An intelligent management system for after-sales maintenance work orders

Abstract

The present invention relates to the technical field of maintenance work order management, and in particular to an intelligent management system for after-sales maintenance work orders, which implements equipment clustering using the K-means algorithm, takes into account the differences in different user habits and usage environments, and improves the accuracy of fault diagnosis. By combining association rule mining with graph convolutional networks to construct a fault propagation network, the fault impact path between hardware is characterized, thereby achieving accurate identification of target faulty hardware, and intelligently determining alternative hardware in combination with recent maintenance records to avoid misjudgment and repeated repairs. The intelligent dispatching module matches the location information of maintenance personnel with real-time spare parts inventory, giving priority to dispatching nearby maintenance personnel with the required spare parts, thereby improving maintenance response efficiency. The model error correction module continuously monitors the operating status of the equipment after the maintenance is completed. If the predicted result does not match the actual maintenance, the node features and edge weight parameters are automatically adjusted, and the graph convolutional model is updated through incremental learning to improve the model's adaptability and evolution capabilities.

Description

Technical Field

[0001] The present invention relates to the technical field of maintenance work order management, and in particular to an intelligent management system for after-sales maintenance work orders. Background Art

[0002] Fault detection and after-sales repair of household appliances have become a major issue affecting user experience and after-sales service efficiency. Traditional after-sales repair services often rely on users actively reporting problems and maintenance personnel manually troubleshooting. This is not only inefficient, but also suffers from problems such as insufficient fault diagnosis accuracy, high rates of repeated repairs, and slow repair response speeds. To improve the service quality and response speed of after-sales repairs, existing technical solutions have attempted to use Internet of Things technology to achieve preliminary intelligent dispatching and equipment status monitoring. However, existing intelligent maintenance solutions generally only make fault judgments based on simple device parameter threshold warnings or general artificial intelligence classification models, ignoring the differences between device groups under different usage habits and scenarios, and the correlation between fault propagation paths and specific hardware. As a result, in actual applications, there are problems such as inaccurate fault identification, frequent repeated repairs, and unreasonable repair dispatching.

[0003] Furthermore, existing technologies fail to fully consider the recent repair history of the hardware being repaired. This often results in multiple repairs for the same hardware, but the actual fault occurs in other related hardware, leading to increased repair costs and reduced user satisfaction. Furthermore, current technical solutions lack effective model feedback and error correction mechanisms. Once a misdiagnosis occurs, it is difficult to automatically update the model, resulting in a high probability of continued misjudgment, which limits the system's adaptability and sustainable optimization capabilities. Summary of the Invention

[0004] In order to solve the above problems, the present invention provides an intelligent management system for after-sales maintenance work orders.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] An intelligent management system for after-sales maintenance work orders, including: an equipment clustering module, a fault network model building module, a fault hardware positioning module, an intelligent dispatching module, and a model error correction module;

[0007] The device clustering module is used to collect the operating parameters, environmental parameters and error logs of several devices through the cloud, and perform cluster analysis on the devices based on the K-means clustering algorithm to obtain several device clusters;

[0008] The fault network model construction module is used to establish an initial fault propagation network based on the operating parameters, environmental parameters and error logs of the devices in the device cluster through an association rule mining algorithm, and optimize the node features and edge weights of the initial fault propagation network through a graph convolutional network to obtain a fault network model;

[0009] The fault hardware location module is used to extract fault characteristics based on the current operating parameters and error logs of the faulty device when a faulty device is discovered in real time, and to determine the target faulty hardware through the fault network model;

[0010] The intelligent dispatching module is used to determine the alternative faulty hardware associated with the target faulty hardware through the fault network model based on whether the target faulty hardware has been repaired or replaced recently, and to dispatch work orders by comparing the real-time spare parts inventory of maintenance personnel within a preset distance, and to update the spare parts inventory information after the repair is completed;

[0011] The model error correction module is used to detect the status of the monitoring equipment after maintenance and to correct the node and edge weights of the fault network model.

[0012] Furthermore, the collection of operating parameters, environmental parameters and error logs of several devices includes the following steps:

[0013] The built-in monitoring device of the acquisition equipment collects voltage, current, temperature, operating time, operating power and start-stop frequency parameters;

[0014] The environmental sensors deployed on the equipment collect the ambient temperature and humidity parameters in real time;

[0015] The monitoring code records the error logs generated by the device in real time, including the error code, occurrence time and fault description information.

[0016] Furthermore, the clustering analysis of the devices based on the K-means clustering algorithm includes the following steps:

[0017] Convert the device's operating parameters, environmental parameters, and error logs into feature vectors to form a feature dataset to be clustered.

[0018] Initialize the cluster center, treat each device sample in the cluster feature dataset, calculate its distance to each cluster center, and assign it to the corresponding device cluster based on the minimum distance;

[0019] Update the cluster center based on the current device cluster, recalculate the distance between all device samples and the cluster center, and update the allocation results;

[0020] Repeat the cluster center update and device cluster reallocation operations until the cluster center converges and no longer changes, and obtain the device cluster division result.

[0021] Furthermore, establishing an initial fault propagation network by using an association rule mining algorithm based on the operating parameters, environmental parameters, and error logs of the devices in the device cluster includes the following steps:

[0022] Standardize the operating parameters, environmental parameters, and error logs of each device in the device cluster to build a structured transaction data set;

[0023] Extracting high-frequency association patterns between different hardware components when failures occur by using an Apriori algorithm or mining frequent item sets from the transaction data set;

[0024] Based on frequent itemsets, an initial fault propagation network is constructed with hardware components as nodes and fault co-occurrence relationships as edges. The support and confidence of the itemsets are used as the initial weights of the edges, and the initial fault propagation network is output.

[0025] Furthermore, the optimizing training of node features and edge weights of the initial fault propagation network through a graph convolutional network includes the following steps:

[0026] For each hardware node in the initial fault propagation network, we perform statistical calculations based on the abnormality level of the corresponding device's operating parameters and environmental parameters, as well as the frequency of specific error logs when any hardware failure occurs in the device cluster. This generates a historical abnormality feature vector that characterizes the hardware node and constructs the input feature matrix of the graph convolutional network.

