Full-channel work order repeated appeal automatic checking method

Abstract

The invention relates to the technical field of power supply services, in particular to a method for automatically checking repeated appeals of all-channel work orders. The method comprises the following steps: S1, constructing an omni-channel work order information base; s2, querying customer appeals in the acceptance content through a natural language processing technology, and screening target customers through screening the total number of work orders; s3, screening sensitive customers by taking the customer telephone number as a screening target; s4, setting a query condition through a select query statement of the object relational database management system, and querying a historical work order of the sensitive customer; s5, in combination with historical work order records, complaint problems are classified and arranged by introducing time intervals and using a random forest model optimized by an XGBoost algorithm; and S6, displaying the repeated appeals of the sensitive customers according to a work order type classification interface. Work order information is collected through multiple channels, and comprehensiveness and diversity of data are ensured. PRA grabbing and manual importing are combined, so that the efficiency and accuracy of data acquisition are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power supply services, and in particular to a method for automatically troubleshooting repeated requests for omni-channel work orders. Background Art

[0002] The statistics of repeated customer calls aims to accurately solve customer needs, find the root causes and influencing factors of the problem, and avoid the problem from happening again. It is an important part of providing quality service. In the past, finding repeated customer demands required manually exporting all customer demand work orders, involving multiple systems such as allocation and grabbing, Marketing 2.0, 12345 Mayor Hotline, and fault repair. The current work order and the historical work order were compared one by one to see if there was the same customer account number or mobile phone number, and the customer repair address and customer demands were manually analyzed and judged, and the work orders with repeated repair reports were screened with the related historical work order information. The work process was mechanical, cumbersome, and had a certain degree of repetitiveness.

[0003] An automatic troubleshooting method for repeated requests for omni-channel work orders, which integrates data from multiple systems such as marketing, distribution, and dispatching, automatically compares and analyzes omni-channel work orders received the previous day with historical work orders. By simply clicking on the query time period, the number of times customers repeatedly call 95598, 12398, 12345, etc. can be clearly seen, and historical requests can be linked with one click, eliminating the need for manual export of work orders and repeated judgment and analysis, greatly improving analysis efficiency. It is especially suitable for scenarios such as when customers report hundreds of repairs in severe weather, and during special and important sensitive periods such as major power outages. It achieves fast, simple, and accurate statistics of customers' repeated requests, timely strategy adjustments, and rapid triggering of corresponding intervention measures, predicting what customers want and need, reasonably avoiding service risks, and improving customer management levels.

[0004] In an existing method for automatically troubleshooting duplicate claims in omni-channel work orders, the existing system does not have a deep enough understanding of semantics, especially when faced with complex or unconventional expressions, it is easy to make misjudgments or missed judgments. This is mainly because users have a variety of ways of expressing themselves, and the data set trained by the system may not be able to fully cover all situations. There are large differences in the information formats and user behavior patterns of different channels, which makes it difficult for a single model to adapt to the needs of multi-channel claim troubleshooting. It is necessary to customize the model for each channel separately, which increases the complexity and maintenance cost of the system. Due to privacy protection and technical limitations, the amount of data and data dimensions obtained by the system are limited, which limits the learning ability and prediction accuracy of the model. The existing automatic troubleshooting methods are slow to respond to new types of problems or emergencies, and require manual intervention to update the model or adjust the parameters, which to some extent weakens the advantages of automation. Therefore, a method for automatically troubleshooting duplicate claims in omni-channel work orders is provided. Summary of the invention

[0005] The purpose of the present invention is to provide an automatic troubleshooting method for duplicate claims in omni-channel work orders, so as to solve the problem raised in the above background technology that in an existing automatic troubleshooting method for duplicate claims in omni-channel work orders, the existing system does not have enough depth of semantic understanding, especially when facing complex or unconventional expressions, it is easy to misjudge or miss the case, and there are large differences in information formats and user behavior patterns of different channels, which makes it difficult for a single model to adapt to the demand troubleshooting of multiple channels, and it is necessary to customize the model for each channel separately, which increases the complexity of the system and the maintenance cost.

[0006] To achieve the above object, the present invention provides a method for automatically troubleshooting duplicate requests for omni-channel work orders, comprising the following steps:

[0007] S1. Build an omni-channel work order information database;

[0008] Among them, the work order information database includes the number of telephone complaints, the total number of work orders, and the content of acceptance;

[0009] S2. Use natural language processing technology to query customer demands in the accepted content and select target customers by screening the total number of work orders;

[0010] S3. Use customer phone numbers as screening targets to screen sensitive customers;

[0011] S4. Set query conditions through the select query statement of the object-relational database management system and query the historical work orders of sensitive customers;

[0012] S5. Combined with historical work order records, the complaint issues are classified and arranged by introducing time intervals and using the random forest model optimized by the XGBoost algorithm;

[0013] S6. Display the repeated demands of this sensitive customer on the work order type classification interface.

[0014] As a further improvement of the technical solution, in S1, the specific steps of building an omni-channel work order information database are:

[0015] S1.1. Use the PRA system to automatically capture work order data from 95598, 12398, and 12345 telephone channels;

[0016] S1.2. For data that cannot be automatically captured by the PRA system, a manual import process is established to ensure the integrity and accuracy of the data.

