Automatic business expansion work order receiving and dispatching method based on full process

Abstract

The invention discloses a work order receiving and dispatching method based on full-process automation business expansion, and relates to the field of work order full-process processing automation. The invention provides a full-process-based automatic business expansion work order receiving and dispatching method, which comprises the following steps of: compiling a python script, and setting an initial URL (Uniform Resource Locator) of a browser as a login page of a power supply service command system; identifying the verification code through OCR and CNN-RNN algorithms to realize automatic login of the system; writing a python script and automatically jumping to an unaudited work order page; identifying the work order to be processed by using the BERT model; distributing the work order to the corresponding worker according to the distribution rule; and recording a distribution log. The invention aims to solve the problem of man-made repeated, tedious and large-batch work tasks in the business expansion business through an automatic process technology, and realizes the whole process automation of the business expansion business from order receiving, order sending and return visit.

Description

Technical Field

[0001] The present application relates to the field of full-process automation of work order processing, and is applied to the full automation of business expansion order acceptance, dispatching, and intelligent follow-up visits, and in particular to a method for business expansion work order acceptance and dispatching based on full-process automation. Background Art

[0002] In order to improve service quality, the marketing department of the power company needs to control the communication service process of fault repair, power supply quality opinions and other 95598 work orders. For unreviewed business expansion work orders, according to the work order processing flow and distribution rules, the work orders will be distributed to the corresponding power supply unit or corresponding personnel. After the distribution is completed, manual return visits are required and customer satisfaction surveys are conducted. This business involves a lot of work that requires logging into the system, querying pages, downloading call records, checking work orders, and business return visits. It is time-consuming and labor-intensive and it is difficult to guarantee results. It is necessary to improve work efficiency and customer satisfaction through full-process automation. Summary of the invention

[0003] The purpose of the embodiments of the present invention is to provide a method for receiving and dispatching business expansion work orders based on full-process automation, aiming to solve the manual repetitive, tedious, and large-scale work tasks in business expansion business through automated process technology, and to realize the full process automation of business expansion business from order receiving, dispatching, and follow-up visits.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] The embodiment of the present invention provides a method for receiving and dispatching work orders based on full-process automation, including the following steps:

[0006] Write a python script and set the initial URL of the browser to the login page of the power supply service command system;

[0007] Automatically log into the system by recognizing verification codes through OCR and CNN-RNN algorithms;

[0008] Write a python script to automatically jump to the unreviewed work order page;

[0009] Use the BERT model to identify pending work orders;

[0010] Dispatching work orders to corresponding staff members according to the dispatching rules;

[0011] Record the dispatch log.

[0012] In some embodiments, identifying the verification code through OCR includes the following steps:

[0013] Grayscale, binarize and remove noise preprocessing of verification code image;

[0014] Locate the verification code image and separate the text area, and divide the text area into single characters;

[0015] Perform discrete Fourier transform on the character image to obtain the frequency domain representation of the character shape.

[0016] In some embodiments, identifying the verification code by using the CNN-RNN algorithm includes the following steps:

[0017] CNN part: The input layer is used to receive the preprocessed verification code image; the convolution layer is used to extract features from the image using multiple convolution layers; the pooling layer is used to reduce the spatial size of the feature map, reduce the amount of calculation, and prevent overfitting; the activation function and the fully connected layer are used to further extract features;

[0018] RNN part: The input layer is used to receive the feature vector output by the CNN part; the recurrent layer is used to process sequence data and learn the temporal relationship between characters; the fully connected layer is used to map the output of the RNN layer to the final classification result.

[0019] In some embodiments, using a BERT model to identify work orders to be processed to process work order information includes the following steps:

[0020] Collect historical work order information, clean and preprocess it, remove irrelevant characters, and unify the text format;

[0021] Delete irrelevant characters in the address; convert the address to a standard format; convert the power supply unit to a standard unified format; classify users into high-voltage, low-voltage residents, and low-voltage non-residents; annotate the data and specify the staff to be assigned to each work order;

[0022] Extract features from the model through the analysis layer, define the loss function and optimizer, load data using DataLoader, and iteratively train the model;

[0023] Use the trained model to classify and predict the new work order information text, obtain the category label of each word, and extract key information based on the label;

[0024] Determine the user's location, power supply department, and power supply unit based on the work order information to be processed;

[0025] The identified work orders will be dispatched to the staff according to the preset dispatching rules.

