Deployment strategy method fusing model knowledge graph and artificial intelligence model

Abstract

The invention provides a deployment strategy method for fusing a model knowledge graph and an artificial intelligence model. The deployment strategy method comprises the steps that a center node receives an edge fusion intention vector, adopts a time synchronization algorithm to align timestamps, fuses multi-node vector data and generates a consistency intention vector; extracting historical intention features from the knowledge graph according to the consistency intention vector, and generating a historical intention feature set; performing fine tuning on the pre-training language model by adopting a transfer learning algorithm to generate a scene adaptation intention classification model; processing the real-time operation and maintenance instruction data through a scene adaptive intention classification model to generate an intention classification tag set; distributing the updated model parameter set to edge nodes, and generating a node synchronization model deployment state; the edge node generates a new edge fusion intention vector according to the real-time operation and maintenance data, and transmits the new edge fusion intention vector to the center node; and the center node dynamically adjusts the scene adaptation intention classification model by adopting an online learning algorithm to generate a dynamic adjustment intention classification model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence data processing technology, and in particular to a deployment strategy method for integrating a model knowledge graph with an artificial intelligence model. Background Art

[0002] Deployment strategy of operation and maintenance intention and fusion model knowledge graph and artificial intelligence model; Operation and maintenance intention understanding aims to accurately understand the operation intention of operation and maintenance personnel through knowledge graph and natural language processing technology, and provide accurate operation and maintenance assistance and decision support. However, in the hierarchical deployment of the fusion model, operation and maintenance knowledge and intention understanding models may be distributed at different levels, and cross-level information transmission and state synchronization may affect the accuracy and real-time nature of intention understanding. At the same time, operation and maintenance scenarios are complex and changeable, and the generalization ability and adaptability of the intention understanding model are also facing challenges, requiring continuous improvement in knowledge precipitation, model optimization, tool integration, etc. Summary of the Invention

[0003] The present invention provides a deployment strategy method for integrating model knowledge graph and artificial intelligence model, which mainly includes: Event data, device data, and operation data are obtained from operation and maintenance data, and a knowledge graph with events and devices as nodes and operation relationships as edges is constructed to generate an initial intent semantic vector. In a hierarchical deployment architecture, the initial intent semantic vector is distributed to edge nodes, and an edge fusion intent vector is generated through local operation and maintenance data and transmitted to the central node. The central node receives the edge fusion intent vector, aligns the timestamps using a time synchronization algorithm, fuses multi-node vector data, and generates a consistent intent vector. Based on the consistent intent vector, historical intent features are extracted from the knowledge graph to generate a historical intent feature set. A transfer learning algorithm is used to train the pre-trained language. The model is fine-tuned to generate a scenario-adaptive intent classification model; real-time operation and maintenance instruction data is processed through the scenario-adaptive intent classification model to generate an intent classification label set; based on the intent classification label set, the operation and maintenance optimization strategy is retrieved from the knowledge base, and the scenario-adaptive intent classification model is updated using an online learning algorithm to generate an updated model parameter set; the updated model parameter set is distributed to the edge nodes to generate the node synchronization model deployment status; the edge nodes generate new edge fusion intent vectors based on real-time operation and maintenance data and pass them to the central node; the central node uses an online learning algorithm to dynamically adjust the scenario-adaptive intent classification model to generate a dynamically adjusted intent classification model.

[0004] Furthermore, a knowledge graph with events and devices as nodes and operation relationships as edges is constructed to generate an initial intent semantic vector, including: extracting operating status, device attributes and action records from operation and maintenance logs and operation records through preset log parsing rules to construct initial graph structure data; using a graph convolutional neural network to calculate the weighted sum of neighbor features of each node, extracting relationship features between nodes, and obtaining a node relationship feature vector; if the dimension of the node relationship feature vector exceeds a preset threshold, the node relationship feature vector is reduced in dimension through principal component analysis to obtain a feature vector after dimensionality reduction; based on the feature vector after dimensionality reduction, the cosine similarity is used to calculate the matching degree with the preset intent template to obtain the initial intent semantic vector.

[0005] Furthermore, the initial intent semantic vector is distributed to the edge node, and an edge fusion intent vector is generated through local operation and maintenance data, including: distributing the initial intent semantic vector to at least one edge node through a distributed message middleware; obtaining environmental feature data from the local operation and maintenance data of the edge node, and using a weighted average algorithm to perform feature fusion on the initial intent semantic vector and the environmental feature data to generate an edge fusion intent vector; if the dimension of the edge fusion intention vector exceeds a preset threshold, the edge fusion intention vector is reduced in dimension through a principal component analysis algorithm to obtain a reduced-dimensional edge fusion intention vector; and the reduced-dimensional edge fusion intention vector is transmitted to the central node through a message queue.

[0006] Furthermore, the central node receives the edge fusion intention vector, uses the time synchronization algorithm to align the timestamps, and fuses the multi-node vector data, including: obtaining the fusion intention vector from each edge node, using the network time protocol to calibrate the timestamp, and if the timestamp deviation exceeds the preset threshold, performing a resynchronization operation to obtain a time-aligned intention vector set; based on the time-aligned intention vector set, using the weighted average algorithm to fuse the multi-node vector data, if the integrity of the vector data is lower than the preset threshold, reducing its weight to obtain a preliminary fusion vector; through consistency verification, if the verification result shows that the deviation between vectors exceeds the preset threshold, the k-means clustering algorithm is used to group and adjust the preliminary fusion vector to obtain a consistent intention vector.

[0007] Furthermore, based on the consistent intention vector, historical intent features are extracted from the knowledge graph to generate a historical intent feature set, including: using cosine similarity calculation to extract historical intent features associated with the operation and maintenance scenarios from the pre-established knowledge graph to obtain a historical intent feature set; performing vector normalization on the historical intent feature set, and if the module length deviation of the feature vector exceeds a preset threshold, performing standardization adjustment to obtain a normalized feature set; based on the normalized feature set, obtaining scene description data from the operation and maintenance scenario database through scene association rules to obtain a scene enhancement feature set; using the principal component analysis algorithm to extract the main feature dimensions, and if the contribution rate of the feature dimension is lower than the preset threshold, eliminating the low-contribution dimension to obtain a reduced-dimensionality feature set.

