Router, gateway and camera operation and maintenance method and system

Abstract

The invention provides a router, gateway and camera operation and maintenance method and system in the technical field of network equipment management. The method comprises the following steps: S1, creating an equipment operation and maintenance strategy generation model; s2, acquiring a large amount of historical equipment monitoring data to construct a data set; s3, training the equipment operation and maintenance strategy generation model through the data set; s4, compressing the equipment operation and maintenance strategy generation model, deploying the compressed equipment operation and maintenance strategy generation model to network equipment of which the equipment type is a router, a gateway or a camera, and performing drift compensation training on the deployed equipment operation and maintenance strategy generation model; and S5, the network equipment obtains the equipment operation and maintenance strategy generation model deployed by inputting the real-time equipment monitoring data, a real-time operation and maintenance strategy is obtained, and operation and maintenance operation is automatically executed based on the real-time operation and maintenance strategy. The method has the advantages that the reliability, timeliness, convenience and stability of operation and maintenance of the network equipment are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of network device management technology, and in particular to a method and system for the operation and maintenance of routers, gateways and cameras. Background Technology

[0002] With the widespread adoption of smart home, security, and IoT technologies, network devices such as routers, gateways, and cameras have become core infrastructure supporting modern digital society. These network devices undertake critical tasks such as data transmission, security monitoring, and real-time communication, and serve emerging technology ecosystems such as 5G and edge computing. Their operational status directly determines the stability, security, and user experience of the overall network; therefore, ensuring reliable operation and maintenance of network devices has become a pressing technical challenge that needs to be addressed.

[0003] Current mainstream operation and maintenance (O&M) models primarily rely on manual intervention and passive response mechanisms, resulting in significant efficiency bottlenecks and timeliness deficiencies. Specifically: 1. Low inspection efficiency: O&M personnel need to periodically check hardware status on-site, but manual inspection intervals are typically several weeks or even months long, making it difficult to promptly detect sudden faults. 2. Delayed response mechanisms: For software-level faults, current solutions employ a reactive repair strategy (i.e., the handling process is only initiated after a user reports a fault), often resulting in a high-risk window of several hours between the occurrence and resolution of the fault. 3. Insufficient monitoring capabilities: Device operating parameters are mostly recorded locally through isolated tools (such as command-line interfaces (CLI) or log systems), lacking comprehensive real-time monitoring methods. 4. Data silo problem: Device health status information is scattered across various independent terminals or local systems, making it impossible to integrate and form a global view. This leads to difficulties in trend analysis and strategy optimization, a problem particularly prominent in scenarios with massive IoT devices and distributed deployments.

[0004] While advancements in IoT sensor technology and AI algorithms have made intelligent operations and maintenance (O&M) possible—for example, enabling real-time cloud access to device parameters via 5G networks and utilizing AI for data stream analysis—existing technologies still have significant shortcomings in practical applications: 1. Limited monitoring dimensions: Only basic indicators such as CPU temperature and bandwidth utilization are acquired, making it difficult to comprehensively assess device health status and thus affecting the reliability of O&M. 2. Lack of fault prediction: Relying primarily on post-incident retrospective analysis, there is a lack of proactive early warning mechanisms, affecting the timeliness of O&M. 3. Weak remote repair capabilities: Current remote maintenance is mostly limited to basic operations (such as device restart or patch installation); complex faults still require manual intervention, affecting the convenience of O&M. 4. Dispersed management architecture: The lack of centralized management and analysis methods for device health status makes it difficult to achieve global optimization, thus affecting the stability of O&M.

[0005] Therefore, how to provide a method and system for the operation and maintenance of routers, gateways and cameras to improve the reliability, timeliness, convenience and stability of network equipment operation and maintenance has become an urgent technical problem to be solved. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and system for the operation and maintenance of routers, gateways and cameras, so as to improve the reliability, timeliness, convenience and stability of network equipment operation and maintenance.

[0007] In a first aspect, the present invention provides a method for operating and maintaining a router, gateway, and camera, comprising the following steps:

[0008] Step S1: Create an equipment operation and maintenance strategy generation model for outputting operation and maintenance strategies carrying predicted fault types, predicted fault times, and fault handling suggestions, and set the loss function of the equipment operation and maintenance strategy generation model;

[0009] Step S2: Obtain a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data; preprocess and label the historical device monitoring data to construct a dataset.

[0010] Step S3: Divide the dataset into a training set, a validation set, and a test set based on stratified sampling. Train the equipment operation and maintenance strategy generation model using the training set and the loss function. Validate the trained equipment operation and maintenance strategy generation model using the validation set. Test the validated equipment operation and maintenance strategy generation model using the test set.

[0011] Step S4: After compressing the device operation and maintenance strategy generation model that has passed the test, deploy it to network devices of the device type, such as routers, gateways, or cameras, and perform drift compensation training on the deployed device operation and maintenance strategy generation model.

[0012] Step S5: The network device obtains real-time device monitoring data and inputs it into the deployed device operation and maintenance strategy generation model to obtain a real-time operation and maintenance strategy, and automatically executes operation and maintenance operations based on the real-time operation and maintenance strategy.

[0013] Step S6: The network device records operation and maintenance logs in real time, including at least real-time operation and maintenance policies, real-time device monitoring data, and policy execution feedback. Based on the federated learning mechanism, the deployed device operation and maintenance policy generation model is iteratively optimized through the operation and maintenance logs.

[0014] Furthermore, in step S1, the equipment operation and maintenance strategy generation model is constructed based on an input adaptation layer, a temporal feature extraction layer, and a multi-task output layer.

[0015] The input adaptation layer is constructed based on a data normalization unit and a type embedding unit. The data normalization unit is used to normalize the input router monitoring data, gateway monitoring data, or camera monitoring data through min-max normalization and Z-score normalization to obtain normalized data. The type embedding unit is used to encode the device type corresponding to the router monitoring data, gateway monitoring data, or camera monitoring data into a low-dimensional device feature vector through the embedding layer. The low-dimensional device feature vector is then concatenated with the corresponding normalized data to output a time series feature matrix.

[0016] The temporal feature extraction layer is constructed based on a bidirectional long short-term memory module and a multi-head self-attention module. The bidirectional long short-term memory module is used to extract long-range temporal dependency features and local event-related features from the time series feature matrix. The multi-head self-attention module is used to extract global augmented dependency features from the time series feature matrix, and the global augmented dependency features, long-range temporal dependency features, and local event-related features are concatenated and fused to obtain a unified contextual feature representation.

[0017] The multi-task output layer is constructed based on a fault type prediction module, a fault time prediction module, and a processing suggestion generation module. The fault type prediction module classifies faults based on a unified context feature representation and generates predicted fault types. The fault time prediction module maps the unified context feature representation to the predicted fault time. The processing suggestion generation module infers from the unified context feature representation, generates fault processing suggestions, and outputs an operation and maintenance strategy carrying the predicted fault type, predicted fault time, and fault processing suggestions.

[0018] Furthermore, step S2 specifically includes:

[0019] Acquire a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data; the router monitoring data includes at least CPU utilization, memory utilization, interface status, packet loss rate, throughput, and temperature; the gateway monitoring data includes at least the number of sessions, NAT table size, firewall rule hit rate, DNS resolution success rate, link status, and power status; the camera monitoring data includes at least video stream quality, connection status, storage status, temperature, and alarm status.

[0020] The monitoring data of each of the historical devices are preprocessed, including at least data cleaning, data normalization and data standardization. The preprocessed monitoring data of each of the historical devices is labeled, including at least the fault type, fault time and fault handling suggestions. A dataset is constructed based on the labeled monitoring data of each of the historical devices.

[0021] Furthermore, step S3 specifically includes:

[0022] Based on stratified sampling, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. The equipment operation and maintenance strategy generation model is trained using the training set and a loss function. During the training process, the hyperparameters of the equipment operation and maintenance strategy generation model are continuously optimized until the loss value of the loss function is less than a preset loss threshold or a preset early shutdown condition is met.

