Metro abnormal event identification method, system, device and medium based on self-evolution

Abstract

The present application relates to the technical field of anomaly identification, and discloses a subway abnormal event identification method based on self-evolution, which comprises the following steps: obtaining historical subway video streams and user natural language requirements, and generating a structured task blueprint through cloud large model analysis; based on the blueprint, the cloud large model generates an initial set composed of multiple candidate prompt words; according to the blueprint, a corresponding test set is extracted from the historical video, and an edge small model image is started in the cloud to load the test set and build a test environment; the inference of each candidate word is carried out in the test environment, and evaluation indexes and an error sample set are obtained; the cloud large model analyzes the errors and generates an updated prompt word set, and through multiple rounds of iteration, the optimal prompt word is obtained until the indexes converge; finally, the optimal word is issued to the edge device to identify the abnormality of the real-time video stream. The present application also proposes a subway abnormal event identification system, equipment and storage medium based on self-evolution. The present application can improve the identification accuracy of subway abnormal events.

Description

Technical Field

[0001] This invention relates to the field of anomaly recognition technology, and in particular to a method, system, device and medium for recognizing subway anomaly events based on self-evolution. Background Technology

[0002] Traditional video surveillance technology in subway security primarily relies on manual monitoring and rule-based computer vision methods. Manual monitoring depends on personnel at the monitoring center spending extended periods staring at video feeds, leading to inefficiency, fatigue, and response delays. Studies show that monitoring personnel's attention span decreases by over 70% after 20 minutes of continuous work. Rule-based computer vision techniques, such as background modeling, optical flow, and frame difference methods, while capable of basic motion detection, suffer from drawbacks in the complex, crowded, and lighting-variable subway environment, including high false alarm rates, low generalization ability, strong environmental sensitivity, and high computational resource consumption.

[0003] In recent years, deep learning technologies, especially large-scale vision-language models (such as GPT-4V and Claude), have shown potential in video anomaly detection. However, their deployment is limited by real-time requirements, making it impossible to achieve low-latency responses in subway scenarios. While lightweight models can run in real-time at the edge, their semantic understanding capabilities are limited, making it difficult to accurately identify complex abnormal behaviors. Traditional model fine-tuning solutions for subway anomaly recognition require a large amount of labeled data and high computing power, resulting in high maintenance costs and difficulty in quickly adapting to new scenarios. Summary of the Invention

[0004] This invention provides a self-evolving method, system, device, and medium for identifying subway anomalies, with the main objective of improving the accuracy of subway anomaly identification.

[0005] To achieve the above objectives, this invention provides a self-evolutionary method for identifying subway anomaly events, comprising:

[0006] Step A: Obtain historical subway video streams and natural language requirements input by the target user, and convert the natural language requirements into a task blueprint to obtain the user task blueprint;

[0007] Step B: Based on the user task blueprint, generate multiple candidate prompt words using a preset cloud-based large model to obtain a candidate prompt word set;

[0008] Step C: Obtain the prompt word test set corresponding to the user task blueprint based on the historical subway video stream, and load the prompt word test set using a preset edge mini-model to construct a prompt word test environment;

[0009] Step D: Using the aforementioned prompt word testing environment, test each candidate prompt word in the candidate prompt word set to obtain the evaluation index and prompt word sample set for each candidate prompt word;

[0010] Step E: Determine whether the preset convergence condition is met based on the evaluation index of each candidate prompt word. If the preset convergence condition is not met, use the cloud-based big model to analyze the evaluation index of each candidate prompt word and the prompt word sample set to obtain the analysis results. Update the multiple candidate prompt words based on the analysis results to obtain the updated prompt word set.

[0011] Step F: Use the updated prompt word set as the candidate prompt word set, and return to step D until the evaluation index meets the preset convergence condition, and use the candidate prompt word corresponding to the evaluation index as the optimal prompt word;

[0012] Step G: Based on the optimal prompt words, use preset edge devices to identify abnormal events from preset real-time subway video streams.

[0013] Optionally, converting the natural language requirements into a task blueprint to obtain a user task blueprint includes:

[0014] Extract the core event phrases from the natural language requirements and perform semantic parsing on the core event phrases to obtain the initial event intent;

[0015] Based on the initial event intent, semantic completion is performed using a preset common sense base to generate the subject of interest, spatial relationship, action features and potential exclusions related to the initial event intent, thus obtaining supplementary semantic elements;

[0016] Based on the supplementary semantic elements, a task blueprint is constructed using a preset structured blueprint template to obtain the user task blueprint.

[0017] Optionally, based on the user task blueprint, multiple candidate prompt words are generated using a preset cloud-based large model to obtain a candidate prompt word set, including:

[0018] Extract the subject of interest, spatial relationships, action features, and potential exclusion items from the user task blueprint to obtain structured semantic elements;

[0019] The structured semantic elements are combined with each prompt template in the preset prompt template set to construct multiple meta-prompt words, thus obtaining a meta-prompt word set;

[0020] The cloud-based big data model is used to generate candidate prompt words corresponding to each meta prompt word in the meta prompt word set, thus obtaining the original candidate prompt word set;

[0021] The candidate prompt words in the original candidate prompt word set are subjected to format verification and similarity deduplication to obtain the candidate prompt word set.

[0022] Optionally, the step of loading the prompt word test set using a preset edge mini-model to construct the prompt word test environment includes:

[0023] Based on a preset cloud server, a software image instance of the edge mini-model is constructed, and the model weight file of the edge mini-model is loaded to obtain a standard mini-model instance;

[0024] The prompt word test set is imported into the memory space of the standard small model instance, and the inference parameters of the standard small model instance are configured to obtain the prompt word test environment.

[0025] Optionally, the analysis of the evaluation metrics of each candidate prompt word and the prompt word sample set using the cloud-based big data model to obtain the analysis results includes:

[0026] The evaluation metrics of each candidate prompt word are analyzed to identify the metrics items below a preset threshold, resulting in multiple screening metrics. Prompt word samples corresponding to the multiple screening metrics are extracted from the prompt word sample set to obtain the error sample set to be analyzed.

