Automatic control method and system for steering lamp

Abstract

The invention relates to the technical field of automatic control, and discloses an automatic control method and system for a steering lamp, and the method comprises the steps: carrying out the orthogonal basis decomposition of three-axis acceleration data and three-axis angular velocity data of a seat scooter, and obtaining intention-related subspace data and environment-related subspace data; carrying out harmonic attitude decomposition and attention mechanism fusion to obtain a fusion feature vector; performing spatio-temporal context steering mode recognition to obtain an initial steering probability vector and a time continuity index; according to the method, accurate separation of a steering intention signal and an environment interference signal is achieved, the technical problem that intentional steering and road surface vibration interference cannot be accurately distinguished through a traditional method is effectively solved, and the method has the advantages that the steering intention signal and the environment interference signal can be accurately distinguished. And the false triggering rate is obviously reduced.

Description

Technical Field

[0001] The present invention relates to the field of automatic control technology, and particularly to an automatic control method and system for a turn signal. Background Art

[0002] Existing attitude sensing technologies face many technical problems when applied to the automatic control of turn signals for seat scooters. The traditional threshold judgment method cannot accurately distinguish the intentional turning of riders from environmental interferences such as road vibrations and bumps, resulting in frequent false triggers or missed detections. In addition, existing technologies often adopt an attitude analysis method with a single time window, which cannot balance the response speed and stability of the system at the same time. Especially in complex road conditions, the judgment accuracy of the system significantly decreases. Existing turn signal control systems usually adopt a binary logic judgment mode, ignoring the impact of dynamic changes in the riding scenario on turn judgment. For example, in different stages such as starting, uniform speed, and deceleration, the same body posture may represent different turning intentions. This judgment mechanism lacking situational awareness significantly limits the applicability and reliability of the automatic turn signal control system in complex environments. Summary of the Invention

[0003] The present invention provides an automatic control method and system for a turn signal. The present invention realizes the precise separation of turn intention signals and environmental interference signals, effectively solves the technical problem that traditional methods cannot accurately distinguish intentional turning from road vibration interference, and significantly reduces the false trigger rate.

[0004] In a first aspect, the present invention provides an automatic control method for a turn signal. The automatic control method for the turn signal includes: Performing orthogonal basis decomposition on the three-axis acceleration data and three-axis angular velocity data of the seat scooter to obtain intention-related subspace data and environment-related subspace data; Performing harmonic attitude decomposition and attention mechanism fusion on the intention-related subspace data to obtain a fused feature vector; Combining the environment-related subspace data to perform spatio-temporal context turn pattern recognition on the fused feature vector to obtain an initial turn probability vector and a time persistence index; Calculating the left turn probability, straight-ahead probability, right turn probability, and decision confidence according to the initial turn probability vector and the time persistence index, and sending a turn signal control instruction to the control unit of the seat scooter.

[0005] Combined with the first aspect, in the first implementation manner of the first aspect of the present invention, the performing orthogonal basis decomposition on the three-axis acceleration data and three-axis angular velocity data of the seat scooter to obtain intention-related subspace data and environment-related subspace data includes: Fix a six-axis attitude sensor at the center position of the seat scooter chassis and collect three-axis acceleration data and three-axis angular velocity data; Perform band-pass filtering on the three-axis acceleration data and the three-axis angular velocity data to obtain filtered attitude data, and perform segmentation processing and feature calculation on the filtered attitude data to obtain a target feature space; Based on the target feature space, perform Gram-Schmidt orthogonalization construction to obtain an intention-related orthogonal basis and an environment-related orthogonal basis; Based on the intention-related orthogonal basis and the environment-related orthogonal basis, perform orthogonal decomposition projection on the filtered attitude data to obtain intention-related subspace data and environment-related subspace data.

