Power station customer service system construction method and device and power station customer service system

Abstract

The embodiment of the invention provides a power station customer service system construction method and device and a power station customer service system, and the method comprises the steps: building a fault data set according to an obtained energy storage power station technical file; performing fine tuning training on a preset large model based on the fault data set, so that the preset large model learns a corresponding relation among an operation scene, a fault phenomenon and a fault reason of the energy storage power station through fine tuning training; vectorizing the obtained energy storage power station technical file to obtain a vector database in which text information in the energy storage power station technical file corresponds to numerical vectors; and setting the vector database and the preset large model subjected to fine tuning training in a workflow of the energy storage power station customer service system as a reasoning layer, so as to perform reasoning on an input consultation statement to obtain a fault reason. According to the power station customer service system construction method provided by the invention, the customer service system capable of automatically determining the fault reason of the energy storage power station according to the description of the customer can be constructed, manual intervention is reduced, and the problem solving speed is improved.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence and deep learning, and in particular to a method, apparatus and system for constructing a power station customer service system. Background Technology

[0002] During the operation of energy storage power stations, tasks such as equipment monitoring, alarm functions, and fault diagnosis are crucial for ensuring the normal operation of the equipment. However, in recent years, with the rapid development of energy storage power station technology, the complexity of power station equipment has been continuously increasing. Existing customer service systems that rely on manual judgment or simple rule-based question-and-answer systems are unable to quickly provide efficient and accurate solutions when equipment problems occur. As a result, when faced with equipment failures and alarm information, staff often cannot quickly determine the cause of the failure, which in turn affects the efficiency of equipment operation and maintenance and customer experience. In order to handle equipment failures and alarm issues more efficiently, energy storage power stations urgently need an intelligent customer service system that can automatically analyze and provide feedback on fault information, reduce manual intervention, and improve response speed and the accuracy of problem solving. Summary of the Invention

[0003] This disclosure provides a method, apparatus, and system for constructing a power station customer service system, which can automatically determine the cause of a power storage power station failure based on customer descriptions.

[0004] In view of the above problems, firstly, this disclosure provides a method for constructing a power station customer service system, including: Based on the obtained technical documents of the energy storage power station, a fault dataset is established; wherein, the technical documents include the operating scenarios of the energy storage power station, the correspondence between fault phenomena and fault causes; Based on the fault dataset, the preset large model is fine-tuned and trained so that the preset large model learns the correspondence between the operating scenarios, fault phenomena and fault causes of the energy storage power station through the fine-tuning training. The acquired energy storage power station technical documents are vectorized to obtain a vector database that corresponds the text information in the energy storage power station technical documents with numerical vectors. The vector database and the pre-set large model with fine-tuning training are set as the inference layer in the workflow of the energy storage power station customer service system so as to infer the cause of the fault from the input inquiry statement.

[0005] In conjunction with the first aspect, in one possible implementation, establishing a fault dataset based on the acquired energy storage power station technical documents includes: Extract key information from the fault cases in the energy storage power station technical documents; wherein, the key information includes: fault phenomenon, operating scenario and fault cause; Determine the operating scenario corresponding to each fault case, as well as the correspondence between the fault phenomenon and the fault cause; Using a semantic analysis model, based on the correspondence between the fault phenomena and the fault causes, duplicate fault cases in the same operating scenario are identified; Based on the number of repeated failure cases, reduce the number of failure cases that exceed the first threshold and increase the number of cases that are less than the second threshold. After adjusting the number of cases, a fault dataset is established based on the correspondence between the fault phenomena, operating scenarios, and fault causes.

[0006] In conjunction with the first aspect, in one possible implementation, the step of fine-tuning the preset large model based on the fault dataset, so that the preset large model learns the correspondence between the operating scenarios, fault phenomena, and fault causes of the energy storage power station through the fine-tuning training, includes: The fault dataset is divided into a training set and a test set according to a preset ratio; Freeze the pre-trained parameters of the preset large model, excluding the adapter layer and the output layer; The operating scenarios and fault phenomena in the training set are input into the preset large model. Based on the deviation between the fault cause output by the output layer and the corresponding fault cause in the training set, the parameters of the adapter layer are adjusted until the fault diagnosis accuracy of the preset large model reaches a first threshold. The fault diagnosis accuracy is determined based on the deviation between the first fault cause output by the preset large model and the corresponding fault cause in the test set. The first fault cause is the fault cause output by the preset large model based on the operating scenarios and fault phenomena in the test set.

