High-speed rail driving monitoring model-oriented single-sample rapid optimization system and method

Abstract

The invention discloses a single-sample rapid optimization system and method for a high-speed rail driving monitoring model, and relates to the technical field of artificial intelligence and rail traffic safety. The system comprises a data acquisition module which acquires multi-source original data in a train operation process in real time; the feature extraction module processes the multi-source original data to generate a continuous feature sequence; the driving behavior monitoring module monitors the driving behavior to obtain a single abnormal sample; when a single abnormal sample is detected, a candidate parameter generation module generates a difference scheme for a model local parameter subset; the time consistency verification module performs consistency verification on the difference scheme; the local freezing type updating module executes the differential scheme, and micro-updating of the driving behavior monitoring module is completed; and the updating management module manages the updating version of the driving behavior monitoring module and records an updating log. According to the method, low-delay, self-adaptive and supervisible rapid optimization can be carried out on a single rare abnormal sample in the driving process, and the abnormal behavior recognition capability is improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and rail transit safety technology, and in particular to a single-sample rapid optimization system and method for high-speed rail driving monitoring models. Background Technology

[0002] The Engineer's Operation Analysis System (EOAS) is a crucial platform in the high-speed rail field for collecting, storing, and analyzing driver operational behavior and train operating status. Its onboard equipment can acquire multiple operational data sources, including ATP, WTD, and CIR signals, and simultaneously record audio and video from the driver's cab and track video, thus providing comprehensive data for operational quality assessment, anomaly review, and safety management. Existing EOAS systems typically rely on onboard recording devices and data transfer cards for data collection and transmission. Most data generated during the driving process must be manually transferred after the train stops before it can be read and used by the ground system.

[0003] While manual data dumping mechanisms ensure data integrity, they also introduce significant timeliness limitations: ground systems cannot obtain the latest driving behavior information in real time during train operation, and model training and analysis can often only be conducted afterward. Furthermore, current driving behavior monitoring models primarily rely on centralized training with static historical data, while factors such as lighting, weather, track structure, and individual driver differences during actual driving can cause model performance to decline to varying degrees as the operating environment changes. Particularly when the number of abnormal driving behavior samples is scarce and their distribution is highly uneven, the model's ability to identify rare abnormal events is often insufficient.

[0004] In high-speed railway operation scenarios, anomalous samples typically occur with low frequency, and many behavioral types even occur only under a very few specific circumstances. Due to the lack of sufficient training instances, existing models struggle to accurately describe these behavioral characteristics. As new driving states or anomalous segments continuously emerge during train operation, if the model cannot utilize these individual, isolated, rare samples for rapid, adaptive incremental optimization, its recognition performance will further decline. Furthermore, due to factors such as vehicle-to-ground communication bandwidth limitations, edge computing power constraints, and high security requirements in the EOAS system, traditional large-scale model retraining or high-latency cloud-based backhaul solutions are insufficient to meet the real-time and reliability requirements of high-speed railways.

[0005] Currently, in the field of driving behavior recognition and monitoring, the main model optimization techniques include the following categories: (1) Data-driven models based on static large-scale offline training. This type of method usually relies on a large amount of historical audio and video and control signal data. After centralized training at the ground center, the trained model is then deployed to the vehicle terminal for operation. This method can achieve good recognition results in common driving scenarios, but once the model is deployed, it is difficult to quickly adjust according to rare behaviors that appear in actual operation. The update cycle is long and the response delay is large. (2) Incremental learning methods based on centralized retraining in the cloud. Some studies have proposed to collect new data in the cloud and retrain the model regularly to improve the model's performance in new scenarios. However, in the existing vehicle-to-ground data interaction mode of EOAS, new data must be obtained through a transfer card after driving ends, which cannot meet the real-time learning needs of abnormal behavior. In addition, cloud retraining usually requires a large number of samples and high computing power, which cannot effectively establish reliable anomaly discrimination capabilities under sparse abnormal sample conditions. (3) Small-scale online update technology based on edge devices. In other industries, some methods have tried to use online learning or fine-tuning technology on edge devices to quickly update models. However, most of these methods require batches of samples or continuous data over a long time window to achieve stable optimization. They are difficult to process a very small number of individual abnormal samples that only appear under specific circumstances in the high-speed railway scenario. At the same time, the limited computing power at the edge makes it difficult to implement complex update mechanisms.

