Electric power customer service work order type identification method and related equipment

Abstract

The invention provides an electric power customer service work order type identification method and related equipment. The method comprises the following steps: acquiring an electric power customer service work order; performing semantic feature extraction on the electric power customer service work order based on the content information, the relative position information and the absolute position information of the electric power customer service work order to obtain a first semantic feature; performing feature extraction of long-distance text information on the first semantic feature to obtain a second semantic feature; and performing type identification based on the second semantic feature to obtain a type identification result of the electric power customer service work order.

Description

Technical Field

[0001] The present disclosure relates to the field of data processing, and in particular to a method for identifying the type of an electric power customer service work order and related equipment. Background Art

[0002] The traditional work order processing process requires customer service staff to manually organize customer requests into work orders. This process is time-consuming, labor-intensive, and prone to errors. Existing work order classification technologies based on machine learning models usually use a combination of feature engineering and machine learning. They rely too much on professional knowledge and have difficulty capturing deep semantic features in the context. The mainstream models used in current deep learning work order recognition technology have difficulty effectively capturing the fine-grained semantic differences in power professional expressions, and effectively processing long-distance dependencies between words, resulting in low accuracy in power work order recognition. In addition, with the rapid development of artificial intelligence technology, hackers are constantly creating new attacks against neural networks. Adversarial sample attacks are becoming a major threat to various neural network models. Most traditional work order recognition technologies do not consider the robustness of their recognition technology in the face of attacks, resulting in the high vulnerability of existing work order recognition methods when facing adversarial sample attacks. Summary of the Invention

[0003] The present disclosure proposes a method for identifying the type of electric power customer service work order and related equipment to solve the above-mentioned technical problems to a certain extent.

[0004] In a first aspect of the present disclosure, a method for identifying the type of an electric power customer service work order is provided, comprising:

[0005] Obtain power customer service work orders;

[0006] Extracting semantic features of the electric power customer service work order based on the content information, relative position information, and absolute position information of the electric power customer service work order to obtain a first semantic feature;

[0007] performing feature extraction of long-distance text information on the first semantic feature to obtain a second semantic feature;

[0008] Type identification is performed based on the second semantic feature to obtain a type identification result of the power customer service work order.

[0009] In a second aspect of the present disclosure, a device for identifying the type of an electric power customer service work order is provided, comprising:

[0010] Acquisition module, used to obtain power customer service work orders;

[0011] A first feature extraction module is configured to extract semantic features of the electric power customer service work order based on the content information, relative position information, and absolute position information of the electric power customer service work order to obtain a first semantic feature;

[0012] A second feature extraction module is used to extract the feature of long-distance text information from the first semantic feature to obtain a second semantic feature;

[0013] An identification module is used to perform type identification based on the second semantic feature to obtain a type identification result of the power customer service work order.

[0014] In a third aspect of the present disclosure, an electronic device is provided, comprising one or more processors, a memory; and one or more programs, wherein the one or more programs are stored in the memory and executed by the one or more processors, and the programs include instructions for executing the method described in the first aspect.

[0015] According to a fourth aspect of the present disclosure, a non-volatile computer-readable storage medium containing a computer program is provided. When the computer program is executed by one or more processors, the processors are caused to execute the method described in the first aspect.

[0016] In a fifth aspect of the present disclosure, a computer program product is provided, comprising computer program instructions, which, when executed on a computer, enable the computer to execute the method described in the first aspect.

[0017] From the above, it can be seen that the present disclosure provides a method and related equipment for identifying the type of electric customer service work orders. By fusing and extracting the first semantic feature through multi-dimensional information (content, relative position, absolute position), it can more comprehensively capture the semantic information of the work order and avoid the semantic loss caused by a single information dimension. The long-distance text information feature extraction of the first semantic feature effectively solves the limitation of traditional methods that are difficult to handle long text dependencies and improves the understanding of complex semantic structures. Finally, type identification is performed based on the second semantic feature, which can significantly improve the accuracy and comprehensiveness of electric customer service work order type identification, and provide more accurate and efficient work order classification support for electric customer service business. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present disclosure or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 A schematic diagram of a type identification architecture for electric customer service work orders according to an embodiment of the present disclosure.

[0020] Figure 2 Schematic diagram of the hardware structure of an exemplary electronic device according to an embodiment of the present disclosure.

[0021] Figure 3 Schematic diagram of a flow chart of a method for identifying the type of an electric power customer service work order according to an embodiment of the present disclosure.

[0022] Figure 4 Schematic diagram of a method for identifying the type of an electric power customer service work order according to an embodiment of the present disclosure.

[0023] Figure 5 A schematic diagram of standardized labeling and preprocessing of electric power customer service work orders according to an embodiment of the present disclosure.

[0024] Figure 6 This is a schematic diagram of the first semantic feature extraction according to an embodiment of the present disclosure.

[0025] Figure 7 This is a schematic diagram of the second semantic feature extraction according to an embodiment of the present disclosure.

[0026] Figure 8 Schematic diagram of a device for identifying the type of an electric customer service work order according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0027] In order to make the objectives, technical solutions and advantages of the present disclosure more clearly understood, the present disclosure is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0028] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present disclosure should have the usual meanings understood by people with ordinary skills in the field to which the present disclosure belongs. The "first", "second" and similar words used in the embodiments of the present disclosure do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative position relationships. When the absolute position of the described object changes, the relative position relationship may also change accordingly.

[0029] It is understandable that before using the technical solutions disclosed in the various embodiments of this disclosure, the type, scope of use, usage scenarios, etc. of the personal information involved in this disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0030] For example, in response to a user's active request, a prompt message is sent to the user to clearly inform the user that the operation requested will require the acquisition and use of the user's personal information. This allows the user to independently choose whether to provide personal information to the electronic device, application, server, storage medium, or other software or hardware that performs the operations of the disclosed technical solution based on the prompt message.

[0031] It is understandable that the above notification and user authorization process are merely illustrative and do not limit the implementation of the present disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0032] Figure 1 A schematic diagram showing the type identification architecture of the power customer service work order according to an embodiment of the present disclosure is shown. Figure 1 The type identification architecture 100 for electric power customer service work orders may include a server 110, a terminal 120, and a network 130 that provides a communication link. The server 110 and the terminal 120 may be connected via a wired or wireless network 130. The server 110 may be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, security services, and CDN.