[0006] Combined with the first aspect, in the second implementation manner of the first aspect of the present invention, the performing harmonic attitude decomposition and attention mechanism fusion on the intention-related subspace data to obtain a fusion feature vector includes: Perform Gabor transform on the intention-related subspace data to obtain a time-frequency spectrum of attitude changes; Use a subspace projection algorithm to perform frequency band separation on the time-frequency spectrum of the attitude changes to obtain a steering intention signal component; Based on an intention enhancement adaptive operator, perform signal enhancement on the steering intention signal component to obtain an enhanced steering intention feature signal; Perform inverse Gabor transform processing on the enhanced steering intention feature signal to obtain a steering intention time series feature; Based on the steering intention time series feature, construct a steering intention feature vector including signal energy, main frequency, phase coherence, duration, waveform steepness, energy distribution skewness, kurtosis, and spectral entropy; Perform multi-scale feature analysis and attention mechanism fusion on the steering intention feature vector to obtain a fusion feature vector.

[0007] Combined with the first aspect, in the third implementation manner of the first aspect of the present invention, the performing multi-scale feature analysis and attention mechanism fusion on the steering intention feature vector to obtain a fusion feature vector includes: By setting a 0.5-second short-time window, a 2-second medium-time window, and a 5-second long-time window, perform sliding analysis on the steering intention feature vector to respectively obtain an initial short-time feature vector, an initial medium-time feature vector, and an initial long-time feature vector; Perform local feature enhancement on the initial short-time feature vector, the initial medium-time feature vector, and the initial long-time feature vector to obtain an enhanced short-time feature vector, an enhanced medium-time feature vector, and an enhanced long-time feature vector; Perform a non - linear transformation on the enhanced short - term feature vector, the enhanced medium - term feature vector, and the enhanced long - term feature vector to obtain a transformed short - term feature vector, a transformed medium - term feature vector, and a transformed long - term feature vector; Based on a soft attention model, calculate dynamic weight coefficients and perform weighted fusion on the transformed short - term feature vector, the transformed medium - term feature vector, and the transformed long - term feature vector to obtain a weighted fusion vector; Calculate the cross - correlation coefficient matrix between different time - scale features in the initial short - term feature vector, the initial medium - term feature vector, and the initial long - term feature vector, and perform feature consistency verification on the weighted fusion vector based on the cross - correlation coefficient matrix to obtain a fused feature vector.

[0008] Combined with the first aspect, in the fourth implementation manner of the first aspect of the present invention, the performing spatio - temporal context turning pattern recognition on the fused feature vector by combining the environment - related subspace data to obtain an initial turning probability vector and a time persistence index includes: Collect the speed data and vibration characteristic data of the seat scooter, and perform multi - feature analysis on the speed data and the vibration characteristic data through a DBSCAN clustering algorithm to obtain the current riding scenario category; Perform joint representation processing on the fused feature vector and the environment - related subspace data to obtain a comprehensive feature representation; Perform manifold space mapping on the comprehensive feature representation and the current riding scenario category to obtain a low - dimensional turning intention representation; According to the current riding scenario category, select a corresponding pre - trained deep neural network classifier to classify the low - dimensional turning intention representation to obtain an initial turning probability vector; Calculate the cosine similarity of the initial turning probability vector between consecutive frames. When the similarity is greater than a preset threshold and the duration exceeds a set threshold, it is confirmed as a valid turning intention to obtain a time persistence index.

