Work order fault intelligent processing and dynamic recommendation method and device

Abstract

The invention relates to the technical field of intelligent business process management, and particularly provides a work order fault intelligent processing and dynamic recommendation method and device, and the method comprises the following steps: S1, driving a process engine; s2, large model data processing and optimization; and S3, dynamically constructing the knowledge base and the small model. Compared with the prior art, the service processing time can be shortened by 50%, the recommendation accuracy can be improved by 35%, the knowledge base is automatically updated every week, new problem types are covered, small model memory occupation is reduced by 70%, and the method is suitable for edge device deployment.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent business process management, and specifically provides a method and device for intelligent processing and dynamic recommendation of work order faults. Background Art

[0002] Traditional business process systems have the following defects:

[0003] High reliance on manual labor: Business orders require manual input of problem descriptions, lacking intelligent classification and recommendations. Expert experience is difficult to accumulate and pass on, and key decisions rely on a small number of senior personnel.

[0004] Static knowledge base: Historical data is not effectively used to optimize recommendation strategies, and knowledge updates rely on manual maintenance, resulting in knowledge obsolescence and redundancy, making it unable to adapt to rapid business changes.

[0005] Weak model capabilities: Existing systems are unable to improve recommendation accuracy through data iteration. Algorithms generally use simple rule matching and lack deep learning and semantic understanding capabilities.

[0006] Strong process rigidity: Once a business process is established, it is difficult to flexibly adjust it. It is impossible to dynamically optimize the path according to actual conditions, resulting in wasted resources and low efficiency.

[0007] System silos: Data barriers between business systems are obvious, preventing effective information sharing and transfer, leading to duplication of work and consistency issues.

[0008] Efficiency bottleneck: As business volume grows, the scalability of traditional process systems is limited, and processing capacity grows linearly, making it difficult to meet the needs of exponential business growth.

[0009] How to solve the problems of intelligent decision-making and process optimization during the digital transformation of enterprises is an urgent issue that needs to be solved by technical personnel in this field. Summary of the Invention

[0010] The present invention aims to address the deficiencies of the above-mentioned prior art and provides a highly practical method for intelligently processing and dynamically recommending work order faults.

[0011] A further technical task of the present invention is to provide a work order fault intelligent processing and dynamic recommendation device with a reasonable design, safety and applicability.

[0012] The technical solution adopted by the present invention to solve its technical problem is:

[0013] A method for intelligently processing and dynamically recommending work order faults includes the following steps:

[0014] S1, process engine driver;

[0015] S2, large model data processing and optimization;

[0016] S3, knowledge base and small model are dynamically constructed.

[0017] Furthermore, in step S1, the user inputs a problem description and calls the NLP model for semantic analysis. Based on the analysis results, similar cases are screened from the historical database, and problem types and solutions are recommended. The lightweight model provides real-time solution priority sorting.

[0018] Furthermore, the recommended solutions and processing steps are displayed, and the knowledge base can be called up with one click to supplement information. Operations and maintenance personnel mark the effectiveness of the solutions, and the data is transmitted back to the big model in real time.

[0019] Invalid data is automatically filtered out and stored in the database according to the dimensions of problem type, resolution time and customer rating. When the data volume reaches the threshold, the large model training process is automatically started.

[0020] Furthermore, step S2 includes two parts: data regression analysis and natural language processing. The data regression analysis includes:

[0021] (1) Feature engineering: converting user feedback into binary label extraction scheme feature vectors to incorporate environmental features;

[0022] (2) Model training: L2 regularized logistic regression is used to prevent overfitting, cross-validation is used to determine the optimal regularization parameter C, and stratified sampling is used to ensure balanced data for each type of problem.

[0023] (3) Use the ROC curve to analyze the model performance, calculate the importance coefficient of each feature, dynamically adjust the recommendation weight, and convert the weight coefficient into a recommendation score correction factor:

[0024] High-quality solution, accuracy rate>90%: basic score × (1+0.2);

[0025] Conventional solution, accuracy rate 70%-90%: basic score remains unchanged;

[0026] Inefficient solution, accuracy <70%: basic score × (1-0.1);

[0027] (4) Real-time application mechanism:

[0028] Redis caches the latest weight coefficients, sets a sliding window, dynamically updates the weight model, and locks the weights during business peak periods.

[0029] Furthermore, the natural language processing includes:

[0030] A1. Data preparation and preprocessing;

[0031] A2, large model training cycle;

[0032] A3,knowledge distillation process;

[0033] A4. Model evaluation and deployment;

[0034] A5. Continuous optimization cycle.

