A method and system for predicting autonomous driving takeover time based on physical-driver-system three-factor assessment and gating selection.

Abstract

This invention provides a method and system for predicting takeover time in autonomous driving based on a three-factor assessment and gating selection of physical-driver-system elements, relating to the field of autonomous driving system technology. The method includes: collecting synchronous time-series data of the physical-driver-system elements within a historical time window; extracting features from the collected synchronous time-series data using expert modules; globally encoding the extracted features and performing gating decisions; activating the takeover timing prediction path corresponding to the feature vector with the largest weight value based on the output gating decision signal, and outputting the final takeover time. This invention constructs three independent expert modules to conduct in-depth analysis of physical interaction risks, driver readiness, and system reliability, and introduces a gating selection mechanism to explicitly determine the most urgent limiting factor, directly using its corresponding time estimate as the final output, thus achieving accurate, safe, and interpretable takeover time prediction.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving system technology, and in particular to a method and system for predicting the time required for driver takeover after an autonomous driving system issues a takeover request. Background Technology

[0002] Autonomous driving technology is gradually evolving from conditional automation to high automation. In Level 3 autonomous driving as defined by the SAE standard, the system needs to issue a takeover request to the driver when it is unable to handle the driving conditions. In this scenario, the accuracy of the system's prediction of the takeover timing directly affects the safety and comfort of practical applications: too early a takeover request will reduce the user experience, while too late a request may lead to safety hazards.

[0003] Existing technologies primarily predict takeover times by constructing mathematical models, but this approach has significant drawbacks. Firstly, existing methods often employ a single information source (e.g., considering only the degree of vehicle lane departure) or a simple multi-factor weighted fusion strategy, making it difficult to effectively handle dynamic information across different dimensions (the coupling relationship between environmental risk level and driver reaction time changes non-linearly over time). Secondly, the decision-making process of these methods lacks interpretability; once the system outputs the takeover time, maintenance personnel cannot determine whether the result stems from a sudden road obstacle or driver inattention, which complicates accident accountability and system improvement. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides an autonomous driving takeover time prediction method and system based on physical-driver-system three-factor assessment and gating selection. This method constructs three independent expert modules to conduct in-depth analysis of physical interaction risks, driver readiness, and system reliability, and innovatively introduces a gating selection mechanism to explicitly determine the most urgent limiting factor and directly use its corresponding time estimate as the final output, thereby achieving accurate, safe, and interpretable takeover time prediction.

[0005] Therefore, the present invention provides the following technical solution:

[0006] On the one hand, this invention provides a method for predicting autonomous driving takeover time based on physical-driver-system three-factor evaluation and gating selection, including: Collect synchronous time-series data of three elements within a historical time window; the three elements include: physical interaction element, driver readiness element, and system reliability element. Three expert modules are used to extract features from the collected synchronous time-series data of the three elements. These expert modules include: a physical interaction situation analysis module, a driver readiness analysis module, and a system reliability degradation analysis module. The physical interaction situation analysis module receives physical interaction element data, extracts primary feature vectors from environmental perception data and vehicle state data, and uses a cross-attention mechanism to deeply fuse these two sets of primary features, outputting a feature vector representing the dynamic evolution of the comprehensive safety boundary at the physical level. The driver readiness analysis module analyzes the driver's attention, behavior, and physiological signals, predicts the driver's recovery process from the current state to a takeover-ready state, and outputs a feature vector describing the evolution of driver readiness. The system reliability degradation analysis module, based on a degradation model and a failure propagation model, predicts the remaining system reliability time and outputs a feature vector describing the evolution of the system's health state. The extracted features are globally encoded, and gating decisions are made based on the global encoding results; Based on the output gating decision signal, the takeover timing prediction path corresponding to the feature vector with the largest weight value is activated, and the final takeover time is output; the takeover timing prediction path adopts a multilayer perceptron.