[0009] Combined with the first aspect, in the fifth implementation manner of the first aspect of the present invention, the calculating the left - turn probability, the straight - ahead probability, the right - turn probability, and the decision confidence according to the initial turning probability vector and the time persistence index, and sending a turn - signal control instruction to the control unit of the seat scooter includes: Construct a time - series conditional probability model of the turning intention based on the initial turning probability vector and the time persistence index, and use a mixture Gaussian process to model the time - series conditional probability model to obtain a turning - intention conditional probability distribution in the continuous angle space; Perform kernel function enhancement processing on the turning - intention conditional probability distribution in the continuous angle space to obtain a target feature mapping relationship; Perform probability density inference based on the target feature mapping relationship to obtain the probability density function of the continuous steering angle; Discretize the probability density function of the continuous steering angle within the range of [-π / 2, π / 2] to obtain a steering intention density vector, and calculate the left-turn probability, straight-ahead probability, and right-turn probability through interval integration; Calculate the probability distribution entropy value of the steering intention density vector, evaluate the uncertainty degree of the steering decision, and obtain the decision confidence; Determine the optimal turn signal control strategy based on the left-turn probability, the straight-ahead probability, the right-turn probability, and the decision confidence, and send a turn signal control instruction to the control unit of the seat scooter.

[0010] Combined with the first aspect, in the sixth implementation manner of the first aspect of the present invention, the determining the optimal turn signal control strategy based on the left-turn probability, the straight-ahead probability, the right-turn probability, and the decision confidence, and sending a turn signal control instruction to the control unit of the seat scooter includes: Construct a turn signal control state transition function, where the turn signal control state transition function includes a state space of three states: left turn signal on, off, and right turn signal on, and an action space of three operations: activate the left turn signal, turn off the turn signal, and activate the right turn signal; Construct a composite reward function based on the left-turn probability, the straight-ahead probability, and the right-turn probability, where the composite reward function includes a steering intention matching degree reward component, an energy consumption penalty component, and a state switching frequency penalty component; Dynamically adjust the steering angle threshold according to the decision confidence to obtain a dynamic steering threshold; Based on the turn signal control state transition function, the composite reward function, and the dynamic steering threshold, solve the optimal turn signal control strategy through the Q-learning algorithm; Determine the turn signal control action to be executed according to the optimal turn signal control strategy, and send the turn signal control instruction corresponding to the turn signal control action to the control unit of the seat scooter through the CAN bus.

[0011] In the second aspect, the present invention provides an automatic control system for turn signals, and the automatic control system for turn signals includes: A decomposition module for performing orthogonal basis decomposition on the three-axis acceleration data and three-axis angular velocity data of the seat scooter to obtain intention-related subspace data and environment-related subspace data; A fusion module for performing harmonic attitude decomposition and attention mechanism fusion on the intention-related subspace data to obtain a fusion feature vector; An identification module, configured to perform spatio-temporal context turning pattern recognition on the fused feature vector by combining the environment-related subspace data, so as to obtain an initial turning probability vector and a time persistence index; A calculation module, configured to calculate a left-turn probability, a straight-ahead probability, a right-turn probability and a decision confidence level according to the initial turning probability vector and the time persistence index, and send a turn signal control instruction to a control unit of the seat scooter.

[0012] In the technical solution provided by the present invention, through the adoption of a harmonic attitude decomposition technique and a biorthogonal basis mapping algorithm, an accurate separation of a turning intention signal and an environmental interference signal is realized, effectively solving the technical problem that a traditional method cannot accurately distinguish an intentional turn from a road surface vibration interference, and significantly reducing a false triggering rate. Based on a multi-level dynamic time window feature fusion technique, attitude change patterns of a short-time window, a medium-time window and a long-time window are analyzed simultaneously, overcoming the technical contradiction that a single time scale cannot take into account both instantaneous responsiveness and anti-interference stability at the same time, enabling the system to quickly respond to a clear turning intention and resist short-term interference. Through a spatio-temporal context-aware turning pattern recognition framework, scene adaptive recognition is performed by combining riding state information, solving the problem of inaccurate judgment caused by ignoring riding context in the prior art, and maintaining a high turning intention recognition accuracy in different riding scenarios. By adopting a turn signal control system driven by probability density, a turning intention is represented as a continuous probability density function instead of a simple yes / no judgment, breaking through the limitation of a traditional binary logic control mode and providing a more refined and progressive expression of a turning intention. Through a dynamic threshold adjustment mechanism driven by decision confidence level, a judgment threshold is increased when uncertainty is high, and a response threshold is decreased when certainty is high, enhancing the robustness and adaptability of the system in various complex environments. Description of the Drawings