[0007] In conjunction with the first aspect, in one possible implementation, the step of vectorizing the acquired energy storage power station technical documents to obtain a vector database that corresponds the text information in the energy storage power station technical documents with numerical vectors includes: The text information in the technical documents of the energy storage power station is sorted into text that characterizes the operating scenario, fault phenomenon and fault cause; The text is divided into multiple text blocks according to semantics; The text blocks are converted into numerical vectors using a preset encoding model, and a vector database is constructed based on the correspondence between the text blocks and the numerical vectors.

[0008] Secondly, this disclosure provides a power station customer service system construction apparatus, including: A module is established to create a fault dataset based on the acquired technical documents of the energy storage power station; wherein, the technical documents include the operating scenarios of the energy storage power station, the correspondence between fault phenomena and fault causes; The training module is used to fine-tune the preset large model based on the fault dataset, so that the preset large model can learn the correspondence between the operation scenarios, fault phenomena and fault causes of the energy storage power station through the fine-tuning training. The processing module is used to vectorize the acquired energy storage power station technical documents to obtain a vector database that corresponds the text information in the energy storage power station technical documents with numerical vectors. The configuration module is used to set the vector database and the preset large model with fine-tuning training as the inference layer in the workflow of the energy storage power station customer service system, so as to infer the cause of the fault from the input inquiry statement.

[0009] In conjunction with the second aspect, in one possible implementation, the establishment module is further configured to extract key information of fault cases from the energy storage power station technical documents; determine the operating scenario corresponding to each fault case, as well as the correspondence between fault phenomena and fault causes; use a semantic analysis model to determine duplicate fault cases in the same operating scenario based on the correspondence between fault phenomena and fault causes; and reduce the number of fault cases exceeding a first threshold and increase the number of cases less than a second threshold based on the number of duplicate fault cases. After adjusting the number of cases, a fault dataset is established based on the correspondence between the fault phenomena, operating scenarios, and fault causes.

[0010] In conjunction with the second aspect, in one possible implementation, the training module is further configured to divide the fault dataset into a training set and a test set according to a preset ratio; freeze the pre-training parameters of the preset large model except for the adapter layer and the output layer; input the running scenarios and fault phenomena in the training set into the preset large model; and adjust the parameters of the adapter layer according to the deviation between the fault causes output by the output layer and the corresponding fault causes in the training set, until the fault diagnosis accuracy of the preset large model reaches a first threshold.

[0011] In conjunction with the second aspect, in one possible implementation, the processing module is further configured to sort the text information in the energy storage power station technical documents into text representing the operating scenario, fault phenomena and fault causes; split the text into multiple text blocks according to semantics; convert the text blocks into numerical vectors through a preset encoding model; and construct a vector database based on the correspondence between the text blocks and the numerical vectors.

[0012] Thirdly, this disclosure provides a power station customer service system, including: an access layer and an inference layer; The access layer includes a data interface for communicating with the user side, used to obtain the consultation statement sent by the user and transmit the consultation statement to the inference layer; The inference layer is constructed based on any one of the methods in the first aspect, and is used to vectorize the consultation statement through a vector database, match the corresponding text blocks, infer the corresponding fault cause from a fine-tuned preset large model, and output the fault cause to the client side through the interface layer.

[0013] In conjunction with the third aspect, in one possible implementation, the power station customer service system further includes: a preprocessing layer and a collaboration layer; The preprocessing layer includes a preset denoising model, which is used to process the inquiry statements input to the energy storage power station customer service system and output the denoised inquiry statements to the inference layer; wherein, the preset denoising model is constructed based on the energy storage power station domain lexicon and is used to eliminate ambiguity in the input inquiry statements; The collaboration layer is used to establish a communication connection between the user side and the human customer service side through the access layer when the inference layer matches a text block in the consultation statement containing collaboration keywords; and / or If the relevance of the text block matched in the consultation statement by the inference layer is lower than a second threshold, a communication connection is established between the user side and the human customer service side through the access layer; wherein, the relevance of the text block is determined based on the numerical vector of the consultation statement and the numerical vector of the text block in the vector database.

