Operation and maintenance technical service problem classification method and system based on deep learning

Abstract

This invention provides a method and system for classifying operation and maintenance (O&M) technical service issues based on deep learning. First, it acquires a set of O&M technical service issue description texts with accompanying collection time information. Next, it performs dynamic updating of the O&M scenario feature library and feature evolution analysis to generate scenario feature evolution trajectory data. Then, it inputs the O&M issue text units, scenario feature data, and scenario feature evolution trajectory data into a pre-trained deep learning classification model to obtain initial classification criteria. Based on historical O&M issue classification results, O&M processing effect data, and scenario feature evolution trajectory data, it adjusts the feature weights and feature association rules in the initial classification criteria in stages to obtain optimized classification criteria. Finally, it performs multi-level association matching between the optimized classification criteria and a dynamically updated O&M technical service issue classification case library to generate O&M issue classification results. This allows it to adapt to dynamically changing O&M scenarios and achieve efficient and accurate issue classification.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a method and system for classifying operation and maintenance technical service issues based on deep learning. Background Technology

[0002] In the current field of operations and maintenance (O&M) technical services, with the rapid development of information technology and the continuous improvement of enterprise informatization, O&M work faces increasingly complex and diverse challenges. O&M technical service issues cover many aspects such as software failures, system performance problems, and network connectivity anomalies, and these issues continue to evolve over time and with business development.

[0003] Traditional methods for classifying operational and maintenance (O&M) technical service issues primarily rely on human experience or simple keyword matching rules. While human experience-based classification can achieve a certain degree of accuracy by leveraging the expertise of O&M personnel, it suffers from low efficiency when dealing with massive amounts of O&M issue data. Furthermore, the varying experience levels among different O&M personnel make it difficult to guarantee the consistency and accuracy of classification results. Classification methods based on simple keyword matching rules lack a deep understanding of the semantics of the issues and fail to accurately grasp their essential characteristics. For issues with complex descriptions or ambiguous meanings, the classification results are often unsatisfactory. In addition, existing methods do not fully consider the dynamic changes in O&M scenarios and the guiding role of historical O&M data in issue classification, making them ill-suited to adapt to constantly changing O&M environments and unable to meet the demands of modern O&M work for efficient and accurate issue classification. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for classifying operational and maintenance technical service problems based on deep learning, the method comprising:

[0005] Obtain a set of text describing operation and maintenance technical service issues. The set of text describing operation and maintenance technical service issues contains multiple text units of operation and maintenance issues, and each text unit of operation and maintenance issues is accompanied by collection time information.

[0006] The system performs dynamic updates and feature evolution analysis of the operation and maintenance scenario feature library to generate scenario feature evolution trajectory data. The operation and maintenance scenario feature library stores scenario feature data corresponding to the operation and maintenance scenarios.

[0007] Each operation and maintenance problem text unit, scenario feature data, and scenario feature evolution trajectory data are input into a pre-trained deep learning classification model to obtain the initial classification basis for each operation and maintenance problem text unit.

[0008] Based on historical operation and maintenance problem classification results, corresponding operation and maintenance processing effect data, and scenario feature evolution trajectory data, the feature weights and feature association rules in the initial classification criteria are adjusted in stages to obtain the optimized classification criteria.

[0009] The optimized classification criteria are matched with the dynamically updated O&M technical service problem classification case library at multiple levels to generate the O&M technical service problem classification result corresponding to each O&M problem text unit.

[0010] In another aspect, embodiments of the present invention also provide a deep learning-based operation and maintenance technical service problem classification system, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0011] Based on the above, this embodiment of the invention acquires a set of operation and maintenance (O&M) technical service problem description texts containing collection time information, performs dynamic updates and feature evolution analysis of the O&M scenario feature library, and generates scenario feature evolution trajectory data. This allows for real-time tracking of changes in the O&M scenario, enabling the classification process to adapt to the constantly evolving O&M environment and enhancing the flexibility and adaptability of the classification. O&M problem text units, scenario feature data, and scenario feature evolution trajectory data are input into a pre-trained deep learning classification model to obtain initial classification criteria. Leveraging the powerful capabilities of deep learning, the complex relationship between problem texts and scenario features is deeply explored, improving the accuracy of the initial classification. Based on historical O&M problem classification results, corresponding O&M processing effect data, and scenario feature evolution trajectory data, the feature weights and feature association rules in the initial classification criteria are adjusted in stages. This fully utilizes the experience of historical data to further optimize the classification criteria, making the classification results more closely aligned with actual O&M needs. Finally, the optimized classification criteria are matched with a dynamically updated O&M technical service problem classification case library at multiple levels to generate O&M technical service problem classification results corresponding to each O&M problem text unit, effectively improving the efficiency and quality of O&M work. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the execution flow of the deep learning-based operation and maintenance technical service problem classification method provided in the embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of exemplary hardware and software components of the deep learning-based operation and maintenance technical service problem classification system provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1This is a flowchart illustrating a deep learning-based method for classifying operational and maintenance technical service issues, as provided in one embodiment of the present invention. The following is a detailed description of this deep learning-based method for classifying operational and maintenance technical service issues.

[0015] Step S110: Obtain a set of operation and maintenance technical service problem description texts. The set of operation and maintenance technical service problem description texts contains multiple operation and maintenance problem text units, and each operation and maintenance problem text unit is accompanied by collection time information.

[0016] This embodiment uses IoT smart home control and maintenance services as a scenario. Through the text acquisition component of the smart home control platform, data is collected from user APP maintenance work orders, device voice interaction fault descriptions, device log text-based anomaly prompts, and text entered by maintenance personnel. A combination of real-time monitoring (work orders, voice interactions) and scheduled retrieval (device logs, maintenance personnel text) is employed. After removing blank, duplicate, and irrelevant text, a collection of maintenance technical service problem description texts is formed. Each maintenance problem text unit corresponds to a specific maintenance event, such as "Smart socket APP control unresponsive, physical switch normal," and is bound to a timestamp generated by the data source as the collection time information. Fields such as "text content," "collection time," and "device identifier" are stored in JSON format.

[0017] Step S120: Perform dynamic update and feature evolution analysis of the operation and maintenance scenario feature library to generate scenario feature evolution trajectory data. The operation and maintenance scenario feature library stores scenario feature data corresponding to the operation and maintenance scenario.

[0018] The operation and maintenance scenario feature library pre-stores features of smart home devices (type, brand, communication protocol), network features (network type, router model), control scenario features (lighting, security control), and user habit features (common control periods). New features are dynamically updated and supplemented, and evolutionary analysis is used to capture the temporal changes in these features, generating scenario feature evolution trajectory data.

[0019] Step S121: Retrieve the business operation log data and historical scenario feature data of the current operation and maintenance business scenario. The business operation log data includes device operation status description data, business process execution description data and abnormal event record data under the current operation and maintenance business scenario. The historical scenario feature data is the scenario feature data stored in the operation and maintenance scenario feature library within a recent period.

[0020] The platform retrieves business operation log data from the distributed log system via a log retrieval interface: device operation status description data includes device online status, sensor data (such as temperature and humidity), and actuator feedback (such as motor action); business process execution description data includes records of linkage rule execution and remote command issuance; and abnormal event record data includes device alarms (low battery, communication interruption) and command execution failure logs. Simultaneously, it retrieves recent historical scenario feature data from the operation and maintenance scenario feature library, including records of device changes and network parameter adjustments.

[0021] Step S122: Extract scenario feature elements from the business operation log data. The scenario feature elements are descriptions that can characterize the current operation and maintenance business scenario, including device type association descriptions, business process stage descriptions, and abnormal event type descriptions.

[0022] A rule engine combined with the BERT semantic extraction model is used to extract scene feature elements. The rule engine matches candidate expressions based on preset keywords (such as "smart lighting - LED", "command issuance", "communication interruption"); the BERT model performs semantic verification on the candidate expressions, eliminates invalid content, and finally determines the scene feature elements, such as the device type association expression "smart door lock - fingerprint type", the business process stage expression "linkage rule execution stage", and the abnormal event type expression "WiFi communication abnormality".

[0023] Step S123: Compare the scene feature elements with the scene feature data already stored in the operation and maintenance scene feature library, filter out scene feature elements whose similarity to the description of the stored scene feature data does not reach the preset correlation level, and treat the scene feature elements as new scene feature data, and classify them into the corresponding scene category according to the storage structure of the operation and maintenance scene feature library.

[0024] The Word2Vec model is used to convert scene feature elements and existing features in the library into semantic vectors of a unified dimension. Cosine similarity is calculated and compared with a preset threshold. Elements that do not reach the threshold are added as new scene feature data and are classified into the corresponding category according to the hierarchical structure of the operation and maintenance scene feature library (equipment feature class - secondary category of lighting / security equipment, etc. - tertiary category of specific equipment features). For example, "intelligent bathroom heater - fan heating type" is classified into "equipment feature class - environmental regulation equipment feature - bathroom heater feature" and metadata such as timestamp and source identifier are added.

[0025] Step S124: Arrange the historical scene feature data and the newly added scene feature data in the order of collection time to construct a scene feature time series dataset. The scene feature time series dataset contains scene feature data at different time points and corresponding time markers.

[0026] Extract the entry timestamps of historical features and the log timestamps of newly added features, sort them by time (with the same timestamp sorted again by "device feature class > network feature class > control scenario feature class > user habit feature class"), associate the timestamps to construct a scenario feature time series dataset, and store fields such as "time stamp", "feature category", "feature content" and "feature status (new / changed)" in a time series database.

[0027] Step S125: Based on the scene feature time series dataset, identify the changing trend of each scene feature data in the time dimension through the feature evolution analysis algorithm, including the changing trend of feature occurrence frequency, the changing trend of feature description details, and the changing trend of feature correlation.

[0028] The time series dataset is divided into sliding windows. The frequency of statistical features appearing in each window is used to identify the trend of frequency change (rising / falling / stable). The description of the same feature in different windows is compared to identify the trend of adjustment of description details (e.g., "communication abnormality" -- "2.4G WiFi abnormality"). The change of feature co-occurrence frequency is calculated to identify the trend of association change (e.g., if the co-occurrence of "smart curtain failure" and "router restart" increases, the association is strengthened).

