Returned intelligent electric meter sorting management system based on artificial intelligence

Abstract

The invention discloses a returned intelligent electric meter sorting management system based on artificial intelligence, and the system comprises the steps: obtaining the appearance, operation state, environmental parameters, historical records and other information of an electric meter through a multi-modal data collection and fusion technology, and generating a unified feature vector; generating a classification rule based on the recursive decision tree and performing dynamic optimization; a mechanical arm and a conveying belt are combined to realize efficient classification and sorting, and task allocation and resource allocation are optimized through a task scheduling and resource management module; a variational auto-encoder and a causal inference technology are used to detect abnormities in real time and generate early warnings; and the data tracing and analyzing module is used for recording whole-process data and generating a statistical report, so that support is provided for rule optimization and resource adjustment. The sorting efficiency and the classification accuracy of the returned intelligent electric meters are remarkably improved, and full-process intelligent management is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of smart meter processing, and in particular relates to a sorting and management system for returned smart meters based on artificial intelligence. Background Art

[0002] With the construction and promotion of smart grids, smart meters, as important metering equipment in power systems, have been widely used in residential, commercial and industrial scenarios. Due to equipment aging, function upgrades, usage failures and other reasons, a large number of returned smart meters are generated every year. These returned meters need to be classified, tested, repaired or scrapped. However, the existing technology has the following technical difficulties in the processing of returned smart meters:

[0003] 1. Insufficient sorting efficiency and accuracy: There are many types of returned electric meters, with different appearance specifications and performance parameters. Traditional manual sorting or semi-automatic sorting methods are inefficient and prone to classification errors due to human judgment errors. Especially when returning in large quantities, the labor cost is high, and the deviation of subjective judgment also affects the overall accuracy of the sorting process.

[0004] 2. Insufficient information management and process intelligence: The process of returning electric meters involves multiple links, including information entry, classification, maintenance, scrapping decisions, etc. However, the existing system lacks a unified management platform, resulting in scattered and incomplete information records, making it difficult to achieve dynamic scheduling of tasks and real-time tracking of data. In addition, the links between the various links are not smooth enough, and there is a lack of automation and intelligent support, which affects the overall efficiency of the process.

[0005] 3. Inadequate resource utilization and dynamic allocation: In the process of handling returned meters, the allocation of resources such as manpower, equipment and sites is not optimized, making it difficult to quickly respond to fluctuations in the number of returned meters. At the same time, decisions on maintenance and scrapping often rely on experience and lack scientific data support, resulting in inefficient resource utilization or even waste.

[0006] 4. Weak quality traceability and predictive maintenance capabilities: In the current process of returning electric meters, it is difficult to trace the entire life cycle information of the electric meter, including its usage history, fault causes and return characteristics. At the same time, there is a lack of predictive maintenance capabilities for potential problems of electric meters, and it is impossible to proactively identify possible return risks, which increases the uncertainty of processing.

[0007] To sum up, the existing technology has significant shortcomings in the sorting management of returned smart meters. There is an urgent need for an intelligent sorting management system based on artificial intelligence technology that can optimize the entire process from data collection, classification and sorting, resource scheduling to quality traceability and maintenance decision-making, and comprehensively improve the efficiency and quality of returned meter processing. Summary of the invention

[0008] The present invention provides a sorting and management system for returned smart meters based on artificial intelligence, which solves the technical problems of low efficiency, insufficient accuracy, unreasonable resource utilization and high management difficulty in the sorting process of returned meters. The system is composed of the following:

[0009] 1. Electricity meter feature collection and fusion module: used to collect multimodal data of returned smart meters, extract and fuse features of the data, and provide a unified feature vector for the sorting rule generation and optimization module; collect the appearance image of the meter through an industrial camera, use a vibration sensor to monitor the operating status, combine the temperature and humidity sensor to record environmental parameters, and read the historical task records and batch information of the meter to form a multimodal data set; use a convolutional neural network to extract image features, generate feature patterns through a self-organizing map network, and combine the multimodal attention mechanism to weightedly integrate various feature data to generate a unified feature vector to provide input for subsequent modules.

[0010] 2. Sorting rule generation and optimization module: Generate classification rules based on feature vectors, and dynamically optimize rule logic through real-time feedback; build an initial classification rule base through a recursive decision tree, and clarify the classification standards, including three types of results: normal use, waiting for repair, and scrapping; analyze the adaptability and usage frequency of rules in combination with power-law distribution, and dynamically adjust the rule weights and priorities; through real-time sorting task feedback data, the module optimizes the classification logic to ensure that the rules adapt to the characteristic changes of different batches of meters; generate decision support rules for repair or scrapping in combination with specific judgment conditions (including the degree of damage to the appearance, the integrity of label information, and abnormal performance parameters), and pass the optimized classification rules to the meter sorting and execution module.