[0014] The beneficial effects of the embodiments disclosed herein include: This disclosure provides a method, apparatus, and system for constructing a power station customer service system, comprising: establishing a fault dataset based on acquired energy storage power station technical documents; wherein the technical documents include the correspondence between the operating scenarios of the energy storage power station, fault phenomena, and fault causes; fine-tuning and training a preset large model based on the fault dataset, enabling the preset large model to learn the correspondence between the operating scenarios of the energy storage power station, fault phenomena, and fault causes through the fine-tuning training; vectorizing the acquired energy storage power station technical documents to obtain a vector database that corresponds text information in the energy storage power station technical documents with numerical vectors; and setting the vector database and the fine-tuned preset large model as an inference layer in the workflow of the energy storage power station customer service system, so as to infer the fault cause from the input inquiry statement. The power station customer service system construction method provided by this disclosure can construct a customer service system that can automatically determine the fault cause of an energy storage power station based on the customer's description, reducing manual intervention and improving the speed of problem resolution. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the method for constructing a power station customer service system according to an embodiment of this disclosure; Figure 2 This is a schematic diagram of the structure of the power station customer service system construction device provided in the embodiments of this disclosure; Figure 3 This is a schematic diagram of the structure of a power station customer service system provided in an embodiment of this disclosure. Detailed Implementation

[0016] This disclosure provides a method, apparatus, and system for constructing a power plant customer service system. Preferred embodiments of this disclosure are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of this disclosure. Furthermore, the embodiments and features described herein can be combined with each other unless otherwise specified.

[0017] This disclosure provides a method for constructing a power plant customer service system, such as... Figure 1 As shown, it can be implemented as follows: S101, Based on the obtained technical documents of the energy storage power station, establish a fault dataset; The technical documents include the operating scenarios of the energy storage power station, the correspondence between fault phenomena and fault causes; S102, Based on the fault dataset, the preset large model is fine-tuned and trained so that the preset large model learns the correspondence between the operating scenarios, fault phenomena and fault causes of the energy storage power station through the fine-tuning training. S103, the obtained energy storage power station technical documents are vectorized to obtain a vector database that corresponds the text information and numerical vectors in the energy storage power station technical documents. S104, The vector database and the preset large model with fine-tuning training are set as the inference layer in the workflow of the energy storage power station customer service system so as to infer the cause of the fault from the input inquiry statement.

[0018] In this embodiment of the disclosure, the technical documents for the energy storage power station may be technical materials related to the maintenance, operation, and malfunctions of the equipment in the energy storage power station, such as operation manuals, maintenance manuals, and fault repair records. These technical documents can be cleaned and deduplicated to obtain key information, and a fault dataset can be obtained based on this key information. The fault dataset may include the operating scenario of the energy storage power station, the fault phenomena, and the corresponding causes of the faults.

[0019] Specifically, the operational scenario can include: environmental data at the time of the fault (e.g., ambient temperature and humidity) and specific parameters of the energy storage power station (e.g., rated energy storage capacity, rated power, and operating cycle). The fault phenomenon can include: equipment information of the faulty device (e.g., specific type, model, and operating time) and specific fault manifestations. The fault cause can include: various factors that led to the fault and corresponding solutions.

[0020] Fine-tuning a pre-defined large model using such a fault dataset enables the model to better learn the correspondence between operating scenarios, fault phenomena, and fault causes in energy storage power stations, thereby improving the accuracy and efficiency of fault diagnosis. The pre-defined large model can be a deep learning model with language understanding, contextual understanding, and reasoning capabilities, and capable of machine learning, such as DeepSeek, LLaMA, and Claude. After fine-tuning the training specifically for energy storage power stations using the fault dataset, the large model can infer the corresponding fault causes based on the input operating scenarios and fault phenomena.

[0021] Vector databases can be built based on an embedding model. This database converts input text information into machine-understandable numerical vectors, which reflect the semantics, features, and relationships of the text. The vector database can also store these numerical vectors and corresponding text information. When a user needs to ask a question to the customer service system, the vector database can convert the user's query into a numerical vector, match the closest numerical vector in the database, and determine the user's intent based on the text information corresponding to the matched numerical vector.

[0022] In this disclosure, the customer service system is a process-oriented system composed of specific workflows, which include multiple work nodes and processing flows between nodes. The inference layer is a work node within this workflow, comprising a vector database obtained from the aforementioned steps and a preset large model. Text information matched by the vector database can be input into the preset large model, and the cause of the fault can be inferred from the preset large model, thereby enabling the automatic output of the fault cause based on the customer's inquiry.