[0023] The accuracy of the strategy is calculated using the validation set to validate the trained equipment operation and maintenance strategy generation model. If the validation fails, the training set is expanded and training continues; if the validation passes, then:

[0024] The F1 score and confidence level are calculated using the test set to test the validated equipment operation and maintenance strategy generation model. If the test fails, the training set is expanded and training continues; if the test passes, training ends.

[0025] Furthermore, step S6 specifically includes:

[0026] Network devices record real-time operation and maintenance logs, including at least real-time operation and maintenance policies, real-time device monitoring data, and policy execution feedback. Based on a federated learning mechanism, an incremental dataset is constructed using these operation and maintenance logs to train and validate the deployed device operation and maintenance policy generation model. The model parameters of the validated device operation and maintenance policy generation model are uploaded to the server as local parameters. The server aggregates the local parameters uploaded by each network device to obtain global parameters. Based on these global parameters, the deployed device operation and maintenance policy generation models are iteratively optimized.

[0027] Secondly, the present invention provides a router, gateway, and camera operation and maintenance system, comprising the following modules:

[0028] The equipment operation and maintenance strategy generation model creation module is used to create an equipment operation and maintenance strategy generation model that outputs operation and maintenance strategies carrying predicted fault types, predicted fault times, and fault handling suggestions, and sets the loss function of the equipment operation and maintenance strategy generation model.

[0029] The dataset construction module is used to acquire a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data, and to construct the dataset after preprocessing and labeling the historical device monitoring data.

[0030] The equipment operation and maintenance strategy generation model training module is used to divide the dataset into a training set, a validation set, and a test set based on stratified sampling. The equipment operation and maintenance strategy generation model is trained using the training set and a loss function. The trained equipment operation and maintenance strategy generation model is validated using the validation set. The validated equipment operation and maintenance strategy generation model is tested using the test set.

[0031] The device operation and maintenance strategy generation model deployment module is used to compress the tested device operation and maintenance strategy generation model and deploy it to network devices of the device type, such as routers, gateways or cameras, and to perform drift compensation training on the deployed device operation and maintenance strategy generation model.

[0032] The device operation and maintenance module is used to obtain real-time device monitoring data from network devices and input it into the deployed device operation and maintenance strategy generation model to obtain real-time operation and maintenance strategies, and automatically execute operation and maintenance operations based on the real-time operation and maintenance strategies.

[0033] The model iteration and optimization module is used to record operation and maintenance logs in real time for network devices, including at least real-time operation and maintenance policies, real-time device monitoring data, and policy execution feedback. Based on the federated learning mechanism, the module generates a model for the deployed device operation and maintenance policies through the operation and maintenance logs and iteratively optimizes it.

[0034] Furthermore, in the device operation and maintenance strategy generation model creation module, the device operation and maintenance strategy generation model is constructed based on an input adaptation layer, a temporal feature extraction layer, and a multi-task output layer.

[0035] The input adaptation layer is constructed based on a data normalization unit and a type embedding unit. The data normalization unit is used to normalize the input router monitoring data, gateway monitoring data, or camera monitoring data through min-max normalization and Z-score normalization to obtain normalized data. The type embedding unit is used to encode the device type corresponding to the router monitoring data, gateway monitoring data, or camera monitoring data into a low-dimensional device feature vector through the embedding layer. The low-dimensional device feature vector is then concatenated with the corresponding normalized data to output a time series feature matrix.

[0036] The temporal feature extraction layer is constructed based on a bidirectional long short-term memory module and a multi-head self-attention module. The bidirectional long short-term memory module is used to extract long-range temporal dependency features and local event-related features from the time series feature matrix. The multi-head self-attention module is used to extract global augmented dependency features from the time series feature matrix, and the global augmented dependency features, long-range temporal dependency features, and local event-related features are concatenated and fused to obtain a unified contextual feature representation.

[0037] The multi-task output layer is constructed based on a fault type prediction module, a fault time prediction module, and a processing suggestion generation module. The fault type prediction module classifies faults based on a unified context feature representation and generates predicted fault types. The fault time prediction module maps the unified context feature representation to the predicted fault time. The processing suggestion generation module infers from the unified context feature representation, generates fault processing suggestions, and outputs an operation and maintenance strategy carrying the predicted fault type, predicted fault time, and fault processing suggestions.

[0038] Furthermore, the dataset construction module is specifically used for:

[0039] Acquire a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data; the router monitoring data includes at least CPU utilization, memory utilization, interface status, packet loss rate, throughput, and temperature; the gateway monitoring data includes at least the number of sessions, NAT table size, firewall rule hit rate, DNS resolution success rate, link status, and power status; the camera monitoring data includes at least video stream quality, connection status, storage status, temperature, and alarm status.

[0040] The monitoring data of each of the historical devices are preprocessed, including at least data cleaning, data normalization and data standardization. The preprocessed monitoring data of each of the historical devices is labeled, including at least the fault type, fault time and fault handling suggestions. A dataset is constructed based on the labeled monitoring data of each of the historical devices.

[0041] Furthermore, the equipment operation and maintenance strategy generation model training module is specifically used for:

[0042] Based on stratified sampling, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. The equipment operation and maintenance strategy generation model is trained using the training set and a loss function. During the training process, the hyperparameters of the equipment operation and maintenance strategy generation model are continuously optimized until the loss value of the loss function is less than a preset loss threshold or a preset early shutdown condition is met.

[0043] The accuracy of the strategy is calculated using the validation set to validate the trained equipment operation and maintenance strategy generation model. If the validation fails, the training set is expanded and training continues; if the validation passes, then:

[0044] The F1 score and confidence level are calculated using the test set to test the validated equipment operation and maintenance strategy generation model. If the test fails, the training set is expanded and training continues; if the test passes, training ends.

[0045] Furthermore, the model iterative optimization module is specifically used for:

[0046] Network devices record real-time operation and maintenance logs, including at least real-time operation and maintenance policies, real-time device monitoring data, and policy execution feedback. Based on a federated learning mechanism, an incremental dataset is constructed using these operation and maintenance logs to train and validate the deployed device operation and maintenance policy generation model. The model parameters of the validated device operation and maintenance policy generation model are uploaded to the server as local parameters. The server aggregates the local parameters uploaded by each network device to obtain global parameters. Based on these global parameters, the deployed device operation and maintenance policy generation models are iteratively optimized.

[0047] The advantages of the present invention are:

[0048] 1. A device operation and maintenance strategy generation model is created to output operation and maintenance strategies carrying predicted fault types, predicted fault times, and fault handling suggestions. A loss function for the device operation and maintenance strategy generation model is defined. Then, a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data, is acquired to construct a dataset. Based on stratified sampling, the dataset is divided into a training set, a validation set, and a test set. The device operation and maintenance strategy generation model is trained using the training set and the loss function. The trained model is validated using the validation set. The validated model is tested using the test set. The tested model is compressed and deployed to network devices of router, gateway, or camera types. Drift compensation training is then performed on the deployed model. The network devices receive real-time device monitoring data as input. The deployed device operation and maintenance (O&M) strategy generation model generates real-time O&M strategies. Based on these strategies, O&M operations are automatically executed, and O&M logs, including at least the real-time O&M strategies, real-time device monitoring data, and strategy execution feedback, are recorded in real time. Based on a federated learning mechanism, the deployed O&M strategy generation model is iteratively optimized using these logs. In other words, the network device O&M is automatically maintained based on the pre-trained and deployed O&M strategy generation model. The network device inputs multi-dimensional router monitoring data, gateway monitoring data, or camera monitoring data into the O&M strategy generation model to predict fault types and fault times, provide fault handling suggestions, and repair faults before they actually occur. This overcomes the shortcomings of traditional post-event retrospective analysis, eliminating the need for manual intervention. Combined with iterative optimization using the federated learning mechanism, global optimization is achieved, ultimately greatly improving the reliability, timeliness, convenience, and stability of network device O&M.

