Deep learning-based operation and maintenance technology service problem classification method and system

Abstract

The invention provides an operation and maintenance technology service problem classification method and system based on deep learning. The method comprises the steps that firstly, an operation and maintenance technology service problem description text set with collection time information is acquired; executing dynamic updating and feature evolution analysis processing of the operation and maintenance scene feature library, and generating scene feature evolution trajectory data; inputting the operation and maintenance problem text unit, the scene feature data and the scene feature evolution trajectory data into a pre-trained deep learning classification model to obtain an initial classification basis; based on a historical operation and maintenance problem classification result, operation and maintenance processing effect data and scene feature evolution trajectory data, feature weights and feature association rules in the initial classification basis are adjusted in stages, and an optimized classification basis is obtained; and finally, performing multi-level association matching on the optimized classification basis and a dynamically updated operation and maintenance technical service problem classification case library, and generating an operation and maintenance problem classification result, thereby being capable of adapting to a dynamically changing operation and maintenance scene, and realizing efficient and accurate problem classification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a deep learning-based operation and maintenance technical service problem classification method and system. BACKGROUND

[0002] In the current operation and maintenance technical service field, with the rapid development of information technology and the continuous improvement of enterprise informatization, operation and maintenance work is facing increasingly complex and diversified challenges. Operation and maintenance technical service problems cover software failures, system performance problems, network connection abnormalities and many other aspects, and these problems evolve over time and with business development.

[0003] Traditional operation and maintenance technical service problem classification methods mainly rely on manual experience or simple keyword matching rules. Although manual experience classification can make a certain degree of accurate judgment on the problem based on the professional knowledge of operation and maintenance personnel, manual processing is low in efficiency when facing massive operation and maintenance problem data, and the experience levels of different operation and maintenance personnel differ, making it difficult to guarantee the consistency and accuracy of the classification results. The classification method based on simple keyword matching rules lacks in-depth understanding of the semantics of the problem and cannot accurately grasp the essential characteristics of the problem. For some problems with complex expression or ambiguous semantics, the classification effect is often unsatisfactory. In addition, existing methods do not fully consider the dynamic changes of the operation and maintenance scene and the guiding role of historical operation and maintenance data on problem classification, making it difficult to adapt to the changing operation and maintenance environment and unable to meet the demand of modern operation and maintenance work for efficient and accurate problem classification. SUMMARY

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present application, the embodiments of the present application provide a deep learning-based operation and maintenance technical service problem classification method, which comprises: obtaining an operation and maintenance technical service problem description text set, the operation and maintenance technical service problem description text set containing a plurality of operation and maintenance problem text units, each operation and maintenance problem text unit being attached with collection time information; performing dynamic updating and feature evolution analysis processing of an operation and maintenance scene feature library to generate scene feature evolution trajectory data, the operation and maintenance scene feature library storing scene feature data corresponding to the operation and maintenance scene; inputting each operation and maintenance problem text unit, scene feature data and scene feature evolution trajectory data into a pre-trained deep learning classification model to obtain 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 scene feature evolution trajectory data, adjusting the feature weights and feature association rules in the initial classification basis in stages to obtain optimized classification basis; The optimized classification basis is matched with the dynamically updated operation and maintenance technical service problem classification case library in multiple levels to generate an operation and maintenance technical service problem classification result corresponding to each operation and maintenance problem text unit.

[0005] In another aspect, the embodiment of the present application also provides a deep learning-based operation and maintenance technical service problem classification system, which comprises a processor and a machine readable storage medium, the machine readable storage medium is connected with the processor, the machine readable storage medium is used for storing programs, instructions or codes, and the processor is used for executing the programs, instructions or codes in the machine readable storage medium to realize the above-mentioned method.

[0006] Based on the above aspects, the embodiment of the present application can obtain an operation and maintenance technical service problem description text set containing collection time information, perform operation and maintenance scene feature library dynamic updating and feature evolution analysis processing, generate scene feature evolution trajectory data, can track the change of the operation and maintenance scene in real time, make the classification process adapt to the continuously evolving operation and maintenance environment, and enhance the flexibility and adaptability of the classification. The operation and maintenance problem text unit, scene feature data and scene feature evolution trajectory data are input into the pre-trained deep learning classification model to obtain the initial classification basis, with the help of the powerful ability of deep learning, the complex relationship between the problem text and the scene feature is deeply mined, and the accuracy of the initial classification is improved. Based on the historical operation and maintenance problem classification result, the corresponding operation and maintenance processing effect data and the scene feature evolution trajectory data, the feature weight and the feature association rule in the initial classification basis are adjusted in stages, the experience of the historical data can be fully utilized, the classification basis is further optimized, and the classification result is more in line with the actual operation and maintenance demand. Finally, the optimized classification basis is matched with the dynamically updated operation and maintenance technical service problem classification case library in multiple levels to generate an operation and maintenance technical service problem classification result corresponding to each operation and maintenance problem text unit, and the efficiency and quality of the operation and maintenance work are effectively improved. BRIEF DESCRIPTION OF DRAWINGS

[0007] Figure 1 is an execution flow diagram of the deep learning-based operation and maintenance technical service problem classification method provided by the embodiment of the present application.

[0008] Figure 2 is a schematic diagram of exemplary hardware and software components of the deep learning-based operation and maintenance technical service problem classification system provided by the embodiment of the present application. DETAILED DESCRIPTION

[0009] The present application will be described in detail below with reference to the accompanying drawings, Figure 1 is a flow diagram of the deep learning-based operation and maintenance technical service problem classification method provided by an embodiment of the present application, and the deep learning-based operation and maintenance technical service problem classification method will be described in detail below.

[0010] Step S110: Obtain a set of operation and maintenance technical service problem description texts, which contains a plurality of operation and maintenance problem text units, each of which is attached with collection time information.

[0011] In this embodiment, the Internet of Things smart home control operation and maintenance service is taken as a scene, and data is collected from user APP operation and maintenance work orders, device voice interaction fault descriptions, device log text abnormality prompts, and operation and maintenance personnel input texts through a text collection component of a smart home control platform. In combination of real-time monitoring (work order, voice interaction) and regular pulling (device log, operation and maintenance input text), after blank, duplicate, and irrelevant texts are removed, an operation and maintenance technical service problem description text set is formed. Each operation and maintenance problem text unit corresponds to a specific operation and maintenance event, such as "smart socket APP control is unresponsive, and the physical switch is normal", and is bound with a timestamp generated by a data source as collection time information, so as to store fields such as "text content", "collection time", and "device identifier" in a JSON format.

[0012] Step S120: Perform dynamic updating and feature evolution analysis processing of an operation and maintenance scene feature library to generate scene feature evolution trajectory data, wherein the operation and maintenance scene feature library stores scene feature data corresponding to an operation and maintenance scene.

[0013] The operation and maintenance scene feature library pre-stores smart home device features (type, brand, communication protocol), network features (network type, router model), control scene features (lighting, security control), and user habit features (common control period), etc. New features are supplemented through dynamic updating, and scene feature evolution trajectory data is generated by combining evolution analysis to capture the time variation law of the features.

[0014] Step S121: Call business operation log data and historical scene feature data of a current operation and maintenance business scene, wherein the business operation log data contains device operation state description data, business process execution description data, and abnormal event record data in the current operation and maintenance business scene, and the historical scene feature data is scene feature data in the operation and maintenance scene feature library in a recent period of time.

[0015] The business operation log data is obtained from a distributed log system through a platform log calling interface: the device operation state description data contains device online state, sensor data (such as temperature and humidity), and actuator feedback (such as motor action); the business process execution description data contains linkage rule execution and remote instruction issuing record; and the abnormal event record data contains device alarm (low power, communication interruption) and instruction execution failure log. At the same time, historical scene feature data in a recent period of time is called from the operation and maintenance scene feature library, which contains device change and network parameter adjustment records.

[0016] Step S122: Extract the scene feature elements in the operation log data, which are expressions that can represent the characteristics of the current operation scene, including device type association expressions, business process stage expressions, and abnormal event type expressions.

[0017] The scene feature elements are extracted using a rule engine combined with a BERT semantic extraction model. The rule engine matches candidate expressions based on pre-set keywords such as "smart lighting-LED", "instruction issuance", and "communication interruption". The BERT model performs semantic verification on the candidate expressions, eliminating 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 anomaly".

[0018] Step S123: Compare the scene feature elements with the stored scene feature data in the operation scene feature library, and filter out the scene feature elements that do not reach the pre-set association degree of similarity with the stored scene feature data expressions. The scene feature elements are treated as new scene feature data and are classified into the corresponding scene category according to the storage structure of the operation scene feature library.

[0019] The scene feature elements and the stored features in the library are converted into unified dimension semantic vectors through the Word2Vec model, the cosine similarity is calculated and compared with the pre-set threshold. The elements that do not reach the threshold are treated as new scene feature data and are classified into the corresponding category according to the hierarchical structure of the operation scene feature library (device feature class-secondary classification of lighting / security devices-third classification of specific device features), such as "smart bath heater-warm type" into "device feature class-environmental regulation device feature-bath heater feature", and adding metadata such as timestamp and source identifier.

[0020] Step S124: Arrange the historical scene feature data and the new scene feature data in chronological order to construct a scene feature time series data set, which contains scene feature data at different time points and corresponding time markers.

[0021] Extract the storage timestamps of historical features and the log timestamps of new features, sort them by time (sort by "device feature class> network feature class> control scene feature class> user habit feature class" twice), and associate the timestamps to construct a scene feature time series data set. Store the "time marker", "feature category", "feature content", and "feature status (new / changed)" fields in a time series database.

