Driver potential physiological risk detection method and system based on physiological energy manifold

Abstract

The invention discloses a driver potential physiological risk detection method and system based on a physiological energy manifold. Relates to the technical field of intelligent traffic, artificial intelligence and physiological signal processing, in particular to a driver potential physiological risk detection method and system based on a physiological energy manifold. Through the deep learning technology, the physiological energy manifold of the driver in the stable driving state is directly learned from the original physiological signals, and therefore potential physiological risks which cannot be sensed by a physical sensor and are caused by psychological tension are recognized. The method comprises the following steps: firstly, providing an event-driven physiological signal slice and single sample normalization method, and eliminating individual baseline difference; secondly, designing and training a physiological energy model, and learning a low-energy manifold of stable driving by comparing an energy loss function; and finally, in combination with physical rules and physiological energy scores, constructing double-flow judgment logic, and realizing accurate detection of potential physiological risks of physical safety but psychological tension.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation, artificial intelligence, and physiological signal processing, specifically to a method and system for detecting potential physiological risks to drivers based on physiological energy manifolds. Background Technology

[0002] With the rapid popularization of Intelligent Transportation Systems (ITS) and Advanced Driver Assistance Systems (ADAS), human-machine co-driving has become the mainstream form of current autonomous driving technology development. In this form, real-time and accurate perception of the driver's physical and mental state (state monitoring) is not only a prerequisite for achieving a smooth handover of driving power, but also the key to building a safe trust mechanism.

[0003] However, existing driver risk detection and condition analysis technologies still face significant challenges in practical applications, mainly in the following two technical dimensions: First, regarding the dimensions and depth of risk perception, existing mainstream solutions rely excessively on vehicle dynamics parameters and external environment perception, resulting in significant "perception blind spots." Current technologies typically use inertial measurement units (IMUs) to collect vehicle acceleration and jerk, or use radar and visual sensors to calculate time-to-collision (TTC) to determine driving risk. This detection logic, based on "explicit physical rules," can only identify moments of physical danger where the vehicle has already moved violently or a collision is imminent. However, in complex real-world traffic environments, there are often "potential physiological risk" scenarios (such as the anticipated moment of meeting oncoming traffic on a narrow road or the sudden appearance of a pedestrian). In these situations, the vehicle may maintain a low and stable speed, and physical sensor readings may indicate safety, but the driver may be in a state of extreme tension or panic. Existing physical rule-based methods cannot perceive this cognitive-level risk of "the mind reacting before the car moves," leading to delayed intervention or misjudgments by driver assistance systems, severely impacting system safety redundancy and user experience. Secondly, regarding the modeling and processing paradigm of physiological signals, although some existing technologies attempt to introduce physiological signals (such as ECG and EDA) to assist in detection, their processing methods are still limited by traditional shallow feature engineering and lack the ability to generalize to unknown scenarios. The closest existing implementation (e.g., Analysis of Driver Behavior in Various Events Using Electrodermal Activity Signal, IEEE BSN, 20 Aug 2025) typically employs the following process: first, denoising the original signal; then, manually extracting a series of statistical features (such as kurtosis, skewness, mean, variance, etc.); and finally, feeding it into traditional classifiers such as Support Vector Machine (SVM) or Random Forest (RF) for discrete category classification. This processing paradigm has the following technical limitations: firstly, the manual feature extraction process leads to a significant loss of deep temporal microstructure and nonlinear dynamic information contained in the bioelectrical signals, limiting the upper limit of the model's detection accuracy; secondly, these supervised learning methods model risk detection as a simple "binary classification" or "multi-class classification" problem, heavily relying on training data containing various hazard labels. This means that the model can only identify specific types of hazards seen in the training set (Closed-set), and cannot effectively measure unknown anomalies that have not appeared in the training set (Open-set Anomaly) based on the "safety manifold".