[0033] Terminal 120 can be implemented in hardware or software. For example, when implemented in hardware, terminal 120 can be any electronic device with a display screen that supports page display, including but not limited to smartphones, tablet computers, e-book readers, laptop computers, and desktop computers. When terminal 120 is implemented in software, it can be installed in the electronic devices listed above; it can be implemented as multiple software or software modules (such as software or software modules used to provide distributed services), or it can be implemented as a single software or software module, and no specific limitations are given here.

[0034] It should be noted that the method for identifying the type of the power customer service work order provided in the embodiment of the present application can be executed by the terminal 120 or by the server 110. It should be understood that Figure 1 The number of terminals, networks, and servers in the embodiment is for illustration only and is not intended to limit the number of terminals, networks, and servers.

[0035] Figure 2 FIG. 2 shows a schematic diagram of the hardware structure of an exemplary electronic device 200 provided in an embodiment of the present disclosure. Figure 2As shown, electronic device 200 may include: processor 202, memory 204, network module 206, peripheral interface 208 and bus 210. Processor 202, memory 204, network module 206 and peripheral interface 208 are connected to each other through bus 210 in communication with each other within electronic device 200.

[0036] The processor 202 may be a central processing unit (CPU), a neural network processor (NPU), a microcontroller (MCU), a programmable logic device, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), or one or more integrated circuits. The processor 202 may be used to perform functions related to the technology described in this disclosure. In some embodiments, the processor 202 may also include multiple processors integrated into a single logical component. For example, Figure 2 As shown, the processor 202 may include a plurality of processors 202a, 202b, and 202c.

[0037] The memory 204 may be configured to store data (eg, instructions, computer code, etc.). Figure 2 As shown, the data stored in the memory 204 may include program instructions (for example, program instructions for implementing the type identification method of the power customer service work order of the embodiment of the present disclosure) and data to be processed (for example, the memory may store configuration files of other modules, etc.). The processor 202 may also access the program instructions and data stored in the memory 204, and execute the program instructions to operate on the data to be processed. The memory 204 may include a volatile storage device or a non-volatile storage device. In some embodiments, the memory 204 may include a random access memory (RAM), a read-only memory (ROM), an optical disk, a magnetic disk, a hard disk, a solid-state drive (SSD), a flash memory, a memory stick, etc.

[0038] The network module 206 can be configured to provide the electronic device 200 with communication with other external devices via a network. The network can be any wired or wireless network capable of transmitting and receiving data. For example, the network can be a wired network, a local wireless network (e.g., Bluetooth, WiFi, near field communication (NFC)), a cellular network, the Internet, or a combination thereof. It will be appreciated that the type of network is not limited to the specific examples above. In some embodiments, the network module 206 can include any number of network interface controllers (NICs), radio frequency modules, transceivers, modems, routers, gateways, adapters, cellular network chips, and the like.

[0039] The peripheral interface 208 can be configured to connect the electronic device 200 to one or more peripheral devices to implement information input and output. For example, the peripheral devices can include input devices such as a keyboard, a mouse, a touchpad, a touch screen, a microphone, and various sensors, and output devices such as a display, a speaker, a vibrator, and an indicator light.

[0040] The bus 210 can be configured to transmit information between the various components of the electronic device 200 (e.g., the processor 202, the memory 204, the network module 206, and the peripheral interface 208), such as an internal bus (e.g., a processor-memory bus), an external bus (USB port, PCI-E bus), etc.

[0041] It should be noted that although the architecture of the electronic device 200 shown above only shows the processor 202, the memory 204, the network module 206, the peripheral interface 208, and the bus 210, in a specific implementation, the architecture of the electronic device 200 may also include other components necessary for normal execution. In addition, those skilled in the art will understand that the architecture of the electronic device 200 may also include only the components necessary to implement the embodiments of the present disclosure, and does not necessarily include all the components shown in the figure.

[0042] Automated identification of power work orders is an inevitable requirement for the digital transformation of power systems. In recent years, the massive growth of work order data and the increase in the variety of work order data types associated with the construction of new power systems have posed new challenges to work order classification and identification. Currently, the processing of work orders by power customer service personnel requires manual organization of customer requests into work orders, which is time-consuming, labor-intensive, and prone to errors. Furthermore, existing work order recognition technologies based on machine learning and deep learning lack adaptability to power industry terminology, fail to fully extract local and global features of work order data, and suffer from low robustness against noisy data and adversarial attacks. Traditional work order processing requires customer service personnel to manually organize customer requests into work orders, a time-consuming, labor-intensive, and error-prone process. Machine learning-based work order classification techniques typically combine feature engineering and machine learning. These techniques first extract data features based on power industry expertise, then use models such as word embeddings to vectorize the text and train a classifier. However, these approaches suffer from a heavy reliance on specialized knowledge and struggle to capture deep semantic features within the context. Mainstream models used in deep learning-based work order recognition technology include recurrent neural networks, transformer models (Transformer, BERT, RoBERTa), and others. However, the power customer service work order recognition method based on the self-attention mechanism of the standard Transformer architecture has difficulty effectively capturing the fine-grained semantic differences in power professional expressions and effectively processing long-distance dependencies between words, resulting in low accuracy in power work order recognition. In addition, with the rapid development of artificial intelligence technology, hackers are constantly creating new attacks against neural networks. Adversarial sample attacks are becoming a major threat to various neural network models. Most traditional work order recognition technologies do not consider the robustness of their recognition technology in the face of attacks, resulting in high vulnerability of existing work order recognition methods to adversarial sample attacks.

[0043] Therefore, how to improve the accuracy and robustness of automatic identification of power work orders has become a technical problem that needs to be solved urgently.

[0044] In view of this, the embodiments of the present disclosure provide a method and related equipment for identifying the type of electric power customer service work orders. By fusing and extracting the first semantic feature through multi-dimensional information (content, relative position, absolute position), the semantic information of the work order can be more comprehensively captured, avoiding the semantic loss caused by a single information dimension. The first semantic feature is subjected to long-distance text information feature extraction, which effectively solves the limitation of traditional methods in handling long text dependencies and improves the understanding of complex semantic structures. Finally, type identification is performed based on the second semantic feature, which can significantly improve the accuracy and comprehensiveness of electric power customer service work order type identification, and provide more accurate and efficient work order classification support for electric power customer service business.

[0045] See also Figure 3 , Figure 3A schematic flow chart of a method for identifying the type of a power customer service work order according to an embodiment of the present disclosure is shown. The method for identifying the type of a power customer service work order according to an embodiment of the present disclosure can be deployed on a terminal or a server. Figure 3 In the embodiment, the method 300 for identifying the type of the electric power customer service work order may further include the following steps.

[0046] In step S310, obtain the power customer service work order.