[0008] Furthermore, a transfer learning algorithm is used to fine-tune the pre-trained language model to generate a scenario-adaptive intent classification model, including: obtaining an intent feature set from historical data, preprocessing it with a feature extraction algorithm, and obtaining structured historical intent features; loading the pre-trained language model through a transfer learning algorithm, fine-tuning the model based on the structured historical intent features, and determining the initial parameters; iteratively updating the model parameters using a gradient descent optimization algorithm, calculating the gradient according to the loss function, and obtaining an optimized parameter set; if the classification accuracy is lower than the preset threshold, adjusting the learning rate and re-executing the gradient descent optimization to generate a scenario-adaptive intent classification model.

[0009] Furthermore, the real-time operation and maintenance instruction data is processed by a scene-adapted intent classification model to generate an intent classification label set, including: obtaining instruction data in real-time operation and maintenance, cleaning invalid characters and noise through preprocessing, and obtaining normalized instruction text; extracting semantic units from the normalized instruction text through text analysis, and using word segmentation and part-of-speech tagging to obtain a structured text sequence; using an attention mechanism to assign weights to the semantic units in the structured text sequence, enhancing the weights of key features, and obtaining a weighted feature vector; if the dimension of the weighted feature vector exceeds a preset threshold, principal component analysis is used to reduce the dimension to obtain a compressed feature vector; based on the compressed feature vector, the intent classification model is used for classification to obtain an intent classification label set.

[0010] Furthermore, the updated model parameter set is distributed to the edge nodes to generate the node synchronization model deployment status, including: receiving the model parameter update package through the message queue, and using hash value comparison to confirm the integrity of the parameter package; the edge node listens to the update event in the message queue, filters the parameter package according to the node identifier, and applies the model parameter change in an incremental update manner; a polling algorithm is used to calculate the resource occupancy rate of the edge node, and the task allocation plan is obtained in combination with the node processing capacity score; the inference request is distributed to the edge node through the load balancer, the model inference calculation is performed, the inference time and resource consumption indicators are recorded, and the node synchronization model deployment status is generated.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: The present invention discloses a method for intelligent operation and maintenance intention recognition based on a distributed architecture, which extracts the initial intent semantic vector by constructing a knowledge graph, performs feature fusion at the edge node to generate an edge fusion intent vector, fuses multi-node data at the central node to generate a consistent intent vector, matches historical intent features for transfer learning, and dynamically updates the scenario adaptation intention classification model. The present invention solves the problem that traditional centralized operation and maintenance systems are difficult to cope with complex and changeable operation and maintenance scenarios in large-scale distributed environments, realizes a hierarchical deployment architecture of edge intelligence and center collaboration, continuously optimizes model parameters through online learning, improves the accuracy and real-time performance of operation and maintenance intention recognition, and provides strong support for intelligent operation and maintenance decision-making. The technical effect of the present invention is to significantly improve the intelligence level of operation and maintenance intention recognition in a distributed environment and enhance the adaptability of the operation and maintenance system to complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flowchart of a deployment strategy method for integrating model knowledge graph and artificial intelligence model of the present invention.

[0013] Figure 2 This is a schematic diagram of a deployment strategy method for integrating a model knowledge graph with an artificial intelligence model according to the present invention.

[0014] Figure 3 This is another schematic diagram of a deployment strategy method for integrating model knowledge graph and artificial intelligence model according to the present invention.

[0015] Figure 4 This is another schematic diagram of a deployment strategy method for integrating model knowledge graph and artificial intelligence model according to the present invention.

[0016] Figure 5 This is another schematic diagram of a deployment strategy method for integrating model knowledge graph and artificial intelligence model according to the present invention.

[0017] Figure 6 This is another schematic diagram of a deployment strategy method for integrating model knowledge graph and artificial intelligence model according to the present invention.

[0018] Figure 7 This is another schematic diagram of a deployment strategy method for integrating model knowledge graph and artificial intelligence model according to the present invention. DETAILED DESCRIPTION

[0019] To further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and examples. The present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0020] like Figure 1-7 In this embodiment, a deployment strategy method for integrating a model knowledge graph with an artificial intelligence model specifically includes: S101. The first step is to extract event data, device data, and operation data from operation and maintenance logs and operation records, build a knowledge graph, use events and devices as nodes, and operation relationships as edges, use a graph convolutional neural network to extract the relationship features between nodes, and generate an initial intent semantic vector.

[0021] Event data, device data, and operation data are collected from operation and maintenance logs and operation records. Pre-set log parsing rules are used to extract operating status, device attributes, and action records to generate structured data. Based on this structured data, a knowledge graph is constructed with events and devices as nodes and operation relationships as edges to generate initial graph data. For this initial graph data, a graph convolutional neural network is used to calculate the weighted sum of each node's neighbor features, extracting inter-node relationship features and generating a node relationship feature vector. If the dimension of the node relationship feature vector exceeds a preset threshold, principal component analysis is used to reduce the dimension of the node relationship feature vector to generate a reduced feature vector. Based on the reduced feature vector, cosine similarity is used to calculate the degree of match with the preset intent template to generate an initial intent semantic vector. New event and operation features are extracted from real-time operation and maintenance data, and a new feature set is generated using a feature extraction algorithm to generate edge feature data. If the integrity of the edge feature data exceeds a preset threshold, principal component analysis is used to generate an edge fusion intent, resulting in an edge intent vector. The edge intent vector is transmitted to the central node via a message queue protocol to determine the completion status of the transmission. Receive edge intention vectors from the central node, use clustering algorithm to analyze vector distribution, and obtain node status classification.