[0004] In summary, current driving risk detection technologies have fundamental limitations: physical sensors cannot quantify "potential psychological risks," while traditional physiological calculation methods are limited by manual features and classification paradigms, failing to achieve end-to-end, continuous risk energy measurement. Therefore, constructing a detection method that can integrate deep physiological features, overcome the blind spots of physical rules, and quantify the driver's intrinsic "physiological cost" in an unsupervised or self-supervised manner is a key technical challenge for improving the anthropomorphism and safety of autonomous driving systems. Summary of the Invention

[0005] To address the shortcomings of existing technologies in effectively detecting potential physiological risks within the blind spots of physical rules, and the lack of an end-to-end processing paradigm capable of preserving deep physiological characteristics and performing continuous energy measurement, this invention aims to propose a method and system for detecting potential physiological risks to drivers based on physiological energy manifolds. By employing deep learning technology, the physiological energy manifold of a driver under stable driving conditions is learned directly from raw physiological signals, thereby identifying potential physiological risks caused by psychological stress that are imperceptible to physical sensors.

[0006] The method includes the following steps: S1. Acquire and preprocess the driver's raw physiological signals. The preprocessed signal is normalized using single-sample Z-score normalization to obtain normalized physiological signal samples. ; S2. Constructing and training the PhysioEBM physiological signal energy model: PhysioEBM goes through the following layers from input to output: feature extraction layer, global aggregation layer, and energy output layer. PhysioEBM also includes a contrastive energy loss function; pass Training PhysioEBM; S3. Obtain real-time normalized physiological signal samples The real-time energy score is calculated by forward propagation using the trained PhysioEBM. ,in, This is the index value for the time. S4. Acquire vehicle dynamics data. When the vehicle dynamics data is less than the physical safety threshold, trigger the potential physiological risk detection step. If real-time energy fraction Greater than the physiological alert threshold If so, it is determined to be a state of potential physiological risk. Conversely, if the conditions are not met, the condition is considered safe.

[0007] Furthermore, the preprocessing specifically involves: based on vehicle driving events, analyzing the raw physiological signals... Extract a fixed-length time series segment from the data, denoted as . ,in, The index value of the time series segment; The formula for calculating the single-sample Z-Score normalization is: ,in, Let be the mean of the i-th time series segment in the time dimension. For the first The standard deviation of each time series segment in the time dimension.

[0008] Furthermore, normalized physiological signal samples Includes: positive samples of smooth driving conditions Negative samples of risky driving conditions ; Several positive samples of stable driving conditions To form a smooth driving dataset ; Negative samples of several risky driving conditions To form a risky driving dataset .

[0009] Furthermore, the feature extraction layer passes through the following layers from input to output in sequence: a one-dimensional convolutional layer, a batch normalization layer, a Mish activation function, and a regularization layer; The global aggregation layer includes: an adaptive global average pooling layer; The energy output layer includes: a fully connected layer.

[0010] Furthermore, the formula for calculating the energy loss function is as follows: ,in, To obtain data from the smooth driving dataset Randomly select a number of positive samples To calculate The mean, The energy function of PhysioEBM. for energy, To obtain data from the risk driving dataset Randomly select a number of negative samples To calculate The mean, for energy, Energy boundary threshold, This represents the ReLU activation function.

[0011] Furthermore, the training method of PhysioEBM is as follows: with the aim of lowering the energy of stable samples and simultaneously increasing the energy of risky samples, the backpropagation algorithm is used to minimize the contrastive energy loss function L.

[0012] Furthermore, real-time normalized physiological signal samples The real-time physiological signals are processed by inputting them into step S1.

[0013] A driver potential physiological risk detection system based on physiological energy manifold, characterized in that the system is used to implement the above method, and the system includes: a signal acquisition module, an on-board computing unit, a dual-flow punching module, and a human-machine interface.

[0014] Furthermore, the signal acquisition module is used to acquire the driver's real-time physiological signals; The onboard computing unit includes: a signal preprocessing module and a physiological signal energy model (PhysioEBM); The signal preprocessing module is used to preprocess real-time physiological signals and perform single-sample Z-Score normalization to obtain real-time normalized physiological signal samples. ; PhysioEBM, a physiological signal energy model, is used to analyze real-time normalized physiological signal samples. Calculate real-time energy fraction ; The dual-flow punching module connects to the vehicle's CAN bus to acquire vehicle dynamics data; The dual-flow punching module is used to calculate the energy fraction in real time. Detect risks using vehicle dynamics data and output judgment results; The human-machine interface is used to provide feedback on the judgment results to the driver.