[0029] Step S126: Based on the identified change trend, construct a feature evolution path for each scene feature data, integrate the evolution paths of all scene feature data, and generate scene feature evolution trajectory data. The scene feature evolution trajectory data includes the time change nodes of the scene features, the description of the change content, and the description of the scope of change impact.

[0030] Based on a timeline, each feature is marked with a time change node, recording the changes (e.g., "T1: Increased frequency of smart lock malfunctions"), and the scope of impact is described from the dimensions of the number of affected devices, user groups, and control scenarios (e.g., "Affects 1 smart lock, 1 user, not involving security scenarios"). All evolution paths are integrated according to feature categories to form a graph structure of scene feature evolution trajectory data, where nodes represent features, edges represent relationships, and time and impact information are labeled.

[0031] Step S130: Input each operation and maintenance problem text unit, scenario feature data, and scenario feature evolution trajectory data into the pre-trained deep learning classification model to obtain the initial classification basis for each operation and maintenance problem text unit.

[0032] The pre-trained deep learning classification model is based on the Transformer architecture and includes modules for text encoding, cross-scene feature transfer, feature fusion, and classification criteria generation. After inputting three types of data, each module processes the data and outputs initial classification criteria, including information such as the problem category and associated scene.

[0033] Step S131: Input the text unit of the operation and maintenance problem into the text encoding module of the pre-trained deep learning classification model, perform semantic parsing and vector transformation on the text unit of the operation and maintenance problem, and generate text semantic feature data corresponding to the text unit of the operation and maintenance problem. The text semantic feature data includes the core expression vector of the operation and maintenance problem, the abnormal phenomenon description vector, and the associated device expression vector.

[0034] The text encoding module consists of an embedding layer, a multi-head attention layer, and a feedforward neural network layer. The embedding layer segments words based on a domain dictionary and converts them into word embedding vectors; the multi-head attention layer calculates semantic relationships between words (e.g., the relationship between "no response" and "device offline"); the feedforward neural network layer generates text semantic feature data through linear transformations and activation functions, including core expression vectors (e.g., "control is not responding"), abnormal phenomenon description vectors (e.g., "device offline"), and associated device expression vectors (e.g., "smart curtains"), all of which are fixed-dimensional multi-value vectors.

[0035] Step S132: Input the scene feature data and scene feature evolution trajectory data into the cross-scene feature transfer module of the pre-trained deep learning classification model, perform feature alignment processing on the scene feature data and scene feature evolution trajectory data, and calculate the transfer adaptation coefficient between the scene feature data and the historical training scene feature data.

[0036] The cross-scene feature transfer module first unifies the feature dimensions (e.g., categorical and sequential features are converted to vectors of the same dimension) and matches semantically relevant features through the feature alignment submodule, and then calculates the adaptation coefficient through the adaptation coefficient calculation submodule.

[0037] Step S1321: Retrieve the historical training scene feature data of the pre-trained deep learning classification model. The historical training scene feature data of the pre-trained deep learning classification model is the scene feature data used in the training process of the pre-trained deep learning classification model, which includes the feature content and feature correlation relationship corresponding to multiple training scenes.

[0038] Historical training scenario feature data is retrieved through the model parameter interface, including smart home operation and maintenance training scenarios such as residences, apartments, and villas. Each scenario contains features such as devices and networks, as well as feature relationships (such as the causal relationship between "router failure" and "device offline"). Fields such as "scenario identifier", "feature category", and "relationship" are stored in a structured format.

[0039] Step S1322: Extract all feature elements from the current scene feature data to form the current feature element set, and extract all feature elements from the historical training scene feature data of the pre-trained deep learning classification model to form the historical feature element set.

[0040] Iterate through the "feature content" field of the current scene feature data and extract elements such as "smart door lock" and "WiFi connection" to form the current feature element set; similarly, iterate through the historical training data to extract elements to form the historical feature element set, and store them using a unique set structure.

[0041] Step S1323: Calculate the representation similarity between the current feature element set and the corresponding feature elements in the historical feature element set. The representation similarity is determined based on the semantic distance, lexical overlap and representation structure similarity of the feature elements, and generate a feature representation similarity set.

[0042] The feature elements are converted into semantic vectors, and the Euclidean distance (semantic distance) is calculated. The elements are split into basic words, and the proportion of identical words is counted (lexical overlap). The structural similarity is determined by comparing the expression structure (such as the "device-function" structure). Weights are assigned according to lexical overlap > semantic distance > structural similarity, and the corresponding element expression similarity is calculated by weighting. The results are then integrated into a feature expression similarity set.

[0043] Step S1324: Analyze the relationships between feature elements in the current scene feature data to form the current feature association network, and analyze the relationships between feature elements in the historical training scene feature data of the pre-trained deep learning classification model to form the historical feature association network.

[0044] Determine the pairwise relationships between current feature elements (based on co-occurrence frequency, logical association, and domain rules), such as the association between "smart lock malfunction" and "gateway offline". Construct the current feature association network using nodes (feature elements, size reflects frequency) and edges (associations, thickness reflects strength). Similarly, construct feature association sub-networks for each training scenario and integrate them into a historical feature association network.

[0045] Step S1325: Calculate the structural similarity between the current feature association network and the historical feature association network, including the similarity of network node connection methods, network hierarchical structure, and network core feature location, and generate feature association similarity.

[0046] The network connection matrix is ​​extracted to calculate the similarity of node connection methods; the network is divided into a core layer (with the most connections), an intermediate layer, and an edge layer, and the number and type of nodes in each layer are compared to determine the similarity of the hierarchical structure; core nodes are identified and their location coordinates are calculated to determine the similarity of core feature locations. The similarity of feature association relationships is obtained by weighted summation of the three indicators.

[0047] Step S1326: Extract the evolution trend parameters of the current scene features based on the scene feature evolution trajectory data, including feature update frequency, feature content adjustment magnitude and feature correlation change rate; extract the evolution trend parameters of historical scene features based on the time series of historical training scene feature data of the pre-trained deep learning classification model.

[0048] Extract current feature evolution trend parameters from scene feature evolution trajectory data: number of feature additions / changes per unit time (update frequency), semantic differences in expression (adjustment magnitude), and number of changes in correlation per unit time (rate of change). Construct a time series of historical training data and extract historical feature evolution trend parameters using the same method.

[0049] Step S1327: Calculate the similarity between the evolution trend parameters of the current scene features and the evolution trend parameters of the historical scene features, and generate feature evolution trend similarity.

[0050] The two types of parameters are processed using Min-Max standardization. The cosine similarity of feature update frequency, content adjustment magnitude and correlation change rate is calculated. The feature evolution trend similarity is obtained by weighted summation with equal weights.

[0051] Step S1328: Assign corresponding computational weights to the feature representation similarity set, feature association similarity, and feature evolution trend similarity. Based on the computational weights, perform a weighted summation of the average value of the feature representation similarity set, feature association similarity, and feature evolution trend similarity to obtain the transfer adaptation coefficient between the scene feature data and the historical training scene feature data of the pre-trained deep learning classification model.

[0052] The weights are assigned according to the similarity set of feature descriptions > similarity of feature association > similarity of feature evolution trend. The average value of the similarity set of feature descriptions is calculated, and then summed after being weighted separately with the other two indicators to obtain the migration adaptation coefficient in the 0-1 interval.

[0053] Step S133: Adjust the scene feature data based on the transfer adaptation coefficient to generate transfer adaptation features that are adapted to the distribution of training data of the pre-trained deep learning classification model.

[0054] If the adaptation coefficient is greater than or equal to the preset threshold, fine-tune the description details; if it is less than the threshold, optimize the feature description (refer to historical data), adjust the association strength, correct the evolution trend parameters, and generate transfer adaptation features so that their data distribution is close to the model training data.

[0055] Step S134: Input the transfer adaptation features and text semantic feature data into the feature fusion module of the pre-trained deep learning classification model to establish the relationship between the transfer adaptation features and text semantic feature data, including the correspondence of feature elements, the influence of feature weights, and the complementary relationship of feature semantics.

[0056] The feature fusion module employs an attention mechanism and a graph neural network architecture. The attention mechanism layer calculates the attention weights of feature elements, matches semantically related elements to establish correspondences (e.g., "smart curtain malfunction" and "curtain unresponsive"); adjusts the text semantic feature weights based on the frequency of transferred feature elements (e.g., if "gateway offline" is frequent, the weight of "device offline" is increased); and identifies complementary semantic information (e.g., the text "lights unresponsive" and the transferred feature "prone to network peak periods") to establish semantic complementarity.

[0057] Step S135: Based on the established association, perform weighted fusion processing on the migration adaptation features and text semantic features to generate fused feature data containing scene migration information and text semantic information.

[0058] The fusion weights are assigned according to the relationship (corresponding and complementary features have higher weights), the feature elements are sorted by weight and then concatenated, and the fusion feature data is generated by dimensional unification, which also includes scene and semantic information.

[0059] Step S136: Input the fused feature data into the classification basis generation module of the pre-trained deep learning classification model, perform feature extraction and rule matching on the fused feature data, and extract key classification information from the fused feature data, including operation and maintenance problem type association features, anomaly level corresponding features, and processing scheme matching features.

[0060] The classification is based on a feature extraction sublayer with a convolutional neural network structure (extracting abstract features using multi-size convolutional kernels), a rule matching sublayer (matching a classification rule base, such as "including communication interruption - network-related issues"), and a key information filtering sublayer, which extracts key classification information such as the type of operation and maintenance problem (network / hardware / configuration), the level of the anomaly (high / medium / low), and the direction of the handling solution (restart / check the network).

[0061] Step S137: Organize the key classification information according to the preset classification criteria structure to generate the initial classification criteria for each operation and maintenance problem text unit.

[0062] Key information is organized according to the structure of "problem type", "abnormality level", "related scenario characteristics", "possible cause" and "handling direction". For example, for "smart curtain APP is unresponsive and shows offline", the initial classification is "problem type: network connection; abnormality level: medium; related scenario characteristics: smart curtain - WiFi type, weak router signal; possible cause: WiFi interruption, gateway connection abnormality; handling direction: check the router, restart the curtain device".

