Train operation risk identification method, computing device and storage medium

Abstract

Provided are a train operation risk identification method, a computing device and a storage medium, involving the following steps: acquiring train real-time monitoring and control data; collecting train dispatching operation log data; generating a prompt instruction based on the train real-time monitoring and control data and the train dispatching operation log data; sending the prompt instruction to a train operation risk detection large language model; and acquiring an output result of the train operation risk detection large language model, wherein the output result includes a risk identification result for train dispatching behavior. It can be seen that the present application proposes a train dispatching operation safety risk identification technology with high intelligence and strong reliability, which can detect risks hidden in dispatching behavior in real time and has important value for preventing train operation safety problems.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of train intelligent scheduling and control technology, and particularly relates to a train operation risk identification method, a computing device and a storage medium. BACKGROUND

[0002] The compliance, timeliness and accuracy of train scheduling directly determine the safety of train operation. Incorrect scheduling operations are likely to cause major train accidents. At present, stations equipped with centralized control systems (CTC) have realized the automation of scheduling operations, which has greatly reduced the possibility of human error. However, in stations without CTC systems or in the very station control mode of CTC systems, train scheduling operations still need to be operated manually by dispatchers. SUMMARY

[0003] In view of the above problems, the present application is proposed in order to provide a train operation risk identification method, a computing device and a storage medium which overcome the above problems or at least partially solve the above problems.

[0004] According to one aspect of the present application, a train operation risk identification method is provided, comprising:

[0005] acquiring train real-time monitoring and control data;

[0006] collecting train scheduling operation log data;

[0007] generating a prompt instruction based on the train real-time monitoring and control data and the train scheduling operation log data;

[0008] sending the prompt instruction to a train operation risk detection large language model, and acquiring an output result of the train operation risk detection large language model, the output result including a risk identification result of train scheduling behavior.

[0009] According to another aspect of the present application, a computing device is provided, comprising a processor, a memory, a communication interface and a communication bus, the processor, the memory and the communication interface completing communication with each other through the communication bus;

[0010] The memory is used to store at least one executable instruction, and the executable instruction causes the processor to perform the following operations:

[0011] acquiring train real-time monitoring and control data;

[0012] collecting train scheduling operation log data;

[0013] generating a prompt instruction based on the train real-time monitoring and control data and the train scheduling operation log data;

[0014] The prompt instruction is sent to a train operation risk detection large language model, and an output result of the train operation risk detection large language model is acquired, and the output result includes a risk identification result of a train scheduling behavior.

[0015] According to another aspect of the embodiment of the present application, a computer storage medium is provided, and the computer storage medium stores at least one executable instruction, and the executable instruction causes a processor to perform operations corresponding to the train operation risk identification method.

[0016] According to the technical solution provided by the present application, the train real-time monitoring and control data is acquired, the train scheduling operation log data is collected, the prompt instruction is generated based on the train real-time monitoring and control data and the train scheduling operation log data, finally, the prompt instruction is sent to the train operation risk detection large language model, and the output result of the train operation risk detection large language model is acquired, and the output result includes the risk identification result of the train scheduling behavior. It can be seen that the present application provides a train scheduling operation safety risk identification technology with high intelligence and strong reliability, which can detect the hidden risks of scheduling behavior in real time and has important value for preventing train operation safety problems.

[0017] The above description is only a summary of the technical solution of the present application. In order to more clearly understand the technical means of the present application, the specific embodiments of the present application can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0018] Various other advantages and benefits will become apparent to those of ordinary skill in the art, upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of preferred embodiments, and are not meant to limit the present application. Furthermore, the same reference numerals are used throughout the several drawings to designate the same or similar parts. In the drawings:

[0019] Figure 1 A flow chart of a train operation risk identification method according to an embodiment of the present application is shown

[0020] Figure 2 A flow chart of a train operation risk identification method according to an embodiment of the present application is shown

[0021] Figure 3 A flow chart of a train operation risk identification method according to an embodiment of the present application is shown

[0022] Figure 4 A flow chart of a train operation risk identification method according to an embodiment of the present application is shown

[0023] Figure 5 A structural schematic diagram of a computing device according to Embodiment Six of the present application is shown. DETAILED DESCRIPTION

[0024] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0025] It should be noted that the running safety of the train is a key link of railway safety prevention and control. The train running safety needs to consider multiple information dimensions, wide involvement, large time span, and is an extremely complex thinking decision-making process, which requires very high ability and quality of the management personnel. Although the rule-based decision support system can quickly identify and process most of the structured data represented risks, effectively ensuring the train running safety. However, the traditional risk prevention and control system still faces many challenges in the aspects of global risk perception, real-time analysis of voice, text and other unstructured data, real-time evaluation of the rationality of the operation of the operating personnel, automatic generation of risk disposal schemes, etc.

[0026] Embodiment One

[0027] Figure 1 A flowchart of a train running risk identification method according to Embodiment One of the present application is shown, as shown in Figure 1 The method includes the following steps:

[0028] Step S101: Obtain real-time monitoring and control data of the train.

[0029] In this embodiment, the train real-time monitoring and control data belongs to a combination of multiple data, including train control system data, departure and arrival plan data, risk monitoring data, etc. Among them, the train control system data refers to the data generated by the train control system, which is mainly responsible for monitoring the running state of the train and ensuring the safe operation of the train. Train control system data usually includes train position, speed, acceleration, communication information between train and track, signal system state, etc. These data can be obtained through on-board equipment (such as transponders, sensors, GPS, etc.) on the train and ground equipment (such as track circuits, wireless communication networks); departure and arrival plan data includes train departure and arrival schedule, dispatching instructions, train marshalling information, shunting operation plan, etc., which are usually formulated by dispatchers in the dispatching center according to real-time conditions and preset dispatching rules, and issued to relevant stations and trains through CTC (Centralized Traffic Control), i.e. centralized train dispatching system; and risk monitoring data involves monitoring various risk factors that may occur during train operation, including monitoring of train technical state, line condition, natural disasters, human errors, etc.; and risk monitoring data can be obtained in various ways, such as through on-board monitoring systems, trackside monitoring equipment, weather information acquisition systems, etc. The above data can be obtained from multiple information sources through pre-set software and hardware interfaces, and the specific acquisition method is known in the art, which will not be described here.

[0030] Obtaining real-time train control system data, departure and arrival plan data and risk monitoring data is crucial for the artificial intelligence large model involved in the present application to identify train dispatching risks. These data enable the model to update inputs in real time, dynamically adjust predictions, improve accuracy, prevent potential risks, and optimize dispatching decisions.

[0031] Step S102: Collect train dispatching operation log data.