[0023] In another embodiment provided in this disclosure, the step S101 above, "establishing a fault dataset based on the obtained energy storage power station technical documents," can be implemented as follows: Step 1: Extract key information from the fault cases in the energy storage power station technical documents; The key information includes: fault symptoms, operating scenario, and fault cause; Step 2: Determine the operating scenario corresponding to each fault case, as well as the correspondence between the fault phenomenon and the fault cause; Step 3: Using a semantic analysis model, based on the correspondence between the fault phenomena and fault causes, identify repeated fault cases in the same operating scenario; Step 4: Based on the number of repeated failure cases, reduce the number of failure cases that exceed the first threshold and increase the number of cases that are less than the second threshold. Step 5: After adjusting the number of cases, establish a fault dataset based on the correspondence between the fault phenomena, operating scenarios, and fault causes.

[0024] In this embodiment, the acquired technical documents may contain a large amount of content unrelated to fault handling, and this content does not need to be included in the fault dataset. For this content, the technical documents can first be cleaned. In practical applications, regular expressions can be used to extract this key information and remove useless information from the technical documents. After extracting the key information, the extracted information can be integrated to obtain the corresponding operating scenario, fault phenomenon, and fault cause for each fault case.

[0025] The semantic analysis model disclosed herein can be a model capable of analyzing the semantics of text. Using this model, semantic analysis can be performed on each failure case, analyzing the types of failure cases in the same operating scenario and determining the number of failure cases in each type. When the number of cases in a certain type exceeds a first threshold, it can be considered that the number of cases in that type is excessive and needs to be reduced to avoid overfitting during training. For case types with a large number of failure cases, a document hash value deduplication method can be used. The hash value of the key information of each failure case is obtained, these hash values ​​are compared, and cases with duplicate hash values ​​are removed to reduce the number of cases in that type and avoid overfitting during training.

[0026] When the number of cases in a certain category is less than a second threshold, it can be considered that the number of cases in that category is too small, which will affect the generalization ability of the model during fine-tuning training. For case categories with a small number of cases, data can be expanded to increase the number of cases for that category. Specifically, synonyms can be generated based on existing key information (e.g., "overheating" can be generated as "excessive temperature," and "restart" can be generated as "reset"); scenarios can also be expanded based on existing cases to add new cases (e.g., adding "insufficient fan speed leading to overheating" based on "fan stopped causing overheating"). This ensures that each fault case has a sufficient number of cases to fine-tune the preset large model, thereby improving the model's generalization ability.

[0027] It should be noted that the first and second quantity thresholds mentioned above can be set according to the total number of all failure cases to ensure that fine-tuning training does not result in overfitting and that the model has sufficient generalization ability.

[0028] Finally, after adjusting the number of cases as described above, the key information of all the failure cases is matched to obtain the failure dataset.

[0029] In another embodiment provided in this disclosure, step 102 above, "fine-tuning the preset large model based on the fault dataset, so that the preset large model learns the correspondence between the operating scenarios, fault phenomena and fault causes of the energy storage power station through the fine-tuning training," can be implemented as follows: Step 1: Divide the fault dataset into a training set and a test set according to a preset ratio; Step 2: Freeze the pre-trained parameters of the preset large model, excluding the adapter layer and the output layer; Step 3: Input the running scenarios and fault phenomena in the training set into the preset large model. Based on the deviation between the fault causes output by the output layer and the corresponding fault causes in the training set, adjust the parameters of the adapter layer until the fault diagnosis accuracy of the preset large model reaches the first threshold. The fault diagnosis accuracy is determined based on the deviation between the first fault cause output by the preset large model and the corresponding fault cause in the test set; the first fault cause is the fault cause output by the preset large model based on the operating scenario and fault phenomenon in the test set.

[0030] In this embodiment, the preset ratio can be set according to actual needs. For example, the fault dataset can be divided into a training set and a test set in a 7:3 ratio. It should be noted that the training set and the test set cannot contain the same fault cases. The first threshold can be preset and is a value that can judge whether the fault diagnosis accuracy of the preset large model meets the target.

[0031] During training, to preserve the original language understanding and reasoning capabilities of the pre-designed large model and reduce training computation costs and time, only the parameters of the adapter layer and output layer can be adjusted, while the pre-trained parameters of other layers in the large model are frozen. The pre-designed large model in this disclosure can be a large language model based on a Transformer architecture, which may also include an input layer, encoder layer, attention layer, and feedforward neural network layer. During fine-tuning training, the parameters of these layers can be frozen, and only the parameters of the adapter layer and output layer can be adjusted.

[0032] By continuously adjusting the parameters of the adapter layer, the preset large model can correctly output the corresponding fault cause after inputting the running scenario and fault phenomenon. The adapter layer can be a low-rank adapter (LoRA) introduced into the preset large model.