[0015] Furthermore, the dual-flow punching module is used to calculate the energy fraction in real time. The specific risks associated with vehicle dynamics data detection are as follows: S101. If the vehicle dynamics data is greater than the physical safety threshold If so, it is determined to be: explicit physical risk; Conversely, proceed to step S102; S102, If the real-time energy fraction Greater than the physiological alert threshold If so, it is determined to be a state of potential physiological risk. Conversely, if the condition is not met, it is determined to be a safe state.

[0016] The beneficial effects of the method described in this invention are as follows: (1) Existing technologies often directly use the original signal or perform simple global normalization. Due to the significant differences in individual driver physiological baselines (such as resting heart rate and skin conductance), the model is difficult to transfer between different drivers. The method described in this invention adopts the strategy of "event-driven slicing + single-sample Z-score independent normalization" to transform absolute amplitude into relative change. This technique effectively eliminates the influence of individual baseline drift, enabling the model to focus on the intrinsic law of physiological signal changes with driving stress, thereby significantly improving the detection accuracy and robustness across subjects. Furthermore, since the method described in this invention uses single-sample independent normalization, the model does not need to perform tedious baseline calibration or collect a large amount of resting data for each new driver, greatly reducing the time cost of system deployment and use.

[0017] (2) Existing technologies mostly employ traditional classification paradigms (such as SVM and RF), which can only identify specific risk types defined in the training set (closed set detection). The method described in this invention constructs a physiological signal energy model, PhysioEBM, and learns the low-energy manifold of "smooth driving" by comparing energy loss functions. The method described in this invention enables the model to have powerful anomaly detection capabilities, not only identifying known risks such as sudden braking and cutting in, but also making high-energy responses to unknown abnormal states that deviate from the normal manifold and have never appeared in the training set (such as sudden illness or extreme emotional fluctuations), significantly improving the safety redundancy of the system.

[0018] (3) Compared with existing technologies that rely on manual extraction of statistical features (such as kurtosis and skewness) leading to information loss, this invention uses one-dimensional convolutional layers (1D-CNN) for end-to-end feature learning, which can automatically capture microsecond-level temporal fluctuations and nonlinear dynamic features in physiological signals. Experimental results show that the AUC index (0.64) of this method on the standard classification task is significantly better than the state-of-the-art method (0.58) based on manual features, demonstrating its superior feature representation ability.

[0019] (4) Existing technologies rely excessively on vehicle dynamics parameters (physical rules) and cannot perceive the "potential physiological risk" of a driver experiencing psychological panic while the vehicle is stable. The dual-flow judgment logic proposed in this invention combines the objectivity of physical rules with the subjectivity of physiological energy, and for the first time realizes the quantification and detection of the potential physiological risk state of "the mind is already agitated before the vehicle moves." This is of decisive significance for improving the anthropomorphism and safety of autonomous driving systems in complex game scenarios (such as meeting oncoming traffic on narrow roads and making unprotected left turns).

[0020] Unlike existing technologies that only output discrete labels, the method described in this invention outputs a continuous "physiological energy score," which not only intuitively reflects the driver's psychological load level, but can also be directly used as a "physiological cost" or negative reward signal in reinforcement learning algorithms. It can be seamlessly integrated into the planning and control closed loop of autonomous driving, providing a quantitative basis for constructing a "human-like" driving strategy that takes into account both efficiency and comfort.

[0021] (5) The system described in this invention integrates functional modules such as physiological signal acquisition, data preprocessing, energy calculation, detection result judgment, and result feedback into the same in-vehicle closed-loop system, realizing real-time perception and dynamic response to the driver's physiological risk state. The system described in this invention can continuously complete data acquisition and risk judgment during vehicle operation and promptly feed back the judgment results to the human-machine interaction or vehicle control unit, thereby forming a closed-loop mechanism of perception-judgment-feedback. Through system-level integrated design, it not only improves the real-time performance and stability of physiological risk detection in the in-vehicle environment, but also avoids the delay and consistency problems caused by the fragmented deployment of multiple systems, enabling the risk identification results to effectively support the decision-making and execution of the driving assistance or autonomous driving system, and comprehensively improving the safety, reliability, and engineering feasibility in the in-vehicle human-machine co-driving scenario. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method described in this invention; Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0024] Example 1 This embodiment provides a method for detecting potential physiological risks to drivers based on physiological energy manifolds. The flowchart of the method is as follows: Figure 1 As shown.