[0032] In this embodiment, an agent can be set up to obtain train dispatching operation log data. The agent can be implemented as a software program or system that can perform specific tasks such as data collection, monitoring, analysis, etc., and the specific implementation method is not limited in the present application. Train dispatching log is a detailed document or data record that records various operations and events during train dispatching. For example, it generally includes train operation information, dispatching decision information, device status information, fault and abnormal information, maintenance activity information, timestamp information, etc.

[0033] According to the historical data provided by the train dispatching operation log data, the model can deeply understand the historical scheduling mode that has occurred, and enhance its risk prediction ability. In particular, according to steps S101 and S102, train real-time monitoring and control data and train dispatching operation log data can be obtained at the same time, so that the artificial intelligence large model of the present application brings the advantages of comprehensiveness, real-time and history combination in train scheduling risk identification, which enhances the analysis ability of the model to the scheduling mode and trend, optimizes the decision support, speeds up the abnormal detection speed, integrates the time series and multi-dimensional data, and can further enhance the advanced reasoning and early warning ability of the model, and provides a more reliable error prevention mechanism for train scheduling.

[0034] Step S103: generating a prompt instruction based on the train real-time monitoring and control data and the train dispatching operation log data.

[0035] It should be noted that the sending frequency of the train dispatching operation log data generated prompt instruction can be diversified, and can be adjusted according to system resource configuration or user demand. Specifically, the setting of the frequency needs to consider the ability of the system to process large data and the processing speed of the large language model, and too high frequency may cause the system load to be too large, while too low frequency may cause the risk identification to lag. In some embodiments, the system can dynamically adjust the sending frequency according to the risk assessment result and the current running state, for example, increase the sending frequency during high risk period or abnormality occurs, and reduce the sending frequency during low risk period.

[0036] In this embodiment, the train real-time monitoring and control data and the train dispatching operation log data obtained and used to generate the prompt instruction can be pure text, or can include audio, voice, image and other information expression forms. The data type that can be accepted and parsed by the artificial intelligence large model is used as the standard, and further, the train real-time monitoring and control data and the train dispatching operation log data are assembled into the prompt instruction for the artificial intelligence large model, and other information is combined to generate the prompt instruction. Exemplarily, a complete prompt instruction usually includes several key parts, such as background information, task description, input data, expected output and limitation or rule description, to ensure the clarity of the information and the integrity of the instruction. The following is an example of a complete prompt instruction according to an embodiment of the present application:

[0037] Background information: In order to comprehensively analyze the risk of scheduling operation, you need to perform risk assessment based on the data of train operation scheduling.

[0038] Task description: Please carefully analyze the following submitted scheduling operation log information to determine whether there is potential risk in the scheduling operation.

[0039] Input data:

[0040] Basic information of the train (e.g., train number, type, origin and destination station, planned departure and arrival time, etc.);

[0041] Train dispatching plan information (including route, stop stations, scheduled timetable, etc.);

[0042] Dispatcher operation log information (records of specific operations, such as dispatching commands, time markers, abnormal event handling, etc.);

[0043] Information perceived by the risk perception system (such as weather changes, equipment status, operational abnormalities, etc.).

[0044] Expected output: Your output should be in two parts:

[0045] First, determine whether there is a risk, starting with "At risk" or "No risk";

[0046] Combine the input information to give a cause analysis of the risk (or lack of risk).

[0047] Limitations or rules: Please ensure comprehensive analysis, covering all input information, and ensure accurate reasoning process.

[0048] Example output:

[0049] At risk.

[0050] Cause analysis: According to the dispatcher operation log information, it is found that train J123 at XX station has a stay time of more than 7 minutes beyond the planned timetable, while the risk perception system reports that the station section has been affected by adverse weather in recent days, causing unstable equipment operation. Combined with the train dispatching plan information, it can be known that the train is already close to the subsequent class, and there is a possibility of further delay, so it is determined that there is a risk of operation.

[0051] Please generate the analysis results according to the above prompts.

[0052] As can be seen from the example, in this prompt instruction, the train real-time monitoring and control data and the train dispatching operation log data obtained or collected are integrated into the input data for the artificial intelligence large model to analyze and judge the risk.

[0053] Step S104: Send the prompt instruction to the train operation risk detection large language model, obtain the output result of the train operation risk detection large language model, and the output result includes the risk identification result of the train dispatching behavior.

[0054] In this embodiment, the prompt instruction generated in step S103 is sent to the artificial intelligence large model, which will automatically analyze and process the content in the prompt instruction, and finally give the output result of whether there is hidden danger or risk in the current train scheduling behavior. For example, the artificial intelligence large model used in this embodiment can be Llama, ChatGLM, ERNIE, GPT series large model, etc. These large models have been widely used in the field and are also suitable for implementing the technical solution of this embodiment. The working mechanism and architecture of each large model are not limited in this embodiment. In addition, as the output result form of the desired large model included in the prompt instruction in step S103, the output result of the large language model in this embodiment can be the following examples:

[0055] Risk conclusion: there is risk.

[0056] Cause analysis: according to the XXX operation log information, it is found that the train XX times in XX station stays for more than XX minutes of the planned timetable, and the risk perception system reports that the station section is affected by bad weather in recent days, causing the equipment to run unstably. Combined with the train scheduling plan information, it can be known that the train has been close to the subsequent shift, and there is a possibility of further delay, so it is determined that there is XX risk.

[0057] By using the train operation risk identification method of this embodiment, by integrating real-time monitoring data and historical operation log data in the prompt instruction sent to the large language model, not only comprehensive and multi-dimensional train scheduling information is provided, but also time series and multi-dimensional data are integrated, which significantly improves the accuracy of risk prediction and early warning ability, enabling the model to deeply understand the scheduling mode and optimize decision support and abnormal detection speed. The train operation risk identification realizes automatic processing, reduces manual intervention, and improves efficiency; in addition, the clear output format is convenient for quick understanding and decision-making.

[0058] Embodiment two

[0059] Figure 2 A flowchart of a train operation risk identification method according to embodiment two of the present application is shown. Embodiment two is a further refinement of step S102 in embodiment one.

[0060] As Figure 2 shown, the method comprises the following steps:

[0061] Step S201: setting a proxy module, the proxy module is configured to collect train scheduling operation log data; the train scheduling operation log data includes train scheduling log text data and train scheduling instruction audio data.

[0062] As mentioned in step S102, the train dispatching operation log data can be obtained by setting an agent. Voice data is an important part of the dispatching log data. In particular, in some emergency and very normal situations, dispatching instructions are transmitted in the form of voice, and some matters that need to be confirmed by the dispatcher and the train driver are also transmitted in the form of voice. Therefore, the processing of voice data is very important for the identification of dispatching behavior risks and must not be ignored.