[0011] 3. Electric meter sorting and execution module: responsible for classifying the returned smart meters according to the classification rules and implementing the sorting tasks; the module receives the classification rules provided by the sorting rule generation and optimization module, classifies the feature vectors, and converts the classification results into specific sorting operations; the electric meter sorting and execution module uses a robotic arm and a conveyor belt to sort the meters to the corresponding areas according to the classification results, including the storage area for normally used meters, the processing area for meters to be repaired, and the processing area for scrapped meters; the module monitors the status of the sorting equipment in real time, dynamically adjusts the conveyor belt speed and the gripping force of the robotic arm to ensure the stability and accuracy of the sorting tasks; the sorting results and equipment operation status data are synchronously fed back to the task scheduling and resource management module and the anomaly detection and early warning module.

[0012] 4. Task scheduling and resource management module: responsible for the optimal allocation of sorting tasks and system resources; dynamically adjusts task priorities and resource allocation strategies by modeling the relationship between tasks, equipment and resources through graph attention networks, and combining multi-agent reinforcement learning; for meters to be repaired, the module assigns maintenance tasks based on classification results and the current status of maintenance equipment; for scrapped meters, the module optimizes task paths based on processing area capacity constraints; the scheduling data generated by the module is synchronously fed back to the sorting rule generation and optimization module for rule optimization.

[0013] 5. Anomaly detection and early warning module: used to monitor the processing of returned smart meters and the operating status of sorting equipment in real time, identify potential anomalies and provide maintenance suggestions; detect abnormal patterns in operating data through variational autoencoders, analyze the root causes of anomalies in combination with causal inference technology, and generate targeted maintenance plans; for meters to be repaired, further analyze the fault type by combining historical data with real-time features to provide auxiliary decision-making for maintenance tasks; identify signals that the system is close to failure based on multi-scale entropy analysis and critical slowing phenomenon detection, generate early warning information and provide adjustment suggestions to the task scheduling and resource management modules.

[0014] 6. Data tracing and analysis module: used to record the data of the entire process of returned smart meters from classification, sorting to maintenance or scrapping, and conduct statistical analysis on the completion of the task; record the classification results, equipment operation records and the execution of maintenance or scrapping to form a complete data chain; generate statistical reports based on historical data to provide a basis for the sorting rule generation and optimization module to adjust the rule logic, and provide support for the task scheduling and resource management module to optimize resource allocation.

[0015] Beneficial effects of the present invention:

[0016] 1. Improve sorting efficiency and classification accuracy:

[0017] Through the synergy of the meter feature collection and fusion module and the sorting rule generation and optimization module, the system can accurately generate classification rules and dynamically optimize them, significantly improving the sorting efficiency and classification accuracy of returned smart meters and reducing human operational errors.

[0018] 2. Realize dynamic optimization of resource allocation:

[0019] The task scheduling and resource management module is based on real-time monitoring and optimization algorithms, and can dynamically allocate task priorities and equipment resources, effectively avoiding resource waste or operation bottlenecks, and improving the overall operation efficiency of the system.

[0020] 3. Enhance system stability and exception handling capabilities:

[0021] The anomaly detection and early warning module detects potential problems in advance and provides maintenance suggestions through abnormal pattern recognition and early warning mechanism, which significantly improves the stability of sorting equipment and processing processes and reduces the risk of failure.

[0022] 4. Provide scientific maintenance and scrapping decision support:

[0023] The sorting rule generation and optimization module combines the judgment conditions such as appearance damage and performance abnormality to provide clear classification rules for maintenance and scrapping. Combined with the task scheduling and resource management module, the system can reasonably allocate maintenance resources and scrapping tasks.

[0024] 5. Realize data traceability throughout the entire life cycle:

[0025] The data tracing and analysis module records the entire process of returned smart meters from classification to processing, and provides analysis results such as the distribution of return reasons, repair success rate and scrap processing efficiency, providing comprehensive data support for system optimization and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A schematic diagram of the system architecture is shown;

[0027] Figure 2 The system workflow diagram is shown. DETAILED DESCRIPTION

[0028] Exemplary embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0029] Combination Figure 1 The present invention provides a monitoring system for sorting production lines of returned smart meters, and the system composition is as follows:

[0030] 1. Electricity meter feature collection and fusion module:

[0031] The meter feature collection and fusion module is responsible for the multimodal data collection, feature extraction and feature fusion of the returned smart meter, and finally generates a unified feature vector to provide high-quality input support for the sorting rule generation and optimization module. The implementation of this module includes four links: data collection, feature extraction, feature pattern generation and feature fusion, which respectively process the multimodal data of appearance images, operating status, environmental parameters and historical task records.

[0032] Furthermore, the data collection link of the meter feature collection and fusion module obtains the appearance image, operation status signal, environmental parameters and historical task records of the returned meter through a variety of sensor devices, specifically including:

[0033] The appearance image is collected by an industrial camera, and the resolution is set to Pixels, frame rate 30fps, can clearly capture the surface features of the meter (such as damage, pollution, blurred labels, etc.). The captured image is normalized after grayscale processing, and the normalization formula is as follows:

[0034]

[0035] in, is the original image data; is the mean value of the image data, which is used to eliminate brightness deviation; is the standard deviation of the image data, used to standardize the contrast. The processed image set is marked as ; The image data is timestamped , used to ensure synchronization with other modality data.