[0017] As a further improvement of the technical solution, in S2, the specific steps of querying customer demands in the accepted content by natural language processing technology and screening target customers by screening the total number of work orders are as follows:

[0018] S2.1, build a sensitive vocabulary database;

[0019] Among them, the sensitive word library does not include the overdue payment keyword;

[0020] S2.2. Use natural language processing technology to analyze the acceptance content in the work order;

[0021] S2.3. Count the number of work orders for each customer and filter out work order records that have been submitted twice or more.

[0022] As a further improvement of the technical solution, in S2.2, the specific steps of using natural language processing technology to analyze the acceptance content in the work order are:

[0023] S2.21. Preprocess the work order content;

[0024] Among them, the average length of each work order is L characters, and the content length after preprocessing is reduced to L′;

[0025] Compression ratio CR after preprocessing:

[0026]

[0027] For N work orders, the total preprocessing time is:

[0028] T prre N × O (f (L));

[0029] Where N represents the number of work orders; T pre represents the total preprocessing time; O(f(L)) represents the preprocessing time complexity;

[0030] S2.22, matching the preprocessed work order content with the sensitive word library to find the work orders containing sensitive words;

[0031] Among them, the sensitive word library contains V words, and each work order needs to be checked W times on average;

[0032] Matching calculation amount Z:

[0033] Z = V × W;

[0034] S2.23. Check whether the work order content contains keywords related to arrears;

[0035] Among them, the keyword library related to arrears contains K keywords, and each work order needs to be checked X times on average;

[0036] Calculation amount of overdue payment keyword check C:

[0037] C = K × X;

[0038] S2.24. Filter out qualified work orders based on the results of sensitive word matching and overdue payment keyword exclusion;

[0039] Among them, M is the total number of initial work orders, M s is the number of work orders remaining after sensitive word matching, M r The number of work orders that meet the conditions that are finally screened out;

[0040] Screening ratio:

[0041]

[0042] Among them, P represents the screening ratio;

[0043] Screening efficiency:

[0044]

[0045] Where E represents the screening efficiency.

[0046] As a further improvement of the technical solution, in S4, the specific steps of setting the query conditions through the select query statement of the object-relational database management system and querying the historical work orders of sensitive customers are:

[0047] S4.1. Create a work order table in the database;

[0048] S4.2. Write an SQL query statement in the object-relational database management system, using the user number and mobile phone number as query conditions to query the user's historical work order records from the work order table;

[0049] S4.3. Preprocess the query results and perform data analysis.

[0050] As a further improvement of the technical solution, in S4.3, the specific steps of preprocessing the query results and performing data analysis are as follows:

[0051] S4.31. Extract key information of the work order record from the query results and format it;

[0052] Among them, the key information of the work order record includes the work order ID, account number, mobile phone number, acceptance content and creation time;

[0053] S4.32, merge the work order records and vibration data records in chronological order;

[0054] S4.33, calculating the time interval between adjacent records;

[0055] The time interval between adjacent records is calculated as:

[0056]

[0057] Where V represents the total number of customers; M i represents the number of work orders of the i-th customer; i represents the index variable; T interval Indicates the time required for each time interval; A time representing the total time interval.

[0058] As a further improvement of this technical solution, combined with historical work order records, the specific steps of classifying and arranging complaint issues by introducing time intervals and using the random forest model optimized by the XGBoost algorithm are as follows:

[0059] S5.1, extracting work order record features;

[0060] The specific steps for extracting work order record features are as follows:

[0061] Extract feature vector X=[X 1 , X 2 , ..., X n ];

[0062] S5.2, obtaining the time interval between adjacent records;

[0063] The expression for obtaining the time interval between adjacent records is:

[0064] Δt i =t i+1 —t i ;

[0065] Among them, t i Indicates the timestamp of the i-th work order; Δt i Indicates the time interval characteristics; Δt i+1 Indicates the timestamp of the i+1 work order;

[0066] S5.3, labeling based on keywords in the work order content;

[0067] The process of labeling based on keywords in the work order content is as follows:

[0068] Generate a label y for each work order based on the content keywords:

[0069] Establish a keyword set K = {k 1 , k 2 , …, k m};

[0070] For work order text T i Process and mark whether it contains keywords:

[0071] y i =f(Ti , K);

[0072] Where f represents the function of text matching and word vector similarity;

[0073] S5.4. Build a random forest model;

[0074] S5.5, by introducing time interval features and using XGBoost to optimize the random forest model;

[0075] S5.6. Use the optimized random forest model to classify and arrange the complaints and output the results of repeated complaints.

[0076] As a further improvement of the technical solution, in S5.4, the specific process of constructing the random forest model is:

[0077] S5.41, training data preparation;

[0078] The training data preparation is as follows:

[0079] Prepare dataset D:

[0080]

[0081] Among them, x i Represents the characteristics of the i-th record; y i represents its label; N represents the total number of samples in the dataset D;

[0082] S5.42. Generate M sub-datasets by random sampling, each of which is N′ in size. Each sub-dataset randomly selects a feature subset for training and constructs a decision tree T j ;

[0083] S5.43, Generation of decision trees;

[0084] The generation of decision tree is as follows:

[0085] Decision Tree T j The generation of is based on recursive partitioning, selecting feature X k and threshold θ to minimize the Gini index Gini(S):

[0086]

[0087] Among them, S represents the sample set in the node, p c represents the sample proportion of category c; c represents the classification category; C represents the total number of categories in the classification task;

[0088] S5.44. Combine the output results of all decision trees into a classification result by voting;

[0089] Among them, the classification results for:

[0090]

[0091] Among them, Mode represents the mode operation; T M (x) represents the predicted output of the Mth decision tree among the M decision trees generated by the random forest algorithm for the input X;

[0092] S5.45. Use the test set to evaluate the performance of the random forest model.