[0027] Construct an adjacency matrix based on the fault co-occurrence relationships between hardware nodes in the initial fault propagation network;

[0028] The input feature matrix and adjacency matrix are input into the graph convolutional network model, and the graph convolution operation is performed to aggregate and propagate node features, and the target fault hardware is output;

[0029] Build a labeled dataset based on historically known faulty hardware, use supervised learning to define the loss function, and train and optimize the graph convolutional network parameters through the backpropagation algorithm;

[0030] Generate a fault network model after training and optimization.

[0031] Furthermore, the convolution operation formula of the graph convolutional network is as follows:

[0032] ;

[0033] in, For the The node feature matrix of the layer graph convolutional network, For the The node feature matrix of the layer graph convolutional network, when When it is 0, is the input feature matrix; Add the self-connected matrix to the k-th order adjacency matrix; is the corresponding degree matrix; For the The trainable weight matrix of the k-th propagation layer; For the The trainable weight matrix of the layer residual connection path; is the k-th order neighbor propagation weight coefficient; is the total number of orders; is a non-linear activation function.

[0034] Furthermore, the extraction of fault features based on the current operating parameters and error logs of the faulty device and determination of target faulty hardware through the fault network model includes the following steps:

[0035] For devices that experience real-time malfunctions, collect their current operating and environmental parameters, and extract the current error types and frequencies from the error logs.

[0036] Based on the historical average values of the normal operating parameters and environmental parameters of the device cluster to which the device belongs, the abnormal deviation degree of the current device operating parameters and environmental parameters is calculated to form the feature vector of the current faulty device;

[0037] The feature vector of the current faulty device is converted into the input feature matrix of the current faulty device, which is then input into the trained and optimized fault network model. The fault probability of each hardware node is calculated through graph convolutional network propagation.

[0038] The hardware node with the highest failure probability is output as the target failure hardware.

[0039] Furthermore, determining candidate faulty hardware associated with the target faulty hardware through the fault network model according to whether the target faulty hardware has been recently repaired or replaced includes the following steps:

[0040] After determining the target faulty hardware, query the recent repair or replacement records of the target faulty hardware to determine whether the target faulty hardware has been repaired or replaced within a preset period;

[0041] If the target faulty hardware has been repaired or replaced, then the fault network model is used to obtain the associated hardware nodes that have a direct edge connection with the target faulty hardware and whose edge weight is higher than a preset threshold;

[0042] The associated hardware nodes are sorted according to their failure probabilities, and the top two associated hardware nodes are selected as candidate failure hardware outputs.

[0043] Furthermore, the dispatching of work orders based on the real-time spare parts inventory of maintenance personnel within a preset distance and updating of spare parts inventory information after maintenance is completed include the following steps:

[0044] Based on the real-time location of the faulty equipment, the location information of maintenance personnel within a preset distance and real-time spare parts inventory information are obtained;

[0045] If more than one qualified maintenance personnel are screened out, they are sorted according to their distance from the faulty equipment, and the work order is assigned to the maintenance personnel closest to them.

[0046] After the maintenance personnel complete the maintenance, the maintenance personnel's spare parts inventory information is updated in real time through the mobile terminal, and the updated spare parts inventory data is uploaded for synchronous update.

[0047] Furthermore, the model error correction module is used to perform the following steps:

[0048] After the maintenance is completed, the equipment's operating parameters, environmental parameters and error logs are collected in real time, and the equipment's operating status is monitored within a preset time period;

[0049] If the device does not report the same type of error or abnormality again within the preset time period, and the hardware actually repaired is inconsistent with the target faulty hardware predicted by the fault network model, then execute S1-S2; if the device still reports the same type of error or abnormality again within the preset time period, or the hardware actually repaired is consistent with the target faulty hardware predicted by the fault network model, then no operation is performed;

[0050] S1. Based on the node positions of the actual repaired hardware and the target faulty hardware in the network, error correction is performed on the feature vector of the target faulty hardware node, including reducing the abnormality value in the feature vector corresponding to the target faulty hardware node and increasing the abnormality value in the feature vector corresponding to the actual repaired hardware node;

[0051] S2. Directively adjust the edge weights of the adjacency matrix between the target faulty hardware and the actual maintenance hardware nodes based on the current prediction deviation, including reducing the edge weights with the target faulty hardware and increasing the edge weights with the actual maintenance hardware nodes.

[0052] The present invention has the following beneficial effects: by collecting operating parameters, environmental parameters, and error logs of household appliances, it accurately divides device groups using the K-means clustering algorithm, taking into account the differences in user habits and usage scenarios. Then, based on the characteristic data of the device group, it uses an association rule mining algorithm and a graph convolutional network to accurately construct a device fault propagation network, effectively identifying the correlations between hardware components and the fault propagation paths, and achieving high-precision localization of the target hardware fault. Integrating the recent repair history of the target faulty hardware, the fault network model intelligently determines alternative faulty hardware, avoiding repeated repairs of the same hardware and optimizing dispatch decisions. The intelligent dispatch module further intelligently matches the most suitable repair personnel based on real-time location information and the repair personnel's spare parts inventory, minimizing response time and improving repair efficiency. Furthermore, this solution incorporates a model error correction module that monitors the operating status of the device after repair and automatically corrects the model's feature vectors and edge weight parameters, continuously improving model accuracy and adaptability, and effectively reducing the model's false positive rate. Compared with traditional after-sales repair services, this solution significantly improves fault diagnosis accuracy and repair response speed, reduces the number of repeated repairs, lowers repair costs, and effectively improves user experience and satisfaction, thus possessing broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a structural diagram of an intelligent management system for after-sales maintenance work orders in the present invention.

[0054] Figure 2 This is a flowchart of the steps of optimizing the node features and edge weights of the initial fault propagation network through a graph convolutional network in the present invention. DETAILED DESCRIPTION

[0055] See also Figure 1-Figure 2 As shown, the present invention relates to an intelligent management system for after-sales maintenance work orders, comprising: an equipment clustering module, a fault network model building module, a fault hardware positioning module, an intelligent dispatching module, and a model error correction module;

[0056] The device clustering module is used to collect the operating parameters, environmental parameters and error logs of several devices through the cloud, and perform cluster analysis on the devices based on the K-means clustering algorithm to obtain several device clusters;

[0057] The fault network model construction module is used to establish an initial fault propagation network based on the operating parameters, environmental parameters and error logs of the devices in the device cluster through an association rule mining algorithm, and optimize the node features and edge weights of the initial fault propagation network through a graph convolutional network to obtain a fault network model;

[0058] The fault hardware location module is used to extract fault characteristics based on the current operating parameters and error logs of the faulty device when a faulty device is discovered in real time, and to determine the target faulty hardware through the fault network model;

[0059] The intelligent dispatching module is used to determine the alternative faulty hardware associated with the target faulty hardware through the fault network model based on whether the target faulty hardware has been repaired or replaced recently, and to dispatch work orders by comparing the real-time spare parts inventory of maintenance personnel within a preset distance, and to update the spare parts inventory information after the repair is completed;

[0060] The model error correction module is used to detect the status of the monitoring equipment after maintenance and to correct the node and edge weights of the fault network model.