[0006] Therefore, proposing a single-sample fast optimization system and method for high-speed rail driving monitoring models to solve the difficulties of existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a single-sample fast optimization system and method for high-speed rail driving monitoring models, which can perform low-latency, adaptive, and supervised fast optimization of a single rare abnormal sample during train operation, thereby improving the ability to identify abnormal behavior; and solves the problem that existing high-speed rail driving behavior monitoring models are difficult to achieve real-time adaptive updates under conditions of scarce rare abnormal samples, high vehicle-to-ground communication latency, and limited edge computing power.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A single-sample fast optimization system for high-speed rail driving monitoring models includes: The data acquisition module is used to acquire multi-source raw data reflecting the driver's operating status and the train control status in real time during train operation; The feature extraction module, connected to the data acquisition module, is used to perform time alignment and structured processing on multi-source raw data to generate continuous feature sequences. The driving behavior monitoring module is connected to the feature extraction module and has a built-in high-speed rail driving monitoring model to monitor driving behavior and obtain individual abnormal samples. The candidate parameter generation module, connected to the driving behavior monitoring module, is used to generate a candidate parameter update scheme for a local parameter subset of the high-speed rail driving monitoring model when a single abnormal sample is detected. The time consistency verification module, connected to the candidate parameter generation module, is used to perform time consistency verification on the candidate parameter update schemes generated for a local parameter subset of the high-speed rail driving monitoring model based on short-time series behavioral data. The local freeze update module is connected to the time consistency verification module and the driving behavior monitoring module respectively. It is used to execute the candidate parameter update scheme for the local parameter subset of the high-speed rail driving monitoring model that has passed the time consistency verification, and complete the micro update of the driving behavior monitoring module. The update management module, connected to the partial freeze update module, is used to manage the update version of the driving behavior monitoring module, record update logs, and perform rollback operations.

[0009] The above-mentioned system may optionally include a data acquisition module, comprising: a camera, an operation signal acquisition unit, a control variable acquisition unit, an operating parameter acquisition unit, and a driving safety signal acquisition unit; The camera, installed in the driver's cab, captures continuous video footage of the driver's cab to capture the driver's posture, gaze direction, changes in attention, and hand movements; The operation signal acquisition unit collects operation signals related to the driver's control behavior from the control handle, buttons and key control devices, and collects real-time changes in control variables through the vehicle input interface, including traction, braking, mode switching and confirmation operations. The operating parameter acquisition unit collects train operating parameters through the train control system, including: operating speed, acceleration, traction force, and braking force; The train safety signal acquisition unit collects train safety signals, including the train's current safety control status, limitations, and response status, through the output signals of the Automatic Train Protection (ATP) system, the Wireless Transmission Device (WTD), and the Locomotive Integrated Wireless Transmission Device (CIR).

[0010] The above-mentioned system may optionally include a feature extraction module comprising: a preprocessing unit, a visual feature extraction unit, an operation signal feature extraction unit, an operating parameter feature extraction unit, and a driving safety signal feature extraction unit; The preprocessing unit matches the timestamps of video footage, operation signals, changes in control variables, train operation parameters, and train safety signals based on the timestamps, and maps different modal data to a unified time axis through a fixed time step. The visual feature extraction unit preprocesses the video footage from the driver's cab, including: image distortion correction, illumination compensation, face and upper body region localization, hand key point extraction, and gaze direction estimation. The extracted key points, posture vectors, and motion trajectories are then transformed into structured visual features. The operation signal feature extraction unit analyzes the changes in operation signals and control variables related to the driver's control behavior on the control handle, button and key control device, and extracts high-dimensional temporal features of the driver's behavioral intention; The operating parameter feature extraction unit statistically analyzes the train's operating speed, acceleration, traction force, and braking force to extract continuous features of changes in the train's operating status. The driving safety signal feature extraction unit analyzes driving safety signals and extracts key event features such as system alarms, speed limit changes, and control mode switching. All features are normalized, denoised, and temporally smoothed, and then combined into a multimodal continuous feature sequence according to a fixed time window.

[0011] The above-mentioned system may optionally include a driving behavior monitoring module, which has a built-in high-speed rail driving monitoring model. Based on the multimodal continuous feature sequence output by the feature extraction module, it classifies driving behavior and determines anomalies, identifying individual abnormal samples.

[0012] The above system, optionally, includes a candidate parameter generation module that analyzes the forward output and backward gradient information of the high-speed rail driving monitoring model on a single abnormal sample to identify sensitive channels, key convolutional kernels, and high-response weight regions, and selects the parameter subset most relevant to the current abnormal behavior as the region to be updated. Candidate parameter update schemes with different optimization strengths and parameter influence ranges are generated by using micro-step gradient perturbation, feature attention scaling, and local layer weight redistribution. Each scheme corresponds to a set of local parameter differences. Real-time simulation tests were conducted on each candidate parameter update scheme, and a multi-dimensional evaluation index system was used to comprehensively score the candidate schemes in terms of classification accuracy, inference speed and resource consumption. Feasible candidate parameter update schemes under edge computing power constraints are selected, and the selected candidate parameter update schemes are sorted from high to low according to the comprehensive score and packaged and output.