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for description in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0014] Figure 1 It is a step schematic diagram of an automatic control method for a turn signal in an embodiment of the present invention; Figure 2 It is a structural schematic diagram of an automatic control system for a turn signal in an embodiment of the present invention. Detailed Embodiments

[0015] An embodiment of the present invention provides an automatic control method and system for a turn signal. The terms "first", "second", "third", "fourth", etc. (if any) in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described herein can be implemented in an order different from that shown or described herein. In addition, the term "comprising" or "having" and any variation thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0016] For ease of understanding, the specific process of the embodiment of the present invention will be described below. Please refer to Figure 1 , an embodiment of the automatic control method for the turn signal in the embodiment of the present invention includes: Step S1: Perform orthogonal basis decomposition on the three-axis acceleration data and three-axis angular velocity data of the seat scooter to obtain intention-related subspace data and environment-related subspace data; It can be understood that the execution subject of the present invention can be an automatic control system for the turn signal, or a terminal or a server. Specifically, it is not limited here. The embodiment of the present invention takes the server as the execution subject as an example for illustration.

[0017] Specifically, a six-axis attitude sensor is installed at the center of the seat scooter chassis. During operation, this sensor continuously synchronously collects triaxial acceleration and triaxial angular velocity information, constituting a dynamic data stream describing the full-body attitude changes of the rider. Since such raw data simultaneously contains various mixed components from the rider's active operations, ground bumps, braking impacts, inertial responses, and electrical noise, frequency-domain cleaning is performed on the triaxial acceleration data and triaxial angular velocity data to exclude irrelevant disturbances. On this basis, band-pass filtering technology is used for processing. The passband design of the filter has been specifically debugged to be able to retain the slow and continuous attitude movements generated by the rider while effectively suppressing extremely low-frequency gravity drift and high-frequency electronic interference. The filtered attitude data is processed in segments. Each segment of data contains a fixed number of sampling points and is in the form of a sliding window to ensure time continuity. On each segment of data, statistical analysis and feature extraction are performed, including key indicators such as energy change trends, main direction stability, and co-movement amplitude. These indicators together constitute the target feature space of the attitude behavior during this time period. Based on the target feature space, an orthogonal basis set is constructed. Using an improved Gram-Schmidt orthogonalization method, through a step-by-step vector orthogonalization and direction correction process, two sets of mutually uncorrelated attitude change bases are constructed. One set is highly correlated with the rider's body movement characteristics and is regarded as the intention-related orthogonal basis, while the other set is more closely related to environmental disturbances and equipment structure responses and is thus defined as the environment-related orthogonal basis. Based on the intention-related orthogonal basis and the environment-related orthogonal basis, orthogonal decomposition projection is performed on the filtered attitude data to achieve spatial decoupling and property classification of the original mixed signal, obtaining two mutually orthogonal but information-complementary data subsets. The intention-related subspace reflects the possible current turning behavior tendency of the user, while the environment-related subspace contains background information from road conditions, vehicle vibrations, and inertial disturbances.