[0033] During training, the running scenarios and fault phenomena in the training set can be input into the input layer of a pre-defined large model. The adapter layer can first call the initial parameters of the low-rank adapter and use these parameters for inference, outputting the inferred cause of the fault. Based on the difference between the inferred fault cause and the fault causes in the training set, a loss function is used to calculate the loss value. Based on the calculated loss value, backpropagation can be performed to adjust the initial parameters of the low-rank adapter until the loss value reaches a pre-set value. The above steps are repeated multiple times until all the content in the training set has been trained using the above method.

[0034] It should be noted that if the adjusted loss value continues to rise during training, the step size of the adapter layer parameter adjustment can be reduced accordingly, allowing the parameters to gradually approach the target parameters and the loss value to decrease again, thus avoiding overfitting.

[0035] After the training process is complete, the fine-tuned large model can be tested using a test set. The operating scenarios and fault phenomena from the test set are input into the large model, which outputs the first cause of the fault. Comparing the first cause of the fault with the corresponding cause in the test set yields the degree of deviation (i.e., the fault diagnosis accuracy). Only when the fault diagnosis accuracy is not less than a first threshold can the pre-set large model be considered to have completed fine-tuning training.

[0036] In practical applications, response speed and robustness can also be used as criteria to determine whether fine-tuning is complete. Response speed can be determined based on the time from the input running scenario and fault phenomenon to the output fault cause of the large model. By adding noise to the input running scenario and fault phenomenon and determining whether the output fault cause is correct, the robustness of the large model can be determined.

[0037] In another embodiment provided in this disclosure, step S103 above, "vectorizing the acquired energy storage power station technical documents to obtain a vector database that corresponds the text information and numerical vectors in the energy storage power station technical documents," can be implemented as follows: Step 1: Organize the text information in the technical documents of the energy storage power station into text that represents the operating scenario, fault phenomenon and fault cause; Step 2: Divide the text into multiple text blocks according to semantics; Step 3: Convert the text block into a numerical vector using a preset encoding model, and construct a vector database based on the correspondence between the text block and the numerical vector.

[0038] In this embodiment of the disclosure, the operating scenario, fault phenomenon, and fault cause can be organized into a text segment according to natural language logic. Furthermore, the content of this text segment can be semantically divided into different text blocks.

[0039] These text blocks can be converted into numerical vectors using a preset encoding model. This preset encoding model can be an embedding model, which can convert each text block into a numerical vector. Text blocks with similar semantics have similar numerical vectors, and the semantic relationship between different text blocks can be captured by the distance between the vectors.

[0040] After converting all text blocks into numerical vectors, a vector database can be constructed by matching the text blocks with the numerical vectors according to their correspondence.

[0041] It should be noted that due to the large variety of models and types of energy storage power stations, using a single vector database to encompass all models and types could result in an excessively large database and excessively long retrieval times. Therefore, a separate vector database can be built for each model or type of energy storage power station. Based on the description of the energy storage power station's situation in the operational scenario, the corresponding vector database can be used for retrieval, thereby improving retrieval speed.

[0042] This disclosure also provides a power plant customer service system construction apparatus, such as... Figure 2 As shown, it includes: Module 201 is used to establish a fault dataset based on the acquired technical documents of the energy storage power station; wherein, the technical documents include the operation scenarios of the energy storage power station, the correspondence between fault phenomena and fault causes; Training module 202 is used to fine-tune the preset large model based on the fault dataset, so that the preset large model learns the correspondence between the operation scenarios, fault phenomena and fault causes of the energy storage power station through the fine-tuning training. The processing module 203 is used to perform vectorization processing on the acquired energy storage power station technical documents to obtain a vector database that corresponds the text information and numerical vectors in the energy storage power station technical documents. The setting module 204 is used to set the vector database and the preset large model with fine-tuning training as the inference layer in the workflow of the energy storage power station customer service system, so as to infer the cause of the fault from the input inquiry statement.

[0043] In another embodiment provided in this disclosure, the establishment module 201 is further configured to extract key information of fault cases in the technical documents of the energy storage power station; determine the operating scenario corresponding to each fault case, as well as the correspondence between fault phenomena and fault causes; use a semantic analysis model to determine duplicate fault cases in the same operating scenario based on the correspondence between fault phenomena and fault causes; and reduce the number of fault cases with a number greater than a first threshold and increase the number of cases with a number less than a second threshold based on the number of duplicate fault cases. After adjusting the number of cases, a fault dataset is established based on the correspondence between the fault phenomena, operating scenarios, and fault causes.