[0013] The above system, optionally, includes a time consistency verification module that performs time-series consistency verification on multiple sets of candidate parameter update schemes packaged and output by the candidate parameter generation module. Specifically: Based on recent continuous window driving behavior data, short-term behavior recognition inference is re-executed in a simulated update environment. Dynamic time warping technology and hidden Markov model are used to compare and analyze the changes in classification sequence before and after the update, detect whether there are abnormal jumps in behavior categories, whether the continuity of recognition segments is damaged, and the degree of matching with the driver's actual operation rhythm. The module has a built-in professional knowledge rule base, which contains normal behavior patterns under typical high-speed rail driving scenarios. By analyzing the continuity of driving behavior, operational stability and state consistency in a short time series after a single abnormal sample, the effectiveness of the candidate parameter update scheme is dynamically verified. Candidate parameter update schemes that fail the timing consistency verification are directly eliminated, while candidate parameter update schemes that pass the verification are prioritized and output according to their timing preservation capabilities.

[0014] The above system, optionally, includes a partial freeze update module that employs a partial freeze micro-update strategy. This strategy selects the optimal solution from a set of validated candidate parameter update schemes and executes a controlled partial update. Specifically: First, by using hierarchical sensitivity analysis, the correlation between parameters of each layer of the model and their weights in terms of overall performance are identified. The core structure of the high-speed rail driving monitoring model is set to a frozen state, and only the region related to the current single abnormal sample is opened as the set of parameters that can be updated. The update process adopts a combination of small step weight adjustment, gradient pruning, and parameter smoothing mechanism. Parameter differences are injected into the local regions that have passed the validation, and the performance changes of the high-speed rail driving monitoring model on the validation set are monitored in real time. Record the parameter difference components, affected area range, and performance change curve for each update.

[0015] The above system optionally includes an update management module that receives the parameter difference components output by the partial freeze update module, performs integrity verification, version identifier generation, and update process recording on the updated model, and archives the model state before and after optimization in a structured manner. The update management module has a built-in automatic rollback mechanism. When the system detects a sharp increase in false alarms, abnormal behavior sequences, or decreased stability in the updated high-speed rail driving monitoring model during subsequent real-time identification, it will restore to the previous stable version based on the recorded difference components. The update management module is responsible for uploading the current update results, security verification indicators, and anonymized features of abnormal samples to the cloud.

[0016] A single-sample fast optimization method for high-speed rail driving monitoring models, applied to the single-sample fast optimization system for high-speed rail driving monitoring models described in any of the above-mentioned embodiments, includes: S1. Real-time acquisition and processing of multi-source raw data reflecting the driver's operating status and train control status during high-speed rail operation, generating multimodal continuous feature sequences; S2. Based on the multimodal continuous feature sequence, classify and determine the abnormality of driving behavior. When a single abnormal sample is detected, analyze the forward output and backward gradient information of the high-speed rail driving monitoring model on the single abnormal sample, and generate and screen candidate parameter update schemes. S3. Perform time-series consistency verification on the selected candidate parameter update schemes and output the candidate parameter update schemes that pass the verification. S4. Set the core structure of the high-speed rail driving monitoring model to a frozen state, and only open the region related to the current single abnormal sample as the updatable parameter set. Inject the local parameter difference of the verified candidate parameter update scheme into the updatable parameter set to complete the rapid adaptive optimization of the model. S5. Perform integrity verification, version identifier generation, and update process recording on the updated model. The model status before and after optimization is archived in a structured manner. It also supports version rollback when a performance degradation is detected. Based on the recorded difference, it restores to the previous stable version and uploads the current update results, safety verification indicators, and anonymized features of abnormal samples to the cloud.