[0007] Furthermore, the physical interaction element data includes: environmental perception data collected by environmental perception sensors and vehicle status data collected by vehicle bus or inertial measurement unit; The driver readiness data is driver status data collected by the driver monitoring system, including the driver's attention signals, behavioral signals, and physiological signals. The system reliability data is the health and performance data of each system component collected by the self-monitoring unit of the autonomous driving system, including: sensor status, positioning information, and computing unit status.

[0008] Furthermore, the environmental perception data includes: a sequence of images from a forward-looking camera, a sequence of target lists from a forward-looking millimeter-wave radar, and a sequence of point clouds from a solid-state lidar; the vehicle status includes: the vehicle's longitudinal speed, longitudinal / lateral acceleration, yaw rate, steering wheel angle, wheel speeds, and brake master cylinder pressure; The driver readiness data includes: driver eyelid opening percentage, gaze point screen coordinates, head posture angle, and hand grip strength distribution; The system reliability data includes: camera exposure anomaly indicator, radar signal-to-noise ratio, average lidar echo intensity; GNSS positioning mode, horizontal protection level, inertial navigation system position drift rate; CPU / GPU utilization of the autonomous driving domain controller, core temperature, and execution cycle jitter of key perception and planning threads.

[0009] Furthermore, the physical interaction situation analysis module includes: an environment coding sub-network, a vehicle coding sub-network, and a bidirectional interactive attention unit; wherein, the environment coding sub-network includes an image branch, a point cloud branch, and a radar branch, and the image branch adopts a lightweight SlowOnly 3D... The CNN extracts spatiotemporal features from video clips and outputs image feature vectors. The point cloud branch uses a PointPillar network to process the point cloud sequence, extracting features of dynamic targets and static structures, and outputting point cloud feature vectors. The radar branch normalizes the radar target sequence and extracts motion pattern features through an LSTM network, outputting radar feature vectors. The features output from the three branches are concatenated and passed through a fully connected layer to obtain the primary environmental feature encoding. The vehicle encoding sub-network standardizes the vehicle state time series, inputs it into a one-dimensional temporal convolutional network, and outputs primary vehicle feature encoding. The bidirectional interactive attention unit includes a first interactive layer representation, a second interactive layer representation, and a multilayer perceptron structure. The first interactive layer representation identifies which information in the environment is most critical under vehicle capability limitations. The second interactive layer representation identifies which vehicle capabilities are most important under current environmental risks and challenges. Finally, the first and second interactive layer representations are concatenated and output as a first feature vector through the multilayer perceptron structure.

[0010] Furthermore, the driver readiness analysis module takes driver state time series data as input, uses a bidirectional gated recurrent unit to learn the temporal dependencies of driver state, and concatenates the forward and backward hidden states of the last time step, then passes them through a fully connected layer to obtain the second feature vector.

[0011] Furthermore, the construction of the system reliability degradation analysis module includes: establishing a degradation model based on a stochastic process for different key performance indicators; updating the degradation model parameters online based on real-time collected system reliability element data, and predicting the remaining time before each indicator first exceeds the preset failure threshold; and generating a reliability degradation prediction by combining the prediction results of each indicator, and encoding it as a third feature vector.

[0012] Furthermore, a degradation model based on stochastic processes is established, including: establishing a Wiener process degradation model with a drift term for each key indicator: ; in, These are performance index values. The drift coefficient, The diffusion coefficient is... For the standard Wiener process, This represents the initial state value of the performance metric.

[0013] Furthermore, the degradation model parameters are updated online, including: using a particle filter algorithm to update the parameters of each degradation model online based on real-time observation data. , and the current status .

[0014] Furthermore, the extracted features are globally encoded, and gating decisions are made based on the global encoding results, including: The three extracted time-series feature vector sequences are aligned and concatenated at time steps to form a multi-factor joint feature sequence; The multi-factor joint feature sequence is input into a temporal context modeling network to generate a global security situation encoding vector; the temporal context modeling network is a transformer encoder. The global security situation encoding vector is input into the gating decision network; based on the current comprehensive situation, the gating decision network determines which of the three safety elements—physical, driver, and system—constitutes the most urgent takeover constraint, and outputs a three-element gating decision signal to indicate which element's takeover timing prediction path should be activated; the gating decision network is composed of a multi-layer sensing mechanism.