[0022] Step S125: Based on the scene feature time series dataset, the change trend of each scene feature data in the time dimension is identified by a feature evolution analysis algorithm, including the change trend of feature occurrence frequency, the change trend of feature expression details adjustment, and the change trend of feature association.

[0023] The time series dataset is divided by a sliding window, the occurrence frequency of the feature in each window is counted to identify the occurrence frequency change trend (up / down / stable), the same feature in different windows is compared to identify the expression detail adjustment trend (such as "communication anomaly" - "2.4G WiFi anomaly"), and the co-occurrence number of features is calculated to identify the association change trend (such as "smart curtain failure" and "router restart" co-occurrence increase, then the association is enhanced).

[0024] Step S126: According to the identified change trend, a feature evolution path is constructed for each scene feature data, all scene feature data evolution paths are integrated, and scene feature evolution trajectory data is generated, which contains time change nodes, change content description, and change impact range description of scene features.

[0025] Based on the time axis, each feature is labeled with a time change node, the change content is recorded (such as "T1: smart door lock failure frequency increases"), and the impact range is described from the number of affected devices, user groups, and control scene dimensions (such as "1 smart door lock, 1 user, no involvement in security scene"). According to the feature category, all evolution paths are integrated to form scene feature evolution trajectory data with a graph structure, nodes are features, edges are association relationships, and time and impact information are labeled.

[0026] Step S130: Each operation and maintenance problem text unit, scene feature data, and scene feature evolution trajectory data are input into a pre-trained deep learning classification model to obtain the initial classification basis of each operation and maintenance problem text unit.

[0027] The pre-trained deep learning classification model is based on the Transformer architecture, including text encoding, cross-scene feature migration, feature fusion, and classification basis generation modules. After inputting the three types of data, the initial classification basis is output after processing by each module, including problem category, associated scene, etc.

[0028] Step S131: The operation and maintenance problem text unit is input into the text encoding module of the pre-trained deep learning classification model for semantic analysis and vector conversion, generating text semantic feature data corresponding to the operation and maintenance problem text unit, which contains core expression vectors, abnormal phenomenon description vectors, and associated device expression vectors of the operation and maintenance problem.

[0029] The text encoding module is composed of an embedding layer, a multi-head attention layer, and a feedforward neural network layer. The embedding layer is based on domain dictionary segmentation and converted into word embedding vectors; the multi-head attention layer calculates the semantic association between words (such as the association between “no response” and “device offline”); the feedforward neural network layer generates text semantic feature data through linear transformation and activation function, including core expression vectors (such as “control no response”), abnormal phenomenon description vectors (such as “device offline”), and associated device expression vectors (such as “smart curtain”), all of which are fixed-dimension multi-value vectors.

[0030] Step S132: input the scene feature data and the scene feature evolution trajectory data into the cross-scene feature migration module of the pre-trained deep learning classification model, and after feature alignment processing of the scene feature data and the scene feature evolution trajectory data, calculate the migration adaptation coefficient of the scene feature data and the historical training scene feature data.

[0031] The cross-scene feature migration module first unifies the feature dimensions (such as converting both category type and sequence type features into the same dimension vector) and matches the semantic related features through the feature alignment submodule, and then calculates the adaptation coefficient through the adaptation coefficient calculation submodule.

[0032] Step S1321: retrieve the historical training scene feature data of the pre-trained deep learning classification model, which is the scene feature data used in the training process of the pre-trained deep learning classification model, containing feature content and feature association relationship corresponding to multiple training scenes.

[0033] The historical training scene feature data is retrieved through the model parameter interface, including smart home operation and maintenance training scenes such as residential buildings, apartments, and villas, each scene containing feature content and feature association relationship (such as the causal association between “router failure” and “device offline”), and storing fields such as “scene identifier”, “feature category”, and “association relationship” in a structured format.

[0034] Step S1322: extract all feature elements in the current scene feature data to form a current feature element set, and extract all feature elements in the historical training scene feature data of the pre-trained deep learning classification model to form a historical feature element set.

[0035] Traverse the “feature content” field of the current scene feature data to extract elements such as “smart door lock” and “WiFi connection” to form the current feature element set; similarly, traverse the historical training data to extract elements to form the historical feature element set, both of which are stored in a non-repeating set structure.

[0036] Step S1323: Calculate the representation similarity of the corresponding feature elements in the current feature element set and the historical feature element set, the representation similarity is determined based on the semantic distance, the lexical overlap and the representation structure similarity of the feature elements, and a feature representation similarity set is generated.

[0037] The feature elements are converted into semantic vectors to calculate the Euclidean distance (semantic distance); the elements are split into basic words to count the proportion of the same words (lexical overlap); and the structure similarity is determined by comparing the representation structure (such as the “device-function” structure). The lexical overlap, the semantic distance and the structure similarity are weighted according to the lexical overlap > the semantic distance > the structure similarity, and the corresponding element representation similarity is calculated by weighting, and integrated into the feature representation similarity set.

[0038] Step S1324: Analyze the association relationship between each feature element in the current scene feature data to form a current feature association network, and analyze the association relationship between each feature element in the historical training scene feature data of the pre-trained deep learning classification model to form a historical feature association network.

[0039] Determine the association relationship between each feature element (based on co-occurrence frequency, logical association and domain rule), such as the association between “smart door lock failure” and “gateway offline”, and construct a current feature association network with nodes (feature elements, size reflecting frequency) and edges (association relationship, thickness reflecting strength). Similarly, construct each training scene feature association sub-network, and integrate into a historical feature association network.

[0040] Step S1325: Calculate the structure similarity of the current feature association network and the historical feature association network, including the network node connection method similarity, the network hierarchical structure similarity and the network core feature position similarity, and generate the feature association relationship similarity.

[0041] Extract the network connection matrix to calculate the node connection method similarity; divide the core layer (most associated), the middle layer and the edge layer, compare the number and type of hierarchical nodes to determine the hierarchical structure similarity; identify the core node and calculate the position coordinate distance to determine the core feature position similarity. Weighted sum of the three indicators to obtain the feature association relationship similarity.

[0042] Step S1326: Extract the evolution trend parameters of the current scene feature based on the scene feature evolution trajectory data, including the feature update frequency, the feature content adjustment amplitude and the feature association relationship change rate, and extract the evolution trend parameters of the historical scene feature based on the time sequence of the historical training scene feature data of the pre-trained deep learning classification model.

[0043] Extract the current feature evolution trend parameters from the scene feature evolution trajectory data: the number of new / changed features per unit time (update frequency), semantic difference (adjustment amplitude), and the number of associated relationship changes per unit time (change rate). Construct a time series of historical training data to extract the historical feature evolution trend parameters in the same way.

[0044] Step S1327: Calculate the similarity between the evolution trend parameters of the current scene feature and the evolution trend parameters of the historical scene feature, and generate the feature evolution trend similarity.

[0045] The two types of parameters are processed using Min-Max standardization, and the cosine similarity of feature update frequency, content adjustment amplitude, and association change rate is calculated. The feature evolution trend similarity is obtained by equal-weighted weighted summation.

[0046] Step S1328: Assign corresponding calculation weights to the feature expression similarity set, feature association relationship similarity, and feature evolution trend similarity. Based on the calculation weights, the average value of the feature expression similarity set, the feature association relationship similarity, and the feature evolution trend similarity are weighted and summed to obtain the transfer adaptation coefficient of the scene feature data and the historical training scene feature data of the pre-trained deep learning classification model.

[0047] The weights are assigned according to the feature expression similarity set > feature association relationship similarity > feature evolution trend similarity. The average value of the feature expression similarity set is calculated, and the other two indicators are weighted and summed to obtain the transfer adaptation coefficient in the interval [0, 1].

[0048] Step S133: Based on the transfer adaptation coefficient, adjust the scene feature data to generate a transfer adaptation feature that is adapted to the training data distribution of the pre-trained deep learning classification model.

[0049] If the adaptation coefficient ≥ the preset threshold, fine-tune the expression details; if < the threshold, optimize the feature expression (refer to historical data), adjust the association strength, and correct the evolution trend parameters to generate a transfer adaptation feature that makes its data distribution close to the model training data.

[0050] Step S134: Input the transfer adaptation feature and the text semantic feature data into the feature fusion module of the pre-trained deep learning classification model to establish the association relationship between the transfer adaptation feature and the text semantic feature data, including the feature element correspondence relationship, the feature weight influence relationship, and the feature semantic complementary relationship.

[0051] The feature fusion module adopts an attention mechanism and a graph neural network architecture. The attention mechanism layer calculates feature element attention weights, matches semantically related elements to establish a corresponding relationship (such as "smart curtain failure" and "curtain non-response"), adjusts text semantic feature weights according to the frequency of migrated feature elements (such as "gateway offline" with high frequency to increase the weight of "device offline"), and identifies complementary semantic information (such as the text "light non-response" and the migrated feature "network peak period prone to occur") to establish a semantic complementary relationship.

[0052] Step S135: Based on the established association relationship, the migrated adaptive features and the text semantic feature data are weighted and fused to generate fusion feature data containing scene migration information and text semantic information.

[0053] According to the association relationship, the fusion weight is distributed (the corresponding and complementary feature weight is high), the feature elements are sorted and spliced according to the weight, and the fusion feature data is generated after dimension unification processing, which contains scene and semantic information.

[0054] Step S136: The fusion feature data is input into the classification basis generation module of the pre-trained deep learning classification model, the fusion feature data is feature extracted and rule matched, and the key classification information in the fusion feature data is extracted, including operation and maintenance problem type associated features, abnormal level corresponding features and processing scheme matching features.