[0063] To this end, further optionally, step S202 is proposed: converting the train dispatching instruction audio data into second train dispatching instruction text data.

[0064] In order to better integrate the text data and audio data in the train dispatching log data, so that the artificial intelligence large model can better understand and respond, it is necessary to convert the train dispatching instruction audio data into train dispatching instruction text data. This step can be implemented using a commonly used voice recognition model with TTS (text to speech). Here, it is not specifically limited, and it can be understood that the voice recognition model can also be a pre-trained voice recognition model that meets the train dispatching scene, so that the recognition result is more accurate.

[0065] Further, step S203 is proposed: converting the train dispatching instruction audio data into second train dispatching instruction text data further comprises: after converting the train dispatching instruction audio data into first train dispatching instruction text data, formatting the first train dispatching instruction text data to obtain the second train dispatching instruction text data after formatting.

[0066] Train dispatching instruction audio data often has a large amount of data and also contains a lot of useless information. If the text data converted from these audio data is directly used, not only will it waste system resources, but it may also seriously affect the accuracy of the artificial intelligence large model in risk identification. In view of this, the first train instruction text data obtained after conversion also needs to be formatted to obtain the second train dispatching instruction text data after formatting. That is, the first train instruction text data converted from the audio needs to be summarized and refined, and useful information needs to be extracted and integrated, and useless information needs to be removed.

[0067] In order to implement the above steps, step S204 should also be provided: converting the train dispatching instruction audio data into second train dispatching instruction text data further comprises:

[0068] The train dispatching instruction audio data is converted into first train dispatching instruction text data using a speech recognition model, and semantic extraction is performed on the first train dispatching instruction text data using a semantic extraction large language model, to obtain the second train dispatching instruction text data after formatting. In some embodiments, the formatting operation includes syntax checking, text data format unification, text data structuring, data integrity checking, and the like, ensuring the standardization and consistency of the train dispatching instruction text data and improving the efficiency and accuracy of data processing.

[0069] Specifically, in the audio data processing flow of the train dispatching instruction, a speech recognition model and a semantic extraction large language model can be used to process the data, respectively. For example, the speech recognition model can be implemented using common TTS models or tools, such as Google Cloud Speech-to-Text, Amazon Transcribe, Baidu DeepSpeech, etc., and the semantic extraction large language model can be BERT, ERNIE, GPT, etc., which are not limited here.

[0070] In embodiment two, by converting the voice data in the dispatch log into text and using a large language model (LLM) for semantic extraction, first, it is ensured that the key dispatching instructions transmitted by voice and the matters confirmed between the train driver in emergency and abnormal situations can be accurately captured, second, the speech recognition model converts these voice data into analyzable text form, laying a foundation for further semantic understanding and risk assessment, then, the LLM analyzes the text in depth, extracts key information, and realizes the formatted output of data, which not only improves the readability and operability of the dispatching instructions, but also enhances the accuracy and response speed of risk identification, improves the utilization efficiency of the dispatch log data and the quality of decision support.

[0071] Embodiment three

[0072] Figure 3 A flowchart of a train operation risk identification method according to embodiment three of the present application is shown. Embodiment three is a further refinement of step S103 in embodiment one.

[0073] As shown in Figure 3 The method includes the following steps:

[0074] Step S301: generating a prompt instruction based on the train real-time monitoring and control data and the train dispatching operation log data includes:

[0075] The train real-time monitoring and control data and the train dispatching operation log data are configured as at least part of the input data in the prompt instruction, and are used as at least part of the data analysis basis for indicating the execution result output of the train operation risk detection large language model.

[0076] The train real-time monitoring and control data and the train dispatching operation log data are assembled into the prompt instruction as part of the input data in the prompt instruction, and need to be explicitly informed in the prompt instruction that the artificial intelligence large model should be based on these input data for reasoning and analysis in the train operation risk identification and judgment, so as to be used as at least part of the data analysis basis for indicating the execution result output of the train operation risk detection large language model.

[0077] Optionally, the step S302 of generating a prompt instruction based on the train real-time monitoring and control data and the train dispatching operation log data further comprises:

[0078] The train real-time monitoring and control data and the train dispatching operation log data are configured to retrieve and obtain text data matched with the train real-time monitoring and control data and the train dispatching operation log data from a train dispatching related knowledge database, and the text data matched with the train real-time monitoring and control data and the train dispatching operation log data are configured as at least part of the input data in the prompt instruction, and are used as at least part of the data analysis basis for indicating the execution result output of the train operation risk detection large language model.

[0079] In this step, the obtained or collected train real-time monitoring and control data and the train dispatching operation log data are used to retrieve a pre-set train dispatching related knowledge database to obtain text data matched with the above data, and these text data are also assembled into the prompt instruction as another part of the input data in the prompt instruction, and need to be explicitly informed in the prompt instruction that the artificial intelligence large model should be based on these input data for reasoning and analysis in the train operation risk identification and judgment, so as to be used as at least part of the data analysis basis for indicating the execution result output of the train operation risk detection large language model. The train dispatching related knowledge database may be a database specially constructed to provide general train dispatching related knowledge, or an Internet database, etc., which is not limited in this step.

[0080] By searching the train scheduling related knowledge database, the artificial intelligence model can obtain highly relevant supplementary information about the current scheduling situation, enhancing its understanding of complex scheduling scenarios and its generalization ability in information processing. In addition, clear prompt instructions can guide the model on how to effectively reason using this data, ensuring the relevance and accuracy of the analysis. This comprehensive data-driven approach not only improves the accuracy of abnormal behavior identification but also provides strong support for scheduling decisions, thereby optimizing the safety and efficiency of train operation.

[0081] Optionally, the step S303 is further included, wherein the train real-time monitoring and control data and the train scheduling operation log data are configured to search and obtain text data matching the train real-time monitoring and control data and the train scheduling operation log data from a train scheduling related knowledge database, including:

[0082] The train scheduling related knowledge database includes an unstructured data vector database and / or a structured data database.

[0083] In this embodiment, the knowledge data stored in the database includes both structured and unstructured data. For example, structured data includes basic information of trains and stations, equipment information, risk identification information, and scheduling plan information, while unstructured information includes scheduling operation regulations, scheduling responsibilities, and risk disposal manuals. Optionally, the unstructured data vector database and the structured data database can be constructed in the following ways, including:

[0084] Data preparation, distinguishing between structured data (including train and station basic information, equipment information, risk identification information, etc.) and unstructured data (such as scheduling operation regulations, risk disposal manuals, etc.) required for train scheduling;

[0085] Vector database construction, natural language processing of unstructured data, conversion to vector representation, and storage in the vector database to support efficient fuzzy retrieval and similarity analysis;

[0086] Structured database construction, importing structured data into the structured database to facilitate fast and accurate data retrieval.