[0061] In some embodiments, the device clustering module first collects operating parameters of household appliances such as air conditioners, washing machines, and refrigerators, such as current, voltage, and temperature; environmental parameters such as indoor temperature and humidity; and error logs such as startup failures and overload protection. The system uses feature engineering methods to map this raw data into high-dimensional feature vectors, and then performs unsupervised classification of the devices using the K-means clustering algorithm. The K value is determined by combining the silhouette coefficient and the elbow method to ensure that the clustering results are compact and separable in terms of classification. The resulting device clusters are used to construct a more similar fault propagation network structure. Next, the fault network model construction module first uses the Apriori algorithm to mine frequent item set relationships between faulty components in a structured transaction dataset based on the data distribution within the device cluster. This is used to construct an initial fault propagation network, where nodes represent specific hardware components, edges represent co-occurrence relationships between faults, and edge weights are determined by support and confidence. This network structure, along with historical anomaly statistics corresponding to each node (hardware) (such as the failure frequency of a component in the device cluster, the average deviation of corresponding device parameters, and the frequency of associated error logs), is then used as input to the graph convolutional network (GCN). The GCN employs a multi-layer propagation structure to aggregate and propagate node features. During training, the labels of target hardware nodes are optimized through supervised learning. The loss function is designed to be a cross-entropy plus an L2 regularization term to prevent overfitting. Ultimately, a fault network model with reasoning capabilities is output. When the system detects a device failure, the faulty hardware localization module collects its operating parameters and error logs in real time and extracts current state features. By calculating the deviation of these state features from the historical distribution of the device cluster, it generates the current input feature vector and inputs it into the trained fault network model. Through multi-layer adjacency matrix propagation and node weight iteration, the GCN outputs a failure probability distribution for each node (i.e., each hardware component). The system ultimately identifies the node with the highest probability as the target faulty hardware. After obtaining the target faulty hardware, the intelligent dispatching module first determines whether the hardware has been repaired or replaced within a preset timeframe. If so, it indicates a possible misjudgment or secondary fault. The system will call the fault network model to output its adjacent nodes with high edge weights as candidate faulty hardware and rank them based on their current failure probability. The system then obtains the real-time location information and spare parts list of all maintenance personnel within the geographic range. It calculates the maintenance scheduling priority by combining spatial distance and inventory matching functions, prioritizing work orders to personnel who are closer and carrying the target spare parts, improving response efficiency and reducing resource waste. After the maintenance is completed, the system automatically collects and uploads the used spare parts and current remaining inventory via mobile terminals, updating the central inventory information database in real time.To enhance the system's self-learning capabilities and predictive reliability, the model error correction module performs state retrospective monitoring on all repaired equipment. If the target faulty hardware predicted by the system is inconsistent with the actual repaired hardware, and the equipment does not experience recurring anomalies within the preset time window, it is considered a valid error correction signal. The system will automatically reduce the abnormality intensity of the original prediction node, enhance the abnormal characteristics of the actual repair node, and trigger incremental training of the graph convolutional network based on the current correction sample, updating the node weights and edge weight structure, so that the model gradually converges to the actual fault logic. Through the collaborative operation of the above modules, this system significantly outperforms existing solutions in terms of equipment fault identification accuracy, work order dispatching efficiency, and model adaptive optimization capabilities. The equipment clustering module builds a differentiated foundation and enhances the generalization ability of the model; the fault network model depicts the fault logic path, realizing predictive reasoning from a global perspective; intelligent dispatching and inventory linkage reduce scheduling costs, and form a data closed-loop feedback loop between modules.

[0062] Furthermore, the collection of operating parameters, environmental parameters and error logs of several devices includes the following steps:

[0063] The built-in monitoring device of the acquisition equipment collects voltage, current, temperature, operating time, operating power and start-stop frequency parameters;

[0064] The environmental sensors deployed on the equipment collect the ambient temperature and humidity parameters in real time;

[0065] The monitoring code records the error logs generated by the device in real time, including the error code, occurrence time and fault description information.

[0066] Specifically, the system first built a multi-source data acquisition architecture to support high-frequency and high-precision acquisition of equipment operating parameters, environmental parameters and error logs. Taking an embedded intelligent air-conditioning device as an example, the device's built-in operation monitoring module includes multiple analog signal acquisition channels and digital signal processing units. The air-conditioning device contains multiple hardware such as main control chips, power modules, compressors, motors and other key components. Specifically, the system uses high-precision sensor modules to collect the voltage, current and temperature parameters of the device. The voltage and current can be used to reflect the load status and power consumption, and the temperature can be used to monitor whether there are thermal protection risks. In addition, the system also periodically reads the operating time and start-stop frequency, which are used to determine the continuous working time and frequent start-stop behavior of the equipment, thereby assisting in determining whether there are any operational anomalies.

[0067] In terms of environmental awareness, the system relies on environmental sensor units deployed around the equipment to collect real-time data on the ambient temperature and relative humidity of the space where the equipment is located. This data is crucial for determining the cause of a fault. For example, compressor overheating may be caused by poor heat dissipation due to excessively high indoor temperature or low humidity. Combining analysis of environmental parameter trends with equipment operating behavior provides a critical baseline for subsequent clustering and anomaly identification.

[0068] Furthermore, at the software level, the system uses an embedded monitoring code module to collect logs of abnormal events during device operation. These logs include the error code, the time the fault occurred, and the fault description returned by the system. This monitoring module, integrated into the device's operating system or application-layer program, automatically records and pushes event information to the cloud platform upon detection of a fault event, forming a critical source of sequential fault records. Error codes adhere to a unified coding standard and can be directly mapped to component identifiers in the device structure diagram, enabling subsequent annotation and label extraction of fault network models.