[0055] The classification basis generation module contains a feature extraction sublayer (multi-size convolution kernel extraction abstract feature) with a convolutional neural network structure, a rule matching sublayer (matching classification rule library, such as "containing communication interruption - network class problem"), and a key information screening sublayer, which extracts key classification information such as operation and maintenance problem type (network / hardware / configuration class), abnormal level (high / medium / low level), and processing scheme direction (restart / check network).

[0056] Step S137: The key classification information is arranged according to the preset classification basis structure to generate the initial classification basis of each operation and maintenance problem text unit.

[0057] The key information is arranged according to the "problem type" "abnormal level" "associated scene feature" "possible cause" "processing direction" structure. For example, for "smart curtain APP non-response, display offline", the initial classification basis is "problem type: network connection class; abnormal level: medium level; associated scene feature: smart curtain-WiFi type, weak router signal; possible cause: WiFi interruption, gateway connection abnormality; processing direction: check router, restart curtain device".

[0058] Step S138: Before calling the pre-trained deep learning classification model, the cross-scene feature migration module of the pre-trained deep learning classification model and the feature fusion module of the pre-trained deep learning classification model are pre-adapted, and the cross-scene migration parameters and the feature fusion parameters used in the training process of the pre-trained deep learning classification model are called.

[0059] Before calling the model, the cross-scene migration parameters (feature alignment, adaptation coefficient weight, etc.) and the feature fusion parameters (attention weight, fusion coefficient, etc.) are called through the parameter management interface, and pre-adaptation adjustment is performed to adapt to the current scene.

[0060] Step S1381: Input the scene feature data sample of the current operation and maintenance scene into the cross-scene feature migration module of the pre-trained deep learning classification model, test the matching degree of the migration adaptation feature output by the cross-scene feature migration module of the pre-trained deep learning classification model and the sample expected feature, and if the matching degree does not reach the preset adaptation threshold, adjust the migration coefficient and feature adjustment weight in the cross-scene migration parameters.

[0061] Select each type of feature data in the current scene as a sample (including feature elements and associated relationships), and set the sample expected feature (element content, dimension, and association strength) by an expert. Input the sample into the cross-scene feature migration module, calculate the matching degree (element coincidence degree + dimension consistency + association strength similarity weighted sum) of the migration adaptation feature output and the expected feature. When the adaptation threshold is not reached, adjust the migration coefficient for missing elements, adjust the feature adjustment weight for dimension mismatch, and correct the association coefficient for association deviation. Repeat the test until the standard is met.

[0062] For example, step S13811: select different types of scene feature data in the current operation and maintenance scene as scene feature data samples, and each type of scene feature data sample includes multiple feature elements and corresponding association relationship descriptions.

[0063] Select samples from device feature classes (such as smart door locks, lights), network feature classes (routers, WiFi), control scene feature classes (lighting linkage, security control), and user habit feature classes (common light-on time period), each sample including multiple feature elements and association relationships (such as "smart door lock - fingerprint type" and "gateway - WiFi type association").

[0064] Step S13812: Set the sample expected feature for each scene feature data sample, which is determined based on the experience of domain experts and the training target of the pre-trained deep learning classification model, and includes expected feature element content, feature dimension, and feature association strength.

[0065] The domain expert sets the expected features for each sample in combination with the current scene and the model training target: clearly includes the feature elements, unified feature dimensions, and reasonable correlation strength (for example, the correlation strength between "smart light-LED type" and "router dual-frequency type" is medium).

[0066] Step S13813: input the scene feature data sample into the cross-scene feature migration module of the pre-trained deep learning classification model, perform feature migration processing on the sample based on the current cross-scene migration parameter, and output the migration adapted feature.

[0067] The sample is input into the cross-scene feature migration module, which completes feature alignment, adaptation coefficient calculation, and feature adjustment based on the current migration parameter, and outputs the migration adapted feature.

[0068] Step S13814: calculate the feature element coincidence degree, feature dimension consistency, and feature correlation strength similarity of the migration adapted feature and the sample expected feature, and determine the matching degree based on the standardized weighted sum of the three.

[0069] The same element proportion (element coincidence degree) of the migration adapted feature and the expected feature is calculated, whether the feature dimensions are consistent (dimension consistency) is compared, and the cosine similarity of the correlation strength (correlation strength similarity) is calculated. The matching degree is obtained by standardizing and weighted summing the three indicators.

[0070] Step S13815: if the calculated matching degree does not reach the preset adaptation threshold, analyze the difference points between the migration adapted feature and the sample expected feature, including missing feature elements, dimensionally unmatched feature items, and feature relationships with large correlation strength deviations.

[0071] When the adaptation threshold is not reached, compare the migration adapted feature and the expected feature to identify the difference points: for example, the "router signal strength" element is missing, the feature dimension is 128 dimensions and the expected dimension is 256 dimensions, the correlation strength between "smart window curtain" and "gateway" is weak and the expected correlation strength is strong, and the deviation is large.

[0072] Step S13816: for the missing feature elements, adjust the migration coefficient in the cross-scene migration parameter, and increase the migration weight of the historical training scene feature data related to the missing feature elements. Specifically, in the cross-scene migration parameter, the value of the migration coefficient is increased for the historical feature items related to the semantic of the missing feature elements, so that the cross-scene feature migration module can refer to these related historical features more when processing the current scene feature data sample, thereby including the missing feature elements in the output migration adapted feature.

[0073] Step S13817: For the feature items with dimension mismatch, adjust the feature adjustment weight in the cross-scene migration parameters, and increase the weight proportion of the target dimension in the dimension conversion process.

[0074] When there is a dimension mismatch between the migration adaptation feature and the sample expected feature, in the feature adjustment weight configuration of the cross-scene migration parameters, find the parameter item related to dimension conversion, and increase the weight proportion of the target dimension (i.e., the dimension of the sample expected feature) in the dimension conversion calculation. At the same time, adjust the mapping relationship parameters in the dimension conversion algorithm, so that the current scene feature data sample can adapt to the dimension requirements of the sample expected feature after dimension conversion, and eliminate the differences caused by dimension mismatch.

[0075] Step S13818: For the feature relationships with large correlation strength deviation, adjust the correlation relationship migration coefficient in the cross-scene migration parameters, and correct the calculation rule of the feature correlation strength.

[0076] For the feature relationship with large correlation strength deviation, locate the corresponding correlation relationship migration coefficient configuration in the cross-scene migration parameters, and adjust the coefficient value according to the deviation direction. If the correlation strength in the migration adaptation feature is lower than expected, increase the correlation relationship migration coefficient; if it is higher than expected, decrease the coefficient. At the same time, correct the calculation rule of the feature correlation strength, for example, increase the weight of the business operation log data related to the feature relationship in the current scene when calculating the correlation strength, so that the calculation result is closer to the correlation strength requirements of the sample expected feature.

[0077] Step S13819: After the adjustment is completed, the scene feature data sample is re-input into the cross-scene feature migration module of the pre-trained deep learning classification model, the new matching degree is calculated, and the adjustment and test steps are repeated until the matching degree reaches the preset adaptation threshold.

[0078] After the above parameter adjustment is completed, the same batch of scene feature data samples is input into the cross-scene feature migration module again, and the module outputs new migration adaptation features based on the adjusted cross-scene migration parameters. According to the calculation method of step S13814, the matching degree of the new migration adaptation feature and the sample expected feature is calculated, and compared with the preset adaptation threshold. If the threshold is still not reached, the difference point analysis, parameter adjustment and matching degree calculation steps are repeated until the matching degree meets the preset requirements.

[0079] Step S1382: Input the text semantic feature data sample and the migration adaptation feature sample into the feature fusion module of the pre-trained deep learning classification model, test the matching degree of the fusion feature data output by the feature fusion module of the pre-trained deep learning classification model and the fusion expected feature, and if the matching degree does not reach the preset fusion threshold, adjust the fusion weight distribution rule and the correlation algorithm parameters in the feature fusion parameters.

[0080] The text semantic feature data samples (such as the core expression vector corresponding to "smart light no response", the abnormal phenomenon description vector, etc.) in the intelligent home operation and maintenance scene and the migration adaptation feature samples (such as the feature vectors corresponding to "smart light-LED type" and "router signal weak") processed through cross-scene feature migration are combined and input into the feature fusion module.

[0081] The domain experts set the fusion expected features for the combined samples according to the feature fusion target, and clearly define the scene and semantic information, feature dimension, and feature association relationship to be contained after fusion. The feature fusion module processes the input samples based on the current feature fusion parameters and outputs the fusion feature data.

[0082] The matching degree of the fusion feature data and the fusion expected features is calculated, which comprehensively considers the information integrity (whether the expected scene and semantic information is contained), dimension consistency (whether the dimension is consistent with the fusion expected features), and feature association rationality (whether the association relationship between the features meets the expectation), and is obtained by standardizing and weighting sum.

[0083] If the matching degree does not reach the preset fusion threshold, the differences between the two are analyzed, such as missing part of the semantic information, inconsistent dimension, unreasonable feature association logic, etc. For the information missing problem, the fusion weight distribution rule is adjusted to increase the fusion weight of the feature elements corresponding to the missing information; for the dimension inconsistency problem, the dimension unification parameter in the feature fusion parameter is adjusted to correct the dimension conversion logic; for the unreasonable association logic problem, the association algorithm parameter is adjusted to optimize the calculation method of the association relationship between the features.

[0084] Step S1383: Apply the adjusted feature fusion parameters to the feature fusion module of the pre-trained deep learning classification model, and re-input the text semantic feature data samples and the migration adaptation feature samples for testing until the matching degree of the fusion feature data and the fusion expected features reaches the preset fusion threshold.

[0085] The adjusted feature fusion parameters are updated to the parameter configuration of the feature fusion module, and the text semantic feature data samples and the migration adaptation feature samples are input again, and the module outputs new fusion feature data. The new matching degree is obtained according to the same calculation method, and if it still does not reach the preset fusion threshold, the difference point analysis, parameter adjustment and testing steps are repeated until the fusion feature data meets the requirements of the fusion expected features.