[0035] Furthermore, in A1, new business data is automatically extracted at regular intervals, and samples with missing labels, outliers, and low-quality samples are removed. Oversampling techniques (SMOTE) and data augmentation are applied to rare problem types.

[0036] In A2, the entire model is retrained using the accumulated dataset once a month, and small-batch incremental learning is performed every week, updating only the parameters of the last few layers. A progressive learning rate and early stopping mechanism are used, and a GPU cluster is used to automatically execute training tasks at night for breakpoint resumption.

[0037] Furthermore, in A3, the large model serves as a teacher and TinyBERT as a student, optimizing both soft label matching and intermediate layer representations. T=2.5 is set to smooth the soft label distribution, enhance knowledge transfer, and perform distillation in stages. In the first stage, general language understanding ability transfer is performed, in the second stage, domain-specific knowledge transfer is performed, and in the third stage, task-specific fine-tuning is performed.

[0038] Furthermore, in A4, offline evaluation, online evaluation, phased release, and rollback mechanisms are included. The offline evaluation includes accuracy, recall, and F1 score.

[0039] The online evaluation question is to compare the effects of new and old models through A / B testing;

[0040] The grayscale release will first provide new model services to 10% of users and gradually expand the coverage;

[0041] The rollback mechanism is to set a performance monitoring threshold.

[0042] Furthermore, in A5, user operation trajectories, clicks, and satisfaction evaluations are recorded, model failure cases are regularly analyzed, and improvement topics are formed. After the model is updated, the knowledge base index and label system are automatically updated, and the model parameters of the last three versions are retained for quick switching.

[0043] Furthermore, in step S3, keywords are extracted based on the TF-IDF algorithm, a fast retrieval index is constructed to build a knowledge base, and the large model is compressed into a lightweight TinyBERT using knowledge distillation technology, and the lightweight TinyBERT is deployed.

[0044] A device for intelligent processing and dynamic recommendation of work order faults, comprising: at least one memory and at least one processor;

[0045] The at least one memory is configured to store a machine-readable program;

[0046] The at least one processor is used to call the machine-readable program to execute a method for intelligent processing and dynamic recommendation of work order faults.

[0047] Compared with the prior art, the method and device for intelligent processing and dynamic recommendation of work order faults of the present invention have the following outstanding beneficial effects:

[0048] The business processing time of the present invention is shortened by 50%, the recommendation accuracy is improved by 35%, the knowledge base is automatically updated every week to cover new problem types, the memory usage of small models is reduced by 70%, and it is suitable for edge device deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 It is a flowchart of an intelligent processing and dynamic recommendation method for work order faults. DETAILED DESCRIPTION

[0051] In order to enable those skilled in the art to better understand the solutions of the present invention, the present invention will be further described in detail below in conjunction with specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0052] A best embodiment is given below:

[0053] like Figure 1 As shown, a method for intelligently processing and dynamically recommending work order faults in this embodiment includes the following steps:

[0054] S1, process engine driver;

[0055] The user enters a problem description (text or voice), and the system calls the NLP model for semantic analysis. Based on the analysis results, similar cases are screened from the historical database, and problem types and solutions are recommended (for example, "PON resource anomaly - contact the manufacturer for resource confirmation and data supplementation"). The lightweight model provides real-time solution priority ranking (for example, TOP3 recommendation).

[0056] Recommended solutions and processing steps are displayed, with one-click access to supplementary information from the knowledge base (e.g., equipment configuration manuals). Operations and maintenance personnel can mark solution validity, with data transmitted back to the large model module in real time. Invalid data (e.g., duplicate cases, uncompleted work orders) is automatically filtered out.

[0057] The data is stored in the database according to dimensions such as problem type, resolution time, and customer rating; when the training is triggered and the data volume reaches the threshold, the large model training process is automatically started.

[0058] S2, large model data processing and optimization;

[0059] It includes two parts: data regression analysis and natural language processing. In data regression analysis, random forest and XGBoost algorithms are used to predict problem-solving time and resource allocation priority. The effectiveness of the solution is analyzed based on logistic regression, and the recommendation weight is optimized (for example: the weight of the high-scoring solution is +20%).

[0060] The specific operations are:

[0061] (1) Feature engineering: convert user feedback (satisfaction rating 1-5 stars) into binary labels (≥4 stars is an effective solution) to extract solution feature vectors (dimensions such as solution time, number of operation steps, resource consumption, etc.) and incorporate them into environmental features (time period when the problem occurred, device type, and customer level).

[0062] (2) Model training: Use L2 regularized logistic regression to prevent overfitting, cross-validation to determine the optimal regularization parameter C (recommended value: 0.1-10), and stratified sampling to ensure balanced data for various problem types (SMOTE technology is used to handle rare categories).