[0093] As a further improvement of the technical solution, in S5.5, by introducing the time interval feature and using XGBoost to optimize the random forest model, the specific process involved is:

[0094] The time interval feature is taken as part of the input data and combined with other work order features X i Together they form the new input vector:

[0095]

[0096] in, represents the input vector that introduces the time interval feature;

[0097] Use XGBoost to optimize the model. XGBoost is based on gradient boosting decision trees for optimization.

[0098] Define the objective function L, where the objective function includes the loss function and the regularization term:

[0099]

[0100] Where L(φ) represents the objective function; Represents the loss function The sum of Represents the model's predicted output for the i-th input sample; represents the regularization term Ω(f k ); t represents the number of basic models; f k Represents the basic model;

[0101] Optimize using the gradient of the loss function g i and the second-order derivative h i ;

[0102]

[0103] in, represents partial derivative;

[0104] Calculate the split gain for each node:

[0105]

[0106] Where Gain represents the split gain; G L represents the sum of the gradients of all samples of the left child node; G R represents the sum of the gradients of all samples of the right child node; H L represents the sum of the second-order derivatives of all samples of the left child node; H R represents the sum of the second-order derivatives of all samples of the right child node; λ represents the regularization parameter; γ represents the regularization parameter that controls the complexity of the tree;

[0107] Considering that data transmission between multiple nodes introduces delays, data synchronization delays affect the accuracy of model prediction. Therefore, by quantifying data synchronization delays and combining delay-sensitive features to design local models, we can ultimately achieve error compensation between the global model and the local model and optimize the final prediction effect.

[0108] In a distributed environment, data synchronization latency refers to the time difference between when work order data arrives from one node to another;

[0109] The synchronization delay is defined as:

[0110] Δt sync，i =t′ i -t i ;

[0111] Among them, Δt sync，i represents data synchronization delay; t′ i Indicates data synchronization time;

[0112] Features i The time-dependent part f(t i ) will be replaced by f(t′ i ), resulting in input deviation:

[0113] Δx i =f(t′ i )-f(t i );

[0114] Where Δx i represents the input feature deviation due to time change; f(t′ i ) represents the time-dependent eigenvalue after replacement; f(t i ) indicates that at time t i The characteristic value of

[0115] Input bias introduced by delay may cause the model to predict Deviation:

[0116]

[0117] in, represents the deviation of the model prediction value due to time change; g(·) represents the model prediction function;

[0118] Introduce a sliding window W to smooth delay-sensitive features x i , used to reduce short-term fluctuations introduced by latency;

[0119]

[0120] Where W represents the window size; x j represents the window feature value; j represents the index variable;

[0121] The local model is based on X delay And target variable Y training:

[0122]

[0123] Among them, X delay Represents the features after time delay processing; Represents the predicted value of the local model for the target variable Y;

[0124] For delay-sensitive features, train a lightweight local model h local ;

[0125]

[0126] in, Represents the predicted value of the local model for the i-th sample;

[0127] The objective function L(h local )for:

[0128]

[0129] Among them, ||h local || 2 Indicates the complexity of the model;

[0130] Based on feature X, the global model outputs the predicted value:

[0131]

[0132] in, represents the global model prediction value;

[0133] Combination of global model predictions and local model predictions:

[0134]

[0135] in, It represents the combined prediction value of the global model prediction value and the local model prediction value; a represents the weight coefficient.

[0136] As a further improvement of the technical solution, in S6, the specific steps of displaying the repeated demands of the sensitive customer according to the work order type classification interface are:

[0137] S6.1. Use the data visualization tool Matplotlib to display the results of repeated appeals;

[0138] S6.2. Use the Web framework to develop a user interface to display the classification results of repeated demands;

[0139] S6.3. Create an HTML template file to display the classification results of repeated demands.

[0140] Compared with the prior art, the present invention has the following beneficial effects:

[0141] 1. In this method of automatic troubleshooting of duplicate requests for work orders through all channels, work order information is collected through multiple channels to ensure the comprehensiveness and diversity of data. PRA crawling and manual import are combined to improve the efficiency and accuracy of data collection. Natural language processing technology is used to screen out work orders containing sensitive words but not involving arrears, reducing false positives and missed reports. The customer phone number is used as the unique identifier to ensure the uniqueness and accuracy of the customer.

[0142] 2. In this method for automatically troubleshooting repeated requests for omni-channel work orders, the user's historical work order records and vibration data records are integrated through SQL query statements to form a complete historical data set. The time interval feature is introduced to capture the temporal change trend and improve the timeliness of the model. The random forest model is optimized using the XGBoost algorithm to improve the training efficiency and prediction performance of the model. The results of repeated requests are displayed through the data visualization tool Matplotlib, so that managers can intuitively understand the distribution and type of problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0143] Figure 1 The figure is a flow chart of the overall method of the present invention. DETAILED DESCRIPTION

[0144] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0145] Example:

[0146] See also Figure 1 As shown, this embodiment provides an automatic troubleshooting method for repeated requests in omni-channel work orders, including the following steps:

[0147] S1. Build an omni-channel work order information database;

[0148] Among them, the work order information database includes the number of telephone complaints, the total number of work orders, and the content of acceptance;

[0149] In this example, the specific steps to build an omni-channel work order information library are:

[0150] S1.1. Use the PRA system to automate the process of capturing work order data from 95598, 12398, and 12345 telephone channels;

[0151] Specifically, the PRA (Process Robotic Automation) system is a technology that uses software robots to automatically execute repetitive and regular business processes. It can simulate human operations and automatically complete tasks such as data entry, file processing, and system interaction, significantly improving work efficiency, reducing human errors, and reducing operating costs. The PRA system is widely used in customer service, financial processing, supply chain management and other fields to help companies achieve digitalization and intelligence of business processes.