[0087] For this embodiment, integrating the structured database and the vector database can significantly optimize the efficiency and quality of train scheduling risk identification.

[0088] Optionally, further comprising step S304: before the train real-time monitoring and control data and the train dispatching operation log data are configured to retrieve and obtain text data matching the train real-time monitoring and control data and the train dispatching operation log data from a train dispatching related knowledge database, further comprising:

[0089] Performing text vectorization on the train real-time monitoring and control data and the train dispatching operation log data.

[0090] In order to achieve good retrieval effect and information matching, the method involved in this step can include:

[0091] Building a knowledge index: vector encoding knowledge from databases, documents, web pages, and knowledge graphs, and storing them in a vector database. This step is to convert knowledge into a machine processable format and store it for subsequent retrieval.

[0092] Retrieval fusion: text vectorization of train real-time monitoring and control data and train dispatching operation log data, further based on semantic vector similarity, retrieve the most relevant document blocks in the vector database, such as the top n document blocks. Fuse the original question and the retrieved document blocks to generate the final retrieval result and output.

[0093] In the embodiments of the present application, the RAG (Retrieval-Augmented Generation) technology is used to enhance the accuracy and reliability of the large model answer. RAG, i.e. Retrieval-Augmented Generation model, is a method of jointly using retrieval mechanism and generation model in the field of natural language processing (NLP), which aims to retrieve relevant information to assist in generating more accurate and rich text content. Its core steps include knowledge index, retrieval fusion, and content generation. Therefore, the RAG module is used to generate rules and related basic knowledge according to the characteristics of the scene.

[0094] For example, the knowledge required for train operation safety includes both structured and unstructured data. The output of the retrieval process in this step can be in the following form: structured data includes basic information of trains and stations, equipment information, risk identification information, dispatching plan information, etc. Unstructured information includes dispatching operation regulations, dispatching responsibilities, risk disposal manuals, etc.

[0095] Therefore, RAG combines the capabilities of retrieval and generation, which can effectively extract key information from a large amount of data and generate accurate risk assessment, greatly improving the accuracy of risk identification and the real-time updating ability of the model. Not only improves the safety guarantee level, but also enhances the transparency and credibility of decision-making, which can effectively solve the illusion problem of large models.

[0096] In this embodiment, on the basis of integrating train real-time monitoring and control data and scheduling operation logs, supplementary information highly related to scheduling conditions can be provided by using knowledge database retrieval, which enhances the model's understanding of complex scenarios. In addition, clear prompt instructions ensure the accuracy of analysis, improve the identification precision of abnormal behavior, and provide strong support for scheduling decisions, optimizing the safety and efficiency of train operation. The method of integrating structured and unstructured data improves the generalization ability of the model, while the automated retrieval and data processing process improves efficiency and reduces human error. Furthermore, through text vectorization and knowledge index construction, the efficiency and quality of retrieval are further improved, ensuring that the model can quickly adapt to new scheduling conditions and knowledge updates.

[0097] Embodiment Four

[0098] Figure 4 A flowchart of a train operation risk identification method according to Embodiment Four of the present application is shown. Embodiment Four is a further refinement of step S104 in Embodiment One.

[0099] As Figure 4 shown, the method includes the following steps:

[0100] Step S401: sending the prompt instruction to the train operation risk detection large language model, and obtaining the output result of the train operation risk detection large language model, wherein the output result includes the risk identification result of train scheduling behavior, including:

[0101] The train operation risk detection large language model is a fine-tuned large language model.

[0102] In this step, by adjusting the parameters of the pre-trained model on the data set in the train scheduling field, the performance of the model on the specific task of vehicle operation risk identification in the present application can be improved. This step can refine and adapt the model, and by optimizing the parameters on new data related to the target task, errors can be corrected.

[0103] In some examples of the embodiments of the present application, a Train Security Large Model (TSLM) based on LLM for train operation risk prevention and control is provided. Specifically, the safety control of railway operation is divided into different scenarios, and TSLM is applied to analyze each scenario in real time to determine whether the train operation scenario is at risk. The edge perception system transmits the monitored data to the LLM in real time, and the TSLM encapsulates the detection data into a specific scenario set according to the characteristics of the detection data, which includes the rules and basic knowledge required for processing the detection data. Then, the TSLM fuses the COT (Chain-of-Thought) technology with the scenario data to generate instruction data that can be used for LLM reasoning, and finally outputs the risk identification result of the scenario and the disposal suggestion.

[0104] Specifically, an open-source LLM is fine-tuned to support the reasoning of TSLM, which is called TSLM-backbone. On the one hand, TSLM-backbone can solve the problem that general LLM lacks professional knowledge in the vertical field of train operation safety control, and can improve the understanding of domain knowledge by LLM. On the other hand, the COT method is used to fine-tune the LLM, which makes the reasoning ability of LLM more prominent. In addition to effectively enhancing the reasoning effect, TSLM-backbone also provides support for the private deployment of TSLM.

[0105] Here, the train operation risk identification method based on scenarios divides the risks of train operation into different scenarios and controls the risks according to the characteristics of different scenarios. The risk identification framework TSLM realized by LLM reasoning can identify the risks of train operation scenarios.

[0106] Optionally, it further includes step S402: the train operation risk detection large language model is a fine-tuned large language model, including:

[0107] A small amount of sample thought chain prompt method is used to generate a training data set for the fine-tuning.

[0108] Train dispatching is a comprehensive work, and the knowledge required for the work covers various railway business knowledge. In order to make LLM fully understand the knowledge in the field of dispatching, it is necessary to fine-tune with various railway business knowledge. In the present application, a large amount of text content related to railway dispatching is collected for model training, which includes, but is not limited to, the following content:

[0109]

[0110] In some examples of the embodiments of the present application, for the collection of fine-tuning data, a large model such as GPT-4 can be used to generate high-quality and widely-covered training data, ensuring the diversity and annotation accuracy of the data.

[0111] However, the traditional manual construction method is very costly, so the present application proposes a method of using a large model to construct a fine-tuning instruction data set. This method uses a few-shot chain-of-thought prompting method to guide the LLM to generate training data with a chain-of-thought, so that the model can discover problems in the scheduling behavior through reasoning. The following is an example of the method of constructing a fine-tuning data set according to the present application:

[0112] In some embodiments, the accuracy of model reasoning is improved by comprehensively using RAG, fine-tuning, COT, and other methods. The accuracy and reliability of the large model's answers are enhanced by using RAG technology.

