An ultrasonic water meter intelligent monitoring system based on edge computing

Abstract

The application discloses an ultrasonic water meter intelligent monitoring system based on edge computing, comprising: a collection module for collecting and preprocessing multi-source measurement signals of the ultrasonic water meter; a feature extraction module for extracting a feature vector and generating a joint feature representation; an end-side identification module for multi-layer feature extraction and multi-branch determination through an end-side intelligent identification model; an event generation module for determining a risk level and a monitoring event type and generating a monitoring event record; a backhaul module for matching a trigger condition and sending a report data frame in a trigger state; a cache module for maintaining end-side identification operations and updating the monitoring event record in a communication unreachable state; and an update module for synchronizing the monitoring event record and updating the end-side intelligent identification model in a communication reachable state. The application combines multi-source features and end-side intelligent identification, realizes real-time and robust determination of ultrasonic water meter anomaly monitoring, and has the advantages of high precision, high reliability and communication self-adaptation.

Description

Technical Field

[0001] This invention relates to the field of intelligent water meter monitoring, and in particular to an ultrasonic water meter intelligent monitoring system based on edge computing. Background Technology

[0002] Existing ultrasonic water meter monitoring systems typically rely on centralized servers for data processing. Ultrasonic time-difference signals, echo amplitude signals, and flow measurement signals are only simply collected at the terminal, lacking effective end-side feature extraction and intelligent analysis capabilities. Because the processing link mainly depends on a remote platform, the system cannot continuously generate effective monitoring results when communication quality fluctuates or the link is unavailable, leading to delays in identifying critical events such as leakage, pipe burst trends, and abnormal backflow. Furthermore, existing systems often use single features or single-modal data as the basis for judgment, making it difficult to achieve joint analysis of acoustic characteristics and flow time-series characteristics, thus limiting monitoring accuracy.

[0003] Meanwhile, traditional monitoring systems lack event-level caching and synchronization mechanisms. After communication is restored, they cannot accurately retransmit events in chronological order, and it is difficult to dynamically correct the end-side identification model based on event confirmation information. This leads to the model deviating from real-world conditions over a long period, resulting in a gradual decline in monitoring performance. Existing technologies also generally lack structures such as adaptive convolution, cross-modal feature fusion, and multi-branch inference at the end-side, making it unable to effectively cope with complex water usage behaviors and multi-dimensional acoustic pattern changes.

[0004] Therefore, how to provide an intelligent monitoring system for ultrasonic water meters based on edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent monitoring system for ultrasonic water meters based on edge computing. This invention integrates multi-source features and edge-side intelligent identification to achieve real-time robust judgment of anomalies in ultrasonic water meters, and has the advantages of high precision, high reliability and adaptive communication.

[0006] An edge computing-based intelligent monitoring system for ultrasonic water meters according to an embodiment of the present invention includes: The acquisition module is used to acquire and preprocess multi-source measurement signals from ultrasonic water meters to generate standardized measurement data sequences; The feature extraction module is used to extract acoustic feature vectors and flow time-series feature vectors, generate fused feature vectors based on feature reliability, and perform joint splicing to generate joint feature representations; The end-side identification module is used to perform tensor quantization mapping on the joint feature representation and input it into the end-side intelligent identification model to generate a water meter status result set; The event generation module is used to determine the risk level and monitoring event type based on the water meter status result set and generate monitoring event records; The feedback module is used to perform trigger condition matching and feedback status determination, and to send up report data frames through the communication link when the trigger status is triggered. The caching module is used to maintain end-side identification operations and update monitoring event records in the local event cache when communication is unreachable. The update module is used to synchronize the monitoring event records in the local event cache when communication is reachable and update the end-side intelligent recognition model based on the event confirmation information.

[0007] Optionally, modules can be integrated using the following methods: Multi-source measurement signals from ultrasonic water meters are collected, preprocessed, and standardized measurement data sequences are generated. Acoustic feature vectors and flow time-series feature vectors are extracted from the ultrasonic water meter end based on standardized measurement data sequences. Based on feature reliability, fused feature vectors are generated and joint splicing is performed to generate joint feature representations. After performing tensor quantization mapping on the joint feature representation, it is input into the end-side intelligent recognition model for multi-layer feature extraction and multi-branch decision processing to generate a water meter status result set. Based on the water meter status result set, determine the risk level and monitoring event type, generate monitoring event records on the terminal side and store them in the local event cache area; The system performs trigger condition matching and feedback status determination on the monitoring event records, and sends reporting data frames through the communication link when the trigger status is triggered. When the communication link is in a communication unreachable state, maintain the end-side identification operation and synchronously update the monitoring event records in the local event buffer; When the communication link is reachable, the monitoring event records in the local event cache are synchronized to the remote monitoring platform in batches according to the event timestamp order, and the end-side intelligent identification model is updated based on the event confirmation information to generate a new water meter status result set.

[0008] Optionally, the multi-source measurement signals include ultrasonic time difference signals, echo amplitude signals, temperature signals, and flow measurement signals, and the preprocessing includes noise reduction processing, time sequence alignment, outlier removal, and time window segmentation processing.

[0009] Optionally, the generation of the joint feature representation includes the following specific steps: The ultrasonic water meter receives a standardized measurement data sequence at the end and performs time alignment and segmentation according to a preset sampling period to generate acoustic analysis data segments and flow analysis data segments. Based on the acoustic analysis data segment, time-domain statistical extraction, frequency-domain statistical extraction, and morphological stability extraction of the acoustic signal are performed, and the extracted features are integrated to obtain the acoustic feature vector. Based on the traffic analysis data segment, perform trend feature extraction, fluctuation feature extraction, periodic feature extraction and abnormal jump feature extraction of traffic time series, and perform feature integration processing on the extracted features to obtain the traffic time series feature vector; Based on the reliability of acoustic features and the reliability of traffic flow time-series features, acoustic fusion weight value and time-series fusion weight value are calculated, and the acoustic feature vector and traffic flow time-series feature vector are weighted and fused based on the weight values ​​to obtain the fused feature vector; The fused feature vector is dimensionally normalized and scaled, and then jointly concatenated with the acoustic feature vector and the traffic temporal feature vector to generate a joint feature representation.