[0026] In some embodiments, it also includes a smart phone return visit; the smart phone return visit includes the following steps:

[0027] Deployment of smartphone development environment;

[0028] Write a script, set a scheduled task based on the work order log, and automatically call the work order phone number at 3 pm the next day;

[0029] Define voice prompt text, conduct satisfaction surveys, set up intelligent voice, and record satisfaction scores;

[0030] Recognize user voice input, convert customer voice into text, train keywords for conversion to manual and trigger the conversion to manual function.

[0031] Under the current background of rapid development of the power market and science and technology, power supply companies are facing major challenges and opportunities for transformation and upgrading. As an important part of the power supply system, the efficiency and quality of the business expansion application process are directly related to the company's service level and market competitiveness. With the increasing maturity of digital and intelligent technologies, how to use these advanced technologies to transform and upgrade the traditional business expansion application process and build a lean management model that meets the requirements of modern enterprise management has become a key issue that power supply companies need to solve urgently. In this process, we must not only give full play to the advantages of digital technology, but also combine the actual situation of power supply companies to ensure that the new model can effectively integrate resources, optimize processes, and improve service quality, and ultimately achieve the sustainable development of power supply companies and industry leadership.

[0032] Lean management emphasizes the overall optimization of the process. Through the careful combing of the business expansion and installation process, it effectively identifies and eliminates redundant links and inefficient operations in the process, and simplifies and standardizes the operation process. This not only shortens the time for processing a single business, but also improves the overall service quality by reducing errors and rework. Lean management advocates customer demand-oriented, and formulates more flexible and personalized service plans through accurate market analysis and demand forecasting. This customer-centric service model can ensure that power supply companies accurately understand and meet customers' specific needs in the first time, thereby greatly improving customer satisfaction and loyalty. For example, through real-time data analysis and efficient customer communication mechanisms, power supply companies can quickly understand and predict customers' electricity consumption patterns, and then prepare or adjust relevant resource allocation in advance to ensure that they can provide rapid and accurate service responses when needs arise. The lean management model emphasizes continuous improvement and innovation. By establishing a continuous improvement mechanism, power supply companies can continuously optimize and upgrade their business expansion and installation processes to make them more in line with market development trends and technological progress requirements. For example, the use of advanced information technology and automated equipment can effectively improve the speed and accuracy of business processing, reduce human errors, and reduce labor intensity.

[0033] Lean management effectively reduces the risks caused by operational errors or improper processes through standardized processes and standardized operations. In the business expansion application process, each link is strictly designed and reviewed to ensure that all operations comply with industry standards and safety specifications, reducing the probability of safety accidents from the source. Lean management of the entire process emphasizes comprehensive monitoring and real-time analysis of data. By establishing a comprehensive data collection and analysis system, power supply companies can implement real-time monitoring of all links in business expansion applications, detect abnormal situations in a timely manner, and take early warning and intervention measures. This data-based risk control mechanism can not only respond to emergencies in a timely manner, but also predict and prevent potential risks, greatly improving the company's risk management capabilities. The continuous improvement and problem-solving mechanism advocated by lean management also provides strong support for risk control. In the process of implementing lean management, power supply companies will continuously collect feedback, analyze problems and optimize processes. This continuous optimization and improvement not only improves the efficiency and effectiveness of the process, but also enhances the company's ability to adapt to external changes and internal problems. By establishing such a dynamic risk management system, power supply companies can respond to various risks and challenges more flexibly and effectively.

[0034] In terms of human resource allocation, power supply enterprises often have an unbalanced allocation of technical personnel and on-site workers in the expansion and installation process. Some key positions such as project planning and quality control are understaffed, while some auxiliary or low-skilled positions are overstaffed. This unbalanced human resource allocation directly affects the efficiency and quality of the expansion and installation process. Key links may be delayed due to insufficient staff, while overstaffed links cause waste of human resources. For example, in the project planning and design stage, there is a lack of experienced engineers to carry out reasonable design and evaluation, while in the construction stage, there may be a shortage of technical personnel and an oversupply of ordinary workers, resulting in obstruction of project progress and idle resources. In terms of material resource allocation, various material resources required in the expansion and installation process of power supply enterprises, such as cables, transformers, and various maintenance tools, are often in short supply or oversupply due to inaccurate procurement plans or imperfect supply chain management. For example, some necessary high-voltage cables and transformers may cause project delays due to long supplier delivery cycles or logistics problems, or because procurement plans do not match actual needs, resulting in oversupply of materials, increased capital occupation, and waste of storage space. Due to the lack of an effective material management system, the allocation and recycling of materials are inefficient, making it difficult to transfer resources between different projects, further exacerbating the problem of resource waste.