[0049] 2. By setting up the equipment operation and maintenance strategy generation model, a multi-task output layer (fault type prediction module, fault time prediction module, and handling suggestion generation module) is included. Through the temporal feature extraction layer (such as bidirectional long short-term memory module and multi-head self-attention module), long-term temporal dependency features, local event correlation features, and global enhanced dependency features are integrated. This enables simultaneous prediction of fault type, fault time, and handling suggestions, covering the entire lifecycle of fault diagnosis. It reduces the error accumulation of single-task models, and the multi-task output layer ensures the collaborative optimization of prediction results, thereby improving the accuracy of fault identification.

[0050] 3. By setting up data preprocessing including data cleaning, normalization (min-max normalization and Z-score standardization), and type embedding (encoding device types into low-dimensional feature vectors), and combining it with a temporal feature extraction layer to process heterogeneous data (such as router CPU utilization, gateway firewall rule hit rate, camera video stream quality, etc.); at the same time, stratified sampling (dividing the dataset in an 8:1:1 ratio) is used to construct training, validation, and test sets; normalization and embedding techniques solve the heterogeneity problem of multi-device data (routers, gateways, cameras), simplifying feature engineering; the temporal feature extraction layer efficiently integrates global and local features, enhancing the model's generalization ability to complex device behaviors and avoiding overfitting; stratified sampling ensures the balance of data distribution, improves model training efficiency, adapts to different device types, and reduces implementation costs.

[0051] 4. By compressing the tested model and deploying it to routers, gateways, or cameras, and combining it with drift compensation training to adapt to environmental changes, real-time monitoring data is acquired on network devices and maintenance operations are performed. Model compression technology (such as quantization into lightweight models) reduces the consumption of computing resources, enabling maintenance functions to run locally on edge devices without relying on cloud servers, reducing communication latency and bandwidth consumption, achieving an efficient edge computing architecture, improving response speed, and enhancing system scalability.

[0052] 5. By using a federated learning mechanism (building an incremental dataset based on operation and maintenance logs and aggregating local parameters to obtain global parameters) and drift compensation training, the model is iteratively optimized to adapt to equipment aging or environmental drift (such as temperature changes or equipment wear), ensuring the model's performance remains stable over time. The federated learning mechanism only uploads model parameters rather than raw data (such as real-time equipment monitoring data), protecting data privacy while reducing data transmission overhead. Through log feedback, continuous optimization is achieved, maintaining a high F1 score for the model and improving the adaptability and robustness of the operation and maintenance system.

[0053] 6. Automatically execute operation and maintenance operations based on real-time operation and maintenance policies, record operation and maintenance logs in real time, and provide policy execution feedback; full-process automation (from data input to operation execution) replaces manual intervention, significantly reducing operation and maintenance manpower and time costs. Through real-time policy execution (such as restarting equipment or adjusting configurations), predictive maintenance reduces unexpected downtime and improves network equipment availability and business operation continuity.

[0054] 7. By constructing a multi-task device operation and maintenance strategy generation model and combining it with a time-series feature extraction layer to accurately predict fault types, times, and handling suggestions, the accuracy and response efficiency of operation and maintenance are significantly improved. Data preprocessing and type embedding technologies are used to uniformly process data from heterogeneous devices such as routers, gateways, and cameras, enhancing the model's generalization ability. During deployment, model compression and drift compensation are used to achieve lightweight operation on edge devices, significantly reducing latency and resource consumption. At the same time, relying on a federated learning mechanism, incremental optimization of real-time operation and maintenance logs is performed to protect privacy, achieving dynamic model iteration and long-term stability. Ultimately, this achieves fully automated operation and maintenance, effectively reducing labor costs, shortening fault recovery time, and ensuring high reliability in complex network environments. Attached Figure Description

[0055] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0056] Figure 1 This is a flowchart of a router, gateway, and camera maintenance method according to the present invention.

[0057] Figure 2 This is a schematic diagram of the operation and maintenance system for a router, gateway, and camera according to the present invention. Detailed Implementation

[0058] The overall idea of ​​the technical solution in this application embodiment is as follows: Automatic operation and maintenance of network devices is performed based on a pre-trained and deployed device operation and maintenance strategy generation model. The network devices input multi-dimensional router monitoring data, gateway monitoring data, or camera monitoring data into the device operation and maintenance strategy generation model to predict fault types and fault times, provide fault handling suggestions, and repair faults before they actually occur. This overcomes the shortcomings of traditional post-event backtracking, eliminates the need for manual intervention, and, combined with iterative optimization through a federated learning mechanism, achieves global optimization, thereby improving the reliability, timeliness, convenience, and stability of network device operation and maintenance.

[0059] Please refer to Figures 1 to 2 As shown, a preferred embodiment of the router, gateway, and camera operation and maintenance method of the present invention includes the following steps:

[0060] Step S1: Create an equipment operation and maintenance strategy generation model for outputting operation and maintenance strategies carrying predicted fault types, predicted fault times, and fault handling suggestions, and set the loss function of the equipment operation and maintenance strategy generation model;

[0061] Step S2: Obtain a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data; preprocess and label the historical device monitoring data to construct a dataset.

[0062] Step S3: Divide the dataset into a training set, a validation set, and a test set based on stratified sampling. Train the equipment operation and maintenance strategy generation model using the training set and the loss function. Validate the trained equipment operation and maintenance strategy generation model using the validation set. Test the validated equipment operation and maintenance strategy generation model using the test set.

[0063] Step S4: After compressing the device operation and maintenance strategy generation model that has passed the test, deploy it to network devices of the device type, such as routers, gateways, or cameras, and perform drift compensation training on the deployed device operation and maintenance strategy generation model.

[0064] By compressing the tested models and deploying them to routers, gateways, or cameras, and combining drift compensation training to adapt to environmental changes, real-time monitoring data is acquired on network devices and maintenance operations are performed. Model compression technology (such as quantization into lightweight models) reduces the consumption of computing resources, enabling maintenance functions to run locally on edge devices without relying on cloud servers, reducing communication latency and bandwidth consumption, achieving an efficient edge computing architecture, improving response speed, and enhancing system scalability.

[0065] Step S5: The network device obtains real-time device monitoring data and inputs it into the deployed device operation and maintenance strategy generation model to obtain a real-time operation and maintenance strategy, and automatically executes operation and maintenance operations based on the real-time operation and maintenance strategy.

[0066] Based on real-time operation and maintenance policies, operation and maintenance operations are automatically executed, operation and maintenance logs are recorded in real time, and policy execution feedback is provided. The entire process is automated (from data input to operation execution), replacing manual intervention, which significantly reduces operation and maintenance manpower and time costs. Through real-time policy execution (such as restarting equipment or adjusting configurations), predictive maintenance reduces unexpected downtime and improves network equipment availability and business operation continuity.

[0067] Step S6: The network device records operation and maintenance logs in real time, including at least real-time operation and maintenance policies, real-time device monitoring data, and policy execution feedback. Based on the federated learning mechanism, the deployed device operation and maintenance policy generation model is iteratively optimized through the operation and maintenance logs.

[0068] By constructing a multi-task device operation and maintenance strategy generation model and combining it with a time-series feature extraction layer to accurately predict fault types, times, and handling suggestions, the accuracy and response efficiency of operation and maintenance are significantly improved. Data preprocessing and type embedding technologies are used to uniformly process data from heterogeneous devices such as routers, gateways, and cameras, enhancing the model's generalization ability. During deployment, model compression and drift compensation are used to achieve lightweight operation on edge devices, significantly reducing latency and resource consumption. At the same time, relying on a federated learning mechanism, incremental optimization of real-time operation and maintenance logs is performed to protect privacy, achieving dynamic model iteration and long-term stability. Ultimately, this achieves fully automated operation and maintenance, effectively reducing labor costs, shortening fault recovery time, and ensuring high reliability in complex network environments.

[0069] In step S1, the equipment operation and maintenance strategy generation model is constructed based on an input adaptation layer, a temporal feature extraction layer, and a multi-task output layer.

[0070] The input adaptation layer unifies router monitoring data, gateway monitoring data, or camera monitoring data into a standardized feature sequence. Specifically, it receives monitoring data from different devices (routers include CPU utilization, memory utilization, interface status, packet loss rate, throughput, and temperature; gateways include session count, NAT table size, firewall rule hit rate, DNS resolution success rate, link status, and power status; cameras include video stream quality, connection status, storage status, temperature, and alarm status), normalizes and embeds it to form a unified-dimensional feature vector sequence, which serves as the input to the temporal feature extraction layer.