[0113] Fine-tuning can make the large model perform better in vertical fields. By adjusting the parameters of the pre-trained model on a specific data set, the model can master professional terminology and knowledge, improve performance accuracy and relevance on tasks, enhance problem-solving ability, and quickly adapt to field requirements, effectively reducing development costs and time. This method can refine and adapt the model through fine-tuning, and optimize parameters on new data related to the target task to correct errors.

[0114] Through COT technology, the reasoning ability and complex problem-solving ability of LLM can be improved. By gradually decomposing problems, COT guides the model to solve tasks in logical steps, thereby improving the accuracy and consistency of answers. In addition, COT enhances the model's performance in multi-step tasks, making it more effective in handling problems that require reasoning and multi-level analysis.

[0115] Therefore, the combination of RAG, Fine-tuning, and COT can significantly improve the application effect of large models. This combination method not only improves the accuracy and reliability of the model in vertical fields, but also enhances its reasoning ability and problem-solving ability, making the model more efficient and flexible in dealing with complex tasks, thereby better meeting various practical needs. Therefore, TSLM uses a scenario-based risk identification method to divide the risks of train operation into different scenarios for fine-grained prevention and control, and combines RAG, Fine-tuning, and COT to enhance the model's reasoning and risk identification capabilities.

[0116] First, construct the scene set. All scheduling jobs occur in a certain scenario and must strictly follow the rules of the on-time rules library. Scheduling jobs that do not meet the rules will pose a safety hazard. Based on this consideration, the application uses scheduling rules as seed data to generate corresponding fine-tuning instruction data.

[0117] Further, determine the structure of the training sample. One of the purposes of the application is to use fine-tuning data to train a scheduling behavior risk identification model to infer and analyze the scheduling log of the dispatcher to identify whether there is a safety risk in the scheduling behavior. Therefore, when designing the fine-tuning data set, the special scenario requirements must be fully considered to ensure that the model can effectively learn and predict potential risks. The following is the preferred format:

[0118] The input part corresponds to the scene set, and the input part of the scene must include a series of train scheduling related feature sequences: 1) rules required for inference, and basic information such as station basic information, train basic information, scheduling plan, etc. This part of information is the basis for LLM inference and judgment. 2) scheduling log. Time-ordered scheduling behavior log. Including scheduling job instruction log obtained from the scheduling system, monitoring data log provided by the natural disaster prevention system, and global data.

[0119] In some examples of the embodiments of the present application, the scene is a description of the state of the train operation environment. The scene is composed of the interaction of multiple dimensions of different entities. Entities can be described by static attribute data and dynamic behavior data. Therefore, the scene can be defined as a set of these elements:

[0120] (1)

[0121] Where: represents the attribute data of the i-th entity, (t) represents the behavior data of the i-th entity at time t, represents the i-th element composed of attribute data and behavior data , and n is the total number of entities in the scene.

[0122] As can be seen from formula (1), the key to constructing S(t) of the scene is to construct . Data can be collected from the logs of existing control systems. Since real-world data containing risk information is very sparse, the generation of behavioral data should consider generating sample data that can trigger risks to train the model's ability to cope with risks. Trains operate according to certain rules; operation within the permitted range is safe, while exceeding the permitted range will generate risks. Therefore, this study uses train operation rules as seed data to generate behavioral data for each S(t).

[0123] Furthermore, regarding the construction of real-time scenes, as shown in formula (1), since the attributes of entities in the scene are static, this part is easy to collect. The difficulty lies in obtaining real-time data of the scene. . A set of behavioral events that may lead to risk can be represented as a combination of multiple elements:

[0124] (2)

[0125] In the formula, A represents the actor, the entity performing the action, including both people and equipment. People include train drivers, dispatchers, station duty officers, and signalmen, while equipment mainly consists of various control systems and edge sensing systems. T represents the time when the action occurs. B represents the specific action, the actual action or activity performed by the actor, which is the core element of the event.

[0126] The following illustrates a prompt word format used in this invention for rainy scenarios:

[0127] Rule: In rainy weather, when the hourly rainfall reaches 45mm or more in key flood prevention areas, the train speed is limited to 120km / h...

[0128] Task Instructions: Based on the input rules and log data, generate the scheduler's operation log. The scheduler will need to process each system log entry. The format of the processed log is [Behavior Subject], [Time], [Action Occurred].

[0129] A more specific example is this: Here's an example of generating scheduling behavior based on rainfall log data:

[0130] [{{"Subject": "Rainfall Monitoring System", "Time": "18:10:54", "Behavior": "Monitored 1-hour rainfall: 47.1mm"}}, ......

[0131] ].

[0132] Therefore, the scheduler's processing log data is generated as follows: [

[0134] {{"Main Body": "Rainfall Monitoring System", "Time": "18:10:54", "Monitored 1-hour rainfall: 47.1mm"}},

[0135] {{"Subject": "Train Dispatcher", "Time": "18:10:54", "Action": "Issues train speed limit order, speed limit 45km / h"}}, ......

[0136] ].

[0137] The core of this prompt is to allow the LLM to simulate the scheduler's decision-making behavior based on rules and log data from the monitoring system, and to randomly generate some non-standard behavioral data.

[0138] As a further optimization of the embodiments of this application, in the scenario construction based on RAG, RAG is a key component of RSLM, providing institutional regulations and necessary basic knowledge for RSLM inference, and serving as the foundation for model inference. To optimize the performance of the RAG model in train operation risk identification, TSLM uses both a structured database and a vector database. This "dual-database" structure has significant advantages in several aspects. First, it enhances retrieval capabilities. By storing structured and unstructured data separately, the model can more effectively retrieve the required information, improving the relevance and accuracy of the answers. Second, it improves flexibility and adaptability, enabling the model to flexibly handle different types of data, adapt to changing information needs, and facilitate optimization and updates for specific data types. Furthermore, combining the two databases improves the comprehensive utilization rate of various types of data in train scheduling, providing a richer and more accurate knowledge background for risk identification. Finally, the rich information retrieved from the two databases enables the RAG model to generate more accurate risk assessment results and scheduling recommendations, supporting high-quality decision-making. These advantages collectively and significantly optimize the efficiency and quality of train operation risk identification.