[0063] (3) Evaluation and optimization: Use ROC curve to analyze model performance (target AUC>0.85), calculate the importance coefficient of each feature, dynamically adjust the recommendation weight, and convert the weight coefficient into a recommendation score correction factor:

[0064] High-quality solution (accuracy>90%): basic score × (1+0.2);

[0065] Conventional plan (accuracy 70%-90%): the basic score remains unchanged;

[0066] Inefficient solution (accuracy <70%): basic score × (1-0.1);

[0067] (4) Real-time application mechanism: Redis caches the latest weight coefficients to achieve millisecond-level response, sets a sliding window (configurable from 7 days to 30 days), dynamically updates the weight model, and locks the weight during business peak hours to prevent model fluctuations from affecting user experience.

[0068] In natural language processing, the BERT model is used to semantically encode problem descriptions, construct a problem-solution matching matrix, and use clustering algorithms (such as K-means) to mine potential problem categories and expand the knowledge base label system.

[0069] Regular incremental training to update model parameters;

[0070] The specific steps are as follows:

[0071] A1. Data preparation and preprocessing;

[0072] New business data (work orders, feedback, operation records) are automatically extracted every 24 hours, and samples with missing labels, outliers, and low-quality samples (accounting for <5%) are eliminated. Oversampling technology (SMOTE) and data enhancement (synonymous sentence generation) are applied to rare question types.

[0073] A2, large model training cycle;

[0074] The entire model is retrained using the accumulated dataset once a month, and small-batch incremental learning is performed weekly, updating only the parameters of the last few layers. A progressive learning rate (0.0001-0.00001) and early stopping mechanism (validation set loss increases three consecutive times) are used. A GPU cluster is used to automatically execute training tasks at night, and breakpoint resumption is supported.

[0075] A3,knowledge distillation process;

[0076] The large model (BERT-base, 110M parameters) is used as the teacher and TinyBERT (14M parameters) as the student. Both soft label matching (KL divergence) and intermediate layer representation (MSE loss) are optimized. T = 2.5 is set to smooth the soft label distribution and enhance knowledge transfer.

[0077] Distillation in stages:

[0078] The first stage is the transfer of general language comprehension ability (using general corpus);

[0079] The second phase is domain-specific knowledge transfer (using business datasets);

[0080] The third phase is task-specific fine-tuning (using annotated work order data).

[0081] A4. Model evaluation and deployment:

[0082] Offline evaluation: precision, recall, F1 score (target: F1>0.85);

[0083] Online evaluation: A / B testing to compare the effectiveness of new and old models (comparing user satisfaction and resolution time);

[0084] Grayscale release: First provide new model services to 10% of users, and gradually expand coverage;

[0085] Rollback mechanism: Set performance monitoring thresholds and automatically roll back if the recommendation accuracy drops by more than 5%.

[0086] A5. Continuous optimization cycle;

[0087] Feedback collection: record user operation traces, clicks and satisfaction evaluations;

[0088] Error analysis: Regularly analyze model failure cases and form improvement topics;

[0089] Knowledge base synchronization: automatically update the knowledge base index and label system after the model is updated;

[0090] Version management: retain model parameters of the last three versions and support quick switching.

[0091] S3, dynamic construction of knowledge base and small models;

[0092] B1. Knowledge base construction:

[0093] Index optimization, extract keywords based on the TF-IDF algorithm and build a fast search index;

[0094] Version management supports the coexistence of multiple versions of knowledge bases (such as: V1.0 basic version, V2.0 enhanced version).

[0095] B2. Small model generation:

[0096] Model compression, using knowledge distillation technology to compress large models (such as BERT) into lightweight TinyBERT;

[0097] Edge deployment, small model embedding process, supports real-time recommendation in offline environment.

[0098] Based on the above method, a work order fault intelligent processing and dynamic recommendation device in this embodiment includes: at least one memory and at least one processor;

[0099] The at least one memory is configured to store a machine-readable program;

[0100] The at least one processor is used to call the machine-readable program to execute a method for intelligent processing and dynamic recommendation of work order faults.

[0101] The above-mentioned specific implementation methods are only specific cases of the present invention. The patent protection scope of the present invention includes but is not limited to the above-mentioned specific implementation methods. Any technical solutions that conform to the above-mentioned specific implementation methods of the present invention and any appropriate changes or substitutions made thereto by ordinary technicians in the relevant technical field shall fall within the patent protection scope of the present invention.