[0015] In another aspect, the present invention also provides an autonomous driving takeover time prediction system based on physical-driver-system three-factor evaluation and gating selection, comprising: The information acquisition module collects synchronous time-series data of three elements within a historical time window; the three elements include: physical interaction element, driver readiness element, and system reliability element. The system employs three expert modules for feature extraction, each dedicated to extracting features from the collected synchronous time-series data of the three key elements. These modules include: a physical interaction situation analysis module, a driver readiness analysis module, and a system reliability degradation analysis module. The physical interaction situation analysis module receives physical interaction element data, extracts primary feature vectors from environmental perception data and vehicle state data, and uses a cross-attention mechanism to deeply fuse these two sets of primary features, outputting a feature vector representing the dynamic evolution of the comprehensive safety boundary at the physical level. The driver readiness analysis module analyzes the driver's attention, behavior, and physiological signals to predict the driver's recovery process from the current state to a takeover-ready state, outputting a feature vector describing the evolution of driver readiness. The system reliability degradation analysis module, based on degradation and failure propagation models, predicts the remaining system reliability time and outputs a feature vector describing the evolution of the system's health state. The global situation analysis and gating decision module performs global encoding on the extracted features and makes gating decisions based on the global encoding results. The takeover time prediction module activates the takeover timing prediction path corresponding to the feature vector with the largest weight value based on the output gating decision signal, and outputs the final takeover time; the takeover timing prediction path adopts a multilayer perceptron.

[0016] Advantages and positive effects of the present invention: 1. This invention processes heterogeneous information such as physical environment, vehicle status, driver behavior and system health through independent expert modules, and performs dynamic interactive fusion at the decision-making level, realizing deep fusion of multi-source information. It overcomes the limitations of traditional methods that simply splice or fuse with fixed weights, and significantly improves the prediction accuracy in complex scenarios.

[0017] 2. This invention introduces a gating selection decision mechanism to dynamically focus on key constraints. It can identify the most pressing safety constraints (such as sudden collision risk, severe driver distraction, or sudden system failure) in real time based on the global situation, making the prediction logic clear and transparent.

[0018] 3. This invention adopts a selective architecture of "decision first, calculation later", which only activates the most relevant time prediction path, avoiding the waste of resources in parallel computing schemes and improving computing efficiency and practicality. Attached Figure Description

[0019] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of an autonomous driving takeover time prediction method based on physical-driver-system three-factor evaluation and gating selection in an embodiment of the present invention; Figure 2 This is a schematic diagram of an autonomous driving takeover time prediction method based on physical-driver-system three-factor evaluation and gating selection in an embodiment of the present invention; Figure 3 This is an architecture diagram of an autonomous driving takeover time prediction system based on physical-driver-system three-factor evaluation and gating selection in an embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] like Figure 1 , Figure 2 As shown, an autonomous driving takeover time prediction method based on physical-driver-system three-factor evaluation and gating selection specifically includes the following steps: S1. Collect synchronous time-series data of the three elements within the historical time window; The three key elements are: physical interaction elements, driver readiness elements, and system reliability elements. Physical interaction element data primarily consists of physical risk data, including environmental perception data and vehicle status data. Environmental perception data is collected through environmental perception sensors, which in practice can be forward-facing cameras, forward-facing millimeter-wave radar, and / or solid-state lidar. Correspondingly, environmental perception data includes forward-facing camera image sequences, forward-facing millimeter-wave radar target list sequences, and solid-state lidar point cloud sequences. Vehicle status data is collected through the vehicle bus or inertial measurement unit, including vehicle longitudinal velocity, longitudinal / lateral acceleration, yaw rate, steering wheel angle, wheel speeds, and brake master cylinder pressure. Driver readiness element data mainly consists of driver status data collected by the driver monitoring system, including physiological signals and behavioral characteristics. In practice, this is acquired through a near-infrared camera-based driver monitoring system and a capacitively sensitive steering wheel, including driver eyelid opening percentage, gaze point screen coordinates, head posture angle, and hand grip force distribution. System reliability factors are mainly the health and performance data of various system components collected by the self-monitoring unit of the autonomous driving system, including: sensor status, positioning information and computing unit status. Sensor status includes camera exposure anomaly indicators, radar signal-to-noise ratio, average LiDAR echo intensity, etc. Positioning information includes GNSS positioning mode, horizontal protection level, inertial navigation system position drift rate, etc. Computing unit status includes CPU / GPU utilization of autonomous driving domain controller, core temperature, execution cycle jitter of key perception and planning threads, etc.