[0139] In some examples of embodiments of this application, the risk of a scenario can be defined as the potential danger arising after the scenario's state characteristics reach a certain threshold. Combining this with the previous definition of a scenario, the risk of a scenario can be defined as follows:

[0140] Let Fs be the state characteristics in a specific scenario. The state characteristics of the scenario can be represented as:

[0141] , (3)

[0142] where f is the risk identification function of entity E. T is the safety threshold, which is generally determined by business rules. When Fs is within the safety threshold T, the scenario S is safe, and when Fs exceeds the range of T, a risk R is generated. That is, when a risk is generated . Assuming that there are m risk scenarios in the process of train operation, the risk R faced by the train operation can be represented as:

[0143] (4)

[0144] where: m represents the number of scenarios of train operation, Tj represents the safety threshold of the jth scenario, Xj represents the characteristics of the jth scenario. The formula represents the set of risk states that meet the condition , that is, as long as the condition is met, that is, the risk of the jth scenario is triggered, then will be included in the risk set .

[0145] The Target part includes the risk process and conclusion. The risk analysis strictly follows the materials provided by the scenario description, applies the thinking chain method for reasoning analysis, analyzes whether the behavior reflected in the log exists risk, and finally concludes whether the behavior exists risk.

[0146] More specifically, the risk prevention and control of train operation is divided into identification and control. Usually, the two parts are carried out separately, first identification, and then processing. At present, many researches only focus on risk identification and early warning, and the research on risk processing is less. Based on the TSLM framework proposed in the embodiments of the present application, the COT reasoning mechanism is used to provide an "end-to-end" risk prevention and control scheme, which can complete the risk identification and risk control processes simultaneously in one reasoning process. TSLM obtains the behavior data of the scenario S in real time, constructs the real-time scenario S(t), combines the rules of the scenario and the related basic knowledge to form the input of TSLM, and obtains the identification result after reasoning by the basic LLM. It can be expressed by the formula as:

[0147] (5)

[0148] where: S(t) represents the real-time scenario set, R represents the scenario rule, K represents the necessary basic knowledge, R and K are generated by the RAG module according to the characteristics of the scenario, represents the construction method of the risk reasoning Prompt. represents the application of a large model for risk identification reasoning. Re is the result of reasoning, where d represents the chain of large model reasoning processes, r represents the risk identification result, and c represents the risk treatment measures. The core of formula (5) is to apply the reasoning ability of LLM within the scene space defined by the scene, to reason on the multi-dimensional information provided by the scene, and output the corresponding reasoning result, in order to maximize the accuracy of risk control.

[0149] Finally, the prompt word generation fine-tuning instruction data is created. A prompt word (Prompt) is created based on the Few-shot Chain-of-Thought Prompting method to generate training data set for the model of scheduling behavior risk identification. The instruction part mainly clarifies the generated instructions. The input part includes control rules and S(t) two parts. The output part includes three parts of reasoning process, risk identification result and risk response scheme. An exemplary,

[0150] The following shows an example of a prompt word form of the COT fine-tuning data generation Prompt used by the present application:

[0151] Task instruction: please analyze the behavior events in the scene according to the following rule scene knowledge, and judge whether the system currently exists risk.

[0152] Requirements: adopt the way of thinking chain, strictly according to the provided rules, first carry on the logical reasoning, then draw the identification result, and give the suggestion of disposing the risk, the output result format is:

[0153] {

[0154] "Risk reasoning": reasoning process,

[0155] "Identification result": [risk exists, no risk]

[0156] "Disposal suggestion": if there is risk, output the disposal suggestion of the risk, if there is no risk, you can not output.

[0157] }

[0158] Rule knowledge: when it rains, the key flood control section 1-hour rainfall reaches 45mm and above and less than 60mm, the train speed limit is 120km / h;…

[0159] Scene: [

[0161] {"subject": "rainfall monitoring system", "time": "2024-05-14 18:10:54", "behavior": "observed 1-hour rainfall: 47.1mm"},

[0162] {"Subject": "Train Dispatcher", "Time": "2024-05-14 18:10:54", "Action": "Issued a train speed limit order, limiting the speed to 45km / h"},

[0163] ...

[0164] ].

[0165] In this embodiment, the TSLM framework utilizes the thought chain reasoning mechanism of LLM to simultaneously complete risk identification and risk processing outputs within a single reasoning process. The COT mechanism of LLM can simulate the step-by-step reasoning process of human thought, arriving at conclusions through multi-step reasoning. TSLM uses this mechanism to process multi-source input information step-by-step, progressively, to achieve comprehensive analysis of complex problems. To guide the large model in step-by-step reasoning, identification, and processing of risks during train operation, the following formula (5) is used... Input and Output Based on the definition, this study proposes a template for constructing this PROMPT.

[0166] Below is an example template for a COT-based Prompt:

[0167]

[0168] Through the above methods, the TSLM framework enables risk identification and handling reasoning to be completed in a single reasoning process by leveraging the thought chain reasoning mechanism of LLM.

[0169] In this step, a large model is used to construct a fine-tuning instruction dataset, providing an efficient and accurate solution for processing large amounts of complex data. This is particularly suitable for applications such as the train operation risk identification described in this invention, which require extracting information from rich but unstructured text content and constructing a training dataset. Therefore, a train operation risk monitoring model is trained using the fine-tuning data to infer the potential safety risks of the train under state S(t).

[0170] By introducing RAG nodes and COT structures, the TSLM model significantly improves the accuracy of knowledge involved in the Prompt, thereby comprehensively optimizing the analytical performance of TSLM. Specifically, this architectural improvement enables the model to more accurately integrate and apply relevant knowledge, enhancing the overall performance of risk identification and interpretation.

[0171] Furthermore, it also includes step S403: the train operation risk detection large language model is a fine-tuned large language model, including:

[0172] The fine-tuning is performed using the LORA method.

[0173] An important technical problem to be solved by the present application is to fine-tune a large model for train scheduling behavior risk identification using fine-tuning technology. Although various embodiments of the present application obtain a large amount of fine-tuning data from public information, the amount of data is still small compared to the amount of data required for large model training.

[0174] By using the Low-Rank Adaptation (LoRA) fine-tuning method, the adaptability of the original model to specific tasks is enhanced while maintaining the original model's ability, improving the fine-tuning efficiency and avoiding overfitting, and ensuring the original inference ability of the basic LLM.

[0175] Lora fine-tuning technology provides an effective means for optimizing large pre-trained models under data constraints. By introducing low-rank updates to specific layers of the model, Lora can achieve efficient parameter adaptation rather than full parameter adjustment, thereby reducing computational cost and improving fine-tuning efficiency. This method is particularly suitable for addressing the problem of limited computational resources and the risk of overfitting in large model fine-tuning. Due to the complexity of train scheduling behavior risk identification and the limitations of data available for fine-tuning, the present application uses Lora to fine-tune the benchmark large model.

[0176] Optionally, the method of fine-tuning a large model using Lora in the present application includes:

[0177] Fine-tuning data preparation: Prepare the fine-tuning dataset in combination with the business data required for the fine-tuning task. The quality of fine-tuning data is a decisive factor in the effectiveness of model fine-tuning.