[0022] Specifically, event data, device data, and operation data are collected from operation and maintenance logs and operation records. Pre-set log parsing rules are used to extract operational status, device attributes, and action records. For example, device IDs, event types, timestamps, and operation instructions are extracted from the logs to generate structured data. Based on this structured data, a knowledge graph is constructed with events and devices as nodes and operation relationships as edges. For example, device A and event B are used as nodes, and operation C is used as an edge to generate initial graph data. For this initial graph data, a graph convolutional neural network is used to calculate the weighted sum of neighbor features for each node. For example, a GCN algorithm is used to weightedly sum the neighbor features of device nodes and event nodes. Internode relationship features are extracted to generate a node relationship feature vector. If the dimension of the node relationship feature vector exceeds a preset threshold, for example, if the dimension is greater than 100, principal component analysis is used to reduce the node relationship feature vector to, for example, 50, to obtain a reduced feature vector. Based on this reduced feature vector, cosine similarity is used to calculate its match with a preset intent template, for example, by calculating the cosine similarity between the feature vector and the template vector to generate an initial intent semantic vector. New event features and operation features are extracted from real-time operation and maintenance data, and new feature sets are generated through feature extraction algorithms. For example, device status changes and operation frequencies are extracted from real-time logs to obtain edge feature data. If the integrity of the edge feature data is higher than the preset threshold, for example, the integrity reaches 90%, the edge fusion intent is generated through the principal component analysis algorithm, for example, the feature set is reduced in dimension and fused to obtain an edge intent vector. Based on the edge intent vector, it is transmitted to the central node through the message queue protocol, for example, the vector is transmitted to the central server using the Kafka protocol to determine the completion status of the transmission. The edge intent vector is received from the central node, and the vector distribution is analyzed using a clustering algorithm, for example, the vector is clustered using the K-means algorithm to obtain the node status classification.

[0023] S102. In the second step, in the hierarchical deployment architecture, the initial intent semantic vector is distributed to the edge nodes. The edge nodes perform feature fusion based on local operation and maintenance data, generate edge fusion intent vectors, and transmit them to the central node through the message queue.

[0024] The distributed message middleware obtains an initial intent semantic vector from the central node and distributes it to at least one edge node, resulting in a distributed initial intent semantic vector. At the edge node, environmental feature data is obtained from local operation and maintenance data. A weighted average algorithm is used to fuse the distributed initial intent semantic vector and the environmental feature data to generate an edge-fused intent vector. If the dimension of the edge-fused intent vector exceeds a preset threshold, the edge-fused intent vector is reduced using a principal component analysis algorithm to obtain a reduced-dimensional edge-fused intent vector. The reduced-dimensional edge-fused intent vector is transmitted to the central node via a message queue protocol, and the transmission completion status is determined. The reduced-dimensional edge-fused intent vector is received from the central node, and a clustering algorithm is used to analyze the vector distribution to obtain an intent category distribution. Based on the intent category distribution, a classification algorithm is used to identify the intent in the reduced-dimensional edge-fused intent vector and determine the final intent category. Event features and operation features are obtained from real-time operation and maintenance data, and a feature extraction algorithm is used to generate a new feature set to generate a new edge-fused intent vector. If the feature set integrity of the new edge-fused intent vector exceeds a preset threshold, the new edge-fused intent vector is transmitted to the central node via a message queue protocol, and the new transmission completion status is determined. According to the new edge fusion intention vector, the clustering algorithm is used to update the intention recognition model of the central node to obtain the optimized intention recognition result.

[0025] Specifically, an initial intent semantic vector is obtained from the central node through a distributed messaging middleware. For example, Kafka is used to distribute the initial intent semantic vector with a dimension of 128 to the edge nodes. At the edge node, environmental feature data is obtained from local operation and maintenance data, such as 10-dimensional features such as temperature and humidity. A weighted average algorithm is used to fuse the initial intent semantic vector and environmental feature data to generate an edge-fused intent vector with a dimension of 138. If the dimension of the edge-fused intent vector exceeds the preset threshold of 100, the dimension is reduced to 80 using a principal component analysis algorithm to obtain a reduced-dimensional edge-fused intent vector. The reduced-dimensional edge-fused intent vector is transmitted to the central node via a message queue protocol, for example, using RabbitMQ to confirm the transmission completion status. The reduced-dimensional edge-fused intent vector is received from the central node, and the vector distribution is analyzed using the K-means clustering algorithm to obtain a distribution of five intent categories. Based on the intent category distribution, the support vector machine algorithm is used to perform intent recognition on the reduced-dimensional edge-fused intent vector, determining the final intent category as "equipment maintenance." Event and operation features are extracted from real-time operation and maintenance data, such as equipment fault logs and operation records. A new feature set is generated using the TF-IDF feature extraction algorithm to obtain a new edge fusion intent vector. If the feature set integrity of the new edge fusion intent vector exceeds a preset threshold of 90%, the new edge fusion intent vector is transmitted to the central node via the message queue protocol to confirm the new transmission completion status. Based on the new edge fusion intent vector, the DBSCAN clustering algorithm is used to update the central node's intent recognition model, resulting in optimized intent recognition results.

[0026] S103, the third step: the central node receives the edge fusion intention vector, uses the time synchronization algorithm to align the timestamps of each node data, fuses the multi-node vector data, and generates a consistent intention vector.

[0027] An edge-fused intent vector is obtained from each edge node. The timestamps of the edge-fused intent vectors are calibrated using the Network Time Protocol to obtain a set of time-calibrated vectors. If the timestamp deviation of any vector in the time-calibrated vector set exceeds a preset threshold, a resynchronization operation is performed to obtain a time-aligned intent vector set. Based on the time-aligned intent vector set, a weighted average algorithm is used to fuse the vector data from multiple nodes to obtain a preliminary fused vector. If the integrity of the preliminary fused vector falls below a preset threshold, the weights of the vectors with insufficient integrity are reduced to obtain an adjusted preliminary fused vector. A consistency check is performed on the adjusted preliminary fused vector to determine whether the inter-vector deviation exceeds a preset threshold, obtaining a consistency check result. If the consistency check result indicates that the inter-vector deviation exceeds the preset threshold, the adjusted preliminary fused vectors are grouped and adjusted using the k-means clustering algorithm to obtain a consistent intent vector. A data integrity check algorithm is used to verify the integrity of the consistent intent vector, obtaining an integrity check result. If the integrity check result meets the preset integrity standard, the consistent intent vector is distributed to each edge node via the distributed message middleware to determine the final fusion result. Based on the final fusion result, a clustering algorithm is used to identify the intent of the consistent intent vector to obtain the final intent category.