[0027] The cloud-based big model is used to perform visual content understanding and false alarm reasoning on each error sample in the set of error samples to be analyzed, and a single sample diagnosis result is obtained.

[0028] Cluster analysis was performed on the single-sample diagnostic results, and the sample category with the highest error frequency was counted to obtain the failure sample classification;

[0029] Based on the classification of the failed samples, the cloud-based large model is used to generate correction suggestions, and the analysis results are generated based on the classification of the failed samples and the correction suggestions.

[0030] Optionally, updating the plurality of candidate prompt words based on the analysis results to obtain an updated prompt word set includes:

[0031] Based on the classification of failed samples in the analysis results, the corresponding prompt word correction template is selected from the preset optimization strategy library to obtain the target correction template;

[0032] Based on the correction suggestions in the analysis results and the target correction template, the multiple candidate prompt words are updated to obtain an updated prompt word set.

[0033] Optionally, the step of using the prompt word testing environment to test each candidate prompt word in the candidate prompt word set to obtain the evaluation index and prompt word sample set for each candidate prompt word includes:

[0034] Based on the aforementioned prompt word testing environment, the preset edge small model mirror is used to infer each candidate prompt word in the candidate prompt word set, and the prediction result corresponding to each candidate prompt word is obtained.

[0035] Obtain the true labels of each candidate prompt word, and compare the prediction results corresponding to each candidate prompt word with the true labels of each candidate prompt word to obtain the comparison results corresponding to multiple candidate prompt words;

[0036] Based on the comparison results, the evaluation indicators of each candidate prompt word are statistically analyzed, and candidate prompt words whose prediction results are inconsistent with the actual labels are selected from the candidate prompt word set to generate a prompt word sample set.

[0037] To address the aforementioned problems, this invention also provides a self-evolving subway anomaly event identification system, the system comprising:

[0038] The prompt word generation module is used to obtain historical subway video streams and natural language requirements input by the target user, and convert the natural language requirements into a task blueprint to obtain the user task blueprint;

[0039] Based on the user task blueprint, multiple candidate prompt words are generated using a preset cloud-based large model to obtain a candidate prompt word set;

[0040] The prompt word testing module is used to obtain the prompt word test set corresponding to the user task blueprint based on the historical subway video stream, and to load the prompt word test set using a preset edge mini-model to construct a prompt word testing environment;

[0041] Using the aforementioned prompt word testing environment, each candidate prompt word in the candidate prompt word set is tested to obtain the evaluation index and prompt word sample set for each candidate prompt word;

[0042] The prompt word update module is used in step E to determine whether the preset convergence condition is met based on the evaluation index of each candidate prompt word.

[0043] If the preset convergence condition is not met, the evaluation index of each candidate prompt word and the prompt word sample set are analyzed using the cloud-based big model to obtain the analysis results, and the multiple candidate prompt words are updated according to the analysis results to obtain the updated prompt word set;

[0044] An abnormal event identification module is used to take the updated prompt word set as the candidate prompt word set, and return to use the prompt word testing environment to test each candidate prompt word in the candidate prompt word set to obtain the evaluation index and prompt word sample set of each candidate prompt word until the evaluation index meets the preset convergence condition, and the candidate prompt word corresponding to the evaluation index is taken as the optimal prompt word.

[0045] Based on the optimal prompt words, abnormal events are identified from a preset real-time subway video stream using preset edge devices.

[0046] To address the above problems, the present invention also provides an electronic device, the electronic device comprising:

[0047] At least one processor; and,

[0048] A memory communicatively connected to the at least one processor; wherein,

[0049] The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the self-evolutionary subway anomaly identification method as described above.

[0050] To address the aforementioned problems, the present invention also provides a computer-readable storage medium, characterized in that it includes a data storage area and a program storage area, wherein the data storage area stores created data and the program storage area stores a computer program; wherein, when the computer program is executed by a processor, it implements the self-evolutionary subway anomaly event identification method described above.

[0051] This invention acquires historical subway video streams and user natural language requirements, and generates a structured task blueprint through cloud-based large-scale model parsing. Based on the blueprint, the cloud-based large-scale model generates multiple candidate prompt words to form an initial set. According to the blueprint, a corresponding test set is extracted from historical videos, and a test environment is constructed by mirroring the test set in a cloud-based edge small-scale model. In the test environment, reasoning is performed on each candidate word to obtain evaluation indicators and an error sample set. The cloud-based large-scale model analyzes errors and generates an updated prompt word set, iterating through multiple rounds until the indicators converge to obtain the optimal prompt word. Finally, the optimal word is sent to edge devices for anomaly identification of the real-time video stream. Therefore, the self-evolving subway anomaly event identification method, system, electronic device, and computer-readable storage medium proposed in this invention, by constructing an edge-cloud collaborative architecture of "cloud-optimized prompt words and edge-executed reasoning," enables the system to quickly adapt to new anomaly types with "zero samples" without modifying model weights. This solves the contradiction between limited edge computing power and complex semantic understanding, and improves the accuracy of subway anomaly event identification. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating a self-evolutionary subway anomaly event identification method according to an embodiment of the present invention.

[0053] Figure 2 A schematic diagram of a self-evolving subway anomaly event recognition system provided in an embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram of the internal structure of an electronic device that implements a self-evolving subway anomaly event recognition method according to an embodiment of the present invention.

[0055] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0056] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0057] This application provides a self-evolving method for identifying subway anomaly events. The executing entity of this self-evolving subway anomaly event identification method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. The server can be a standalone server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. In other words, the self-evolving subway anomaly event identification method can be executed by software or hardware installed on remote devices or server-side devices, and the software can be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0058] Reference Figure 1 The diagram shown is a flowchart illustrating a self-evolutionary subway anomaly event identification method according to an embodiment of the present invention. In this embodiment, the self-evolutionary subway anomaly event identification method includes the following steps S1-S5:

[0059] S1. Obtain historical subway video streams and natural language requirements input by the target user, and convert the natural language requirements into a task blueprint to obtain the user task blueprint.