[0035] At the technical level, the information systems of various departments in power supply enterprises are often relatively independent and lack effective interconnection mechanisms, resulting in a lag in the transmission of key information between different systems, and the timeliness and accuracy of information are difficult to ensure. For example, the demand information of the customer service department, the progress information of the engineering department, and the inventory information of the material management department cannot be shared in real time, which affects the timeliness and correctness of decision-making. At the organizational structure level, there may be barriers in the internal structure of power supply enterprises, and the communication and collaboration mechanism between departments is not smooth enough. Each department often sticks to its own information resources and lacks the awareness and mechanism of active sharing and collaborative work. This "information island" phenomenon weakens the overall coordination of the entire process of business expansion and registration, resulting in information breaks in the circulation process, hindering the effective transmission and use of information. At the cultural level, the corporate culture of power supply enterprises and the mentality of employees may also affect the adequacy of information sharing. If the corporate culture fails to emphasize the importance of information sharing, or employees have concerns about sharing the information they have, such as worrying that information sharing will affect their job stability or work performance evaluation, then even if there are no obstacles in technology and organizational structure, the actual effect of information sharing will be greatly reduced. In the process of business expansion and project implementation, power supply companies often need to deal with a large amount of complex data and information, such as customer information, project progress, equipment status and market dynamics. Insufficient information sharing will lead to a lack of sufficient information support for key decisions, resulting in inaccurate and slow decisions, which restricts the efficiency and effectiveness of power supply companies in the entire process of business expansion and installation.

[0036] In the whole process of business expansion and installation of power supply enterprises, the connection between some process links is not tight, resulting in too isolated task division and lack of effective information transmission and collaboration mechanism in each link. For example, after the design link is completed, its information may not be transmitted to the procurement or construction link in a timely and accurate manner, resulting in the inability to effectively connect the subsequent work, and even the need for re-adjustment or redesign, resulting in a waste of time and resources. In the business expansion and installation process of power supply enterprises, multiple departments and teams need to work closely together to complete complex tasks. Although each department of the power supply enterprise may have its own tasks and goals, the lack of a unified and efficient collaboration and communication platform makes it difficult to coordinate the work between departments and form a joint force. For example, the construction plan of the engineering department needs to be closely integrated with the material supply plan of the supply chain department, but due to the lack of effective coordination and communication, the materials may not be in place in time, affecting the progress of the entire project. The business expansion and installation process of power supply enterprises needs to face changes in the external environment and adjustments to internal conditions, which requires the process to have certain flexibility and adaptability. However, in actual operation, since power supply projects often have long cycles and complex environments, it is inevitable to encounter various unexpected changes in the process, such as changes in customer needs, adjustments to policies and regulations, and changes in market conditions. Due to the rigidity of process design and the lack of change management mechanisms, power supply companies often find it difficult to adjust processes in a timely manner to respond to these changes, resulting in project delays or cost overruns.

[0037] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0038] The method for receiving and dispatching business expansion work orders based on full-process automation provided by the present invention can realize innovation in the power field based on full-process automation technology, including full-process automation of automatic receiving, dispatching, and return visits of business expansion work orders. The present invention solves repetitive operations such as repeated system entry, multi-system data acquisition, process monitoring and urging, and event monitoring, recording, and reminders through full-process automation applications, improves the degree of automation of business expansion receiving and dispatching work, improves production efficiency, and eliminates human errors to the greatest extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the present disclosure, the following briefly introduces the drawings required for use in some embodiments of the present disclosure. Obviously, the drawings described below are only drawings of some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can also be obtained based on these drawings. In addition, the drawings described below can be regarded as schematic diagrams, and are not limitations on the actual size of the products involved in the embodiments of the present disclosure, the actual process of the method, etc.

[0040] Figure 1 A flowchart of receiving and dispatching business expansion orders according to some embodiments of the present disclosure;

[0041] Figure 2 The present invention is a flowchart of a smartphone callback process according to some embodiments of the present disclosure. DETAILED DESCRIPTION

[0042] The following will be combined with the accompanying drawings to clearly and completely describe the technical solutions in some embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all the embodiments. Based on the embodiments provided by the present disclosure, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present disclosure.