[0028] Specifically, edge-fused intent vectors are obtained from each edge node. Timestamps are calibrated using the Network Time Protocol (NTP), with a time deviation threshold of 50 milliseconds. If a node's timestamp deviation reaches 60 milliseconds, a resynchronization mechanism is triggered to align the timestamps of all nodes. Based on the aligned intent vectors, a weighted averaging algorithm is used for fusion. The weight coefficients are dynamically adjusted based on data integrity. If a vector is missing more than 20% of its feature dimensions, its weight is reduced to 0.5. A consistency check is performed on the preliminary fused vectors, and the Euclidean distance deviation is calculated. If the deviation exceeds a threshold of 0.3, the k-means clustering algorithm is used for grouping and adjustment, with the number of cluster centers set to 3 and an upper limit of 100 iterations. The consistent intent vectors are verified using a data integrity check algorithm to check whether the missing value ratio is less than 5%. If this condition is met, the vectors are distributed to edge nodes via a Kafka message queue. Finally, the DBSCAN clustering algorithm is used to identify the consistent intent vectors, with a neighborhood radius of ε = 0.5 and a minimum sample size of min_samples = 5, to output the final intent category.

[0029] S104. The fourth step is to extract the historical intent features corresponding to the current operation and maintenance scenario from the pre-established knowledge graph through cosine similarity matching based on the consistent intent vector, and generate a historical intent feature set.

[0030] Based on the consistent intent vector, cosine similarity is calculated to match historical intent features associated with the current operation and maintenance scenario from a pre-established knowledge graph to obtain a historical intent feature set. Vector normalization is applied to the historical intent feature set. If the modulus deviation of the feature vector exceeds a preset threshold, it is normalized to obtain a normalized feature set. Based on the normalized feature set, scenario association rules are used to retrieve corresponding scenario description data from the operation and maintenance scenario database to obtain a scenario enhancement feature set. Principal component analysis is used to extract the primary feature dimensions from the scenario enhancement feature set. If the contribution rate of a feature dimension falls below a preset threshold, the low-contributing dimension is removed to obtain a reduced-dimensionality feature set. Using the reduced-dimensionality feature set, the k-nearest neighbor algorithm is used to cluster the feature set into groups. If the feature distance between groups exceeds a preset threshold, the group boundaries are adjusted to determine the feature grouping set. For each feature grouping set, extended feature data corresponding to each group is retrieved from the knowledge graph to obtain an extended feature set. Based on the extended feature set, a weighted fusion algorithm is used to integrate the feature data. If the completeness of the feature data falls below a preset threshold, the weight is reduced to obtain the final feature set. For the final feature set, an intent classification algorithm is used to generate an intent classification label set corresponding to the operation and maintenance scenario, resulting in an intent classification label set. Based on the intent classification label set, matching operation and maintenance optimization strategy data is retrieved from a pre-established knowledge base. An online learning algorithm is used to update the scenario-adaptive intent classification model, resulting in an updated model parameter set.

[0031] Specifically, based on the consistent intent vectors, cosine similarity is calculated with a similarity threshold of 0.85. The historical intent features with the highest matching scores are selected from the knowledge graph. For example, five sets of historical operation and maintenance scenario feature vectors are matched to form a historical intent feature set. This feature set is then normalized using the L2 norm. If the modulus deviation of a feature vector exceeds a preset threshold of 0.3, the Z-score normalization method is used to adjust all feature vectors to fall within the range [0, 1], generating a normalized feature set. Based on this normalized feature set, matching log description data is retrieved from the operation and maintenance scenario database using predefined scenario association rules, such as "CPU usage > 90% associated with an alarm scenario." 20 scenario description fields are supplemented to construct a scenario-enhanced feature set. Principal component analysis is applied to this feature set, retaining dimensions with a cumulative contribution exceeding 90% and removing minor dimensions with a contribution below 5%. The original 50-dimensional features are compressed to 15 dimensions, resulting in a reduced-dimensionality feature set. The k-nearest neighbor algorithm (k=5) is used to cluster the reduced feature set. If the Euclidean distance between groups exceeds a threshold of 2.0, the cluster center is dynamically adjusted, ultimately resulting in three feature groups. For each group, associated extended features are extracted from the knowledge graph. For example, 10 pieces of relevant device model and topology data are added to the "network latency" group to form an extended feature set. Feature weights are calculated using the entropy weight method. If a feature's missing rate exceeds 30%, its weight is reduced to 0.1. The remaining features are weighted based on their information content. After fusion, a final feature set consisting of eight core features is generated. Classification training is performed on the final feature set using the random forest algorithm, outputting six intent labels, such as "hardware failure" and "configuration error." Based on the label set, matching optimization strategies are retrieved from the knowledge base. For example, the strategy ID for "memory leak" is P-203. The model parameters are updated using the FTRL online learning algorithm with a learning rate of 0.01.

[0032] S105. The fifth step uses the historical intent feature set and the transfer learning algorithm to fine-tune the pre-trained language model, optimizes the model parameters through gradient descent, and generates a scene-adapted intent classification model.

[0033] An intent feature set is obtained from historical data. A feature extraction algorithm is used to preprocess the data to obtain structured historical intent features. A pretrained language model is loaded using a transfer learning algorithm. The model is fine-tuned based on the structured historical intent features to determine the initial model parameters. A gradient descent optimization algorithm is used to iteratively update the model parameters. The gradient of the loss function is calculated to obtain an optimized parameter set. If the classification accuracy falls below a preset threshold, the learning rate is adjusted and gradient descent optimization is rerun to determine whether the expected accuracy has been achieved. The model output is evaluated based on its scenario adaptability. Classification results on a test dataset are obtained to determine the model's intent classification performance. Based on the performance evaluation results, the feature extraction strategy is adjusted to generate new historical intent features, resulting in an updated feature set. The transfer learning algorithm is rerun using the updated feature set to fine-tune the model and determine new model parameters. An online learning algorithm receives new edge-fused intent vectors and dynamically adjusts model parameters to obtain a dynamically adjusted intent classification model. If the model performance is stable, the final results of the parameter optimization process are saved to generate a scenario-adapted intent classification model.