[0060] Natural language requirements refer to the colloquial instructions that subway maintenance personnel input through the human-computer interaction interface to describe the need to identify abnormal events, such as "keep an eye on whether anyone climbs over the turnstile" or "check if anyone is fighting." These requirements are expressed in natural language and may contain contextual information, but they lack the precise semantic structure that machines can directly execute.

[0061] Among them, historical subway video streams refer to surveillance video data collected and stored at subway stations within past time periods. These videos have been annotated and include typical segments of various abnormal events (such as climbing over turnstiles, fighting, and track intrusion) and normal scenes. This serves as the original resource library for constructing the prompt word test set, providing standardized test samples for subsequent offline verification and optimization of candidate prompt words.

[0062] The task blueprint is a standardized data representation generated after the above natural language requirements are parsed, semantically completed, and structured by a large cloud model. It usually includes predefined fields such as the subject of interest, spatial relationships, action features, exclusions, and contextual clues, which can completely and unambiguously depict the abnormal events that the user expects to identify.

[0063] Furthermore, the task blueprint needs to be structured because structuring can eliminate the ambiguity of natural language, break down vague intentions into machine-combinable and searchable semantic units, provide a unified input for the subsequent automated generation of diverse candidate prompts, provide search conditions for accurately matching the corresponding test dataset, and ensure the standardization and consistency of data exchange between multiple modules such as large cloud models, small edge models, and test environments.

[0064] Understandably, by transforming users' vague spoken instructions into structured task descriptions that machines can understand, and by introducing historical video data as a verification basis for subsequent optimization, accurate semantic evidence and test materials are provided for the subsequent automatic generation and iterative optimization of prompt words.

[0065] Furthermore, the step of converting the natural language requirements into a task blueprint to obtain a user task blueprint includes:

[0066] Extract the core event phrases from the natural language requirements and perform semantic parsing on the core event phrases to obtain the initial event intent;

[0067] Based on the initial event intent, semantic completion is performed using a preset common sense base to generate the subject of interest, spatial relationship, action features and potential exclusions related to the initial event intent, thus obtaining supplementary semantic elements;

[0068] Based on the supplementary semantic elements, a task blueprint is constructed using a preset structured blueprint template to obtain the user task blueprint.

[0069] Among them, the core event phrase refers to the smallest semantic unit that expresses the core of the user's attention, extracted from the natural language requirements input by the user. It is usually composed of the core verb and its key arguments. It is the original linguistic carrier of the user's intention but has not yet been structured.

[0070] The initial event intent refers to the machine-readable semantic skeleton in a structured form obtained after semantic parsing of the core event phrase. It typically includes standardized semantic roles such as action elements and object elements, providing a standardized version of the core event phrase that can be directly processed by subsequent steps.

[0071] Among them, the common sense base refers to the general knowledge system embedded in the model parameters obtained by the cloud-based large model through massive pre-training. It covers knowledge such as entity attributes, action features, spatial relationships and excluded scenarios, and is used to intelligently complete the details and expand the context of the initial event intent.

[0072] Among them, the structured blueprint template refers to a predefined, standardized data format specification used to organize the complete semantic elements of abnormal events. It specifies fixed fields such as the subject of concern, spatial relationships, action characteristics, and exclusion items to ensure that the generated user task blueprints have a consistent machine-resolvable structure.

[0073] S2. Based on the user task blueprint, generate multiple candidate prompt words using a preset cloud-based large model to obtain a candidate prompt word set.

[0074] Understandably, generating multiple candidate prompts based on the user task blueprint is equivalent to transforming the structured task definition into multiple natural language instruction versions with different expression styles and reasoning logics. This provides diverse initial inputs for subsequent automated testing and iterative optimization in the mirror environment, ensuring that the optimal instruction most suitable for the edge small model can be selected from multiple candidate versions.

[0075] Among them, the cloud-based large model refers to a large-parameter model deployed in a data center with powerful semantic understanding and logical reasoning capabilities. In the implementation of this invention, it plays the roles of "teacher" and "optimizer", responsible for generating candidate prompt words according to the task blueprint and reflecting on and analyzing the test results to drive the evolution of prompt words.

[0076] Among them, candidate prompts refer to multiple natural language instructions with different expression styles generated by the large cloud model based on the task blueprint, which are used to guide the small edge model to identify anomalies. They are the initial input versions for subsequent automated testing and iterative optimization in the mirror environment.

[0077] Furthermore, based on the user task blueprint, multiple candidate prompt words are generated using a preset cloud-based large model to obtain a candidate prompt word set, including:

[0078] Extract the subject of interest, spatial relationships, action features, and potential exclusion items from the user task blueprint to obtain structured semantic elements;

[0079] The structured semantic elements are combined with each prompt template in the preset prompt template set to construct multiple meta-prompt words, thus obtaining a meta-prompt word set;

[0080] The cloud-based big data model is used to generate candidate prompt words corresponding to each meta prompt word in the meta prompt word set, thus obtaining the original candidate prompt word set;

[0081] The candidate prompt words in the original candidate prompt word set are subjected to format verification and similarity deduplication to obtain the candidate prompt word set.

[0082] Among them, structured semantic elements refer to the set of core features of abnormal events extracted from the user task blueprint and processed in a standardized manner. Structured semantic elements present key information such as the subject of interest, spatial relationships, action features and exclusion items in a unified format, providing standardized input data for subsequent prompt word generation.

[0083] The prompt template set refers to a collection of pre-defined prompt word writing frameworks, including templates with different styles such as concise description, step-by-step reasoning, and constraint emphasis, which are used to guide the cloud-based large model to generate diverse candidate prompt words from different perspectives.

[0084] Among them, meta-prompt words refer to composite instructions constructed by combining structured semantic elements with specific prompt templates, used to guide large cloud models to generate specific prompt words. It tells the large model "what style and what elements to generate a prompt word based on".