[0063] Step S138: Before calling the pre-trained deep learning classification model, perform pre-adaptation processing on the cross-scene feature transfer module and the feature fusion module of the pre-trained deep learning classification model, and retrieve the cross-scene transfer parameters and feature fusion parameters used in the training process of the pre-trained deep learning classification model.

[0064] Before calling the model, cross-scene transfer parameters (feature alignment, adaptation coefficient weights, etc.) and feature fusion parameters (attention weights, fusion coefficients, etc.) are retrieved through the parameter management interface for pre-adaptation and adjustment to suit the current scene.

[0065] Step S1381: Input the scene feature data sample of the current operation and maintenance scenario into the cross-scene feature transfer module of the pre-trained deep learning classification model, test the matching degree between the transfer adaptation feature output by the cross-scene feature transfer module of the pre-trained deep learning classification model and the expected feature of the sample. If the matching degree does not reach the preset adaptation threshold, adjust the transfer coefficient and feature adjustment weight in the cross-scene transfer parameters.

[0066] Select various types of feature data from the current scenario as samples (including feature elements and their relationships). Experts set the expected features for the samples (element content, dimensions, and relationship strength). Input the samples into the cross-scenario feature transfer module, and calculate the matching degree between the output transfer-adapted features and the expected features (weighted sum of element overlap, dimension consistency, and relationship strength similarity). If the matching threshold is not reached, adjust the transfer coefficient for missing elements, adjust the feature weight for dimension mismatch, and correct the relationship coefficient for relationship deviation. Repeat the test until the target is met.

[0067] For example, step S13811: Select different types of scenario feature data under the current operation and maintenance scenario as scenario feature data samples. Each type of scenario feature data sample contains multiple feature elements and corresponding relationship descriptions.

[0068] Samples are selected from device feature categories (such as smart door locks and lights), network feature categories (routers and WiFi), control scenario feature categories (lighting linkage and security control), and user habit feature categories (common times when lights are turned on). Each sample contains multiple feature elements and relationships (such as the association between "smart door lock - fingerprint type" and "gateway - WiFi type").

[0069] Step S13812: Set expected features for each scene feature data sample. The expected features are determined based on domain expert experience and the training objective of the pre-trained deep learning classification model, and include the expected feature element content, feature dimension and feature correlation strength.

[0070] Domain experts, in conjunction with the current scenario and model training objectives, set expected features for each sample: clearly defining the feature elements to be included, a unified feature dimension, and a reasonable correlation strength (e.g., the correlation strength between "intelligent lighting - LED type" and "router dual-band type" is medium).

[0071] Step S13813: Input the scene feature data sample into the cross-scene feature transfer module of the pre-trained deep learning classification model. The cross-scene feature transfer module of the pre-trained deep learning classification model performs feature transfer processing on the sample based on the current cross-scene transfer parameters and outputs the transfer adaptation features.

[0072] Input the sample into the cross-scene feature transfer module. The module completes feature alignment, adaptation coefficient calculation and feature adjustment based on the current transfer parameters, and outputs the transfer and adaptation features.

[0073] Step S13814: Calculate the overlap of feature elements, consistency of feature dimensions, and similarity of feature association strength between the transfer adaptation features and the expected features of the samples, and determine the matching degree based on the standardized weighted sum of the three.

[0074] The matching degree is obtained by statistically analyzing the percentage of identical elements between the adapted and expected features (element overlap), comparing the consistency of feature dimensions (dimensional consistency), and calculating the cosine similarity of the association strength (association strength similarity). The matching degree is then obtained by standardizing the three indicators and then weighting and summing them.

[0075] Step S13815: If the calculated matching degree does not reach the preset adaptation threshold, analyze the differences between the migration adaptation features and the expected features of the sample. The differences include missing feature elements, feature items with mismatched dimensions, and feature relationships with large deviations in association strength.

[0076] When the adaptation threshold is not reached, compare the migration adaptation features with the expected features to identify the differences: such as the missing "router signal strength" element, the feature dimension of 128 dimensions not matching the expected 256 dimensions, and the weak association strength between "smart curtains" and "gateway" deviating greatly from the expected strong association strength.

[0077] Step S13816: For missing feature elements, adjust the transfer coefficient in the cross-scene transfer parameters to increase the transfer weight of historical training scene feature data related to the missing feature elements. Specifically, in the cross-scene transfer parameters, increase the transfer coefficient value for historical feature entries that are semantically related to the missing feature elements. This allows the cross-scene feature transfer module to refer more to these relevant historical features when processing current scene feature data samples, thereby including the missing feature element in the output transfer adaptation features.

[0078] Step S13817: For feature items with mismatched dimensions, adjust the feature adjustment weights in the cross-scene migration parameters to increase the weight ratio of the target dimension during the dimension transformation process.

[0079] When there is a dimensionality mismatch between the transferred adaptive features and the expected features of the samples, the feature adjustment weight configuration of the cross-scene transfer parameters identifies the parameter items related to dimensionality transformation and increases the weight ratio of the target dimension (i.e., the dimension of the expected features of the samples) in the dimensionality transformation calculation. Simultaneously, the mapping relationship parameters in the dimensionality transformation algorithm are adjusted so that the current scene feature data samples, after dimensionality transformation, can adapt to the dimensionality requirements of the expected features of the samples, eliminating the differences caused by dimensionality mismatch.

[0080] Step S13818: For feature relationships with large deviations in association strength, adjust the association migration coefficient in the cross-scene migration parameters and correct the calculation rules for feature association strength.

[0081] For feature pairs with significant deviations in association strength, the corresponding association migration coefficient configuration is located in the cross-scenario migration parameters, and the coefficient value is adjusted according to the direction of deviation. If the association strength in the migration adaptation feature is lower than expected, the association migration coefficient is increased; if it is higher than expected, the coefficient is decreased. Simultaneously, the calculation rules for feature association strength are modified. For example, when calculating association strength, the weight of business operation log data related to the feature relationship in the current scenario is increased, making the calculation results closer to the expected association strength requirements of the sample features.

[0082] Step S13819: After the adjustment is completed, re-input the scene feature data samples into the cross-scene feature transfer module of the pre-trained deep learning classification model, calculate the new matching degree, and repeat the adjustment and testing steps until the matching degree reaches the preset adaptation threshold.

[0083] After adjusting the parameters as described above, the same batch of scene feature data samples are input into the cross-scene feature transfer module again. The module outputs new transfer adaptation features based on the adjusted cross-scene transfer parameters. Following the calculation method in step S13814, the matching degree between the new transfer adaptation features and the expected features of the samples is calculated and compared with a preset adaptation threshold. If the threshold is still not met, the difference point analysis, parameter adjustment, and matching degree calculation steps are repeated until the matching degree meets the preset requirements.

[0084] Step S1382: Input the text semantic feature data samples and transfer adaptation feature samples into the feature fusion module of the pre-trained deep learning classification model, test the matching degree between the fused feature data output by the feature fusion module of the pre-trained deep learning classification model and the expected fusion features. If the matching degree does not reach the preset fusion threshold, adjust the fusion weight allocation rule and the association algorithm parameters in the feature fusion parameters.

[0085] Select text semantic feature data samples from smart home operation and maintenance scenarios (such as the core expression vector and abnormal phenomenon description vector corresponding to "smart lights not responding") and transfer adaptation feature samples after cross-scenario feature transfer processing (such as the feature vectors corresponding to "smart lights - LED type" and "weak router signal"), and input the two types of samples into the feature fusion module.

[0086] Domain experts define the expected features for the combined samples based on the feature fusion objectives, clarifying the scenario and semantic information, feature dimensions, and feature relationships that should be included after fusion. The feature fusion module processes the input samples based on the current feature fusion parameters and outputs fused feature data.

[0087] The matching degree between the fused feature data and the expected fused features is calculated. This matching degree comprehensively considers three indicators: the information completeness of the fused features (whether it contains the expected scene and semantic information), the dimensional consistency (whether it is consistent with the dimensions of the expected fused features), and the reasonableness of feature association (whether the relationship between features meets expectations). It is obtained by standardized weighted summation.

[0088] If the matching degree does not reach the preset fusion threshold, analyze the differences between the two, such as missing semantic information in the fused features, dimensions not matching expectations, and unreasonable feature association logic. To address the missing information issue, adjust the fusion weight allocation rules to increase the fusion weight of feature elements corresponding to the missing information; to address the dimension mismatch issue, adjust the dimension unification parameter in the feature fusion parameters and correct the dimension transformation logic; to address the unreasonable association logic issue, adjust the association algorithm parameters and optimize the calculation method of the association relationship between features.

[0089] Step S1383: Apply the adjusted feature fusion parameters to the feature fusion module of the pre-trained deep learning classification model, re-input text semantic feature data samples and transfer adaptation feature samples for testing, until the matching degree between the fused feature data and the expected fused features reaches the preset fusion threshold.

[0090] Update the adjusted feature fusion parameters in the feature fusion module's parameter configuration, and input the text semantic feature data samples and transfer adaptation feature samples again. The module outputs new fused feature data. Obtain the new matching degree using the same calculation method. If the preset fusion threshold is still not reached, repeat the difference point analysis, parameter adjustment, and testing steps until the fused feature data meets the requirements of the expected fusion features.

[0091] Step S1384: After completing the pre-adaptation process, save the adjusted cross-scene transfer parameters and feature fusion parameters, and use them as the running parameters of the currently called pre-trained deep learning classification model before executing the subsequent feature transfer adaptation and association fusion steps.

[0092] Once the output matching degree of both the cross-scenario feature transfer module and the feature fusion module reaches a preset threshold, the adjusted cross-scenario transfer parameters and feature fusion parameters are saved through the model parameter storage interface and marked as the model running parameters for the current smart home operation and maintenance scenario. Subsequently, when processing text units of actual operation and maintenance problems, the pre-trained deep learning classification model uses this set of running parameters to perform feature transfer adaptation and association fusion operations.