[0086] Step S1384: After completing the pre-adaptation processing, save the adjusted cross-scene migration parameters and feature fusion parameters as the running parameters of the current called pre-trained deep learning classification model, and then perform the subsequent feature migration adaptation and association fusion steps.

[0087] When the output matching degrees of the cross-scene feature migration module and the feature fusion module both reach the preset threshold, the adjusted cross-scene migration parameters and the feature fusion parameters are saved through the model parameter storage interface, and are marked as model running parameters in the current smart home operation and maintenance scene. Subsequently, when processing an actual operation and maintenance problem text unit, the pre-trained deep learning classification model adopts the set of running parameters to perform feature migration adaptation and associated fusion operations.

[0088] Step S140: Based on the historical operation and maintenance problem classification results, the corresponding operation and maintenance processing effect data and the scene feature evolution trajectory data, the feature weights and the feature association rules in the initial classification basis are adjusted in stages to obtain an optimized classification basis.

[0089] The historical operation and maintenance data in the smart home operation and maintenance scene are called, including the classification results of historical operation and maintenance problems, the corresponding processing effects and the historical scene feature evolution trajectories. Through a staged iterative manner, the feature weights and the feature association rules in the initial classification basis generated in step S137 are adjusted and optimized, so that the classification basis is more suitable for the needs of the actual operation and maintenance scene, and the classification accuracy is improved.

[0090] Step S141: The historical operation and maintenance problem classification result data set, the corresponding operation and maintenance processing effect data set and the historical scene feature evolution trajectory data are called. The historical operation and maintenance problem classification result data set includes historical operation and maintenance problem text units, corresponding historical classification results and classification bases. The operation and maintenance processing effect data set includes operation and maintenance processing schemes corresponding to the historical classification results, business recovery state data after processing and processing efficiency data. The historical scene feature evolution trajectory data is a scene feature evolution trajectory record in a historical time period.

[0091] Through the query interface of the operation and maintenance data management platform, the historical operation and maintenance problem classification result data set is called. The historical operation and maintenance problem classification result data set is stored in chronological order, and each record includes a historical operation and maintenance problem text unit (such as “smart socket cannot be connected to the network”), a corresponding historical classification result (such as “network connection type - device networking fault”), and a classification basis used at the time.

[0092] The operation and maintenance processing effect data set is also called. The operation and maintenance processing effect data set is associated with the historical operation and maintenance problem classification result data set through a unique identifier. Each record includes an operation and maintenance processing scheme (such as “restart the router and smart socket”) adopted for the historical classification result, business recovery state data after processing (such as “device recovery networking, control normal”), and processing efficiency data (such as “processing time consumption” and “resource consumption”).

[0093] In addition, historical scene feature evolution track data is invoked, which is scene feature evolution track records in the past period of time (such as the past year), and contains scene feature changes and influence range information of each historical time node.

[0094] Step S142: Data correlation is performed on the historical operation and maintenance problem classification result data set and the operation and maintenance processing effect data set, a corresponding relationship between each historical classification result and corresponding business recovery state data and processing efficiency data is established, and a classification effect correlation data set is generated.

[0095] Based on the unique identifier (such as the operation and maintenance work order number) in the historical operation and maintenance problem classification result data set and the operation and maintenance processing effect data set, the two data sets are associated and matched, so that each historical classification result corresponds to its processed business recovery state data and processing efficiency data. For example, the historical classification result "network connection type - device networking failure" corresponds to the business recovery state data "device networking recovery" and the processing efficiency data "short processing time".

[0096] The associated data set is grouped according to the type of historical classification result to form a classification effect correlation data set, each group of data corresponds to a historical classification result type, which facilitates subsequent analysis of the processing effect of different classification results.

[0097] Step S143: Extract the effect features corresponding to the business recovery state data and the efficiency features corresponding to the processing efficiency data in the classification effect correlation data set. The effect features include business recovery integrity expression, recovery time expression and post-recovery stability expression. The efficiency features include processing step number expression, processing resource consumption expression and processing response speed expression.

[0098] Effect features are extracted from the business recovery state data in the classification effect correlation data set: business recovery integrity expression is used to describe the completeness of device or business recovery, such as "complete recovery" and "partial function recovery"; recovery time expression is used to describe the time from the start of processing to the recovery of normal, such as "short time recovery" and "long time recovery"; post-recovery stability expression is used to describe the running stability after recovery, such as "stable operation without exception" and "frequent recurrence".

[0099] Efficiency features are extracted from the processing efficiency data: processing step number expression is used to describe how many steps the processing process contains, such as "simple steps" and "complex steps"; processing resource consumption expression is used to describe the resource consumption in the processing process, such as "low resource consumption" and "high resource consumption"; processing response speed expression is used to describe the response speed from receiving the problem to starting processing, such as "rapid response" and "delayed response".

[0100] Step S144: input the classification basis in the historical classification result, the corresponding effect characteristics, efficiency characteristics and historical scene characteristic evolution track data into the iterative optimization module, analyze the association strength between the feature elements in the classification basis and the effect characteristics and efficiency characteristics, and the change law of the feature elements with the scene characteristics evolution.

[0101] The classification basis (including feature elements and feature association rules) in the historical classification result, the corresponding effect characteristics, efficiency characteristics and historical scene characteristic evolution track data are input into the iterative optimization module. The iterative optimization module first calculates the association strength between each feature element in the classification basis and the effect characteristics and efficiency characteristics through an association analysis algorithm, for example, analyzes the association closeness between the feature element "router signal weak" and the effect characteristic "long business recovery time".

[0102] At the same time, combined with the historical scene characteristic evolution track data, the appearance frequency, expression change and association relationship change of each feature element in the classification basis at different historical scene evolution stages are tracked, and the change law of the feature elements with the scene evolution is summarized, such as the feature element "smart door lock fingerprint identification fault" whose appearance frequency significantly increases at the router upgrade stage.

[0103] Step S145: determine the initial adjustment direction of each feature element in the classification basis based on the association strength, determine the weight dynamic adjustment amplitude of the feature element based on the change law, and generate a first round of feature weight adjustment scheme.

[0104] According to the association strength between the feature elements and the effect characteristics and efficiency characteristics, the initial adjustment direction of each feature element is determined. If a feature element has a high association strength with positive features such as "complete business recovery" and "high processing efficiency", the initial adjustment direction of the weight of the feature element is to increase; if it has a high association strength with negative features such as "incomplete business recovery" and "low processing efficiency", the initial adjustment direction is to decrease.

[0105] Based on the change law of the feature elements with the scene evolution, the weight dynamic adjustment amplitude is determined. If the appearance frequency of a feature element at the current scene evolution stage shows an upward trend, and the associated positive features are obvious, the adjustment amplitude is set to be larger; if the appearance frequency is stable or decreases, the adjustment amplitude is set to be smaller. The first round of feature weight adjustment scheme is generated by comprehensively considering the initial adjustment direction and the adjustment amplitude, which clearly defines the weight adjustment value and adjustment method of each feature element.

[0106] 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 unit into the pre-trained deep learning classification model to generate the first round of simulated classification result.

[0107] According to the first round feature weight adjustment scheme, the weight values of each feature element in the initial classification basis are modified to form a first round optimized classification basis. The historical operation and maintenance problem text units corresponding to the classification basis are selected, and both are input into the pre-trained deep learning classification model (using the running parameters saved in step S1384). The model classifies the historical operation and maintenance problem text units based on the optimized classification basis, and outputs the first round simulated classification result.

[0108] Step S147: Compare the first round simulated classification result with the actual classification result in the historical classification result, calculate the classification accuracy, classification recall rate and classification matching degree, and generate first round optimization evaluation data.

[0109] The first round simulated classification result and the actual classification result in the historical classification result are compared one by one, the proportion of the number of correct classifications to the total number is counted to obtain the classification accuracy; the proportion of the number of correct classifications to the total number of the actual classification result is counted to obtain the classification accuracy; the matching degree of the feature elements and the association rules in the simulated classification result and the actual classification result is counted to obtain the classification matching degree. The three indicators are integrated into the first round optimization evaluation data.

[0110] Step S148: If the first round optimization evaluation data does not reach the preset optimization target, adjust the parameters of the feature weight adjustment scheme based on the first round optimization evaluation data, including the adjustment amplitude coefficient and the adjustment direction, generate a second round feature weight adjustment scheme, and repeat the weight adjustment, simulated classification and evaluation steps.

[0111] The first round optimization evaluation data is compared with the preset optimization target (such as classification accuracy ≥ preset value, classification recall rate ≥ preset value, classification matching degree ≥ preset value). If the target is not reached, the reason for not meeting the standard is analyzed. For example, if the classification recall rate of a certain category is low, it means that the weight of the feature element corresponding to the category may be insufficient.

[0112] Based on the analysis result, the parameters of the feature weight adjustment scheme are adjusted: if the influence of a certain feature element needs to be further enhanced, the adjustment amplitude coefficient is increased; if it is found that the initial adjustment direction is wrong (such as the classification effect decreases after the weight of a certain feature element is increased), the adjustment direction is reversed. According to this, a second round feature weight adjustment scheme is generated, and the operations of steps S146 to S147 are repeated to obtain second round optimization evaluation data.

[0113] Step S149: In the process of multiple iterations, the feature association rules in the classification basis are adjusted based on the scene feature evolution trajectory data, including the feature element combination rule, the feature priority ordering rule and the feature conflict resolution rule.

[0114] Meanwhile, the feature correlation rules in the classification basis are adjusted in combination with the scene feature evolution track data, so as to ensure that the correlation rules are adapted to the evolution state of the current scene feature.