[0036] The operating status signal is recorded by a vibration sensor at a sampling frequency of 100 Hz, and the collected time series data is expressed as In order to remove high frequency noise, the signal is processed by a low pass filter, whose transfer function is:

[0037]

[0038] in, is the transfer function of the filter; f is the sampling frequency; is the cutoff frequency of the filter, which is set to 50 Hz to retain the low-frequency characteristics.

[0039] The filtered signal is used to reflect the operating characteristics of the meter and is marked as .

[0040] The environmental parameters are collected by temperature and humidity sensors, including the ambient temperature and humidity The acquisition ranges of the temperature and humidity data are to For the convenience of subsequent analysis, the data were normalized, and the normalization formula is as follows:

[0041]

[0042] in, They are original temperature and humidity data respectively; They are the normalized temperature and humidity data, which are used to match the numerical range of other modal features.

[0043] The historical task records are collected through the meter data interface, and the record content includes usage time, fault cause and batch information, etc. The records are stored in the form of structured data sets as , and used for subsequent feature extraction and analysis.

[0044] Furthermore, the feature extraction link of the electric meter feature acquisition and fusion module extracts features of the appearance image, operating status, environmental parameters and historical records respectively based on the collected multimodal data through deep learning and frequency domain analysis technology, specifically including:

[0045] The appearance features are extracted by a convolutional neural network (CNN). Specifically, the convolutional neural network architecture consists of three convolutional layers and two fully connected layers:

[0046] The size of the first convolution kernel is , the stride is 1, which is used to extract the edge features of the image;

[0047] The size of the second convolution kernel is ,, the stride is 1, which is used to extract local detail features;

[0048] The size of the third convolution kernel is ,, the stride is 1, and batch normalization is combined to improve the robustness of the features.

[0049] Each convolutional layer is followed by a maximum pooling layer with a pooling kernel size of The convolutional features are mapped to a 128-dimensional feature vector through a fully connected layer, and the formula is as follows:

[0050]

[0051] in, is the appearance feature vector, which represents the main characteristics of the image; is the set of real numbers.

[0052] The operating status characteristics are extracted by wavelet packet decomposition (WPD) to extract frequency domain energy characteristics. Specifically, after the signal is decomposed to the third layer, the energy characteristics of each frequency band are calculated:

[0053]

[0054] in, is the wavelet packet decomposition coefficient of the vibration signal; For the The number of subbands in the layer; the extracted vibration feature vector is .

[0055] The environmental features are generated by adaptive cluster analysis. Specifically, cluster centers are calculated based on historical data distribution. , and calculate the weighted distance from the sample to the cluster center. The formula is as follows:

[0056]

[0057] in, is the clustering weight; is the number of cluster centers, and its value is 4.

[0058] The calculation result generates an environmental feature vector .

[0059] The historical record data is processed by vectorization to generate a feature vector with a dimension of 32 .

[0060] Furthermore, the feature pattern generation link of the electric meter feature collection and fusion module generates a low-dimensional topological pattern of multimodal features through a self-organizing map (SOM) network, specifically including:

[0061] A). Structure of SOM network

[0062] The grid size of the SOM network is , the network nodes are represented by points in the feature space; the weight update of the network nodes adopts the Gaussian neighborhood function, and its formula is as follows:

[0063]

[0064] in, is the coordinate of the current winning node (when the data vector is input, the node with the smallest Euclidean distance to the input data vector is found through calculation, that is, the winning node); For Node The coordinates of is the neighborhood width, over time Gradually decrease.

[0065] B). Feature pattern generation

[0066] The weight update formula is:

[0067]

[0068] in, For Node The weight vector of is the learning rate; is the input feature vector.

[0069] Furthermore, the feature fusion link of the electric meter feature collection and fusion module realizes weighted integration of multimodal features through a multimodal attention mechanism to generate a unified feature vector, which specifically includes:

[0070]

[0071] in, is the weight coefficient, satisfying .

[0072] The weight initialization is calculated by standardized variance, and the formula is as follows:

[0073]

[0074] in, , The eigenvectors are , The standard deviation of; The dynamic adjustment of weights is optimized by gradient descent method, and the optimization objective function is:

[0075]

[0076] in, They represent the fused feature vector and target feature vector respectively.

[0077] The dimension of the fused feature vector is .

[0078] 2. Sorting rule generation and optimization module

[0079] The sorting rule generation and optimization module generates classification rules based on the unified feature vector generated by the meter feature collection and fusion module, and dynamically optimizes the rule logic through real-time feedback to ensure the accuracy and adaptability of the sorting task. The implementation of this module includes rule generation, rule optimization and feedback processing.