[0044] In another embodiment provided in this disclosure, the training module 202 is further configured to divide the fault dataset into a training set and a test set according to a preset ratio; freeze the pre-training parameters of the preset large model except for the adapter layer and the output layer; input the running scenarios and fault phenomena in the training set into the preset large model; and adjust the parameters of the adapter layer according to the deviation between the fault causes output by the output layer and the corresponding fault causes in the training set, until the fault diagnosis accuracy of the preset large model reaches a first threshold.

[0045] In another embodiment provided in this disclosure, the processing module 203 is further configured to sort the text information in the technical documents of the energy storage power station into text representing the operating scenario, fault phenomenon and fault cause; split the text into multiple text blocks according to semantics; convert the text blocks into numerical vectors through a preset encoding model; and construct a vector database according to the correspondence between the text blocks and the numerical vectors.

[0046] This disclosure also provides a power plant customer service system, such as... Figure 3 As shown, it includes: access layer 301 and inference layer 302; The access layer 301 includes a data interface for communicating with the user side, used to obtain the inquiry statement sent by the user and transmit the inquiry statement to the inference layer 302; The inference layer 302 is constructed based on the power station customer service system construction method provided in any of the above embodiments. It is used to vectorize the consultation statement through a vector database, match the corresponding text block, infer the corresponding fault cause from the fine-tuned preset large model, and output the fault cause to the customer side through the interface layer.

[0047] In this embodiment of the disclosure, the workflow of the customer service system can adopt a four-layer architecture consisting of an access layer 301, a preprocessing layer 303, an inference layer 302, and a collaboration layer 304. Each layer is a working node in the workflow, and the layers are independent of each other.

[0048] Access layer 301 can serve as the interface for the power station customer service system to interact with the user side, access various user terminals, obtain the inquiry statements sent by users through the terminals, and input the inquiry statements into the customer service system.

[0049] In the workflow of the customer service system, the access layer 301 can serve as the starting point of the workflow. After the inquiry statement is entered into the customer service system through the access layer 301, it can be transmitted to other layers in the customer service system according to the specific situation and the workflow.

[0050] Specifically, the access layer 301 can be implemented as a programming interface using a RESTful API architecture. This interface allows data to be transferred between different software systems (e.g., WeChat customer service, power plant maintenance software, and local device software). This interface integrates various channels, enabling staff to communicate with the customer service system promptly in different work scenarios, ensuring timely maintenance support.

[0051] The inference layer 302 may include a vector database constructed using the aforementioned power station customer service system construction method, and a finely tuned preset large model. After the inquiry statement is processed by the inference layer 302, it is first converted into a numerical vector from the vector database, and then matched with the corresponding text block in the vector database. This matching process can employ a similarity threshold matching method based on cosine similarity, where cosine similarity is calculated using the cosine similarity formula based on the dot product and norm of each vector. When the cosine similarity between the numerical vector of the inquiry statement and the vector of the text block reaches a certain range, the numerical vector can be considered a match between the two. Specifically, a calculated cosine similarity of 1 indicates that the two vectors completely overlap, representing the highest similarity; a calculated cosine similarity of 0 indicates that the two vectors are perpendicular, representing no similarity; and a calculated cosine similarity of -1 indicates that the two vectors are completely opposite, representing the lowest similarity.

[0052] The matched text blocks and the user's inquiry statements are concatenated into an inquiry statement and a text block as reference information. This concatenation is then input into a pre-defined large model, which performs inference. The inference result obtained by the large model can be output to the user through the access layer 301. The user can ask follow-up questions based on the inference result. During a consultation, the customer service system can retain the context of the consultation as a basis for inference until the user ends the consultation.

[0053] In yet another embodiment provided in this disclosure, such as Figure 3 As shown, it also includes: a preprocessing layer 303 and a collaboration layer 304; The preprocessing layer 303 includes a preset denoising model, which is used to process the inquiry statements input to the energy storage power station customer service system and output the denoised inquiry statements to the inference layer 302; wherein, the preset denoising model is constructed based on the energy storage power station domain lexicon and is used to eliminate ambiguity in the input inquiry statements; The collaboration layer 304 is used to establish a communication connection between the user side and the human customer service side through the access layer 301 when the reasoning layer 302 matches a text block in the consultation statement containing collaboration keywords; and / or If the relevance of the text block matched in the consultation statement by the inference layer 302 is lower than a second threshold, the access layer 301 enables the user side to establish a communication connection with the human customer service side; wherein, the relevance of the text block is determined based on the numerical vector of the consultation statement and the numerical vector of the text block in the vector database.