[0115] Step S1491: Evolution stage division is performed on the scene feature evolution track data, and the evolution process is divided into multiple evolution stages based on the change amplitude and change frequency of the scene feature, each evolution stage corresponding to a continuous time interval and the scene feature state in the time interval.

[0116] The scene feature change amplitude threshold and change frequency threshold are set, and the scene feature evolution track data is analyzed. When the change amplitude of the scene feature in a continuous time interval exceeds the threshold or the change frequency exceeds the threshold, the time interval is divided into an independent evolution stage. Each evolution stage corresponds to a clear time start and end point and a stable scene feature state in the stage, such as a "router upgrade stage" and a "new device access stage".

[0117] Step S1492: Core scene feature elements in each evolution stage are extracted, which are the scene feature elements with the highest appearance frequency and the most associated features in the evolution stage.

[0118] The scene feature elements in each evolution stage are counted, the appearance frequency and the number of associated features of each element are calculated, and the top-ranked feature elements in both aspects are selected as the core scene feature elements in the stage. For example, in the "router upgrade stage", the "router model change" and "device networking anomaly" have the highest appearance frequency and the most associated features, which are the core scene feature elements in the stage.

[0119] Step S1493: The combination mode of the core scene feature elements and other scene feature elements in each evolution stage is analyzed, and the typical combination patterns of the scene feature elements in different evolution stages are summarized, and the feature element combination rules in the classification basis are adjusted based on the typical combination patterns, so that the feature element combination rules are matched with the scene feature combination patterns in the current evolution stage.

[0120] For each evolution stage, the combination of the core scene feature elements and other scene feature elements is counted, and the typical combination patterns of the stage are summarized, such as "router model change + smart device networking anomaly" and "new device access + linkage rule failure". According to these typical combination patterns, the feature element combination rules in the classification basis are adjusted, for example, in the "router upgrade stage", the combination rule priority of "router model change" and "device networking anomaly" is improved, so that the classification basis can accurately capture the problem features in the stage.

[0121] Step S1494: Based on the importance change of the core scene feature elements in each evolution stage, adjust the feature priority ranking rules in the classification basis, and promote the priority of the classification basis feature elements corresponding to the core scene feature elements in the current evolution stage, and reduce the priority of the classification basis feature elements corresponding to the non-core scene feature elements.

[0122] According to the importance change of the core scene feature elements in different evolution stages (judged comprehensively by the appearance frequency, the number of associated features and the influence degree on operation and maintenance problems), the priority ranking rules of the feature elements in the classification basis are adjusted. In the current evolution stage, the priority of the classification basis feature elements corresponding to the core scene feature elements (such as the feature elements corresponding to “router model change”) is promoted to the front row, and the priority of the classification basis feature elements corresponding to the non-core scene feature elements is reduced, so as to ensure that the key features in the current scene are considered first in classification.

[0123] Step S1495: Identify the conflict types between the scene feature elements in each evolution stage, including expression conflict, association relationship conflict and influence range conflict, and develop corresponding conflict resolution strategies based on the conflict types, such as using semantic fusion strategy for expression conflict and using primary and secondary association strategy for association relationship conflict.

[0124] In each evolution stage, the conflicts between the scene feature elements are analyzed: expression conflict refers to the inconsistency of different feature elements in expressing the same phenomenon, such as “device offline” and “device not connected”; association relationship conflict refers to the contradictory association logic between feature elements, such as “A feature is associated with B feature” and “A feature is not associated with B feature” existing at the same time; influence range conflict refers to the overlapping or contradictory description of the influence range of feature elements.

[0125] Different conflict types are formulated to solve the strategies: expression conflict adopts semantic fusion strategy to integrate different expressions into a unified standard expression; association relationship conflict adopts primary and secondary association strategy to determine the primary and secondary relationship according to the association relationship of the core scene feature elements; influence range conflict adopts range division strategy to clearly define the influence boundary of each feature element.

[0126] Step S1496: Integrate the conflict resolution strategy into the feature conflict resolution rules in the classification basis, and replace the original conflict resolution rule content.

[0127] The formulated conflict resolution strategy is updated to the feature conflict resolution rules in the classification basis, replacing the original rule content that does not conform to the current evolution stage. For example, in the “new device access stage”, if there is an expression conflict, semantic fusion strategy is adopted for processing to ensure that the classification basis can solve the feature conflict according to the rules adapted to the current scene.

[0128] Step S1497: In each round of iterative optimization, based on the current scene feature evolution stage, the corresponding feature element combination rule, feature priority sorting rule and feature conflict resolution rule are called and applied to the classification basis adjustment process.

[0129] Before each round of feature weight iterative optimization, the current scene feature evolution stage is determined, and the feature element combination rule, feature priority sorting rule and feature conflict resolution rule corresponding to the 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 rule of the classification basis is synchronized with the current scene evolution state.

[0130] Step S1410: When the optimization evaluation data of any round reaches the preset optimization target, stop iteration, and determine the classification basis obtained in this round as the optimized classification basis.

[0131] After each round of optimization evaluation, if the classification accuracy, classification recall rate and classification matching degree in the optimization evaluation data all reach the preset optimization target, stop the iteration process, and determine the classification basis obtained by adjusting in this round as the optimized classification basis for subsequent operation and maintenance problem classification matching operation.

[0132] Step S150: Perform multi-level association matching between the optimized classification basis and the dynamically updated operation and maintenance technical service problem classification case library to generate an operation and maintenance technical service problem classification result corresponding to each operation and maintenance problem text unit.

[0133] In this embodiment, the dynamically updated operation and maintenance technical service problem classification case library is the core case resource of the intelligent home control operation and maintenance service, and contains historical accumulated and newly supplemented operation and maintenance problem cases. By performing multi-level matching between the optimized classification basis and the case library, the most suitable case classification result is accurately located to ensure the accuracy and practicality of operation and maintenance problem classification.

[0134] Step S151: Call the dynamically updated operation and maintenance technical service problem classification case library, which contains historical operation and maintenance problem case data and latest operation and maintenance problem case data. The historical operation and maintenance problem case data contains historical case problem text, historical case scene features, historical case classification results and historical case classification basis. The latest operation and maintenance problem case data is newly added operation and maintenance problem case data in recent period of time, which contains latest case problem text, latest case scene features, latest case classification results and latest case classification basis.

[0135] The case base access interface calls the dynamically updated operation and maintenance technical service problem classification case base. The operation and maintenance technical service problem classification case base adopts a distributed storage architecture and is divided into storage areas according to a “historical case area” and a “latest case area”. Historical operation and maintenance problem case data covers operation and maintenance cases in a long period of time in the past, and each piece of data completely records historical case problem text (such as “2024 X month intelligent light linkage failure, triggering condition is human body sensing”), historical case scene features (such as “router is dual-frequency type, linkage rule is ‘human body sensing triggering-light on’”), historical case classification results (such as “control scene configuration class-linkage rule failure”), and historical case classification basis (including feature weight, association rule, and other detailed information). The latest operation and maintenance problem case data is the newly added case in the past month, and the data structure is consistent with that of the historical case, ensuring that the case base can reflect the latest operation and maintenance problem features and classification logic.

[0136] Step S152: The optimized classification basis is subjected to feature hierarchical processing, and the key classification information in the optimized classification basis is divided into a core feature layer, a secondary feature layer, and an auxiliary feature layer according to the association dimension. The core feature layer contains feature content directly associated with the operation and maintenance problem type, the secondary feature layer contains feature content associated with the operation and maintenance problem level, and the auxiliary feature layer contains feature content associated with the operation and maintenance problem processing scheme.

[0137] Based on the fields of “problem type”, “abnormal level”, “processing direction” and the like of the optimized classification basis, feature hierarchical processing is performed. The core feature layer focuses on feature content directly related to the operation and maintenance problem type, such as “WiFi communication abnormality” and “gateway offline” corresponding to “network connection class” problems. These features are the core basis for judging the problem category; the secondary feature layer revolves around the operation and maintenance problem level, such as “affecting a single device” and “can be repaired through remote operation” corresponding to “intermediate abnormality”. These features are used to define the problem severity; and the auxiliary feature layer is associated with the operation and maintenance processing scheme direction, such as “router model” and “network signal strength” corresponding to “check router”. Each layer of features is stored in the form of a vector, retaining the feature weight and association relationship information in the optimized classification basis.

[0138] Step S153: The feature content of the core feature layer is subjected to first-level matching with the core features in the case classification basis of all cases in the dynamically updated operation and maintenance technical service problem classification case base, the core feature similarity is calculated, and a candidate case set with a core feature similarity reaching a first preset threshold is selected.

[0139] The cosine similarity algorithm is used to calculate the similarity of the core feature layer of the optimized classification basis and the core feature of each case in the case library. Before calculation, the core feature vectors of the two are converted to a unified dimension to ensure the consistency of the vector space. For example, the core feature vector of the optimized classification basis is “[WiFi communication anomaly, gateway offline, device offline]”, and the core feature vector of a case is “[WiFi communication anomaly, gateway offline, smart socket offline]”. The core feature similarity is obtained by calculating the cosine value of the angle between the two vectors.

[0140] The calculated core feature similarity is compared with the first preset threshold (such as 0.7) to filter out cases with a similarity reaching the threshold, and a candidate case set is formed by integrating the cases. If the core features of a case and the optimized classification basis core feature are highly coincident, such as both containing “WiFi communication anomaly” and “gateway offline”, the similarity is easy to reach the threshold, and the candidate set is entered.

[0141] Step S154: The feature content of the secondary feature layer is matched with the secondary feature in the case classification basis of each case in the candidate case set at the second level, the secondary feature similarity is calculated, and a secondary candidate case set with a secondary feature similarity reaching a second preset threshold is selected.