[0025] In terms of the dimensions and depth of risk perception, existing technologies typically rely excessively on vehicle dynamics parameters (such as acceleration, jerk, and time-to-collision (TTC)) to determine risk. This detection logic, based on explicit physical rules, is inherently lagging, only able to identify moments of physical danger when the vehicle has already undergone violent movement or a collision is imminent. For "potential physiological risks" in complex traffic scenarios (such as when meeting oncoming traffic on a narrow road or anticipating the sudden appearance of a pedestrian), where the vehicle may maintain a low and stable speed, but the driver is already in a state of extreme tension or panic, existing technologies have significant blind spots. This blind spot often leads to delayed intervention or misjudgments by driver assistance systems, severely impacting the system's safety redundancy and user trust in autonomous driving.

[0026] To address the problem that existing technologies cannot detect "potential physiological risks," the present invention provides a detection method based on the Physiological Energy Model (EBM). This method, by learning the physiological energy manifold under stable driving conditions, can independently and sensitively identify abnormal states of increased driver psychological load, thus effectively filling the perception blind spots of physical sensors and improving the safety and anthropomorphism of human-machine co-driving systems. The method includes the following steps: S1. To address the issue of significant individual differences (baseline drift) in driver physiological signals, the method described in this invention employs an "event-driven slicing + single-sample independent normalization" strategy. This allows the model to focus on the relative changes in signal characteristics rather than absolute amplitude, thereby enabling cross-individual generalization. The events include, but are not limited to, time window triggering, vehicle control command triggering, or environmental perception triggering.

[0027] The specific implementation of step S1 is as follows: acquiring the driver's raw physiological signals through onboard or wearable sensors (ECG / EDA). The original physiological signals are preprocessed, and then each preprocessed signal (slice sample) is normalized using a single-sample Z-Score to eliminate individual differences and obtain normalized physiological signal samples. ; Preprocessing specifically involves: based on vehicle driving events (including events that cause extreme tension or panic in the driver, such as sudden braking, sudden appearance of a pedestrian, sudden illness, extreme emotional fluctuations, and meeting oncoming traffic on narrow roads), processing continuous raw physiological signals... Extract a fixed-length time series segment from the data, denoted as . ,in, The index value of the time series segment is a fixed length in this embodiment. For 2000 sampling points; The formula for calculating the single-sample Z-Score normalization is: ,in, Let be the mean of the i-th time series segment in the time dimension. Let be the standard deviation of the i-th time series segment in the time dimension.

[0028] Normalized physiological signal samples Includes: positive samples of smooth driving conditions Negative samples of risky driving conditions ; Several positive samples of stable driving conditions To form a smooth driving dataset ; Negative samples of several risky driving conditions To form a risky driving dataset .

[0029] In the method described in this invention, single-sample Z-score normalization calculates the mean and standard deviation independently for each truncated sample and performs standardization. This differs from traditional global normalization and is key to eliminating individual differences. This step makes the model input sensitive to the waveform morphology of the signal but insensitive to baseline drift.

[0030] Regarding the modeling and processing paradigms of physiological signals, while existing technologies attempt to incorporate physiological signals to assist detection, they mostly employ a processing mode of "manual feature extraction + traditional classifiers." Manual feature extraction (such as statistical kurtosis and skewness) leads to a significant loss of deep temporal microstructure and nonlinear dynamic information contained in bioelectrical signals, limiting the upper limit of the model's detection accuracy. Furthermore, existing technologies typically model risk detection as a simple supervised binary or multi-class classification problem, ignoring the continuity of physiological state changes. Such models can only identify specific hazard types seen in the training set (Closed-set) and cannot effectively measure unknown anomalies not seen in the training set (Open-set Anomalies) based on the "safety manifold," resulting in poor generalization ability and low robustness when facing unknown risks in the open world.