[0047] A power customer service work order refers to a work document or electronic record generated by a power company during the customer service process to record, track, and address customer issues and related business needs. A power customer service work order can include basic customer information, such as name, address, and contact information, as well as detailed information about the issue, work order status, and processing history.

[0048] Specifically, within power companies, there are multiple ways to access customer service work orders. For example, the customer service business system, as a core platform, undertakes the important task of recording the entire communication process between customers and customer service. When customers initiate inquiries, complaints, or repair requests through various channels such as telephone, online customer service, and mobile applications, the system will respond quickly and automatically generate a customer service work order that contains the conversation content, customer details, and specific business needs. By logging into the system and filtering queries in the work order management module based on precise conditions such as time range, customer information, and business type, you can quickly obtain the required work order.

[0049] Power companies also collect customer feedback through various external channels and generate customer service tickets. The customer service hotline is a crucial window for direct communication with customers. Call takers carefully record the content of customer calls, including customer information, detailed descriptions of issues, and specific needs. This information is then entered into the customer service system to generate a ticket. Online customer service and feedback sections on official websites also provide convenient feedback channels. Companies assign dedicated personnel to regularly review and address this feedback, ensuring that customer issues receive timely attention and generate tickets. At the same time, social media platforms have become a new platform for companies to interact with their customers. Companies monitor customer comments and private messages on official accounts, organizing customer issues into tickets for subsequent processing. Furthermore, power companies can also obtain valuable customer service tickets by collaborating with third-party platforms, such as consumer protection agencies and industry regulators, to collect electricity-related customer feedback.

[0050] Data interfaces provide a more efficient and flexible way for power companies to access customer service work orders. On the one hand, power companies connect with other related systems, such as marketing and metering systems, enabling data transmission and sharing between these systems through data interfaces. For example, when a customer's outstanding bills appear in the marketing system, this information can be quickly transmitted to the customer service system via the data interface, automatically generating a work order for outstanding bills, improving the timeliness and accuracy of work order processing. On the other hand, power companies have developed and opened API interfaces, allowing external partners or developers to access customer service work order data in accordance with specifications and standards. This initiative not only promotes business innovation and collaboration among power companies but also opens up more channels for obtaining work order data. External partners or developers can develop their own applications based on the provided development documentation and sample code, efficiently accessing customer service work order data through the API interface, and achieving broader business collaboration.

[0051] In step S320 , semantic features of the electric power customer service work order are extracted based on the content information, relative position information, and absolute position information of the electric power customer service work order to obtain a first semantic feature.

[0052] Content information refers to the specific information contained in the text directly presented in a power customer service work order, providing a detailed description and record of the events and issues involved in the work order. Relative position information refers to the relative positional relationships between different words in the power customer service work order, including the order of sentences and the connections between paragraphs. Different relative positional relationships often imply different semantic logic. By analyzing relative position information, we can better understand the overall structure and semantic context of the work order content. Absolute position information refers to the exact position of each character, word, or sentence in the power customer service work order, typically expressed in the form of character index, line number, column number, etc. Absolute position information can provide more refined contextual clues for semantic feature extraction. For example, in a long text, the position of a keyword may affect its semantic understanding. If the word "fault" appears at the beginning of the work order, it may be more likely to describe the initial state of the problem; however, if it appears at the end, it may be more likely to summarize the problem resolution results. By considering absolute position information, we can more comprehensively capture the semantic details in the text and improve the accuracy of semantic feature extraction. The first semantic feature can refer to the result of semantic feature extraction based on the content, relative position, and absolute position information of the power customer service work order. It is a preliminary abstraction and representation of the work order's semantics. The first semantic feature integrates multiple aspects of the work order information and can more comprehensively reflect the semantic connotation of the work order. It provides a foundation for subsequent semantic analysis and processing. For example, in the type identification of power customer service work orders, the first semantic feature can be used as input data for further analysis and classification using machine learning algorithms or deep learning models, thereby accurately determining the type of work order and supporting efficient work order processing.

[0053] In some embodiments, semantic feature extraction is performed on the electric power customer service work order based on the content information, relative location information, and absolute location information of the electric power customer service work order to obtain a first semantic feature, including:

[0054] Determining an attention weight of each word in the content information and the relative position information;

[0055] Semantic features are extracted from the text content based on the attention weight and the absolute position information to obtain the first semantic feature.

[0056] Based on the content information (each word in the text), relative position information (such as word order and paragraph association), and absolute position information (such as character index and line number) of a power customer service work order, primary semantic features that accurately reflect the work order's semantics are extracted. This provides a foundation for subsequent semantic analysis (such as work order classification and intent recognition). An attention mechanism can be used to combine content information and relative position information to assign an attention weight to each word. The attention weight reflects the importance of the word in expressing the work order's semantics. The attention weight provides an important basis for subsequent semantic feature extraction. When extracting semantic features, different words are weighted according to the attention weight, highlighting the semantic contribution of important words and suppressing unimportant words. The incorporation of absolute position information can more accurately capture the semantic characteristics of words at different positions. By combining attention weights and absolute position information, semantic features are extracted from text content. Deep learning models (such as Transformer and BERT) can be used to implement this process. These models can learn semantic patterns and features in text and represent the text as high-dimensional semantic vectors, namely primary semantic features.

[0057] Specifically, see Figure 4 , Figure 4 A schematic diagram of a method for identifying the type of an electric power customer service work order according to an embodiment of the present disclosure is shown. Figure 4In this paper, we first collect customer service work order data from real-world work environments and standardize the training data based on actual production requirements. We then clean and preprocess the work order data to remove special symbols, repeated fields, and stop words. Secondly, we extract semantic features from the customer service work order data. The DeBERTa-Chinese model, optimized for long-distance semantic dependencies and complex semantic understanding tasks, is selected for data semantic feature extraction. The input data is the customer service work order data that has undergone the previous preprocessing and data cleaning steps. After fine-tuning the DeBERTa-Chinese model, the output data is the feature vector for each customer service work order. DeBERTa (Decoding-enhanced BERT with disentangled attention) is a Transformer-based neural language model pre-trained on a large raw text corpus using self-supervised learning. DeBERTa significantly improves model performance and generalization by introducing two new techniques: a disentangled attention mechanism and an enhanced mask decoder. The DeBERTa-Chinese model is a Chinese DeBERTa model based on the DeBERTa model, pre-trained using a Chinese domain dataset, making it more suitable for Chinese data processing.