[0071] The temporal feature extraction layer extracts time-dependent features from the feature sequence output by the input adaptation layer, capturing long-term trends, abnormal patterns, and device state correlations. It outputs a unified contextual feature representation, which is then input into the multi-task output layer. This layer optimizes temporal dependency modeling and enhances the attention weights for key events (such as temperature fluctuations or connection interruptions) to improve the accuracy of fault prediction.

[0072] The multi-task output layer generates three components for operational strategies: predicting fault types using contextual feature representations (as classification output), predicting fault times (as regression output), and generating fault handling suggestions (as sequence generation output). The output layer's design incorporates multi-task collaborative optimization for real-world application scenarios, avoiding information redundancy.

[0073] The input adaptation layer serves as the starting point of the model, and its output is directly connected to the input of the temporal feature extraction layer. The temporal feature extraction layer outputs a shared contextual feature representation, which is fed in parallel into three sub-modules of the multi-task output layer (fault type prediction module, fault time prediction module, and processing suggestion generation module), ensuring feature sharing and reducing inference overhead. The overall model adopts an end-to-end serial-parallel structure: Input adaptation layer → Temporal feature extraction layer → (parallel) multi-task output layer (fault type prediction module, fault time prediction module, and processing suggestion generation module).

[0074] The input adaptation layer is constructed based on a data normalization unit and a type embedding unit. The data normalization unit is used to normalize the input router monitoring data, gateway monitoring data, or camera monitoring data through min-max normalization and Z-score normalization to obtain normalized data. The type embedding unit is used to encode the device type corresponding to the router monitoring data, gateway monitoring data, or camera monitoring data into a low-dimensional device feature vector through the embedding layer. The low-dimensional device feature vector is then concatenated with the corresponding normalized data to output a time series feature matrix (time steps × feature dimensions).

[0075] The data normalization unit converts monitoring data from different devices (with varying numerical ranges) into a uniform scale (-1 to 1 or mean 0 and variance 1), eliminating dimensional differences between devices and improving the model's ability to generalize to input data.

[0076] The type embedding unit uses a trainable embedding layer (with the number of device types as the input dimension and a fixed embedding vector as the output dimension) to encode device types (router, gateway, or camera) into low-dimensional feature vectors and concatenate them onto a normalized feature sequence. This is used to capture device category information and enhance the model's ability to distinguish between different devices.

[0077] The temporal feature extraction layer is constructed based on a bidirectional long short-term memory module and a multi-head self-attention module. The bidirectional long short-term memory module is used to extract long-range temporal dependency features and local event-related features from the time series feature matrix. The multi-head self-attention module is used to extract global augmented dependency features from the time series feature matrix, and the global augmented dependency features, long-range temporal dependency features, and local event-related features are concatenated and fused to obtain a unified contextual feature representation.

[0078] The bidirectional long short-term memory module consists of two stacked bidirectional LSTM layers (the number of hidden units is adaptive according to the data dimension). It processes feature sequences from the forward and backward directions to extract long-term temporal dependent features (such as trend changes in CPU utilization or temperature anomalies) and capture local event correlations (such as the chain reaction between packet loss rate and interface status).

[0079] The multi-head self-attention module uses a Transformer encoder structure (multi-head attention layer, with 8 heads and output dimension matching the input) to calculate the attention weights at each time step within the sequence. This is used to enhance globally dependent features (e.g., highlighting key events such as high packet loss rates or alarm states) and to compensate for the long-term information forgetting problem of BiLSTM. The attention mechanism is specifically optimized for sparse fault events.

[0080] The time-series feature extraction layer optimizes the time sliding window partitioning (e.g., the window size is set to 24 hours of history) to adapt to periodic monitoring data in real operation and maintenance scenarios.

[0081] The multi-task output layer is constructed based on a fault type prediction module, a fault time prediction module, and a processing suggestion generation module. The fault type prediction module classifies faults based on a unified context feature representation and generates predicted fault types. The fault time prediction module maps the unified context feature representation to the predicted fault time. The processing suggestion generation module infers from the unified context feature representation, generates fault processing suggestions, and outputs an operation and maintenance strategy carrying the predicted fault type, predicted fault time, and fault processing suggestions.

[0082] The fault type prediction module, fault time prediction module, and processing suggestion generation module share the unified context feature representation of the temporal feature extraction layer;

[0083] The fault type prediction module consists of a fully connected layer (ReLU activation) and a softmax output layer (the output dimension is the number of fault type categories, such as hardware faults, software errors, network interruptions, etc.), which is used to classify based on contextual features and predict potential fault types.

[0084] The failure time prediction module consists of a fully connected layer (ReLU activation) and a linear regression output layer (outputting a single value representing the time offset of the predicted failure occurrence, such as within the next hour), used to map contextual features to point-in-time predictions and support regression modeling.

[0085] The suggestion generation module consists of an LSTM decoder and an attention mechanism layer (based on contextual features). The decoder outputs a variable-length text sequence (such as the suggestion "restart the device" or "check the temperature sensor") to generate natural language processing suggestions. The attention mechanism ensures that the suggestions are relevant to the fault type and time.

[0086] By setting up a device operation and maintenance strategy generation model that includes a multi-task output layer (fault type prediction module, fault time prediction module, and handling suggestion generation module), and by using a temporal feature extraction layer (such as a bidirectional long short-term memory module and a multi-head self-attention module) to fuse long-term temporal dependency features, local event-related features, and global enhanced dependency features, it can simultaneously predict fault type, fault time, and handling suggestions, covering the entire lifecycle of fault diagnosis. This reduces the error accumulation of a single-task model, and the multi-task output layer ensures the collaborative optimization of prediction results, thereby improving the accuracy of fault identification.

[0087] The equipment operation and maintenance strategy generation model employs the Adam optimizer (learning rate set to 0.001, beta1 = 0.9, beta2 = 0.999), combined with a learning rate scheduling strategy (such as ReduceLROnPlateau dynamically adjusted based on validation set performance) to accelerate convergence and prevent getting trapped in local optima. Optimizations for practical applications include: adding gradient pruning (threshold set to 1.0) to avoid gradient explosion, and using class-weighted sampling in batch training (oversampling fault samples) to address the imbalance problem of equipment fault data.

[0088] The formula for the loss function is:

[0089] L_total=α*L_type+β*L_time+γ*L_advice;

[0090] Wherein, L_total represents the loss value of the loss function; L_type is the fault type classification loss: using categorical cross-entropy loss with a weight α set to 0.4, emphasizing the learning of high-frequency events; L_time is the fault time regression loss: using mean squared error loss with a weight β set to 0.3, adapting to the continuity of time-shifted prediction; L_advice is the processing suggestion generation loss: using sequence cross-entropy loss with a weight γ set to 0.3, and adjusting the weights based on BLEU scores to ensure the semantic accuracy of the text.

[0091] Step S2 specifically involves:

[0092] Acquire a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data; the router monitoring data includes at least CPU utilization, memory utilization, interface status, packet loss rate, throughput, and temperature; the gateway monitoring data includes at least the number of sessions, NAT table size, firewall rule hit rate, DNS resolution success rate, link status, and power status; the camera monitoring data includes at least video stream quality, connection status, storage status, temperature, and alarm status.

[0093] The monitoring data of each of the historical devices are preprocessed, including at least data cleaning, data normalization and data standardization. The preprocessed monitoring data of each of the historical devices is labeled, including at least the fault type, fault time and fault handling suggestions. A dataset is constructed based on the labeled monitoring data of each of the historical devices.

[0094] By setting up data preprocessing including data cleaning, normalization (min-max normalization and Z-score standardization), and type embedding (encoding device types into low-dimensional feature vectors), and combining it with a temporal feature extraction layer to process heterogeneous data (such as router CPU utilization, gateway firewall rule hit rate, camera video stream quality, etc.), and using stratified sampling (dividing the dataset in an 8:1:1 ratio) to construct training, validation, and test sets, the normalization and embedding techniques solve the heterogeneity problem of multi-device data (routers, gateways, cameras) and simplify feature engineering. The temporal feature extraction layer efficiently integrates global and local features, enhancing the model's generalization ability to complex device behaviors and avoiding overfitting. Stratified sampling ensures the balance of data distribution, improves model training efficiency, adapts to different device types, and reduces implementation costs.

