] The embodiment of the present application solves the problems of insufficient accuracy in identifying water use events, limited ability to express behavioral patterns, and fuzzy positioning of abnormal propagation paths in existing methods through a method for classifying urban residents' water use behavior based on smart water meters, thereby improving the behavioral perception accuracy and spatial decision-making capabilities of urban water supply systems in multi-user scenarios.

[0020] The overall idea of ​​the solution in the embodiments of this application is as follows:

[0021] Firstly, by using the high-frequency sampling data of smart water meters, a sequence of water flow change directions is dynamically generated by calculating the flow difference between adjacent sampling points. Combined with the range of changes in users' historical behavior during the same period, a personalized trigger mechanism is constructed to accurately identify water use events. After the event is captured, the flow gradient change characteristics and timestamp density information are further extracted.

[0022] Secondly, the noise suppression parameters are adaptively adjusted according to the timestamp distribution characteristics, water hammer effect filtering is performed, and the waveform curvature mutation points are accurately located in combination with the water flow gradient structure to construct a stable key turning point sequence; the sequence is then compared with the standard template in the behavioral pattern library for topological structure consistency to generate a dual-channel feature vector that simultaneously characterizes waveform similarity and structural difference, thereby enhancing the robustness and accuracy of classification.

[0023] Subsequently, the metering grid units were divided according to the valve layout of the building water supply network, and a distributed clustering analysis driven by pressure gradient was carried out on the characteristic vectors within the grid. The pressure sensor data was integrated to construct a cross-grid abnormal conduction map.

[0024] Finally, a graph convolutional network is used to perform multi-layer modeling of the propagation path of abnormal features, and the source of pressure attenuation mutation is reversely located based on the pipeline network topology. The leakage points at the building component level and their corresponding abnormal water use behavior types are output, realizing a closed-loop analysis from individual behavior identification to system-level risk diagnosis.

[0025] See also Figure 1 , an embodiment of the present invention provides a technical solution: a classification method for urban residents' water use behavior based on smart water meters, comprising the following steps: S1. using a smart water meter terminal to calculate the flow difference between adjacent sampling points in real time to generate a water flow change direction sequence, triggering event capture when the direction sequence is continuously stable for a period exceeding the user's historical extreme value range for the same period, and extracting the flow change gradient distribution and timestamp density features within the event; S2. adaptively configuring the noise elimination window size based on the timestamp distribution density features, performing water hammer effect filtering on the water use event waveform, locating the waveform curvature mutation point in combination with the flow change gradient distribution features, constructing a key turning point sequence, and connecting the node connection relationship of the sequence with the behavior. Perform topological consistency comparison for typical waveform templates in the pattern library to generate dual-channel feature vectors that describe waveform similarity and topological connection integrity; S3. Divide independent metering grid units according to the physical valve distribution of the building water supply network, perform pressure gradient-driven distribution clustering analysis on the dual-channel feature vectors of users in the same grid unit, and construct cross-grid abnormal feature transmission paths based on the real-time fluctuation data of the pipe network pressure sensor; S4. Analyze the diffusion direction of abnormal features in the transmission path through a multi-layer graph convolutional network, trace the water pressure attenuation mutation point in reverse along the physical direction of the water supply pipeline, and output leakage point identification and abnormal water use behavior pattern markers specific to the location of building components.