[0085] Among them, the original candidate prompt words refer to the initial version prompt words generated directly by the cloud-based large model based on the meta prompt words, without any post-processing. These original outputs may have problems such as non-standard format or duplicate content, and need to be verified and deduplicated in the future to form a formal set of candidate prompt words.

[0086] S3. Based on the historical subway video stream, obtain the prompt word test set corresponding to the user task blueprint, and load the prompt word test set using a preset edge mini-model to construct a prompt word test environment.

[0087] Understandably, by providing a repeatable offline verification platform that is completely consistent with the physical equipment for the automated testing and iterative optimization of subsequent prompt words without affecting the real-time monitoring tasks of subway stations, it ensures that all trial and error processes are completed securely in the cloud mirror, and only the optimal instructions that have been verified will be finally sent to the real equipment.

[0088] The prompt word test set refers to the labeled video dataset corresponding to the user task blueprint, which is retrieved from historical subway video streams. It contains positive samples of target abnormal events and negative samples that are prone to false alarms, and is used to objectively evaluate the recognition performance of each candidate prompt word in a cloud mirror environment.

[0089] Among them, the edge small model refers to the lightweight visual language model deployed on the edge computing gateway of the subway station. In this step, it specifically refers to its software image instance in the cloud server. This image instance has the same software environment and model weights as the real physical device.

[0090] The prompt word testing environment refers to a closed testing platform built in the cloud, consisting of edge small model images and prompt word test sets. It simulates the operating state of real edge devices and is used to perform security verification and performance evaluation of candidate prompt words without affecting real-time monitoring.

[0091] Furthermore, the step of loading the prompt word test set using a preset edge mini-model to construct the prompt word test environment includes:

[0092] Based on a preset cloud server, a software image instance of the edge mini-model is constructed, and the model weight file of the edge mini-model is loaded to obtain a standard mini-model instance;

[0093] The prompt word test set is imported into the memory space of the standard small model instance, and the inference parameters of the standard small model instance are configured to obtain the prompt word test environment.

[0094] Among them, the cloud server refers to a remote server cluster deployed in a data center with powerful computing and storage capabilities. In this solution, it is used to run the digital twin image, execute the prompt word optimization algorithm, and store the abnormal event sample library, providing cloud-based intelligent support for edge devices.

[0095] Among them, a software image instance refers to a virtual copy running on a cloud server that is completely identical to the software environment of a real edge device. It contains the same operating system, model version, and runtime configuration, and is used to conduct offline testing and verification of prompt words without affecting real monitoring tasks.

[0096] Among them, the model weight file refers to the parameter file obtained after the edge small model is trained. It stores the knowledge that the model has learned from the data. Only after loading this file can the model have the ability to understand and recognize video content. It is the core technology carrier for realizing anomaly detection.

[0097] Among them, the standard small model instance refers to a software image instance running entity that has successfully loaded the model weight file and completed the inference parameter configuration on the cloud server and is in a ready state. It acts as a "shadow student" to perform actual inference test tasks in the cloud environment.

[0098] S4. Using the aforementioned prompt word testing environment, test each candidate prompt word in the candidate prompt word set to obtain the evaluation index and prompt word sample set for each candidate prompt word.

[0099] Understandably, by objectively evaluating each candidate suggestion in a standardized testing environment, the system can quantify the performance of each suggestion in terms of accuracy and recall, and accurately locate which samples were misidentified, providing a reliable data foundation for subsequent in-depth analysis and targeted optimization of large cloud models.

[0100] The evaluation metrics refer to the quantitative performance data obtained after performing inference tests on the prompt word test set for each candidate prompt word. These metrics typically include accuracy, recall, and false positive rate, and are used to measure the prompt word's ability to identify abnormal events.

[0101] The prompt word sample set refers to the set of samples selected from the prompt word test set during the testing process where the predicted results of the current candidate prompt word are inconsistent with the true label. It includes missed alarm samples (should have triggered an alarm but did not) and false alarm samples (should not have triggered an alarm but did). These samples are the core basis for subsequent failure mode analysis and targeted optimization of the cloud-based large model.

[0102] Furthermore, the step of using the prompt word testing environment to test each candidate prompt word in the candidate prompt word set to obtain the evaluation index and prompt word sample set for each candidate prompt word includes:

[0103] Based on the aforementioned prompt word testing environment, the preset edge small model mirror is used to infer each candidate prompt word in the candidate prompt word set, and the prediction result corresponding to each candidate prompt word is obtained.

[0104] Obtain the true labels of each candidate prompt word, and compare the prediction results corresponding to each candidate prompt word with the true labels of each candidate prompt word to obtain the comparison results corresponding to multiple candidate prompt words;

[0105] Based on the comparison results, the evaluation indicators of each candidate prompt word are statistically analyzed, and candidate prompt words whose prediction results are inconsistent with the actual labels are selected from the candidate prompt word set to generate a prompt word sample set.

[0106] Among them, the real label refers to the category identifier that is pre-labeled for each video sample in the prompt word test set and reflects its real situation. It is used to mark whether the sample contains the target abnormal event (such as "climbing over the gate") or belongs to the normal scene, and serves as an objective reference standard for evaluating the accuracy of the prediction results.

[0107] The comparison result refers to the matching record obtained by comparing the prediction result corresponding to each candidate prompt word with the real label of the test sample one by one. It identifies whether the prediction of each sample is correct or incorrect, as well as the error type (missed or false alarm), and provides basic data for subsequent statistical evaluation indicators and screening of erroneous samples.

[0108] Among them, the edge small model image refers to a virtual copy running on a cloud server that is completely identical to the software environment of the real edge device. It has loaded the model weight file of the edge small model and completed the inference parameter configuration, and is in a ready state to perform the inference test task of candidate prompt words.

[0109] Evaluation metrics are numerical measures used to quantify the performance of candidate prompt word recognition, which are derived from the statistical results of the comparison. They typically include accuracy (the proportion of correctly predicted words in all samples), recall (the proportion of correctly identified words in abnormal samples), and false alarm rate (the proportion of false alarms in normal samples). These metrics together constitute the objective basis for evaluating the quality of candidate prompt words.