[0058] In some embodiments, semantic feature extraction is performed on the electric power customer service work order based on the content information, relative location information, and absolute location information of the electric power customer service work order to obtain a first semantic feature, including:

[0059] Extracting semantic features of the electric power customer service work order based on the trained work order recognition model to obtain a first semantic feature;

[0060] The work order recognition model is obtained by pre-training based on a preset vocabulary and adjusting model parameters based on adversarial training, including:

[0061] The work order recognition model is obtained by pre-training an initial language model based on a preset vocabulary, wherein the preset vocabulary is updated based on a first corpus in historical electric power customer service work orders that meets preset requirements and a second corpus whose semantic similarity with the first corpus is within a preset range;

[0062] In response to the amount of words updated in the preset vocabulary reaching a preset number, the work order recognition model is fine-tuned based on the updated corpus to update the model parameters of the work order recognition model.

[0063] The trained work order recognition model is used to extract semantic features from the power customer service work order to obtain the first semantic feature. The initial language model can be pre-trained based on a preset vocabulary. The preset vocabulary is dynamically updated based on historical power customer service work orders. Specifically, the preset vocabulary is continuously enriched by incorporating first corpus that meets preset requirements and second corpus whose semantic similarity with the first corpus is within a preset range. When the number of words in the preset vocabulary reaches a preset number, the updated corpus can be used to fine-tune the work order recognition model and adjust the model parameters. At the same time, an adversarial training mechanism is introduced during the model training process to further optimize the model parameters, giving the model stronger semantic understanding and feature extraction capabilities, and ultimately achieving accurate extraction of semantic features of power customer service work orders.

[0064] In some embodiments, in response to the amount of words updated in the preset vocabulary reaching a preset number, fine-tuning the work order recognition model based on the updated corpus to update the model parameters of the work order recognition model includes:

[0065] Normalize the updated corpus to obtain a standardized vector;

[0066] determining an output difference based on a first output of the work order recognition model for the normalized vector before adding the adversarial perturbation and a second output of the work order recognition model for the normalized vector after adding the adversarial perturbation;

[0067] The model parameters are updated to maximize the adversarial disturbance while keeping the output difference within a preset threshold.

[0068] Among them, when the number of updated words in the preset vocabulary reaches a preset number, the updated corpus in the vocabulary is standardized and converted into a standardized vector. Then, the work order recognition model processes the standardized vector before and after adding the adversarial disturbance, obtains the first output and the second output, and determines the output difference accordingly. Then, by updating the model parameters, the adversarial disturbance is maximized while keeping the output difference within the preset threshold, so as to adjust and optimize the model parameters and improve the robustness and generalization ability of the model. This can enhance the adaptability of the work order recognition model to different input changes. By maximizing the adversarial disturbance training, the model can still maintain a stable output when facing small but disruptive input changes, reducing the performance degradation caused by input disturbances, thereby improving the accuracy and reliability of the model in extracting semantic features of power customer service work orders, and better adapting to the actual complex and changeable work order environment.

[0069] In some embodiments, pre-training an initial language model based on a preset vocabulary to obtain the work order recognition model includes:

[0070] Standardizing the fault type labeling and data preprocessing of the historical power customer service work orders to obtain training samples; wherein the preprocessing includes at least one of deleting blank work order records, deleting work orders containing only special symbols or numbers, deleting personal sensitive data such as user account numbers, telephone numbers, and detailed addresses in work order records, deleting duplicate work order records, and deleting stop words;

[0071] The initial language model is pre-trained based on the training samples to minimize the loss function of the pre-training to obtain the work order recognition model.

[0072] The process begins with standardized labeling of historical power customer service work orders to clarify the fault type of the work order. A series of data preprocessing operations are also performed, such as removing blank spaces, special symbols or numbers, personal sensitive data, duplicate work order records, and stop words. This generates standardized and high-quality training samples. These training samples are then used to pre-train the initial language model. During the training process, the model parameters are continuously adjusted to gradually minimize the pre-training loss function, ultimately resulting in a work order recognition model with specific semantic understanding capabilities. This significantly improves the quality and standardization of the training samples, effectively removes noise data and irrelevant information, and allows model training to focus more on key content. The pre-trained work order recognition model can better capture the semantic features of power customer service work orders, improve the recognition accuracy of different fault types, and provide more reliable support for subsequent power customer service business applications based on this model (such as work order classification and fault prediction), helping to improve the intelligence level and service efficiency of power customer service.

[0073] Specifically, see Figure 5 , Figure 5 A schematic diagram of standardized labeling and preprocessing of electric power customer service work orders according to an embodiment of the present disclosure is shown. Standardized labeling can refer to labeling electric power customer service work order data into eight types, such as power anomaly, poor line contact, electric energy meter anomaly, meter box failure, faulty power outage, voltage instability, phase loss, and power outage due to arrears; preprocessing electric power customer service data includes six specific operations: deleting blank work order records, deleting work orders containing only special symbols or numbers, deleting personal sensitive data such as the user's account number, telephone number, and detailed address in the work order record, deleting duplicate work order records, and deleting stop words. The word frequency statistics of the electric power customer service work order data can be performed and an incremental electric power field dictionary can be constructed, and the vocabulary of the current Chinese DeBERTa model can be expanded in real time using electric power professional vocabulary; then the Chinese DeBERTa model is fine-tuned, and the fine-tuned model is used to generate word embeddings for the electric power customer service work order text as the feature vector of the text. As Figure 6 As shown, Figure 6 A schematic diagram of first semantic feature extraction according to an embodiment of the present disclosure is shown. Figure 6In the process, since the traditional professional vocabulary expansion relies on manual sorting of high-frequency words, it is difficult to cover emerging equipment terms. Therefore, based on the traditional vocabulary expansion method, the present invention proposes an incremental domain dictionary construction technology. First, the pre-processed electric customer service work order text is segmented. After segmentation, the frequency of electric customer service work order text is counted, and the top 100 electric professional vocabulary with the highest frequency in the existing electric customer service work order data is used to expand the original vocabulary of the Chinese DeBERTa model. Then, the electric customer service work orders received in the current month are pre-processed and segmented on a monthly basis, and the semantic similarity between the newly collected electric customer service work orders and the existing electric professional terms in the dictionary is calculated. New vocabulary with high semantic similarity is automatically added to the existing dictionary to realize the dynamic evolution of the domain dictionary. The formula for calculating semantic similarity is:

[0074]

[0075] Among them, w new Indicates the newly collected power professional terms (such as "virtual power plant"), w base Indicates the existing electric power professional terms in the dictionary (such as "power plant"), V w The word vector representing the vocabulary used to calculate similarity. For example, the word semantic similarity threshold can be set to 0.7. If the similarity (Similarity(w new ,w base )) is greater than 0.7, it is recorded as a new power term (w new ), when the cumulative number of new power terms reaches 50, the model lightweight fine-tuning is triggered. By using the dictionary data to fine-tune the model, the resource overhead of full training is avoided.