[0010] Optionally, the generation of the water meter status result set includes the following specific steps: The joint feature representation is received at the end of the ultrasonic water meter and the input format conversion process is performed to map the joint feature representation into the joint feature tensor of the end-side intelligent recognition model. The joint feature tensor is input into the intelligent recognition model at the end. In the convolutional layer, the directionality adjustment processing of the convolutional kernel structure is performed. The sampling direction and sensing direction of the convolutional kernel are adjusted according to the direction component representing the direction information of the sound wave in the joint feature tensor, and the direction adjustment convolutional kernel is generated. The edge-side intelligent recognition model is a multi-branch recognition model built on the LCNN architecture, including an input structure, a convolutional layer structure, a feature extraction structure and a multi-branch output structure. An adaptive convolutional kernel adjustment mechanism for acoustic direction information is introduced in the convolutional layer structure, and a cross-modal joint feature processing structure is configured in the middle layer of the feature extraction structure. The orientation field weights are calculated based on the orientation field information representing the intensity distribution of the incident direction of the sound wave in the joint feature tensor, and then the orientation field weights are loaded into the convolution channels corresponding to the orientation adjustment convolution kernel to generate an orientation-weighted convolution feature map. Multi-layer convolutional feature extraction and hierarchical feature convergence processing are performed on the directional weighted convolutional feature map to obtain acoustic pattern feature map and flow temporal pattern feature map; In the intermediate layer, cross-modal feature fusion processing is performed on the acoustic mode feature map and the traffic temporal mode feature map to generate a cross-modal fused feature map; The cross-modal fusion feature map is processed by feature compression, feature normalization and decision branch mapping to obtain the input features of each branch. Based on the input features of each branch, end-side identification and reasoning processing is performed to output leakage judgment results, pipe burst trend judgment results, backflow anomaly judgment results and equipment operation status judgment results. The four types of judgment results are then aggregated to generate a water meter status result set.

[0011] Optionally, the generation of the monitoring event record includes the following specific steps: Read the water meter status result set at the end and perform judgment field parsing processing on the leakage judgment result, pipe burst trend judgment result, backflow anomaly judgment result and equipment operation status judgment result; Based on the parsed judgment fields, the event candidate generation process is performed, and each judgment field is respectively assigned to leakage event candidates, pipe burst trend event candidates, backflow abnormality event candidates, and equipment operation event candidates. Based on the risk value mapping table, the judgment level of each event candidate is mapped to a risk score, and a weighted calculation is performed on the risk scores according to the corresponding judgment confidence level to generate a comprehensive risk score; Based on the comprehensive risk score, the risk level is determined by locating the corresponding risk level within the risk level range. Then, based on the risk score of each candidate event, the event type is determined and the monitoring event type is generated. Obtain the current time on the client side to form an event timestamp, and aggregate the parsed judgment fields, event timestamp, risk level and monitoring event type to generate a monitoring event record, and store it in the local event cache.

[0012] Optionally, the generation of the reported data frame includes the following specific steps: Read the monitoring event records on the device side, perform back transmission trigger condition matching processing according to the monitoring event type and risk level, and determine the back transmission trigger status based on the matching result; Based on the current trigger status of the return, the return status determination process is executed. If the return status is triggered, the summary feature data generation process is entered; if the return status is not triggered, the current return process is terminated. In the triggered state, feature components corresponding to the monitoring event type are selected from the acoustic feature vector, flow time-series feature vector and joint feature representation associated with the monitoring event record to generate summary feature data; The monitoring event records and summary feature data are sorted by field, length-marked, and segmented according to the data frame structure of the communication link to generate a reporting data frame; The monitoring event records and corresponding summary feature data are reported to the remote monitoring platform by sending reporting data frames through the communication link.

[0013] Optionally, the communication unreachable state includes a link unavailable state and a communication quality insufficiency state. The determination of the communication unreachable state includes: reading the link availability field of the communication link, performing a state comparison process based on the availability identifier and the unavailability identifier of the availability field, and mapping the unavailability identifier to the link unavailable state; performing quality value processing on the communication quality indicators of the communication link, and performing a numerical comparison process between the communication quality value and the communication quality threshold, and determining the communication quality insufficiency state when the communication quality value is lower than the communication quality threshold; and determining that the communication link is in a communication unreachable state when either the link unavailable state or the communication quality insufficiency state is met. The communication quality indicators include signal strength, number of bit errors, and round-trip time. Optionally, the update of the edge-side intelligent recognition model includes: reading the event confirmation information returned by the remote monitoring platform, and writing the event confirmation label field, event result field, and event deviation information field into the confirmation data set in the order of the fields; performing deviation calculation processing on the event deviation information field in the confirmation data set to generate a deviation sequence; performing weight adjustment processing on the convolutional layer structure of the edge-side intelligent recognition model according to the deviation sequence, and performing component update operation according to the deviation ratio on the weight components corresponding to the deviation positions in the convolutional kernel weight matrix; performing feature path correction processing on the feature extraction structure of the edge-side intelligent recognition model, and performing channel weight increase or decrease processing on the feature channels corresponding to the deviation sequence according to the deviation direction; performing judgment boundary adjustment processing on the multi-branch output structure of the edge-side intelligent recognition model, and performing numerical adjustment processing on the judgment threshold of each branch according to the deviation signal in the deviation sequence; and generating the updated edge-side intelligent recognition model after the weight adjustment processing, feature path correction processing, and judgment boundary adjustment processing are all written into the edge-side intelligent recognition model.

[0014] The beneficial effects of this invention are: This invention transforms the traditional monitoring mode, which relies on remote platforms, into a local real-time intelligent processing mode by constructing an edge computing system at the ultrasonic water meter end, integrating data acquisition, feature extraction, intelligent recognition, event generation, cache management, and model updating. By performing high-dimensional feature extraction, fusion weight generation, and joint splicing processing on acoustic feature vectors and flow time-series feature vectors at the end, a joint feature representation capable of simultaneously characterizing hydraulic state changes and sound wave propagation characteristics is formed, giving the recognition model higher sensitivity and sparse event detection capabilities. Based on a multi-branch recognition structure, the system can simultaneously and independently determine leakage, pipe burst trends, backflow anomalies, and equipment operating status at the end, significantly improving the precision and comprehensiveness of monitoring and effectively reducing monitoring errors caused by traditional systems with single features and delayed judgments.