[0080] Furthermore, the rule generation link of the sorting rule generation and optimization module constructs an initial classification rule base through a recursive decision tree, which specifically includes:

[0081] A). Setting of classification goals

[0082] The classification targets of the sorting rules include the following three categories:

[0083] a. Normal use of electric meter: The appearance and performance are normal, no further processing is required, and its characteristics meet the following ranges:

[0084]

[0085] in, is the appearance feature vector; is the running state feature vector; and Calculated by the following formula:

[0086]

[0087]

[0088] in, is the mean, is the standard deviation, It is an adjustment factor used to adapt to the characteristics of different batches of electricity meters.

[0089] b. Meters to be repaired: Meters with abnormal appearance or performance but repairable, with the following characteristics:

[0090]

[0091] And the repair conditions are met, including appearance damage (the proportion of the label blurred area exceeds 10%) can be repaired by replacing components, or abnormal vibration signal values ​​can be restored to the normal range through adjustment.

[0092] c. Scrapped electric meter: The appearance is severely damaged or the performance is irreparable, and its characteristics meet the following conditions:

[0093]

[0094] in, is the environmental feature vector; , It is an adjustment factor used to adapt to specific environmental conditions.

[0095] B). Construction of recursive decision tree

[0096] The recursive decision tree is based on a unified feature vector As input, the partitioning features are selected according to the principle of maximizing information gain. The formula is as follows:

[0097]

[0098] in, is the current node dataset, To divide the features, Features The value set of ; is the data set entropy, defined as:

[0099]

[0100] in, For Category probability.

[0101] The classification conditions generated by each node include: If the label fuzziness ratio exceeds the threshold (such as 10%), it is classified as "meter to be repaired"; if The frequency domain energy outlier value exceeds , and cannot be repaired, it is classified as a "scrapped meter".

[0102] C) Generation of rule base

[0103] The constructed recursive decision tree is stored in a tree structure to form an initial rule base:

[0104]

[0105] Among them, each Contains classification criteria, weights, and priorities.

[0106] Furthermore, the rule optimization link of the sorting rule generation and optimization module combines the power law distribution to analyze the rule adaptability and usage frequency, and optimizes the rule logic through real-time feedback, specifically including:

[0107] A) Analysis of rule adaptability

[0108] Calculate the rule fitness score using feedback data:

[0109]

[0110] The power law distribution is used to analyze the priority of rules, and the formula is as follows:

[0111]

[0112] in, Indicates that it is proportional, is the frequency of rule usage, is the power law distribution parameter, which is fitted by the following formula:

[0113]

[0114] B). Dynamic adjustment of rule weights

[0115] Dynamically adjust rule weights:

[0116]

[0117] in is the average adaptability score of the rule base; is the weight value for the next step; is the current weight value; is the learning rate.

[0118] C). Rule amendment and deletion

[0119] Rules with adaptability scores lower than the threshold are deleted from the rule base. At the same time, new rule nodes are added and classification conditions are regenerated.

[0120] Furthermore, the feedback processing link of the sorting rule generation and optimization module realizes rule optimization through the interaction between real-time data and sorting results, which specifically includes:

[0121] A). Collection of real-time feedback data

[0122] The meter sorting and execution module returns the sorting results and equipment operation status, which are recorded as a feedback set:

[0123]

[0124] in, Feedback result of a single sorting task.

[0125] B) Iterative update of rule base

[0126] Update the rule base according to the feedback data. The optimized rule base is expressed as:

[0127]

[0128] in, Optimize functions for rules, including rule weight adjustment, condition modification, and new rule addition.

[0129] C) Maintaining the stability of the rule base

[0130] Ensure the integrity of classification logic after rule updates and avoid rule conflicts through consistency checks.

[0131] 3. Electricity meter sorting and execution module

[0132] The meter sorting and execution module classifies the returned smart meters and performs sorting operations according to the classification rules provided by the sorting rule generation and optimization module; by receiving the classification rules, the feature vectors are parsed and matched, and the classification results are converted into sorting instructions; the meters are sorted to the storage area for normal use meters, the processing area for meters to be repaired, and the processing area for scrapped meters using a robotic arm and a conveyor belt; this module has real-time monitoring capabilities and can dynamically adjust the gripping force and conveyor belt speed to ensure the efficiency and stability of the sorting task. The sorting results and equipment operation status data are synchronously fed back to the task scheduling and resource management module and the anomaly detection and warning module to provide support for system optimization and maintenance. The implementation of this module includes rule reception, feature classification, execution control, data synchronization and feedback.

[0133] Furthermore, the rule receiving link of the electric meter sorting and execution module is responsible for receiving and parsing the classification rules provided by the sorting rule generation and optimization module, specifically including:

[0134] A). Rule delivery

[0135] Receive the sorting rule base, which is represented by a collection of nodes:

[0136]

[0137] in, is a rule node, each node contains classification conditions and target sorting area identification ; n is the number of nodes.