[0026] In this implementation, step S1 is used to identify whether a user's specific water use behavior constitutes an independent "event." The so-called "flow change direction sequence" refers to determining whether the water flow is continuously increasing, decreasing, or remaining stable based on the flow difference between adjacent sampling points (i.e., current flow minus the flow at the previous sampling point), thereby forming a set of directional change indicators (e.g., rising, falling, or flat). This sequence can reflect whether the user has entered a typical water use cycle (e.g., bathing, washing clothes). When this directional change is continuous and the duration of stability exceeds the extreme range of the user's historical behavior during the same period (i.e., an "unusually long" water use behavior for this user), the system determines it as a water use event and records it. The "flow change gradient distribution" refers to the pattern of changes in the water flow rate throughout the event, further describing the intensity and trend of the behavior. The "timestamp density" refers to the concentration of sampling time points within the water use event, reflecting whether the event occurred in a concentrated manner, how long it lasted, and whether there were interruptions, which facilitates the identification of subsequent behavioral patterns. Step S2 aims to extract patterns and encode structures from the waveform data corresponding to the water use event. First, the length of the noise filtering window is dynamically adjusted based on the timestamp distribution density (i.e., the concentration of sampling points during an event). This ensures that noise suppression neither oversmoothes the actual waveform nor misses important details. "Water hammer filtering" identifies and removes atypical spike waveforms caused by sudden tap closures and sudden pressure changes, preventing them from interfering with actual behavior recognition. Subsequently, the "flow rate gradient distribution features" are used to help locate points of sudden curvature changes in the waveform—i.e., behavioral turning points (e.g., from peak water usage to decay), forming a "key turning point sequence." These turning points form a graph structure, with nodes representing turning points and edges representing logical connections between adjacent turning points. This graph is then compared with a template graph in a pre-defined "behavioral pattern library" for topological consistency, comparing the structural similarities between the two graphs. This ultimately generates a "two-channel feature vector": the first channel represents the morphological similarity of the waveform itself, while the second channel represents the difference in waveform structure (i.e., the connection between turning points) from the template structure. Step S3 introduces spatial distribution constraints, making user behavior analysis more closely aligned with the physical structure of the pipe network. Specifically, "physical valve distribution" refers to the water supply control structures within a building (such as floor valves, household branch valves, etc.), which naturally divide the water supply system into multiple independent or semi-independent "metering grid units." In these units, users share part of the water supply path, and there may be coupling between behaviors. In each divided grid unit, the collected "dual-channel feature vectors" are fed into the "pressure gradient-driven distribution cluster analysis" module, where "pressure gradient" refers to the change in water pressure in the pipe network caused by users turning on and off water-using equipment. This analysis method comprehensively considers the similarity of the feature space and the impact path driven by physical pressure to more accurately distinguish between local common behaviors and sudden abnormal behaviors.Furthermore, combined with real-time pressure sensor fluctuation data, an "abnormal feature transmission path" is established. This describes how abnormal behavior propagates across the pipe network and between grids, facilitating subsequent identification of the source of the problem. This mechanism is more systematic and predictive than traditional isolated user behavior analysis. Step S4 is an advanced pattern analysis phase based on graph neural networks. Graph convolutional networks (GCNs) are a type of deep learning model suitable for graph-structured data. They achieve feature extraction and pattern learning within the graph structure through feature propagation and weighted computation between nodes and their neighbors. In this step, a multi-layer graph convolutional network is used to analyze the "abnormal feature transmission path graph" constructed in step S3, identifying how the abnormal behavior spreads within the network, which nodes experience the abnormality first, and which nodes are affected, thereby establishing a chain of causal relationships across time and space. The system then "traces backwards along the physical path of the water supply pipeline" to identify the source of the sudden change in water pressure. This involves determining which component or pipe section first experienced the abnormal water pressure, potentially indicating a leak or illegal water use. The "output leakage point identification" is the spatial location of the abnormal source, and the "behavior pattern marker" is used to classify which type of water use behavior the abnormality belongs to (such as sudden leakage, long-term high-frequency water use, etc.), providing a basis for subsequent management.

[0027] Specifically, step S1 includes the following steps: using the smart water meter terminal to calculate the flow difference between adjacent sampling points in real time to generate a water flow change direction sequence, marking the flow increase as a positive state, the flow decrease as a negative state, and the flow stability as a zero state; triggering the event capture mechanism when the direction sequence continuously maintains the same state for a duration that exceeds the maximum duration threshold of the user's history during the same period; extracting the gradient distribution characteristics of the flow change in the water use event, calculating the distribution law of the flow change rate in each time period through sliding window analysis, extracting the timestamp density characteristics, and statistically analyzing the distribution characteristics of the number of sampling points in the unit time window.

[0028] In this implementation, step S1.1: The flow difference between adjacent sampling points is calculated in real time through the smart water meter terminal to generate a water flow change direction sequence, marking an increase in flow as a positive state, a decrease in flow as a negative state, and a stable flow as a zero state; this step is the basis for identifying water use behavior events. The continuous flow data obtained through the smart water meter terminal is recorded as a sampling sequence. ,in Indicates the The instantaneous flow value at a time point, the sampling time interval is a fixed value The flow difference calculation formula is: ; Direction status determination rule is: if , then it is marked as positive (+1); if , then it is marked as negative; if , then it is marked as stable (0); where, The minimum effective change threshold is set to avoid misjudgment caused by small random fluctuations. The generated direction sequence is recorded as: Step S1.2: When the duration of the direction sequence maintaining the same state continuously exceeds the historical maximum duration threshold of the user in the same period, the event capture mechanism is triggered; this step is to determine whether the water use behavior has sufficient duration and consistency to constitute an independent event. The historical maximum duration threshold is denoted as , for users The statistical upper bound of the longest single direction state duration in the same time period (such as 6:00 to 9:00 every morning). The judgment rule is: if there is a continuous segment ,and ,and , it is considered to have entered an effective water use event, and the time window is recorded as the event range. Step S1.3: Extract the gradient distribution characteristics of flow changes in the water use event, and calculate the distribution law of flow change rate in each period through sliding window analysis; set the event occurrence interval to be , divide it into overlapping sliding windows, each window length is , the step size is , remember The window is . Calculate the average gradient change rate within each window: ; The gradient distribution of the flow change of the entire event is recorded as a vector: ;in, It can be used to describe the rate trend and phase change of flow increase or decrease in the event. Step S1.4: Extract the time-interval density feature and count the distribution characteristics of the number of sampling points in the unit time window; this step is used to measure the time series density and fluctuation intensity of the sampling data. For the same total time, different fluctuation frequencies may represent different water use behaviors. Assume that the event period Divided into consecutive fixed-length time slices, each of which is , remember The time slice is , count the number of sampling points in each time slice . Define the density sequence as: The statistical characteristics of density include maximum value, mean value, and standard deviation parameters, which can be used for subsequent waveform structure extraction and feature encoding.