[0095] Step S3 specifically involves:

[0096] Based on stratified sampling, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. The equipment operation and maintenance strategy generation model is trained using the training set and a loss function. During the training process, the hyperparameters of the equipment operation and maintenance strategy generation model are continuously optimized until the loss value of the loss function is less than a preset loss threshold or a preset early shutdown condition is met.

[0097] The accuracy of the strategy is calculated using the validation set to validate the trained equipment operation and maintenance strategy generation model. If the validation fails, the training set is expanded and training continues; if the validation passes, then:

[0098] The F1 score and confidence level are calculated using the test set to test the validated equipment operation and maintenance strategy generation model. If the test fails, the training set is expanded and training continues; if the test passes, training ends.

[0099] Step S6 specifically involves:

[0100] Network devices record real-time operation and maintenance logs, including at least real-time operation and maintenance policies, real-time device monitoring data, and policy execution feedback. Based on a federated learning mechanism, an incremental dataset is constructed using these operation and maintenance logs to train and validate the deployed device operation and maintenance policy generation model. The model parameters of the validated device operation and maintenance policy generation model are uploaded to the server as local parameters. The server aggregates the local parameters uploaded by each network device to obtain global parameters. Based on these global parameters, the deployed device operation and maintenance policy generation models are iteratively optimized.

[0101] By employing a federated learning mechanism (building an incremental dataset based on operation and maintenance logs and aggregating local parameters to obtain global parameters) and drift compensation training, the model is iteratively optimized to adapt to equipment aging or environmental drift (such as temperature changes or equipment wear), ensuring the model's performance remains stable over time. The federated learning mechanism only uploads model parameters rather than raw data (such as real-time equipment monitoring data), protecting data privacy while reducing data transmission overhead. Through continuous optimization via log feedback, the model maintains a high F1 score, improving the adaptability and robustness of the operation and maintenance system.

[0102] A preferred embodiment of the router, gateway, and camera maintenance system of the present invention includes the following modules:

[0103] The equipment operation and maintenance strategy generation model creation module is used to create an equipment operation and maintenance strategy generation model that outputs operation and maintenance strategies carrying predicted fault types, predicted fault times, and fault handling suggestions, and sets the loss function of the equipment operation and maintenance strategy generation model.

[0104] The dataset construction module is used to acquire a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data, and to construct the dataset after preprocessing and labeling the historical device monitoring data.

[0105] The equipment operation and maintenance strategy generation model training module is used to divide the dataset into a training set, a validation set, and a test set based on stratified sampling. The equipment operation and maintenance strategy generation model is trained using the training set and a loss function. The trained equipment operation and maintenance strategy generation model is validated using the validation set. The validated equipment operation and maintenance strategy generation model is tested using the test set.

[0106] The device operation and maintenance strategy generation model deployment module is used to compress the tested device operation and maintenance strategy generation model and deploy it to network devices of the device type, such as routers, gateways or cameras, and to perform drift compensation training on the deployed device operation and maintenance strategy generation model.

[0107] By compressing the tested models and deploying them to routers, gateways, or cameras, and combining drift compensation training to adapt to environmental changes, real-time monitoring data is acquired on network devices and maintenance operations are performed. Model compression technology (such as quantization into lightweight models) reduces the consumption of computing resources, enabling maintenance functions to run locally on edge devices without relying on cloud servers, reducing communication latency and bandwidth consumption, achieving an efficient edge computing architecture, improving response speed, and enhancing system scalability.

[0108] The device operation and maintenance module is used to obtain real-time device monitoring data from network devices and input it into the deployed device operation and maintenance strategy generation model to obtain real-time operation and maintenance strategies, and automatically execute operation and maintenance operations based on the real-time operation and maintenance strategies.

[0109] Based on real-time operation and maintenance policies, operation and maintenance operations are automatically executed, operation and maintenance logs are recorded in real time, and policy execution feedback is provided. The entire process is automated (from data input to operation execution), replacing manual intervention, which significantly reduces operation and maintenance manpower and time costs. Through real-time policy execution (such as restarting equipment or adjusting configurations), predictive maintenance reduces unexpected downtime and improves network equipment availability and business operation continuity.

[0110] The model iteration and optimization module is used to record operation and maintenance logs in real time for network devices, including at least real-time operation and maintenance policies, real-time device monitoring data, and policy execution feedback. Based on the federated learning mechanism, the module generates a model for the deployed device operation and maintenance policies through the operation and maintenance logs and iteratively optimizes it.

[0111] By constructing a multi-task device operation and maintenance strategy generation model and combining it with a time-series feature extraction layer to accurately predict fault types, times, and handling suggestions, the accuracy and response efficiency of operation and maintenance are significantly improved. Data preprocessing and type embedding technologies are used to uniformly process data from heterogeneous devices such as routers, gateways, and cameras, enhancing the model's generalization ability. During deployment, model compression and drift compensation are used to achieve lightweight operation on edge devices, significantly reducing latency and resource consumption. At the same time, relying on a federated learning mechanism, incremental optimization of real-time operation and maintenance logs is performed to protect privacy, achieving dynamic model iteration and long-term stability. Ultimately, this achieves fully automated operation and maintenance, effectively reducing labor costs, shortening fault recovery time, and ensuring high reliability in complex network environments.

[0112] In the device operation and maintenance strategy generation model creation module, the device operation and maintenance strategy generation model is constructed based on an input adaptation layer, a temporal feature extraction layer, and a multi-task output layer.

[0113] The input adaptation layer unifies router monitoring data, gateway monitoring data, or camera monitoring data into a standardized feature sequence. Specifically, it receives monitoring data from different devices (routers include CPU utilization, memory utilization, interface status, packet loss rate, throughput, and temperature; gateways include session count, NAT table size, firewall rule hit rate, DNS resolution success rate, link status, and power status; cameras include video stream quality, connection status, storage status, temperature, and alarm status), normalizes and embeds it to form a unified-dimensional feature vector sequence, which serves as the input to the temporal feature extraction layer.

[0114] The temporal feature extraction layer extracts time-dependent features from the feature sequence output by the input adaptation layer, capturing long-term trends, abnormal patterns, and device state correlations. It outputs a unified contextual feature representation, which is then input into the multi-task output layer. This layer optimizes temporal dependency modeling and enhances the attention weights for key events (such as temperature fluctuations or connection interruptions) to improve the accuracy of fault prediction.

[0115] The multi-task output layer generates three components for operational strategies: predicting fault types using contextual feature representations (as classification output), predicting fault times (as regression output), and generating fault handling suggestions (as sequence generation output). The output layer's design incorporates multi-task collaborative optimization for real-world application scenarios, avoiding information redundancy.

[0116] The input adaptation layer serves as the starting point of the model, and its output is directly connected to the input of the temporal feature extraction layer. The temporal feature extraction layer outputs a shared contextual feature representation, which is fed in parallel into three sub-modules of the multi-task output layer (fault type prediction module, fault time prediction module, and processing suggestion generation module), ensuring feature sharing and reducing inference overhead. The overall model adopts an end-to-end serial-parallel structure: Input adaptation layer → Temporal feature extraction layer → (parallel) multi-task output layer (fault type prediction module, fault time prediction module, and processing suggestion generation module).

[0117] The input adaptation layer is constructed based on a data normalization unit and a type embedding unit. The data normalization unit is used to normalize the input router monitoring data, gateway monitoring data, or camera monitoring data through min-max normalization and Z-score normalization to obtain normalized data. The type embedding unit is used to encode the device type corresponding to the router monitoring data, gateway monitoring data, or camera monitoring data into a low-dimensional device feature vector through the embedding layer. The low-dimensional device feature vector is then concatenated with the corresponding normalized data to output a time series feature matrix (time steps × feature dimensions).