[0076] The incremental domain dictionary can be applied to the pre-trained Chinese DeBERTa model, which can then be used for further research. The Chinese DeBERTa model is pre-trained using the Masked Language Model (MLM) approach on the WuDaoCorpora Chinese corpus. During pre-training, the DeBERTa model addresses attention bias in the self-attention mechanism using two techniques: attention decoupling and enhanced masked decoders, while also improving the performance of the pre-trained model.

[0077] Attention decoupling is a unique technology of the DeBERTa model for solving the attention bias in the self-attention mechanism. It uses two independent attention mechanisms to encode the content and position of a word respectively. Through the attention decoupling mechanism, the model can better process text and process different types of information independently. It enhances the adaptability to various natural language understanding tasks. This position attention mechanism enables the model to effectively understand the structure of the text and efficiently capture long-distance dependencies. Specifically, the DeBERTa model can calculate the attention weight of each word in the input sequence through the decoupling matrix of content and position. For example, the calculation formula for the attention weight of the word pair (i, j) is:

[0078]

[0079] Among them, H is the content vector, P is the relative position vector of word i relative to word j, and the cross-attention score between words can be decomposed into: content to content, content to position, and position to content.

[0080] Furthermore, the formula for calculating the relative distance between two words is:

[0081]

[0082] Here, k is a hyperparameter that controls the maximum possible relative distance between i and j. For example, the value of k can be 512.

[0083] Since the decoupled attention only considers content and relative position, but not absolute position information, DeBERTa uses an enhanced masked decoder (EMD) to introduce absolute position information. At the same time, the enhanced masked decoder can improve the performance of the masked language model pre-training method used by DeBERTa. The enhanced masked decoder has two input blocks, labeled H and I, where H represents the hidden state of the previous Transformer layer and I represents all the information required for decoding. A DeBERTa model can have multiple EMD blocks. The EMD block is connected after all the Transformer layers of the DeBERTa model and before the Softmax layer to add absolute position information to the model. In DeBERTa, the number of EMD blocks is generally n=2, where I of the first layer of EMD is the absolute position code, and the output of the second and subsequent EMD layers serves as the input I of the next EMD layer.

[0084] Furthermore, the Chinese DeBERTa model can be fine-tuned using the Scale-Invariant Fine-Tuning (SiFT) method. This method incorporates adversarial training algorithms in addition to traditional fine-tuning techniques to improve the generalization and robustness of deep learning fusion models. Specifically, the main implementation algorithms of Scale-Invariant Fine-Tuning (SiFT) include embedding normalization and adversarial perturbation constraints.

[0085] Embedding normalization: The embedding vector of the input word is set to x. In the embedding normalization step, x is normalized to eliminate the scale differences of different word embeddings. The calculation formula is:

[0086]

[0087] In formula (4), μ x is the mean of the embedding vector x, δ x is the standard deviation, is the normalized embedding vector.

[0088] Adversarial perturbation constraints: Adversarial perturbation generation constraint technology in the standardized embedding space In , we generate an adversarial perturbation δ that maximizes the model prediction difference, and set a threshold ε to ensure that the constraint norm does not exceed the specified threshold. The formula is expressed as:

[0089]

[0090] Among them, ||δ|| p Indicates L p norm constraint, L is the loss function, f is the forward calculation process of the model, and θ is the model parameter.

[0091] SiFT addresses the instability of traditional adversarial training in large models by combining embedding normalization and adversarial perturbation constraints. This allows perturbation generation to be restricted to the normalized embedding space, rather than relying on the absolute scale of the original embeddings, thereby improving model generalization and training efficiency.

[0092] Finally, this module adds a fully connected layer after the fine-tuned Chinese DeBERTa to extract the feature vector of each power customer service work order. Formula (6) shows the process of extracting the intrinsic semantic features of the power customer service work order, where Fine-Tuning represents the fine-tuning operation of the model, and W i (i=1,2,3...n) represents a pre-processed work order record, STInFeat i (i=1,2,3...n) represents the feature vector of a work order record after being processed by this module, and DeB-C represents the Chinese DeBERTa model used in this study.

[0093] STInFeat i =Fine-Tuning(DeB-C(W i )) (6)

[0094] In step S330 , feature extraction of long-distance text information is performed on the first semantic feature to obtain a second semantic feature.

[0095] The second semantic feature is derived by extracting long-range textual information from the first semantic feature. This is typically achieved using the deep learning attention mechanism and recurrent neural network (RNN) variants (such as long short-term memory (LSTM) networks and gated recurrent units (GRU)), or the Transformer architecture. The attention mechanism enables the model to focus on key parts of the text relevant to the current semantic understanding, even if these parts are located farther away in the text. For example, when processing an electric customer service ticket such as "A customer reported frequent power outages last month, which were restored after repairs, but similar issues recurred this month, accompanied by voltage fluctuations," the attention mechanism allows the model to focus on key information such as "frequent outages," "restored after repairs," "recurred," and "voltage fluctuations," even if these are scattered across the text. LSTMs and GRUs use gating mechanisms to mitigate the vanishing gradient problem of traditional RNNs, effectively capturing long-range dependencies and preserving and transmitting information from distant locations. The self-attention mechanism of the Transformer architecture handles long-range dependencies more directly, eliminating the sequential processing required by RNNs and enabling parallel computation for improved efficiency. Through these techniques, the model integrates long-range textual information from the first semantic feature to extract a more comprehensive and advanced second semantic feature. This can comprehensively consider long-distance information in the text and avoid semantic understanding deviations caused by ignoring key information. For example, in the above-mentioned work order, the persistence and changes of the customer's power outage problem can be accurately grasped, and the accuracy of the semantic understanding of the work order can be improved. The second semantic feature integrates long-distance text information and contains richer semantic connotations, which helps subsequent tasks (such as work order classification and intent recognition) to more comprehensively understand the work order content and improve task performance. Through the effective processing of long-distance text information, the model can work better on different types and styles of power customer service work orders, improve the generalization ability of the model, and adapt to a wider range of practical application scenarios.