[0138] B). Rule analysis

[0139] Analyze the classification conditions and target area identification of the rule nodes to determine the applicable scope of each rule. satisfy , then assign to (Normal use of meter storage area); if satisfied , then assign to (Processing area of ​​the meter to be repaired); if satisfied , then assign to (Scrapped meter processing area); if none of them meet the requirements, then allocate to (Abnormal area).

[0140] C) Rule synchronization and fault tolerance mechanism

[0141] The maximum delay time of the rule synchronization time setting is set to 1 second; and the cached rules are verified. When the previous version of the rule base cached locally is loaded, its validity is verified. If the rule version number is consistent with the historical classification result, the cached rule is valid, otherwise it is invalid. The default rule base will be enabled and manual intervention will be triggered.

[0142] Furthermore, the feature classification link of the electric meter sorting and execution module performs classification operations according to the input feature vector and classification rules, specifically including:

[0143] A). Feature matching

[0144] For feature vector , matching the classification conditions of the rule nodes one by one :

[0145]

[0146] B). Classification priority decision

[0147] If the feature vector matches multiple rule nodes, the classification result is determined based on the priority. The priority calculation formula is:

[0148]

[0149] in, is the initial weight of the rule node, the range , set when the rule is generated; is the rule matching frequency (unit: times / second); is the adjustment coefficient, set to 0.5.

[0150] C). Exception handling

[0151] If the feature vector cannot match any rule node, it is marked as an "abnormal sample" and assigned to the abnormal area . Abnormal sample feedback includes:

[0152]

[0153] in, The classification criteria to try to match; Is the exception type.

[0154] Furthermore, the execution control link of the electric meter sorting and execution module completes the specific sorting operation according to the classification result, which specifically includes the following steps:

[0155] A). Hardware instruction generation

[0156] Generate operating instructions for sorting equipment:

[0157]

[0158] in: Identify the target area; are the gripping force of the robot and the conveyor belt speed parameters.

[0159] B). Dynamic adjustment of the gripping force of the robot arm

[0160] Grasping strength Combined with the meter characteristics calculation:

[0161]

[0162] in, is the mass of the meter (unit: kg), range ; is the acceleration due to gravity; is the adjustment coefficient, range , determined by the task load; is the appearance size characteristic, and the calculation formula is:

[0163]

[0164] in, , are height and width respectively (unit: cm); , is the thickness, is the projection angle; is the label clarity weight, ranging from .

[0165] C). Dynamic adjustment of conveyor belt speed

[0166] The formula for conveyor belt speed is:

[0167]

[0168] in, is the conveyor belt length (unit: m), range ; is the task sorting time (unit: s); The maximum safe speed.

[0169] D). Real-time monitoring and adjustment

[0170] Monitor the equipment's operating status, including the grasping success rate and speed deviation. If the grasping success rate is less than 90%, adjust the grasping force; if the speed deviation exceeds 5%, adjust the conveyor belt speed.

[0171] Furthermore, the data synchronization and feedback link of the meter sorting and execution module synchronizes the sorting results and equipment operation status data to the task scheduling and resource management module and the anomaly detection and early warning module. The data structure is as follows:

[0172]

[0173] Includes, sample number ; Sorting area identification ; Execution status ;Device status .

[0174] 4.Task scheduling and resource management module

[0175] The task scheduling and resource management module uses the graph attention network (GAT) to model the dynamic relationship between tasks, equipment and resources, and combines multi-agent reinforcement learning (MARL) to optimize task priorities and resource allocation strategies to solve problems such as equipment overload and low resource utilization in the task of sorting returned smart meters. The specific implementation of this module includes task modeling, dynamic adjustment of task priorities, resource allocation optimization, and scheduling data synchronization and feedback.

[0176] Furthermore, the task modeling phase of the task scheduling and resource management module uses GAT to construct a dynamic relationship diagram of tasks, equipment and resources, and clarifies the strength of association between each node, including:

[0177] A). Graph construction and node definition

[0178] Task Map , node set ,in, is a task node, representing each meter sorting task; It is a device node, which indicates a specific execution device (such as a robotic arm, maintenance platform, etc.); is a resource node, representing system resources (such as maintenance station capacity, scrap area storage space). ,in, It is the edge between the task node and the device node, indicating the degree of demand of the task for the device; It is the edge between the device node and the resource node, indicating the resource consumption intensity of the device.