[0069] Furthermore, the clustering analysis of the devices based on the K-means clustering algorithm includes the following steps:

[0070] Convert the device's operating parameters, environmental parameters, and error logs into feature vectors to form a feature dataset to be clustered.

[0071] Initialize the cluster center, treat each device sample in the cluster feature dataset, calculate its distance to each cluster center, and assign it to the corresponding device cluster based on the minimum distance;

[0072] Update the cluster center based on the current device cluster, recalculate the distance between all device samples and the cluster center, and update the allocation results;

[0073] Repeat the cluster center update and device cluster reallocation operations until the cluster center converges and no longer changes, and obtain the device cluster division result.

[0074] In a specific embodiment, taking a certain brand of smart washing machine as an example, its raw collected data includes operating parameters (such as current, voltage, drum motor load, heating power, program duration, etc.), environmental parameters (such as indoor temperature and humidity), and error logs (such as abnormal water inflow, dehydration failure, door lock abnormality, etc.). To achieve numerical consistency and feature representation accuracy, the system uses Z-score normalization to normalize continuous parameters to ensure comparability between different dimensions. At the same time, error logs are converted into sparse vectors using one-hot encoding, and the TF-IDF (TermFrequency-Inverse Document Frequency) weighting is then used to enhance the discriminative power of fault information in the clustering process. Ultimately, each device is represented as a set of fixed-dimensional feature vectors, forming the feature dataset input to the K-means algorithm.

[0075] During the initialization phase of the K-means clustering algorithm, the system selects initial cluster centers using the K-means++ strategy to mitigate local optima caused by random initialization. For each device sample in the feature dataset, the system uses Euclidean distance as a similarity metric to calculate its distance from all cluster centers and assigns the sample to the device cluster represented by the closest cluster center. After the initial cluster assignment is complete, the system averages the features of the member samples in each cluster and updates the cluster center coordinates. The distances between all samples and the new cluster centers are then recalculated and the cluster assignments are re-performed. This process is repeated until the position change of all cluster centers in two consecutive iterations is less than the preset convergence threshold, or the maximum number of iterations is reached. The final device clustering results are then output. In actual deployments, the system sets an appropriate K value for different device categories. The K value is selected based on a comprehensive consideration of the Silhouette Coefficient, Calinski-Harabasz Index, and the Elbow method, and is optimized offline before deployment. The final output equipment cluster not only improves the accuracy of subsequent graph neural network modeling of fault relationships, but also serves as structural support for upper-level functions such as maintenance cycle prediction and failure mode migration analysis under differentiated strategies.

[0076] Furthermore, establishing an initial fault propagation network by using an association rule mining algorithm based on the operating parameters, environmental parameters, and error logs of the devices in the device cluster includes the following steps:

[0077] Standardize the operating parameters, environmental parameters, and error logs of each device in the device cluster to build a structured transaction data set;

[0078] Extracting high-frequency association patterns between different hardware components when failures occur by using an Apriori algorithm or mining frequent item sets from the transaction data set;

[0079] Based on frequent itemsets, an initial fault propagation network is constructed with hardware components as nodes and fault co-occurrence relationships as edges. The support and confidence of the itemsets are used as the initial weights of the edges, and the initial fault propagation network is output.

[0080] It should be noted that each device's operating parameters, environmental parameters, and error logs are first uniformly coded and standardized. The operating and environmental parameters are discretized using an interval discretization strategy, classifying continuous numerical variables into discrete labels such as "normal," "high," or "low" based on their statistical distribution within the device cluster. This enhances the identifiability of transaction items. For example, if the mean current in a device cluster is 2.5A and the standard deviation is 0.3A, a device recording a value of 3.1A would be labeled "high current" in that dimension. Error logs are converted into independent fault event labels based on error type (e.g., E01 - temperature control failure, E02 - fan abnormality). These labels collectively constitute the device's transaction item set, resulting in a complete, structured transaction dataset for the device cluster. After constructing the transaction dataset, the system uses the Apriori algorithm to mine frequent itemsets within the transaction set. The Apriori algorithm uses an iterative pruning method to ensure frequent item set coverage while improving computational efficiency. In each iteration, the system evaluates the support and confidence metrics of item sets to identify frequently co-occurring fault event combinations. For example, if "compressor current is too high" and "main control board fault alarm" frequently co-occur in transactions across multiple devices and meet the set thresholds of support greater than 10% and confidence greater than 60%, this combination is considered a valid fault association rule. Based on the mined frequent item sets, the system constructs an initial fault propagation network diagram. Each node in the network represents a specific hardware component or its corresponding typical fault event, and edges between nodes represent co-occurrence or potential causal relationships between fault events. Edge weights are initially assigned based on the support and confidence of the corresponding frequent item set. Higher support indicates a more common fault mode within the device cluster, while higher confidence indicates a greater likelihood that one fault will trigger another, thereby enhancing the authority and reliability of the propagation path. Taking a cluster of electric water heaters as an example, transaction mining revealed that "temperature sensor reading is too low" and "heating pipe abnormal alarm" co-occur frequently, with a support of 12% and a confidence of 78%. The system establishes a directed edge between the "temperature sensor node" and the "heating tube node", and the edge weight is initialized to the weighted function value of the group of data (such as support × confidence) for subsequent training and optimization of the graph convolutional network model.

[0081] Furthermore, the optimizing training of node features and edge weights of the initial fault propagation network through a graph convolutional network includes the following steps:

[0082] For each hardware node in the initial fault propagation network, we perform statistical calculations based on the abnormality level of the corresponding device's operating parameters and environmental parameters, as well as the frequency of specific error logs when any hardware failure occurs in the device cluster. This generates a historical abnormality feature vector that characterizes the hardware node and constructs the input feature matrix of the graph convolutional network.

[0083] Construct an adjacency matrix based on the fault co-occurrence relationships between hardware nodes in the initial fault propagation network;

[0084] The input feature matrix and adjacency matrix are input into the graph convolutional network model, and the graph convolution operation is performed to aggregate and propagate node features, and the target fault hardware is output;

[0085] Build a labeled dataset based on historically known faulty hardware, use supervised learning to define the loss function, and train and optimize the graph convolutional network parameters through the backpropagation algorithm;

[0086] Generate a fault network model after training and optimization.