[0015] This invention further introduces a communication state adaptive mechanism. By jointly determining the link availability field and communication quality indicators, the system can maintain stable operation of the end-side identification process even when communication is unreachable, and continuously write monitoring event records to the local event cache, achieving uninterrupted information and no monitoring loss. After the communication link is restored to a reachable state, the system batch synchronizes monitoring event records to the remote monitoring platform according to the event timestamp order, ensuring data time sequence consistency and the integrity of retransmission, avoiding the event loss, out-of-order, and missed reporting problems caused by communication interruptions in traditional systems. Simultaneously, this invention utilizes the event confirmation information returned by the remote platform to perform convolutional layer weight adjustment, feature path correction, and multi-branch decision boundary optimization on the end-side intelligent identification model, enabling the model to continuously iterate and update based on real feedback, maintaining long-term identification accuracy and stability.

[0016] In summary, this invention enables the ultrasonic water meter monitoring mode to evolve from traditional passive acquisition and centralized identification to end-side intelligence, real-time judgment, and adaptive updates. Through multi-source feature fusion, cross-modal convolutional processing, event-level buffering and backhaul mechanisms, and dynamic model correction based on deviation information, the system achieves stronger environmental robustness, communication adaptability, and identification stability during monitoring. Compared to existing technologies, this invention significantly improves the timeliness of leakage identification, the foresight of pipe burst trend judgment, and the accuracy of backflow anomaly detection, while ensuring the integrity and continuity of the monitoring data chain even in the event of communication failures. Therefore, it provides a more reliable and practically valuable technical solution for the field of smart water management. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for an ultrasonic water meter intelligent monitoring system based on edge computing proposed in this invention. Figure 2 This is a flowchart illustrating the generation process of a water meter status result set in an ultrasonic water meter intelligent monitoring system based on edge computing, as proposed in this invention. Figure 3 This is a flowchart illustrating the event processing and model update process of an edge computing-based intelligent monitoring system for ultrasonic water meters proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1-3An edge computing-based intelligent monitoring system for ultrasonic water meters includes: The acquisition module is used to acquire and preprocess multi-source measurement signals from ultrasonic water meters to generate standardized measurement data sequences; The feature extraction module is used to extract acoustic feature vectors and flow time-series feature vectors, generate fused feature vectors based on feature reliability, and perform joint splicing to generate joint feature representations; The end-side identification module is used to perform tensor quantization mapping on the joint feature representation and input it into the end-side intelligent identification model to generate a water meter status result set; The event generation module is used to determine the risk level and monitoring event type based on the water meter status result set and generate monitoring event records; The feedback module is used to perform trigger condition matching and feedback status determination, and to send up report data frames through the communication link when the trigger status is triggered. The caching module is used to maintain end-side identification operations and update monitoring event records in the local event cache when communication is unreachable. The update module is used to synchronize the monitoring event records in the local event cache when communication is reachable and update the end-side intelligent recognition model based on the event confirmation information.

[0020] In this embodiment, the modules are interconnected using the following method: Multi-source measurement signals from ultrasonic water meters are collected, preprocessed, and standardized measurement data sequences are generated. Acoustic feature vectors and flow time-series feature vectors are extracted from the ultrasonic water meter end based on standardized measurement data sequences. Based on feature reliability, fused feature vectors are generated and joint splicing is performed to generate joint feature representations. After performing tensor quantization mapping on the joint feature representation, it is input into the end-side intelligent recognition model for multi-layer feature extraction and multi-branch decision processing to generate a water meter status result set. Based on the water meter status result set, determine the risk level and monitoring event type, generate monitoring event records on the terminal side and store them in the local event cache area; The system performs trigger condition matching and feedback status determination on the monitoring event records, and sends reporting data frames through the communication link when the trigger status is triggered. When the communication link is in a communication unreachable state, maintain the end-side identification operation and synchronously update the monitoring event records in the local event buffer; When the communication link is reachable, the monitoring event records in the local event cache are synchronized to the remote monitoring platform in batches according to the event timestamp order, and the end-side intelligent identification model is updated based on the event confirmation information to generate a new water meter status result set.

[0021] In this embodiment, the multi-source measurement signals include ultrasonic time difference signals, echo amplitude signals, temperature signals, and flow measurement signals. The preprocessing includes noise reduction, time alignment, outlier removal, and time window segmentation.