[0096] See also Figure 4In order to further extract the deep semantic features of the power customer service work order text, the Deep Pyramid Convolutional Neural Networks for Text Categorization (DPCNN) that can obtain long-distance text dependencies is selected to further process the work order data from the extracted feature vectors, complete the deep semantic information extraction and ultimately achieve work order recognition and classification. DPCNN extracts long-distance dependencies by superimposing the network, solving the problem that TextCNN cannot obtain long-distance text dependencies. In order to complete the recognition of power customer service work order data, the present invention adds a fully connected layer after the DPCNN model, sets the output dimension to 8, sets the cross-entropy loss function to calculate the loss, and completes the recognition of the data through the fully connected layer. Finally, the result of the deep learning fusion model for power customer service work order recognition is output.

[0097] In some embodiments, extracting features of long-distance text information from the first semantic feature to obtain a second semantic feature includes:

[0098] Performing feature alignment on the first semantic feature to obtain a corresponding word embedding vector;

[0099] Based on the word embedding vector, multiple levels of equal-length convolution and pooling are alternately stacked to obtain the second semantic feature.

[0100] First, feature alignment is performed on the first semantic features. This process aims to align first semantic features from different sources or representations in the semantic space, making them comparable and fusible. After alignment, the corresponding word embedding vector is obtained. A word embedding vector maps words in a text into a high-dimensional vector space. It captures the semantic information of the words and places words with similar semantics close together in the vector space. Multi-level equal-length convolutions are then performed based on the word embedding vectors. Convolution extracts local semantic features by sliding a convolution kernel across the word embedding vectors. Multi-level convolutions use convolution kernels of different sizes or perform multiple convolutions to capture semantic information at different scales. Equal-length convolutions ensure that the size of the feature map remains relatively stable during the convolution process, facilitating subsequent processing. Pooling is performed after the convolution operation. Pooling reduces the dimensionality of the feature map, reducing computational effort while extracting the most representative features. Multi-level equal-length convolutions and pooling are stacked alternately, performing one level of convolution, followed by pooling, followed by the next level of convolution and pooling, and so on. In this way, the model can gradually extract hierarchical features from long-distance text information, ultimately obtaining secondary semantic features. Feature alignment and word embedding vectors enable the model to better integrate the long-distance text information from the first semantic features into a unified semantic space, enabling the model to more comprehensively understand the overall semantics of the text. For example, when processing power customer service work orders, the model can accurately capture long-distance information such as the different fault phenomena and time spans involved in the work order, improving the accuracy of the work order semantic understanding. The alternating stacking of multi-level equal-length convolutions and pooling extracts richer and more advanced semantic features. These features better reflect the long-distance dependencies of the text, providing stronger support for subsequent tasks such as work order classification and intent recognition. Effective processing of long-distance text information enables the model to better extract key semantic features for power customer service work orders of various types and styles, improving the model's generalization ability. For example, the model can accurately analyze and process new work order expressions that have not appeared in the training set based on the extracted secondary semantic features.

[0101] Specifically, see Figure 7 , Figure 7A schematic diagram of the second semantic feature extraction according to an embodiment of the present disclosure is shown. The original electric customer service work order data is first subjected to a fine-tuned Chinese DeBERTa model to obtain the first semantic feature vector of the electric customer service work order data. In order to further obtain the deep semantic information of the electric customer service work order data, the electric customer service work order feature vector is input into the deep pyramid convolutional neural network (DPCNN). Through the hierarchical convolution and downsampling operations of DPCNN, the long-distance dependencies and global semantic features of the work order text are extracted, which solves the problem of insufficient modeling capabilities of traditional CNN models (such as TextCNN) for long texts and improves the accuracy of work order recognition. The first layer of DPCNN is the text embedding layer (Embedding Layer), and the input is the work order feature vector STInFeat of the fine-tuned Chinese DeBERTa. i (i=1,2,3...n), the text embedding layer adjusts the dimension and aligns the features of the input work order feature vector, maps the high-dimensional features to the semantic space suitable for the convolution operation, and outputs the word embedding vector X i (i = 1, 2, 3...n). To extract features from long-distance text information, the DPCNN model places equal-length convolutional layers and 1 / 2 pooling layers after the text embedding layer. By alternating multiple levels of equal-length convolutional layers and 1 / 2 pooling layers, it gradually extracts local and global semantic features from the work order text.

[0102] For multiple word embedding vectors X i The work order is composed of the same length, the convolution layer uses a sliding window of size h, selects a feature sequence to perform the convolution operation, and outputs a feature map C i The equal-length convolution method requires adding zero padding symmetrically at the beginning and end of the input sequence to ensure dimension alignment. The convolution kernel size in the convolution operation is W c ∈R h×k , for example, from word X i Calculate the feature C after the convolution layer i The formula is:

[0103] C i =SeLU(W c *X i +b) (7)

[0104] As shown in formula (5), the ReLU activation function used in the original DPCNN model can be replaced with the SeLU activation function. The ReLU activation function is commonly used in neural network models. However, the hard truncation mechanism of ReLU easily leads to gradient information loss, making the model more sensitive to noise and outliers during training, easily leading to neuron inactivation, and affecting the stability of weight updates. Therefore, the present invention modifies the activation function of DPCNN to a SeLU function with adaptive normalization capability. The SeLU function formula is:

[0105]

[0106] Where x is the input from the previous neural network layer, and λ and α are pre-set weight values, for example, λ = 1.0507, α = 1.6733.

[0107] After the equal-length convolutional layers, the feature vectors of the work order data are passed through a 1 / 2 pooling layer to reduce dimensionality, reduce the number of parameters, improve robustness and computational efficiency, and enhance generalization and noise immunity. DPCNN uses a max pooling layer of size 3 and stride 2 to achieve downsampling, compressing the output of the convolutional layer and halving the sequence length. The formula for calculating the i-th vector output by a 1 / 2 pooling layer is:

[0108]

[0109] The output of the pooling layer is m represents the number of convolution filters.

[0110] The feature vector of the Chinese electric power customer service work order is processed through the embedding layer, equal-length convolution layer, and 1 / 2 pooling layer of DPCNN to obtain the deep semantic features of the electric power work order containing the long-distance dependencies in the text, the electric power work order text feature STInFeat i (i=1,2,3...n) The deep semantic feature output of DPCNN is represented as ST i , as shown in formula (10):

[0111] ST i =DPCNN(STInFeat i ) (10)

[0112] You can add a fully connected layer at the end of the DPCNN model, set the output dimension to 8, and use the cross entropy loss function to calculate the loss.