[0093] Step S140: Based on the historical operation and maintenance problem classification results, the corresponding operation and maintenance processing effect data, and the scenario feature evolution trajectory data, adjust the feature weights and feature association rules in the initial classification criteria in stages to obtain the optimized classification criteria.

[0094] Retrieve historical operation and maintenance data in the smart home operation and maintenance scenario, including the classification results of historical operation and maintenance issues, the corresponding processing effects, and the historical scenario feature evolution trajectory. Through a phased iterative approach, adjust and optimize the feature weights and feature association rules in the initial classification criteria generated in step S137, so that the classification criteria are more in line with the needs of the actual operation and maintenance scenario and improve the classification accuracy.

[0095] Step S141: Retrieve the historical operation and maintenance problem classification result dataset, the corresponding operation and maintenance processing effect dataset, and the historical scene feature evolution trajectory data. The historical operation and maintenance problem classification result dataset contains historical operation and maintenance problem text units, corresponding historical classification results, and classification basis. The operation and maintenance processing effect dataset contains the operation and maintenance processing schemes corresponding to the historical classification results, the processed business recovery status data, and the processing efficiency data. The historical scene feature evolution trajectory data is a record of the scene feature evolution trajectory within a historical time period.

[0096] By using the query interface of the operation and maintenance data management platform, the historical operation and maintenance problem classification result dataset can be retrieved. This historical operation and maintenance problem classification result dataset is stored in chronological order. Each record contains a historical operation and maintenance problem text unit (such as "smart socket cannot connect to the network"), the corresponding historical classification result (such as "network connection type - device network failure"), and the classification basis used at that time.

[0097] At the same time, the operation and maintenance processing effect dataset is retrieved. This operation and maintenance processing effect dataset is associated with the historical operation and maintenance problem classification result dataset through a unique identifier. Each record contains the operation and maintenance processing solution adopted for the historical classification results (such as "restarting the router and smart socket"), the business recovery status data after processing (such as "the device has restored network access and control is normal"), and the processing efficiency data (such as "processing time" and "resource consumption").

[0098] In addition, historical scene feature evolution trajectory data is retrieved. This historical scene feature evolution trajectory data is a record of the scene feature evolution trajectory over a period of time (such as the past year), including information on scene feature changes and the scope of impact at each historical time node.

[0099] Step S142: Associate the historical operation and maintenance problem classification result dataset with the operation and maintenance processing effect dataset, establish the correspondence between each historical classification result and the corresponding business recovery status data and processing efficiency data, and generate the classification effect association dataset.

[0100] Based on unique identifiers (such as maintenance work order numbers) in the historical maintenance problem classification result dataset and the maintenance processing effect dataset, the two datasets are associated and matched so that each historical classification result corresponds to its processed business recovery status data and processing efficiency data. For example, the historical classification result "Network connectivity - Device network failure" corresponds to the business recovery status data "Device network recovery" and the processing efficiency data "Short processing time".

[0101] The associated datasets are grouped according to the types of historical classification results to form a classification effect associated dataset. Each group of data corresponds to a type of historical classification result, which facilitates subsequent analysis of the processing effect of different classification results.

[0102] Step S143: Extract the effect features corresponding to the business recovery status data and the efficiency features corresponding to the processing efficiency data in the classification effect association dataset. The effect features include a description of business recovery integrity, a description of recovery time, and a description of post-recovery stability. The efficiency features include a description of the number of processing steps, a description of processing resource consumption, and a description of processing response speed.

[0103] The following performance features are extracted from the business recovery status data in the classification performance association dataset: Business recovery integrity description describes the degree of completeness of device or business recovery, such as "complete recovery" or "partial function recovery"; recovery time description describes the time from the start of processing to normal recovery, such as "short-time recovery" or "long-time recovery"; and post-recovery stability description describes the operational stability after recovery, such as "stable operation without abnormalities" or "frequent recurrence".

[0104] Efficiency features are extracted from processing efficiency data: the number of processing steps describes the number of steps in the processing process, such as "simple steps" or "complex steps"; the resource consumption description describes the resources consumed in the processing process, such as "low resource consumption" or "high resource consumption"; and the response speed description describes how fast the response is from receiving the question to starting processing, such as "rapid response" or "slow response".

[0105] Step S144: Input the classification criteria, corresponding effect features, efficiency features, and historical scene feature evolution trajectory data from the historical classification results into the iterative optimization module to analyze the correlation strength between the feature elements in the classification criteria and the effect features and efficiency features, as well as the changing pattern of feature elements with scene feature evolution.

[0106] The classification criteria (including feature elements and feature association rules), corresponding effect features, efficiency features, and historical scenario feature evolution trajectory data from the historical classification results are input into the iterative optimization module. The iterative optimization module first uses an association analysis algorithm to calculate the association strength between each feature element in the classification criteria and the effect features and efficiency features. For example, it analyzes the closeness of the association between the feature element "weak router signal" and the effect feature "long service recovery time".

[0107] Meanwhile, by combining historical scene feature evolution trajectory data, we tracked the frequency of occurrence, changes in expression, and changes in correlation of each feature element in the classification criteria at different historical scene evolution stages, and summarized the change pattern of feature elements with scene evolution. For example, the feature element "smart door lock fingerprint recognition failure" appeared significantly more frequently during the router upgrade stage.

[0108] Step S145: Determine the initial adjustment direction of each feature element in the classification criteria based on the correlation strength, determine the dynamic adjustment range of the feature element weights based on the change pattern, and generate the first round of feature weight adjustment scheme.

[0109] The initial adjustment direction of each feature element is determined based on the correlation strength between the feature elements and the effect and efficiency features. If a feature element has a high correlation strength with positive features such as "complete business recovery" and "high processing efficiency", the initial adjustment direction of the feature element's weight is to increase; if it has a high correlation strength with negative features such as "incomplete business recovery" and "low processing efficiency", the initial adjustment direction is to decrease.

[0110] Based on the changing patterns of feature elements as the scene evolves, the dynamic adjustment range of the weights is determined. If the frequency of a feature element's occurrence in the current scene evolution stage shows an upward trend, and the associated positive features are obvious, then the adjustment range is set to be larger; if the frequency of occurrence is stable or decreasing, then the adjustment range is set to be smaller. Combining the initial adjustment direction and adjustment range, a first-round feature weight adjustment scheme is generated, clarifying the weight adjustment value and adjustment method for each feature element.

[0111] Step S146: Apply the first round of feature weight adjustment scheme to the initial classification basis to obtain the first round of optimized classification basis. Input the first round of optimized classification basis and the corresponding historical operation and maintenance problem text units into the pre-trained deep learning classification model to generate the first round of simulated classification results.

[0112] Following the first round of feature weight adjustment scheme, the weight values ​​of each feature element in the initial classification criteria are modified to form the first round of optimized classification criteria. Historical maintenance problem text units corresponding to this classification criteria are selected, and both are input into a pre-trained deep learning classification model (using the running parameters saved in step S1384). The model classifies the historical maintenance problem text units based on the optimized classification criteria and outputs the first round of simulated classification results.

[0113] Step S147: Compare the first round of simulated classification results with the actual classification results in the historical classification results, calculate the classification accuracy, classification recall and classification matching degree, and generate the first round of optimization evaluation data.

[0114] The simulation results from the first round of classification are compared one by one with the actual classification results from the historical classification results. The proportion of correctly classified results out of the total number of results is used to obtain the classification accuracy. The proportion of correctly classified results out of the total number of results in a certain category is used to obtain the classification recall. The degree of matching between feature elements and association rules in the simulation results and the actual classification results is used to obtain the classification matching degree. These three indicators are integrated into the first round of optimization evaluation data.

[0115] Step S148: If the first round of optimization evaluation data does not reach the preset optimization target, then adjust the parameters of the feature weight adjustment scheme based on the first round of optimization evaluation data, including the adjustment magnitude coefficient and adjustment direction, generate the second round of feature weight adjustment scheme, and repeat the weight adjustment, simulated classification and evaluation steps.

[0116] The first round of optimization evaluation data is compared with the preset optimization targets (such as classification accuracy ≥ preset value, classification recall ≥ preset value, and classification matching degree ≥ preset value). If the target is not achieved, the reasons for the failure are analyzed. For example, if the classification recall of a certain category is low, it indicates that the weight of the feature elements corresponding to that category may be insufficient.

[0117] Based on the analysis results, the parameters of the feature weight adjustment scheme are adjusted: if it is necessary to further enhance the influence of a certain feature element, its adjustment magnitude coefficient is increased; if it is found that the initial adjustment direction is incorrect (such as the classification effect decreasing after the weight of a certain feature element is increased), the adjustment direction is reversed. A second-round feature weight adjustment scheme is then generated, and steps S146 to S147 are repeated to obtain the second-round optimization evaluation data.

[0118] Step S149: During multiple iterations, the feature association rules in the classification criteria are adjusted synchronously based on the scene feature evolution trajectory data, including feature element combination rules, feature priority ranking rules, and feature conflict resolution rules.

[0119] While adjusting the feature weights in each round, the feature association rules in the classification criteria are adjusted in conjunction with the scene feature evolution trajectory data to ensure that the association rules are adapted to the current scene feature evolution state.

[0120] Step S1491: Divide the scene feature evolution trajectory data into evolution stages. Based on the change magnitude and frequency of scene features, divide the evolution process into multiple evolution stages. Each evolution stage corresponds to a continuous time interval and the scene feature state within that time interval.

[0121] Set thresholds for the magnitude and frequency of scene feature changes, and analyze the scene feature evolution trajectory data. When the magnitude or frequency of scene feature changes exceeds the threshold within a certain continuous time interval, that time interval is divided into an independent evolution stage. Each evolution stage corresponds to a clear start and end point and a stable scene feature state within that stage, such as "router upgrade stage" or "new device access stage".

[0122] Step S1492: Extract the core scene feature elements within each evolutionary stage. The core scene feature elements are the scene feature elements that appear most frequently and have the most associated features within that evolutionary stage.