[0110] S5. Determine whether the preset convergence condition is met based on the evaluation index of each candidate prompt word. If the preset convergence condition is not met, use the cloud-based big data model to analyze the evaluation index of each candidate prompt word and the prompt word sample set to obtain the analysis results. Update the multiple candidate prompt words based on the analysis results to obtain the updated prompt word set.

[0111] Understandably, by comparing the performance of the current prompt with the preset standard, the system can autonomously determine whether the expected goal has been achieved. This allows it to terminate the iteration in a timely manner while ensuring recognition quality, avoiding ineffective optimization, and ensuring that the prompts finally sent to edge devices are the optimal versions that have been objectively verified.

[0112] The preset convergence condition refers to the quantitative standard used during the iterative process to determine whether the prompt word optimization has achieved the expected goal and when iteration can be stopped. This condition is usually set based on the evaluation index obtained in step D. For example, when the recognition accuracy of the prompt word on the prompt word test set exceeds 95%, the index improvement of three consecutive iterations is less than 1%, or the number of error samples drops below a preset threshold, the convergence condition is considered to be met. The setting of this condition aims to balance the optimization effect and the iteration cost, and avoid infinite loops.

[0113] Understandably, by analyzing error cases in test results using a large cloud-based model, the system diagnoses the defects in the current prompt words and then generates improved new prompt words accordingly. This allows the system to learn from each test and continuously optimize the quality of instructions without manual intervention or model retraining.

[0114] In this embodiment of the invention, the diagnostic information output by the cloud-based large model after in-depth analysis of the evaluation indicators and the prompt word sample set typically includes failure sample classification (such as "misclassifying an object as a person") and targeted optimization suggestions (such as "adding leg movement constraints"), which are used to guide subsequent prompt word updates.

[0115] Furthermore, the analysis of the evaluation metrics of each candidate prompt word and the prompt word sample set using the cloud-based big data model yields the following analysis results:

[0116] The evaluation metrics of each candidate prompt word are analyzed to identify the metrics items below a preset threshold, resulting in multiple screening metrics. Prompt word samples corresponding to the multiple screening metrics are extracted from the prompt word sample set to obtain the error sample set to be analyzed.

[0117] The cloud-based big model is used to perform visual content understanding and false alarm reasoning on each error sample in the set of error samples to be analyzed, and a single sample diagnosis result is obtained.

[0118] Cluster analysis was performed on the single-sample diagnostic results, and the sample category with the highest error frequency was counted to obtain the failure sample classification;

[0119] Based on the classification of the failed samples, the cloud-based large model is used to generate correction suggestions, and the analysis results are generated based on the classification of the failed samples and the correction suggestions.

[0120] Among them, the preset threshold refers to the pre-set quantitative standard boundary value used to measure whether the performance of candidate prompt words is qualified. For example, the recall threshold is set to 0.9 and the false positive rate threshold is set to 0.05. When the evaluation index is lower or higher than the boundary, the subsequent optimization process is triggered.

[0121] Among them, the screening index refers to the specific index item that fails to meet the preset threshold requirements from the evaluation index of each candidate suggestion word. For example, if the recall rate of a certain candidate suggestion word is 0.85, which is lower than the threshold of 0.9, the recall rate index is marked as the screening index and is used to locate the performance shortcomings that need to be optimized.

[0122] Among them, the error samples to be analyzed refer to the set of error samples that are selected from the set of prompt words and correspond to the screening indicators. For example, in response to the problem of low recall, those missed samples that should have triggered an alarm but did not are specifically extracted as the objects to be analyzed for subsequent in-depth diagnosis.

[0123] Among them, the single-sample diagnostic result refers to the individual analysis conclusion output by the cloud-based large model after performing visual content understanding and false alarm reasoning on a single erroneous sample. For example, "the large luggage carried by the pedestrian in this sample was misidentified by the model as a human torso," which reveals the specific reason for the failure of a single sample.

[0124] Among them, failure sample classification refers to the common error types summarized after cluster analysis of multiple single sample diagnostic results. For example, multiple individual diagnostic results of "luggage misidentified as a person" are grouped into the failure pattern of "failure to distinguish between objects and human bodies" to guide the generation of targeted correction suggestions in the future.

[0125] Further, updating the plurality of candidate prompt words based on the analysis results to obtain an updated prompt word set includes:

[0126] Based on the classification of failed samples in the analysis results, the corresponding prompt word correction template is selected from the preset optimization strategy library to obtain the target correction template;

[0127] Based on the correction suggestions in the analysis results and the target correction template, the multiple candidate prompt words are updated to obtain an updated prompt word set.

[0128] The target correction template refers to a prompt word correction framework customized for specific failure modes, matched from a pre-defined optimization strategy library. It has built-in correction logic and instruction format for this type of error (such as "object misidentified as a person"), which is used to guide the large cloud model to generate new prompt words that can compensate for the defect.

[0129] The optimization strategy library refers to a pre-built knowledge base that stores various failure modes and their corresponding correction strategies. It contains multiple prompt word correction templates and optimization rules for different error types (such as "object misidentified as a person" or "ignoring contextual features"), which are used to quickly match the corresponding correction scheme after the failure mode is analyzed, and guide the cloud-based large model to generate targeted optimization prompt words.

[0130] S6. Use the updated prompt word set as the candidate prompt word set, and return to S4 until the evaluation index meets the preset convergence condition, and use the candidate prompt word corresponding to the evaluation index as the optimal prompt word.

[0131] Understandably, by analyzing error cases in test results using a large cloud-based model, the system diagnoses the defects in the current prompts, then generates improved prompts and re-tests them in an edge computing environment. This allows the system to learn from each test, continuously optimizing instruction quality without manual intervention or model retraining, thus automatically approaching the optimal prompts for specific scenarios in iterations.