[0102] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent processing and dynamic recommendation of work order faults, characterized in that: The steps are as follows: S1, process engine driver; S2, large model data processing and optimization; S3, knowledge base and small model are dynamically constructed.

2. A method for intelligent processing and dynamic recommendation of work order faults according to claim 1, characterized in that: In step S1, the user enters a problem description and calls the NLP model for semantic analysis. Based on the analysis results, similar cases are screened from the historical database, and problem types and solutions are recommended. The lightweight model provides real-time solution priority sorting.

3. A method for intelligent processing and dynamic recommendation of work order faults according to claim 2, characterized in that: Recommended solutions and processing steps are displayed, and the knowledge base can be called up for supplementary information with one click. Operations and maintenance personnel mark the effectiveness of the solutions, and data is transmitted back to the big model in real time. Invalid data is automatically filtered out and stored in the database according to the dimensions of problem type, resolution time and customer rating. When the data volume reaches the threshold, the large model training process is automatically started.

4. A method for intelligent processing and dynamic recommendation of work order faults according to claim 3, characterized in that: In step S2, there are two parts: data regression analysis and natural language processing. In the data regression analysis, it includes: (1) Feature engineering: converting user feedback into binary label extraction scheme feature vectors to incorporate environmental features; (2) Model training: L2 regularized logistic regression is used to prevent overfitting, cross-validation is used to determine the optimal regularization parameter C, and stratified sampling is used to ensure balanced data for each type of problem. (3) Use the ROC curve to analyze the model performance, calculate the importance coefficient of each feature, dynamically adjust the recommendation weight, and convert the weight coefficient into a recommendation score correction factor: High-quality solution, accuracy rate>90%: basic score × (1+0.2); Conventional solution, accuracy rate 70%-90%: basic score remains unchanged; Inefficient solution, accuracy <70%: basic score × (1-0.1); (4) Real-time application mechanism: Redis caches the latest weight coefficients, sets a sliding window, dynamically updates the weight model, and locks the weights during business peak periods.

5. A method for intelligent processing and dynamic recommendation of work order faults according to claim 4, characterized in that: The natural language processing includes: A1. Data preparation and preprocessing; A2, large model training cycle; A3,knowledge distillation process; A4. Model evaluation and deployment; A5. Continuous optimization cycle.

6. A method for intelligent processing and dynamic recommendation of work order faults according to claim 5, characterized in that: In A1, new business data is automatically extracted at regular intervals, and samples with missing labels, outliers, and low-quality samples are removed. Oversampling techniques (SMOTE) and data augmentation are applied to rare problem types. In A2, the entire model is retrained using the accumulated dataset once a month, and small-batch incremental learning is performed every week, updating only the parameters of the last few layers. A progressive learning rate and early stopping mechanism are used, and a GPU cluster is used to automatically execute training tasks at night for breakpoint resumption.

7. A method for intelligent processing and dynamic recommendation of work order faults according to claim 6, characterized in that: In A3, the large model serves as a teacher and TinyBERT as a student, optimizing both soft label matching and intermediate layer representations. Setting T = 2.5 smoothes the soft label distribution, enhances knowledge transfer, and performs distillation in stages. In the first stage, general language understanding capabilities are transferred, in the second stage, domain-specific knowledge is transferred, and in the third stage, task-specific fine-tuning is performed.

8. A method for intelligent processing and dynamic recommendation of work order faults according to claim 7, characterized in that: A4 includes offline evaluation, online evaluation, phased release, and rollback mechanisms. The offline evaluation includes accuracy, recall, and F1 score. The online evaluation question is to compare the effects of new and old models through A / B testing; The grayscale release will first provide new model services to 10% of users and gradually expand the coverage; The rollback mechanism is to set a performance monitoring threshold.

9. A method for intelligent processing and dynamic recommendation of work order faults according to claim 8, characterized in that: In A5, user operation trajectories, clicks, and satisfaction ratings are recorded, model failure cases are regularly analyzed, and improvement topics are formed. After the model is updated, the knowledge base index and label system are automatically updated, and the model parameters of the last three versions are retained for quick switching.

10. A method for intelligent processing and dynamic recommendation of work order faults according to claim 9, characterized in that: In step S3, keywords are extracted based on the TF-IDF algorithm, a fast retrieval index is constructed to build the knowledge base, and the knowledge distillation technology is used to compress the large model into a lightweight TinyBERT, which is then deployed.

11. A device for intelligent processing and dynamic recommendation of work order faults, characterized in that: include: at least one memory and at least one processor; The at least one memory is configured to store a machine-readable program; The at least one processor is configured to call the machine-readable program to execute the method according to any one of claims 1 to 10.

Images