[0024] S2. Use three expert modules to extract features from the collected synchronous time-series data of the three elements respectively; The three expert modules are: physical interaction situation analysis module, driver readiness analysis module, and system reliability degradation analysis module.

[0025] The physical risk analysis module receives physical interaction element data and extracts primary feature vectors from environmental perception data and vehicle state data through parallel environmental coding sub-networks and vehicle coding sub-networks, respectively. Subsequently, a cross-attention mechanism is used to perform deep fusion processing on these two sets of primary features, and outputs feature vectors representing the dynamic evolution of the comprehensive safety boundary at the physical level.

[0026] Specifically, the physical interaction situation analysis module includes: an environment coding subnetwork, a vehicle coding subnetwork, and a bidirectional interaction attention unit; among them, the environment coding subnetwork includes an image branch, a point cloud branch, and a radar branch. The image branch uses a lightweight SlowOnly 3D CNN to extract spatiotemporal features from video clips and outputs feature vectors. The point cloud branch uses a PointPillar network to process the point cloud sequence, extracting features of dynamic targets and static structures, and outputting... The radar branch normalizes the radar target sequence (range, velocity, azimuth) and then extracts motion pattern features through an LSTM network, outputting the result. The features output from the three branches are concatenated and passed through a fully connected layer to obtain the primary environment feature encoding. The vehicle coding subnetwork standardizes the vehicle state time series, inputs it into a one-dimensional temporal convolutional network, and outputs a primary vehicle feature code. The bidirectional interactive attention unit comprises a first interaction layer representation, a second interaction layer representation, and a multi-layer perceptron structure. The first interaction layer representation identifies which information in the environment is most critical under vehicle capability limitations. ; in, , , Learnable weight matrix when environmental features are used as queries and vehicle features are used as keys and values. This is the scaling factor.

[0027] The second interaction layer represents which vehicle capabilities are most important under the current environmental risks and challenges: ; in, , , A learnable weight matrix when vehicle features are used as queries and environmental features are used as keys and values. This is the scaling factor.

[0028] Finally, and The first feature vector is output by concatenating the features and passing them through a multilayer perceptron structure. .

[0029] The driver readiness analysis module analyzes the driver's attention, behavior, and physiological signals to predict the driver's recovery process from the current state to a takeover state, outputting a feature vector describing the evolution of driver readiness. The module takes as input time-series data of the driver's state, including eye movements, head posture, and steering wheel grip force, and uses a bidirectional gated recurrent unit (Bi-GRU) to learn the temporal dependencies of the driver's state. After concatenating the forward and backward hidden states of the last time step, the result is passed through a fully connected layer to obtain the second feature vector. .

[0030] The system reliability degradation analysis module, based on degradation and failure propagation models, predicts the remaining reliability time of the system and outputs a feature vector describing the evolution of the system's health state. Specifically, the construction of the system reliability degradation analysis module includes: establishing degradation models based on stochastic processes (such as Wiener processes and gamma processes) for different key performance indicators; updating degradation model parameters online based on real-time collected system reliability element data and predicting the remaining time before each indicator first exceeds the preset failure threshold; and generating a reliability degradation prediction by combining the prediction results of each indicator and encoding it as a third feature vector.

[0031] A Wiener process degradation model with a drift term is established for each key indicator: ; in, These are performance index values. The drift coefficient, The diffusion coefficient is... For the standard Wiener process, This represents the initial state value of the performance metric.