[0178] Model selection: Select a pre-trained benchmark large model, such as Llama, ChatGLM, ERNIE, etc., which can be selected according to the characteristics of the task to be fine-tuned.

[0179] Fine-tuning training: Mainly includes identifying adjustment layers, initializing a set of low-rank matrices in the selected layers, and updating the low-rank matrices with the help of gradient descent optimization algorithms.

[0180] Evaluation and optimization: Evaluate the model performance on the validation set, monitor the fine-tuning process and avoid overfitting. According to the performance of the model on the validation set, make necessary adjustments and repeat the fine-tuning process to optimize the performance.

[0181] The optimization of Lora under data constraints provides an effective means. Lora introduces low-rank updates at the model-specific layer, allowing Lora to achieve efficient parameter adaptation rather than full parameter adjustment, thereby reducing computational cost and improving fine-tuning efficiency. This method is particularly suitable for addressing the computational resource constraints and overfitting risk problems in large model fine-tuning. Due to the complexity of train operation risk identification and the limitations of data available for fine-tuning, this study fine-tunes the benchmark large model using Lora. The main parameter configurations for this fine-tuning are shown as follows.

[0182]

[0183] In this embodiment, for the train operation risk identification scenario, through fine-tuning on specific datasets, the model can more accurately understand and adapt to the characteristics and rules of the train scheduling field; the few sample thought chain prompting method solves the problem of insufficient labeled data, and the integration of rich railway scheduling knowledge ensures the comprehensive understanding of the model to safety standards; in addition, the LORA fine-tuning technology optimizes the model under data constraints, reduces the computational cost, reduces the risk of overfitting, and through continuous evaluation and optimization, ensures the generalization ability of the model and the reliability of decision support.

[0184] As a further optimization of the embodiments of the present application, in order to ensure the best performance of the large-scale language model after fine-tuning in actual application, a dual optimization strategy combining automatic evaluation of large models and manual evaluation is adopted. First, the performance of the fine-tuned model is evaluated using automatic evaluation of large models, mainly by quantifying the performance of the model through accuracy, rationality score of reasoning, etc. Automatic evaluation of large models can quickly identify the performance of the model under different conditions, providing guidance for preliminary optimization. Second, manual evaluation is performed to supplement the details that may be missed by automatic evaluation. Specifically, an evaluation team composed of domain experts is established to strictly review the model output results. Experts evaluate the performance of the model in actual application scenarios according to their professional knowledge, such as whether the reasoning process strictly implements the requirements of the established regulations, and whether the generated scheme omits key links. Combined with the automatic evaluation results and expert rating results, the model is repeatedly iterated and optimized to gradually improve the accuracy and reliability of the model in actual application.

[0185] The train operation risk identification method based on a large language model (LLM) provided by the embodiment of the present application uses a hybrid enhancement strategy, that is, combines RAG and domain-specific fine-tuning (Fine-Tuning), to construct a model TSLM specially used for train risk monitoring. The TSLM has high accuracy, reasonable reasoning process and effective risk disposal scheme in various risk scenarios. In untrained scenarios, the TSLM can still maintain high risk identification accuracy and reasonableness, demonstrating its strong generalization ability and reliability. Through the risk identification system framework provided by the TSLM, more accurate data support is provided for train dispatch safety evaluation, and a new technical approach is provided for safety management. Through continuous optimization and improvement, the TSLM model is expected to play an important role in more practical scenarios, further ensuring the safety and efficiency of railway transportation.

[0186] Embodiment five

[0187] Embodiment five of the present application provides a non-volatile storage medium, which stores at least one executable instruction, and the executable instruction can execute the virtual conversation processing method in any method embodiment described above.

[0188] The executable instruction can be used to enable the processor to perform the operations of embodiments one to five described above.

[0189] Embodiment six

[0190] Figure 5 A structural schematic diagram of a computing device according to embodiment four of the present application is shown, and the specific embodiments of the present application do not limit the specific implementation of the computing device.

[0191] As Figure 5 shown, the computing device can include a processor 702, a communications interface 704, a memory 706, and a communications bus 708.

[0192] Among them:

[0193] The processor 702, the communications interface 704, and the memory 706 complete mutual communication through the communications bus 708.

[0194] The communications interface 704 is used to communicate with network elements of other devices such as clients or other servers.

[0195] The processor 702 is used to execute the program 710, and specifically can execute the related steps in the virtual conversation processing method embodiments described above.

[0196] In particular, the program 710 can include program code including computer operation instructions.

[0197] The processor 702 can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to perform the operations of embodiments of the application. The one or more processors included in the computing device can be of the same type or different types, such as one or more CPUs and one or more ASICs.

[0198] The memory 706 is configured to store the program 710. The memory 706 can include a high-speed RAM memory, and can also include a non-volatile memory such as at least one disk memory.

[0199] The program 710 can be specifically configured to cause the processor 702 to perform the operations of the above-mentioned embodiments one to five.

[0200] Through the scheme provided by the embodiment, train real-time monitoring and control data is acquired, train dispatching operation log data is collected, and a prompt instruction is generated based on the train real-time monitoring and control data and the train dispatching operation log data. Finally, the prompt instruction is sent to a train operation risk detection large language model, and an output result of the train operation risk detection large language model is acquired. The output result includes a risk identification result of train dispatching behavior. It can be seen that the embodiment provides a computing device that integrates a train dispatching operation safety risk identification technology with high intelligence and strong reliability, can detect risks hidden in dispatching behavior in real time, and has important value for preventing train operation safety problems.

[0201] The algorithms and displays presented herein are not inherently related to any particular computer, virtual system, or other apparatus. Various general purpose systems can be used with programs in accordance with the teachings herein, or it can prove convenient to construct more specialized apparatus to perform the required method steps. The required structure for a variety of these systems will be apparent from the description above. In addition, the present application is not intended to be limited to a particular programming language. It will be appreciated that there are many programming languages that can be used to implement the teachings herein, and any specific language can be chosen for use in this application.

[0202] In the specification provided herein, a large number of specific details are described. However, it can be understood that embodiments of the application can be practiced without these specific details. In some examples, well-known methods, structures and techniques are not described in detail in order not to obscure the understanding of the specification.

[0203] Similarly, it is to be understood that the embodiments of the present application can be altered, combined, or modified and constitute matters of application within the scope of the inventive subject matter embodied in the present application. Accordingly, the scope of the present application is not intended to be limited to the described embodiments thereof but is to be accorded the widest scope consistent with the principles and novel features to the known art.