[0031] To address the issues of information loss and poor generalization ability in traditional manual feature extraction, this invention aims to provide an end-to-end risk energy measurement method. By constructing a deep neural network model (preferably a lightweight one-dimensional convolutional neural network 1D-CNN) and introducing a contrastive energy loss function, a direct mapping from the original physiological signal to the risk energy value is achieved. This method not only fully preserves the deep temporal microstructure of the signal but also, through manifold learning, endows the model with strong generalization ability for open-set anomalies, significantly improving detection robustness. The model is constructed and trained in step S2.

[0032] S2. The method described in this invention abandons the traditional classification paradigm and instead adopts the Energy Basis Model (EBM) paradigm. By designing a special contrastive energy loss function, "smooth driving" is modeled as a low energy valley in the energy landscape, pushing abnormal states to high-energy regions, thereby achieving open-set detection capability for unknown risks. Note that physiological signals are not limited to ECG / EDA, but should also include PPG, EMG, EEG, and physiological indicators extracted through non-contact visual sensors. This step utilizes positive samples from relaxed driving (Relaxed, ) and high-risk driving negative samples (Risky, Training a deep neural network. (Note: Negative samples during the training phase include not only physically dangerous driving data, but also, crucially, physiological data collected in driving simulators through psychologically induced stress (such as cognitive tasks and simulated emergencies), showing the vehicle remaining stable while the driver is psychologically tense.) The specific implementation of step S2 is as follows: Construct and train a deep neural network consisting of a one-dimensional convolutional layer (1D-CNN), a global average pooling layer (GAP), and a fully connected layer (PhysioEBM (Offline Phase), used to map temporal physiological signals into scalar energy values).

[0033] PhysioEBM goes through the following layers from input to output: feature extraction layer, global aggregation layer, and energy output layer. The feature extraction layer passes through the following layers from input to output: a one-dimensional convolutional layer (Conv1d), a batch normalization layer (BN), a Mish activation function, and a dropout regularization layer. N); the feature extraction layer is used to extract deep temporal features.

[0034] The global aggregation layer includes: an adaptive global average pooling layer (AdaptiveAvgPool1d). The global aggregation layer is used to compress deep temporal features.

[0035] The energy output layer includes: a fully connected layer (Linear); The energy output layer is used to map compressed deep temporal features to scalar energy values ​​E.

[0036] PhysioEBM also includes a contrastive energy loss function; the contrastive energy loss function consists of two parts: minimizing the energy of positive samples (smooth driving) while maximizing the energy of negative samples (high-risk driving) to a preset margin.

[0037] The design of the contrastive energy loss function: abandoning the traditional classification loss, it adopts an energy-based manifold learning strategy and defines the loss function. as follows: ,in, To obtain data from the smooth driving dataset Randomly select a number of positive samples To calculate The mean, The energy function of PhysioEBM. for energy, To obtain data from the risk driving dataset Randomly select a number of negative samples To calculate The mean, for energy, In this embodiment, the energy boundary threshold is used. , This represents the ReLU activation function.

[0038] pass PhysioEBM is trained end-to-end using smooth and high-risk driving data to enable it to learn the safe driving manifold.

[0039] The training method of PhysioEBM is to construct a low-energy "safe driving manifold" in the feature space by lowering the energy of stationary samples (making them approach 0) and increasing the energy of risky samples (making them greater than m). The backpropagation algorithm is used to minimize the contrastive energy loss function L.

[0040] Regarding the continuity and interpretability of risk quantification, existing technologies typically output discrete binary or multi-valued labels (such as "dangerous / safe"), lacking a continuous quantitative description of the degree of risk. Discrete output signals make it difficult for autonomous driving systems to adopt tiered control strategies based on the severity of risks (such as achieving a smooth transition from mild warnings to emergency takeover), and also make it difficult to directly feed physiological risks as continuous reward signals into reinforcement learning algorithms for policy optimization, thus limiting their application value in the closed loop of autonomous driving planning and control.