[0043] Unless the context requires otherwise, throughout the specification and claims, the term "including" is to be interpreted as an open, inclusive meaning, that is, "including, but not limited to". In the description of the specification, the terms "one embodiment", "some embodiments", "exemplary embodiments", "examples" or "some examples" and the like are intended to indicate that specific features, structures, materials or characteristics associated with the embodiment or example are included in at least one embodiment or example of the present disclosure. The schematic representations of the above terms do not necessarily refer to the same embodiment or example. In addition, the specific features, structures, materials or characteristics described may be included in any one or more embodiments or examples in any appropriate manner.

[0044] The embodiment of the present invention provides a method for receiving and dispatching work orders based on full-process automation. Figure 1 As shown, it includes steps: S100~S400.

[0045] S100, write a python script, and set the initial URL of the browser to the login page of the power supply service command system.

[0046] The login information should include the login account and password. Use the element positioning function to locate the username and password input boxes on the login page with the XPath selector, add the steps to fill in the text field, and enter the preset username and password into the corresponding input boxes.

[0047] S200, uses OCR and CNN-RNN algorithm to recognize verification codes and realize automatic login system.

[0048] When logging into the system, the verification code is relatively simple and regular. Based on the CNN-RNN model, OCR is used to recognize digital verification codes, and an algorithm is written to generate or predict verification codes. OCR (Optical Character Recognition) technology is used to automatically recognize verification codes.

[0049] To recognize verification codes through OCR, first, sufficient verification codes need to be collected for training and testing. Second, the data needs to be preprocessed to improve the recognition rate through steps such as resizing, graying, binarization, and denoising, including steps: S210-S230.

[0050] S210, gray-scale, binarize, and remove noise preprocessing on the verification code image.

[0051] Specifically, the color verification code image is converted into a grayscale image, and then the grayscale image is further converted into a binary image, and finally the complex information in the image is removed to improve the clarity of the text.

[0052] S220, locating the verification code image and separating the text area, and dividing the text area into single characters.

[0053] S230, performing discrete Fourier transform on the character image to obtain a frequency domain representation of the character shape.

[0054] Recognizing verification codes through the CNN-RNN algorithm includes a CNN part and a RNN part.

[0055] Among them, the CNN part includes steps: S241-S245.

[0056] S241, input layer: receiving the pre-processed verification code image.

[0057] S242, Convolutional layer: Use multiple convolutional layers to extract features from the image.

[0058] S243, pooling layer: reduces the spatial size of the feature map, reduces the amount of calculation, and prevents overfitting.

[0059] S244, activation function: Use ReLU nonlinear activation function to increase the nonlinearity of the network.

[0060] S245, fully connected layer (optional): CNN is followed by one or more fully connected layers to further extract features.

[0061] The CNN-RNN algorithm treats the OCR problem as a multi-label learning problem. The 4-digit verification code is equivalent to a 4-label image recognition problem. First, an image is input through the ConvNet network to output the image representation I. At the same time, the label is mapped to the low-dimensional space w in the form of a one-hot vector through label embedding. k (t), and then input w at each time step k(t), and the previous hidden layer state r(t-1), and finally the RNN network outputs o(t), then the image representation I and the RNN output o(t) are mapped to the low-dimensional space through the mapping layer to obtain x(t), and then the prediction layer is used to obtain the probability of the predicted label, and the largest one is taken as the output.

[0062] The probability of the predicted path can be calculated by the RNN network. The image, label, and recursive representation are projected into the same low-dimensional space to model image-text relationships as well as label redundancy. The RNN model is used as a compact and powerful representation of label co-occurrence dependencies in this space. It embeds the predicted label at each time step and maintains a hidden state to model label co-occurrence information. Given the previously predicted label, the prior probability of the current label can be calculated based on the dot product of the image and the periodic embedding. The probability of the predicted path can be obtained by multiplying the prior probability of each label in the given previous label in the given predicted path.