[0087] In some embodiments, the system uses a graph convolutional network to optimize and train the node features and edge weights in the initial fault propagation network to improve fault identification accuracy and model generalization. Specifically, a feature representation is first constructed for each hardware node in the initial fault propagation network. This feature vector is not derived from data collected from the hardware itself, but rather from statistics on the degree of abnormality in the overall operating parameters and the degree of deviation in environmental parameters when the hardware fault occurred within the device cluster in which it resides, as well as the frequency of occurrence of error logs corresponding to the hardware fault. Taking air conditioning compressors as an example, the system selects sample devices in the device cluster that have previously undergone maintenance due to compressor failures and extracts the deviations in operating current, voltage, and power, the trends in indoor and outdoor temperature differences, and the concentration of related error codes before the failure. By digitizing and standardizing these abnormal features, a feature vector is generated that represents the historical abnormal behavior of the hardware. Next, the system constructs an adjacency matrix based on the co-occurrence relationships between the nodes in the initial fault propagation network. Each edge represents the frequent co-occurrence of faults between two hardware components in the historical device cluster, and its weight is determined by the support and confidence obtained from frequent item set mining. This adjacency matrix not only describes the interconnected structure between hardware components but also serves as the core foundation for subsequent feature aggregation and propagation within the graph neural network. During training, the system inputs the constructed node feature matrix and adjacency matrix into the graph convolutional network. The network achieves feature enhancement by aggregating information from the node's neighbors layer by layer, ensuring that each node's final representation not only incorporates its own historical anomaly information but also incorporates the impact of failures on its associated nodes. Through supervised learning, the network is trained to predict faulty nodes using a fault label dataset constructed from historical maintenance information. The loss function uses cross-entropy to minimize the error between the fault prediction and the actual label. The training process uses backpropagation and gradient descent algorithms for iterative parameter updates. This approach significantly differs from existing technologies in several ways. Firstly, the system does not rely on the traditional method of determining whether a device is faulty based on overall device thresholds. Instead, it builds a graph network model with structural recognition capabilities based on group behavior patterns between devices and components. Secondly, the introduction of structural dependencies within the graph convolutional neural network enables the system to maintain a certain level of reasoning capabilities even when data is incomplete or individual hardware signals are not significant, effectively improving diagnostic accuracy in complex scenarios. Through the deep fusion of graph structure information and node features, this embodiment achieves efficient integration and cross-modeling of multi-source information, further enhances the context sensitivity of fault identification, and forms a high-dimensional reasoning model that is different from the traditional flat processing method.

[0088] Furthermore, the convolution operation formula of the graph convolutional network is as follows:

[0089] ;

[0090] in, For the The node feature matrix of the layer graph convolutional network, For the The node feature matrix of the layer graph convolutional network, when When it is 0, is the input feature matrix; Add the self-connected matrix to the k-th order adjacency matrix; is the corresponding degree matrix; For the The trainable weight matrix of the k-th propagation layer; For the The trainable weight matrix of the layer residual connection path; is the k-th order neighbor propagation weight coefficient; is the total number of orders; is a non-linear activation function.

[0091] It should be noted that For graph convolutional network The node feature matrix of the layer. Each row corresponds to a hardware component node, and each column is an eigenvalue of the node. In this scheme, the initial It is a vector matrix constructed based on statistical characteristics such as the degree of abnormality of the entire equipment's operating parameters, the degree of abnormality of environmental parameters, and the frequency of error logs when the component represented by the node fails. For the The node feature matrix of a layer graph convolutional network is the output of the current layer and the input of the next layer. This matrix reflects the aggregation result of node features after considering neighbor information. The k-th order adjacency matrix represents the connections between each node in the graph and its k-th order neighbors. It also adds self-connections (i.e., 1s are added to the diagonal) to ensure that each node retains its own information during propagation. In this scheme, order 0 represents the node itself, order 1 represents directly connected hardware components, and order 2 represents indirect fault paths that propagate through an intermediate hardware component. is the k-th order adjacency matrix The degree matrix of , which corresponds to the diagonal matrix of the number of connections (i.e., the number of neighbors) of each node. It is used to normalize , to prevent nodes with many connections from having too large a weight during propagation. is the propagation weight coefficient of the kth order adjacent path, which can be trained to control the contribution of different order propagations to the final feature. In this solution, this parameter reflects the different importance of fault propagation on direct paths and indirect paths. For example, when the first order adjacent propagation is more representative than the second order propagation, the system can automatically learn to increase The weight of . Control the conversion of neighbor information in the feature dimension. In actual training, the system will automatically learn the optimal weights based on the fault propagation patterns between different hardware in historical data. It has input features Directly apply a linear transformation and then add it to the convolution result. Residual connections help preserve original features, alleviate the vanishing gradient problem in deep neural networks, and improve training stability. It is a ReLU nonlinear activation function, which is used to introduce nonlinear feature transformation to make the model have stronger fitting ability.

[0092] Furthermore, the extraction of fault features based on the current operating parameters and error logs of the faulty device and determination of target faulty hardware through the fault network model includes the following steps:

[0093] For devices that experience real-time malfunctions, collect their current operating and environmental parameters, and extract the current error types and frequencies from the error logs.

[0094] Based on the historical average values of the normal operating parameters and environmental parameters of the device cluster to which the device belongs, the abnormal deviation degree of the current device operating parameters and environmental parameters is calculated to form the feature vector of the current faulty device;

[0095] The feature vector of the current faulty device is converted into the input feature matrix of the current faulty device, which is then input into the trained and optimized fault network model. The fault probability of each hardware node is calculated through graph convolutional network propagation.

[0096] The hardware node with the highest failure probability is output as the target failure hardware.