[0142] For each case in the candidate case set, the secondary feature in the case classification basis is extracted, and the same cosine similarity algorithm as the core feature matching is used to calculate the similarity of the optimized classification basis secondary feature layer and the case secondary feature. For example, the secondary feature of the optimized classification basis is “[medium-level anomaly, affecting a single smart window, remote repairable]”, and the secondary feature of a candidate case is “[medium-level anomaly, affecting a single smart light, remote repairable]”. The secondary feature similarity is obtained by vector calculation.

[0143] A second preset threshold (such as 0.6) is set to filter out cases with a secondary feature similarity reaching the threshold to form a secondary candidate case set. This level of matching can eliminate cases with similar core features but different anomaly levels and impact ranges, further narrowing the matching range.

[0144] Step S155: The feature content of the auxiliary feature layer is matched with the auxiliary feature in the case classification basis of each case in the secondary candidate case set at the third level, the auxiliary feature similarity is calculated, the comprehensive similarity of each case is generated, and the comprehensive similarity is the weighted sum of the core feature similarity, the secondary feature similarity, and the auxiliary feature similarity.

[0145] The auxiliary features of each case in the secondary candidate case set are extracted, and the similarity between the optimized classification auxiliary features and the case auxiliary features is calculated. For example, the optimized classification auxiliary features are “[check router, restart device, signal strength detection]”, and the auxiliary features of a secondary candidate case are “[check router, restart device, WiFi channel adjustment]”. The auxiliary feature similarity is calculated by vector calculation.

[0146] The core feature similarity, the secondary feature similarity, and the auxiliary feature similarity are assigned weights, wherein the core feature similarity has the highest weight (such as 0.5), the secondary feature similarity has the second highest weight (such as 0.3), and the auxiliary feature similarity has the lowest weight (such as 0.2). The three similarities are multiplied by the corresponding weights and summed to obtain the comprehensive similarity of each case. For example, the core similarity of a case is 0.8, the secondary similarity is 0.7, and the auxiliary similarity is 0.6. The comprehensive similarity is 0.8*0.5+0.7*0.3+0.6*0.2=0.73.

[0147] Step S156: Based on the comprehensive similarity, the cases in the secondary candidate case set are sorted, and the case with the highest comprehensive similarity is selected as the optimal matching case.

[0148] The cases in the secondary candidate case set are sorted in descending order of comprehensive similarity. If there are multiple cases with the same highest comprehensive similarity, the scene features of the cases are further compared with the current operation and maintenance scene features to select the case with higher compatibility as the optimal matching case. For example, if the comprehensive similarity of two cases is 0.85, and the scene features of one of the cases include “router model consistent with the current scene”, this case is selected as the optimal matching case.

[0149] Step S157: The case classification result corresponding to the optimal matching case is extracted, and the case classification result is taken as the preliminary classification result corresponding to the current operation and maintenance problem text unit.

[0150] The case classification result is extracted from the metadata of the optimal matching case, and the case classification result includes the specific category, sub-type, and associated operation and maintenance label to which the problem belongs. For example, the classification result of the optimal matching case is “network connection class-WiFi communication failure-router signal weak”, and this result is taken as the preliminary classification result of the current operation and maintenance problem text unit (such as “smart window curtain APP unresponsive, showing offline”).

[0151] Step S158: The case scene features and case problem text corresponding to the optimal matching case are retrieved and compared with the current operation and maintenance problem text unit and the corresponding scene feature data to confirm the scene adaptability and text consistency.

[0152] The case scenario features of the optimal matching case (such as "smart curtain-WiFi type, router is dual-band A type") and the case problem text (such as "smart curtain APP control is unresponsive, APP displays WiFi connection failure") are compared with the current operation and maintenance problem text unit and the corresponding scene feature data (such as the current smart curtain is WiFi type, the router is dual-band A type, and the problem text is "smart curtain APP is unresponsive, displays offline") in each dimension.

[0153] The scene adaptability comparison focuses on whether the core scene elements such as equipment type, network environment, and control scene are consistent; the text expression consistency comparison focuses on the semantic similarity of the problem phenomenon description, such as the semantic correlation degree of "WiFi connection failure" and "display offline". Through comparison, it is confirmed whether there are obvious differences, such as equipment type inconsistency, network environment difference is too large, etc.

[0154] Step S159: If the scene adaptability and the text expression consistency both meet the preset requirements, the preliminary classification result is determined as the operation and maintenance technical service problem classification result corresponding to each operation and maintenance problem text unit; if not, the case with the second highest comprehensive similarity is selected to repeat the supplementary comparison step until the operation and maintenance technical service problem classification result that meets the requirements is determined.

[0155] The scene adaptability threshold (such as core scene element matching degree ≥0.9) and the text expression consistency threshold (such as semantic similarity ≥0.6) are set. If the comparison results of the optimal matching case all reach the threshold, the preliminary classification result is directly determined as the final operation and maintenance technical service problem classification result.

[0156] If the threshold is not reached, such as the router model of the optimal matching case and the current scene have large differences, the scene adaptability does not meet the requirements, then the case is excluded, the case with the second highest comprehensive similarity is selected as the new candidate case, and the extraction and comparison operations of steps S157 to S158 are repeated. If no case that meets the requirements is found after traversing the secondary candidate case set, the "problem type" in the optimized classification basis is taken as the basic classification result, and a "to be manually reviewed" label is marked, and the classification result is further confirmed by the operation and maintenance personnel in combination with the actual situation.

[0157] Step S160: The dynamic updating of the operation and maintenance technical service problem classification case library is triggered by the number or time interval of newly added operation and maintenance problem classification results. When the number of newly added operation and maintenance problem classification results reaches a preset number threshold, or the time interval from the last update reaches a preset time threshold, the dynamic updating of the operation and maintenance technical service problem classification case library is triggered.

[0158] The dynamic updating trigger mechanism is built in the case library management system, including quantity trigger and time trigger. In the quantity trigger mode, a preset threshold of the number of newly added operation and maintenance problem classification results (such as 50) is set, and the case library management system real-time counts the number of newly added classification results, and automatically triggers the update when the number reaches the threshold. In the time trigger mode, a preset update time interval (such as 7 days) is set, and if the time interval from the last update time has been reached and the quantity update has not been triggered, the system automatically starts the update process. The two trigger modes complement each other to ensure that new cases are timely incorporated into the case library.

[0159] Step S161: After triggering the dynamic updating 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 called, and the operation and maintenance problem classification result data includes newly added operation and maintenance problem text units, corresponding classification results, classification basis and associated scene feature data.

[0160] After triggering the update, the case library management system calls all operation and maintenance problem classification result data generated in the recent period (from the last update to the present) through a data interface. The operation and maintenance problem classification result data is pushed by the operation and maintenance classification module to the data buffer in real time, including newly added operation and maintenance problem text units (such as “smart door lock fingerprint recognition frequent failure, password unlocking normal”), corresponding final classification results (such as “device hardware class-fingerprint recognition module failure”), optimized classification basis and associated scene feature data (such as “smart door lock-fingerprint type, use time 1 year, firmware version V2.0”).

[0161] Step S162: The newly added operation and maintenance problem classification result data is processed for case format conversion, the newly added operation and maintenance problem text units are converted into case problem text format, the associated scene feature data are converted into case scene feature format, the classification results are converted into case classification result format, and the classification basis is converted into case classification basis format, to generate to-be-supplemented case data.

[0162] The case library management system has built-in format conversion templates, which standardize the new data according to the templates. The newly added operation and maintenance problem text units are converted into case problem text format according to the structure of “problem phenomenon + occurrence time + equipment identification”; the associated scene feature data are converted into case scene feature format according to the hierarchical structure of “device feature + network feature + control scene feature”; the classification results are converted into case classification result format according to the three-level structure of “main category-subcategory-label”; and the classification basis retains the original feature layering, weight allocation and other information, and is converted into case classification basis format. After conversion, the to-be-supplemented case data with unified structure is generated.

[0163] Step S163: The generated to-be-supplemented case data is subjected to case uniqueness verification, and the similarity of the to-be-supplemented case data and the case data stored in the dynamically updated operation and maintenance technical service problem classification case library is compared. If the similarity does not reach the case repetition threshold, the to-be-supplemented case data is marked as the supplementable case data.

[0164] The uniqueness verification method based on semantic hashing is adopted to calculate the similarity of the to-be-supplemented case data and the case data stored in the case library. The case problem text and case scene features of the to-be-supplemented case are extracted to generate a semantic hash value. Meanwhile, the corresponding fields of the stored case are extracted to generate a semantic hash value, and the similarity of the two hash values is calculated.

[0165] A preset case repetition threshold (such as 0.85) is set. If the similarity is lower than the threshold, it is determined as a new case and marked as supplementable case data. If the similarity is higher than the threshold, it is determined as a repeated case and discarded and a repeated log is recorded. For example, if the semantic similarity of the problem text and scene features of the to-be-supplemented case and the stored case is 0.7, it is marked as supplementable case data.

[0166] 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 hierarchically divided according to the operation and maintenance problem type, scene category, and classification time.

[0167] The storage structure of the case library adopts a three-level classification directory. The first-level directory is divided according to the operation and maintenance problem type (such as network connection class, device hardware class, and control scene configuration class). The second-level directory is divided according to the scene category under the first-level directory (such as WiFi fault and gateway exception under the network connection class). The third-level directory is divided according to the classification time under the second-level directory (such as by month).

[0168] According to the classification result and classification time of the supplementable case data, it is classified into the corresponding three-level directory. For example, the supplementable case with the classification result of “network connection class-WiFi fault” and the classification time of September 2025 is classified into the “network connection class-WiFi fault-September 2025” directory, and is assigned a unique case ID for indexing.