[0041] To address the problem of discrete risk quantification and difficulty in closed-loop control, the method described in this invention outputs a continuous and interpretable Energy Score. This score not only quantifies the current level of risk, supporting a smooth transition of control strategies, but can also be directly integrated into the reinforcement learning framework of autonomous driving as a "physiological cost" or "negative reward" signal, providing a quantitative basis for constructing intelligent driving strategies that balance driving efficiency and passenger comfort.

[0042] This invention is the first to clearly define and detect the state of "potential physiological risk," which is a special scenario where the vehicle's dynamic parameters are within limits (physical safety) but the driver's physiological energy is abnormally elevated (psychological danger), filling the perception blind spot of existing ADAS systems.

[0043] S3. Using the trained PhysioEBM model, calculate the current driver's physiological energy score in real time.

[0044] The specific implementation of step S3 is as follows: acquire real-time physiological signals, and input the real-time physiological signals into step S1 for processing to obtain real-time normalized physiological signal samples. ; Normalized physiological signal samples The real-time energy score is calculated by forward propagation using the trained PhysioEBM (Online Phase). ,in, The index value is the time; the online stage indicates the stage of the vehicle's journey. Real-time normalized physiological signal samples The real-time physiological signals are processed by inputting them into step S1.

[0045] The method described in this invention is the first to clearly define and detect the state of "potential physiological risk", that is, a special scenario in which the vehicle's dynamic parameters are not exceeded (physical safety) but the driver's physiological energy is abnormally increased (psychological danger), filling the perception blind spot of the existing ADAS system.

[0046] S4. Construct a parallel judgment mechanism that includes "physical channels" and "physiological channels", and clearly define the specific judgment rules for protection: output a "potential physiological risk" warning signal if and only if "physical rules determine it to be safe" and "physiological energy determines it to be dangerous".

[0047] The specific implementation of step S4 is as follows: acquire vehicle dynamics data (such as Jeck), and trigger the potential physiological risk detection step when the vehicle dynamics data is less than the physical safety threshold; If real-time energy fraction Greater than the physiological alert threshold If so, it is determined to be a state of potential physiological risk. Conversely, if the conditions are not met, the condition is considered safe.

[0048] The core of the method described in this invention lies in constructing a physiological signal anomaly detection framework based on the "Energy Manifold" to solve the problem that traditional physical sensors cannot detect the "potential physiological risks" of drivers. First, an event-driven physiological signal slicing and single-sample normalization method (S1) is proposed to eliminate individual baseline differences; Secondly, a physiological energy model (PhysioEBM) based on 1D-CNN is designed and trained to learn the low-energy manifold for smooth driving by comparing the energy loss function (S2). Finally, by combining physical rules and physiological energy scores, a dual-flow judgment logic (S4) is constructed to accurately detect potential physiological risks of "physical safety but psychological tension".

[0049] Example 2 This embodiment further defines Embodiment 1. This embodiment provides a driver potential physiological risk detection system based on physiological energy manifold. The structure of the system is as follows: Figure 2 As shown, the system is a hardware carrier and system architecture for implementing the method described in Embodiment 1.

[0050] The system includes: a signal acquisition module, an on-board computing unit, a dual-stream punching module, and a human-machine interface.

[0051] The signal acquisition module is a non-contact bioelectrode integrated into the steering wheel or seat to acquire the driver's real-time physiological (ECG / EDA) signals; The onboard computing unit includes: a signal preprocessing module and a physiological signal energy model (PhysioEBM); The signal preprocessing module is used to preprocess real-time physiological signals and perform single-sample Z-Score normalization to obtain real-time normalized physiological signal samples. ; PhysioEBM, a physiological signal energy model, is used to analyze real-time normalized physiological signal samples. Calculate real-time energy fraction ; The dual-flow punching module connects to the vehicle's CAN bus to acquire vehicle dynamics data (such as...). ); The dual-flow punching module is used to calculate the energy fraction in real time. The specific risks associated with vehicle dynamics data detection are as follows: S101. If the vehicle dynamics data is greater than the physical safety threshold (Right now If the risk is ), then it is classified as: explicit physical risk (other risks); Conversely (i.e.) If the physical safety is confirmed, then proceed to step S102, the physiological energy determination step. 102. If the real-time energy fraction Greater than the physiological alert threshold ( > If the vehicle is physically stable but the driver's psychological load is too high, it is judged as: potential physiological risk (hidden risk state). Conversely, if the condition is not met, it is determined to be a safe state.