[0022] In this embodiment, the generation of the joint feature representation includes the following specific steps: The ultrasonic water meter receives a standardized measurement data sequence at the end and performs time alignment and segmentation according to a preset sampling period to generate acoustic analysis data segments and flow analysis data segments. The preset sampling period is determined based on the ultrasonic propagation path length, transducer excitation timing, and the internal flow stability range of the water meter. Based on the acoustic analysis data segment, time-domain statistical extraction, frequency-domain statistical extraction, and morphological stability extraction of the acoustic signal are performed, and the extracted features are integrated to obtain the acoustic feature vector. The time-domain statistical extraction includes: reading the acoustic wave amplitude sequence of the acoustic analysis data segment according to the sampling order; calculating the amplitude change based on the amplitude difference between adjacent sampling points; generating the acoustic wave energy accumulation based on the sum of the amplitudes of all sampling points in the acoustic analysis data segment; and calculating the envelope fluctuation based on the difference between adjacent envelope sampling points extracted from the envelope curve of the amplitude sequence. Numerical normalization, time dimension alignment, and sequence splicing are then performed on the amplitude change, acoustic wave energy accumulation, and envelope fluctuation to form time-domain statistical features characterizing the acoustic wave propagation state. The frequency domain statistical extraction includes: performing a frequency domain transformation operation on the acoustic analysis data segment to generate a frequency domain energy distribution sequence; generating the position of the main energy concentration interval based on the frequency position corresponding to the energy peak in the frequency domain energy distribution sequence; generating the frequency domain energy diffusion based on the distribution range of frequency domain energy in different frequency intervals; and generating the frequency domain energy peak change based on the trend of the frequency domain energy distribution sequence peak changing with frequency. Numerical mapping, frequency axis unification, and feature arrangement processing are then performed on the position of the main energy concentration interval, the frequency domain energy diffusion, and the frequency domain energy peak change to form frequency domain statistical features characterizing the sound wave energy distribution state. The morphological stability extraction includes: performing morphological segmentation on the acoustic analysis data segment to divide the acoustic wave sequence into multiple adjacent acoustic wave morphological segments; calculating the acoustic wave morphological offset based on the difference between corresponding sampling points of adjacent acoustic wave morphological segments; generating the acoustic wave morphological stability interval ratio based on the proportion of sampling intervals that maintain waveform stability in each acoustic wave morphological segment; and calculating the acoustic wave morphological continuity change based on the curve change trend of adjacent acoustic wave morphological segments. Segment-level statistics, morphological feature alignment, and component aggregation are then performed on the acoustic wave morphological offset, the acoustic wave morphological stability interval ratio, and the acoustic wave morphological continuity change to form morphological stability features characterizing the stability state of the acoustic mode. Based on the traffic analysis data segment, perform trend feature extraction, fluctuation feature extraction, periodic feature extraction and abnormal jump feature extraction of traffic time series, and perform feature integration processing on the extracted features to obtain the traffic time series feature vector; The trend feature extraction includes: reading the flow sequence of the flow analysis data segment in chronological order, calculating the flow change increment based on the flow difference between adjacent sampling points, generating the flow trend change amount based on the cumulative change trend of the flow change increment, generating the flow mean trend amount based on the mean change of the flow sequence within a preset time sliding window, and performing sequence alignment processing and feature component normalization processing on the flow trend change amount and the flow mean trend amount to form a trend feature characterizing the changing state of water use behavior; The fluctuation feature extraction includes: performing sliding window variance calculation on the flow analysis data segment to generate flow fluctuation variance; generating rapid flow fluctuation based on the absolute value of the difference between adjacent sampling points in the flow sequence; generating flow fluctuation amplitude based on the range of change of continuous fluctuation intervals in the flow sequence; and performing numerical normalization and time axis consistency processing on the flow fluctuation variance, rapid flow fluctuation, and flow fluctuation amplitude to form fluctuation features characterizing transient impact change states. The periodic feature extraction includes: performing a periodic detection operation on the flow analysis data segment, generating a flow repetition period based on the repetitive change pattern of the flow sequence within multiple sliding windows, generating a flow period peak position based on the peak position of the repetitive change pattern, generating a flow period duration based on the duration of the periodic change pattern, and performing feature extraction and alignment processing and component recombination processing on the flow repetition period, flow period peak position, and flow period duration to form a periodic feature characterizing the continuous deviation change state. The abnormal jump feature extraction includes: performing a difference calculation operation on the flow analysis data segment, generating a flow jump amplitude based on the jump difference between adjacent sampling points, generating a flow jump duration based on the continuous time segment in which the flow jump occurs, generating a flow jump frequency based on the frequency of the flow jump in the sequence, and performing segment-level statistical processing and feature aggregation processing on the flow jump amplitude, flow jump duration and flow jump frequency to form abnormal jump features used to characterize the abnormal change state of water use behavior; Based on the reliability of acoustic features and the reliability of traffic flow time-series features, acoustic fusion weight value and time-series fusion weight value are calculated, and the acoustic feature vector and traffic flow time-series feature vector are weighted and fused based on the weight values ​​to obtain the fused feature vector; The acoustic feature reliability is obtained by: performing feature component sampling segment division operations on the time-domain statistical features, frequency-domain statistical features, and morphological stability features in the acoustic feature vector, dividing each type of feature component into multiple continuous sampling segments of fixed length; performing component difference calculation operations on each sampling segment to generate a segment difference sequence; performing variance calculation operations on the segment difference sequence to generate a segment difference variance; performing numerical range calculation operations on each sampling segment to generate a segment numerical interval; performing difference matrix calculation operations between segments of the same type of feature component to generate a segment correlation matrix; and performing normalization processing and component-level aggregation processing on the segment difference variance, segment numerical interval, and segment correlation matrix to generate the acoustic feature reliability. The reliability of the traffic flow time series features is obtained by: performing feature component sampling segment division operations on the trend features, fluctuation features, periodic features, and abnormal jump features in the traffic flow time series feature vector, dividing each feature component into multiple continuous sampling segments of fixed length; performing adjacent component difference calculation operations on each sampling segment to generate a segment change difference sequence, and performing range calculation operations on the segment change difference sequence to generate a segment change range quantity; performing numerical distribution statistical operations on each sampling segment to generate a segment distribution interval quantity; performing difference sequence comparison operations on segments of the same category of feature components to generate a segment correlation quantity; and performing normalization processing and feature aggregation processing on the segment change range quantity, segment distribution interval quantity, and segment correlation quantity to generate the traffic flow time series feature reliability. The calculation of the acoustic fusion weight value and the temporal fusion weight value includes: performing segmentation operations on acoustic feature reliability and flow temporal feature reliability respectively; performing segment mean calculation, segment fluctuation calculation and segment correlation calculation on each segment in sequence; wherein the segment fluctuation of acoustic feature reliability is generated by variance calculation and the segment fluctuation of flow temporal feature reliability is generated by variation range calculation; and performing normalization processing and feature aggregation processing on the generated segment mean, segment fluctuation and segment correlation to form acoustic fusion weight value and temporal fusion weight value respectively. The fused feature vector is dimensionally normalized and scaled, and then jointly concatenated with the acoustic feature vector and the traffic temporal feature vector to generate a joint feature representation.