[0132] S7. Based on the optimal prompt words, use preset edge devices to identify abnormal events from preset real-time subway video streams.

[0133] Understandably, by using the optimal prompt words obtained through iterative optimization and employing pre-set edge devices to identify abnormal events in real-time subway video streams, the semantic understanding capabilities of cloud-based large-scale models can be combined with the low-latency response advantages of edge devices. This enables the system to accurately identify various pre-set abnormal events (such as passenger falls, left-behind items, and area intrusions) in real-world scenarios, while also eliminating continuous reliance on network bandwidth. This allows for localized real-time monitoring and rapid alarms, effectively improving the intelligent assurance level of subway operation safety.

[0134] The optimal prompt word refers to the final version of the prompt word obtained after multiple rounds of iterative cycles of "generation-testing-feedback-optimization" when the evaluation metrics meet the preset convergence conditions. This prompt word has been fully validated in the prompt word testing environment and has been proven to be the most accurate and stable tool for identifying abnormal events in current edge devices and specific subway video scenarios. Therefore, it was selected as the instruction template for final deployment in practical applications.

[0135] This invention, based on a fixed sampling period, collects data from the capacitive touch sensors of the target seats. It acquires historical subway video streams and user natural language requirements, and generates a structured task blueprint through cloud-based large-scale model analysis. Based on the blueprint, the cloud-based large-scale model generates multiple candidate prompt words to form an initial set. According to the blueprint, corresponding test sets are extracted from historical videos, and a software image instance is launched in the cloud to load the test sets and construct a test environment. In the test environment, reasoning is performed on each candidate word to obtain evaluation indicators and an error sample set. The cloud-based large-scale model analyzes errors and generates an updated prompt word set, iterating through multiple rounds until the indicators converge to obtain the optimal prompt word. Finally, the optimal word is sent to edge devices for anomaly identification of the real-time video stream. Therefore, the self-evolving subway anomaly event identification method, system, electronic device, and computer-readable storage medium proposed in this invention, by constructing an edge-cloud collaborative architecture of "cloud-optimized prompt words and edge-executed reasoning," enables the system to quickly adapt to new anomaly types with "zero samples" without modifying model weights. This resolves the contradiction between limited edge computing power and complex semantic understanding, and improves the accuracy of subway anomaly event identification.

[0136] like Figure 2 The diagram shown is a schematic of the modules of the self-evolving subway abnormal event recognition system of the present invention.

[0137] The self-evolving subway anomaly event recognition system 100 described in this invention can be installed in an electronic device. Depending on the functions implemented, the self-evolving subway anomaly event recognition system may include a prompt word generation module 101, a prompt word testing module 102, a prompt word updating module 103, and an anomaly event recognition module 104. The module described in this invention can also be called a unit, referring to a series of computer program segments that can be executed by the processor of an electronic device and perform a fixed function, stored in the memory of the electronic device.

[0138] In this embodiment, the functions of each module / unit are as follows:

[0139] The prompt word generation module 101 is used to obtain historical subway video streams and natural language requirements input by the target user, and convert the natural language requirements into a task blueprint to obtain a user task blueprint.

[0140] Based on the user task blueprint, multiple candidate prompt words are generated using a preset cloud-based large model to obtain a candidate prompt word set;

[0141] The prompt word testing module 102 is used to obtain the prompt word test set corresponding to the user task blueprint based on the historical subway video stream, and to load the prompt word test set using a preset edge mini-model to construct a prompt word testing environment;

[0142] Using the aforementioned prompt word testing environment, each candidate prompt word in the candidate prompt word set is tested to obtain the evaluation index and prompt word sample set for each candidate prompt word;

[0143] The prompt word update module 103 is used in step E to determine whether the preset convergence condition is met based on the evaluation index of each candidate prompt word.

[0144] If the preset convergence condition is not met, the evaluation index of each candidate prompt word and the prompt word sample set are analyzed using the cloud-based big model to obtain the analysis results, and the multiple candidate prompt words are updated according to the analysis results to obtain the updated prompt word set;

[0145] The abnormal event identification module 104 is used to take the updated prompt word set as the candidate prompt word set, and return to use the prompt word test environment to test each candidate prompt word in the candidate prompt word set to obtain the evaluation index and prompt word sample set of each candidate prompt word until the evaluation index meets the preset convergence condition, and the candidate prompt word corresponding to the evaluation index is taken as the optimal prompt word.

[0146] Based on the optimal prompt words, abnormal events are identified from a preset real-time subway video stream using preset edge devices.

[0147] In detail, the modules in the self-evolving subway anomaly event recognition system 100 described in this embodiment of the invention employ the same methods as described above. Figure 1 The method used is the same as the self-evolutionary subway anomaly event identification method and can produce the same technical effect, so it will not be described in detail here.

[0148] like Figure 3 The diagram shown is a structural schematic of the electronic device that implements the self-evolutionary subway anomaly event recognition method of the present invention.

[0149] The electronic device may include a processor 10, a memory 11, a communication bus 12, and a communication interface 13. It may also include a computer program stored in the memory 11 and capable of running on the processor 10, such as a self-evolving subway anomaly event identification program.

[0150] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory 11 (e.g., executing a self-evolving subway anomaly event recognition program) and calls data stored in the memory 11 to perform various functions of the electronic device and process data.

[0151] The memory 11 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of an electronic device, such as a portable hard drive. In other embodiments, the memory 11 can be an external storage device of the electronic device, such as a plug-in portable hard drive, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc. Furthermore, the memory 11 can include both internal and external storage units of the electronic device. The memory 11 can be used not only to store application software and various types of data installed on the electronic device, such as code for a self-evolving subway anomaly detection program, but also to temporarily store data that has been output or will be output.

[0152] The communication bus 12 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into an address bus, a data bus, a control bus, etc. The bus is configured to enable communication between the memory 11 and at least one processor 10, etc.

[0153] The communication interface 13 is used for communication between the aforementioned electronic device and other devices, including a network interface and a user interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, Bluetooth interface, etc.), typically used to establish communication connections between the electronic device and other electronic devices. The user interface may be a display, an input unit (such as a keyboard), or optionally, a standard wired or wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device and to display a visual user interface.