[0118] The data normalization unit converts monitoring data from different devices (with varying numerical ranges) into a uniform scale (-1 to 1 or mean 0 and variance 1), eliminating dimensional differences between devices and improving the model's ability to generalize to input data.

[0119] The type embedding unit uses a trainable embedding layer (with the number of device types as the input dimension and a fixed embedding vector as the output dimension) to encode device types (router, gateway, or camera) into low-dimensional feature vectors and concatenate them onto a normalized feature sequence. This is used to capture device category information and enhance the model's ability to distinguish between different devices.

[0120] The temporal feature extraction layer is constructed based on a bidirectional long short-term memory module and a multi-head self-attention module. The bidirectional long short-term memory module is used to extract long-range temporal dependency features and local event-related features from the time series feature matrix. The multi-head self-attention module is used to extract global augmented dependency features from the time series feature matrix, and the global augmented dependency features, long-range temporal dependency features, and local event-related features are concatenated and fused to obtain a unified contextual feature representation.

[0121] The bidirectional long short-term memory module consists of two stacked bidirectional LSTM layers (the number of hidden units is adaptive according to the data dimension). It processes feature sequences from the forward and backward directions to extract long-term temporal dependent features (such as trend changes in CPU utilization or temperature anomalies) and capture local event correlations (such as the chain reaction between packet loss rate and interface status).

[0122] The multi-head self-attention module uses a Transformer encoder structure (multi-head attention layer, with 8 heads and output dimension matching the input) to calculate the attention weights at each time step within the sequence. This is used to enhance globally dependent features (e.g., highlighting key events such as high packet loss rates or alarm states) and to compensate for the long-term information forgetting problem of BiLSTM. The attention mechanism is specifically optimized for sparse fault events.

[0123] The time-series feature extraction layer optimizes the time sliding window partitioning (e.g., the window size is set to 24 hours of history) to adapt to periodic monitoring data in real operation and maintenance scenarios.

[0124] The multi-task output layer is constructed based on a fault type prediction module, a fault time prediction module, and a processing suggestion generation module. The fault type prediction module classifies faults based on a unified context feature representation and generates predicted fault types. The fault time prediction module maps the unified context feature representation to the predicted fault time. The processing suggestion generation module infers from the unified context feature representation, generates fault processing suggestions, and outputs an operation and maintenance strategy carrying the predicted fault type, predicted fault time, and fault processing suggestions.

[0125] The fault type prediction module, fault time prediction module, and processing suggestion generation module share the unified context feature representation of the temporal feature extraction layer;

[0126] The fault type prediction module consists of a fully connected layer (ReLU activation) and a softmax output layer (the output dimension is the number of fault type categories, such as hardware faults, software errors, network interruptions, etc.), which is used to classify based on contextual features and predict potential fault types.

[0127] The failure time prediction module consists of a fully connected layer (ReLU activation) and a linear regression output layer (outputting a single value representing the time offset of the predicted failure occurrence, such as within the next hour), used to map contextual features to point-in-time predictions and support regression modeling.

[0128] The suggestion generation module consists of an LSTM decoder and an attention mechanism layer (based on contextual features). The decoder outputs a variable-length text sequence (such as the suggestion "restart the device" or "check the temperature sensor") to generate natural language processing suggestions. The attention mechanism ensures that the suggestions are relevant to the fault type and time.

[0129] By setting up a device operation and maintenance strategy generation model that includes a multi-task output layer (fault type prediction module, fault time prediction module, and handling suggestion generation module), and by using a temporal feature extraction layer (such as a bidirectional long short-term memory module and a multi-head self-attention module) to fuse long-term temporal dependency features, local event-related features, and global enhanced dependency features, it can simultaneously predict fault type, fault time, and handling suggestions, covering the entire lifecycle of fault diagnosis. This reduces the error accumulation of a single-task model, and the multi-task output layer ensures the collaborative optimization of prediction results, thereby improving the accuracy of fault identification.

[0130] The equipment operation and maintenance strategy generation model employs the Adam optimizer (learning rate set to 0.001, beta1 = 0.9, beta2 = 0.999), combined with a learning rate scheduling strategy (such as ReduceLROnPlateau dynamically adjusted based on validation set performance) to accelerate convergence and prevent getting trapped in local optima. Optimizations for practical applications include: adding gradient pruning (threshold set to 1.0) to avoid gradient explosion, and using class-weighted sampling in batch training (oversampling fault samples) to address the imbalance problem of equipment fault data.

[0131] The formula for the loss function is:

[0132] L_total=α*L_type+β*L_time+γ*L_advice;

[0133] Wherein, L_total represents the loss value of the loss function; L_type is the fault type classification loss: using categorical cross-entropy loss with a weight α set to 0.4, emphasizing the learning of high-frequency events; L_time is the fault time regression loss: using mean squared error loss with a weight β set to 0.3, adapting to the continuity of time-shifted prediction; L_advice is the processing suggestion generation loss: using sequence cross-entropy loss with a weight γ set to 0.3, and adjusting the weights based on BLEU scores to ensure the semantic accuracy of the text.

[0134] The dataset construction module is specifically used for:

[0135] Acquire a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data; the router monitoring data includes at least CPU utilization, memory utilization, interface status, packet loss rate, throughput, and temperature; the gateway monitoring data includes at least the number of sessions, NAT table size, firewall rule hit rate, DNS resolution success rate, link status, and power status; the camera monitoring data includes at least video stream quality, connection status, storage status, temperature, and alarm status.

[0136] The monitoring data of each of the historical devices are preprocessed, including at least data cleaning, data normalization and data standardization. The preprocessed monitoring data of each of the historical devices is labeled, including at least the fault type, fault time and fault handling suggestions. A dataset is constructed based on the labeled monitoring data of each of the historical devices.

[0137] By setting up data preprocessing including data cleaning, normalization (min-max normalization and Z-score standardization), and type embedding (encoding device types into low-dimensional feature vectors), and combining it with a temporal feature extraction layer to process heterogeneous data (such as router CPU utilization, gateway firewall rule hit rate, camera video stream quality, etc.), and using stratified sampling (dividing the dataset in an 8:1:1 ratio) to construct training, validation, and test sets, the normalization and embedding techniques solve the heterogeneity problem of multi-device data (routers, gateways, cameras) and simplify feature engineering. The temporal feature extraction layer efficiently integrates global and local features, enhancing the model's generalization ability to complex device behaviors and avoiding overfitting. Stratified sampling ensures the balance of data distribution, improves model training efficiency, adapts to different device types, and reduces implementation costs.

[0138] The equipment operation and maintenance strategy generation model training module is specifically used for:

[0139] Based on stratified sampling, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. The equipment operation and maintenance strategy generation model is trained using the training set and a loss function. During the training process, the hyperparameters of the equipment operation and maintenance strategy generation model are continuously optimized until the loss value of the loss function is less than a preset loss threshold or a preset early shutdown condition is met.

[0140] The accuracy of the strategy is calculated using the validation set to validate the trained equipment operation and maintenance strategy generation model. If the validation fails, the training set is expanded and training continues; if the validation passes, then:

[0141] The F1 score and confidence level are calculated using the test set to test the validated equipment operation and maintenance strategy generation model. If the test fails, the training set is expanded and training continues; if the test passes, training ends.

[0142] The model iterative optimization module is specifically used for:

[0143] Network devices record real-time operation and maintenance logs, including at least real-time operation and maintenance policies, real-time device monitoring data, and policy execution feedback. Based on a federated learning mechanism, an incremental dataset is constructed using these operation and maintenance logs to train and validate the deployed device operation and maintenance policy generation model. The model parameters of the validated device operation and maintenance policy generation model are uploaded to the server as local parameters. The server aggregates the local parameters uploaded by each network device to obtain global parameters. Based on these global parameters, the deployed device operation and maintenance policy generation models are iteratively optimized.