[0023] In this embodiment, the generation of the water meter status result set includes the following specific steps: The joint feature representation is received at the end of the ultrasonic water meter and the input format conversion process is performed to map the joint feature representation into the joint feature tensor of the end-side intelligent recognition model. The process of obtaining the joint feature tensor includes: reading the acoustic feature vector and the flow time-series feature vector from the joint feature representation; performing component order rearrangement on the acoustic feature vector and the flow time-series feature vector according to their component arrangement order; performing component length alignment on the rearranged acoustic feature components and the flow time-series feature components to ensure that the two types of feature components maintain the same data length in the same dimension; performing dimension superposition on the length-aligned acoustic feature components and the flow time-series feature components to form a multi-dimensional feature matrix including multiple rows and columns of feature components; and performing tensor mapping on the multi-dimensional feature matrix to map the row dimension, column dimension, and channel dimension of the matrix to the three-dimensional tensor dimension corresponding to the input structure of the edge intelligent recognition model, thereby generating the joint feature tensor. The joint feature tensor is input into the intelligent recognition model at the end. In the convolutional layer, the directionality adjustment processing of the convolutional kernel structure is performed. The sampling direction and sensing direction of the convolutional kernel are adjusted according to the direction component representing the direction information of the sound wave in the joint feature tensor, and the direction adjustment convolutional kernel is generated. The edge-side intelligent recognition model is a multi-branch recognition model built on the LCNN architecture, including an input structure, a convolutional layer structure, a feature extraction structure and a multi-branch output structure. An adaptive convolutional kernel adjustment mechanism for acoustic direction information is introduced in the convolutional layer structure, and a cross-modal joint feature processing structure is configured in the middle layer of the feature extraction structure. The generation of the orientation-adjusted convolutional kernel includes: reading the orientation components representing the acoustic wave orientation information point by point in the joint feature tensor, constructing an orientation sampling sequence according to the arrangement order of the orientation components; performing orientation difference calculation on the orientation sampling sequence to generate an orientation difference sequence; performing summation and integer division on the difference of each sampling point with the difference of adjacent sampling points within the window according to the smoothing window, replacing the original sampling point difference with the result to form an orientation smoothing sequence; and adjusting the sampling row position and sampling column position of the original weight matrix of the convolutional kernel according to the orientation increment of each sampling point in the orientation smoothing sequence. Element-wise offset processing is performed, and the corresponding row and column offsets are adjusted for each sampling position according to the direction increment. The receptive area shape adjustment processing is performed on the convolution kernel weight matrix after offset processing. By performing continuous position rearrangement processing on each weight position in the convolution kernel according to the direction path represented by the direction smoothing sequence, the adjacent weight positions in the convolution kernel are arranged sequentially along the direction of the direction smoothing sequence. Boundary filling processing and matrix size correction processing are performed on the rearranged weight positions, so that the convolution kernel as a whole forms a direction-adjusted convolution kernel that extends along the direction of the direction smoothing sequence. The orientation field weights are calculated based on the orientation field information representing the intensity distribution of the incident direction of the sound wave in the joint feature tensor, and then the orientation field weights are loaded into the convolution channels corresponding to the orientation adjustment convolution kernel to generate an orientation-weighted convolution feature map. The generation of the orientation-weighted convolution feature map includes: performing orientation field sampling on the orientation field information representing the intensity distribution of the incident sound wave in the joint feature tensor, and constructing an orientation field sampling sequence according to the sampling order of the orientation field; performing amplitude reading processing on each sampling point in the orientation field sampling sequence to obtain the corresponding orientation field amplitude components, and performing channel weighting processing on the orientation field amplitude components according to the channel arrangement order of the convolutional layer to form an orientation field weighting sequence; performing channel element-wise reading processing on the convolution weights of each convolutional channel in the orientation-adjusted convolutional kernel, and performing multiplication operation processing on the convolution weights at the corresponding positions and the element-wise weights of the orientation field weighting sequence to generate an orientation-weighted convolution weight matrix; performing convolution operation processing on the orientation-weighted convolution weight matrix and the joint feature tensor, calculating the convolution response value of each convolutional region according to the convolution stride and the convolution region coverage relationship, and writing each convolution response value into the output matrix according to the convolution scan order to generate an orientation-weighted convolution feature map; Multi-layer convolutional feature extraction and hierarchical feature convergence processing are performed on the directional weighted convolutional feature map to obtain acoustic pattern feature map and flow temporal pattern feature map; The acoustic pattern feature map and the traffic flow time-series pattern feature map are obtained by: performing layer-by-layer convolution processing on the directional weighted convolution feature map; sampling regions of the directional weighted convolution feature map according to the convolution kernel size and stride of the convolution layer; and performing convolution operations on the sampled data of each region and the corresponding convolution kernel weights to generate a convolution response matrix; performing activation processing, downsampling processing, and channel concatenation processing on each convolution response matrix to form a multi-level convolution feature matrix; performing hierarchical feature convergence processing on the multi-level convolution feature matrix; performing feature component alignment, feature component superposition, and feature component aggregation processing on the output of each convolution layer according to the feature dimension; and mapping the converged feature matrix to the acoustic pattern feature map and the traffic flow time-series pattern feature map respectively according to the acoustic feature extraction path and the time-series feature extraction path. In the intermediate layer, cross-modal feature fusion processing is performed on the acoustic mode feature map and the traffic temporal mode feature map to generate a cross-modal fused feature map; The cross-modal feature fusion processing includes: performing channel alignment processing on the acoustic mode feature map and the flow time series mode feature map according to the channel dimension; performing component-level fusion processing on the processed acoustic mode feature map and the flow time series mode feature map, performing element-wise summation and element-wise multiplication processing on the feature components at corresponding positions to generate a fusion basis matrix; performing feature dimension expansion processing on the fusion basis matrix to form a multi-dimensional fusion matrix, and performing channel recombination and channel compression processing on the multi-dimensional fusion matrix; writing the recombinated fusion matrix into the corresponding dimension positions according to the input structure of the model intermediate layer to generate a cross-modal fusion feature map; The cross-modal fusion feature map is processed by feature compression, feature normalization and decision branch mapping to obtain the input features of each branch. The process of obtaining the input features of each branch includes: performing a feature matrix reading operation on the cross-modal fusion feature map, and performing matrix compression processing on the feature matrix according to the row and column dimensions. A compressed matrix is ​​generated by performing summation and integer division processing on the matrix components within each compression window; performing matrix normalization processing on the compressed matrix, and performing position rearrangement and dimension adjustment processing on the matrix components according to the normalization order to generate a normalized matrix; performing leakage judgment mapping operation, pipe burst trend judgment mapping operation, reflux anomaly judgment mapping operation, and equipment operating status judgment mapping operation on the normalized matrix respectively, performing component selection processing and component recombination processing on the corresponding components of the normalized matrix according to the component selection rules of each judgment mapping, and writing the component recombination results into the corresponding branch input structure to form leakage judgment branch input features, pipe burst trend judgment branch input features, reflux anomaly judgment branch input features, and equipment operating status judgment branch input features respectively; Based on the input features of each branch, end-side identification and reasoning processing is performed to output leakage judgment results, pipe burst trend judgment results, backflow anomaly judgment results and equipment operation status judgment results. The four types of judgment results are then aggregated to generate a water meter status result set. The end-side recognition reasoning process includes: performing component arrangement processing on each feature component in the input features of each branch according to the input dimension of the end-side intelligent recognition model to generate a reasoning input matrix; performing layer-by-layer reasoning calculation processing on the reasoning input matrix, performing matrix multiplication and addition operations, matrix activation operations, and matrix compression operations on the reasoning input matrix in sequence according to the hierarchical structure of the end-side intelligent recognition model to generate a reasoning intermediate matrix; performing output mapping processing on the reasoning intermediate matrix, and performing component selection processing and component value writing processing on the components at the corresponding positions in the reasoning intermediate matrix according to the output dimension requirements of the leakage judgment branch, the pipe burst trend judgment branch, the backflow anomaly judgment branch, and the equipment operating status judgment branch, respectively forming leakage judgment results, pipe burst trend judgment results, backflow anomaly judgment results, and equipment operating status judgment results.