[0097] Specifically, the system first automatically retrieves the device's operating and environmental parameter data at the time of the fault. Operating parameters include but are not limited to real-time voltage, current, temperature, power, operating time, and start / stop frequency; environmental parameters primarily include the temperature and humidity of the device's operating environment. Simultaneously, the system extracts error log data recorded at the time of the device fault, including the type of error code, recurrence frequency, and time series distribution characteristics. Next, the system compares the currently collected operating and environmental parameters of the faulty device with the mean and variance of the operating parameters under historical normal conditions within its device cluster. By calculating statistical deviation metrics such as the Z-score, the system quantifies the degree of abnormality of the current state compared to the average state of similar devices, thereby constructing a multidimensional feature vector representing the device's current fault state. This feature vector not only includes the numerical offset of the operating metrics but also incorporates the impact of environmental fluctuations on device performance and the temporal and frequency patterns of log events, ensuring a timely and discriminative input data structure. This feature vector is encoded as the input feature matrix for the currently faulty device and fed into the trained graph convolutional network model. This graph convolutional network constructs a fault propagation graph based on historical multi-device, multi-fault collaborative data. Each node in the graph represents a hardware component, and edge weights represent its co-occurrence probability with other hardware components in a fault event. The graph convolution operation combines the feature vector of the current faulty device with the structural connectivity of each node in the graph, performing layer-by-layer information propagation and aggregation, ultimately generating a fault probability prediction for each hardware node at the output layer. The system ranks the failure probabilities of all hardware nodes and selects the node with the highest probability as the target faulty hardware. This fault location method combines the manifestation of current device state anomalies with the commonality of historical failure patterns across multiple devices, accurately identifying the source of internal faults without requiring individual sensor monitoring of each hardware component. The approach adopted in this embodiment offers significant advantages over existing solutions that rely on static thresholds and a single classification output based on a full-device model. By comparing and normalizing the current device state with the evolutionary characteristics of a large-scale device cluster, and then embedding it into a graph structure model for associative reasoning, this method not only improves fault location accuracy but also enhances the system's adaptability to complex, multi-source, heterogeneous device operating environments. The algorithmic structure and application results demonstrate the innovation and practical value of this invention.

[0098] Furthermore, determining candidate faulty hardware associated with the target faulty hardware through the fault network model according to whether the target faulty hardware has been recently repaired or replaced includes the following steps:

[0099] After determining the target faulty hardware, query the recent repair or replacement records of the target faulty hardware to determine whether the target faulty hardware has been repaired or replaced within a preset period;

[0100] If the target faulty hardware has been repaired or replaced, then the fault network model is used to obtain the associated hardware nodes that have a direct edge connection with the target faulty hardware and whose edge weight is higher than a preset threshold;

[0101] The associated hardware nodes are sorted according to their failure probabilities, and the top two associated hardware nodes are selected as candidate failure hardware outputs.

[0102] Specifically, after the system preliminarily identifies the target faulty hardware node based on the output of the fault network model, it accesses the historical equipment maintenance database to search for repair and replacement records for that hardware within a preset time window (e.g., the past 30 days). If the target faulty hardware node is found to have been recently replaced or repaired, the system further initiates the associated reasoning process for candidate hardware to mitigate the risk of ineffective replacement or misidentification of the primary cause. During the associated hardware identification process, the system first retrieves other hardware nodes in the fault network model that have direct edges connected to the target faulty hardware node and reads the weight of these edges. These edge weights, optimized during graph convolution training, effectively represent the strength of the interconnectedness between the two hardware components in historical failure samples. The system compares these edge weights with a preset threshold (e.g., 0.6) and retains only those hardware nodes with edge weights greater than the threshold as candidates for the current target faulty hardware node. The system then re-evaluates the failure probability scores of these candidate hardware nodes based on the input features of the current faulty device using the fault network model. This evaluation utilizes the aforementioned graph convolution propagation mechanism, that is, the candidate node will calculate the probability of failure in the current scenario based on the similarity between itself and the current feature matrix of the faulty device, the adjacency weight on the propagation path, and its historical anomaly vector. Ultimately, the system outputs the hardware nodes with the top two failure probabilities in the candidate set as candidate faulty hardware. This method solves the problem of the inability to identify replacement failures or fault transfer situations in traditional solutions by introducing the strategy of "proximate cause denial + graph association confirmation". It is particularly suitable for scenarios with high coupling between modules in complex home appliance systems. Compared with the existing dispatching logic that only relies on single-point prediction of fault probability or does not consider historical maintenance behavior, this embodiment fully demonstrates the advantages of combining graph structure modeling with state dynamic analysis, and realizes the closed loop of "fault prediction + misjudgment avoidance" from the perspective of algorithm collaboration and actual application effect, significantly improving the system's intelligent decision-making capabilities and user satisfaction.

[0103] Furthermore, the dispatching of work orders based on the real-time spare parts inventory of maintenance personnel within a preset distance and updating of spare parts inventory information after maintenance is completed include the following steps:

[0104] Based on the real-time location of the faulty equipment, the location information of maintenance personnel within a preset distance and real-time spare parts inventory information are obtained;

[0105] If more than one qualified maintenance personnel are screened out, they are sorted according to their distance from the faulty equipment, and the work order is assigned to the maintenance personnel closest to them.

[0106] After the maintenance personnel complete the maintenance, the maintenance personnel's spare parts inventory information is updated in real time through the mobile terminal, and the updated spare parts inventory data is uploaded for synchronous update.

[0107] In some embodiments, once faulty hardware or candidate faulty hardware is identified, the system first determines the spatial location of the faulty device based on the location coordinates uploaded by the device terminal, combined with GIS positioning services or deployed geocoding modules. The system then retrieves the real-time location information of all current maintenance personnel and the spare parts inventory list uploaded by their mobile terminals or the system. This inventory data is dynamically synchronized with the personnel's smart spare parts box via their mobile device, containing the quantity, model, and batch information of each hardware module, and can be uploaded to the cloud in real time for centralized management. The system then selects maintenance personnel within a preset service radius (e.g., 5 kilometers, 8 kilometers, etc.) and determines whether they are carrying the target faulty hardware and its candidate hardware. If multiple maintenance personnel meet the requirements, they are further ranked based on their current spatial distance to the faulty device, with work orders being assigned to those closest to the faulty device, minimizing response time and avoiding resource waste. After the maintenance task is completed, the maintenance personnel enter the type and quantity of hardware actually replaced via their mobile device, and the system deducts the amount from the maintenance personnel's local inventory record in real time. Updated inventory data is also synchronized to the platform backend for inventory verification and replenishment warning management for subsequent tasks, ensuring the consistency and timeliness of inventory data. Through this implementation path, the module implements a dual-factor dispatch optimization strategy based on location constraints and inventory matching. This differs from existing technologies that rely solely on extensive task scheduling based on regional division or manual assignment. This module offers the advantages of strong real-time performance, high decision-making accuracy, and efficient resource utilization, effectively improving overall after-sales maintenance efficiency and customer satisfaction.