[0179] B).Node feature definition

[0180] a.Task Node , including: Task priority , set according to the urgency of the task; task type , maintenance tasks ( ) or scrap tasks ( ); Type of resources required for the task ; Estimated completion time , combined with dynamic estimation of historical task data, the calculation formula is:

[0181]

[0182] in:

[0183] , represents the task workload, which is quantified by combining task complexity and external resource requirements. is the operational complexity of the task (unit: standard task unit), obtained by analyzing historical task data; The resource coordination complexity of the task (unit: standard task unit), reflecting the task's dependence on external resources (such as equipment, space). For example, a maintenance task requires the use of a specific maintenance station, and a scrapping task requires transportation equipment and storage space; is the weight factor;

[0184] , represents the actual processing capacity of the device, where The maximum processing capacity of the device (unit: task unit / second), which means the task units that the device can process per second under no-load conditions; The current load of the device (unit: number of tasks), indicating the number of tasks currently assigned to the device; is the maximum load of the device (unit: number of tasks), indicating the maximum number of tasks that the device can bear;

[0185] , represents the task preparation time, where, is the scale of the task (unit: task unit), such as the number of meters or the amount of data; The device initialization time (in seconds), which is the fixed time required for the device to load tasks and prepare for execution; is a weight factor used to adjust the relative impact of task size and initialization time on task preparation time;

[0186] , represents the cumulative delay time of the task, where For the task The waiting time for each time period (in seconds); The total number of time slots spent waiting for the task.

[0187] b. Device Node , including: device status , 1 means available, 0 means unavailable; current load of the device , the number of assigned tasks, ranging from , The maximum load of the device; processing capacity per unit time (tasks / second) .

[0188] c. Resource nodes , including: remaining capacity of resources , such as the remaining processing capacity of the maintenance area; resource utilization efficiency , defined as the ratio of the time a resource is effectively occupied to the total time.

[0189] C). Edge weight calculation

[0190] Edge Weight Represents the dependency strength between nodes, and the calculation formula is as follows:

[0191]

[0192] in, For Node The eigenvector of is the feature mapping weight matrix; is the attention weight vector; For Node The set of neighbors.

[0193] Furthermore, the task scheduling and resource management module dynamically adjusts the task priority through MARL, and dynamically adjusts the priority according to the real-time status of the task. A). Task status definition Task state vector The definition is as follows:

[0194]

[0195] in is the current task queue length; is the remaining capacity of the resource; is the task type; The accumulated delay time for the task (in seconds).

[0196] B) Reward function design

[0197] Reward Function Aims to optimize task completion efficiency and reduce load pressure. The formula is as follows:

[0198]

[0199] in is the reward weight factor; Fluctuations in resource utilization; is the number of completed tasks; Accumulates delay time for the task.

[0200] C). Priority update formula

[0201] Task priority is dynamically updated using the following formula:

[0202]

[0203] in, is the priority adjustment of the agent output, ranging from .

[0204] Furthermore, the resource allocation optimization of the task scheduling and resource management module dynamically adjusts the allocation scheme according to the task priority and the device status.

[0205] A). Resource allocation of meters to be repaired

[0206] The resource allocation formula for the meter to be repaired is as follows:

[0207]

[0208] Among them, argmax means finding the value of the independent variable that maximizes the function value in the function; Prioritize the tasks; is the remaining capacity of the resource; for resource utilization efficiency; Provide an estimated completion time for the task.

[0209] B). Optimization of scrapped electricity meter paths

[0210] The optimization formula for scrapped electricity meters is as follows:

[0211]

[0212] Among them, argmin means finding the value of the independent variable that minimizes the function value in the function; , is the path length, is the unit transportation cost; , where is the resource demand, is the resource capacity, is the penalty factor.

[0213] Furthermore, the task scheduling and resource management module synchronously feeds back the scheduling data to the sorting rule generation and optimization module.

[0214] 5. Anomaly detection and warning module

[0215] The anomaly detection and early warning module is used to monitor the processing of returned smart meters and the operating status of sorting equipment in real time. It detects anomalies through variational autoencoders (VAE), analyzes the root causes of anomalies by combining causal inference technology, and uses multi-scale entropy and critical slowing down detection to achieve failure risk prediction. The module finally generates early warning information and feeds it back to the task scheduling and resource management module to provide dynamic adjustment suggestions. The specific implementation of this module includes four links: anomaly detection, abnormal root cause analysis, failure risk prediction, and maintenance suggestions and feedback.

[0216] Furthermore, the anomaly detection link of the anomaly detection and early warning module includes:

[0217] A). Input data and feature vector

[0218] The module receives the multimodal feature vector provided by the meter feature collection and fusion module , including: Appearance features: indicating the degree of damage to the appearance of the meter, label ambiguity, etc., extracted through convolutional neural networks; Operation status features: obtained by decomposing vibration signals through wavelet packets, used to evaluate the stability of equipment operation; Environmental features: including aggregated data of environmental conditions such as current temperature and humidity; Historical record features: indicating task history, including the fault type of the meter, task completion rate and cumulative number of maintenance times.

[0219] B). Anomaly Detection Model

[0220] The model consists of an encoder and a decoder.

[0221] The encoder maps the input feature vector to latent variables:

[0222]

[0223] The decoder reconstructs the input features through hidden variables:

[0224]

[0225] The optimization goal of the model is to minimize the difference between the reconstruction error and the latent variable distribution. The loss function is:

[0226]

[0227] in: is the reconstruction error, which is used to measure the difference between the input and reconstructed data; is the divergence, which is used to constrain the distribution of latent variables to be close to the standard normal distribution.