[0204] Those skilled in the art will appreciate that the modules in the apparatuses in the embodiments can be adapted and placed in one or more apparatuses other than the embodiments. The modules or units or components in the embodiments can be combined into one module or unit or component, and further can be divided into more sub-modules or sub-units or sub-components. Any combination of all the features disclosed in the specification (including the accompanying claims, abstract and drawings), and any method or of the apparatuses disclosed in the specification (including the accompanying claims, abstract and drawings) can be taken, except that at least some of such features and / or processes or units are mutually exclusive. Unless explicitly stated, each feature disclosed in the specification (including the accompanying claims, abstract and drawings) can be replaced by alternative features that serve the same, equivalent or similar purpose.

[0205] Further, those skilled in the art will appreciate that a combination of features of different embodiments can mean within the scope of the application and form a different embodiment. For example, in the claims, any one of the claimed embodiments can be used in any combination.

[0206] It is noted that the foregoing examples have been provided merely for the purposes of illustration and are not intended to limit the application of the application. Alternative embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. As will be apparent, certain features of the application can be utilized to advantage without necessarily following the application in all respects as can be apparent to those with ordinary skill in the art. Various modifications and changes can be made thereto by those skilled in the art which freely retain the spirit of the application, which is defined by the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps other than those listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The application can be implemented by means of both hardware and software, and the present application is not limited to hardware or software implementations. The use of the words first, second and third, etc. do not imply any ordering, but rather are used for naming purposes only.

Claims

1. A train operation risk identification method, comprising: acquiring train real-time monitoring and control data; collecting train dispatching operation log data; wherein the train dispatching operation log data is a detailed document or data record recording various operations and events in the train dispatching process; generating a prompt instruction based on the train real-time monitoring and control data and the train dispatching operation log data; sending the prompt instruction to a train operation risk detection large language model, obtaining the output result of the train operation risk detection large language model, and the output result including the risk identification result of the train dispatching behavior; wherein the train real-time monitoring and control data and the train dispatching operation log data are configured as at least part of the input data in the prompt instruction, and are used as at least part of the data analysis basis for indicating the train operation risk detection large language model to execute result output; the train real-time monitoring and control data and the train dispatching operation log data are configured to retrieve and obtain text data matching the train real-time monitoring and control data and the train dispatching operation log data from a train dispatching related knowledge database, and the text data matching the train real-time monitoring and control data and the train dispatching operation log data are configured as at least part of the input data in the prompt instruction, and are used as at least part of the data analysis basis for indicating the train operation risk detection large language model to execute result output.

2. The method of claim 1, wherein the train real-time monitoring and control data comprises: train control system data, and / or, arrival and departure plan data, and / or, real-time risk monitoring data.

3. The method of claim 1, wherein the collecting train dispatching operation log data comprises: setting a proxy module configured to collect the train dispatching operation log data; the train dispatching operation log data comprises train dispatching log text data and train dispatching instruction audio data.

4. The method of claim 3, wherein the collecting train dispatching operation log data further comprises: converting the train dispatching instruction audio data into second train dispatching instruction text data; the converting the train dispatching instruction audio data into second train dispatching instruction text data further comprises: after converting the train dispatching instruction audio data into first train dispatching instruction text data, formatting the first train dispatching instruction text data to obtain the formatted second train dispatching instruction text data.

5. The method of claim 4, wherein the converting the train dispatching instruction audio data into second train dispatching instruction text data further comprises: using a speech recognition model to convert the train dispatching instruction audio data into first train dispatching instruction text data, and using a semantic extraction large language model to perform semantic extraction on the first train dispatching instruction text data to obtain the formatted second train dispatching instruction text data.

6. The method of claim 1, wherein the train real-time monitoring and control data and the train dispatching operation log data are configured to be used to retrieve and obtain text data matching the train real-time monitoring and control data and the train dispatching operation log data from a train dispatching related knowledge database, comprising: the train dispatching related knowledge database comprises an unstructured data vector database and / or a structured data database.

7. The method of claim 1 or 6, further comprising, before the train real-time monitoring and control data and the train dispatching operation log data are configured to be used to retrieve and obtain text data matching the train real-time monitoring and control data and the train dispatching operation log data from a train dispatching related knowledge database: performing text vectorization on the train real-time monitoring and control data and the train dispatching operation log data.

8. The method of claim 1, wherein the sending of the prompt instruction to the train operation risk detection large language model and the obtaining of the output result of the train operation risk detection large language model, the output result comprising a risk identification result of train dispatching behavior, comprises: the train operation risk detection large language model is a fine-tuned large language model.

9. The method of claim 8, wherein the train operation risk detection large language model is a fine-tuned large language model, comprising: using a small sample thinking chain prompt method to generate a training data set for the fine-tuning.

10. The method of claim 8 or 9, wherein the train operation risk detection large language model is a fine-tuned large language model, comprising: performing the fine-tuning using LORA method.

11. The method of any one of claims 1-3, 5-6, 8-9, further comprising, after the sending of the prompt instruction to the train operation risk detection large language model and the obtaining of the output result of the train operation risk detection large language model, the output result comprising a risk identification result of train dispatching behavior: the risk identification result comprises a risk conclusion and a reasoning analysis process for obtaining the risk conclusion.

12. A computing device comprising: a processor, a memory, a communication interface, and a communication bus, the processor, the memory, and the communication interface being in communication with each other through the communication bus; the memory is configured to store at least one executable instruction, the executable instruction causing the processor to perform the following operations: obtaining train real-time monitoring and control data; collecting train dispatching operation log data; wherein the train dispatching operation log data is a detailed document or data record recording various operations and events in the train dispatching process; generating a prompt instruction based on the train real-time monitoring and control data and the train dispatching operation log data; sending the prompt instruction to a train operation risk detection large language model and obtaining an output result of the train operation risk detection large language model, the output result comprising a risk identification result of train dispatching behavior; The train real-time monitoring and control data and the train dispatching operation log data are configured as at least part of input data in the prompt instruction, and are used as at least part of data analysis basis for indicating the execution result output of the train operation risk detection large language model; the train real-time monitoring and control data and the train dispatching operation log data are configured to retrieve and obtain text data matched with the train real-time monitoring and control data and the train dispatching operation log data from a train dispatching related knowledge database, and the text data matched with the train real-time monitoring and control data and the train dispatching operation log data are configured as at least part of input data in the prompt instruction, and are used as at least part of data analysis basis for indicating the execution result output of the train operation risk detection large language model. 13.A computer storage medium, wherein at least one executable instruction is stored in the storage medium, and the executable instruction causes a processor to perform operations corresponding to a train operation risk identification method according to any one of claims 1-11.

Images