[0152] The omni-channel work order information database includes the 95598 power service hotline, 12398 energy supervision hotline, 12345 citizen service hotline and other work order information databases;

[0153] S1.2. For data that cannot be automatically captured by the PRA system, a manual import process is established to ensure the integrity and accuracy of the data.

[0154] Specifically, the manual import process refers to the process of importing external data or files into the system through manual operations. This process usually involves manually filling out forms, uploading files, copying and pasting data, etc. to ensure the accuracy and completeness of the data. Manual import is often used to process unstructured data, one-time data migration, or data transmission between systems. Although its efficiency is relatively low, it is still indispensable in some cases.

[0155] S2. Use natural language processing technology to query customer demands in the accepted content and select target customers by screening the total number of work orders;

[0156] In this example, natural language processing technology is used to query customer demands in the acceptance content, and the specific steps for filtering target customers by filtering the total number of work orders are as follows:

[0157] S2.1, build a sensitive vocabulary database;

[0158] Among them, the sensitive word library does not include the overdue payment keyword;

[0159] Specifically, the sensitive word library includes, but is not limited to, words such as complaint, dissatisfaction, and power outage, but excludes keywords related to arrears;

[0160] A sensitive word library is a predefined collection of words that usually contains words that may cause concern or have specific meanings, such as complaints, dissatisfaction, power outages, service interruptions, etc. These words are used to identify and mark possible problems or negative emotions during text analysis, helping companies and organizations quickly locate and respond to customer needs and concerns. In an automatic troubleshooting method for repeated requests in omni-channel work orders, the sensitive word library is used to filter and screen the content of the work order to more accurately identify customers and issues that require special attention.

[0161] S2.2. Use natural language processing technology to analyze the acceptance content in the work order;

[0162] Specifically, Natural Language Processing (NLP) is a branch of computer science and artificial intelligence that is dedicated to enabling computers to understand, interpret, and generate human natural language. NLP technology covers multiple aspects, from text segmentation, part-of-speech tagging, named entity recognition, grammatical analysis to semantic understanding and sentiment analysis. It aims to enable computers to not only recognize the formal structure of language, but also deeply understand the meaning behind the language, thereby achieving more intelligent human-computer interaction, information retrieval, machine translation, sentiment analysis and other functions. Through NLP technology, enterprises and organizations can effectively process large amounts of text data, extract valuable information from it, and improve work efficiency and service quality.

[0163] In this example, the specific steps for using natural language processing technology to analyze the acceptance content in the work order are as follows:

[0164] S2.21. Preprocess the work order content;

[0165] Among them, the average length of each work order is L characters, and the content length after preprocessing is reduced to L′;

[0166] Compression ratio CR after preprocessing:

[0167]

[0168] For N work orders, the total preprocessing time is:

[0169] T pre =N×O(f(L));

[0170] Where N represents the number of work orders; Tpre represents the total preprocessing time; O(f(L)) represents the preprocessing time complexity;

[0171] Specifically, the process of preprocessing the work order content is as follows:

[0172] Eliminate punctuation and special characters to prevent them from interfering with subsequent word segmentation and matching processes;

[0173] Unify the text format to avoid the matching results being affected by different capitalizations;

[0174] Eliminate numbers unless they are significant in a specific context;

[0175] Eliminate extra spaces to ensure neat text formatting;

[0176] Split text into words or phrases for subsequent processing;

[0177] Eliminate common nonsense words and reduce noise.

[0178] The example process is:

[0179]

[0180]

[0181] S2.22, matching the preprocessed work order content with the sensitive word library to find the work orders containing sensitive words;

[0182] Among them, the sensitive word library contains V words, and each work order needs to be checked W times on average;

[0183] Matching calculation amount Z:

[0184] Z = V × W;

[0185] S2.23. Check whether the work order content contains keywords related to arrears;

[0186] Among them, the keyword library related to arrears contains K keywords, and each work order needs to be checked X times on average;

[0187] Calculation amount of overdue payment keyword check C:

[0188] C = K × X;

[0189] Specifically, if a work order contains keywords related to arrears, the work order will not be included in the subsequent screening.

[0190] S2.24. Filter out qualified work orders based on the results of sensitive word matching and overdue payment keyword exclusion.

[0191] Among them, M is the total number of initial work orders, Ms is the number of work orders remaining after sensitive word matching, M r The number of work orders that meet the conditions that are finally screened out;

[0192] Screening ratio:

[0193]

[0194] Among them, P represents the screening ratio;

[0195] Screening efficiency:

[0196]

[0197] Where E represents the screening efficiency.

[0198] S2.3. Count the number of work orders for each customer and filter out work order records that have been submitted twice or more.

[0199] S3. Use customer phone numbers as screening targets to screen sensitive customers;

[0200] Specifically, taking customer phone calls as the screening target, the specific steps for screening sensitive customers are:

[0201] S3.1. Extract the customer's phone number from the work order record;

[0202] S3.2. Count the number of work orders submitted by each customer;

[0203] S3.3. Filter out customers who have submitted work orders twice or more;

[0204] S3.4. The screened customers are designated as sensitive customers.