[0228] Anomaly detection based on reconstruction error and dynamic threshold:

[0229]

[0230] in: is the entropy of the reconstruction error, defined as:

[0231]

[0232] Used to measure the distribution complexity of reconstruction error;

[0233] It is a dynamic threshold that can adapt to the characteristics of different batches of tasks.

[0234] Furthermore, the abnormality root cause analysis link of the abnormality detection and early warning module includes:

[0235] A) Causal Modeling

[0236] Construct a causal graph based on task state variables and equipment operation state variables:

[0237] a. Node collection: includes variables such as task queue length, resource remaining capacity, device load, and operating temperature.

[0238] b. Edge set: represents the causal relationship between variables.

[0239] B). Abnormal propagation path identification

[0240] Causal weight calculation:

[0241] The causal weight of each edge is calculated based on the conditional entropy:

[0242]

[0243] in:

[0244] : Conditional entropy, which indicates the uncertainty of variables under known conditions;

[0245] : The overall uncertainty of the variable.

[0246] The optimal anomaly propagation path is the path that minimizes the causal weight:

[0247]

[0248] Furthermore, the failure risk prediction link of the anomaly detection and early warning module includes:

[0249] A). Multi-scale entropy analysis

[0250] Compute the multiscale entropy of the system state at different time scales:

[0251]

[0252] in: : Time scale The probability distribution of the system state.

[0253] B) Critical slowing down detection

[0254] Calculate the critical slowing rate from time series data:

[0255]

[0256] when When it continues to rise and exceeds the threshold, combined with the entropy increase trend:

[0257]

[0258] Determine that the system is close to failure.

[0259] 6.Data tracing and analysis module

[0260] The data tracing and analysis module is responsible for recording the entire process data of returned smart meters from classification, sorting to repair or scrapping, forming a complete data chain, and generating statistical reports on task completion through analysis. Specific implementation includes:

[0261] A). Data recording and storage

[0262] After the meter sorting and execution module completes the classification, the classification results of the meter are recorded in real time, including:

[0263] Unique identification of the electric meter, marking the independent number of each electric meter;

[0264] Classification category, describing the classification results;

[0265] Classification timestamp, which records the exact time when the classification is completed.

[0266] After each sorting task is completed, the operating status of the sorting equipment is recorded, including:

[0267] conveyor speed, which represents the average speed during the task execution;

[0268] The robot arm operation frequency, describing the frequency of task execution;

[0269] Sorting equipment status, records the real-time operation status of the sorting equipment.

[0270] Maintenance record of the electric meter to be repaired:

[0271] Maintenance task number, which uniquely identifies the maintenance task;

[0272] Maintenance completion status, recording the results of maintenance tasks;

[0273] Maintenance time refers to the time it takes to complete the task.

[0274] Disposal record of scrapped electric meters:

[0275] The scrapped task number uniquely identifies the scrapped task;

[0276] Scrap completion time and record the task completion time.

[0277] The module uses a distributed database to store the above data, and the sub-table design is as follows:

[0278] Table 1: Classification_Log, records the classification results;

[0279] Table 2: Operation_Log, records equipment operations;

[0280] Table 3: Repair_Disposal_Log, records repair and disposal status.

[0281] B) Data Analysis

[0282] The distribution of reasons for returning electric meters is analyzed based on the classification results. The calculation formula for the proportion of classification categories is:

[0283]

[0284] in:

[0285] : Number of meters in the classification category;

[0286] : Total number of returned electricity meters.

[0287] For all maintenance tasks, calculate the maintenance success rate:

[0288]

[0289] in:

[0290] : Number of successful maintenance tasks;

[0291] : Total number of maintenance tasks.

[0292] The average time taken to scrap a task is calculated as:

[0293]

[0294] in:

[0295] : Processing time of scrapped tasks;

[0296] : Total number of scrapped tasks.

[0297] C). Report Generation

[0298] a. Contents of statistical report:

[0299] The module periodically generates the following statistical reports:

[0300] Return reason distribution report: Displays the proportion distribution of normal use, waiting for repair, and scrapped categories;

[0301] Maintenance success rate report: Provides the success rate of each batch of maintenance tasks;

[0302] Scrap processing efficiency report: Analyze the average scrap processing time of different batches.

[0303] b. Data Visualization:

[0304] Use a bar chart to display the distribution of return reasons;

[0305] Use a line chart to show the changing trend of the repair success rate;

[0306] Use a box plot to analyze the distribution of scrap processing times.

[0307] The module adjusts the weight of the classification rules through the analysis results of the return reason distribution and the repair success rate. This includes: when the scrap category accounts for a high proportion, the scrap feature weight is increased; when the repair success rate is reduced, the judgment criteria for the meter to be repaired are optimized.

[0308] Optimize resource allocation based on the statistical data of scrap processing efficiency and maintenance time, including: improving the priority allocation of maintenance resources; adding scrap area processing equipment to reduce task waiting time.