[0029] Specifically, the noise elimination window size is adaptively configured according to the timestamp distribution density characteristics, and the specific process of performing water hammer effect filtering on the water use event waveform is as follows: the noise elimination window size is dynamically matched based on the timestamp distribution density characteristics, a compact filtering window is used in the timestamp high-density area, and an extended filtering window is used in the timestamp low-density area; water hammer effect elimination processing is performed according to the hydraulic characteristics of water-using appliances, a high-frequency oscillation suppression mode is used for the faucet start-stop pulse waveform, and a low-frequency damping mode is used for the long-period waveform of bathing equipment, so as to eliminate the waveform distortion caused by the sudden change of pipe network pressure and retain the core morphological characteristics of water use behavior.

[0030] In this implementation, step S2.1: Dynamically matching the noise removal window size based on the timestamp distribution density feature: This step adjusts the filter window size according to the timestamp density of different time intervals in the water use event, thereby optimizing the granularity of noise removal. Assume that the total duration of the water use event is , the interval is divided into non-overlapping time periods , each segment For each time interval , calculate the timestamp density: ;in: : Time period Timestamp density; : In the time period The number of sampling points recorded in the :The start and end time points of this segment. Define the filter window size as: ;in: : For time period The filter window length; : The initial coefficient of the filter window is used to control the overall window size benchmark; : Density weight index, which controls the degree of compression of high-density areas. Explanation: When Larger (timestamps dense), then becomes smaller, which is suitable for fine noise reduction in high-frequency disturbance areas; when is smaller (timestamps are sparse), then It becomes larger and is suitable for trend extraction in low-frequency slow-changing areas. Step S2.2: Select the filter response mode according to the hydraulic characteristics of the water-using appliance: different types of water-using behaviors have significantly different hydraulic waveform characteristics. In order to avoid the loss of behavioral characteristics caused by the traditional unified filtering strategy, this step adopts different filter response modes according to the known or inferred equipment type. The characteristics of the pulse waveform after the faucet stops: short period, high-frequency oscillation; filtering method: use a median filter or bandpass filter in high-frequency suppression mode to eliminate the high-frequency oscillation at the start / shutdown moment; mathematical expression (bandpass filter kernel): ;in: : After filtering at time point The flow value; : high frequency response filter coefficient; : The first The flow rate value at each sampling point. Characteristics of long-period waveforms for bathing equipment: long duration, large fluctuation amplitude, but low frequency. Filtering method: A weighted moving average filter with low-frequency damping mode is used to suppress abnormal spikes caused by sudden changes in water pressure. Mathematical expression (damped weighted moving average): ;in: : Weighting factor for the low-frequency response filter; other parameters are the same as above. Step S2.3: Preserve the core morphological features of water use behavior: During the filtering process, by adjusting the window scale and filter response mode, the main structural features of the waveform (such as the flow rate change trend, main band, and periodic characteristics) are retained, while effectively weakening high-frequency abnormal vibrations caused by physical disturbances such as water hammer to avoid waveform distortion. In addition, to ensure that the filtering process does not cause misidentification of feature points, differential verification processing can be performed after each round of filtering: ;like Keep the original value, where: : Difference amplitude before and after filtering; : Morphological protection threshold, used to avoid key point distortion.

[0031] Specifically, the waveform curvature mutation points are located in combination with the flow change gradient distribution characteristics, and the specific process of constructing the key turning point sequence is as follows: the curvature detection sensitivity is adaptively adjusted in combination with the flow change gradient distribution characteristics, the detection sensitivity is reduced in the low change intensity area to avoid accidentally touching small fluctuations, and the detection sensitivity is increased in the high change intensity area to accurately capture the state transition point; the key turning position of the waveform is located through curvature extreme value analysis, and an ordered turning sequence including starting point, peak point, valley point, and end point is constructed, and the time coordinates and type attributes of each turning point are recorded.