[0108] Furthermore, the model error correction module is used to perform the following steps:

[0109] After the maintenance is completed, the equipment's operating parameters, environmental parameters and error logs are collected in real time, and the equipment's operating status is monitored within a preset time period;

[0110] If the device does not report the same type of error or abnormality again within the preset time period, and the hardware actually repaired is inconsistent with the target faulty hardware predicted by the fault network model, then execute S1-S2; if the device still reports the same type of error or abnormality again within the preset time period, or the hardware actually repaired is consistent with the target faulty hardware predicted by the fault network model, then no operation is performed;

[0111] S1. Based on the node positions of the actual repaired hardware and the target faulty hardware in the network, error correction is performed on the feature vector of the target faulty hardware node, including reducing the abnormality value in the feature vector corresponding to the target faulty hardware node and increasing the abnormality value in the feature vector corresponding to the actual repaired hardware node;

[0112] S2. Directively adjust the edge weights of the adjacency matrix between the target faulty hardware and the actual maintenance hardware nodes based on the current prediction deviation, including reducing the edge weights with the target faulty hardware and increasing the edge weights with the actual maintenance hardware nodes.

[0113] First, at the node feature vector level, the system weakens the abnormal representation of the target faulty hardware node. This is achieved by reducing the relevant components of the node's features that indicate the degree of abnormality. Specifically, the system identifies abnormal dimension features formed during the node's training process. For example, parameter dimensions with high fault frequency in historical statistics or high correlation with abnormal error logs are reduced by a certain percentage. The system then reduces the feature values of these dimensions by a certain percentage, thereby weakening the abnormal information propagated in subsequent graph convolution operations, thereby reducing the probability of the node being misidentified as a fault source again. Simultaneously, the system also enhances the corresponding abnormal features of the actual repaired hardware node to compensate for abnormal paths that the original model failed to correctly focus on. Second, at the graph structure level, specifically in terms of edge weight processing of the adjacency matrix, the system dynamically adjusts the edge weights between the target faulty hardware node and other nodes in the original model. If there is a connection between the original predicted node and the actual repaired node, the system reduces the edge weight coefficient between them to weaken their propagation influence. At the same time, the edge weight between the actual repaired node and its associated hardware nodes is increased to re-enforce structures that are more likely to constitute fault propagation paths. This directional edge weight fine-tuning not only corrects local propagation relationships, but also enables the graph structure to continuously evolve with actual maintenance feedback after training, gradually approaching the actual fault logic.

[0114] In some specific embodiments, the system first assesses the degree of deviation from the current prediction based on whether the actual maintenance node appears in the initial prediction ranking and the corresponding activation frequency of the node in historical samples. The system uses a three-level scale: Significant Deviation: If the actual maintenance node does not appear in the top 10 predicted nodes and its activation frequency in historical training data is less than 10%, the prediction is considered to be significantly deviating from the actual fault pattern. In this case, the anomaly weight of each dimension in the feature vector of the target faulty hardware node is reduced by 30%, while the weight of the core anomaly dimension corresponding to the actual maintenance node is increased by 40%. Moderate Deviation: If the actual maintenance node is ranked 5th to 10th in the prediction or has been partially misdiagnosed and activated in historical samples, the system considers it a moderate deviation. A milder adjustment strategy is then implemented, reducing the weight of the target node dimension by 15%-20% and increasing the weight of the actual maintenance node by 20%-25%. Minor Deviation: If the actual maintenance node is second only to the original target node or has an associated edge relationship with it, the deviation is considered minor, and only a small adjustment of up to 10% is made to the error-related dimension. The innovation of this multi-level adjustment mechanism lies in that it not only considers the direct difference between the current prediction results and the actual maintenance, but also combines the credibility of the historical characteristics of the nodes in the training samples, ensuring that the adjustment strategy has both punishment and compensation, and has the stability of long-term iterative optimization. In terms of edge weight adjustment, the system further introduces a "corrected edge gradient weight function", using the degree of prediction deviation between the two nodes in the edge connection as the driving factor. If there is a connecting edge between the target node and the actual maintenance node, the weight of the edge will be increased by 15%-25%, and the specific increase value is proportional to the contribution rate of the edge in the historical propagation path; conversely, if the target node has a high-weight edge with other unrelated nodes, the system will reversely reduce the edge weight according to its historical propagation success rate, and the reduction range is set to 10%-30%.

[0115] In summary, the error correction module of this embodiment achieves dynamic adaptive updating of the fault propagation model through the "feature correction + structural adjustment" method, breaking through the limitations of static model prediction and ineffective utilization of misdiagnosis feedback in the existing technology, enabling the system to have the ability of continuous learning and evolution, thereby realizing more efficient and accurate intelligent decision support in after-sales maintenance scenarios.

[0116] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present application may be implemented in various computer languages, such as C, VHDL, Verilog, object-oriented programming language Java, and interpreted scripting language JavaScript.

[0117] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0118] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0120] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. An intelligent management system for after-sales maintenance work orders, characterized by: include: Equipment clustering module, fault network model building module, fault hardware location module, intelligent dispatching module and model error correction module; The device clustering module is used to collect the operating parameters, environmental parameters and error logs of several devices through the cloud, and perform cluster analysis on the devices based on the K-means clustering algorithm to obtain several device clusters; The fault network model construction module is used to establish an initial fault propagation network based on the operating parameters, environmental parameters and error logs of the devices in the device cluster through an association rule mining algorithm, and optimize the node features and edge weights of the initial fault propagation network through a graph convolutional network to obtain a fault network model; The fault hardware location module is used to extract fault characteristics based on the current operating parameters, environmental parameters and error logs of the faulty device when a faulty device is discovered in real time, and to determine the target faulty hardware through the fault network model; The intelligent dispatching module is used to determine the alternative faulty hardware associated with the target faulty hardware through the fault network model based on whether the target faulty hardware has been repaired or replaced recently, and to dispatch work orders by comparing the real-time spare parts inventory of maintenance personnel within a preset distance, and to update the spare parts inventory information after the repair is completed; The model error correction module is used to detect the status of monitoring equipment after maintenance and to correct the node features and edge weights of the fault network model.

2. The intelligent management system for after-sales maintenance work orders according to claim 1, characterized in that: The collection of operating parameters, environmental parameters and error logs of several devices includes the following steps: The voltage, current, temperature, operating time, operating power and start-stop frequency parameters are collected through the equipment's built-in monitoring device; The environmental sensors deployed on the equipment collect the ambient temperature and humidity parameters in real time; The monitoring code records the error logs generated by the device in real time, including the error code, occurrence time and fault description information.