[0018] Step S2: Perform harmonic attitude decomposition and attention mechanism fusion on the intention-related subspace data to obtain a fused feature vector; Specifically, time-frequency characteristic analysis is performed on the attitude data in the intention-related subspace. The Gabor transform is used to process the data. This transform method can retain both time and frequency information simultaneously and is suitable for analyzing the dynamic change characteristics in non-stationary signals. By introducing an adjustable window function, the Gabor transform flexibly adjusts the time and frequency resolution capabilities according to the local characteristics of the data change, thereby generating an attitude time-frequency spectrum covering the entire process. This time-frequency spectrum records the details of the rider's attitude changes at different time periods and is an important basis for identifying the steering intention. The subspace projection algorithm is used to separate the frequency bands of the attitude time-frequency spectrum, and representative steering intention signal components are extracted from it. After separating the target frequency band, an intention-enhancing adaptive operator is introduced to enhance the signal components. In the enhancement process, the consistency of the signal in time and the persistence of energy are comprehensively considered, and the characteristic signals with stable structures and clear trends are enhanced through an adaptive adjustment method, making the steering intention more prominent and significant. The enhanced steering intention characteristic signal is processed by the inverse Gabor transform to restore the time-frequency domain signal to the time domain form, obtaining continuous steering intention time-series characteristics. Based on this time-series signal, several key characteristic parameters are extracted, including indicators such as the overall energy of the signal, the main frequency components, phase consistency, action duration, waveform steepness, energy distribution offset, peak characteristics, and spectral entropy. To improve the system's recognition ability at different time scales, a multi-scale feature analysis mechanism is introduced, and the above features are processed separately in three time windows: short-term, medium-term, and long-term, so as to capture different levels of information such as fast actions, stable trends, and long-term offsets. On this basis, the feature vectors of the three time scales are weighted and fused through an attention mechanism. During the fusion process, the importance of the features of each time scale is evaluated, and different attention weights are assigned, so as to strengthen the key features, suppress redundant interference, and finally generate a fused feature vector.

[0019] Three representative time windows are set, namely a short-time window of 0.5 seconds, a medium-time window of 2 seconds, and a long-time window of 5 seconds, to cover the behavioral characteristics of the rider's posture changes at different time scales. These three time windows continuously analyze the steering intention feature vector in a sliding manner, and respectively extract the initial short-time feature vector, the initial medium-time feature vector, and the initial long-time feature vector, which are used to reflect the feature differences at the three levels of instantaneous response, process dynamics, and long-term trend. The initial feature vectors at the three time scales are respectively subjected to local feature enhancement processing. In the enhancement process, the key elements in the features are prominently expressed by introducing spatial convolution or weighted amplification mechanisms, so as to improve the discrimination ability at this scale, and the enhanced short-time feature vector, the enhanced medium-time feature vector, and the enhanced long-time feature vector are obtained. The enhanced feature vectors at each scale are input into the non-linear transformation module for processing. This module uses an activation function with a threshold to strengthen the non-linear expression ability of the features and modulate the distribution form of the features, so that the transformed short-time, medium-time, and long-time feature vectors are more dispersed in the feature space, making it easier for the attention mechanism to effectively distinguish them. By constructing a soft attention model, the dynamic weight coefficients of the three transformed feature vectors are calculated and fused. The attention mechanism automatically allocates the fusion weights according to the significance of each scale feature in the current task, ensuring that the key scale information is preferentially retained while the influence of redundant or conflicting features is suppressed. In this way, a weighted fusion vector is generated, which fully synthesizes the most representative information components in the three time levels, and realizes overall stability while retaining local sensitivity. To ensure the reliability of the fusion result in terms of time consistency, the correlation relationship between the initial short-time, medium-time, and long-time feature vectors is evaluated, and the cross-correlation coefficient matrix between the three is calculated to quantify the coupling degree and mutual support between the features at different time scales. If the overall cross-correlation coefficient is at a high consistency level, it indicates that the current fusion structure is credible; if there is a low consistency section, the corresponding feature components in the weighted fusion vector are dynamically adjusted, so as to eliminate the influence of local anomalies and enhance the robustness of the overall judgment, and finally the fusion feature vector is obtained.