[0017] As can be seen from the above technical solution, compared with the prior art, the present invention provides a single-sample fast optimization system and method for high-speed rail driving monitoring models, which has the following beneficial effects: This invention addresses the challenge of abnormal behaviors occurring in large batches in high-speed railway scenarios. Instead of relying on grouped samples or long windows, it generates local parameter differences based on single, real-time acquired samples. It then selects the subset of parameters most relevant to the abnormal behavior from the model structure as the update region, avoiding the high computational overhead and instability caused by full model updates. Furthermore, this invention dynamically verifies the effectiveness of candidate parameter updates by analyzing the continuity of driving behavior, operational stability, and state consistency within a short time series following the new sample. Model optimization is only allowed when consistency is achieved, thus improving the safety and accuracy of updates. Finally, this invention freezes most of the model's core structure, injecting parameters only into verified local regions. Differential optimization enables the model to perform rapid optimization with extremely low computational cost, adapting to the real-world behavior characteristics and environmental conditions of the current driver, and is suitable for real-time operation on edge devices under constrained conditions. This invention records the time, triggering sample, and update magnitude of each parameter update and supports rollback operations, ensuring that the model update behavior is controllable and traceable, and that model performance will not be degraded due to erroneous samples or noise. Without changing the data acquisition architecture of the EOAS system, this invention enables the driving behavior monitoring model to perform low-latency, adaptive, and supervised rapid optimization of individual rare abnormal samples during driving, thereby improving the ability to identify abnormal behavior, reducing the model's dependence on static training, and enhancing stability and real-time performance in complex operating environments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This invention provides a structural diagram of a single-sample fast optimization system for high-speed rail driving monitoring models. Figure 2 The flowchart of a single-sample fast optimization method for a high-speed rail driving monitoring model provided by the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0022] Reference Figure 1 As shown, this invention discloses a single-sample fast optimization system for high-speed rail driving monitoring models, comprising: The data acquisition module is used to acquire multi-source raw data reflecting the driver's operating status and the train control status in real time during train operation; The feature extraction module, connected to the data acquisition module, is used to perform time alignment and structured processing on multi-source raw data to generate continuous feature sequences. The driving behavior monitoring module is connected to the feature extraction module and has a built-in high-speed rail driving monitoring model to monitor driving behavior and obtain individual abnormal samples. The candidate parameter generation module, connected to the driving behavior monitoring module, is used to generate a candidate parameter update scheme for a local parameter subset of the high-speed rail driving monitoring model when a single abnormal sample is detected. The time consistency verification module, connected to the candidate parameter generation module, is used to perform time consistency verification on the candidate parameter update schemes generated for a local parameter subset of the high-speed rail driving monitoring model based on short-time series behavioral data. The local freeze update module is connected to the time consistency verification module and the driving behavior monitoring module respectively. It is used to execute the candidate parameter update scheme for the local parameter subset of the high-speed rail driving monitoring model that has passed the time consistency verification, and complete the micro update of the driving behavior monitoring module. The update management module, connected to the partial freeze update module, is used to manage the update version of the driving behavior monitoring module, record update logs, and perform rollback operations.

[0023] Furthermore, the data acquisition module includes: a camera, an operation signal acquisition unit, a control variable acquisition unit, an operating parameter acquisition unit, and a driving safety signal acquisition unit; The camera, installed in the driver's cab, captures continuous video footage of the driver's cab to capture the driver's posture, gaze direction, changes in attention, and hand movements; The operation signal acquisition unit collects operation signals related to the driver's control behavior from the control handle, buttons and key control devices, and collects real-time changes in control variables through the vehicle input interface, including traction, braking, mode switching and confirmation operations. The operating parameter acquisition unit collects train operating parameters through the train control system, including: operating speed, acceleration, traction force, and braking force; The train safety signal acquisition unit collects train safety signals, including the train's current safety control status, limitations, and response status, through the output signals of the Automatic Train Protection (ATP) system, the Wireless Transmission Device (WTD), and the Locomotive Integrated Wireless Transmission Device (CIR).

[0024] The output signals of key driving safety systems such as ATP, WTD, and CIR are transmitted in real time through a dedicated communication link to reflect the current safety control status, constraints, and system response. During the data collection process, the vehicle-mounted host performs time reference alignment to ensure that visual signals, operation signals, running status signals and safety control signals are strictly synchronized. Then, the data is packaged in a unified format and sent to the feature extraction module to provide complete and time-consistent input information for the behavior recognition model.

[0025] Furthermore, the feature extraction module includes: a preprocessing unit, a visual feature extraction unit, an operation signal feature extraction unit, an operating parameter feature extraction unit, and a driving safety signal feature extraction unit; The preprocessing unit matches the timestamps of video footage, operation signals, changes in control variables, train operation parameters, and train safety signals based on the timestamps. By mapping different modal data to a unified time axis through a fixed time step, it ensures the semantic consistency of multi-source data at the same time. The visual feature extraction unit preprocesses the video footage from the driver's cab, including: image distortion correction, illumination compensation, face and upper body region localization, hand key point extraction, and gaze direction estimation. The extracted key points, posture vectors, and motion trajectories are then transformed into structured visual features. The operation signal feature extraction unit analyzes the changes in operation signals and control variables related to the driver's control behavior on the control handle, button and key control device, and extracts high-dimensional temporal features of the driver's behavioral intention; The operating parameter feature extraction unit statistically analyzes the train's operating speed, acceleration, traction force, and braking force to extract continuous features of changes in the train's operating status. The driving safety signal feature extraction unit analyzes driving safety signals and extracts key event features such as system alarms, speed limit changes, and control mode switching. All features are normalized, denoised, and temporally smoothed, and then combined into a multimodal continuous feature sequence according to a fixed time window. Finally, the sequence is sent to the driving behavior monitoring model in a unified format for behavior classification, anomaly detection, and subsequent single-sample rapid optimization.