[0032] In this implementation, the waveform curvature mutation point location and key turning point sequence are constructed, and the curvature detection sensitivity is adaptively adjusted based on the gradient strength. First, the local flow change gradient sequence of the water use event waveform is calculated. Assume that the waveform flow value sequence is , the corresponding sampling time series is , then the local change gradient of point u is: On this basis, in order to avoid misidentification of turning points in weak fluctuation areas, a gradient intensity adjustment factor is constructed: ; Binding regulatory factors , adjust the sensitivity of the curvature detection process, and the subsequent curvature detection threshold Defined as: ;in: : Time point The original flow value; : local gradient estimate of point u; : normalized gradient strength factor; : Dynamic curvature sensitivity threshold of point u; α: Initial threshold coefficient; β: Adjustment factor index. Perform local curvature analysis and locate curvature mutation points. Based on the obtained sensitivity adjustment threshold, calculate the curvature value of each sampling point. , which is used to reflect the local curvature of the waveform. The curvature estimation formula is: ; When the absolute value of the curvature at a point Exceeds its sensitivity threshold If the extreme value condition is met (i.e., it is a local maximum or minimum), it is determined to be a curvature mutation point and retained as a candidate turning point. Construct a key turning point sequence. All curvature mutation points that pass the test are marked according to the logical order of the waveform and the type of characteristic value (rising edge, peak, falling edge, valley, plateau, etc.), and the attributes of each point are recorded. Finally, a key turning point sequence is constructed: ;in: : key turning point sequence; : No. The time coordinates of the turning points; : No. Type identification of a turning point, such as "starting point", "peak point", "valley point", "end point", etc. Type identification It can be automatically determined by the following methods: And the traffic is on the rise, it is marked as the “starting point”; if And it is a local maximum, marked as a "peak point"; if If the flow rate is close to constant within a continuous time period, the end point is marked as the “end point”.

[0033] Specifically, the node connection relationship of the sequence is compared with the typical waveform template in the behavior pattern library for topological consistency, and the specific process of generating a dual-channel feature vector describing the waveform similarity and topological connection integrity is as follows: establish an adjacency matrix of the node connection relationship of the turning sequence, and perform three-level topological consistency verification with the typical waveform template in the behavior pattern library. First, match the key turning point type sequence consistency, then verify the continuity of the connection relationship between adjacent turning points, and finally detect the integrity of the special topological structure; generate a first-channel feature vector that independently characterizes the similarity of the waveform morphology, and simultaneously generate a second-channel feature vector that quantifies the integrity of the topological connection relationship to form a dual-channel behavioral feature code.

[0034] In this implementation, dual-channel behavioral feature vectors are generated based on topological consistency comparison to construct a node connection relationship diagram of the key turning point sequence. Let the key turning point sequence be , convert the sequence into a directed graph form and establish the node connection relationship adjacency matrix : ;in: : The node connection adjacency matrix corresponding to the current waveform; : Time coordinates of key turning points; : Turning point index; the connection relationship is derived based on the time increment and flow transition logic to ensure that the path has behavioral interpretability. Perform a three-level topological consistency comparison with the waveform template in the behavioral pattern library. Assume that there are K types of typical water use behavior templates in the behavioral pattern library, which are represented as: ;in: : The turning point type sequence of the kth template; : The corresponding adjacency matrix reflects the connection structure within the template. The comparison process is divided into the following three levels: Level 1 comparison: Turning point type sequence matching Let the type sequence of the current behavior be , the improved edit distance algorithm is used to calculate the type sequence matching score: ;in: : represents the minimum edit distance between turning point type sequences; : indicates the sequence length; : The type matching score of the kth template, range [0,1]. Secondary comparison: connection relationship continuity check compares the current waveform adjacency matrix Adjacency matrix with template Structural similarity: ;in: : indicator function, 1 if the condition is met, otherwise 0; : Connection relationship consistency score. Level 3 comparison: Special topological structure integrity check identifies whether the behavior template contains a characteristic substructure (such as the three-stage "sharp rise, platform, slow descent", etc.), and determines whether the current waveform has the complete structure. This part uses a rule-driven approach to output the structure matching score: ;in: : Based on the subgraph structure matching function, the structural integrity score in the range of 0 to 1 is output. Generate a dual-channel feature vector and combine the results of the above three-level comparison to construct a dual-channel behavior code: the first channel (waveform morphology similarity) feature vector : ; Second channel (topological connectivity integrity) feature vector : ; Finally, a dual-channel behavioral feature coding pair is formed: . : The first channel feature, reflecting the similarity of waveform shape; : The second channel feature reflects the topological connection integrity.

[0035] Specifically, independent metering grid units are divided according to the physical valve distribution of the building water supply network, and the specific process of pressure gradient-driven distribution clustering analysis of the dual-channel feature vectors of users in the same grid unit is as follows: independent metering grid units are divided according to the actual installation positions of the physical valves in the building water supply network, and each grid unit forms a closed hydraulic monitoring area; the dual-channel feature vectors of all users in the same grid unit are collected, and the real-time pressure difference vector direction between adjacent user nodes is calculated; the position of the user behavior clustering center is dynamically adjusted according to the pressure difference vector direction; the waveform similarity feature of the first channel and the topological connection integrity feature of the second channel are integrated to calculate the user behavior association weight coefficient; and a spatial distribution heat map of abnormal water use behavior in the grid unit is generated.