[0020] Step S3: Combine the environmental related subspace data to perform spatio-temporal context steering pattern recognition on the fusion feature vector, and obtain the initial steering probability vector and the time persistence index; Specifically, speed data and vibration characteristic data are collected in real time during the operation of the seat scooter to reflect the movement amplitude of the current riding state and the road surface condition. Based on these raw data, a set of feature vectors including the rate of speed change, vibration frequency distribution, acceleration change trend, etc. are constructed through multi-dimensional feature extraction means, and the DBSCAN clustering algorithm is used to perform clustering analysis on them. This algorithm divides the data into multiple typical scenario categories through the density connectivity recognition method, such as stationary, straight-line uniform speed, acceleration, deceleration, low-speed turning, high-speed turning, road surface bump, and emergency obstacle avoidance, etc., so as to obtain the specific category of the current riding scenario. While completing the riding scenario recognition, the fused feature vector and the environment-related subspace data are jointly represented and processed. The two are complementary in information structure. The fused feature mainly expresses the behavior intention of the rider, while the environment subspace contains external disturbances and vehicle body response characteristics. The combination of the two forms a more comprehensive integrated feature representation. A mapping mechanism based on manifold learning is introduced, and the integrated feature representation and the currently recognized riding scenario category are input into the mapping model and processed in the non-linear manifold space to generate a low-dimensional steering intention representation with a compact structure and compressed dimensions. Based on the low-dimensional steering intention representation, according to the current riding scenario category, a pre-trained deep neural network classifier matching the scenario is automatically selected. Each scenario category corresponds to a dedicated neural network classification model. All model structures are the same but the parameter weights are different, and they have targeted learning capabilities. The low-dimensional intention representation is input into the selected classifier to perform classification operations, and an initial steering probability vector is output, which respectively contains the probability estimates of turning left, going straight, and turning right. This vector can dynamically reflect the possible operation intention of the user at the current moment. To improve the recognition stability and anti-fluctuation ability, a time consistency judgment mechanism is introduced to calculate the cosine similarity of the initial steering probability vectors between consecutive frames and monitor the time coherence of the intention judgment. When the similarity in a certain direction continuously exceeds the preset threshold and this state lasts for more than the lower limit of the time set by the system, it is confirmed that the steering behavior is a real and effective user intention, and the system generates a time persistence index accordingly.

[0021] Step S4: Calculate the probabilities of turning left, going straight, and turning right and the decision confidence based on the initial steering probability vector and the time persistence index, and send a turn signal control instruction to the control unit of the seat scooter.

[0022] Specifically, a time-series conditional probability model reflecting the dynamic change process of the steering intention is constructed. Based on the initial steering probability vector, this model introduces a time persistence index as an important factor for judging the signal stability to express the continuous trend and consistency of the current behavior in the time dimension. To improve the expression ability of this conditional probability model in non-linear complex behavior changes, a mixture Gaussian process is introduced as a modeling method, and the steering intention is regarded as a conditional probability distribution in the continuous angle space. The mixture Gaussian process allows the model to depict both local trends and global changes simultaneously. Through the joint modeling of multiple Gaussian kernel functions, the model can maintain sufficient sensitivity and generalization ability in different steering situations, so as to describe the behavior probability at continuous angles between left turn and right turn. After obtaining the preliminary conditional probability distribution through modeling, kernel function enhancement processing is performed on this distribution. By introducing a deep feature mapping mechanism, the local correlation in the high-dimensional feature space is modeled in combination with a neural network, and the original input is mapped to a target feature space with higher discriminant ability through non-linear transformation to obtain a more representative target feature mapping relationship. Based on the target feature mapping relationship, probability density inference is carried out to obtain the probability density function of the steering intention over the entire continuous angle interval. This function takes the angle as the independent variable and describes the probability distribution characteristics of the possible steering behavior of the user in each angle direction. To enable the probability density function to serve the generation of specific control instructions, it is discretized within the range of [-π / 2, π / 2] to form a discretized steering intention density vector, and each element represents the probability density value of an angle interval. On this basis, according to the set angle boundary threshold, interval integration is performed on the density vector to calculate the total probability values representing the three behavior trends of left turn, straight ahead, and right turn respectively, and the left turn probability, straight ahead probability, and right turn probability are obtained. At the same time, to judge the uncertainty level of the current decision-making process, the entropy value of the entire steering intention density vector is calculated to obtain its probability distribution entropy value as a measure of the decision confidence. A lower entropy value indicates that the system is more certain about the current judgment, and vice versa indicates ambiguity in the intention or unstable signal. The above-mentioned three types of behavior probabilities and decision confidence are comprehensively analyzed, and the current optimal turn signal control strategy is determined through a preset multi-objective decision function. In the decision function, consider which direction has the largest probability value and adjust the trigger threshold of the control instruction in combination with the confidence to achieve a balance between high accuracy and high stability. When the strategy is determined, the system sends the corresponding turn signal control instruction to the control unit of the seat scooter through the vehicle-mounted communication interface to achieve automatic response of the left turn light, right turn light, or light off.