[0113] Specifically, the collected feature vector data of electric power customer service work orders was divided into training, test, and validation sets in a 6:2:2 ratio. The training and test sets were used for model training and testing, while the validation set was used to verify the model's ability to recognize electric power customer service work order data. The deep learning fusion model proposed in this paper was trained and tested, with the input being the feature vectors of electric power customer service work orders and the output being the recognition results of the deep learning fusion model for electric power customer service work orders. Evaluation metrics such as the model's recognition precision, recall, and F1 score were calculated to assess the model's recognition performance.

[0114] Therefore, the present disclosure proposes for the first time an electric power customer service work order recognition method that integrates the DeBERTa model and the DPCNN model, which significantly improves the accuracy of work order recognition. Traditional work order recognition methods often rely on simple word vectors or local feature extraction, which makes it difficult to fully capture the complex semantics and long-distance dependencies in electric power customer service work orders. The present invention combines the attention decoupling mechanism of the DeBERTa model to accurately distinguish the fine-grained semantics of electric power professional terms, and at the same time utilizes the pyramid convolution and downsampling structure of DPCNN to hierarchically extract local details and deep semantic features of the work order text, thereby overcoming the shortcomings of the existing technology.

[0115] The present disclosure also pre-trains a Chinese DeBERTa model based on the WuDaoCorpora Chinese corpus, which has better adaptability to Chinese than the original DeBERTa model. In addition, since the professional vocabulary in the power field is rich and constantly updated, the vocabulary in the pre-trained model may not cover all professional vocabulary in the power field. Therefore, the present invention proposes an incremental power field dictionary construction method, which detects and automatically expands the dictionary in real time, and triggers a lightweight fine-tuning model at the same time, ensuring that the model can accurately segment words and extract features when processing power customer service work orders, thereby improving the professional adaptability of the model.

[0116] Most current deep learning-based work order recognition technologies focus solely on recognition accuracy, while neglecting the model's robustness in the face of noisy data and attacks. This paper innovatively introduces the scale-invariant fine-tuning (SiFT) algorithm when fine-tuning the Chinese DeBERTa model. By embedding standardization and adversarial perturbation constraint techniques, it enhances the model's generalization and robustness. This enables the deep learning fusion model proposed in this paper to maintain a high recognition accuracy rate in the face of adversarial sample attacks or noisy data, thereby improving the model's practical application value.

[0117] In step S340, type recognition is performed based on the second semantic feature to obtain a type recognition result of the power customer service work order.

[0118] Among them, the second semantic feature contains rich semantic information extracted from the power customer service work order. The second semantic feature is mapped to the preset work order type space, and the semantic information such as "frequent power outages" and "voltage instability" in the work order is encoded in the second semantic feature. The mapped features are matched using the trained work order recognition model. The work order recognition model has mastered the correspondence between different work order types and features during the learning phase. When the second semantic feature is input, the work order recognition model will calculate the similarity or probability of the feature with each work order type based on these correspondences. The work order type with the highest similarity or the highest probability can be selected as the recognition result. For example, if a second semantic feature has the highest similarity with the "fault repair" type, then it is determined that the power customer service work order belongs to the "fault repair" type.

[0119] In some embodiments, performing type recognition based on the second semantic feature to obtain a type recognition result of the electric power customer service work order includes:

[0120] The second semantic feature is classified based on the work order recognition model to obtain the type recognition result; wherein, the type recognition result includes one of multiple fault types, and the fault types include abnormal power, poor line contact, abnormal electricity meter, meter box failure, fault power outage, unstable voltage, phase loss or power outage due to unpaid bills.

[0121] After obtaining the second semantic feature extracted from the power customer service work order, it is provided as input to the work order recognition model. The model conducts in-depth analysis and processing of the second semantic feature, classifying the input feature based on its internally learned semantic patterns and the feature differences between different fault types. By calculating the similarity or probability distribution between the feature and the feature space corresponding to each fault type, the fault type corresponding to the second semantic feature is ultimately determined. The output is one of several fault types, including abnormal power consumption, poor line contact, abnormal energy meter, meter box failure, power outage, voltage instability, phase loss, or power outage due to unpaid bills, as the type recognition result. The second semantic feature contains rich semantic information about the work order. The work order recognition model classifies based on these features, more accurately capturing the essence of the fault described in the work order, significantly improving the accuracy of fault type recognition. For example, for work orders with vague or complex descriptions, such as "My home appliances sometimes won't start, and the voltage seems abnormal," the model can accurately classify the fault type as "voltage instability" based on the second semantic feature. Automation performs type identification based on the second semantic feature. Compared with traditional manual classification methods, it can quickly process a large number of power customer service work orders, greatly shorten the time for work order classification, and improve the processing efficiency of the entire customer service process. Accurate type identification results help to quickly divert different types of work orders to the corresponding processing departments or professionals, realizing the rational allocation of resources. For example, "electricity meter anomaly" work orders are promptly assigned to metering professionals for processing, and "fault power outage" work orders are preferentially assigned to emergency repair teams, improving the timeliness and effectiveness of problem solving. By accurately identifying work order types, power customer service departments can better understand user needs and problems, provide targeted solutions and services, and improve user satisfaction and service quality.

[0122] As can be seen, in order to effectively extract the local and global semantics of short texts of power customer service work orders and improve the accuracy of work order recognition, this paper proposes a power customer service work order feature extraction method that combines the DeBERTa and DPCNN models. DeBERTa's attention decoupling mechanism separates content and position encoding, accurately distinguishing the fine-grained semantics of power professional terminology. DPCNN's pyramid convolution and downsampling structure are used to hierarchically extract local details and deep semantic features of the work order text.

[0123] This paper also uses a Chinese dataset to pre-train a Chinese DeBERTa model. By processing a large amount of power work order data, the Chinese DeBERTa model dictionary is expanded. An incremental domain dictionary construction method is proposed. This method uses cosine similarity of word vectors to detect the semantic relevance of newly appearing words in work orders and power terminology in real time. The method automatically expands the dictionary and triggers lightweight fine-tuning, achieving collaborative optimization of the dictionary and model. Furthermore, this paper uses a scale-invariant fine-tuning method to fine-tune the Chinese DeBERTa model, improving the model's generalization and robustness.

[0124] This paper uses the DPCNN model to extract the deep semantic features of work orders, and reserves a cross-modal alignment interface in the DPCNN backend, which provides support for the dynamic fusion of fault site images and text features of power customer service work orders in the future, and adapts to complex scenarios of mixed image and text work orders.