[0154] Figure 3 Only electronic devices with components are shown; it will be understood by those skilled in the art that... Figure 3The structure shown does not constitute a limitation on the electronic device and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0155] For example, although not shown, the electronic device may also include a power supply (such as a battery) to power the various components. Preferably, the power supply can be logically connected to the at least one processor 10 through a power management device, thereby enabling functions such as charging management, discharging management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0156] It should be understood that the embodiments described are for illustrative purposes only and are not limited to this structure in the scope of the patent application.

[0157] The self-evolving subway anomaly detection program stored in the memory 11 of the electronic device is a combination of multiple computer programs. When run in the processor 10, it can achieve the following:

[0158] Step A: Obtain historical subway video streams and natural language requirements input by the target user, and convert the natural language requirements into a task blueprint to obtain the user task blueprint;

[0159] Step B: Based on the user task blueprint, generate multiple candidate prompt words using a preset cloud-based large model to obtain a candidate prompt word set;

[0160] Step C: Obtain the prompt word test set corresponding to the user task blueprint based on the historical subway video stream, and load the prompt word test set using a preset edge mini-model to construct a prompt word test environment;

[0161] Step D: Using the aforementioned prompt word testing environment, test each candidate prompt word in the candidate prompt word set to obtain the evaluation index and prompt word sample set for each candidate prompt word;

[0162] Step E: Determine whether the preset convergence condition is met based on the evaluation index of each candidate prompt word. If the preset convergence condition is not met, use the cloud-based big model to analyze the evaluation index of each candidate prompt word and the prompt word sample set to obtain the analysis results. Update the multiple candidate prompt words based on the analysis results to obtain the updated prompt word set.

[0163] Step F: Use the updated prompt word set as the candidate prompt word set, and return to step D until the evaluation index meets the preset convergence condition, and use the candidate prompt word corresponding to the evaluation index as the optimal prompt word;

[0164] Step G: Based on the optimal prompt words, use preset edge devices to identify abnormal events from preset real-time subway video streams.

[0165] Specifically, the processor 10's implementation method of the above-mentioned computer program can be found in [reference needed]. Figure 1 The descriptions of the relevant steps in the corresponding embodiments are not repeated here.

[0166] Furthermore, if the modules / units integrated into the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile. For example, the computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).

[0167] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor of an electronic device, can perform the following:

[0168] Step A: Obtain historical subway video streams and natural language requirements input by the target user, and convert the natural language requirements into a task blueprint to obtain the user task blueprint;

[0169] Step B: Based on the user task blueprint, generate multiple candidate prompt words using a preset cloud-based large model to obtain a candidate prompt word set;

[0170] Step C: Obtain the prompt word test set corresponding to the user task blueprint based on the historical subway video stream, and load the prompt word test set using a preset edge mini-model to construct a prompt word test environment;

[0171] Step D: Using the aforementioned prompt word testing environment, test each candidate prompt word in the candidate prompt word set to obtain the evaluation index and prompt word sample set for each candidate prompt word;

[0172] Step E: Determine whether the preset convergence condition is met based on the evaluation index of each candidate prompt word. If the preset convergence condition is not met, use the cloud-based big model to analyze the evaluation index of each candidate prompt word and the prompt word sample set to obtain the analysis results. Update the multiple candidate prompt words based on the analysis results to obtain the updated prompt word set.

[0173] Step F: Use the updated prompt word set as the candidate prompt word set, and return to step D until the evaluation index meets the preset convergence condition, and use the candidate prompt word corresponding to the evaluation index as the optimal prompt word;

[0174] Step G: Based on the optimal prompt words, use preset edge devices to identify abnormal events from preset real-time subway video streams.

[0175] In the several embodiments provided by this invention, it should be understood that the disclosed devices, systems, and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0176] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0177] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0178] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0179] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0180] The blockchain referred to in this invention is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0181] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0182] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or systems stated in a system claim may also be implemented by a single unit or system through software or hardware. The term "second class" is used to indicate names and does not indicate any specific order.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying subway anomaly events based on self-evolution, characterized in that, The method includes: Step A: Obtain historical subway video streams and natural language requirements input by the target user, and convert the natural language requirements into a task blueprint to obtain the user task blueprint; Step B: Based on the user task blueprint, generate multiple candidate prompt words using a preset cloud-based large model to obtain a candidate prompt word set; Step C: Obtain the prompt word test set corresponding to the user task blueprint based on the historical subway video stream, and load the prompt word test set using a preset edge mini-model to construct a prompt word test environment; Step D: Using the aforementioned prompt word testing environment, test each candidate prompt word in the candidate prompt word set to obtain the evaluation index and prompt word sample set for each candidate prompt word; Step E: Determine whether the preset convergence condition is met based on the evaluation index of each candidate prompt word. If the preset convergence condition is not met, use the cloud-based big model to analyze the evaluation index of each candidate prompt word and the prompt word sample set to obtain the analysis results. Update the multiple candidate prompt words based on the analysis results to obtain the updated prompt word set. Step F: Use the updated prompt word set as the candidate prompt word set, and return to step D until the evaluation index meets the preset convergence condition, and use the candidate prompt word corresponding to the evaluation index as the optimal prompt word; Step G: Based on the optimal prompt words, use preset edge devices to identify abnormal events from preset real-time subway video streams; The step of using the cloud-based big data model to analyze the evaluation metrics of each candidate prompt word and the prompt word sample set to obtain analysis results includes: The evaluation metrics of each candidate prompt word are analyzed to identify the metrics items below a preset threshold, resulting in multiple screening metrics. Prompt word samples corresponding to the multiple screening metrics are extracted from the prompt word sample set to obtain the error sample set to be analyzed. The cloud-based big model is used to perform visual content understanding and false alarm reasoning on each error sample in the set of error samples to be analyzed, and a single sample diagnosis result is obtained. Cluster analysis was performed on the single-sample diagnostic results, and the sample category with the highest error frequency was counted to obtain the failure sample classification; Based on the classification of the failed samples, the cloud-based large model is used to generate correction suggestions, and the analysis results are generated based on the classification of the failed samples and the correction suggestions.