[0054] In this embodiment, the preset denoising model in the preprocessing layer 303 can be a model based on a word segmentation lexicon, which can store a lexicon in the field of energy storage power stations. Consultation statements received by the access layer 301 can be input into the preprocessing layer 303. The preset denoising model can replace abbreviations (e.g., replacing PCS with energy storage converter, BMS with battery management system), meaningless words (e.g., interjections and colloquialisms), and ambiguous words (e.g., replacing power station power outage with power station power interruption) in the input consultation statement according to the contextual semantics, thereby improving the accuracy of the consultation statement in pointing to the problem. After denoising preprocessing, the consultation statement can be further input into the inference layer 302 for fault inference according to the workflow.

[0055] The collaboration layer 304 is used to connect users with human customer service in specific situations. One specific situation could be when a collaboration keyword is matched in the query statement, for example, when the user explicitly requests human customer service intervention in their query. Another specific situation could be when the inference layer 302 cannot match a text block corresponding to the query statement in the vector database. Here, relevance can be the cosine similarity between numerical vectors. If, after matching all numerical vectors in the vector database, no numerical vector with a cosine similarity greater than a second threshold is found, it indicates that there is no relevant information in the vector database, and the inference result is likely inaccurate. In this case, a notification suggesting human customer service intervention can be sent to the user, and after the user confirms, a human customer service representative can be connected to the user. The second threshold can be a pre-set value that judges whether two numerical vectors are related.

[0056] After human customer service intervenes, the 304 collaboration layer can first send the previous consultation process to human customer service. Specific formats may include the issue ID, user identity, issue details, and historical interaction records, for example: [Issue ID: 202405200XX] User: Power Plant Operations and Maintenance; Problem: Inverter error code I003, grid connection failed; Historical interaction record: The mains voltage has been checked and found to be normal at 380V.

[0057] Once a human customer service representative confirms that an issue is being processed, the customer service system can automatically mark it as "in progress," preventing duplicate responses from multiple representatives. If no human customer service representative contacts the issue within a certain period, the system can send a timeout reminder to the human representative to ensure that inquiries are resolved promptly.

[0058] Through the above description of the embodiments, those skilled in the art can clearly understand that the embodiments of this disclosure can be implemented in hardware or by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this disclosure.

[0059] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes in the drawings are not necessarily essential for implementing this disclosure.

[0060] Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0061] The sequence numbers of the embodiments disclosed above are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0062] Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, this disclosure is also intended to include such modifications and variations.

Claims

1. A method for constructing a power station customer service system, characterized in that, include: Based on the obtained technical documents of the energy storage power station, a fault dataset is established; wherein, the technical documents include the operating scenarios of the energy storage power station, the correspondence between fault phenomena and fault causes; Based on the fault dataset, the preset large model is fine-tuned and trained so that the preset large model learns the correspondence between the operating scenarios, fault phenomena and fault causes of the energy storage power station through the fine-tuning training. The acquired energy storage power station technical documents are vectorized to obtain a vector database that corresponds the text information in the energy storage power station technical documents with numerical vectors. The vector database and the pre-set large model with fine-tuning training are set as the inference layer in the workflow of the energy storage power station customer service system so as to infer the cause of the fault from the input inquiry statement.

2. The method as described in claim 1, characterized in that, The step of establishing a fault dataset based on the acquired energy storage power station technical documents includes: Extract key information from the fault cases in the energy storage power station technical documents; wherein, the key information includes: fault phenomenon, operating scenario and fault cause; Determine the operating scenario corresponding to each fault case, as well as the correspondence between the fault phenomenon and the fault cause; Using a semantic analysis model, based on the correspondence between the fault phenomena and the fault causes, duplicate fault cases in the same operating scenario are identified; Based on the number of repeated failure cases, reduce the number of failure cases that exceed the first threshold and increase the number of cases that are less than the second threshold. After adjusting the number of cases, a fault dataset is established based on the correspondence between the fault phenomena, operating scenarios, and fault causes.