[0123] The scene feature elements within each evolutionary stage are statistically analyzed, and the frequency of occurrence and the number of associated features for each element are calculated. The feature element with the highest frequency of occurrence and the highest number of associated features is selected as the core scene feature element for that stage. For example, in the "router upgrade stage," the occurrence frequency and the number of associated features for "router model change" and "device network connection abnormality" are the highest, and therefore, they are the core scene feature elements for that stage.

[0124] Step S1493: Analyze the combination of core scene feature elements with other scene feature elements in each evolutionary stage, summarize the typical combination patterns of scene feature elements in different evolutionary stages, and adjust the feature element combination rules in the classification criteria based on the typical combination patterns so that the feature element combination rules match the scene feature combination patterns of the current evolutionary stage.

[0125] For each evolutionary stage, the combinations of core scenario feature elements with other scenario feature elements are statistically analyzed to summarize typical combination patterns for that stage, such as "router model change + smart device network connection anomaly" and "new device access + linkage rule failure." Based on these typical combination patterns, the feature element combination rules in the classification criteria are adjusted. For example, in the "router upgrade stage," the combination rule of "router model change" and "device network connection anomaly" is given higher priority to ensure that the classification criteria can accurately capture the problem characteristics of that stage.

[0126] Step S1494: Based on the changes in the importance of core scene feature elements in each evolutionary stage, adjust the feature priority ranking rules in the classification criteria, increase the priority of the classification criteria feature elements corresponding to the core scene feature elements in the current evolutionary stage, and decrease the priority of the classification criteria feature elements corresponding to the non-core scene feature elements.

[0127] Based on the changing importance of core scenario feature elements at different evolution stages (judged comprehensively by frequency of occurrence, number of associated features, and impact on operational issues), the priority ranking rules for feature elements in the classification criteria are adjusted. At the current evolution stage, the priority of classification criteria feature elements corresponding to core scenario feature elements (such as feature elements corresponding to "router model change") is increased to the forefront, while the priority of classification criteria feature elements corresponding to non-core scenario feature elements is decreased, ensuring that key features in the current scenario are considered first during classification.

[0128] Step S1495: Identify the conflict types between scene feature elements in each evolution stage, including expression conflicts, relationship conflicts, and scope of influence conflicts. Formulate corresponding conflict resolution strategies based on the conflict types, such as using a semantic fusion strategy for expression conflicts and a primary-secondary association strategy for relationship conflicts.

[0129] At each evolution stage, the conflicts between scene feature elements are analyzed: Representation conflict refers to inconsistent descriptions of the same phenomenon by different feature elements, such as "device offline" and "device not connected to the network"; Association conflict refers to contradictory association logic between feature elements, such as "feature A is associated with feature B" and "feature A is not associated with feature B" existing simultaneously; Influence range conflict refers to overlapping or contradictory descriptions of the influence range of feature elements.

[0130] Strategies are developed for different conflict types: for expression conflicts, a semantic fusion strategy is adopted to integrate different expressions into a unified standard expression; for relationship conflicts, a primary-secondary relationship strategy is adopted, based on the relationship between core scene feature elements to determine the primary and secondary relationships; for impact scope conflicts, a scope division strategy is adopted to clarify the impact boundaries of each feature element.

[0131] Step S1496: Integrate the conflict resolution strategy into the feature conflict resolution rules in the classification criteria, replacing the original conflict resolution rule content.

[0132] The established conflict resolution strategy is updated in the feature conflict resolution rules of the classification criteria, replacing the original rules that are inconsistent with the current evolution stage. For example, in the "new device access stage," if a conflict of expression occurs, a semantic fusion strategy is used to handle it, ensuring that the classification criteria can resolve feature conflicts according to rules adapted to the current scenario.

[0133] Step S1497: In each round of iterative optimization, based on the current scene feature evolution stage, call the corresponding feature element combination rules, feature priority ranking rules, and feature conflict resolution rules, and apply them to the classification basis adjustment process.

[0134] Before each round of feature weight iteration optimization, the current scene feature evolution stage is determined. The feature element combination rules, feature priority ranking rules, and feature conflict resolution rules corresponding to this stage are called from the adjusted feature association rule library and applied to the adjustment process of the classification basis in this round, so that the association rules of the classification basis are synchronized with the current scene evolution state.

[0135] Step S1410: When the evaluation data of any round of optimization reaches the preset optimization target, stop the iteration and determine the classification basis obtained in that round as the optimized classification basis.

[0136] After each round of optimization evaluation, if the classification accuracy, classification recall, and classification matching degree in the optimization evaluation data all reach the preset optimization target, the iteration process is stopped, and the classification criteria obtained in this round are determined as the optimized classification criteria for subsequent operation and maintenance problem classification and matching operations.

[0137] Step S150: Perform multi-level association matching between the optimized classification criteria and the dynamically updated O&M technical service problem classification case library to generate the O&M technical service problem classification result corresponding to each O&M problem text unit.

[0138] In this embodiment, the dynamically updated operation and maintenance technical service problem classification case library is the core case resource for smart home control operation and maintenance services, containing both historically accumulated and newly added operation and maintenance problem cases. By performing multi-level matching between the optimized classification criteria and the case library, the most suitable case classification result is accurately located, ensuring the accuracy and practicality of the operation and maintenance problem classification.

[0139] Step S151: Retrieve the dynamically updated O&M technical service problem classification case library. The dynamically updated O&M technical service problem classification case library includes historical O&M problem case data and the latest O&M problem case data. The historical O&M problem case data includes historical case problem text, historical case scenario features, historical case classification results, and historical case classification basis. The latest O&M problem case data is the O&M problem case data added in the recent period, including the latest case problem text, the latest case scenario features, the latest case classification results, and the latest case classification basis.

[0140] The system accesses a dynamically updated O&M technical service issue classification case library via a case library access interface. This case library employs a distributed storage architecture, divided into "Historical Cases" and "Latest Cases" storage areas. Historical O&M issue case data covers O&M cases from a relatively long period. Each record fully documents the historical case issue text (e.g., "Smart lighting linkage failure in Month X, 2024, triggered by human body detection"), historical case scenario characteristics (e.g., "Router is dual-band, linkage rule is 'human body detection triggered -- lights on'"), historical case classification results (e.g., "Control scenario configuration - linkage rule failure"), and historical case classification criteria (including feature weights, association rules, and other detailed information). The latest O&M issue case data comprises cases added within the last month, with a data structure consistent with historical cases, ensuring the case library reflects the latest O&M issue characteristics and classification logic.

[0141] Step S152: Perform feature layering processing on the optimized classification criteria. Divide the key classification information in the optimized classification criteria into a core feature layer, a secondary feature layer, and an auxiliary feature layer according to the correlation dimension. The core feature layer contains feature content directly related to the type of operation and maintenance problem. The secondary feature layer contains feature content related to the level of operation and maintenance problem. The auxiliary feature layer contains feature content related to the operation and maintenance problem handling solution.

[0142] Based on the optimized classification criteria, fields such as "problem type," "anomaly level," and "handling direction," feature layers are established. The core feature layer focuses on features directly related to the type of operational problem, such as "WiFi communication anomaly" and "gateway offline" for "network connectivity" problems. These features are the core basis for determining the problem category. The secondary feature layer revolves around the operational problem level, such as "affecting a single device" and "can be attempted to be repaired remotely" for "medium-level anomaly," used to define the severity of the problem. The auxiliary feature layer associates with the direction of the operational handling plan, such as "router model" and "network signal strength" for "checking the router." Each layer of features is stored in vector form, retaining the feature weights and correlation information from the optimized classification criteria.

[0143] Step S153: Perform first-level matching between the feature content of the core feature layer and the core features in the case classification criteria of all cases in the dynamically updated operation and maintenance technical service problem classification case library, calculate the core feature similarity, and select a set of candidate cases whose core feature similarity reaches the first preset threshold.

[0144] The cosine similarity algorithm is used to calculate the similarity between the optimized classification criteria core feature layer and the core features of each case in the case library. Before calculation, the core feature vectors of both are converted to a unified dimension to ensure consistency of the vector space. For example, if the optimized classification criteria core feature vector is "[WiFi communication abnormal, gateway offline, device offline]", and a case's core feature vector is "[WiFi communication abnormal, gateway offline, smart socket offline]", the core feature similarity is obtained by calculating the cosine of the angle between the two vectors.

[0145] The calculated core feature similarity is compared with a first preset threshold (e.g., 0.7), and cases whose similarity reaches the threshold are selected and integrated to form a candidate case set. If the core features of a case highly overlap with the core features of the optimized classification criteria, such as both containing the features of "WiFi communication abnormality" and "gateway offline", then its similarity is likely to reach the threshold and enter the candidate set.

[0146] Step S154: Perform a second-level matching between the feature content of the secondary feature layer and the secondary features in the case classification criteria of each case in the candidate case set, calculate the secondary feature similarity, and select the secondary candidate case set whose secondary feature similarity reaches the second preset threshold.

[0147] For each case in the candidate case set, secondary features are extracted from its case classification criteria. The same cosine similarity algorithm used for matching core features is employed to calculate the similarity between the optimized classification criteria secondary feature layer and the case's secondary features. For example, if the optimized classification criteria secondary feature is "[Medium anomaly, affecting a single smart curtain, remote repair attempt possible]", and a candidate case's secondary feature is "[Medium anomaly, affecting a single smart light, remote repair attempt possible]", the secondary feature similarity is obtained through vector calculation.

[0148] A second preset threshold (e.g., 0.6) is set to filter out cases whose secondary feature similarity reaches the threshold, forming a secondary candidate case set. This hierarchical matching can eliminate cases with similar core features but significantly different anomaly levels and impact ranges, further narrowing the matching range.

[0149] Step S155: Perform third-level matching between the feature content of the auxiliary feature layer and the auxiliary features in the case classification criteria of each case in the secondary candidate case set, calculate the auxiliary feature similarity, and generate the comprehensive similarity of each case. The comprehensive similarity is the weighted sum of the core feature similarity, secondary feature similarity, and auxiliary feature similarity.