[0063] A label k can be represented as a one-hot vector e k = [0, ... 0, 1, 0, ... , 0], where the kth position is represented by 1 and the rest of the positions are replaced by 0. Label embedding can be done by adding the one-hot vector to the label embedding matrix U l The label embedding of label k is actually the label embedding matrix U l The kth row of:

[0064] w k =U l *e k ;

[0065] The recurrent layer embeds the previously predicted labels and models the co-occurrence dependencies in its hidden recurrent state by learning a nonlinear function:

[0066] o(t)=h o (r(t-1),w k (t)), r(t) = h r (r(t-1),w k (t));

[0067] where r(t) and o(t) are the hidden state and output of the recurrent layer at time step t, respectively, and w k (t) is the label embedding of the kth label in the prediction path, h o (.) represents a nonlinear RNN function.

[0068] The output of the recurrent layer and the image representation are projected into the same low-dimensional space as the label embedding.

[0069]

[0070] in, and They are the projection matrices of the recursive layer output and image representation, respectively. Their number of columns is the same as the label embedding matrix U1. At the same time, it can be seen that x t What is actually represented is the black dots in the joint space, which is the label embedding information obtained by combining the image features and the output state. o(t) represents the state output at time step t, and I is the image representation of the convolutional neural network. In summary, the original label is embedded into the low-dimensional space using the label embedding matrix, and the output of each time step of the RNN and the original image representation are also embedded into the low-dimensional space using the projection matrix. It can be shown later that the learned joint embedding effectively represents the correlation between the image and the label.

[0071] Finally, x can be calculated by transposing U1 and multiplying t The distance between each label embedding is used to calculate the label score:

[0072]

[0073] By normalizing the scores using softmax, the predicted label probability can be calculated, and the largest label can be used as the input of the next RNN.

[0074] The RNN part is used to process sequence data and capture the relationship between characters, and the output layer is used for character classification.

[0075] Among them, the RNN part includes steps: S246-S248.

[0076] S246, input layer: receives the feature vector output by the CNN part.

[0077] S247, recurrent layer: Use RNN variants such as LSTM to process sequence data and learn the temporal relationship between characters. Multiple RNN layers can be stacked to improve the performance of the model.

[0078] S248, fully connected layer: maps the output of the RNN layer to the final classification result, that is, the character category in the verification code.

[0079] Verification code recognition is achieved through the above method.

[0080] S300, write a python script to automatically jump to the unreviewed work order page. The script contains the required username and password information and handles various interactions during the login process.

[0081] Add "mouse click" or similar components to simulate the mouse click operation. Complete the configuration of fault reporting, inquiry consultation, complaint, report, opinion (suggestion), customer reminder, business application work order, specify the click position by entering coordinates, using the element selector to select interface elements, etc., and automatically jump to the unreviewed work order page.

[0082] Use the jieba word segmentation library to segment the unreviewed work orders. Import the jieba word segmentation library, use the precise mode to accurately separate the work order information for word segmentation, and return the word segmentation result in the form of a generator.

[0083] Use string search and regular expressions to check whether these contents contain specific text indicating the existence of a work order (such as "work order number", "status: submitted", etc.). If matching text or elements are found, it can be determined that there is a work order on the page, and the result of the JavaScript code execution is connected to the conditional judgment component.

[0084] S400, using the BERT model to identify pending work orders.

[0085] The BERT model is used to predict the staff to whom the work order needs to be dispatched. The text is encoded, and then the classifier is used to extract keywords to identify the user address, power supply unit, user classification, and demand type.

[0086] Using the BERT model to identify work orders to be processed includes steps: S410 to S450.

[0087] S410, collect historical work order information, including key information such as user address, power supply unit, user classification, demand type, etc., and clean and pre-process it to remove irrelevant characters and unify the text format.

[0088] S420, delete irrelevant characters in the address, such as special symbols, extra spaces, etc.; convert the address to a standard format, such as province-city-district-street-building-house number; convert the power supply unit to a standard unified format, such as "xx Power Supply Company"; classify users into 1, 2, and 3, and the corresponding information is high voltage, low voltage residents, and low voltage non-residents respectively; mark the data to specify the staff who need to be dispatched for each work order.

[0089] S430, extract features from the model through the analysis layer, define the loss function and optimizer, load data using DataLoader, and iteratively train the model.

[0090] S440: Use the trained model to classify and predict the new work order information text, obtain the category label of each word, and extract key information based on the label, such as user address, power supply unit, user classification, demand type, etc.

[0091] S450, determining the user location, power supply department, and power supply unit according to the work order information to be processed.

[0092] S500: Dispatching the work order to the corresponding staff according to the dispatching rules.