[0205] S4. Set query conditions through the select query statement of the object-relational database management system and query the historical work orders of sensitive customers;

[0206] In this example, the specific steps for setting query conditions through the select query statement of the object-relational database management system and querying the historical work orders of sensitive customers are as follows:

[0207] S4.1. Create a work order table in the database;

[0208] S4.2. Write an SQL query statement in the object-relational database management system, using the user number and mobile phone number as query conditions to query the user's historical work order records from the work order table;

[0209] S4.3. Preprocess the query results and perform data analysis.

[0210] In this example, the specific steps for preprocessing the query results and performing data analysis are as follows:

[0211] S4.31. Extract key information of the work order record from the query results and format it;

[0212] Among them, the key information of the work order record includes the work order ID, account number, mobile phone number, acceptance content and creation time;

[0213] S4.32, extract key information of the vibration data record from the query result and format it;

[0214] Among them, the key information of vibration data records includes vibration data ID, account number, mobile phone number, vibration data and recording time;

[0215] S4.33, calculating the time interval between adjacent records;

[0216] The time interval between adjacent records is calculated as:

[0217]

[0218] Where V represents the total number of customers; M i represents the number of work orders of the i-th customer; i represents the index variable; T interval Indicates the time required for each time interval; A time representing the total time interval.

[0219] Specifically, the examples are:

[0220]

[0221] S5. Combined with historical work order records, the complaint issues are classified and arranged by introducing time intervals and using the random forest model optimized by the XGBoost algorithm;

[0222] In this example, the specific steps of classifying and arranging complaint issues by combining historical work order records, introducing time intervals and optimizing the random forest model with the XGBoost algorithm are as follows:

[0223] S5.1, extracting work order record features;

[0224] The specific steps for extracting work order record features are as follows:

[0225] Extract feature vector x=[x 1 , x 2 , ..., x n ];

[0226] S5.2, obtaining the time interval between adjacent records;

[0227] The expression for obtaining the time interval between adjacent records is:

[0228] Δt i =t i+1 —t i ;

[0229] Among them, t i Indicates the timestamp of the i-th work order; Δt i Represents the time interval characteristics; t i+1 Indicates the timestamp of the i+1 work order;

[0230] S5.3, labeling based on keywords in the work order content;

[0231] The process of labeling based on keywords in the work order content is as follows:

[0232] Generate a label y for each work order based on the content keywords:

[0233] Establish a keyword set K = {k 1 , k 2 , ..., k m};

[0234] For work order text T i Process and mark whether it contains keywords:

[0235] y i =f(T i , K);

[0236] Where f represents the function of text matching and word vector similarity;

[0237] S5.4. Build a random forest model;

[0238] In this example, the specific process of building a random forest model is:

[0239] S5.41, training data preparation;

[0240] The training data preparation is as follows:

[0241] Prepare dataset D:

[0242]

[0243] Among them, x i Represents the characteristics of the i-th record; y i represents its label; N represents the total number of samples in the dataset D;

[0244] S5.42. Generate M sub-datasets by random sampling, each of which is N′ in size. Each sub-dataset randomly selects a feature subset for training and constructs a decision tree t j ;

[0245] S5.43, Generation of decision trees;

[0246] The generation of decision tree is as follows:

[0247] Decision Tree T j The generation is based on recursive partitioning, selecting feature x k and threshold θ to minimize the Gini index Gini(S):

[0248]

[0249] Among them, S represents the sample set in the node, p c represents the sample proportion of category c; c represents the classification category; C represents the total number of categories in the classification task;

[0250] S5.44. Combine the output results of all decision trees into a classification result by voting;

[0251] Among them, the classification results for:

[0252]

[0253] Where Mode represents the mode operation; T M (x) represents the predicted output of the Mth decision tree among the M decision trees generated by the random forest algorithm for the input X;

[0254] S5.45. Use the test set to evaluate the performance of the random forest model.

[0255] S5.5, by introducing time interval features and using XGBoost to optimize the random forest model;

[0256] In this example, the specific process of introducing the time interval feature and using XGBoost to optimize the random forest model is as follows:

[0257] The time interval feature is taken as part of the input data and combined with other work order features X i Together they form the new input vector:

[0258]

[0259] in, represents the input vector that introduces the time interval feature;

[0260] Use XGBoost to optimize the model. XGBoost is based on gradient boosting decision trees for optimization.

[0261] Define the objective function L, where the objective function includes the loss function and the regularization term:

[0262]

[0263] Where L(φ) represents the objective function; Represents the loss function The sum of Represents the model's predicted output for the i-th input sample; represents the regularization term Ω(f k ); T represents the number of basic models; f k Represents the basic model;

[0264] Optimize using the gradient of the loss function g i and the second-order derivative h i ;

[0265]

[0266] in, represents partial derivative;

[0267] Calculate the split gain for each node:

[0268]

[0269] Where Gain represents the split gain; G L represents the sum of the gradients of all samples of the left child node; G R represents the sum of the gradients of all samples of the right child node; H L represents the sum of the second-order derivatives of all samples of the left child node; H R represents the sum of the second-order derivatives of all samples of the right child node; λ represents the regularization parameter; γ represents the regularization parameter that controls the complexity of the tree;

[0270] Considering that data transmission between multiple nodes introduces delays, data synchronization delays affect the accuracy of model prediction. Therefore, by quantifying data synchronization delays and combining delay-sensitive features to design local models, we can ultimately achieve error compensation between the global model and the local model and optimize the final prediction effect.