[0034] Specifically, user interaction logs were extracted from a historical database. The TF-IDF algorithm was used to calculate text feature weights, with a minimum word frequency threshold of 5. After filtering out low-frequency words, a 500-dimensional sparse feature matrix was obtained, which was then reduced to a 100-dimensional structured feature vector using PCA. A pre-trained BERT-base model was loaded, and a fully connected layer was added to the last layer. The structured features were fine-tuned using the Adam optimizer with an initial learning rate of 0.001. The parameters of the first eight layers were frozen, and only the top three layers were trained. Backpropagation was performed using the cross-entropy loss function, with 32 samples input per batch. Iterations were terminated when the validation set accuracy fluctuated by less than 0.5% for three consecutive rounds. If the F1 value on the test set fell below 0.85, the learning rate was decayed to 0.1 times the original value and the optimization process was restarted until the accuracy threshold of 0.9 was reached. In an e-commerce customer service scenario, a confusion matrix was used to evaluate 200,000 inquiries. To address the low recall rate in the "return" category, an N-gram feature enhancer was introduced to supplement 2-gram sequence features. During retraining, a dynamic learning rate schedule is used, automatically switching to a learning rate of 0.0001 when the loss value stagnates. During the deployment phase, JSON-formatted intent data uploaded by edge devices is received in real time via a Kafka message queue, and model parameters are updated every 10 minutes using an incremental learning algorithm. If online A / B testing shows that the model's accuracy standard deviation is less than 0.2% over a continuous 24-hour period, the model parameters are permanently stored in the HDFS distributed file system.

[0035] S106, the sixth step is to adapt the intent classification model according to the scenario, process the real-time operation and maintenance instruction data, enhance the key feature weights in the instruction text through the attention mechanism, and generate the intent classification label set.

[0036] Instruction data from real-time operations and maintenance is obtained and preprocessed to remove invalid characters and noise to obtain standardized instruction text. Semantic units are extracted from the standardized instruction text through text analysis, and word segmentation and part-of-speech tagging are used to obtain a structured text sequence. An attention mechanism is used to assign weights to the semantic units in the structured text sequence, enhancing the weights of key features to obtain a weighted feature vector. If the dimension of the weighted feature vector exceeds a preset threshold, principal component analysis is used to reduce the dimensionality to obtain a compressed feature vector. Based on the compressed feature vector, a pre-trained intent classification model is used for classification to obtain a preliminary intent label set. The preliminary intent label set is corrected using pre-set scenario rules to obtain a final intent label set. The final intent label set is used to generate an execution sequence for real-time operations and maintenance instructions, obtaining a classification result. Based on the classification result, matching operation and maintenance optimization strategy data is retrieved from a pre-established knowledge base to obtain an optimization strategy set. An online learning algorithm is used to update the scenario-adaptive intent classification model to obtain an updated model parameter set.

[0037] Specifically, real-time O&M command data is obtained and regular expression matching is used to remove special characters and garbled characters, retaining Chinese, English, and numeric text to obtain standardized command text. Semantic units are extracted through text analysis, and precise pattern segmentation is performed using the Jieba word segmentation tool. LTP part-of-speech tagging is then used to identify key components such as nouns and verbs, resulting in a structured text sequence. A multi-head attention mechanism is used to calculate semantic unit weights, with the number of attention heads set to 8. Key O&M terms such as "restart" and "capacity expansion" are given higher weights to obtain a weighted feature vector. If the weighted feature vector dimension exceeds 512, PCA dimensionality reduction is used to retain the principal components with 95% variance contribution to obtain a compressed feature vector. A pre-trained BERT classification model is loaded onto the compressed feature vector to output the probability distribution of each intent category. Labels with a top-3 probability value exceeding 0.8 are selected to obtain a preliminary intent label set. A scenario rule engine is used to match the O&M scenario library. If the label "hardware failure" conflicts with monitoring data showing a current server load rate greater than 90%, it is corrected to "resource overload," resulting in the final intent label set. An execution sequence is generated based on the final tag set. For example, the tag "database backup" automatically triggers the backup script path / bin / backup.sh. Policy documents with a similarity greater than 0.9 are retrieved from the Elasticsearch knowledge base and matched with historical solutions, such as the "Redis cache cleanup process," to obtain an optimized policy set. The FTRL online learning algorithm is used to update the fully connected layer parameters of the intent classification model at a learning rate of 0.01 to obtain the updated model parameter set.

[0038] S107, the seventh step is to retrieve matching operation and maintenance optimization strategy data from a pre-established knowledge base based on the intent classification label set, use an online learning algorithm to update the scenario adaptation intent classification model, and generate an updated model parameter set.

[0039] The intent classification label set is obtained from the data source and filtered using a label matching mechanism to determine the matching label set associated with the operation and maintenance optimization strategy. Using a knowledge retrieval efficiency metric, the operation and maintenance optimization strategy data corresponding to the matching label set is retrieved from the preset knowledge base to generate a strategy dataset. Based on the strategy generation rules, the strategy dataset is sorted and filtered to obtain a preferred strategy set with high scenario adaptation accuracy. If the scenario adaptation accuracy of the preferred strategy set falls below a preset threshold, the scenario-adapted intent classification model is adjusted using an online learning algorithm to generate a temporary parameter set. The temporary parameter set is then fused with the existing updated parameter set through a parameter optimization path to generate a new parameter set. The scenario-adapted intent classification model is updated using the new parameter set to obtain an updated model structure. Based on the updated model structure, the intent classification label set is reclassified and predicted to generate an optimized intent classification label set. If the confidence score of the optimized intent classification label set falls below a preset threshold, the feature vector corresponding to the label set is reconstructed using a clustering algorithm to generate a reconstructed feature vector. The updated scenario-adapted intent classification model is retrained using the reconstructed feature vector, and the parameters are optimized using a loss function to obtain the final optimized intent classification model.