2. The subway anomaly event identification method based on self-evolution as described in claim 1, characterized in that, The process of converting the natural language requirements into a task blueprint to obtain a user task blueprint includes: Extract the core event phrases from the natural language requirements and perform semantic parsing on the core event phrases to obtain the initial event intent; Based on the initial event intent, semantic completion is performed using a preset common sense base to generate the subject of interest, spatial relationship, action features and potential exclusions related to the initial event intent, thus obtaining supplementary semantic elements; Based on the supplementary semantic elements, a task blueprint is constructed using a preset structured blueprint template to obtain the user task blueprint.

3. The subway anomaly event identification method based on self-evolution as described in claim 1, characterized in that, Based on the user task blueprint, multiple candidate prompt words are generated using a preset cloud-based large model to obtain a candidate prompt word set, including: Extract the subject of interest, spatial relationships, action features, and potential exclusions from the user task blueprint to obtain structured semantic elements; The structured semantic elements are combined with each prompt template in the preset prompt template set to construct multiple meta-prompt words, thus obtaining a meta-prompt word set; The cloud-based big data model is used to generate candidate prompt words corresponding to each meta prompt word in the meta prompt word set, thus obtaining the original candidate prompt word set; The candidate prompt words in the original candidate prompt word set are subjected to format verification and similarity deduplication to obtain the candidate prompt word set.

4. The subway anomaly event identification method based on self-evolution as described in claim 1, characterized in that, The step of loading the prompt word test set using a preset edge mini-model and constructing a prompt word test environment includes: Based on a preset cloud server, a software image instance of the edge mini-model is constructed, and the model weight file of the edge mini-model is loaded to obtain a standard mini-model instance; The prompt word test set is imported into the memory space of the standard small model instance, and the inference parameters of the standard small model instance are configured to obtain the prompt word test environment.

5. The subway anomaly event identification method based on self-evolution as described in claim 1, characterized in that, The step of updating the plurality of candidate prompt words based on the analysis results to obtain an updated prompt word set includes: Based on the classification of failed samples in the analysis results, the corresponding prompt word correction template is selected from the preset optimization strategy library to obtain the target correction template; The multiple candidate prompt words are updated based on the correction suggestions in the analysis results and the target correction template to obtain an updated prompt word set.

6. The subway anomaly event identification method based on self-evolution as described in claim 1, characterized in that, The step of using the prompt word testing environment to test each candidate prompt word in the candidate prompt word set to obtain the evaluation index and prompt word sample set for each candidate prompt word includes: Based on the aforementioned prompt word testing environment, the preset edge small model mirror is used to infer each candidate prompt word in the candidate prompt word set, and the prediction result corresponding to each candidate prompt word is obtained. Obtain the true labels of each candidate prompt word, and compare the prediction results corresponding to each candidate prompt word with the true labels of each candidate prompt word to obtain the comparison results corresponding to multiple candidate prompt words; Based on the comparison results, the evaluation indicators of each candidate prompt word are statistically analyzed, and candidate prompt words whose prediction results are inconsistent with the actual labels are selected from the candidate prompt word set to generate a prompt word sample set.

7. A self-evolving subway anomaly event identification system, characterized in that, The system includes: The prompt word generation module is used to obtain historical subway video streams and natural language requirements input by the target user, and convert the natural language requirements into a task blueprint to obtain the user task blueprint; Based on the user task blueprint, multiple candidate prompt words are generated using a preset cloud-based large model to obtain a candidate prompt word set; The prompt word testing module is used to obtain the prompt word test set corresponding to the user task blueprint based on the historical subway video stream, and to load the prompt word test set using a preset edge mini-model to construct a prompt word testing environment; Using the aforementioned prompt word testing environment, each candidate prompt word in the candidate prompt word set is tested to obtain the evaluation index and prompt word sample set for each candidate prompt word; The prompt word update module is used to determine whether the preset convergence condition is met based on the evaluation index of each candidate prompt word; If the preset convergence condition is not met, the evaluation index of each candidate prompt word and the prompt word sample set are analyzed using the cloud-based big model to obtain the analysis results, and the multiple candidate prompt words are updated according to the analysis results to obtain the updated prompt word set; The step of using the cloud-based big data model to analyze the evaluation metrics of each candidate prompt word and the prompt word sample set to obtain analysis results includes: The evaluation metrics of each candidate prompt word are analyzed to identify the metrics items below a preset threshold, resulting in multiple screening metrics. Prompt word samples corresponding to the multiple screening metrics are extracted from the prompt word sample set to obtain the error sample set to be analyzed. The cloud-based big model is used to perform visual content understanding and false alarm reasoning on each error sample in the set of error samples to be analyzed, and a single sample diagnosis result is obtained. Cluster analysis was performed on the single-sample diagnostic results, and the sample category with the highest error frequency was counted to obtain the failure sample classification; Based on the classification of the failed samples, the cloud-based large model is used to generate correction suggestions, and the analysis results are generated based on the classification of the failed samples and the correction suggestions. An abnormal event identification module is used to take the updated prompt word set as the candidate prompt word set, and return to use the prompt word testing environment to test each candidate prompt word in the candidate prompt word set to obtain the evaluation index and prompt word sample set of each candidate prompt word until the evaluation index meets the preset convergence condition, and the candidate prompt word corresponding to the evaluation index is taken as the optimal prompt word. Based on the optimal prompt words, abnormal events are identified from a preset real-time subway video stream using preset edge devices.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the self-evolutionary subway anomaly event identification method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It includes a data storage area and a program storage area. The data storage area stores the created data, and the program storage area stores the computer program. When the computer program is executed by the processor, it implements the self-evolutionary subway abnormal event identification method as described in any one of claims 1 to 6.

Images