[0271] In a distributed environment, data synchronization latency refers to the time difference between when work order data arrives from one node to another;

[0272] The synchronization delay is defined as:

[0273] Δt sync,i =t′i -t i ;

[0274] Among them, Δt sync，i represents data synchronization delay; t′ i Indicates data synchronization time;

[0275] Feature X i The time-dependent part f(t i ) will be replaced by f(t′ i ), resulting in input deviation:

[0276] Δx i =f(t′ i )-f(t i );

[0277] Where Δx i represents the input feature deviation due to time change; f(t′ i ) represents the time-dependent eigenvalue after replacement; f(t i ) indicates that at time t i The characteristic value of

[0278] Input bias introduced by delay may cause the model to predict Deviation:

[0279]

[0280] in, represents the deviation of the model prediction value due to time change; g(·) represents the model prediction function;

[0281] Introduce a sliding window W to smooth delay-sensitive features x i , used to reduce short-term fluctuations introduced by latency;

[0282]

[0283] Where W represents the window size; x j represents the window feature value; j represents the index variable;

[0284] The local model is based on X delay And target variable Y training:

[0285]

[0286] Among them, X delay Represents the features after time delay processing; Represents the predicted value of the local model for the target variable Y;

[0287] For delay-sensitive features, train a lightweight local model hlocal ;

[0288]

[0289] in, Represents the predicted value of the local model for the i-th sample;

[0290] The objective function L(h local )for:

[0291]

[0292] Among them, ||h local || 2 Indicates the complexity of the model;

[0293] Based on feature X, the global model outputs the predicted value:

[0294]

[0295] in, represents the global model prediction value;

[0296] Combination of global model predictions and local model predictions:

[0297]

[0298] in, It represents the combined prediction value of the global model prediction value and the local model prediction value; α represents the weight coefficient.

[0299] Where α is:

[0300] α=exp(-βΔt sync，i );

[0301] Among them, exp represents the natural exponential function; β represents the decay rate;

[0302] Specifically, by introducing time interval features and sliding window smoothing technology, the model can better capture time-related patterns and reduce prediction deviations caused by data synchronization delays, thereby improving the accuracy of identifying repeated requests for work orders. Lightweight local models can adapt to new data changes in a short period of time and quickly respond to delay-sensitive features, ensuring that the model can maintain efficient real-time prediction performance in a distributed environment. The global model is responsible for processing long-term trends, while the local model focuses on short-term fluctuations. The combination of the two enables the model to reduce computational complexity and resource consumption while maintaining high accuracy, making it suitable for large-scale distributed systems. By more accurately identifying and processing repeated requests for work orders, companies can respond to customer needs more quickly, reduce customer waiting time, and improve service quality and customer satisfaction. Accurate prediction results provide management with reliable data support, helping them better understand problems in work order processing, optimize resource allocation, and develop more effective operational strategies.

[0303] S5.6. Use the optimized random forest model to classify and arrange the complaints and output the results of repeated complaints.

[0304] Specifically, the numerical features are first standardized by the standard scaler to ensure that the feature values ​​are at the same level; then the data set is divided into a training set and a test set to ensure the independence of model training and evaluation. Next, the XGBoost classifier is initialized and the parameters are set, the calculated time interval is added to the feature vector of each vibration data record, and the random forest model is trained using the training set data through the XGBoost classifier to improve the robustness and generalization ability of the model. Finally, the test set is used to evaluate the model performance to ensure that the model can accurately identify and handle repeated requests in actual applications, reduce manual intervention, and improve customer satisfaction and service quality.

[0305] S6. Display the repeated demands of this sensitive customer on the work order type classification interface.

[0306] In this example, the specific steps for displaying the repeated demands of this sensitive customer according to the work order type classification interface are as follows:

[0307] S6.1. Use the data visualization tool Matplotlib to display the results of repeated appeals;

[0308] Specifically, Matplotlib is a widely used Python drawing library that provides rich drawing functions and can generate high-quality static, dynamic, and interactive charts. With Matplotlib, users can easily create various types of charts, such as line charts, scatter charts, bar charts, pie charts, etc., which are helpful for data visualization and analysis. Matplotlib's flexibility and powerful customization capabilities make it an important tool for data scientists, researchers, and engineers in the field of data visualization.

[0309] Furthermore, the implementation process is:

[0310] import matplotlib.pyplot as plt

[0311] import seaborn as sns

[0312] #Assume that y_pred is the prediction result of the model

[0313] sns.countplot(x=y_pred)

[0314] plt.xlabel('Work Order Type')

[0315] plt.ylabel('Quantity')

[0316] plt.title('Results of repeated appeal classification')

[0317] plt.show()

[0318] S6.2. Use the Web framework to develop a user interface to display the classification results of repeated demands;

[0319] Specifically, a web framework is a software framework used to simplify web application development. It provides a set of pre-written modules and tools to help developers build, deploy, and maintain web applications more efficiently. Web frameworks usually include functions such as routing management, request processing, database interaction, and template rendering, allowing developers to focus on the implementation of business logic without having to deal with the underlying HTTP requests and responses from scratch. Common web frameworks include Python's Flask and Django, Java's Spring Boot, Ruby's Rails, etc. These frameworks not only improve development efficiency, but also enhance the security and scalability of applications.