[0036] In this implementation scheme, based on the dual-channel user behavior cluster analysis driven by pressure gradient, the independent metering grid unit is divided according to the physical topology of the building water supply system to identify the actual installation location of the physical valve in the pipe network; with the valve as the boundary, multiple closed water supply subsystems are divided to form independent metering grid units, and the water use behavior in each unit is affected by the same hydraulic source; each grid unit contains a group of user nodes , each user has a generated dual-channel feature vector pair Constructing a pressure difference vector field Using pressure sensors deployed in the pipe network to collect real-time pressure values ​​at each user node ; Calculate the relative pressure difference vector field between users and define adjacent nodes The pressure difference vector between them is: ;in: :user The water pressure value at the current sampling moment; :express arrive The relative physical space position vector of : Indicates the directionality and intensity of the pressure difference. The pressure-driven resultant force vector of each user node is obtained by scanning the entire grid: ;in: : With user There is a set of adjacent nodes with direct water supply connections. Dynamic cluster center adjustment takes the user's dual-channel feature vector as the initial input and adopts an improved density-guided clustering algorithm (such as DBSCANPG); based on the combined pressure vector of each user Direction, offset the cluster center, the adjustment formula is: ;in: :user In the The location of the feature space cluster center in the iteration; : Adjustment coefficient to control the influence of pressure guidance on the offset of cluster center. This step ensures that users who are spatially affected by similar water pressure fluctuations have a higher coupling degree in feature clustering, thereby more accurately identifying behavioral correlation. Calculate the behavioral correlation weight coefficient between users to fuse the dual-channel features into a comprehensive behavioral similarity weight. , and its calculation formula is: ;in: : vector cosine similarity function; : Channel fusion weight coefficient (generally set to 0.5~0.7); : indicates user and The degree of correlation in behavioral patterns. Generate a spatial distribution heat map to map the abnormal degree of clustered user behavior (such as cluster outliers, low-weight users) to their physical locations; based on the relative coordinates and abnormal weights within the grid cells to which they belong, construct a two-dimensional heat map matrix: ;in: :user The physical space coordinates of :user Abnormal water use behavior score (calculated by indicators such as isolation and low correlation); A two-dimensional Gaussian kernel function or Dirac function controls the degree of fuzzy diffusion of the heat map. Finally, a heat map of abnormal behavior of all users within the grid is output for subsequent tracing of abnormal areas or behavioral intervention.

[0037] Specifically, the specific process of constructing cross-grid abnormal characteristic conduction paths by combining the real-time fluctuation data of pipeline network pressure sensors is as follows: real-time fluctuation data of pressure sensors at the boundaries of adjacent grid units are obtained, and the matching degree between the pressure fluctuation characteristics and the abnormal behavior clustering areas in the spatial distribution thermal map is detected; when the matching degree exceeds the set threshold, a directional conduction link from the high-pressure area to the low-pressure area is established, the characteristic value of the second channel topological connection integrity is extracted, and the attenuation correction of the conduction link strength is performed according to the characteristic value to generate a cross-grid abnormal conduction path dataset.

[0038] In this embodiment, the grid boundary pressure fluctuation characteristics are extracted for all adjacent grid cell pairs. , extract the set of pressure sensor nodes from its common edge area ;For each boundary sensor node Extract the pressure change sequence within a continuous time period: ;in: : Time point Time sensor The collected instantaneous pressure value; : Represents the characteristic sequence of pressure fluctuations in a continuous time series. Matching the abnormal cluster area with the pressure fluctuation pattern to analyze the grids on both sides of the boundary sensor Abnormal spatial heat map , calculate the centroid coordinates of the high-heat zone respectively ;Construct pointing vector: ; Further calculate the consistency index of the direction vector and the average fluctuation gradient vector of the boundary pressure sensor: ;in: : Heat map abnormal aggregation direction vector; :sensor The fluctuation direction vector of (can be extracted by the principal component direction); : Indicates the consistency between the pressure fluctuation direction and the abnormal aggregation trend, with a value range of .when ( Setting a threshold for experience) to determine the existence of Towards The potential abnormal behavior transmission trend of the network is established. If the consistency matching threshold condition is met, a directed link is established between the two grid cells: This link records the possible transmission direction and directional consistency of the anomaly diffusion. The topological feature attenuation factor is introduced to further combine the topological features of the second channel in the dual channel and extract the average topological integrity difference value of the boundary users between the two grids: ;in: :user The second channel eigenvector corresponds to the scalar component of topological integrity; : indicates the number of boundary matching user pairs; : Indicates the degree of structural difference. The larger the value, the more abrupt the topological transition and the higher the attenuation. Calculate the strength of the conductive link: ;in: : Cross-grid abnormal characteristic conduction intensity; : Parameter to adjust the attenuation sensitivity (usually set to a positive value). Generate abnormal characteristic conduction path data set The transmission links (with output thresholds set) are uniformly output to form a directed weighted graph structure: ;in: : A set of cross-grid anomaly conduction paths; each edge represents the direction and strength of the conduction path of the anomaly feature in the physical water supply grid.