[0032] The particle filter algorithm is used to update the parameters of each degradation model online based on real-time observation data. , and the current status The prediction indicates that this indicator will exceed the preset failure threshold for the first time. The time, i.e., the remaining useful life (RUL).

[0033] ; The predicted RUL of each indicator and its corresponding prediction uncertainty are concatenated and passed through a fully connected layer to output a third feature vector. .

[0034] S3. Perform global encoding on the extracted features and make gating decisions based on the global encoding results; The three temporal feature vector sequences output from step S2 are time-step aligned and concatenated to form a multi-factor joint feature sequence; this sequence is then input into the temporal context modeling network to generate a global security situation encoding vector. C The global security situation encoding vector is input into a gating decision network. Based on the current comprehensive situation, the gating decision network determines which of the three security elements—physical, driver, and system—constitutes the most urgent takeover constraint. The network outputs a three-element gating decision signal, indicating which element's takeover timing prediction path should be activated.

[0035] The three feature vectors are concatenated to obtain Features are input into the transformer encoder to obtain the global security posture encoding vector.C The vector is input into a gated decision network composed of an MLP, which outputs three logical values. The influence weights of each value are obtained through a softmax function. , , .

[0036] S4. Based on the output gating decision signal, activate the takeover timing prediction path corresponding to the feature vector with the largest weight value, and output the final takeover time.

[0037] The feature vector corresponding to the largest weight value is taken as the dominant vector for this takeover time prediction. It is input into the takeover timing prediction pathway (Multilayer Perceptron MLP) to obtain the final takeover time prediction value.

[0038] ; .

[0039] Specifically, by adopting a selective architecture of "decision first, calculation later", only the most relevant time prediction path is activated, which avoids the waste of resources in parallel computing schemes and improves computing efficiency and practicality.

[0040] The above embodiments have the following technical advantages: (1) This invention processes heterogeneous information such as physical environment, vehicle status, driver behavior and system health through independent expert modules, and performs dynamic interactive fusion at the decision-making level, realizing deep fusion of multi-source information, overcoming the limitations of simple splicing or fixed weight fusion of traditional methods, and significantly improving the prediction accuracy in complex scenarios.

[0041] (2) The present invention introduces a gating selection decision mechanism to dynamically focus on key constraints and can identify the most pressing safety constraints (such as sudden collision risk, serious driver distraction or sudden system failure) in real time based on the global situation, making the prediction logic clear and transparent.

[0042] like Figure 3 As shown, an autonomous driving takeover time prediction system based on physical-driver-system three-factor evaluation and gating selection is disclosed. The system includes: The information acquisition module collects synchronous time-series data of three elements within a historical time window; the three elements include: physical interaction element, driver readiness element, and system reliability element. The system employs three expert modules for feature extraction, each dedicated to extracting features from the collected synchronous time-series data of the three key elements. These modules include: a physical interaction situation analysis module, a driver readiness analysis module, and a system reliability degradation analysis module. The physical interaction situation analysis module receives physical interaction element data, extracts primary feature vectors from environmental perception data and vehicle state data, and uses a cross-attention mechanism to deeply fuse these two sets of primary features, outputting a feature vector representing the dynamic evolution of the comprehensive safety boundary at the physical level. The driver readiness analysis module analyzes the driver's attention, behavior, and physiological signals to predict the driver's recovery process from the current state to a takeover-ready state, outputting a feature vector describing the evolution of driver readiness. The system reliability degradation analysis module, based on degradation and failure propagation models, predicts the remaining system reliability time and outputs a feature vector describing the evolution of the system's health state. The global situation analysis and gating decision module performs global encoding on the extracted features and makes gating decisions based on the global encoding results. The takeover time prediction module activates the takeover timing prediction path corresponding to the feature vector with the largest weight value based on the output gating decision signal, and outputs the final takeover time; the takeover timing prediction path adopts a multilayer perceptron.