3. The intelligent management system for after-sales maintenance work orders according to claim 1, characterized in that: The clustering analysis of devices based on the K-means clustering algorithm includes the following steps: Convert the device's operating parameters, environmental parameters, and error logs into feature vectors to form a feature dataset to be clustered. Initialize the cluster center, treat each device sample in the cluster feature dataset, calculate its distance to each cluster center, and assign it to the corresponding device cluster based on the minimum distance; Update the cluster center based on the current device cluster, recalculate the distance between all device samples and the cluster center, and update the allocation results; Repeat the cluster center update and device cluster reallocation operations until the cluster center converges and no longer changes, and obtain the device cluster division result.

4. The intelligent management system for after-sales maintenance work orders according to claim 1, characterized in that: The method of establishing an initial fault propagation network based on the operating parameters, environmental parameters, and error logs of the devices in the device cluster by using an association rule mining algorithm comprises the following steps: Standardize the operating parameters, environmental parameters, and error logs of each device in the device cluster to build a structured transaction data set; Extracting high-frequency association patterns between different hardware components when failures occur by using an Apriori algorithm or mining frequent item sets from the transaction data set; Based on frequent itemsets, an initial fault propagation network is constructed with hardware components as nodes and fault co-occurrence relationships as edges. The support and confidence of the itemsets are used as the initial weights of the edges, and the initial fault propagation network is output.

5. The intelligent management system for after-sales maintenance work orders according to claim 4, characterized in that: The optimization training of node features and edge weights of the initial fault propagation network using a graph convolutional network includes the following steps: For each hardware node in the initial fault propagation network, we perform statistical calculations based on the abnormality level of the corresponding device's operating parameters and environmental parameters, as well as the frequency of specific error logs when any hardware failure occurs in the device cluster. This generates a historical abnormality feature vector that characterizes the hardware node and constructs the input feature matrix of the graph convolutional network. Construct an adjacency matrix based on the fault co-occurrence relationships between hardware nodes in the initial fault propagation network; The input feature matrix and adjacency matrix are input into the graph convolutional network model, and the graph convolution operation is performed to aggregate and propagate node features, and the target fault hardware is output; Build a labeled dataset based on historically known faulty hardware, use supervised learning to define the loss function, and train and optimize the graph convolutional network parameters through the backpropagation algorithm; Generate a fault network model after training and optimization.

6. The intelligent management system for after-sales maintenance work orders according to claim 5, characterized in that: The convolution operation formula of the graph convolutional network is as follows: ； in, For the The node feature matrix of the layer graph convolutional network, For the The node feature matrix of the layer graph convolutional network, when When it is 0, is the input feature matrix; Add the self-connected matrix to the k-th order adjacency matrix; is the corresponding degree matrix; For the The trainable weight matrix of the k-th propagation layer; For the The trainable weight matrix of the layer residual connection path; is the k-th order neighbor propagation weight coefficient; is the total number of orders; is a non-linear activation function.

7. The intelligent management system for after-sales maintenance work orders according to claim 5, characterized in that: Extracting fault features based on the current operating parameters, environmental parameters, and error logs of the faulty device and determining the target faulty hardware through the fault network model includes the following steps: For devices experiencing real-time faults, the system collects their current operating and environmental parameters. Based on the historical averages of the normal operating and environmental parameters of the device cluster to which the device belongs, the system calculates the degree of abnormal deviation of the current operating and environmental parameters. Furthermore, the system extracts the current error type and frequency from the error log to form a feature vector for the current faulty device. The feature vector of the current faulty device is converted into the input feature matrix of the current faulty device, which is then input into the trained and optimized fault network model. The fault probability of each hardware node is calculated through graph convolutional network propagation. The hardware node with the highest failure probability is output as the target failure hardware.

8. The intelligent management system for after-sales maintenance work orders according to claim 1, characterized in that: The determining of candidate faulty hardware associated with the target faulty hardware by using the fault network model according to whether the target faulty hardware has been recently repaired or replaced comprises the following steps: After determining the target faulty hardware, query the recent repair or replacement records of the target faulty hardware to determine whether the target faulty hardware has been repaired or replaced within a preset period; If the target faulty hardware has been repaired or replaced, then the fault network model is used to obtain the associated hardware nodes that have a direct edge connection with the target faulty hardware and whose edge weight is higher than a preset threshold; The associated hardware nodes are sorted according to their failure probabilities, and the top two associated hardware nodes are selected as candidate failure hardware outputs.

9. The intelligent management system for after-sales maintenance work orders according to claim 1, characterized in that: The method of dispatching work orders by comparing the real-time spare parts inventory of maintenance personnel within a preset distance and updating the spare parts inventory information after the maintenance is completed includes the following steps: Based on the real-time location of the faulty equipment, obtain the location information of maintenance personnel within a preset distance and real-time spare parts inventory information; If more than one qualified maintenance personnel are screened out, they are sorted according to their distance from the faulty equipment, and the work order is assigned to the maintenance personnel closest to them. After the maintenance personnel complete the maintenance, the maintenance personnel's spare parts inventory information is updated in real time through the mobile terminal, and the updated spare parts inventory data is uploaded for synchronous update.

10. The intelligent management system for after-sales maintenance work orders according to claim 1, characterized in that: The model error correction module is used to perform the following steps: After the maintenance is completed, the equipment's operating parameters, environmental parameters and error logs are collected in real time, and the equipment's operating status is monitored within a preset time period; If the device does not report the same type of error or abnormality again within the preset time period, and the hardware actually repaired is inconsistent with the target faulty hardware predicted by the fault network model, then execute S1-S2; if the device still reports the same type of error or abnormality again within the preset time period, or the hardware actually repaired is consistent with the target faulty hardware predicted by the fault network model, then no operation is performed; S1. Based on the node positions of the actual repaired hardware and the target faulty hardware in the network, error correction is performed on the feature vector of the target faulty hardware node, including reducing the abnormality value in the feature vector corresponding to the target faulty hardware node and increasing the abnormality value in the feature vector corresponding to the actual repaired hardware node; S2. Directively adjust the edge weights of the adjacency matrix between the target faulty hardware and the actual maintenance hardware nodes based on the current prediction deviation, including reducing the edge weights with the target faulty hardware and increasing the edge weights with the actual maintenance hardware nodes.

Images