[0093] The identified work orders are dispatched to the staff according to the preset dispatch rules. The work order information is read, the staff information and the dispatch strategy are matched, and the system records are updated to reflect the dispatch status of the work order.

[0094] S600, record the dispatch log.

[0095] Use logging tools to add logging logic at key nodes before dispatch, after dispatch, and when processing results are returned, to ensure that the logging logic can capture all necessary log information and record it in a predetermined format. Write the corresponding code based on the selected logging tool and designed logging logic, add logging statements to the relevant code of the dispatch process, ensure that logs can be generated at key nodes, deploy the logging system to the production environment, and ensure that the dispatch logs can be recorded in real time.

[0096] In some embodiments, the method for receiving and dispatching work orders based on full-process automation provided by the present invention also includes a smart phone follow-up, and the smart phone follow-up includes the steps: S710~S740.

[0097] S710, deploy the smart phone development environment. Use the Speech Recognition library to provide an interface to connect to the microphone and capture audio. Specifically includes: S711 ~ S713.

[0098] S711, use the sudo yum update command to update the YUM source to ensure that the system has the latest software packages, and install common development tools through the sudo yum groupinstall "Development Tools command.

[0099] S712, install wegt to facilitate downloading files from the Internet, and use the pip install virtualenv command to install virtualenv. Virtualenv can create an independent Python environment to avoid dependency conflicts between different projects.

[0100] S713, install the pytorch model framework and download the PaddleSpeech intelligent voice package processing tool.

[0101] S720, write a script, set a scheduled task based on the work order log, and automatically dial the work order phone number at 3 pm the next day. Specific examples include: S721~S724.

[0102] S721, create a database to store dispatch information, including work order number, customer phone number, dispatch status, etc. The storage structure facilitates quick retrieval and update of work order information.

[0103] S722, write code to monitor changes in the order status. When the order status changes to "completed", trigger the logic of automatically making a call.

[0104] S723, use Windows Task Scheduler to set a scheduled task for the next day at 3 pm. In the scheduled task, write code to retrieve the dispatch information that has been completed that day and prepare to make a call.

[0105] S724, uses the speech synthesis module in the intelligent speech development environment to convert the dispatch information into speech, uses the selected library and API (such as Twilio) to automatically make calls, and plays the converted voice information.

[0106] S730, define voice prompt text, conduct satisfaction survey, set intelligent voice, and record satisfaction score. Specifically includes: S731 ~ S733.

[0107] S731, set the intelligent voice text to "Dear user, are you satisfied with the service acceptance result of this area? If you are satisfied, please press 1; if you are relatively satisfied, please press 2; if you are not satisfied, please press 3; if you are not satisfied, please press 4." And configure the corresponding processing logic for the number keys 1, 2, 3, and 4 to record the satisfaction score.

[0108] S732, select FastSpeech speech synthesis technology to convert text into speech. Through the RESTful API interface provided by the cloud service, send the text to the service through an HTTP request and send the URL of the received audio file.

[0109] S733, write code to call the speech synthesis service, initialize the client of the speech synthesis service, construct a request containing the text to be converted, send the request to the speech synthesis service, receive and process the returned audio data, and use the Volp telephone communication protocol to directly send the generated voice data to the audio playback interface of the telephone communication system.

[0110] S740, recognize the user's voice input, convert the customer's voice into text, train the keyword of "transfer to manual" and trigger the function of "transfer to manual". Train and configure the word "transfer to manual" to ensure that the ASR system can accurately recognize it. Specifically, it includes: S741~S743.

[0111] S741, set keywords that need to be transferred to manual service, such as "transfer to manual service", "manual customer service", or trigger the voice transfer function when the user's question cannot be accurately answered by artificial intelligence.

[0112] S742. Input the selected keywords into the training module of the artificial intelligence telephone system, and train the keywords through the TextRank algorithm. During the training process, the system needs to learn how to identify these keywords and trigger the corresponding function of transferring to a human operator when the keywords are recognized.

[0113] S743. After the training is completed, test the system to ensure that the keywords can accurately trigger the function of transferring to a human operator.

[0114] Among them, the steps of voice keyword recognition through TextRank are as follows:

[0115] Collect historical work order telephone recordings and store the voice telephone content as a processable audio file.

[0116] Convert the voice into text through the Speech-to-Text API technology.

[0117] Remove irrelevant characters in the text, such as punctuation marks, numbers, special symbols, etc., and convert the text to lowercase.