[0039] Specifically, a multi-layer graph convolutional network is used to analyze the diffusion direction of abnormal features in the conduction path, and the specific process of reversely tracing the water pressure attenuation mutation point along the physical direction of the water supply pipeline is as follows: the cross-grid abnormal conduction path dataset is loaded into the graph convolution processing module, the bottom-level convolution operation is performed to extract the conduction direction characteristics within the grid, and the top-level convolution operation is performed to capture the cross-grid cascade conduction law; the water pressure loss gradient value of each pipe section is calculated in reverse along the physical connection direction of the water supply pipeline, the sudden inflection point position of the water pressure loss gradient curve is detected, and the inflection point position is mapped to the three-dimensional coordinates of the building in combination with the pipeline network spatial topology database.

[0040] In this implementation, a multi-layer graph convolutional network analyzes the diffusion direction of abnormal features, loads the abnormal conduction path dataset, constructs a graph structure to input the cross-grid abnormal feature conduction path dataset, and ; Construct a weighted directed graph ;in: Represents a set of grid cell nodes; represents a set of directed edges, corresponding to abnormal conduction links; is the edge weight matrix, and the weight is the conduction intensity The underlying graph convolution operation extracts local conduction features and inputs a dual-channel feature vector to each node. ; Apply graph convolution operation: ;in: Indicates the Tier The hidden state of the node; Representation node The set of neighbor nodes of Represents a slave node To Node The edge weight of Representation node The degree of (i.e., the number of neighbor nodes); Indicates the The weight matrix of the layer; Represents an activation function (e.g., ReLU). This operation fuses the features of adjacent nodes to achieve effective aggregation of local abnormal conduction information. Reversely track the location of the water pressure attenuation mutation point and reversely calculate the water pressure loss gradient along the physical connection direction of the water supply pipeline to the water supply network section sequence. ; Calculate the pressure gradient of adjacent pipe sections based on real-time pressure measurement point data: ;in: Indicates the pressure of the kth pipe section; Represents the water pressure loss gradient along the physical water flow direction. The gradient sequence is used to detect the inflection point of the water pressure loss gradient mutation. ; Second-order difference analysis locates the mutation point: ; The conditions for determining the mutation point are: ;in: Represents the preset gradient mutation threshold. Spatial mapping to building 3D coordinates uses the pipe network spatial topology database to map the corresponding pipe segment position of the mutation point to the specific building 3D coordinates, achieving accurate positioning in physical space.

[0041] Specifically, the specific process of outputting leakage point identification and abnormal water use behavior pattern marking specific to the location of building components is as follows: according to the building's three-dimensional coordinates of the inflection point of the water pressure loss gradient mutation, the engineering component database is retrieved, the leakage point identification of the specific building component is output and associated with the flange sealing surface or valve interface number; the dual-channel feature change pattern in the conduction path dataset is simultaneously analyzed, when the waveform similarity distortion value of the first channel exceeds the threshold, the equipment sealing failure is marked, when the topological connection integrity fracture value of the second channel exceeds the threshold, the pipeline structure damage is marked, and when the dual-channel features are abnormal at the same time, the user operation function failure is marked.

[0042] In this implementation, building components are located based on the inflection points of sudden changes in water pressure loss gradients. First, the sudden inflection points detected by the water pressure loss gradient in the previous step, combined with the building's three-dimensional coordinate information, are searched against the corresponding spatial topology database of the pipe network. This database records the actual locations of engineering components corresponding to each pipe segment, such as flange sealing surfaces and valve interfaces. Through this mapping, the leak location is precisely pinpointed to a specific building component, achieving precise physical location and providing a clear target for subsequent repairs. The leak point identifier is output and associated with specific component information. The system automatically generates a leak point identifier, including the component name, type, and unique number (such as the flange sealing surface number or valve interface number), allowing maintenance personnel to quickly identify and take appropriate measures. Simultaneously analyzing the characteristic change patterns of two channels, the distortion value of the first channel (waveform similarity) is analyzed: this channel reflects the degree of similarity between the water usage waveform and the typical pattern. If the waveform similarity distortion value exceeds a preset threshold, it indicates an abnormal waveform morphology, possibly caused by abnormal pressure fluctuations caused by poor equipment sealing. The system then marks the leak as "equipment seal failure." Second channel (topological connection integrity) fracture value analysis: This channel measures the integrity of connections at waveform turning points. Fracture values ​​exceeding the threshold indicate a fracture or damage in the pipe network structure, disrupting the topological relationship of the water usage waveform. The system therefore labels this as "pipe network structure damage." Simultaneous abnormality of dual-channel characteristics: When the abnormality indicators of both channels exceed the threshold, it indicates that the anomaly is not limited to a single physical fault and may involve abnormal user-side operation (such as illegal operation or equipment failure). The system therefore labels this as "user operation dysfunction." Leakage location and anomaly classification based on dual analysis of physical measurements and behavioral characteristics improve the accuracy and reliability of water use anomaly detection, providing solid technical support for the precise maintenance and management of smart water supply systems.