[0043] The autonomous driving takeover time prediction system of this embodiment is described simply because it corresponds to the autonomous driving takeover time prediction method in the above embodiment. For related similarities, please refer to the description of the autonomous driving takeover time prediction method in the above embodiment, which will not be described in detail here.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting autonomous driving takeover time based on physical-driver-system three-factor evaluation and gating selection, characterized in that, include: Collect synchronous time-series data of three elements within a historical time window; the three elements include: physical interaction element, driver readiness element, and system reliability element. Three expert modules are used to extract features from the collected synchronous time-series data of the three elements. These expert modules include: a physical interaction situation analysis module, a driver readiness analysis module, and a system reliability degradation analysis module. The physical interaction situation analysis module receives physical interaction element data, extracts primary feature vectors from environmental perception data and vehicle state data, and uses a cross-attention mechanism to deeply fuse these two sets of primary features, outputting a feature vector representing the dynamic evolution of the comprehensive safety boundary at the physical level. The driver readiness analysis module analyzes the driver's attention, behavior, and physiological signals, predicts the driver's recovery process from the current state to a takeover-ready state, and outputs a feature vector describing the evolution of driver readiness. The system reliability degradation analysis module, based on a degradation model and a failure propagation model, predicts the remaining system reliability time and outputs a feature vector describing the evolution of the system's health state. The extracted features are globally encoded, and gating decisions are made based on the global encoding results; Based on the output gating decision signal, the takeover timing prediction path corresponding to the feature vector with the largest weight value is activated, and the final takeover time is output; the takeover timing prediction path adopts a multilayer perceptron.

2. The method for predicting autonomous driving takeover time based on physical-driver-system three-factor evaluation and gating selection as described in claim 1, characterized in that, The physical interaction element data includes: environmental perception data collected by environmental perception sensors and vehicle status data collected by vehicle bus or inertial measurement unit. The driver readiness data is driver status data collected by the driver monitoring system, including the driver's attention signals, behavioral signals, and physiological signals. The system reliability data is the health and performance data of each system component collected by the self-monitoring unit of the autonomous driving system, including: sensor status, positioning information, and computing unit status.

3. The method for predicting autonomous driving takeover time based on physical-driver-system three-factor evaluation and gating selection as described in claim 2, characterized in that, The environmental perception data includes: a sequence of images from a forward-looking camera, a sequence of target lists from a forward-looking millimeter-wave radar, and a sequence of point clouds from a solid-state lidar; the vehicle status includes: the vehicle's longitudinal speed, longitudinal / lateral acceleration, yaw rate, steering wheel angle, wheel speeds, and brake master cylinder pressure. The driver readiness data includes: driver eyelid opening percentage, gaze point screen coordinates, head posture angle, and hand grip strength distribution; The system reliability data includes: camera exposure anomaly indicator, radar signal-to-noise ratio, average lidar echo intensity; GNSS positioning mode, horizontal protection level, inertial navigation system position drift rate; CPU / GPU utilization of the autonomous driving domain controller, core temperature, and execution cycle jitter of key perception and planning threads.

4. The method for predicting autonomous driving takeover time based on physical-driver-system three-factor evaluation and gating selection as described in claim 1, characterized in that, The physical interaction situation analysis module includes: an environment coding subnetwork, a vehicle coding subnetwork, and a bidirectional interactive attention unit; wherein, the environment coding subnetwork includes an image branch, a point cloud branch, and a radar branch, and the image branch adopts a lightweight SlowOnly 3D... The CNN extracts spatiotemporal features from video clips and outputs image feature vectors. The point cloud branch uses a PointPillar network to process the point cloud sequence, extracting features of dynamic targets and static structures, and outputting point cloud feature vectors. The radar branch normalizes the radar target sequence and extracts motion pattern features through an LSTM network, outputting radar feature vectors. The features output from the three branches are concatenated and passed through a fully connected layer to obtain the primary environmental feature encoding. The vehicle encoding sub-network standardizes the vehicle state time series, inputs it into a one-dimensional temporal convolutional network, and outputs primary vehicle feature encoding. The bidirectional interactive attention unit includes a first interactive layer representation, a second interactive layer representation, and a multilayer perceptron structure. The first interactive layer representation identifies which information in the environment is most critical under vehicle capability limitations. The second interactive layer representation identifies which vehicle capabilities are most important under current environmental risks and challenges. Finally, the first and second interactive layer representations are concatenated and output as a first feature vector through the multilayer perceptron structure.