[0309] Combination Figure 2 , the workflow of the present invention is as follows:

[0310] S1: Electricity meter feature collection and fusion

[0311] Multimodal data of returned smart meters are collected through a variety of sensor devices, feature extraction and fusion are completed, and a unified feature vector is generated.

[0312] S2: Sorting rule generation and optimization

[0313] Based on the collected feature vectors, meter classification rules are generated and the rule logic is dynamically optimized to adapt to the characteristic changes of returned meters in different batches.

[0314] S3: Meter sorting and execution

[0315] According to the classification rules, the returned smart meters are classified and the sorting task is carried out. After completion, the sorting results and equipment operating status are fed back synchronously.

[0316] S4: Task Scheduling and Resource Management

[0317] Optimize task allocation strategies and resource allocation to ensure efficient operation of the sorting process.

[0318] S5: Anomaly detection and warning

[0319] Monitor the meter sorting process and equipment operating status in real time, detect potential anomalies and generate early warning information.

[0320] S6: Data tracing and analysis

[0321] Record the data of the entire process of electricity meters from classification to maintenance or scrapping, generate statistical reports, and provide optimization basis for other modules.

[0322] In summary, the present invention realizes the intelligent management of the entire process of returned smart meters from data collection, classification and sorting to task scheduling, anomaly detection and data tracing through the collaborative work of multiple modules, effectively improving the sorting efficiency, classification accuracy and resource utilization, providing reliable technical support for large-scale processing of smart meters, and has broad application value and development prospects.

[0323] Although this document shows the specific application of the present invention through specific embodiments, these embodiments are only for illustrative purposes and are not intended to limit the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims, and any appropriate modification, equivalent substitution or improvement based on the principle of the present invention should be deemed to fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be given the broadest interpretation to cover all such modifications, equivalent structures and functions.

Claims

1. A sorting and management system for returned smart meters based on artificial intelligence, characterized in that: The following units are included: The meter feature collection and fusion module is used to collect multimodal data of returned smart meters, including appearance images, operating status signals, environmental parameters and historical task records, and generate a unified feature vector through feature extraction and fusion to provide input for subsequent modules; Sorting rule generation and optimization module, used to generate classification rules based on unified feature vectors and dynamically adjust rule weights and priorities through real-time feedback; The meter sorting and execution module is used to sort the returned smart meters according to the classification rules, and synchronously feed back the sorting results and equipment operation status data to the task scheduling and resource management module and the anomaly detection and early warning module for optimizing resource allocation and abnormal status analysis; Task scheduling and resource management module, used to optimize sorting task execution through task priority and resource allocation strategy; Anomaly detection and early warning module, used to monitor the sorting process and equipment operation status, identify potential anomalies and generate early warning information; The data tracing and analysis module is used to record the entire process data of returned smart meters from classification to repair or scrapping, generate statistical reports, and provide a basis for sorting rules and resource allocation optimization.

2. The artificial intelligence-based return smart meter sorting management system according to claim 1 is characterized in that: The sorting rule generation and optimization module constructs an initial classification rule base based on a recursive decision tree. The classification categories include normal use, to be repaired, and scrapped, and dynamically optimizes the rules through the following mechanisms: Use power law distribution to analyze the frequency and adaptability of rule usage; Adjust the weight and priority of classification rules based on real-time feedback data; Combined with specific judgment conditions, including the degree of damage to the meter's appearance, label integrity, and abnormal performance parameters, optimization support rules for repair or scrapping decisions are generated.

3. The artificial intelligence-based return smart meter sorting management system according to claim 1 is characterized in that: The meter sorting and execution module implements the sorting task in the following ways: According to the classification rules, the returned electric meters are sorted into the normal use area, the area to be repaired and the scrapping area using the robotic arm and conveyor belt; Real-time adjustment of the robot arm operating frequency and conveyor belt speed to improve the stability of the sorting task; Record the sorting results and equipment operating status, and synchronously feed back to the task scheduling and resource management module as well as the anomaly detection and early warning module.

4. The artificial intelligence-based return smart meter sorting management system according to claim 1 is characterized in that: The task scheduling and resource management module establishes a dynamic relationship between tasks and resources through a graph attention network, and combines multi-agent reinforcement learning to optimize the scheduling scheme, specifically including: Adjust task allocation strategy based on task priority; Allocate maintenance resources to the meters to be repaired, and optimize the task path of scrapped meters to reduce processing time; Dynamically adjust resource allocation strategies and feed back scheduling results to the sorting rule generation and optimization module.

5. The artificial intelligence-based return smart meter sorting management system according to claim 1 is characterized in that: The anomaly detection and warning module generates anomaly warnings by combining the following mechanisms: Detect abnormal patterns in running data through variational autoencoders and calculate dynamic thresholds for abnormal distributions; Use causal inference models to analyze the root causes of abnormalities and combine multi-scale entropy to identify signals that the system is close to failure; Generates warning information including exception number, exception impact range and recommended adjustment measures, and feeds back to the task scheduling and resource management module.

Images