[0043] In summary, this application has at least the following effects:

[0044] The urban residents' water use behavior classification method based on smart water meters realizes dynamic analysis of real-time sampling data based on smart water meters, can accurately capture key events in water use behavior, and improve the timeliness and accuracy of event identification; through adaptive noise elimination and water hammer effect filtering processing, it effectively suppresses the signal interference caused by pipe network pressure fluctuations, and ensures the authenticity and stability of waveform characteristics; uses flow gradient distribution characteristics to adaptively adjust the curvature detection sensitivity, accurately locates the key turning points of the waveform, and improves the fine-grained recognition ability of water use behavior classification; combines topological consistency comparison technology to achieve efficient matching of waveform features and behavior pattern libraries, and generate multi-dimensional dual-channel feature vectors, Enhance the expressive power of the classification model; divide the metering grid based on the distribution of valves in the building water supply network, integrate the pressure gradient-driven distribution clustering analysis to achieve spatially refined identification of abnormal water use behavior; combine real-time pressure sensor data to build cross-grid abnormal conduction paths, and dynamically capture the propagation law of abnormal behavior in the network; deeply analyze the abnormal conduction characteristics through multi-layer graph convolutional networks, and combine the reverse tracking of physical water pressure gradient mutation points to achieve precise spatial positioning of leakage points; multi-channel abnormal features jointly determine different types of water use anomalies, achieve accurate classification and labeling of leakage, equipment failure and user operation anomalies, and improve the automation and intelligence level of intelligent water management.

[0045] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0046] The present invention is described with reference to flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0047] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0048] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0049] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0050] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A classification method for urban residents’ water use behavior based on smart water meters, characterized by: The following steps are involved: S1. Calculate the flow difference between adjacent sampling points in real time through the smart water meter terminal to generate a flow change direction sequence. When the duration of the direction sequence's continuous stability exceeds the user's historical extreme value range for the same period, trigger an event capture and extract the flow change gradient distribution characteristics and timestamp density characteristics within the event. S2. Adaptively configure the noise elimination window size based on the timestamp density feature, perform water hammer filtering on the water use event waveform, locate the waveform curvature mutation points based on the flow rate gradient distribution characteristics, construct a key turning point sequence, and compare the node connectivity of the sequence with typical waveform templates in the behavior pattern library for topological consistency, generating a dual-channel feature vector that describes waveform similarity and topological connectivity integrity. S3. The building water supply network is divided into independent metering grid units based on the physical valve distribution. A pressure gradient-driven distribution cluster analysis is performed on the dual-channel feature vectors of users within the same grid unit. This analysis is combined with real-time fluctuation data from the network pressure sensors to construct cross-grid abnormal feature transmission paths. S4. A multi-layer graph convolutional network is used to analyze the diffusion direction of abnormal features in the conduction path, and the water pressure attenuation mutation point is tracked in reverse along the physical direction of the water supply pipeline. The leakage point identification and abnormal water use behavior pattern markers specific to the location of the building components are output.

2. The method for classifying urban residents' water usage behavior based on smart water meters according to claim 1 is characterized by: Step S1 includes the following steps: The smart water meter terminal calculates the flow difference between adjacent sampling points in real time to generate a water flow change direction sequence, marking flow increases as positive states, flow decreases as negative states, and stable flow as zero states; When the direction sequence maintains the same state for a duration exceeding the maximum duration threshold of the user's history during the same period, the event capture mechanism is triggered; The gradient distribution characteristics of flow changes within water use events are extracted, the distribution law of flow change rate in each period is calculated through sliding window analysis, the timestamp density characteristics are extracted, and the distribution characteristics of the number of sampling points in the unit time window are counted.

3. The method for classifying urban residents' water usage behavior based on smart water meters according to claim 1 is characterized by: The specific process of adaptively configuring the noise elimination window size based on the timestamp distribution density characteristics and performing water hammer effect filtering on the water use event waveform is as follows: Dynamically matching the noise elimination window size based on the timestamp distribution density characteristics, using a compact filtering window in areas with high timestamp density and an extended filtering window in areas with low timestamp density; Water hammer effect elimination processing is performed based on the hydraulic characteristics of water-using appliances. A high-frequency oscillation suppression mode is used for pulse waveforms such as faucet start and stop. A low-frequency damping mode is used for long-period waveforms such as bathing equipment. This eliminates waveform distortion caused by sudden changes in pipe network pressure and retains the core morphological characteristics of water use behavior.