[0052] Based on the judgment results, the human-computer interaction interface provides graded prompts to the driver through changes in ambient light color or seat vibration.

[0053] Example 3 This embodiment is a further limitation of Embodiment 1.

[0054] This embodiment provides an alternative to physiological signal preprocessing and feature extraction (corresponding to steps S1 and S2 in Embodiment 1): (1) Alternative to Multimodal Fusion: Although ECG / EDA signals were used as an example in Example 1, in practical applications, the present invention can also utilize an "end-to-end multimodal fusion model". For example, a multi-channel parallel input neural network can be constructed, which includes independent encoders for different signals (such as a 2D-CNN for facial expression heatmaps, a visual Transformer (ViT) for eye-tracking data, and a one-dimensional convolutional network (1D-CNN) for ECG / EDA temporal data). The feature vectors output by these encoders are concatenated or fused through a cross-attention mechanism and then input into the energy output layer. This alternative also aligns with the core concept of the present invention of "learning physiological energy manifolds".

[0055] (2) Replacement of network backbone: In step S2, this invention focuses on the PhysioEBM architecture based on 1D-CNN. Other deep neural network models that can effectively extract temporal features, such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), or Transformer (based on Self-Attention mechanism), are also applicable to the energy model construction process proposed in this invention. As long as its output is designed as a scalar energy value and trained using contrastive energy loss, it falls within the scope of protection of this invention.

[0056] Example 4 This embodiment is a further limitation of Embodiment 1.

[0057] This embodiment provides an alternative solution for the logic regarding the determination of potential physiological risks (corresponding to step S4): (1) Probabilistic Model-Based Alternatives: Although a threshold-based dual-stream decision logic was used in Example 1, a "probabilistic graphical model" or "Bayesian inference" scheme can also be adopted. For example, the objective risk probability output by the physical rules can be used. Subjective physiological risk probability output by EBM As input nodes of a Bayesian network, the posterior probability is calculated. To determine potential physiological risks.

[0058] (2) Alternative based on machine learning regression: The physiological energy score output in step S3 can also be concatenated with the vehicle dynamics parameters (speed, acceleration) to form a feature vector, which is then input into a lightweight machine learning regression model (such as support vector regression SVR or gradient boosting tree XGBoost) to directly train the regression model to output the final comprehensive risk index.

[0059] Example 5 This embodiment is a further limitation of Embodiment 1.

[0060] This embodiment provides an alternative regarding downstream applications and reward function construction: (1) Application of Online Reinforcement Learning (Online RL): The physiological energy score generated by this invention It can be used not only for offline dataset construction but also for online training paradigms. For example, it can be used for real-time computation. Negative values ​​are taken as physiological rewards and with task rewards Weighted mixing is performed to provide real-time, human factors-integrated reward signals to online reinforcement learning algorithms (such as Soft Actor-Critic, SAC, or Proximal Policy Optimization, PPO) to guide the real-time optimization of policies in simulated environments or real vehicles.

[0061] (2) Application of Imitation Learning: Furthermore, the physiological energy score can also be used as a data screening criterion. For example, using... "Expert trajectories" that are "low-energy and high-efficiency" (i.e., high-performance and high-comfort) are selected from massive natural driving datasets and then used to train a "behavioral cloning (BC)" model, thereby enabling a rapid cold start of human-like driving strategies.