[0320] Furthermore, the implementation process is:

[0321]

[0322] S6.3. Create an HTML template file to display the classification results of repeated demands.

[0323] Specifically, an HTML template file is a file that contains structured tags and dynamic content placeholders, which is used to generate the final HTML page. These template files are often used in conjunction with backend frameworks (such as Flask, Django, etc.), allowing developers to define the static part of the page in the template while inserting dynamic data through placeholders or variables. When a user requests a page, the backend framework will fill the dynamic data into the template, generate the complete HTML content, and send it to the browser. This approach not only improves development efficiency, but also makes the management and update of page content more flexible and convenient.

[0324] Furthermore, the implementation process is:

[0325]

[0326]

[0327] By using the data visualization tool Matplotlib to display the results of repeated requests, using the Web framework to develop the user interface, and creating an HTML template file to display the classification results, the repeated requests can be intuitively presented and efficiently managed. This process not only helps managers quickly identify and understand the distribution and types of repeated requests, but also improves the efficiency of work order processing and customer satisfaction, ensuring timely problem resolution and reasonable allocation of resources.

[0328] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and descriptions are only preferred examples of the present invention, and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. A method for automatically troubleshooting duplicate requests for omni-channel work orders, characterized by: The following steps are involved: S1. Build an omni-channel work order information database; Among them, the work order information database includes the number of telephone complaints, the total number of work orders, and the content of acceptance; S2. Use natural language processing technology to query customer demands in the accepted content and select target customers by screening the total number of work orders; S3. Use customer phone numbers as screening targets to screen sensitive customers; S4. Set query conditions through the select query statement of the object-relational database management system and query the historical work orders of sensitive customers; S5. Combined with historical work order records, the complaint issues are classified and arranged by introducing time intervals and using the random forest model optimized by the XGBoost algorithm; S6. Display the repeated demands of this sensitive customer on the work order type classification interface.

2. The method for automatically troubleshooting duplicate requests for omni-channel work orders according to claim 1, characterized in that: In S1, the specific steps involved in building the omni-channel work order information database are: S1.

1. Use the PRA system to automatically capture work order data from 95598, 12398, and 12345 telephone channels; S1.

2. For data that cannot be automatically captured by the PRA system, a manual import process is established to ensure the integrity and accuracy of the data.

3. The method for automatically troubleshooting duplicate requests for omni-channel work orders according to claim 1, characterized in that: In S2, the specific steps of querying customer demands in the accepted content by natural language processing technology and screening target customers by screening the total number of work orders are as follows: S2.1, build a sensitive vocabulary library; Among them, the sensitive word library does not include the overdue payment keyword; S2.

2. Use natural language processing technology to analyze the acceptance content in the work order; S2.

3. Count the number of work orders for each customer and filter out work order records that have been submitted twice or more.

4. The method for automatically troubleshooting duplicate requests for omni-channel work orders according to claim 3 is characterized by: In S2.2, the specific steps of using natural language processing technology to analyze the acceptance content in the work order are: S2.

21. Preprocess the work order content; Among them, the average length of each work order is L characters, and the content length after preprocessing is reduced to L′; Compression ratio CR after preprocessing: For N work orders, the total preprocessing time is: T pre =N×O(f(L)); Where N represents the number of work orders; T pre represents the total preprocessing time; O(f(L)) represents the preprocessing time complexity; S2.22, matching the preprocessed work order content with the sensitive word library to find the work orders containing sensitive words; Among them, the sensitive word library contains V words, and each work order needs to be checked W times on average; Matching calculation amount Z: Z = V × W; S2.

23. Check whether the work order content contains keywords related to arrears; Among them, the keyword library related to arrears contains K keywords, and each work order needs to be checked X times on average; Calculation amount of overdue payment keyword check C: C = K × X; S2.

24. Filter out qualified work orders based on the results of sensitive word matching and overdue payment keyword exclusion; Among them, M is the total number of initial work orders, M s is the number of work orders remaining after sensitive word matching, M r The number of work orders that meet the conditions that are finally screened out; Screening ratio: Among them, P represents the screening ratio; Screening efficiency: Where E represents the screening efficiency.

5. The method for automatically troubleshooting duplicate requests for omni-channel work orders according to claim 1, characterized in that: In S4, the specific steps of setting the query conditions through the select query statement of the object-relational database management system and querying the historical work orders of sensitive customers are as follows: S4.

1. Create a work order table in the database; S4.

2. Write an SQL query statement in the object-relational database management system, using the user number and mobile phone number as query conditions to query the user's historical work order records from the work order table; S4.

3. Preprocess the query results and perform data analysis.

6. The method for automatically troubleshooting duplicate requests for omni-channel work orders according to claim 5, characterized in that: In S4.3, the specific steps of preprocessing the query results and performing data analysis are as follows: S4.

31. Extract key information of the work order record from the query results and format it; Among them, the key information of the work order record includes the work order ID, account number, mobile phone number, acceptance content and creation time; S4.32, merge the work order records in chronological order; S4.33, calculating the time interval between adjacent records; The time interval between adjacent records is calculated as: Where V represents the total number of customers; M i represents the number of work orders of the i-th customer; i represents the index variable; T interval Indicates the time required for each time interval; A time representing the total time interval.