4. The method for classifying urban residents' water usage behavior based on smart water meters according to claim 3 is characterized by: The specific process of locating the waveform curvature mutation point by combining the flow change gradient distribution characteristics and constructing the key turning point sequence is as follows: The curvature detection sensitivity is adaptively adjusted based on the gradient distribution characteristics of flow changes. In areas with low change intensity, the detection sensitivity is reduced to avoid false triggering of small fluctuations, while in areas with high change intensity, the detection sensitivity is increased to accurately capture state transition points. The key turning points of the waveform are located through curvature extreme value analysis, and an orderly turning sequence including starting point, peak point, valley point and ending point is constructed, and the time coordinates and type attributes of each turning point are recorded.

5. The method for classifying urban residents' water usage behavior based on smart water meters according to claim 4 is characterized by: The specific process of comparing the topological consistency of the node connection relationship of the sequence with the typical waveform template in the behavior pattern library and generating a dual-channel feature vector describing the waveform similarity and topological connection integrity is as follows: Establish an adjacency matrix of the connection relationship between the turning sequence nodes, and perform three-level topological consistency verification with the typical waveform template in the behavior pattern library. First, match the key turning point type sequence consistency, then verify the continuity of the connection relationship between adjacent turning points, and finally detect the integrity of the special topological structure. A first channel feature vector is generated to independently characterize the similarity of waveform morphology, and a second channel feature vector is generated simultaneously to quantify the integrity of the topological connection relationship, forming a dual-channel feature vector.

6. The method for classifying urban residents' water usage behavior based on smart water meters according to claim 1 is characterized by: The specific process of dividing the independent metering grid units according to the physical valve distribution of the building water supply network and performing pressure gradient-driven distribution cluster analysis on the dual-channel feature vectors of users in the same grid unit is as follows: Divide the building water supply network into independent metering grid units according to the actual installation location of the physical valves, and each grid unit forms a closed hydraulic monitoring area; Collect the dual-channel feature vectors of all users in the same grid unit and calculate the real-time pressure difference vector direction between adjacent user nodes; The center position of user behavior clustering is dynamically adjusted according to the direction of the pressure difference vector, and the user behavior association weight coefficient is calculated by integrating the waveform similarity characteristics of the first channel and the topological connection integrity characteristics of the second channel; a spatial distribution heat map of abnormal water use behavior within the grid unit is generated.

7. The method for classifying urban residents' water usage behavior based on smart water meters according to claim 6 is characterized by: The specific process of constructing cross-grid abnormal characteristic transmission paths by combining real-time fluctuation data from pipe network pressure sensors is as follows: Acquire real-time fluctuation data from pressure sensors at the boundaries of adjacent grid cells and detect the degree of match between pressure fluctuation characteristics and abnormal behavior cluster areas in the spatial distribution thermal map; When the matching degree exceeds the set threshold, a directional conduction link is established from the high-pressure area to the low-pressure area, the characteristic value of the second channel topological connection integrity is extracted, and the attenuation correction of the conduction link strength is performed according to the characteristic value to generate a cross-grid abnormal conduction path dataset.

8. The method for classifying urban residents' water usage behavior based on smart water meters according to claim 1 is characterized by: The specific process of analyzing the diffusion direction of abnormal features in the conduction path through a multi-layer graph convolutional network and tracing the water pressure attenuation mutation point in reverse along the physical direction of the water supply pipeline is as follows: Load the cross-grid abnormal conduction path dataset into the graph convolution processing module, perform the bottom-level convolution operation to extract the grid internal conduction direction characteristics, and perform the top-level convolution operation to capture the cross-grid cascade conduction law; The water pressure loss gradient value of each pipe section is calculated in reverse along the physical connection direction of the water supply pipeline, the sudden inflection point position of the water pressure loss gradient curve is detected, and the inflection point position is mapped to the three-dimensional coordinates of the building in combination with the pipeline network spatial topology database.

9. The method for classifying urban residents' water usage behavior based on smart water meters according to claim 8 is characterized by: The specific process of outputting leakage point identification and abnormal water use behavior pattern markings specific to the location of building components is as follows: Search the engineering component database based on the building's 3D coordinates of the inflection point where the water pressure loss gradient suddenly changes, output the leakage point identification of the specific building component, and associate it with the flange sealing surface or valve interface number; The dual-channel feature change patterns in the conduction path dataset are analyzed synchronously. When the waveform similarity distortion value of the first channel exceeds the threshold, the device sealing failure is marked. When the topological connection integrity fracture value of the second channel exceeds the threshold, the pipeline structure damage is marked. When the dual-channel features are abnormal at the same time, the user operation function failure is marked.

Images