[0169] Step S165: After supplementing the case data, all case data in the dynamically updated operation and maintenance technical service problem classification case library is subjected to indexing update, and feature indexing is established for the newly added supplementable case data, including case problem text indexing, case scene feature indexing, and case classification result indexing.

[0170] The case base management system starts the index updating process, and adopts the inverted index technology to establish multi-dimensional indexes for the newly added replenishable case data. The case problem text index is constructed based on the keywords in the problem text, such as the keywords "smart door lock" and "fingerprint recognition failure", which are associated with the corresponding case ID; the case scene feature index is constructed according to the feature categories, such as the features "device type-smart door lock" and "network type-WiFi", which are associated with the case ID; and the case classification result index is constructed according to the "main category-subcategory-label" hierarchy, and each hierarchical classification item is associated with the corresponding case ID.

[0171] At the same time, the global index table of the case base is updated to ensure that the newly added cases can be quickly queried through any index dimension. After the index updating is completed, the system performs index verification to ensure the consistency of the index and the case data.

[0172] Step S166: After the index updating is completed, a dynamically updated operation and maintenance technical service problem classification case base update log is generated to record the time of this update, the number of replenished cases, and the case classification distribution, thereby forming a dynamically updated operation and maintenance technical service problem classification case base containing historical operation and maintenance problem case data and the latest operation and maintenance problem case data.

[0173] The update log is generated in a structured format and contains update basic information (update start time, end time, and update trigger mode), replenished case statistics (total number and case number of each problem type), case classification distribution (case proportion under each first-level and second-level directory), and abnormal record (such as the number of duplicate cases and the number of format conversion failure cases).

[0174] The update log is stored in the log management module of the case base, and the replenished replenishable case data and the original historical and latest case data are integrated to form the updated dynamic operation and maintenance technical service problem classification case base.

[0175] Figure 2 A schematic diagram of exemplary hardware and software components of the deep learning-based operation and maintenance technical service problem classification system 100 that can implement the idea of the present application is shown. For example, the processor 120 can be used in the deep learning-based operation and maintenance technical service problem classification system 100 and used to perform the functions in the present application.

[0176] The deep learning-based operation and maintenance technical service problem classification system 100 can be a general server or a special-purpose server, both of which can be used to implement the deep learning-based operation and maintenance technical service problem classification method of the present application. Although only one server is shown in the present application, for the sake of convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0177] For example, the deep learning based operation and maintenance technical service problem classification system 100 can include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, a ROM, or a RAM, or any combination thereof. Illustratively, the deep learning based operation and maintenance technical service problem classification system 100 can also include program instructions stored in a ROM, a RAM, or other types of non-transitory storage media, or any combination thereof. The methods of the present 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 and output devices.

[0178] For ease of illustration, 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 in the present application can also include multiple processors, so the steps performed by one processor described in the present application can also be jointly performed or separately performed by multiple processors. For example, if the processor of the deep learning based operation and maintenance technical service problem classification system 100 performs steps A and B, it should be understood that steps A and B can also be jointly performed by two different processors or separately performed in one processor. For example, a first processor performs step A, a second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0179] In addition, the present application also provides a readable storage medium, wherein computer executable instructions are pre-set in the readable storage medium, and when a processor executes the computer executable instructions, the deep learning based operation and maintenance technical service problem classification method is realized.

[0180] It should be noted that, in order to simplify the expression of the present application and to help understand one or more embodiments of the present application, in the foregoing description of the embodiments of the present application, various features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A method for classifying operation and maintenance technical service problems based on deep learning, characterized in that: The method comprises: Obtaining an operation and maintenance technical service problem description text set, wherein the operation and maintenance technical service problem description text set includes multiple operation and maintenance problem text units, and each operation and maintenance problem text unit is accompanied by collection time information; Perform dynamic update and feature evolution analysis processing of the operation and maintenance scenario feature library to generate scenario feature evolution trajectory data, wherein the operation and maintenance scenario feature library stores scenario feature data corresponding to the operation and maintenance scenario; 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; Based on the historical O&M problem classification results, the corresponding O&M processing effect data, and the scenario feature evolution trajectory data, the feature weights and feature association rules in the initial classification basis are adjusted in stages to obtain the optimized classification basis. The optimized classification basis is multi-level correlated and matched with the dynamically updated operation and maintenance technical service problem classification case library to generate the operation and maintenance technical service problem classification results corresponding to each operation and maintenance problem text unit.

2. The method for classifying operation and maintenance technical service problems based on deep learning according to claim 1 is characterized in that: The execution of the dynamic update of the operation and maintenance scenario feature library and the feature evolution analysis and processing to generate scenario feature evolution trajectory data includes: 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 the 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; Extracting scenario feature elements from business operation log data. The scenario feature elements are expressions that can characterize the characteristics of the current operation and maintenance business scenario, including device type association expressions, business process stage expressions, and abnormal event type expressions; Compare the scene feature elements with the scene feature data stored in the operation and maintenance scene feature library, screen out scene feature elements whose expression similarity with the stored scene feature data does not reach a preset correlation level, and use 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; 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, which contains scene feature data at different time points and corresponding time tags; 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 trend of feature description detail adjustment, and the changing trend of feature correlation relationship; Based on the identified change trends, a feature evolution path is constructed for each scene feature data, and 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 change impact range.

3. The method for classifying operation and maintenance technical service problems based on deep learning according to claim 1 is characterized in that: The method of inputting each operation and maintenance problem text unit, scene feature data, and scene feature evolution trajectory data into a pre-trained deep learning classification model to obtain the initial classification basis for each operation and maintenance problem text unit includes: Input the operation and maintenance problem text unit into the text encoding module of the pre-trained deep learning classification model, perform semantic analysis and vector conversion on the operation and maintenance problem text unit, and generate text semantic feature data corresponding to the operation and maintenance problem text unit. The text semantic feature data includes the core expression vector of the operation and maintenance problem, the abnormal phenomenon description vector, and the related equipment expression vector; Input the scene feature data and the scene feature evolution trajectory data into the cross-scene feature migration module of the pre-trained deep learning classification model, perform feature alignment on the scene feature data and the scene feature evolution trajectory data, and then calculate the migration adaptation coefficient between the scene feature data and the historical training scene feature data; Based on the transfer adaptation coefficient, the scene feature data is adjusted to generate transfer adaptation features that are adapted to the training data distribution of 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 correlation between the transfer adaptation features and the text semantic feature data, including the corresponding relationship between feature elements, the influence relationship of feature weights, and the complementary relationship between feature semantics. Based on the established association relationship, the migration adaptation features and text semantic feature data are weightedly 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, abnormality level corresponding features, and treatment solution matching features. The key classification information is sorted according to the preset classification basis structure to generate the initial classification basis for each operation and maintenance problem text unit.

4. The method for classifying operation and maintenance technical service problems based on deep learning according to claim 3 is characterized in that: The calculation of the migration adaptation coefficient between the scene feature data and the historical training scene feature data includes: Retrieving historical training scene feature data of a pre-trained deep learning classification model, where the historical training scene feature data of the pre-trained deep learning classification model is scene feature data used in the training process of the pre-trained deep learning classification model, and includes feature content and feature association relationships corresponding to multiple training scenes; Extract all feature elements from the current scene feature data to form a 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 a historical feature element set; Calculating the expression similarity between the current feature element set and the corresponding feature elements in the historical feature element set, wherein the expression similarity is determined based on the semantic distance, vocabulary overlap and expression structure similarity of the feature elements, and generating a feature expression similarity set; Analyze the correlation between each feature element in the current scene feature data to form a current feature correlation network, and analyze the correlation between each feature element in the historical training scene feature data of the pre-trained deep learning classification model to form a historical feature correlation network; Calculate the structural similarity between the current feature association network and the historical feature association network, including the similarity of network node connection mode, network hierarchical structure similarity and network core feature position similarity, and generate feature association relationship similarity; Extract the evolution trend parameters of the current scene features based on the scene feature evolution trajectory data, including the feature update frequency, feature content adjustment amplitude, and feature correlation change rate. Extract the evolution trend parameters of the historical scene features based on the time series of the historical training scene feature data of the pre-trained deep learning classification model. Calculate the similarity between the evolution trend parameters of the current scene features and the evolution trend parameters of the historical scene features to generate feature evolution trend similarity; Corresponding calculation weights are assigned to the feature description similarity set, feature association relationship similarity, and feature evolution trend similarity. Based on the calculation weights, the average value of the feature description similarity set, the feature association relationship similarity, and the feature evolution trend similarity are weightedly summed to obtain the migration adaptation coefficient between the scene feature data and the historical training scene feature data of the pre-trained deep learning classification model.