[0144] By employing a federated learning mechanism (building an incremental dataset based on operation and maintenance logs and aggregating local parameters to obtain global parameters) and drift compensation training, the model is iteratively optimized to adapt to equipment aging or environmental drift (such as temperature changes or equipment wear), ensuring the model's performance remains stable over time. The federated learning mechanism only uploads model parameters rather than raw data (such as real-time equipment monitoring data), protecting data privacy while reducing data transmission overhead. Through continuous optimization via log feedback, the model maintains a high F1 score, improving the adaptability and robustness of the operation and maintenance system.

[0145] In summary, the advantages of this invention are:

[0146] 1. A device operation and maintenance strategy generation model is created to output operation and maintenance strategies carrying predicted fault types, predicted fault times, and fault handling suggestions. A loss function for the device operation and maintenance strategy generation model is defined. Then, a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data, is acquired to construct a dataset. Based on stratified sampling, the dataset is divided into a training set, a validation set, and a test set. The device operation and maintenance strategy generation model is trained using the training set and the loss function. The trained model is validated using the validation set. The validated model is tested using the test set. The tested model is compressed and deployed to network devices of router, gateway, or camera types. Drift compensation training is then performed on the deployed model. The network devices receive real-time device monitoring data as input. The deployed device operation and maintenance (O&M) strategy generation model generates real-time O&M strategies. Based on these strategies, O&M operations are automatically executed, and O&M logs, including at least the real-time O&M strategies, real-time device monitoring data, and strategy execution feedback, are recorded in real time. Based on a federated learning mechanism, the deployed O&M strategy generation model is iteratively optimized using these logs. In other words, the network device O&M is automatically maintained based on the pre-trained and deployed O&M strategy generation model. The network device inputs multi-dimensional router monitoring data, gateway monitoring data, or camera monitoring data into the O&M strategy generation model to predict fault types and fault times, provide fault handling suggestions, and repair faults before they actually occur. This overcomes the shortcomings of traditional post-event retrospective analysis, eliminating the need for manual intervention. Combined with iterative optimization using the federated learning mechanism, global optimization is achieved, ultimately greatly improving the reliability, timeliness, convenience, and stability of network device O&M.

[0147] 2. By setting up the equipment operation and maintenance strategy generation model, a multi-task output layer (fault type prediction module, fault time prediction module, and handling suggestion generation module) is included. Through the temporal feature extraction layer (such as bidirectional long short-term memory module and multi-head self-attention module), long-term temporal dependency features, local event correlation features, and global enhanced dependency features are integrated. This enables simultaneous prediction of fault type, fault time, and handling suggestions, covering the entire lifecycle of fault diagnosis. It reduces the error accumulation of single-task models, and the multi-task output layer ensures the collaborative optimization of prediction results, thereby improving the accuracy of fault identification.

[0148] 3. By setting up data preprocessing including data cleaning, normalization (min-max normalization and Z-score standardization), and type embedding (encoding device types into low-dimensional feature vectors), and combining it with a temporal feature extraction layer to process heterogeneous data (such as router CPU utilization, gateway firewall rule hit rate, camera video stream quality, etc.); at the same time, stratified sampling (dividing the dataset in an 8:1:1 ratio) is used to construct training, validation, and test sets; normalization and embedding techniques solve the heterogeneity problem of multi-device data (routers, gateways, cameras), simplifying feature engineering; the temporal feature extraction layer efficiently integrates global and local features, enhancing the model's generalization ability to complex device behaviors and avoiding overfitting; stratified sampling ensures the balance of data distribution, improves model training efficiency, adapts to different device types, and reduces implementation costs.

[0149] 4. By compressing the tested model and deploying it to routers, gateways, or cameras, and combining it with drift compensation training to adapt to environmental changes, real-time monitoring data is acquired on network devices and maintenance operations are performed. Model compression technology (such as quantization into lightweight models) reduces the consumption of computing resources, enabling maintenance functions to run locally on edge devices without relying on cloud servers, reducing communication latency and bandwidth consumption, achieving an efficient edge computing architecture, improving response speed, and enhancing system scalability.

[0150] 5. By using a federated learning mechanism (building an incremental dataset based on operation and maintenance logs and aggregating local parameters to obtain global parameters) and drift compensation training, the model is iteratively optimized to adapt to equipment aging or environmental drift (such as temperature changes or equipment wear), ensuring the model's performance remains stable over time. The federated learning mechanism only uploads model parameters rather than raw data (such as real-time equipment monitoring data), protecting data privacy while reducing data transmission overhead. Through log feedback, continuous optimization is achieved, maintaining a high F1 score for the model and improving the adaptability and robustness of the operation and maintenance system.

[0151] 6. Automatically execute operation and maintenance operations based on real-time operation and maintenance policies, record operation and maintenance logs in real time, and provide policy execution feedback; full-process automation (from data input to operation execution) replaces manual intervention, significantly reducing operation and maintenance manpower and time costs. Through real-time policy execution (such as restarting equipment or adjusting configurations), predictive maintenance reduces unexpected downtime and improves network equipment availability and business operation continuity.

[0152] 7. By constructing a multi-task device operation and maintenance strategy generation model and combining it with a time-series feature extraction layer to accurately predict fault types, times, and handling suggestions, the accuracy and response efficiency of operation and maintenance are significantly improved. Data preprocessing and type embedding technologies are used to uniformly process data from heterogeneous devices such as routers, gateways, and cameras, enhancing the model's generalization ability. During deployment, model compression and drift compensation are used to achieve lightweight operation on edge devices, significantly reducing latency and resource consumption. At the same time, relying on a federated learning mechanism, incremental optimization of real-time operation and maintenance logs is performed to protect privacy, achieving dynamic model iteration and long-term stability. Ultimately, this achieves fully automated operation and maintenance, effectively reducing labor costs, shortening fault recovery time, and ensuring high reliability in complex network environments.

[0153] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for maintaining and operating routers, gateways, and cameras, characterized in that: Includes the following steps: Step S1: Create an equipment operation and maintenance strategy generation model for outputting operation and maintenance strategies carrying predicted fault types, predicted fault times, and fault handling suggestions, and set the loss function of the equipment operation and maintenance strategy generation model; Step S2: Obtain a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data; preprocess and label the historical device monitoring data to construct a dataset. Step S3: Divide the dataset into a training set, a validation set, and a test set based on stratified sampling. Train the equipment operation and maintenance strategy generation model using the training set and the loss function. Validate the trained equipment operation and maintenance strategy generation model using the validation set. Test the validated equipment operation and maintenance strategy generation model using the test set. Step S4: After compressing the device operation and maintenance strategy generation model that has passed the test, deploy it to network devices of the device type, such as routers, gateways, or cameras, and perform drift compensation training on the deployed device operation and maintenance strategy generation model. Step S5: The network device obtains real-time device monitoring data and inputs it into the deployed device operation and maintenance strategy generation model to obtain a real-time operation and maintenance strategy, and automatically executes operation and maintenance operations based on the real-time operation and maintenance strategy. Step S6: The network device records operation and maintenance logs in real time, including at least real-time operation and maintenance policies, real-time device monitoring data, and policy execution feedback. Based on the federated learning mechanism, the deployed device operation and maintenance policy generation model is iteratively optimized through the operation and maintenance logs.

2. The method for maintaining a router, gateway, and camera as described in claim 1, characterized in that: In step S1, the equipment operation and maintenance strategy generation model is constructed based on an input adaptation layer, a temporal feature extraction layer, and a multi-task output layer. The input adaptation layer is constructed based on a data normalization unit and a type embedding unit. The data normalization unit is used to normalize the input router monitoring data, gateway monitoring data, or camera monitoring data through min-max normalization and Z-score normalization to obtain normalized data. The type embedding unit is used to encode the device type corresponding to the router monitoring data, gateway monitoring data, or camera monitoring data into a low-dimensional device feature vector through the embedding layer. The low-dimensional device feature vector is then concatenated with the corresponding normalized data to output a time series feature matrix. The temporal feature extraction layer is constructed based on a bidirectional long short-term memory module and a multi-head self-attention module. The bidirectional long short-term memory module is used to extract long-range temporal dependency features and local event-related features from the time series feature matrix. The multi-head self-attention module is used to extract global augmented dependency features from the time series feature matrix, and the global augmented dependency features, long-range temporal dependency features, and local event-related features are concatenated and fused to obtain a unified contextual feature representation. The multi-task output layer is constructed based on a fault type prediction module, a fault time prediction module, and a processing suggestion generation module. The fault type prediction module classifies faults based on a unified context feature representation and generates predicted fault types. The fault time prediction module maps the unified context feature representation to the predicted fault time. The processing suggestion generation module infers from the unified context feature representation, generates fault processing suggestions, and outputs an operation and maintenance strategy carrying the predicted fault type, predicted fault time, and fault processing suggestions.