[0026] Furthermore, the driving behavior monitoring module has a built-in high-speed rail driving monitoring model. Based on the multimodal continuous feature sequence output by the feature extraction module, it classifies driving behavior and determines anomalies, identifying individual abnormal samples.

[0027] Furthermore, the candidate parameter generation module automatically constructs multiple alternative local optimization parameter sets based on the internal feature response and gradient-sensitive region of the current base model when the system detects a new single abnormal sample, thereby realizing a single-sample triggered local candidate parameter update mechanism. By analyzing the forward output and backward gradient information of the high-speed rail driving monitoring model on a single abnormal sample, sensitive channels, key convolutional kernels and high response weight regions are identified, and the parameter subset most relevant to the current abnormal behavior is selected as the region to be updated. We use micro-step gradient perturbation, feature attention scaling, and local layer weight redistribution to generate candidate parameter update schemes with different optimization strengths and parameter influence ranges. Each scheme corresponds to a set of local parameter differences, avoiding the high computational cost and instability caused by full model update. Real-time simulation tests are performed on each candidate parameter update scheme. A multi-dimensional evaluation index system is used to comprehensively score the candidate schemes in terms of classification accuracy, inference speed and resource consumption, ensuring that feasible parameter optimization schemes can still be generated efficiently under the condition of limited edge computing power. Feasible candidate parameter update schemes under edge computing power constraints are selected. The selected candidate parameter update schemes are sorted from high to low according to the comprehensive score and packaged and output to the time consistency verification module, providing multiple optional optimization paths for subsequent model security updates.

[0028] Furthermore, the time consistency verification module is used to construct a time consistency verification mechanism to screen reliable samples. It performs time consistency verification on multiple sets of candidate parameter update schemes packaged and output by the candidate parameter generation module. Specifically: Based on recent continuous window driving behavior data, short-term behavior recognition inference is re-executed in a simulated update environment. Dynamic time warping technology and hidden Markov model are used to compare and analyze the changes in classification sequence before and after the update. The focus is on detecting whether there are abnormal jumps in behavior categories, whether the continuity of recognition segments is impaired, and the degree of matching with the driver's actual operating rhythm. The module has a built-in professional knowledge rule base, which contains normal behavior patterns under typical high-speed rail driving scenarios. By analyzing the continuity of driving behavior, operational stability and state consistency in a short time series after a single abnormal sample, the effectiveness of the candidate parameter update scheme is dynamically verified. Candidate parameter update schemes that fail the temporal consistency verification are directly eliminated. Verified candidate parameter update schemes are prioritized according to their temporal preservation capabilities and output to the local freeze update module. This ensures that model optimization does not disrupt the reasonable rhythmic characteristics of high-speed rail driving behavior, thereby improving the safety and accuracy of the update. Furthermore, the partial freeze update module employs a partial freeze micro-update strategy, selecting the optimal solution from the validated set of candidate parameter update schemes to perform a controlled partial update, specifically: First, by using hierarchical sensitivity analysis, the correlation between parameters of each layer of the model and their weights in terms of overall performance are identified. The core structure of the high-speed rail driving monitoring model is set to a frozen state, and only the region related to the current single abnormal sample is opened as the set of parameters that can be updated. The update process employs a combination of small-step weight adjustment, gradient pruning, and parameter smoothing mechanisms. Parameter differences are injected into the validated local regions to monitor the performance changes of the high-speed rail driving monitoring model on the validation set in real time, ensuring that the update's impact is controllable and does not impair the ability to identify regular samples. This update method enables the model to be rapidly optimized with extremely low computational cost, adapting to the real behavioral characteristics and environmental conditions of current drivers. The module also records the parameter difference components, the range of the affected area, and the performance change curve for each update, providing basic information for the subsequent update management module to perform model backtracking, thus achieving rapid adaptation to abnormal sample characteristics while maintaining the overall stability of the model.