7. The method for automatically troubleshooting duplicate requests for omni-channel work orders according to claim 1, characterized in that: In S5, the specific steps of classifying and arranging the complaint issues by introducing time intervals and using the random forest model optimized by the XGBoost algorithm in combination with the historical work order records are as follows: S5.1, extracting work order record features; The specific steps for extracting work order record features are as follows: Extract feature vector x=[x1, x2, …, x n ]; S5.2, obtaining the time interval between adjacent records; The expression for obtaining the time interval between adjacent records is: Δt i =t i+1 -t i ; Among them, t i Indicates the timestamp of the i-th work order; Δt i Represents the time interval characteristics; t i+1 Indicates the timestamp of the i+1 work order; S5.3, labeling based on keywords in the work order content; The process of labeling based on keywords in the work order content is as follows: Generate a label y for each work order based on the content keywords: Establish a keyword set K = {k1, k2, ..., k m }; For work order text T i Process and mark whether it contains keywords: y i =f(T i ,K); Where f represents the function of text matching and word vector similarity; S5.

4. Build a random forest model; S5.5, by introducing time interval features and using XGBoost to optimize the random forest model; S5.

6. Use the optimized random forest model to classify and arrange the complaints and output the results of repeated complaints.

8. The method for automatically troubleshooting duplicate requests for omni-channel work orders according to claim 7, characterized in that: In S5.4, the specific process of constructing the random forest model is: S5.41, training data preparation; The training data preparation is as follows: Prepare dataset D: Among them, x i Represents the characteristics of the i-th record; y i represents its label; N represents the total number of samples in the dataset D; S5.

42. Generate M sub-datasets by random sampling. The size of each sub-dataset is N′. Each sub-dataset randomly selects a feature subset for training and constructs a decision tree T j ; S5.43, Generation of decision trees; The generation of decision tree is as follows: Decision Tree T j The generation is based on recursive partitioning, selecting feature x k and threshold θ to minimize the Gini index Gini(S): Among them, S represents the sample set in the node, p c represents the sample proportion of category c; c represents the classification category; C represents the total number of categories in the classification task; S5.

44. Combine the output results of all decision trees into a classification result by voting; Among them, the classification results for: Where Mode represents the mode operation; T M (x) represents the predicted output of the Mth decision tree among the M decision trees generated by the random forest algorithm for the input X; S5.

45. Use the test set to evaluate the performance of the random forest model.

9. The method for automatically troubleshooting duplicate requests for omni-channel work orders according to claim 8, characterized in that: In S5.5, the specific process of introducing the time interval feature and using XGBoost to optimize the random forest model is as follows: The time interval feature is taken as part of the input data and together with other work order features, forms a new input vector: in, represents the input vector that introduces the time interval feature; Use XGBoost to optimize the model. XGBoost is based on gradient boosting decision trees for optimization. Define the objective function L, where the objective function includes the loss function and the regularization term: Where L(φ) represents the objective function; Represents the loss function The sum of Represents the model's predicted output for the i-th input sample; represents the regularization term Ω(f k ); T represents the number of basic models; f k Represents the basic model; Optimize using the gradient of the loss function g i and the second-order derivative h i ; in, represents partial derivative; Calculate the split gain for each node: Where Gain represents the split gain; G L represents the sum of the gradients of all samples of the left child node; G R represents the sum of the gradients of all samples of the right child node; H L represents the sum of the second-order derivatives of all samples of the left child node; H R represents the sum of the second-order derivatives of all samples of the right child node; λ represents the regularization parameter; γ represents the regularization parameter that controls the complexity of the tree; Considering that data transmission between multiple nodes introduces delays and data synchronization delays that affect the accuracy of model prediction, we quantify data synchronization delays and design local models in combination with delay-sensitive features to achieve error compensation between the global model and the local model, thereby optimizing the final prediction effect. In a distributed environment, data synchronization latency refers to the time difference between when work order data arrives from one node to another; The synchronization delay is defined as: Δt sync，i =t′ i -t i ; Among them, Δt sync，i represents data synchronization delay; t′ i Indicates data synchronization time; Features i The time-dependent part f(t i ) will be replaced by f(t′ i ), resulting in input deviation: Δx i =f(t′ i )-f(t i ); Where Δx i represents the input feature deviation due to time change; f(t′ i ) represents the time-dependent eigenvalue after replacement; f(t i ) indicates that at time t i The characteristic value of Input bias introduced by delay may cause the model to predict Deviation: in, represents the deviation of the model prediction value due to time change; g(·) represents the model prediction function; Introduce a sliding window W to smooth delay-sensitive features x i , used to reduce short-term fluctuations introduced by latency; Where W represents the window size; x j represents the window feature value; j represents the index variable; The local model is based on X delay And target variable Y training: Among them, X delay Represents the features after time delay processing; Represents the predicted value of the local model for the target variable Y; For delay-sensitive features, train a lightweight local model h local ; in, Represents the predicted value of the local model for the i-th sample; The objective function L(h local )for: Among them, ||h local || 2 Indicates the complexity of the model; Based on feature X, the global model outputs the predicted value: in, represents the global model prediction value; Combination of global model predictions and local model predictions: in, It represents the combined prediction value of the global model prediction value and the local model prediction value; α represents the weight coefficient.

10. The method for automatically troubleshooting duplicate requests for omni-channel work orders according to claim 1, characterized in that: In S6, the specific steps of displaying the repeated demands of the sensitive customer according to the work order type classification interface are: S6.

1. Use the data visualization tool Matplotlib to display the results of repeated appeals; S6.

2. Use the Web framework to develop a user interface to display the classification results of repeated demands; S6.

3. Create an HTML template file to display the classification results of repeated demands.

Images