5. The method for classifying operation and maintenance technical service problems based on deep learning according to claim 1 is characterized in that: Based on the historical operation and maintenance problem classification results, the corresponding operation and maintenance processing effect data and the scene feature evolution trajectory data, the feature weights and feature association rules in the initial classification basis are adjusted in stages to obtain the optimized classification basis, including: Retrieve a historical operation and maintenance problem classification result dataset, a corresponding operation and maintenance processing effect dataset, and historical scenario feature evolution trajectory data. The historical operation and maintenance problem classification result dataset includes historical operation and maintenance problem text units, corresponding historical classification results, and classification basis. The operation and maintenance processing effect dataset includes an operation and maintenance processing solution corresponding to the historical classification result, processed 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 on the historical operation and maintenance problem classification result dataset and the operation and maintenance processing effect dataset, and a corresponding relationship is established between each historical classification result and the corresponding business recovery status data and processing efficiency data to generate a classification effect association dataset; Extracting effect features corresponding to the business recovery status data and efficiency features corresponding to the processing efficiency data in the classification effect association data set. The effect features include business recovery integrity, recovery time, and post-recovery stability. The efficiency features include the number of processing steps, processing resource consumption, and processing response speed. The classification basis in the historical classification results and the corresponding effect characteristics, efficiency characteristics and historical scene characteristic evolution trajectory data are input into the iterative optimization module to analyze the correlation strength between the characteristic elements in the classification basis and the effect characteristics and efficiency characteristics, as well as the change pattern of the characteristic elements as the scene characteristics evolve; Determine the initial adjustment direction of each characteristic element in the classification basis based on the association strength, determine the dynamic adjustment range of the characteristic element weight based on the change law, and generate the first round of characteristic weight adjustment plan; Apply the first-round feature weight adjustment scheme to the initial classification basis to obtain the first-round optimized classification basis. Input the first-round optimized classification basis and the corresponding historical operation and maintenance problem text unit into the pre-trained deep learning classification model to generate the first-round simulated classification results. Compare the first round of simulated classification results with the actual classification results in the historical classification results, calculate the classification accuracy, classification recall rate and classification matching degree, and generate the first round of optimization evaluation data; If the first round of optimization evaluation data does not reach the preset optimization target, the parameters of the feature weight adjustment plan are adjusted based on the first round of optimization evaluation data, including the adjustment amplitude coefficient and adjustment direction, and a second round of feature weight adjustment plan is generated, and the weight adjustment, simulation classification and evaluation steps are repeated; During multiple rounds of iterations, the feature association rules in the classification basis are adjusted synchronously based on the scene feature evolution trajectory data, including feature element combination rules, feature priority sorting rules, and feature conflict resolution rules; When any round of optimization evaluation data reaches the preset optimization target, the iteration is stopped and the classification basis obtained in that round is determined as the classification basis after optimization.

6. The method for classifying operation and maintenance technical service problems based on deep learning according to claim 5 is characterized in that: In the multi-round iteration process, the feature association rules in the classification basis are adjusted synchronously based on the scene feature evolution trajectory data, including: The scene feature evolution trajectory data is divided into evolution stages. Based on the change amplitude and frequency of scene features, the evolution process is divided into multiple evolution stages. Each evolution stage corresponds to a continuous time interval and the state of the scene features within the time interval. Extracting the core scene feature elements in each evolution stage, wherein the core scene feature elements are the scene feature elements with the highest frequency of occurrence and the largest number of associated features in the evolution stage; Analyze the combination mode of core scene feature elements and other scene feature elements in each evolution stage, summarize the typical combination patterns of scene feature elements in different evolution stages, and adjust the feature element combination rules in the classification basis based on the typical combination patterns to make the feature element combination rules match the scene feature combination patterns in the current evolution stage; Based on the changes in the importance of the core scene feature elements in each evolution stage, the feature priority sorting rules in the classification basis are adjusted, and the priority of the classification basis feature elements corresponding to the core scene feature elements in the current evolution stage is increased, and the priority of the classification basis feature elements corresponding to the non-core scene feature elements is reduced; Identify the types of conflicts between scene feature elements in each evolutionary stage, including expression conflicts, association conflicts, and impact scope conflicts. Develop 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 association conflicts. Integrate the conflict resolution strategy into the feature conflict resolution rules in the classification basis, replacing the original conflict resolution rule content; In each round of iterative optimization, based on the current scene feature evolution stage, the corresponding feature element combination rules, feature priority sorting rules and feature conflict resolution rules are called and applied to the classification basis adjustment process.

7. The method for classifying operation and maintenance technical service problems based on deep learning according to claim 1 is characterized in that: The optimized classification basis is matched with the dynamically updated operation and maintenance technical service problem classification case library at multiple levels to generate the operation and maintenance technical service problem classification result corresponding to each operation and maintenance problem text unit, including: Retrieve a dynamically updated operation and maintenance technical service problem classification case library, wherein the dynamically updated operation and maintenance technical service problem classification case library includes historical operation and maintenance problem case data and the latest operation and maintenance problem case data, wherein the historical operation and maintenance problem case data includes historical case problem texts, historical case scenario features, historical case classification results, and historical case classification basis, and the latest operation and maintenance problem case data is operation and maintenance problem case data newly added in a recent period, including the latest case problem texts, the latest case scenario features, the latest case classification results, and the latest case classification basis; Perform feature stratification processing on the optimized classification basis, dividing the key classification information in the optimized classification basis into a core feature layer, a secondary feature layer, and an auxiliary feature layer according to the associated dimensions. The core feature layer contains feature content directly associated with the operation and maintenance problem type, the secondary feature layer contains feature content associated with the operation and maintenance problem level, and the auxiliary feature layer contains feature content associated with the operation and maintenance problem solution; Performing a first-level match between the feature content of the core feature layer and the core features in the case classification basis of all cases in the dynamically updated operation and maintenance technical service problem classification case library, calculating the core feature similarity, and screening out a set of candidate cases whose core feature similarity reaches a first preset threshold; Performing a second-level matching between the feature content of the secondary feature layer and the secondary features in the case classification basis of each case in the candidate case set, calculating the secondary feature similarity, and screening out the secondary candidate case set whose secondary feature similarity reaches a second preset threshold; Performing a third-level matching between the feature content of the auxiliary feature layer and the auxiliary features in the case classification basis of each case in the secondary candidate case set, calculating the auxiliary feature similarity, and generating a comprehensive similarity for each case, wherein the comprehensive similarity is a weighted sum of the core feature similarity, the secondary feature similarity, and the auxiliary feature similarity; Sort the cases in the secondary candidate case set based on comprehensive similarity, and select the case with the highest comprehensive similarity as the best 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 text expression consistency; If both the scenario adaptability and the text expression consistency meet the preset requirements, the preliminary classification result will be determined as the operation and maintenance technical service problem classification result 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 steps will be repeated until the operation and maintenance technical service problem classification result that meets the requirements is determined.

8. The method for classifying operation and maintenance technical service problems based on deep learning according to claim 7 is characterized in that: The dynamically updated operation and maintenance technical service problem classification case library is retrieved, including: Establish a dynamic update trigger mechanism for the dynamically updated operation and maintenance technical service problem classification case library. The dynamic update trigger mechanism for the dynamically updated operation and maintenance technical service problem classification case library triggers an update operation based on the number or time interval of newly added operation and maintenance problem classification results. When the number of newly added operation and maintenance problem classification results reaches a preset number threshold, or the time interval from the last update reaches a preset time threshold, the dynamically updated operation and maintenance technical service problem classification case library is triggered to update; After triggering the update operation of the dynamically updated operation and maintenance technical service problem classification case library, the operation and maintenance problem classification result data generated in the recent period is retrieved, wherein the operation and maintenance problem classification result data includes the newly added operation and maintenance problem text unit, the corresponding classification result, the classification basis and the associated scenario feature data; 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 the case problem text format, convert the associated scenario feature data into the case scenario feature format, convert the classification result into the case classification result format, and convert the classification basis into the case classification basis format to generate the case data to be supplemented; 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 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; The supplementable case data is classified 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 directory is hierarchically divided according to the operation and maintenance problem type, scenario category and classification time; After the case data is supplemented, all case data in the dynamically updated operation and maintenance technical service problem classification case library are indexed and 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 operation and maintenance technical service problem classification case library update log is generated to record the time of this update, the number of supplemented cases and the case classification distribution, forming a dynamically updated operation and maintenance technical service problem classification case library that includes historical operation and maintenance problem case data and the latest operation and maintenance problem case data.

9. The method for classifying operation and maintenance technical service problems based on deep learning according to claim 1 is 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 an initial classification basis for each operation and maintenance problem text unit further includes: Before calling the pre-trained deep learning classification model, pre-adapt the cross-scene feature migration module and the feature fusion module of the pre-trained deep learning classification model, and retrieve the cross-scene migration parameters and feature fusion parameters used in the training process of the pre-trained deep learning classification model; Input the scene feature data sample of the current operation and maintenance scene into the cross-scene feature migration module of the pre-trained deep learning classification model, test the matching degree between the migration adaptation features output by the cross-scene feature migration module of the pre-trained deep learning classification model and the expected features of the sample, and adjust the migration coefficient and feature adjustment weight in the cross-scene migration parameters if the matching degree does not reach the preset adaptation threshold; Apply the adjusted cross-scene migration parameters to the cross-scene feature migration module of the pre-trained deep learning classification model, and re-input the scene feature data sample for testing until the matching degree between the migrated adaptation features and the expected features of the sample reaches the preset adaptation threshold; Input the text semantic feature data samples and the transfer adaptation feature samples into the feature fusion module of the pre-trained deep learning classification model, test the matching degree between the fusion feature data output by the feature fusion module of the pre-trained deep learning classification model and the expected fusion features, and if the matching degree does not reach the preset fusion threshold, adjust the fusion weight distribution rule and association algorithm parameters in the feature fusion parameters; Apply the adjusted feature fusion parameters to the feature fusion module of the pre-trained deep learning classification model, re-input the text semantic feature data sample and the transfer adaptation feature sample for testing until the matching degree between the fused feature data and the expected fusion feature reaches the preset fusion threshold; After completing the pre-adaptation process, save the adjusted cross-scene migration parameters and feature fusion parameters and use them as the operating parameters of the currently called pre-trained deep learning classification model, and then perform subsequent feature migration adaptation and association fusion steps.

10. A deep learning-based operation and maintenance technical service problem classification system, characterized by: The deep learning-based operation and maintenance technical service problem classification system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the deep learning-based operation and maintenance technical service problem classification method described in any one of claims 1 to 9 above.

Images