[0029] Furthermore, the update management module is used to manage the security, version control, rollback mechanism and cloud synchronization during the model update process; it receives the parameter difference components output by the partial freeze update module, performs integrity verification, version identifier generation and update process recording on the updated model, and archives the model status before and after optimization in a structured manner. The update management module has a built-in automatic rollback mechanism. When the system detects a sharp increase in false alarms, abnormal behavior sequences, or decreased stability in the updated high-speed rail driving monitoring model during subsequent real-time identification, it will restore to the previous stable version based on the recorded difference components, ensuring that the high-speed rail driving monitoring can maintain continuous and reliable operation even when potential risks occur. The update management module is responsible for uploading the current update results, safety verification indicators, and anonymized features of abnormal samples to the cloud, providing basic support for global model integration and long-term training. Through the update management module, the entire system achieves end-to-cloud collaborative model lifecycle management, making the rapid optimization process of single samples highly controllable, traceable, and secure.

[0030] Reference Figure 2 As shown, a single-sample fast optimization method for high-speed rail driving monitoring models is applied to the single-sample fast optimization system for high-speed rail driving monitoring models described above, comprising: S1. Real-time acquisition and processing of multi-source raw data reflecting the driver's operating status and train control status during high-speed rail operation, generating multimodal continuous feature sequences; S2. Based on the multimodal continuous feature sequence, classify and determine the abnormality of driving behavior. When a single abnormal sample is detected, analyze the forward output and backward gradient information of the high-speed rail driving monitoring model on the single abnormal sample, and generate and screen candidate parameter update schemes. S3. Perform time-series consistency verification on the selected candidate parameter update schemes and output the candidate parameter update schemes that pass the verification. S4. Set the core structure of the high-speed rail driving monitoring model to a frozen state, and only open the region related to the current single abnormal sample as the updatable parameter set. Inject the local parameter difference of the verified candidate parameter update scheme into the updatable parameter set to complete the rapid adaptive optimization of the model. S5. Perform integrity verification, version identifier generation, and update process recording on the updated model. The model status before and after optimization is archived in a structured manner. It also supports version rollback when a performance degradation is detected. Based on the recorded difference, it restores to the previous stable version and uploads the current update results, safety verification indicators, and anonymized features of abnormal samples to the cloud.

[0031] In one specific embodiment, a single-sample fast optimization system for a high-speed rail driving monitoring model is deployed in an embedded inference device on the high-speed rail train. This device has approximately 8 TOPS of computing power and acquires multi-source real-time data through the train network bus. The system samples multimodal raw data every 50 ms, including driver's cab video, handle position signals, button operation signals, train speed, traction force, braking force, and outputs from the train safety system. The collected data is synchronized using a unified time base and a sliding window mechanism is used to generate a continuous feature sequence of 1 second (20 frames). The obtained continuous feature sequence is input into a built-in lightweight Transformer-LSTM hybrid structure model for driving behavior classification. When a single abnormal sample is detected during a run (such as the handle angle instantly changing from 0 to braking 4), the system calculates its forward output and calculates the gradient while keeping the backbone network frozen. The system identifies the subset of local parameters most relevant to the abnormal behavior based on the gradient-sensitive region. Based on the extracted gradient information, three types of candidate parameter update schemes are generated: 1. A low-rank gradient compression scheme with rank=4; 2. A scheme that selects the top 1% based on the Fisher information matrix. The key parameters are micro-updated; 3. Local reparameterization based on feature similarity; Subsequently, the cached features of the most recent 4 seconds (80 frames) are used to perform short-term consistency verification on each scheme. If the consistency verification is passed, it is considered that the scheme will not disrupt the continuity of the driving behavior sequence. In this embodiment, the low-rank compression scheme has the highest score, followed by the feature similarity reparameterization scheme, while the Fisher information scheme is eliminated due to excessive perturbation; The verified candidate schemes are sorted from high to low according to the comprehensive score, that is, the first is the low-rank compression scheme, the second is the feature similarity reparameterization scheme, and are output to the local frozen update module to provide a priority selection path for subsequent fast updates on the end side; In this embodiment, the model backbone (the first 6 layers of Transformer) is kept completely frozen through hierarchical sensitivity analysis, and only the local FFN layer and attention substructure parameters (about 1200, with trainable parameters accounting for <1%) most relevant to the current abnormal sample are opened; Finally, the model update adopts the parameter differential injection strategy: the optimal scheme after sorting is the low-rank compression scheme, and the entire update process takes about 38 seconds. The update can be completed in real time during train operation. After the update is completed, the system generates a model version number (e.g., "v2.3.17-edge-20251211-4aacde") and records the parameter difference size, anonymized anomalous sample summary (PCA compressed to 32 dimensions), and various validation metrics. If the false alarm rate of the model is subsequently detected to increase by more than 3%, the edge will roll back to the previous stable version (e.g., "v2.3.16-edge") based on the archived difference data. At the same time, the results of this difference update, safety validation metrics, and anomalous sample characteristics will be uploaded to the cloud to provide a basis for global model fusion and long-term training.