[0040] Specifically, the intent classification label set is obtained from the data source. Cosine similarity is used to calculate label matching, and a similarity threshold of 0.85 is set to select matching label sets. An inverted index is used to retrieve policy data from the knowledge base, and retrieval efficiency is calculated using TF-IDF weighting. A dataset of the top 50 policies is returned. The dataset is sorted based on policy priority rules (for example, a weight of 0.6 for troubleshooting timeliness and a weight of 0.4 for resource consumption). Policies with a score above 0.9 are retained to form a preferred set. If the average adaptation accuracy of the preferred set falls below the threshold of 0.8, the FTRL online learning algorithm is used to adjust the model with a learning rate of 0.01 to generate a temporary parameter set. The temporary parameter set (weight 0.3) is fused with the existing parameter set (weight 0.7) using a weighted average method to generate a new parameter set. This new parameter set is loaded into the BERT-based classification model, the hidden layer dimension is updated to 768, and the updated model structure is output. The updated model is used to predict the label set, and labels with a softmax confidence score above 0.95 are used as the optimization results. If the confidence level is insufficient, K-means clustering (k=5) is used to reconstruct the feature vector. PCA is then used to reduce the dimensionality to 300 and generate the reconstructed vector. Incremental training is performed using the reconstructed vector, with the cross-entropy loss function iterated 10 times. The final model accuracy is improved to 92%.

[0041] S108, Step 8: In the hierarchical deployment architecture, the updated model parameter set is distributed to the edge nodes through the message queue, the inference tasks are allocated using the load balancing algorithm, and the node synchronization model deployment status is generated.

[0042] 1. Model parameter update packages sent from the central server are distributed to edge nodes via a message queue. The integrity of the parameter packages is verified, and hash value comparisons are used to confirm that the parameter packages have not been tampered with. 2. Edge nodes listen for model parameter update events in the message queue, select the parameter packages applicable to them based on their node identifiers, and apply the model parameter changes using an incremental update method. 3. The load balancer uses a round-robin algorithm to calculate the current resource utilization of each edge node and, combined with the node processing power score, determines the optimal task allocation. 4. The inference request routing module receives model inference requests initiated by the client, parses the request content type and priority, and distributes them to appropriate edge nodes via the load balancer. 5. The edge node performs model inference calculations, recording inference time and resource consumption metrics. If inference fails, a backup node takes over and confirms the inference result. 6. The state synchronization module regularly collects operational status information from each edge node, including model version number, parameter update timestamp, and resource utilization, to build a global deployment status view. 7. The health monitoring system analyzes edge node status data. If an abnormal node status is detected, it determines the node's recovery status by restarting the service or redistributing model parameters. 8. Based on the global deployment status view, event and operation features are extracted from the real-time O&M data of edge nodes, and a feature extraction algorithm is used to generate a feature set. 9. If the feature set integrity exceeds a preset threshold, a principal component analysis algorithm is used to generate an edge fusion intent, resulting in an intent vector. This intent vector is then transmitted to the central node via a message queue to determine the transmission completion status.

[0043] Specifically, the message queue uses RabbitMQ to distribute model parameter update packages. Each parameter package is accompanied by a SHA-256 hash value. Upon receiving the package, the edge node calculates the local hash value and compares it with the original value. If a match occurs, the package passes verification. The edge node listens for messages on a specific topic through Kafka, filters the parameter packages using the node ID (e.g., Edge-Node-01), and uses a Delta update algorithm to apply only differential parameters, reducing bandwidth consumption. The load balancer, based on Nginx's weighted round-robin strategy, collects real-time CPU utilization (e.g., 70%) and memory usage (e.g., 4GB / 8GB). This information is combined with node performance scores (e.g., 1.2TFlops of computing power) to generate a task allocation weight table. The inference request routing module parses the Content-Type (e.g., application / json) and Priority fields (e.g., P1-P3) in the HTTP request header and calls the load balancer interface to distribute the request to the node with the highest weight. Edge nodes perform TensorRT inference calculations, recording inference latency (e.g., 150ms) and GPU memory usage (e.g., 3.5GB / 6GB). If a timeout (threshold of 500ms) is exceeded, Consul service discovery is triggered to switch to a standby node. The state synchronization module collects node metrics via Prometheus every 30 seconds, aggregating model version (e.g., v2.1.3), parameter update time (UTC timestamp), and CPU utilization (e.g., 65%) to generate a global state tree for Etcd storage. The health monitoring system uses Zabbix to monitor node heartbeat intervals. If three consecutive timeouts occur (default interval is 1 second), an Ansible playbook is triggered to restart the service or re-fetch parameter packages from MinIO storage. The feature extraction module extracts metrics such as QPS (e.g., 200 queries / second) and error rate (e.g., 0.5%) from node logs. Using the PCA algorithm, the module reduces the dimensionality to a 10-dimensional vector. This generated intent vector is then transmitted to the central node via MQTT. Upon completion of the transmission, the ACK status flag in Redis is updated.

[0044] S109, the ninth step is to extract new event features and operation features from real-time operation and maintenance data according to the node synchronization model deployment status, generate new edge fusion intent vectors, and transmit them to the central node through the message queue.

[0045] Based on the node synchronization model deployment status, event data and operation data are extracted from real-time operation and maintenance data. A feature extraction algorithm is used to generate event and operation feature sets, resulting in an initial feature set. If the integrity of the initial feature set exceeds a preset threshold, the event and operation feature sets are reduced in dimension using a principal component analysis algorithm to obtain a reduced feature vector. Based on the reduced feature vector, the cosine similarity algorithm is used to calculate the degree of match with the preset intent template, generating an edge fusion intent vector and determining the intent vector representation. The edge fusion intent vector is encapsulated into a message packet using a message queue protocol and transmitted to the central node, where the transmission completion status is determined. The edge fusion intent vector is received from the central node, and a clustering algorithm is used to analyze its distribution characteristics to determine the node status classification. If the node status classification is consistent with the synchronization model deployment status, the operation and maintenance data analysis model is updated using a feature data processing algorithm to determine the analysis results. Based on the analysis results, a task allocation optimization algorithm is used to adjust the computing task allocation of the edge nodes, resulting in an optimized computing task configuration. Using this optimized computing task configuration, the node synchronization model deployment parameters are updated to determine the new deployment status. According to the new deployment status, updated event features and operation features are extracted from the real-time operation and maintenance data to generate a new edge fusion intention vector, which is transmitted to the central node to determine the transmission completion status.