Claims

1. A method for detecting potential physiological risks of a driver based on a physiological energy manifold, characterized in that, The method comprises the following steps: S1, obtaining and pre-processing the raw physiological signals of the driver performing single-sample Z-Score normalization on the pre-processed signals to obtain normalized physiological signal samples ; S2, constructing and training a physiological signal energy model PhysioEBM: The PhysioEBM sequentially passes through a feature extraction layer, a global aggregation layer and an energy output layer from input to output; The PhysioEBM further comprises a contrast energy loss function; By Training PhysioEBM; S3, obtaining a real-time normalized physiological signal sample calculating a real-time energy score by forward propagation with the trained PhysioEBM wherein, is an index value of time; S4, acquiring vehicle dynamics data, when the vehicle dynamics data is less than a physical safety threshold, triggering a potential physiological risk detection link; if real-time energy fraction greater than physiological alert threshold then determine: potential physiological risk state; Otherwise, it is determined as a safe state.

2. The method of claim 1, wherein the physiological energy landscape is determined by a plurality of physiological sensors. The preprocessing specifically comprises: based on a vehicle driving event, intercepting a time sequence segment of a fixed length from an original physiological signal denoted as wherein, is an index value of the time sequence segment. The formula for calculating the single-sample Z-Score normalization is: ,in, Let be the mean of the i-th time series segment in the time dimension. For the first The standard deviation of each time series segment in the time dimension.

3. The method of claim 2, wherein the physiological energy landscape is determined by a plurality of physiological sensors. Normalized physiological signal samples including positive samples of a steady driving state and negative samples of a risky driving state ; a number of positive samples of steady driving states , to constitute a steady driving dataset ; Negative samples of several risky driving states , to constitute a risky driving dataset .

4. The method of claim 3, wherein the physiological energy landscape is determined by a plurality of physiological sensors. The feature extraction layer sequentially passes through a one-dimensional convolution layer, a batch normalization layer, a Mish activation function and a regularization layer from input to output; The global aggregation layer comprises an adaptive global average pooling layer; The energy output layer comprises a fully connected layer.

5. The physiological energy landscape based driver potential physiological risk detection method according to claim 4, characterized in that, The formula for calculating the energy loss function is as follows: ,in, To obtain data from the smooth driving dataset Randomly select a number of positive samples To calculate The mean, The energy function of PhysioEBM. for energy, To obtain data from the risk driving dataset Randomly select a number of negative samples To calculate The mean, for energy, The energy boundary threshold, This represents the ReLU activation function.

6. The physiological energy landscape based driver potential physiological risk detection method according to claim 5, characterized in that, The training method of the PhysioEBM is: taking the purpose of lowering the energy of stable samples and simultaneously increasing the energy of risk samples, using a back propagation algorithm to minimize the contrast energy loss function L.

7. The physiological energy landscape based driver potential physiological risk detection method according to claim 6, characterized in that, Real-time normalized physiological signal samples By way of example, the real-time physiological signal input step S1 is performed by processing.

8. A driver potential physiological risk detection system based on physiological energy manifold, characterized in that, The system is used to implement the method of any one of claims 1 to 7, and the system comprises a signal acquisition module, a vehicle-mounted computing unit, a double-flow blanking module and a human-computer interaction interface.

9. The physiological energy landscape based driver's potential physiological risk detection system according to claim 8, characterized in that, The signal acquisition module is used to acquire real-time physiological signals of a driver; The vehicle-mounted computing unit comprises a signal preprocessing module and a physiological signal energy model PhysioEBM; The signal preprocessing module is configured to preprocess and perform single-sample Z-Score normalization on the real-time physiological signal to obtain a real-time normalized physiological signal sample ; A physio-signal energy model PhysioEBM is used to compute real-time energy scores from real-time normalized physio-signal samples ;​ The double-flow blanking module is connected with a vehicle CAN bus to acquire vehicle dynamics data; A double flow blanking module is used to detect risks according to real-time energy score and vehicle dynamics data, output a decision result; The human-computer interaction interface is used to feed back the determination result to the driver.

10. The physiological energy landscape based driver potential physiological risk detection system of claim 9, wherein, A double flow blanking module for detecting risks according to real-time energy fraction and vehicle dynamics data, in particular: S101, if the vehicle dynamics data is greater than the physical safety threshold then determine that there is an explicit physical risk; Otherwise, it enters step S102; S102、if real-time energy fraction greater than a physiological guard threshold a potential physiological risk state is determined. Otherwise, it is determined as a safe state.

Images