[0032] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units 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. Those skilled in the art can understand and implement this without creative effort.

[0033] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A single-sample fast optimization system for high-speed rail driving monitoring models, characterized in that, include: The data acquisition module is used to acquire multi-source raw data reflecting the driver's operating status and the train control status in real time during train operation; The feature extraction module, connected to the data acquisition module, is used to perform time alignment and structured processing on multi-source raw data to generate continuous feature sequences. The driving behavior monitoring module is connected to the feature extraction module and has a built-in high-speed rail driving monitoring model to monitor driving behavior and obtain individual abnormal samples. The candidate parameter generation module, connected to the driving behavior monitoring module, is used to generate a candidate parameter update scheme for a local parameter subset of the high-speed rail driving monitoring model when a single abnormal sample is detected. The time consistency verification module, connected to the candidate parameter generation module, is used to perform time consistency verification on the candidate parameter update schemes generated for a local parameter subset of the high-speed rail driving monitoring model based on short-time series behavioral data. The local freeze update module is connected to the time consistency verification module and the driving behavior monitoring module respectively. It is used to execute the candidate parameter update scheme for the local parameter subset of the high-speed rail driving monitoring model that has passed the time consistency verification, and complete the micro update of the driving behavior monitoring module. The update management module, connected to the partial freeze update module, is used to manage the update version of the driving behavior monitoring module, record update logs, and perform rollback operations.

2. The single-sample fast optimization system for high-speed rail driving monitoring models according to claim 1, characterized in that, The data acquisition module includes: a camera, an operation signal acquisition unit, a control variable acquisition unit, an operating parameter acquisition unit, and a driving safety signal acquisition unit; The camera, installed in the driver's cab, captures continuous video footage of the driver's cab to capture the driver's posture, gaze direction, changes in attention, and hand movements; The operation signal acquisition unit collects operation signals related to the driver's control behavior from the control handle, buttons and key control devices, and collects real-time changes in control variables through the vehicle input interface, including traction, braking, mode switching and confirmation operations. The operating parameter acquisition unit collects train operating parameters through the train control system, including: operating speed, acceleration, traction force, and braking force; The train safety signal acquisition unit collects train safety signals, including the train's current safety control status, limitations, and response status, through the output signals of the Automatic Train Protection (ATP) system, the Wireless Transmission Device (WTD), and the Locomotive Integrated Wireless Transmission Device (CIR).

3. The single-sample fast optimization system for high-speed rail driving monitoring models according to claim 2, characterized in that, The feature extraction module includes: a preprocessing unit, a visual feature extraction unit, an operation signal feature extraction unit, an operating parameter feature extraction unit, and a driving safety signal feature extraction unit; The preprocessing unit matches the timestamps of video footage, operation signals, changes in control variables, train operation parameters, and train safety signals based on the timestamps, and maps different modal data to a unified time axis through a fixed time step. The visual feature extraction unit preprocesses the video footage from the driver's cab, including: image distortion correction, illumination compensation, face and upper body region localization, hand key point extraction, and gaze direction estimation. The extracted key points, posture vectors, and motion trajectories are then transformed into structured visual features. The operation signal feature extraction unit analyzes the changes in operation signals and control variables related to the driver's control behavior on the control handle, button and key control device, and extracts high-dimensional temporal features of the driver's behavioral intention; The operating parameter feature extraction unit statistically analyzes the train's operating speed, acceleration, traction force, and braking force to extract continuous features of changes in the train's operating status. The driving safety signal feature extraction unit analyzes driving safety signals and extracts key event features such as system alarms, speed limit changes, and control mode switching. All features are normalized, denoised, and temporally smoothed, and then combined into a multimodal continuous feature sequence according to a fixed time window.

4. The single-sample fast optimization system for high-speed rail driving monitoring models according to claim 3, characterized in that, The driving behavior monitoring module has a built-in high-speed rail driving monitoring model. Based on the multimodal continuous feature sequence output by the feature extraction module, it classifies driving behavior and determines anomalies, identifying individual abnormal samples.