[0046] Specifically, based on the deployment status of the node synchronization model, event data and operation data are obtained from real-time operation and maintenance data. Feature extraction algorithms such as TF-IDF or Word2Vec are used to generate event feature sets and operation feature sets, resulting in an initial feature set containing 100-dimensional features. If the integrity of the initial feature set exceeds a preset threshold of 90%, the event feature set and operation feature set are subjected to dimensionality reduction processing using a principal component analysis algorithm, reducing the feature dimensions from 100 to 20, and obtaining a reduced-dimensional feature vector. Based on the reduced-dimensional feature vector, the cosine similarity algorithm is used to calculate the degree of match with the preset intent template. If the match is greater than 0.8, an edge-fused intent vector is generated and the intent vector representation is determined. The edge-fused intent vector is encapsulated into a message packet using a message queue protocol such as Kafka and transmitted to the central node, where the transmission completion status is determined to be successful. The edge-fused intent vector is received from the central node, and a clustering algorithm such as K-means is used to analyze the distribution characteristics of the intent vector, classifying the node status into three categories: normal, abnormal, and warning. If the node status classification is consistent with the synchronization model deployment status, the operation and maintenance data analysis model is updated using a feature data processing algorithm, such as a random forest algorithm, to determine the analysis results as node performance optimization requirements. Based on the analysis results, a task allocation optimization algorithm, such as a genetic algorithm, is used to adjust the computing task allocation of edge nodes, resulting in an optimized computing task configuration and balancing the task load across all nodes. Using this optimized computing task configuration, the deployment parameters of the node synchronization model are updated, and the new deployment status is determined to be task allocation complete. Based on the new deployment status, updated event features and operation features are extracted from the real-time operation and maintenance data to generate a new edge fusion intent vector, which is transmitted to the central node, and the transmission completion status is determined to be successful.

[0047] S1010, the tenth step: the central node receives the new edge fusion intent vector, uses the online learning algorithm to dynamically adjust the scene adaptation intent classification model, optimizes the model parameters through the loss function, and generates a dynamically adjusted intent classification model.

[0048] The central node receives the fused intent vector transmitted by the edge device and normalizes it using the data preprocessing module to obtain a standardized intent vector. If the dimension of the standardized intent vector does not match the input dimension of the scene-adapted intent classification model, the vector dimension is adjusted using a dimensionality transformation algorithm to obtain an adapted intent vector. The scene-adapted intent classification model is incrementally trained using an online learning algorithm. The gradient is calculated using a loss function, and the model parameters are updated to obtain an updated model. The adapted vector is classified and predicted based on the updated model to generate an intent classification result. If the confidence level of the intent classification result falls below a preset threshold, the adapted vector is reconstructed using a clustering algorithm to obtain a reconstructed feature vector. The updated model is retrained using the reconstructed feature vector, and the parameters are optimized using a loss function to obtain an optimized intent classification model. The reconstructed feature vector is then classified using the optimized intent classification model to obtain the final intent classification result. An intent feature set is obtained from historical data and preprocessed using a feature extraction algorithm to obtain structured historical intent features. A pretrained language model is loaded using a transfer learning algorithm, and the model is fine-tuned based on the structured historical intent features to obtain an intent classification model with optimized initial parameters.

[0049] Specifically, the central node receives the fused intent vector transmitted by the edge device and uses a data preprocessing module to normalize the vector. For example, the Z-score normalization method is used to adjust the vector's mean to 0 and its variance to 1, resulting in a standardized intent vector. If the dimension of the standardized intent vector is 128, while the input dimension of the scene-adapted intent classification model is 256, the PCA dimensionality reduction algorithm is used to reduce the vector's dimension to 256, resulting in an adapted intent vector. The scene-adapted intent classification model is incrementally trained using an online learning algorithm, such as stochastic gradient descent (SGD). The gradient is calculated using the cross-entropy loss function, and the model parameters are updated to obtain an updated model. Based on the updated model, the adapted vector is classified and predicted, for example, using the Softmax function to output the probability distribution of each class and generate an intent classification result. If the confidence level of the intent classification result falls below a preset threshold of 0.8, the adapted vector is reconstructed using the K-means clustering algorithm. The vector is divided into five clusters and the cluster center features are extracted to obtain a reconstructed feature vector. The updated model is retrained using the reconstructed feature vectors, and parameters are optimized using the mean squared error loss function to obtain the optimized intent classification model. The reconstructed feature vectors are then classified using the optimized intent classification model, for example, using a decision tree algorithm to output the final intent classification result. The intent feature set is obtained from the historical data, and the data is preprocessed using a feature extraction algorithm such as TF-IDF to obtain structured historical intent features. A transfer learning algorithm is used to load the pretrained language model BERT. The model is fine-tuned based on the structured historical intent features, for example, using the Adam optimizer with a learning rate of 0.001 to obtain an intent classification model with optimized initial parameters.

[0050] The above embodiment is only one of the preferred implementation methods of the present invention and should not be used to limit the scope of protection of the present invention. Any changes or modifications that have no substantive meaning made to the main design concept and spirit of the present invention, as long as the technical problems they solve are still consistent with the present invention, should be included in the scope of protection of the present invention.

Claims

1. A deployment strategy method for integrating model knowledge graph and artificial intelligence model, characterized in that: include: Obtain event data, device data, and operation data from operation and maintenance data, build a knowledge graph with events and devices as nodes and operation relationships as edges, and generate an initial intent semantic vector; In the hierarchical deployment architecture, the initial intent semantic vector is distributed to the edge nodes, and the edge fusion intent vector is generated through local operation and maintenance data and transmitted to the central node; The central node receives the edge fusion intent vector, uses a time synchronization algorithm to align timestamps, fuses multi-node vector data, and generates a consistent intent vector; According to the consistent intent vector, historical intent features are extracted from the knowledge graph to generate a historical intent feature set; Use transfer learning algorithms to fine-tune the pre-trained language model to generate a scenario-adaptive intent classification model; Process real-time operation and maintenance instruction data through the scenario-adaptive intent classification model to generate an intent classification label set; Based on the intent classification label set, the operation and maintenance optimization strategy is retrieved from the knowledge base, and the scenario-adaptive intent classification model is updated using an online learning algorithm to generate an updated model parameter set. Distribute the updated model parameter set to edge nodes and generate node synchronization model deployment status; The edge node generates a new edge fusion intent vector based on real-time operation and maintenance data and transmits it to the central node; The central node uses an online learning algorithm to dynamically adjust the scene adaptation intention classification model to generate a dynamically adjusted intention classification model.