[0118] Use the jieba word segmentation tool to split the text into words, perform part-of-speech tagging, and filter out non-keyword parts of speech such as auxiliary words and conjunctions; remove common stop words such as "de" and "le" to reduce noise.

[0119] Take the words or phrases in the segmented text as nodes to construct a candidate keyword graph G=(V, E), and use the co-occurrence relationship to construct the edges between the nodes. There is an edge between two nodes only when the corresponding words co-occur in a window of length K, where K represents the window size.

[0120] According to the formula of the TextRank algorithm, iterate and propagate the weights of each node until convergence. Initialize the weight of each node to an equal value (such as 1 / N, where N is the total number of nodes). In each round of iteration, update the weight of the node according to the weights of the neighboring nodes of the node and the weights of the edges.

[0121] Sort the node weights in descending order, and select the top T words or phrases with the highest weights as candidate keywords. Mark these keywords in the original text and check whether adjacent phrases are formed.

[0122] Match the extracted keywords with a preset keyword list (such as "transfer to human", "human customer service", etc.). If the match is successful, trigger the corresponding function or operation (such as transferring the call to a human customer service). If multiple keywords are recognized simultaneously, the priority can be determined according to the business logic or multiple operations can be performed.

[0123] Use the test data set to evaluate the accuracy of keyword recognition, and optimize text preprocessing, word segmentation tools, stop word lists, TextRank algorithm parameters, etc. based on the evaluation results.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Other modifications or equivalent substitutions made to the technical solution of the present invention by ordinary technicians in the field should be included in the scope of the claims of the present invention as long as they do not depart from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for receiving and dispatching work orders based on full-process automation, characterized in that: The following steps are involved: Write a python script and set the initial URL of the browser to the login page of the power supply service command system; Automatically log into the system by recognizing verification codes through OCR and CNN-RNN algorithms; Write a python script to automatically jump to the unreviewed work order page; Use the BERT model to identify pending work orders; Dispatching work orders to corresponding staff members according to the dispatching rules; Record the dispatch log.

2. The method for receiving and dispatching work orders based on full-process automation as claimed in claim 1, characterized in that: Recognizing the verification code through OCR includes the following steps: Grayscale, binarize and remove noise preprocessing of verification code image; Locate the verification code image and separate the text area, and divide the text area into single characters; Perform discrete Fourier transform on the character image to obtain the frequency domain representation of the character shape.

3. The method for receiving and dispatching work orders based on full-process automation as claimed in claim 1, characterized in that: Recognizing the verification code through the CNN-RNN algorithm includes the following steps: CNN part: The input layer is used to receive the preprocessed verification code image; the convolution layer is used to extract features from the image using multiple convolution layers; the pooling layer is used to reduce the spatial size of the feature map, reduce the amount of calculation, and prevent overfitting; the activation function and the fully connected layer are used to further extract features; RNN part: The input layer is used to receive the feature vector output by the CNN part; the recurrent layer is used to process sequence data and learn the temporal relationship between characters; the fully connected layer is used to map the output of the RNN layer to the final classification result.

4. The method for receiving and dispatching work orders based on full-process automation as claimed in claim 1, characterized in that: Using the BERT model to identify pending work orders includes the following steps: Collect historical work order information, clean and preprocess it, remove irrelevant characters, and unify the text format; Delete irrelevant characters in the address; convert the address to a standard format; convert the power supply unit to a standard unified format; classify users into high-voltage, low-voltage residents, and low-voltage non-residents; annotate the data and specify the staff to be assigned to each work order; Extract features from the model through the analysis layer, define the loss function and optimizer, load data using DataLoader, and iteratively train the model; Use the trained model to classify and predict the new work order information text, obtain the category label of each word, and extract key information based on the label; Determine the user's location, power supply department, and power supply unit based on the work order information to be processed; The identified work orders are dispatched to staff members according to the preset dispatching rules.

5. The method for receiving and dispatching work orders based on full-process automation as claimed in claim 1, characterized in that: It also includes smart phone return visits; The smart phone interview includes the following steps: Deployment of smartphone development environment; Write a script, set a scheduled task based on the work order log, and automatically call the work order phone number at 3 pm the next day; Define voice prompt text, conduct satisfaction surveys, set up intelligent voice, and record satisfaction scores; recognize user voice input, convert customer voice into text, train keywords for conversion to manual and trigger the conversion to manual function.

Images