3. The router, gateway, and camera maintenance method as described in claim 1, characterized in that: Step S2 specifically involves: Acquire a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data; the router monitoring data includes at least CPU utilization, memory utilization, interface status, packet loss rate, throughput, and temperature; the gateway monitoring data includes at least the number of sessions, NAT table size, firewall rule hit rate, DNS resolution success rate, link status, and power status; the camera monitoring data includes at least video stream quality, connection status, storage status, temperature, and alarm status. The monitoring data of each of the historical devices are preprocessed, including at least data cleaning, data normalization and data standardization. The preprocessed monitoring data of each of the historical devices is labeled, including at least the fault type, fault time and fault handling suggestions. A dataset is constructed based on the labeled monitoring data of each of the historical devices.

4. The router, gateway, and camera maintenance method as described in claim 1, characterized in that: Step S3 specifically involves: Based on stratified sampling, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:

1. The equipment operation and maintenance strategy generation model is trained using the training set and a loss function. During the training process, the hyperparameters of the equipment operation and maintenance strategy generation model are continuously optimized until the loss value of the loss function is less than a preset loss threshold or a preset early shutdown condition is met. The accuracy of the strategy is calculated using the validation set to validate the trained equipment operation and maintenance strategy generation model. If the validation fails, the training set is expanded and training continues; if the validation passes, then: The F1 score and confidence level are calculated using the test set to test the validated equipment operation and maintenance strategy generation model. If the test fails, the training set is expanded and training continues; if the test passes, training ends.

5. The method for maintaining a router, gateway, and camera as described in claim 1, characterized in that: Step S6 specifically involves: Network devices record real-time operation and maintenance logs, including at least real-time operation and maintenance policies, real-time device monitoring data, and policy execution feedback. Based on a federated learning mechanism, an incremental dataset is constructed using these operation and maintenance logs to train and validate the deployed device operation and maintenance policy generation model. The model parameters of the validated device operation and maintenance policy generation model are uploaded to the server as local parameters. The server aggregates the local parameters uploaded by each network device to obtain global parameters. Based on these global parameters, the deployed device operation and maintenance policy generation models are iteratively optimized.

6. A router, gateway, and camera maintenance system, characterized in that: Includes the following modules: The equipment operation and maintenance strategy generation model creation module is used to create an equipment operation and maintenance strategy generation model that outputs operation and maintenance strategies carrying predicted fault types, predicted fault times, and fault handling suggestions, and sets the loss function of the equipment operation and maintenance strategy generation model. The dataset construction module is used to acquire a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data, and to construct the dataset after preprocessing and labeling the historical device monitoring data. The equipment operation and maintenance strategy generation model training module is used to divide the dataset into a training set, a validation set, and a test set based on stratified sampling. The equipment operation and maintenance strategy generation model is trained using the training set and a loss function. The trained equipment operation and maintenance strategy generation model is validated using the validation set. The validated equipment operation and maintenance strategy generation model is tested using the test set. The device operation and maintenance strategy generation model deployment module is used to compress the tested device operation and maintenance strategy generation model and deploy it to network devices of the device type, such as routers, gateways or cameras, and to perform drift compensation training on the deployed device operation and maintenance strategy generation model. The device operation and maintenance module is used to obtain real-time device monitoring data from network devices and input it into the deployed device operation and maintenance strategy generation model to obtain real-time operation and maintenance strategies, and automatically execute operation and maintenance operations based on the real-time operation and maintenance strategies. The model iteration and optimization module is used to record operation and maintenance logs in real time for network devices, including at least real-time operation and maintenance policies, real-time device monitoring data, and policy execution feedback. Based on the federated learning mechanism, the module generates a model for the deployed device operation and maintenance policies through the operation and maintenance logs and iteratively optimizes it.

7. The router, gateway, and camera maintenance system as described in claim 6, characterized in that: In the device operation and maintenance strategy generation model creation module, the device operation and maintenance strategy generation model is constructed based on an input adaptation layer, a temporal feature extraction layer, and a multi-task output layer. The input adaptation layer is constructed based on a data normalization unit and a type embedding unit. The data normalization unit is used to normalize the input router monitoring data, gateway monitoring data, or camera monitoring data through min-max normalization and Z-score normalization to obtain normalized data. The type embedding unit is used to encode the device type corresponding to the router monitoring data, gateway monitoring data, or camera monitoring data into a low-dimensional device feature vector through the embedding layer. The low-dimensional device feature vector is then concatenated with the corresponding normalized data to output a time series feature matrix. The temporal feature extraction layer is constructed based on a bidirectional long short-term memory module and a multi-head self-attention module. The bidirectional long short-term memory module is used to extract long-range temporal dependency features and local event-related features from the time series feature matrix. The multi-head self-attention module is used to extract global augmented dependency features from the time series feature matrix, and the global augmented dependency features, long-range temporal dependency features, and local event-related features are concatenated and fused to obtain a unified contextual feature representation. The multi-task output layer is constructed based on a fault type prediction module, a fault time prediction module, and a processing suggestion generation module. The fault type prediction module classifies faults based on a unified context feature representation and generates predicted fault types. The fault time prediction module maps the unified context feature representation to the predicted fault time. The processing suggestion generation module infers from the unified context feature representation, generates fault processing suggestions, and outputs an operation and maintenance strategy carrying the predicted fault type, predicted fault time, and fault processing suggestions.

8. The router, gateway, and camera maintenance system as described in claim 6, characterized in that: The dataset construction module is specifically used for: Acquire a large amount of historical device monitoring data, including router monitoring data, gateway monitoring data, and camera monitoring data; the router monitoring data includes at least CPU utilization, memory utilization, interface status, packet loss rate, throughput, and temperature; the gateway monitoring data includes at least the number of sessions, NAT table size, firewall rule hit rate, DNS resolution success rate, link status, and power status; the camera monitoring data includes at least video stream quality, connection status, storage status, temperature, and alarm status. The monitoring data of each of the historical devices are preprocessed, including at least data cleaning, data normalization and data standardization. The preprocessed monitoring data of each of the historical devices is labeled, including at least the fault type, fault time and fault handling suggestions. A dataset is constructed based on the labeled monitoring data of each of the historical devices.

9. The router, gateway, and camera maintenance system as described in claim 6, characterized in that: The equipment operation and maintenance strategy generation model training module is specifically used for: Based on stratified sampling, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:

1. The equipment operation and maintenance strategy generation model is trained using the training set and a loss function. During the training process, the hyperparameters of the equipment operation and maintenance strategy generation model are continuously optimized until the loss value of the loss function is less than a preset loss threshold or a preset early shutdown condition is met. The accuracy of the strategy is calculated using the validation set to validate the trained equipment operation and maintenance strategy generation model. If the validation fails, the training set is expanded and training continues; if the validation passes, then: The F1 score and confidence level are calculated using the test set to test the validated equipment operation and maintenance strategy generation model. If the test fails, the training set is expanded and training continues; if the test passes, training ends.

10. A router, gateway, and camera maintenance system as described in claim 6, characterized in that: The model iterative optimization module is specifically used for: Network devices record real-time operation and maintenance logs, including at least real-time operation and maintenance policies, real-time device monitoring data, and policy execution feedback. Based on a federated learning mechanism, an incremental dataset is constructed using these operation and maintenance logs to train and validate the deployed device operation and maintenance policy generation model. The model parameters of the validated device operation and maintenance policy generation model are uploaded to the server as local parameters. The server aggregates the local parameters uploaded by each network device to obtain global parameters. Based on these global parameters, the deployed device operation and maintenance policy generation models are iteratively optimized.

Images