3. The method as described in claim 1, characterized in that, The step of fine-tuning and training a pre-set large model based on the fault dataset, enabling the pre-set large model to learn the correspondence between the operating scenarios, fault phenomena, and fault causes of the energy storage power station through the fine-tuning training, includes: The fault dataset is divided into a training set and a test set according to a preset ratio; Freeze the pre-trained parameters of the preset large model, excluding the adapter layer and the output layer; The operating scenarios and fault phenomena in the training set are input into the preset large model. Based on the deviation between the fault cause output by the output layer and the corresponding fault cause in the training set, the parameters of the adapter layer are adjusted until the fault diagnosis accuracy of the preset large model reaches a first threshold. The fault diagnosis accuracy is determined based on the deviation between the first fault cause output by the preset large model and the corresponding fault cause in the test set. The first fault cause is the fault cause output by the preset large model based on the operating scenarios and fault phenomena in the test set.

4. The method as described in claim 1, characterized in that, The process of vectorizing the acquired energy storage power station technical documents to obtain a vector database that corresponds the text information in the energy storage power station technical documents with numerical vectors includes: The text information in the technical documents of the energy storage power station is sorted into text that characterizes the operating scenario, fault phenomenon and fault cause; The text is divided into multiple text blocks according to semantics; The text blocks are converted into numerical vectors using a preset encoding model, and a vector database is constructed based on the correspondence between the text blocks and the numerical vectors.

5. A device for constructing a power station customer service system, characterized in that, include: A module is established to create a fault dataset based on the acquired technical documents of the energy storage power station; wherein, the technical documents include the operating scenarios of the energy storage power station, the correspondence between fault phenomena and fault causes; The training module is used to fine-tune the preset large model based on the fault dataset, so that the preset large model can learn the correspondence between the operation scenarios, fault phenomena and fault causes of the energy storage power station through the fine-tuning training. The processing module is used to vectorize the acquired energy storage power station technical documents to obtain a vector database that corresponds the text information in the energy storage power station technical documents with numerical vectors. The configuration module is used to set the vector database and the preset large model with fine-tuning training as the inference layer in the workflow of the energy storage power station customer service system, so as to infer the cause of the fault from the input inquiry statement.

6. The apparatus as claimed in claim 5, characterized in that, The establishment module is also used to extract key information of fault cases from the technical documents of the energy storage power station; and to determine the operating scenario corresponding to each fault case, as well as the correspondence between fault phenomena and fault causes. Using a semantic analysis model, based on the correspondence between the fault phenomena and the fault causes, duplicate fault cases in the same operating scenario are identified; based on the number of duplicate fault cases, the number of fault cases exceeding a first threshold is reduced, and the number of cases less than a second threshold is increased. After adjusting the number of cases, a fault dataset is established based on the correspondence between the fault phenomena, operating scenarios, and fault causes.

7. The apparatus as claimed in claim 5, characterized in that, The training module is further configured to divide the fault dataset into a training set and a test set according to a preset ratio; freeze the pre-training parameters of the preset large model except for the adapter layer and the output layer; input the running scenarios and fault phenomena in the training set into the preset large model; and adjust the parameters of the adapter layer according to the deviation between the fault causes output by the output layer and the corresponding fault causes in the training set, until the fault diagnosis accuracy of the preset large model reaches a first threshold.

8. The apparatus as claimed in claim 5, characterized in that, The processing module is further configured to sort the text information in the technical documents of the energy storage power station into text representing the operating scenario, fault phenomenon and fault cause; split the text into multiple text blocks according to semantics; convert the text blocks into numerical vectors through a preset encoding model; and construct a vector database according to the correspondence between the text blocks and the numerical vectors.

9. A power station customer service system, characterized in that, include: Access layer and inference layer; The access layer includes a data interface for communicating with the user side, used to obtain the consultation statement sent by the user and transmit the consultation statement to the inference layer; The inference layer is constructed based on the method described in any one of claims 1-4, and is used to vectorize the consultation statement through a vector database, match the corresponding text blocks, infer the corresponding fault cause from a finely tuned preset large model, and output the fault cause to the client side through the interface layer.

10. The system as described in claim 9, characterized in that, Also includes: Preprocessing layer and collaboration layer; The preprocessing layer includes a preset denoising model, which is used to process the inquiry statements input to the energy storage power station customer service system and output the denoised inquiry statements to the inference layer; wherein, the preset denoising model is constructed based on the energy storage power station domain lexicon and is used to eliminate ambiguity in the input inquiry statements; The collaboration layer is used to establish a communication connection between the user side and the human customer service side through the access layer when the inference layer matches a text block in the consultation statement containing collaboration keywords; and / or If the relevance of the text block matched in the consultation statement by the inference layer is lower than a second threshold, a communication connection is established between the user side and the human customer service side through the access layer; wherein, the relevance of the text block is determined based on the numerical vector of the consultation statement and the numerical vector of the text block in the vector database.

Images