5. A single-sample fast optimization system for high-speed rail driving monitoring models according to claim 4, characterized in that, The candidate parameter generation module analyzes the forward output and backward gradient information of the high-speed rail driving monitoring model on a single abnormal sample to identify sensitive channels, key convolutional kernels and high response weight regions, and selects the parameter subset most relevant to the current abnormal behavior as the region to be updated. Candidate parameter update schemes with different optimization strengths and parameter influence ranges are generated by using micro-step gradient perturbation, feature attention scaling, and local layer weight redistribution. Each scheme corresponds to a set of local parameter differences. Real-time simulation tests were conducted on each candidate parameter update scheme, and a multi-dimensional evaluation index system was used to comprehensively score the candidate schemes in terms of classification accuracy, inference speed and resource consumption. Feasible candidate parameter update schemes under edge computing power constraints are selected, and the selected candidate parameter update schemes are sorted from high to low according to the comprehensive score and packaged and output.

6. A single-sample fast optimization system for high-speed rail driving monitoring models according to claim 5, characterized in that, The time consistency verification module performs time-series consistency verification on multiple sets of candidate parameter update schemes packaged and output by the candidate parameter generation module. Specifically: Based on recent continuous window driving behavior data, short-term behavior recognition inference is re-executed in a simulated update environment. Dynamic time warping technology and hidden Markov model are used to compare and analyze the changes in classification sequence before and after the update, detect whether there are abnormal jumps in behavior categories, whether the continuity of recognition segments is damaged, and the degree of matching with the driver's actual operation rhythm. The module has a built-in professional knowledge rule base, which contains normal behavior patterns under typical high-speed rail driving scenarios. By analyzing the continuity of driving behavior, operational stability and state consistency in a short time series after a single abnormal sample, the effectiveness of the candidate parameter update scheme is dynamically verified. Candidate parameter update schemes that fail the timing consistency verification are directly eliminated, while candidate parameter update schemes that pass the verification are prioritized and output according to their timing preservation capabilities.

7. A single-sample fast optimization system for high-speed rail driving monitoring models according to claim 6, characterized in that, The partial freeze update module employs a partial freeze micro-update strategy, selecting the optimal solution from a set of validated candidate parameter update schemes to perform a controlled partial update. Specifically: First, by using hierarchical sensitivity analysis, the correlation between parameters of each layer of the model and their weights in terms of overall performance are identified. The core structure of the high-speed rail driving monitoring model is set to a frozen state, and only the region related to the current single abnormal sample is opened as the set of parameters that can be updated. The update process adopts a combination of small step weight adjustment, gradient pruning, and parameter smoothing mechanism. Parameter differences are injected into the local regions that have passed the validation, and the performance changes of the high-speed rail driving monitoring model on the validation set are monitored in real time. Record the parameter difference components, affected area range, and performance change curve for each update.

8. A single-sample fast optimization system for high-speed rail driving monitoring models according to claim 7, characterized in that, The update management module receives the parameter difference components output by the partial freeze update module, performs integrity verification, version identifier generation, and update process recording on the updated model, and archives the model status before and after optimization in a structured manner. The update management module has a built-in automatic rollback mechanism. When the system detects a sharp increase in false alarms, abnormal behavior sequences, or decreased stability in the updated high-speed rail driving monitoring model during subsequent real-time identification, it will restore to the previous stable version based on the recorded difference components. The update management module is responsible for uploading the current update results, security verification indicators, and anonymized features of abnormal samples to the cloud.

9. A single-sample fast optimization method for a high-speed rail driving monitoring model, applied to the single-sample fast optimization system for a high-speed rail driving monitoring model as described in any one of claims 1-8, comprising: S1. Real-time acquisition and processing of multi-source raw data reflecting the driver's operating status and train control status during high-speed rail operation, generating multimodal continuous feature sequences; S2. Based on the multimodal continuous feature sequence, classify and determine the abnormality of driving behavior. When a single abnormal sample is detected, analyze the forward output and backward gradient information of the high-speed rail driving monitoring model on the single abnormal sample, and generate and screen candidate parameter update schemes. S3. Perform time-series consistency verification on the selected candidate parameter update schemes and output the candidate parameter update schemes that pass the verification. S4. Set the core structure of the high-speed rail driving monitoring model to a frozen state, and only open the region related to the current single abnormal sample as the updatable parameter set. Inject the local parameter difference of the verified candidate parameter update scheme into the updatable parameter set to complete the rapid adaptive optimization of the model. S5. Perform integrity verification, version identifier generation, and update process recording on the updated model. The model status before and after optimization is archived in a structured manner. It also supports version rollback when a performance degradation is detected. Based on the recorded difference, it restores to the previous stable version and uploads the current update results, safety verification indicators, and anonymized features of abnormal samples to the cloud.

Images