5. The method for predicting autonomous driving takeover time based on physical-driver-system three-factor evaluation and gating selection as described in claim 1, characterized in that, The driver readiness analysis module takes driver state time-series data as input and uses a bidirectional gated loop unit to learn the temporal dependencies of driver state. After concatenating the forward and backward hidden states of the last time step, the second feature vector is obtained by passing it through a fully connected layer.

6. The method for predicting autonomous driving takeover time based on physical-driver-system three-factor evaluation and gating selection as described in claim 1, characterized in that, The construction of the system reliability degradation analysis module includes: establishing a degradation model based on a stochastic process for different key performance indicators; updating the degradation model parameters online based on real-time collected system reliability element data, and predicting the remaining time before each indicator first exceeds the preset failure threshold; and generating a reliability degradation prediction by combining the prediction results of each indicator, and encoding it as a third feature vector.

7. The method for predicting autonomous driving takeover time based on physical-driver-system three-factor evaluation and gating selection as described in claim 6, characterized in that, Establish a degradation model based on stochastic processes, including: establishing a Wiener process degradation model with a drift term for each key indicator. ； in, These are performance index values. The drift coefficient, The diffusion coefficient is... For the standard Wiener process, This represents the initial state value of the performance metric.

8. The method for predicting autonomous driving takeover time based on physical-driver-system three-factor evaluation and gating selection as described in claim 7, characterized in that, Online updating of degradation model parameters includes: using a particle filter algorithm to update the parameters of each degradation model online based on real-time observation data. , and the current status .

9. The method for predicting autonomous driving takeover time based on physical-driver-system three-factor evaluation and gating selection as described in claim 1, characterized in that, The extracted features are globally encoded, and gating decisions are made based on the global encoding results, including: The three extracted time-series feature vector sequences are aligned and concatenated at time steps to form a multi-factor joint feature sequence; The multi-factor joint feature sequence is input into a temporal context modeling network to generate a global security situation encoding vector; the temporal context modeling network is a transformer encoder. The global security situation encoding vector is input into the gating decision network; based on the current comprehensive situation, the gating decision network determines which of the three safety elements—physical, driver, and system—constitutes the most urgent takeover constraint, and outputs a three-element gating decision signal to indicate which element's takeover timing prediction path should be activated; the gating decision network is composed of a multi-layer sensing mechanism.

10. An automated driving takeover time prediction system based on physical-driver-system three-factor evaluation and gating selection, characterized in that, include: The information acquisition module collects synchronous time-series data of three elements within a historical time window; the three elements include: physical interaction element, driver readiness element, and system reliability element. The system employs three expert modules for feature extraction, each dedicated to extracting features from the collected synchronous time-series data of the three key elements. These modules include: a physical interaction situation analysis module, a driver readiness analysis module, and a system reliability degradation analysis module. The physical interaction situation analysis module receives physical interaction element data, extracts primary feature vectors from environmental perception data and vehicle state data, and uses a cross-attention mechanism to deeply fuse these two sets of primary features, outputting a feature vector representing the dynamic evolution of the comprehensive safety boundary at the physical level. The driver readiness analysis module analyzes the driver's attention, behavior, and physiological signals to predict the driver's recovery process from the current state to a takeover-ready state, outputting a feature vector describing the evolution of driver readiness. The system reliability degradation analysis module, based on degradation and failure propagation models, predicts the remaining system reliability time and outputs a feature vector describing the evolution of the system's health state. The global situation analysis and gating decision module performs global encoding on the extracted features and makes gating decisions based on the global encoding results. The takeover time prediction module activates the takeover timing prediction path corresponding to the feature vector with the largest weight value based on the output gating decision signal, and outputs the final takeover time; the takeover timing prediction path adopts a multilayer perceptron.

Images