[0024] In this embodiment, the generation of the monitoring event record includes the following specific steps: Read the water meter status result set at the end and perform judgment field parsing processing on the leakage judgment result, pipe burst trend judgment result, backflow anomaly judgment result and equipment operation status judgment result; The judgment field parsing process includes: performing field reading processing on each judgment result in the water meter status result set; performing field separation processing on the judgment level field, judgment label field, and judgment confidence field of each judgment result according to the preset field order; performing position rearrangement processing and type unification processing on the separated fields according to the field type; and writing the normalized fields into the corresponding judgment field set. Based on the parsed judgment fields, the event candidate generation process is performed, and each judgment field is respectively assigned to leakage event candidates, pipe burst trend event candidates, backflow abnormality event candidates, and equipment operation event candidates. The event candidate generation process includes: performing event type mapping processing on the parsed judgment fields, performing field selection and field recombination processing according to the event type order, and writing the recombined fields into the event candidate structure; Based on the risk value mapping table, the judgment level of each event candidate is mapped to a risk score, and a weighted calculation is performed on the risk scores according to the corresponding judgment confidence level to generate a comprehensive risk score; Based on the comprehensive risk score, the risk level is determined by locating the corresponding risk level within the risk level range. Then, based on the risk score of each candidate event, the event type is determined and the monitoring event type is generated. Obtain the current time on the client side to form an event timestamp, and aggregate the parsed judgment fields, event timestamp, risk level and monitoring event type to generate a monitoring event record, and store it in the local event cache.

[0025] In this embodiment, the generation of the reporting data frame includes the following specific steps: The monitoring event records are read on the terminal side, and the backhaul trigger condition matching process is performed according to the monitoring event type and risk level. The backhaul trigger state is determined according to the matching result. The backhaul trigger conditions include event type trigger conditions, risk level trigger conditions and time interval trigger conditions. The backhaul trigger state includes a trigger state that meets the trigger conditions and a non-trigger state that does not meet the trigger conditions. Based on the current trigger status of the return, the return status determination process is executed. If the return status is triggered, the summary feature data generation process is entered; if the return status is not triggered, the current return process is terminated. In the triggered state, feature components corresponding to the monitoring event type are selected from the acoustic feature vector, flow time-series feature vector and joint feature representation associated with the monitoring event record to generate summary feature data; The monitoring event records and summary feature data are sorted by field, length-marked, and segmented according to the data frame structure of the communication link to generate a reporting data frame; The monitoring event records and corresponding summary feature data are reported to the remote monitoring platform by sending reporting data frames through the communication link.

[0026] In this embodiment, the communication unreachable state includes a link unavailable state and a communication quality insufficiency state. The determination of the communication unreachable state includes: reading the link availability field of the communication link, performing a state comparison process based on the availability identifier and the unavailability identifier of the availability field, and mapping the unavailability identifier to the link unavailable state; performing quality value processing on the communication quality indicators of the communication link, and performing a numerical comparison process between the communication quality value and the communication quality threshold, and determining the communication quality insufficiency state when the communication quality value is lower than the communication quality threshold; and determining that the communication link is in a communication unreachable state when either the link unavailable state or the communication quality insufficiency state is met. The communication quality indicators include signal strength, number of bit errors, and round-trip time. In this embodiment, the update of the edge-side intelligent recognition model includes: reading the event confirmation information returned by the remote monitoring platform, and writing the event confirmation label field, event result field, and event deviation information field into the confirmation data set in the order of the fields; performing deviation calculation processing on the event deviation information field in the confirmation data set to generate a deviation sequence; performing weight adjustment processing on the convolutional layer structure of the edge-side intelligent recognition model according to the deviation sequence, and performing component update operation according to the deviation ratio on the weight components corresponding to the deviation positions in the convolutional kernel weight matrix; performing feature path correction processing on the feature extraction structure of the edge-side intelligent recognition model, and performing channel weight increase or decrease processing on the feature channels corresponding to the deviation sequence according to the deviation direction; performing judgment boundary adjustment processing on the multi-branch output structure of the edge-side intelligent recognition model, and performing numerical adjustment processing on the judgment threshold of each branch according to the deviation signal in the deviation sequence; and generating the updated edge-side intelligent recognition model after the weight adjustment processing, feature path correction processing, and judgment boundary adjustment processing are all written into the edge-side intelligent recognition model.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to an ultrasonic water meter intelligent monitoring scenario in a residential water supply area of ​​a southern city. This area has long suffered from problems such as difficulty in timely detection of leakage trends, delayed identification of pipe burst precursors, inability to report backflow anomalies in real time, and frequent fluctuations in the communication network during peak water usage periods in the morning and evening. These issues have prevented traditional monitoring methods from continuously and stably acquiring effective status data. The edge computing architecture and end-side identification mechanism of this invention were fully applied in this complex water usage environment.

[0028] In practical deployment, ultrasonic water meters are installed in the water supply pipes of residential buildings, and ultrasonic time difference signals, echo amplitude signals, temperature signals, and flow measurement signals are collected on-site. The end-side equipment continuously preprocesses the multi-source measurement signals and generates standardized measurement data sequences. Due to the complex pipe network structure and significant differences in user water usage behavior in this area, the water meter end-side needs to possess both real-time performance and robustness. Therefore, the joint feature representation method adopted in this invention can simultaneously express acoustic fluctuation characteristics and flow change patterns, enabling the end-side to maintain stable judgment capabilities even under conditions of high noise and frequent signal interference.

[0029] After receiving the joint feature representation, the end-side intelligent recognition model performs directional weighted convolution calculation, cross-modal feature fusion, and multi-branch decision processing locally, generating a water meter status result set in a timely manner. It also classifies and identifies potential leaks, pipe burst trends, backflow anomalies, and equipment operating status. During field testing, continuous operational data over several days showed that even with weakened signal at night and significant signal interference during morning peak hours, the end-side model could still output structurally stable four-category decision results, enabling the monitoring platform to continuously acquire effective status information.