2. The deployment strategy method of the fusion model knowledge graph and artificial intelligence model according to claim 1 is characterized in that: Build a knowledge graph with events and devices as nodes and operation relationships as edges to generate the initial intent semantic vector, including: Extract operating status, device attributes, and action records from operation and maintenance logs and operation records through preset log parsing rules to construct initial graph structure data; A graph convolutional neural network is used to calculate the weighted sum of the neighbor features of each node, extract the relationship features between nodes, and obtain the node relationship feature vector; If the dimension of the node relationship feature vector exceeds the preset threshold, the node relationship feature vector is reduced in dimension through principal component analysis to obtain a feature vector after dimension reduction; According to the feature vector after dimensionality reduction, the cosine similarity is used to calculate the matching degree with the preset intent template to obtain the initial intent semantic vector.

3. The deployment strategy method of the fusion model knowledge graph and artificial intelligence model according to claim 2 is characterized in that: Distribute the initial intent semantic vector to the edge node and generate an edge fusion intent vector based on local operation and maintenance data, including: Distribute the initial intent semantic vector to at least one edge node through a distributed messaging middleware; Obtain environmental feature data from the local operation and maintenance data of the edge node, use the weighted average algorithm to fuse the initial intent semantic vector and environmental feature data, and generate an edge fusion intent vector; If the dimension of the edge fusion intention vector exceeds the preset threshold, the edge fusion intention vector is reduced in dimension using the principal component analysis algorithm to obtain the reduced-dimensional edge fusion intention vector; The reduced-dimensional edge fusion intent vector is transmitted to the central node through the message queue.

4. The deployment strategy method of the fusion model knowledge graph and artificial intelligence model according to claim 3 is characterized in that: The central node receives the edge fusion intent vector, uses a time synchronization algorithm to align timestamps, and fuses multi-node vector data, including: Obtain fused intent vectors from each edge node, calibrate timestamps using the network time protocol, and perform a resynchronization operation if the timestamp deviation exceeds a preset threshold to obtain a time-aligned intent vector set; Based on the time-aligned intention vector set, a weighted average algorithm is used to fuse multi-node vector data. If the integrity of the vector data is lower than the preset threshold, its weight is reduced to obtain a preliminary fusion vector. Through consistency check, if the check result shows that the deviation between vectors exceeds the preset threshold, the k-means clustering algorithm is used to group and adjust the preliminary fusion vector to obtain the consistent intention vector.

5. The deployment strategy method of the fusion model knowledge graph and artificial intelligence model according to claim 4 is characterized in that: According to the consistent intent vector, historical intent features are extracted from the knowledge graph to generate a historical intent feature set, including: Using cosine similarity calculation, we extract historical intent features associated with operation and maintenance scenarios from the pre-established knowledge graph to obtain a historical intent feature set. Perform vector normalization on the historical intent feature set. If the modulus deviation of the feature vector exceeds a preset threshold, perform standardization adjustment to obtain a normalized feature set. According to the normalized feature set, the scenario description data is obtained from the operation and maintenance scenario database through the scenario association rules to obtain the scenario enhanced feature set; The principal component analysis algorithm is used to extract the main feature dimensions. If the contribution rate of the feature dimension is lower than the preset threshold, the low-contribution dimension is eliminated to obtain a reduced-dimensional feature set.

6. The deployment strategy method of the fusion model knowledge graph and artificial intelligence model according to claim 5 is characterized in that: A transfer learning algorithm is used to fine-tune the pre-trained language model to generate a scenario-adapted intent classification model, including: Obtain the intent feature set from historical data, preprocess it using feature extraction algorithm, and obtain structured historical intent features; Load the pre-trained language model through the transfer learning algorithm, fine-tune the model based on the structured historical intent features, and determine the initial parameters; The gradient descent optimization algorithm is used to iteratively update the model parameters, and the gradient is calculated according to the loss function to obtain the optimized parameter set; If the classification accuracy is lower than the preset threshold, the learning rate is adjusted and the gradient descent optimization is re-executed to generate a scene-adapted intent classification model.

7. The deployment strategy method of the fusion model knowledge graph and artificial intelligence model according to claim 1 is characterized in that: The scenario-adaptive intent classification model processes real-time operation and maintenance instruction data and generates an intent classification label set, including: Obtain instruction data from real-time operation and maintenance, and remove invalid characters and noise through preprocessing to obtain standardized instruction text; Extract semantic units from the standardized instruction text through text analysis, and use word segmentation and part-of-speech tagging to obtain a structured text sequence; The attention mechanism is used to assign weights to semantic units in structured text sequences, enhance the weights of key features, and obtain weighted feature vectors; If the dimension of the weighted eigenvector exceeds a preset threshold, the dimension is reduced by principal component analysis to obtain a compressed eigenvector; According to the compressed feature vector, the intent classification model is used for classification to obtain the intent classification label set.

8. The deployment strategy method of the fusion model knowledge graph and artificial intelligence model according to claim 1 is characterized in that: Distribute the updated model parameter set to edge nodes and generate node synchronization model deployment status, including: Receive model parameter update packages through the message queue and use hash value comparison to confirm the integrity of the parameter packages; The edge node listens for update events in the message queue, filters the parameter package based on the node identifier, and applies the model parameter changes using an incremental update method; A polling algorithm is used to calculate the resource occupancy rate of edge nodes, and the task allocation plan is obtained by combining the node processing capacity score; The load balancer distributes inference requests to edge nodes, performs model inference calculations, records inference time and resource consumption indicators, and generates node synchronization model deployment status.

Images