[0125] It should be noted that the method of the embodiments of the present disclosure can be performed by a single device, such as a computer or server. The method of the embodiments of the present disclosure can also be applied in a distributed scenario, where multiple devices cooperate to perform the method. In such a distributed scenario, one of the multiple devices may only perform one or more steps of the method of the embodiments of the present disclosure, and the multiple devices will interact with each other to complete the method.

[0126] It should be noted that the above description is limited to some embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in an order different from that described in the above embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0127] Based on the same technical concept, corresponding to any of the above embodiments and methods, the present disclosure also provides a type identification device for electric customer service work orders, see Figure 8 , the type identification device of the power customer service work order, the device comprising:

[0128] Acquisition module, used to obtain power customer service work orders;

[0129] A first feature extraction module is configured to extract semantic features of the electric power customer service work order based on the content information, relative position information, and absolute position information of the electric power customer service work order to obtain a first semantic feature;

[0130] A second feature extraction module is used to extract the feature of long-distance text information from the first semantic feature to obtain a second semantic feature;

[0131] An identification module is used to perform type identification based on the second semantic feature to obtain a type identification result of the power customer service work order.

[0132] For the convenience of description, the above devices are described as being functionally divided into various modules. Of course, when implementing the present disclosure, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0133] The device of the above embodiment is used to implement the type identification method of the corresponding power customer service work order in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be repeated here.

[0134] Based on the same technical concept, corresponding to any of the above-mentioned embodiment methods, the present disclosure also provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable the computer to execute the type identification method of the electric power customer service work order as described in any of the above embodiments.

[0135] The computer-readable media of this embodiment include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0136] The computer instructions stored in the storage medium of the above embodiment are used to enable the computer to execute the type identification method of the power customer service work order as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0137] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present disclosure (including the claims) is limited to these examples. Within the scope of the present disclosure, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present disclosure as described above, which are not provided in detail for the sake of clarity.

[0138] In addition, to simplify the description and discussion, and so as not to obscure the embodiments of the present disclosure, known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided figures. In addition, devices may be shown in the form of block diagrams to avoid obscuring the embodiments of the present disclosure, and this also takes into account the fact that the details of the implementation of these block diagram devices are highly dependent on the platform on which the embodiments of the present disclosure are to be implemented (i.e., these details should be fully within the purview of those skilled in the art). Where specific details (e.g., circuits) are set forth to describe exemplary embodiments of the present disclosure, it will be apparent to those skilled in the art that the embodiments of the present disclosure may be implemented without these specific details or with variations in these specific details. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0139] Although the present disclosure has been described in conjunction with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may use the embodiments discussed.

[0140] The embodiments of the present disclosure are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present disclosure should be included in the scope of protection of the present disclosure.

Claims

1. A method for identifying the type of an electric power customer service work order, characterized in that: include: Obtain power customer service work orders; Extracting semantic features of the electric power customer service work order based on the content information, relative position information, and absolute position information of the electric power customer service work order to obtain a first semantic feature; performing feature extraction of long-distance text information on the first semantic feature to obtain a second semantic feature; Type identification is performed based on the second semantic feature to obtain a type identification result of the power customer service work order.

2. The method according to claim 1, characterized in that Semantic features of the electric power customer service work order are extracted based on the content information, relative position information, and absolute position information of the electric power customer service work order to obtain a first semantic feature, including: Determining an attention weight of each word in the content information and the relative position information; Semantic features are extracted from the text content based on the attention weight and the absolute position information to obtain the first semantic feature.

3. The method according to claim 1, characterized in that Semantic features of the electric power customer service work order are extracted based on the content information, relative position information, and absolute position information of the electric power customer service work order to obtain a first semantic feature, including: Extracting semantic features of the electric power customer service work order based on the trained work order recognition model to obtain a first semantic feature; The work order recognition model is obtained by pre-training based on a preset vocabulary and adjusting model parameters based on adversarial training, including: The work order recognition model is obtained by pre-training an initial language model based on a preset vocabulary, wherein the preset vocabulary is updated based on a first corpus in historical electric power customer service work orders that meets preset requirements and a second corpus whose semantic similarity with the first corpus is within a preset range; In response to the amount of words updated in the preset vocabulary reaching a preset number, the work order recognition model is fine-tuned based on the updated corpus to update the model parameters of the work order recognition model.

4. The method according to claim 3, characterized in that In response to the amount of words updated in the preset vocabulary reaching a preset number, fine-tuning the work order recognition model based on the updated corpus to update model parameters of the work order recognition model includes: Normalize the updated corpus to obtain a standardized vector; determining an output difference based on a first output of the work order recognition model for the normalized vector before adding the adversarial perturbation and a second output of the work order recognition model for the normalized vector after adding the adversarial perturbation; The model parameters are updated to maximize the adversarial disturbance while keeping the output difference within a preset threshold.

5. The method according to claim 1, wherein Extracting features of long-distance text information from the first semantic feature to obtain a second semantic feature includes: Performing feature alignment on the first semantic feature to obtain a corresponding word embedding vector; Based on the word embedding vector, multiple levels of equal-length convolution and pooling are alternately stacked to obtain the second semantic feature.

6. The method according to claim 3, characterized in that Performing type recognition based on the second semantic feature to obtain a type recognition result of the electric power customer service work order includes: The second semantic feature is classified based on the work order recognition model to obtain the type recognition result; wherein, the type recognition result includes one of multiple fault types, and the fault types include abnormal power, poor line contact, abnormal electricity meter, meter box failure, fault power outage, unstable voltage, phase loss or power outage due to unpaid bills.

7. The method according to claim 3, characterized in that The work order recognition model is obtained by pre-training the initial language model based on a preset vocabulary, including: Standardizing the fault type labeling and data preprocessing of the historical power customer service work orders to obtain training samples; wherein the preprocessing includes at least one of deleting blank work order records, deleting work orders containing only special symbols or numbers, deleting personal sensitive data such as user account numbers, telephone numbers, and detailed addresses in work order records, deleting duplicate work order records, and deleting stop words; The initial language model is pre-trained based on the training samples to minimize the loss function of the pre-training to obtain the work order recognition model.

8. A device for identifying the type of an electric power customer service work order, characterized in that: include: Acquisition module, used to obtain power customer service work orders; A first feature extraction module is configured to extract semantic features of the electric power customer service work order based on the content information, relative position information, and absolute position information of the electric power customer service work order to obtain a first semantic feature; A second feature extraction module is used to extract the feature of long-distance text information from the first semantic feature to obtain a second semantic feature; An identification module is used to perform type identification based on the second semantic feature to obtain a type identification result of the power customer service work order.

9. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to cause a computer to execute the method according to any one of claims 1 to 7.

Images