[0030] In this scenario, the present invention solves the problems of traditional monitoring equipment, such as high network dependence, inability to locally identify status, and inability to continuously provide effective monitoring data during communication interruptions. Especially during periods of frequent short-term communication link unavailability, the device maintains the extraction and identification process of acoustic and traffic features, and writes event timestamps, monitoring event types, risk levels, and summary feature data into a local event cache. When communication is restored, the present invention sends monitoring event records in batches to the monitoring platform according to the event time sequence, and performs convolutional layer weight adjustment, feature path correction, and decision boundary update on the device-side intelligent recognition model based on the event confirmation information returned by the platform, enabling the model to continuously adapt to the real-world situation.

[0031] Comparison of operational data over several consecutive days reveals that, even in real-world environments with fluctuating communication network quality and frequent changes in pipeline status, this invention, through local identification, event caching, and an acknowledgment-driven model update mechanism, enhances the stability and continuity of the monitoring link. On-site maintenance personnel report that the system can output instructive monitoring event types and risk levels before critical anomalies occur, effectively reducing the workload of manual inspections and improving the operational safety of water supply facilities.

[0032] Table 1. Performance Comparison of the Invention and Traditional Ultrasonic Water Meter Monitoring Methods

[0033] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0034] Regarding the accuracy of leakage detection, the traditional method achieves 82.3%, while this invention reaches 94.7%, an improvement of 12.4%. The main reason for this is that this invention uses a joint feature representation and cross-modal feature fusion mechanism, which enables the acoustic mode feature map and the flow time-series mode feature map to be modeled simultaneously at the edge, thus improving the recognition stability under weak leakage signals and mild disturbances.

[0035] In terms of the accuracy of burst pipe trend recognition, the traditional method achieves 76.5%, while the present invention achieves 91.2%, an improvement of 14.7%. This improvement comes from the directional weighted convolution feature map and directional convolution kernel structure introduced in the present invention, which makes the enhanced expression of early microcrack features of burst pipe more obvious due to changes in the direction of sound wave propagation, thus maintaining high recognition capability even in scenarios with large data fluctuations.

[0036] In terms of backflow anomaly identification, this invention achieves a success rate of 93.4%, which is 14.3% higher than the 79.1% of traditional methods. This invention uses a multi-branch decision structure to decouple the backflow anomaly branch from other state branches, allowing the model to establish independent feature paths for backflow behavior, thus reducing the problem of mutual interference between features.

[0037] Regarding the timeliness of event reporting, the traditional method achieves 68.7%, while this invention reaches 96.1%, an improvement of 27.4%. This is because this invention adopts an event-triggered feedback strategy and generates summary feature data in the triggered state, which significantly reduces the communication load and makes the reporting link more stable.

[0038] Regarding data retention during communication outages, traditional methods achieve only 42.8%, while this invention reaches 98.3%. This is because this invention continuously writes event timestamps, event types, risk levels, and summary characteristic data into the event buffer, enabling the system to maintain complete data continuity even during prolonged communication interruptions.

[0039] The success rate of event retransmission after communication recovery reaches 97.9% in this invention, which is 24.5% higher than the 73.4% of the traditional method. This is mainly due to the event timestamp sorting mechanism and batch synchronization mechanism, which enable the retransmission link to have sequential consistency and data integrity.

[0040] Finally, the fault tolerance robustness score of this invention reaches 0.91, while the traditional method only achieves 0.52, representing an improvement of 0.39. This is due to the triple mechanism of local identification, breakpoint caching, and communication recovery synchronization inherent in this invention, which enables the system to maintain stable output under various abnormal conditions.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent monitoring system for ultrasonic water meters based on edge computing, characterized in that, include: The acquisition module is used to acquire and preprocess multi-source measurement signals from ultrasonic water meters to generate standardized measurement data sequences; The feature extraction module is used to extract acoustic feature vectors and flow time-series feature vectors, generate fused feature vectors based on feature reliability, and perform joint splicing to generate joint feature representations; The end-side identification module is used to perform tensor quantization mapping on the joint feature representation and input it into the end-side intelligent identification model to generate a water meter status result set; The event generation module is used to determine the risk level and monitoring event type based on the water meter status result set and generate monitoring event records; The feedback module is used to perform trigger condition matching and feedback status determination, and to send up report data frames through the communication link when the trigger status is triggered. The caching module is used to maintain end-side identification operations and update monitoring event records in the local event cache when communication is unreachable. The update module is used to synchronize the monitoring event records in the local event cache when communication is reachable and update the end-side intelligent recognition model based on the event confirmation information.

2. The intelligent monitoring system for ultrasonic water meters based on edge computing according to claim 1, characterized in that, The modules are connected in the following way: Multi-source measurement signals from ultrasonic water meters are collected, preprocessed, and standardized measurement data sequences are generated. Acoustic feature vectors and flow time-series feature vectors are extracted from the ultrasonic water meter end based on standardized measurement data sequences. Based on feature reliability, fused feature vectors are generated and joint splicing is performed to generate joint feature representations. After performing tensor quantization mapping on the joint feature representation, it is input into the end-side intelligent recognition model for multi-layer feature extraction and multi-branch decision processing to generate a water meter status result set. Based on the water meter status result set, determine the risk level and monitoring event type, generate monitoring event records on the terminal side and store them in the local event cache area; The system performs trigger condition matching and feedback status determination on the monitoring event records, and sends reporting data frames through the communication link when the trigger status is triggered. When the communication link is in a communication unreachable state, maintain the end-side identification operation and synchronously update the monitoring event records in the local event buffer; When the communication link is reachable, the monitoring event records in the local event cache are synchronized to the remote monitoring platform in batches according to the event timestamp order, and the end-side intelligent identification model is updated based on the event confirmation information to generate a new water meter status result set.

3. The ultrasonic water meter intelligent monitoring system based on edge computing according to claim 2, characterized in that, The multi-source measurement signals include ultrasonic time difference signals, echo amplitude signals, temperature signals, and flow measurement signals. The preprocessing includes noise reduction, time alignment, outlier removal, and time window segmentation.

4. The ultrasonic water meter intelligent monitoring system based on edge computing according to claim 2, characterized in that, The generation of the joint feature representation includes the following specific steps: The ultrasonic water meter receives a standardized measurement data sequence at the end and performs time alignment and segmentation according to a preset sampling period to generate acoustic analysis data segments and flow analysis data segments. Based on the acoustic analysis data segment, time-domain statistical extraction, frequency-domain statistical extraction, and morphological stability extraction of the acoustic signal are performed, and the extracted features are integrated to obtain the acoustic feature vector. Based on the traffic analysis data segment, perform trend feature extraction, fluctuation feature extraction, periodic feature extraction and abnormal jump feature extraction of traffic time series, and perform feature integration processing on the extracted features to obtain the traffic time series feature vector; Based on the reliability of acoustic features and the reliability of traffic flow time-series features, acoustic fusion weight value and time-series fusion weight value are calculated, and the acoustic feature vector and traffic flow time-series feature vector are weighted and fused based on the weight values ​​to obtain the fused feature vector; The fused feature vector is dimensionally normalized and scaled, and then jointly concatenated with the acoustic feature vector and the traffic temporal feature vector to generate a joint feature representation.