[0150] Extract auxiliary features from each case in the secondary candidate case set, and calculate the similarity between the optimized classification criteria auxiliary feature layer and the case auxiliary features. For example, if the optimized classification criteria auxiliary features are "[Check router, restart device, signal strength detection]", and a secondary candidate case auxiliary features are "[Check router, restart device, WiFi channel adjustment]", the similarity of the auxiliary features is obtained through vector calculation.

[0151] Weights are assigned to the core feature similarity, secondary feature similarity, and auxiliary feature similarity, with the core feature similarity having the highest weight (e.g., 0.5), secondary feature similarity having the next highest weight (e.g., 0.3), and auxiliary feature similarity having the lowest weight (e.g., 0.2). The three similarities are multiplied by their respective weights and then summed to obtain the overall similarity for each case. For example, if a case has a core similarity of 0.8, a secondary similarity of 0.7, and an auxiliary similarity of 0.6, then the overall similarity is 0.8 × 0.5 + 0.7 × 0.3 + 0.6 × 0.2 = 0.73.

[0152] Step S156: Sort the cases in the secondary candidate case set based on the comprehensive similarity, and select the case with the highest comprehensive similarity as the optimal matching case.

[0153] The cases in the secondary candidate case set are sorted from highest to lowest overall similarity. If multiple cases have the same highest overall similarity, the case's scenario features are further compared with the current operation and maintenance scenario features, and the case with the higher similarity is selected as the optimal matching case. For example, if two cases both have an overall similarity of 0.85, and one case's scenario feature includes "the router model is consistent with the current scenario," then that case is selected as the optimal matching case.

[0154] Step S157: Extract the case classification result corresponding to the best matching case, and use the case classification result as the preliminary classification result corresponding to the current operation and maintenance problem text unit.

[0155] Extract the case classification result from the metadata of the best matching case. This case classification result includes the specific category, sub-type, and associated operation and maintenance tags of the problem. For example, if the classification result of the best matching case is "Network connectivity - WiFi communication failure - caused by weak router signal", then this result is used as the preliminary classification result of the current operation and maintenance problem text unit (such as "Smart curtain APP is unresponsive and shows offline").

[0156] Step S158: Retrieve the case scenario features and case problem text corresponding to the best matching case, and compare them with the current operation and maintenance problem text unit and the corresponding scenario feature data to confirm the scenario adaptability and consistency of text description.

[0157] The case scenario features of the best matching case (e.g., "Smart curtains - WiFi type, router is dual-band A type") and the case problem text (e.g., "Smart curtains are not responding when controlled by APP, APP shows WiFi connection failure") are retrieved and compared with the current operation and maintenance problem text unit and the corresponding scenario feature data (e.g., the current smart curtains are WiFi type, router is dual-band A type, problem text is "Smart curtain APP is not responding, showing offline"), and then compared dimension by dimension.

[0158] Scene compatibility comparison focuses on whether core scene elements such as device type, network environment, and control scenario are consistent; textual expression consistency comparison focuses on the semantic similarity of the description of the problem phenomenon, such as the semantic correlation between "WiFi connection failed" and "displays offline". The comparison confirms whether there are significant differences between the two, such as mismatched device types or significant differences in network environment.

[0159] Step S159: If the scenario adaptability and text expression consistency both meet the preset requirements, the preliminary classification result is determined as the classification result of the operation and maintenance technical service problem corresponding to each operation and maintenance problem text unit; if not, the case with the second highest comprehensive similarity is selected and the supplementary comparison step is repeated until the operation and maintenance technical service problem classification result that meets the requirements is determined.

[0160] Set a scenario adaptability threshold (e.g., core scenario element matching degree ≥ 0.9) and a text expression consistency threshold (e.g., semantic similarity ≥ 0.6). If the comparison results of the best matching cases all reach the threshold, the preliminary classification result is directly determined as the final classification result of the operation and maintenance technical service problem.

[0161] If the threshold is not reached, such as when the router model of the best matching case differs significantly from the current scenario and the scenario adaptability does not meet the requirements, the case is removed, and the case with the second highest overall similarity is selected as a new candidate case. The extraction and comparison operations in steps S157 to S158 are repeated. If no case that meets the requirements is found after traversing the set of secondary candidate cases, the "problem type" in the optimized classification criteria is used as the basic classification result and marked with the "awaiting manual review" label. The classification result will then be further confirmed by the operations and maintenance personnel based on the actual situation.

[0162] Step S160: Retrieve the dynamic update triggering mechanism of the dynamically updated operation and maintenance technical service problem classification case library. The dynamic update triggering mechanism of the dynamically updated operation and maintenance technical service problem classification case library triggers the update operation based on the number of newly added operation and maintenance problem classification results or the time interval. When the number of newly added operation and maintenance problem classification results reaches a preset number threshold, or the time interval since the last update reaches a preset time threshold, the dynamic update of the operation and maintenance technical service problem classification case library is triggered.

[0163] The dynamic update triggering mechanism is built into the case library management system and includes two modes: quantity-triggered and time-triggered. In quantity-triggered mode, a preset threshold for the number of newly added operation and maintenance issue category results is set (e.g., 50). The case library management system counts the number of newly added category results in real time, and automatically triggers an update when the threshold is reached. In time-triggered mode, a preset update interval is set (e.g., 7 days). If the interval has elapsed since the last update and no quantity update has been triggered, the system automatically starts the update process. The two triggering modes complement each other, ensuring that the case library can be updated with new cases in a timely manner.

[0164] Step S161: After triggering the dynamic update operation of the operation and maintenance technical service problem classification case library, retrieve the operation and maintenance problem classification result data generated in the recent period. The operation and maintenance problem classification result data includes newly added operation and maintenance problem text units, corresponding classification results, classification basis and associated scenario feature data.

[0165] After the update is triggered, the case library management system retrieves all maintenance issue classification results data generated in the recent period (from the last update to the present) through the data interface. This maintenance issue classification result data is pushed to the data buffer in real time by the maintenance classification module, including newly added maintenance issue text units (such as "Smart door lock fingerprint recognition frequently fails, password unlocking is normal"), the corresponding final classification results (such as "Device hardware - fingerprint recognition module failure"), optimized classification criteria, and associated scenario feature data (such as "Smart door lock - fingerprint type, usage time 1 year, firmware version V2.0").

[0166] Step S162: Perform case format conversion processing on the newly added operation and maintenance problem classification result data, convert the newly added operation and maintenance problem text unit into case problem text format, convert the associated scenario feature data into case scenario feature format, convert the classification result into case classification result format, convert the classification basis into case classification basis format, and generate case data to be supplemented.

[0167] The case library management system has built-in format conversion templates to standardize newly added data. New maintenance issue text units are converted into case issue text format using the structure "problem phenomenon + occurrence time + device identifier"; associated scenario feature data is converted into case scenario feature format using the hierarchical structure "device features + network features + control scenario features"; classification results are converted into case classification result format using the three-level structure "main category - subcategory - tag"; and the classification criteria retain the original feature hierarchy, weight allocation, and other information, and are converted into a case classification criterion format. After conversion, a structured set of supplementary case data is generated.

[0168] Step S163: Perform case uniqueness verification on the generated case data to be supplemented, and compare the similarity between the case data to be supplemented and the case data already stored in the dynamically updated operation and maintenance technical service problem classification case library. If the similarity does not reach the case duplication threshold, the case data to be supplemented is marked as supplementable case data.

[0169] A semantic hash-based uniqueness verification method is used to calculate the similarity between the supplementary case data and the case data already stored in the case library. The text of the case question and the case scenario features of the case to be supplemented are extracted to generate a semantic hash value; simultaneously, the corresponding fields of the stored cases are extracted to generate semantic hash values, and the similarity between the two hash values ​​is calculated.

[0170] A preset case duplication threshold (e.g., 0.85) is set. If the similarity is below the threshold, it is considered a new case and marked as supplementary case data; if it is above the threshold, it is considered a duplicate case, discarded, and recorded in the duplicate log. For example, if the semantic similarity between the question text and scene features of the case to be supplemented and the already stored case is 0.7, it is marked as supplementary case data.

[0171] Step S164: The supplementable case data is classified into the corresponding case classification directory according to the storage structure of the dynamically updated operation and maintenance technical service problem classification case library. The case classification directory is divided into levels according to the operation and maintenance problem type, scenario category and classification time.

[0172] The case library uses a three-level classification directory structure: the first-level directory is divided by the type of operation and maintenance problem (such as network connection, device hardware, and control scenario configuration); the second-level directory is divided by scenario category under the first-level directory (such as WiFi failure, gateway abnormality, etc. under the network connection category); and the third-level directory is divided by classification time under the second-level directory (such as by month).

[0173] Based on the classification results and classification time of supplementary case data, they are categorized into the corresponding third-level directories. For example, supplementary cases with a classification result of "Network Connection - WiFi Failure" and a classification time of September 2025 are categorized into the "Network Connection - WiFi Failure - September 2025" directory and assigned a unique case ID for indexing.

[0174] Step S165: After supplementing the case data, update the index of all case data in the dynamically updated operation and maintenance technical service problem classification case library, and establish feature indexes for the newly added supplementable case data, including case problem text index, case scenario feature index and case classification result index.

[0175] The case library management system initiates an index update process, employing inverted index technology to build multi-dimensional indexes for newly added supplementary case data. The case issue text index is constructed based on keywords in the issue text; for example, keywords such as "smart door lock" and "fingerprint recognition failure" are associated with corresponding case IDs. The case scenario feature index is constructed according to feature categories; for example, features such as "device type - smart door lock" and "network type - WiFi" are associated with case IDs. The case classification result index is constructed hierarchically according to "main category - subcategory - tag," with each level of classification item associated with a corresponding case ID.

[0176] Simultaneously, the global index table of the case library is updated to ensure that newly added cases can be quickly queried through any index dimension. After the index update is complete, the system will perform index verification to ensure consistency between the index and the case data.

[0177] Step S166: After completing the index update, generate a dynamically updated operation and maintenance technical service problem classification case library update log, recording the update time, the number of cases added, and the case classification distribution, forming a dynamically updated operation and maintenance technical service problem classification case library containing historical operation and maintenance problem case data and the latest operation and maintenance problem case data.

[0178] The update log is generated in a structured format and includes basic update information (update start time, end time, update trigger method), supplementary case statistics (total number, number of cases for each issue type), case classification distribution (percentage of cases under each first-level and second-level directory), and exception records (such as the number of duplicate cases and the number of cases that failed to convert formats).

[0179] The update logs are stored in the log management module of the case library. At the same time, the supplementary case data is integrated with the original historical and latest case data to form an updated dynamic operation and maintenance technical service problem classification case library.

[0180] Figure 2 The diagram illustrates exemplary hardware and software components of a deep learning-based operations and maintenance technical service problem classification system 100, which can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the deep learning-based operations and maintenance technical service problem classification system 100 and to perform the functions in this application.

[0181] The deep learning-based operation and maintenance technical service problem classification system 100 can be a general-purpose server or a special-purpose server; both can be used to implement the deep learning-based operation and maintenance technical service problem classification method of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0182] For example, the deep learning-based operation and maintenance technical service problem classification system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the deep learning-based operation and maintenance technical service problem classification system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The deep learning-based operation and maintenance technical service problem classification system 100 also includes an I / O interface 150 between the computer and other input / output devices.

[0183] For ease of explanation, only one processor is described in the deep learning-based operation and maintenance technical service problem classification system 100. However, it should be noted that the deep learning-based operation and maintenance technical service problem classification system 100 of this application may also include multiple processors. Therefore, the steps executed by one processor described in this application may also be executed jointly by multiple processors or individually. For example, if the processor of the deep learning-based operation and maintenance technical service problem classification system 100 executes steps A and B, it should be understood that steps A and B may also be executed jointly by two different processors or individually by one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor jointly execute steps A and B.

[0184] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned deep learning-based operation and maintenance technical service problem classification method is implemented.

[0185] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for classifying operation and maintenance technical service problems based on deep learning, characterized in that, The method includes: Obtain a set of text describing operation and maintenance technical service issues. The set of text describing operation and maintenance technical service issues contains multiple text units of operation and maintenance issues, and each text unit of operation and maintenance issues is accompanied by collection time information. The system performs dynamic updates and feature evolution analysis of the operation and maintenance scenario feature library to generate scenario feature evolution trajectory data. The operation and maintenance scenario feature library stores scenario feature data corresponding to the operation and maintenance scenarios. Each operation and maintenance problem text unit, scenario feature data, and scenario feature evolution trajectory data are input into a pre-trained deep learning classification model to obtain the initial classification basis for each operation and maintenance problem text unit. Based on historical operation and maintenance problem classification results, corresponding operation and maintenance processing effect data, and scenario feature evolution trajectory data, the feature weights and feature association rules in the initial classification criteria are adjusted in stages to obtain the optimized classification criteria. The optimized classification criteria are matched with the dynamically updated O&M technical service problem classification case library at multiple levels to generate the O&M technical service problem classification result corresponding to each O&M problem text unit.

2. The deep learning-based method for classifying operational and maintenance technical service issues according to claim 1, characterized in that, The dynamic updating and feature evolution analysis of the operation and maintenance scenario feature library generates scenario feature evolution trajectory data, including: Retrieve the business operation log data and historical scenario feature data of the current operation and maintenance business scenario. The business operation log data includes device operation status description data, business process execution description data and abnormal event record data under the current operation and maintenance business scenario. The historical scenario feature data is the scenario feature data of the recent period stored in the operation and maintenance scenario feature library. Extract scenario feature elements from business operation log data. The scenario feature elements are descriptions that can characterize the characteristics of the current operation and maintenance business scenario, including device type association descriptions, business process stage descriptions, and abnormal event type descriptions. The scene feature elements are compared with the scene feature data already stored in the operation and maintenance scene feature library. Scene feature elements whose similarity to the description of the stored scene feature data does not reach the preset correlation level are selected. These scene feature elements are added as new scene feature data and are classified into the corresponding scene category according to the storage structure of the operation and maintenance scene feature library. Historical scene feature data and newly added scene feature data are arranged in chronological order of collection time to construct a scene feature time series dataset, which includes scene feature data at different time points and corresponding time markers; Based on the scene feature time series dataset, the feature evolution analysis algorithm is used to identify the changing trend of each scene feature data in the time dimension, including the changing trend of feature occurrence frequency, the changing trend of feature description details, and the changing trend of feature correlation. Based on the identified trends, a feature evolution path is constructed for each scene feature data. The evolution paths of all scene feature data are integrated to generate scene feature evolution trajectory data. The scene feature evolution trajectory data includes the time change nodes of the scene features, the description of the change content, and the description of the scope of the change's impact.

3. The deep learning-based method for classifying operational and maintenance technical service issues according to claim 1, characterized in that, The process involves inputting each operation and maintenance problem text unit, scenario feature data, and scenario feature evolution trajectory data into a pre-trained deep learning classification model to obtain the initial classification criteria for each operation and maintenance problem text unit, including: The text unit of operation and maintenance problem is input into the text encoding module of the pre-trained deep learning classification model. The text unit of operation and maintenance problem is semantically parsed and vector transformed to generate text semantic feature data corresponding to the text unit of operation and maintenance problem. The text semantic feature data includes the core expression vector of operation and maintenance problem, the abnormal phenomenon description vector and the related device expression vector. The scene feature data and scene feature evolution trajectory data are input into the cross-scene feature transfer module of the pre-trained deep learning classification model. After performing feature alignment processing on the scene feature data and scene feature evolution trajectory data, the transfer adaptation coefficient between the scene feature data and the historical training scene feature data is calculated. Based on the transfer adaptation coefficient, the scene feature data is adjusted to generate transfer adaptation features that are adapted to the distribution of training data for the pre-trained deep learning classification model. The transfer adaptation features and text semantic feature data are input into the feature fusion module of the pre-trained deep learning classification model to establish the relationship between the transfer adaptation features and text semantic feature data, including the correspondence of feature elements, the influence of feature weights, and the complementary relationship of feature semantics. Based on the established relationships, the migration adaptation features and text semantic features are weighted and fused to generate fused feature data containing scene migration information and text semantic information. The fused feature data is input into the classification basis generation module of the pre-trained deep learning classification model. Feature extraction and rule matching are performed on the fused feature data to extract key classification information from the fused feature data, including operation and maintenance problem type association features, anomaly level corresponding features, and handling scheme matching features. The key classification information is organized according to the preset classification criteria structure to generate the initial classification criteria for each operation and maintenance issue text unit.

4. The deep learning-based method for classifying operational and maintenance technical service issues according to claim 3, characterized in that, The calculation of the transfer adaptation coefficient between the scene feature data and the historical training scene feature data includes: Retrieve the historical training scene feature data of the pre-trained deep learning classification model. The historical training scene feature data of the pre-trained deep learning classification model is the scene feature data used in the training process of the pre-trained deep learning classification model, which includes the feature content and feature correlation relationship corresponding to multiple training scenes. Extract all feature elements from the current scene feature data to form the current feature element set; extract all feature elements from the historical training scene feature data of the pre-trained deep learning classification model to form the historical feature element set. Calculate the representation similarity between the current feature element set and the corresponding feature elements in the historical feature element set. The representation similarity is determined based on the semantic distance, lexical overlap and representation structure similarity of the feature elements, and generate a feature representation similarity set. Analyze the relationships between feature elements in the current scene feature data to form a current feature association network; analyze the relationships between feature elements in the historical training scene feature data of the pre-trained deep learning classification model to form a historical feature association network. Calculate the structural similarity between the current feature association network and the historical feature association network, including the similarity of network node connection methods, network hierarchical structure, and the similarity of the location of core network features, and generate feature association similarity. Based on scene feature evolution trajectory data, the evolution trend parameters of current scene features are extracted, including feature update frequency, feature content adjustment magnitude and feature correlation change rate. Based on the time series of historical training scene feature data of pre-trained deep learning classification models, the evolution trend parameters of historical scene features are extracted. Calculate the similarity between the evolution trend parameters of the current scene features and the evolution trend parameters of the historical scene features, and generate feature evolution trend similarity. Assign corresponding computational weights to the feature representation similarity set, feature association similarity, and feature evolution trend similarity. Based on the computational weights, perform a weighted summation of the average value of the feature representation similarity set, feature association similarity, and feature evolution trend similarity to obtain the transfer adaptation coefficient between the scene feature data and the historical training scene feature data of the pre-trained deep learning classification model.

5. The deep learning-based method for classifying operational and maintenance technical service issues according to claim 1, characterized in that, Based on historical operation and maintenance problem classification results, corresponding operation and maintenance processing effect data, and scenario feature evolution trajectory data, the feature weights and feature association rules in the initial classification criteria are adjusted in stages to obtain optimized classification criteria, including: Retrieve historical operation and maintenance problem classification result dataset, corresponding operation and maintenance processing effect dataset, and historical scenario feature evolution trajectory data. The historical operation and maintenance problem classification result dataset contains historical operation and maintenance problem text units, corresponding historical classification results, and classification basis. The operation and maintenance processing effect dataset contains operation and maintenance processing solutions corresponding to historical classification results, post-processing business recovery status data, and processing efficiency data. The historical scenario feature evolution trajectory data is a record of the scenario feature evolution trajectory within a historical time period. Data association is performed between the historical operation and maintenance problem classification result dataset and the operation and maintenance processing effect dataset. The correspondence between each historical classification result and the corresponding business recovery status data and processing efficiency data is established to generate a classification effect association dataset. Extract the performance features corresponding to the business recovery status data and the efficiency features corresponding to the processing efficiency data in the classification performance association dataset. The performance features include descriptions of business recovery integrity, recovery time, and post-recovery stability. The efficiency features include descriptions of the number of processing steps, processing resource consumption, and processing response speed. Input the classification criteria, corresponding effect features, efficiency features, and historical scene feature evolution trajectory data from the historical classification results into the iterative optimization module to analyze the correlation strength between the feature elements in the classification criteria and the effect features and efficiency features, as well as the changing pattern of feature elements with scene feature evolution. The initial adjustment direction of each feature element in the classification criteria is determined based on the correlation strength, and the dynamic adjustment range of the feature element weights is determined based on the change pattern, thus generating the first round of feature weight adjustment scheme. The first round of feature weight adjustment scheme is applied to the initial classification basis to obtain the first round of optimized classification basis. The first round of optimized classification basis and the corresponding historical operation and maintenance problem text units are input into the pre-trained deep learning classification model to generate the first round of simulated classification results. The first round of simulated classification results are compared with the actual classification results in the historical classification results. The classification accuracy, classification recall and classification matching degree are calculated to generate the first round of optimization evaluation data. If the first round of optimization evaluation data does not meet the preset optimization target, the parameters of the feature weight adjustment scheme, including the adjustment magnitude coefficient and adjustment direction, are adjusted based on the first round of optimization evaluation data to generate the second round of feature weight adjustment scheme, and the weight adjustment, simulated classification and evaluation steps are repeated. During multiple iterations, the feature association rules in the classification criteria are adjusted synchronously based on the scene feature evolution trajectory data, including feature element combination rules, feature priority ranking rules, and feature conflict resolution rules; When the evaluation data in any round reaches the preset optimization target, the iteration stops, and the classification criteria obtained in that round are determined as the optimized classification criteria.