5. The intelligent monitoring system for ultrasonic water meters based on edge computing according to claim 2, characterized in that, The generation of the water meter status result set includes the following specific steps: The joint feature representation is received at the end of the ultrasonic water meter and the input format conversion process is performed to map the joint feature representation into the joint feature tensor of the end-side intelligent recognition model. The joint feature tensor is input into the intelligent recognition model at the end. In the convolutional layer, the directionality adjustment processing of the convolutional kernel structure is performed. The sampling direction and sensing direction of the convolutional kernel are adjusted according to the direction component representing the direction information of the sound wave in the joint feature tensor, and the direction adjustment convolutional kernel is generated. The edge-side intelligent recognition model is a multi-branch recognition model built on the LCNN architecture, including an input structure, a convolutional layer structure, a feature extraction structure and a multi-branch output structure. An adaptive convolutional kernel adjustment mechanism for acoustic direction information is introduced in the convolutional layer structure, and a cross-modal joint feature processing structure is configured in the middle layer of the feature extraction structure. The orientation field weights are calculated based on the orientation field information representing the intensity distribution of the incident direction of the sound wave in the joint feature tensor, and then the orientation field weights are loaded into the convolution channels corresponding to the orientation adjustment convolution kernel to generate an orientation-weighted convolution feature map. Multi-layer convolutional feature extraction and hierarchical feature convergence processing are performed on the directional weighted convolutional feature map to obtain acoustic pattern feature map and flow temporal pattern feature map; In the intermediate layer, cross-modal feature fusion processing is performed on the acoustic mode feature map and the traffic temporal mode feature map to generate a cross-modal fused feature map; The cross-modal fusion feature map is processed by feature compression, feature normalization and decision branch mapping to obtain the input features of each branch. Based on the input features of each branch, end-side identification and reasoning processing is performed to output leakage judgment results, pipe burst trend judgment results, backflow anomaly judgment results and equipment operation status judgment results. The four types of judgment results are then aggregated to generate a water meter status result set.

6. The intelligent monitoring system for ultrasonic water meters based on edge computing according to claim 2, characterized in that, The generation of the monitoring event record includes the following specific steps: Read the water meter status result set at the end and perform judgment field parsing processing on the leakage judgment result, pipe burst trend judgment result, backflow anomaly judgment result and equipment operation status judgment result; Based on the parsed judgment fields, the event candidate generation process is performed, and each judgment field is respectively assigned to leakage event candidates, pipe burst trend event candidates, backflow abnormality event candidates, and equipment operation event candidates. Based on the risk value mapping table, the judgment level of each event candidate is mapped to a risk score, and a weighted calculation is performed on the risk scores according to the corresponding judgment confidence level to generate a comprehensive risk score; Based on the comprehensive risk score, the risk level is determined by locating the corresponding risk level within the risk level range. Then, based on the risk score of each candidate event, the event type is determined and the monitoring event type is generated. Obtain the current time on the client side to form an event timestamp, and aggregate the parsed judgment fields, event timestamp, risk level and monitoring event type to generate a monitoring event record, and store it in the local event cache.

7. The intelligent monitoring system for ultrasonic water meters based on edge computing according to claim 2, characterized in that, The generation of the reported data frame includes the following specific steps: Read the monitoring event records on the device side, perform back transmission trigger condition matching processing according to the monitoring event type and risk level, and determine the back transmission trigger status based on the matching result; Based on the current trigger status of the return, the return status determination process is executed. If the return status is triggered, the summary feature data generation process is entered; if the return status is not triggered, the current return process is terminated. In the triggered state, feature components corresponding to the monitoring event type are selected from the acoustic feature vector, flow time-series feature vector and joint feature representation associated with the monitoring event record to generate summary feature data; The monitoring event records and summary feature data are sorted by field, length-marked, and segmented according to the data frame structure of the communication link to generate a reporting data frame; The monitoring event records and corresponding summary feature data are reported to the remote monitoring platform by sending reporting data frames through the communication link.

8. The intelligent monitoring system for ultrasonic water meters based on edge computing according to claim 2, characterized in that, The communication unreachable state includes a link unavailable state and a communication quality insufficiency state. The determination of the communication unreachable state includes: reading the link availability field of the communication link, performing a state comparison process based on the availability identifier and the unavailability identifier of the availability field, and mapping the unavailability identifier to the link unavailable state; performing quality value processing on the communication quality indicators of the communication link, and performing a numerical comparison process between the communication quality value and the communication quality threshold, and determining the communication quality insufficiency state when the communication quality value is lower than the communication quality threshold; and determining that the communication link is in a communication unreachable state when either the link unavailable state or the communication quality insufficiency state is met. The communication quality indicators include signal strength, number of bit errors, and round-trip time. The ultrasonic water meter intelligent monitoring system based on edge computing according to claim 2 is characterized in that, The update of the edge-side intelligent recognition model includes: reading the event confirmation information returned by the remote monitoring platform, and writing the event confirmation label field, event result field, and event deviation information field into the confirmation data set in the order of the fields; performing deviation calculation processing on the event deviation information field in the confirmation data set to generate a deviation sequence; performing weight adjustment processing on the convolutional layer structure of the edge-side intelligent recognition model according to the deviation sequence, and performing component update operation according to the deviation ratio on the weight components corresponding to the deviation positions in the convolutional kernel weight matrix; performing feature path correction processing on the feature extraction structure of the edge-side intelligent recognition model, and performing channel weight increase or decrease processing on the feature channels corresponding to the deviation sequence according to the deviation direction; performing judgment boundary adjustment processing on the multi-branch output structure of the edge-side intelligent recognition model, and performing numerical adjustment processing on the judgment threshold of each branch according to the deviation signal in the deviation sequence; after the weight adjustment processing, feature path correction processing, and judgment boundary adjustment processing are all written into the edge-side intelligent recognition model, the updated edge-side intelligent recognition model is generated.

Images