6. The deep learning-based method for classifying operational and maintenance technical service issues according to claim 5, characterized in that, During multiple iterations, the feature association rules in the classification criteria are adjusted synchronously based on the scene feature evolution trajectory data, including: The evolutionary trajectory data of scene features is divided into evolutionary stages. Based on the magnitude and frequency of change of scene features, the evolutionary process is divided into multiple evolutionary stages. Each evolutionary stage corresponds to a continuous time interval and the state of scene features within that time interval. Extract the core scene feature elements within each evolutionary stage. The core scene feature elements are the scene feature elements that appear most frequently and have the most associated features within that evolutionary stage. Analyze the combination of core scene feature elements with other scene feature elements in each evolutionary stage, summarize the typical combination patterns of scene feature elements in different evolutionary stages, and adjust the feature element combination rules in the classification criteria based on the typical combination patterns so that the feature element combination rules match the scene feature combination patterns in the current evolutionary stage. Based on the changes in the importance of core scene feature elements in each evolutionary stage, the feature priority ranking rules in the classification criteria are adjusted, increasing the priority of the classification criteria feature elements corresponding to the core scene feature elements in the current evolutionary stage and decreasing the priority of the classification criteria feature elements corresponding to the non-core scene feature elements. Identify the conflict types among scene feature elements in each evolutionary stage, including expression conflicts, relationship conflicts, and scope of influence conflicts, and formulate corresponding conflict resolution strategies based on the conflict types, such as using a semantic fusion strategy for expression conflicts and a primary-secondary association strategy for relationship conflicts. Integrate conflict resolution strategies into the feature-based conflict resolution rules in the classification criteria, replacing the original conflict resolution rule content; During each iteration of optimization, based on the current stage of scene feature evolution, the corresponding feature element combination rules, feature priority ranking rules, and feature conflict resolution rules are invoked and applied to the classification criteria adjustment process.

7. The deep learning-based method for classifying operational and maintenance technical service issues according to claim 1, characterized in that, The optimized classification criteria are then matched with a dynamically updated database of operational technical service problem classification cases at multiple levels to generate a classification result for each operational technical service problem text unit, including: The system retrieves a dynamically updated database of operation and maintenance technical service issues, which includes historical and latest operation and maintenance issue case data. The historical issue case data includes historical issue text, historical scenario features, historical classification results, and historical classification criteria. The latest issue case data consists of recently added operation and maintenance issue case data, including the latest issue text, latest scenario features, latest classification results, and latest classification criteria. The optimized classification criteria are processed by feature layering. The key classification information in the optimized classification criteria is divided into a core feature layer, a secondary feature layer and an auxiliary feature layer according to the correlation dimension. The core feature layer contains feature content directly related to the type of operation and maintenance problem. The secondary feature layer contains feature content related to the level of operation and maintenance problem. The auxiliary feature layer contains feature content related to the operation and maintenance problem handling solution. The first-level matching is performed between the feature content of the core feature layer and the core features in the case classification criteria of all cases in the dynamically updated operation and maintenance technical service problem classification case library. The core feature similarity is calculated, and a set of candidate cases whose core feature similarity reaches the first preset threshold is selected. The feature content of the secondary feature layer is matched with the secondary features in the case classification criteria of each case in the candidate case set at the second level. The secondary feature similarity is calculated, and the secondary candidate case set with the secondary feature similarity reaching the second preset threshold is selected. The auxiliary feature content is matched with the auxiliary features in the case classification criteria of each case in the secondary candidate case set at the third level. The auxiliary feature similarity is calculated to generate the comprehensive similarity of each case. The comprehensive similarity is the weighted sum of the core feature similarity, secondary feature similarity and auxiliary feature similarity. The cases in the secondary candidate case set are sorted based on comprehensive similarity, and the case with the highest comprehensive similarity is selected as the optimal matching case. Extract the case classification result corresponding to the best matching case, and use the case classification result as the preliminary classification result corresponding to the current operation and maintenance problem text unit; Retrieve the case scenario features and case problem text corresponding to the best matching case, and compare them with the current operation and maintenance problem text unit and the corresponding scenario feature data to confirm the scenario adaptability and consistency of text description. If the scenario adaptability and textual consistency both meet the preset requirements, the preliminary classification result will be determined as the classification result of the operation and maintenance technical service problem corresponding to each operation and maintenance problem text unit; if not, the case with the second highest comprehensive similarity will be selected and the supplementary comparison step will be repeated until the operation and maintenance technical service problem classification result that meets the requirements is determined.

8. The deep learning-based method for classifying operational and maintenance technical service issues according to claim 7, characterized in that, The dynamically updated operational technical service problem classification case library includes: A dynamic update triggering mechanism is established for the dynamic update of the operation and maintenance technical service problem classification case library. The dynamic update triggering mechanism for the dynamic update of the operation and maintenance technical service problem classification case library is triggered based on the number of newly added operation and maintenance problem classification results or the time interval. When the number of newly added operation and maintenance problem classification results reaches a preset number threshold, or the time interval since the last update reaches a preset time threshold, the dynamic update of the operation and maintenance technical service problem classification case library is triggered. After triggering the dynamic update operation of the operation and maintenance technical service problem classification case library, the operation and maintenance problem classification result data generated in the recent period is retrieved. The operation and maintenance problem classification result data includes newly added operation and maintenance problem text units, corresponding classification results, classification basis and associated scenario feature data. The newly added operation and maintenance problem classification result data is processed to convert the case format, converting the newly added operation and maintenance problem text unit into the case problem text format, converting the associated scenario feature data into the case scenario feature format, converting the classification result into the case classification result format, converting the classification basis into the case classification basis format, and generating case data to be supplemented; The generated case data to be supplemented is checked for uniqueness. The similarity between the case data to be supplemented and the case data already stored in the dynamically updated operation and maintenance technical service problem classification case library is compared. If the similarity does not reach the case duplication threshold, the case data to be supplemented is marked as supplementable case data. Supplementable case data are categorized into corresponding case classification directories according to the storage structure of the dynamically updated operation and maintenance technical service problem classification case library. The case classification directories are hierarchically divided according to operation and maintenance problem type, scenario category, and classification time. After supplementing the case data, the index of all case data in the dynamically updated operation and maintenance technical service problem classification case library is updated, and feature indexes are established for the newly added supplementable case data, including case problem text index, case scenario feature index and case classification result index. After the index update is completed, a dynamically updated O&M technical service issue classification case library update log is generated, recording the time of this update, the number of cases added, and the case classification distribution, forming a dynamically updated O&M technical service issue classification case library containing historical O&M issue case data and the latest O&M issue case data.

9. The deep learning-based method for classifying operational and maintenance technical service issues according to claim 1, characterized in that, The step of inputting each operation and maintenance problem text unit, scenario feature data, and scenario feature evolution trajectory data into a pre-trained deep learning classification model to obtain the initial classification criteria for each operation and maintenance problem text unit also includes: Before calling the pre-trained deep learning classification model, the cross-scene feature transfer module and the feature fusion module of the pre-trained deep learning classification model are pre-adapted, and the cross-scene transfer parameters and feature fusion parameters used in the training process of the pre-trained deep learning classification model are retrieved. Input the scene feature data samples of the current operation and maintenance scenario into the cross-scene feature transfer module of the pre-trained deep learning classification model, test the matching degree between the transfer adaptation features output by the cross-scene feature transfer module of the pre-trained deep learning classification model and the expected features of the samples. If the matching degree does not reach the preset adaptation threshold, adjust the transfer coefficient and feature adjustment weight in the cross-scene transfer parameters. The adjusted cross-scene transfer parameters are applied to the cross-scene feature transfer module of the pre-trained deep learning classification model. Scene feature data samples are re-inputted for testing until the matching degree between the transferred adaptive features and the expected features of the samples reaches the preset adaptation threshold. Input text semantic feature data samples and transfer adaptation feature samples into the feature fusion module of the pre-trained deep learning classification model, test the matching degree between the fused feature data output by the feature fusion module of the pre-trained deep learning classification model and the expected fusion feature. If the matching degree does not reach the preset fusion threshold, adjust the fusion weight allocation rule and the associated algorithm parameters in the feature fusion parameters. The adjusted feature fusion parameters are applied to the feature fusion module of the pre-trained deep learning classification model. Text semantic feature data samples and transfer adaptation feature samples are re-inputted for testing until the matching degree between the fused feature data and the expected fused features reaches the preset fusion threshold. After completing the pre-adaptation process, the adjusted cross-scene transfer parameters and feature fusion parameters are saved and used as the running parameters of the currently invoked pre-trained deep learning classification model before executing the subsequent feature transfer adaptation and association fusion steps.

10. A deep learning-based system for classifying operational and maintenance technical service issues, characterized in that, The deep learning-based operation and maintenance technical service problem classification system includes a processor and a memory, the memory and the processor are connected, the memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the deep learning-based operation and maintenance technical service problem classification method according to any one of claims 1-9.

Images