[0023] Build a state transition model to describe the behavioral evolution process of the turn signal under different control states. This model includes three basic turn signal states, namely the left turn signal on, the light off, and the right turn signal on, which together constitute the state space of the system; at the same time, three corresponding control actions are set, namely activating the left turn signal, turning off the turn signal, and activating the right turn signal, as the action space of the system. Based on this combination relationship between states and actions, establish a state transition function to describe the target state that the current state may transfer to after executing a certain action, and introduce a reinforcement learning mechanism on this structure to optimize the control strategy. To achieve an accurate response to the turning behavior, construct a composite reward function based on the left turn probability, straight-ahead probability, and right turn probability obtained in the previous step. The design of this reward function takes into account multiple factors such as behavior matching degree, system energy efficiency, and control stability. Among them, the steering intention matching degree reward component is used to measure whether the current control state conforms to the user's true intention. If the actual control state is consistent with the highest probability behavior direction, a higher reward is given; the energy consumption penalty component is used to suppress the energy waste caused by frequent switching of the light state, especially in high-frequency control scenarios, where the energy consumption of ineffective actions needs to be reduced; the state switching frequency penalty component focuses on the stability of the system response. When the control state fluctuates frequently in a short period of time, the overall score will be reduced, thus guiding the strategy to be more stable. In the control logic, the decision confidence is used to dynamically adjust the determination threshold of the steering angle, that is, the sensitivity standard for determining that a steering action is effective will be automatically corrected according to the confidence. When the confidence is high, the system uses a lower steering angle threshold to quickly respond to the behavior intention; when the confidence is low, the system will increase the steering angle threshold to avoid misjudgment caused by unstable signals. The dynamic steering threshold mechanism enhances the adaptability of the system under different behavior clarity conditions, ensuring the robustness and reliability of the control strategy. Based on the above state transition function, composite reward function, and dynamic steering threshold, the optimal control strategy is iteratively solved through the Q-learning algorithm. During the learning process, calculate the corresponding Q value for each state-action pair, and continuously update the Q value function to approximate the global optimal strategy. After the strategy converges, select the action with the maximum expected return according to the current state as the control operation to be executed currently. Select the control instruction to be executed according to the action determined by the optimal turn signal control strategy, and send this instruction to the control unit of the seat scooter through the CAN bus to drive the left turn signal, right turn signal, or turn them off, so as to achieve automatic, dynamic, and context-aware turn signal control.

[0024] In the embodiment of the present invention, by adopting the harmonic attitude decomposition technology and the biorthogonal basis mapping algorithm, the accurate separation of the turning intention signal and the environmental interference signal is achieved, which effectively solves the technical problem that the traditional method cannot accurately distinguish between the intentional turning and the road vibration interference, and significantly reduces the false trigger rate. Based on the multi-level dynamic time window feature fusion technology, the attitude change patterns of the short-time window, the medium-time window and the long-time window are analyzed simultaneously, which overcomes the technical contradiction that a single time scale cannot take into account both the instantaneous responsiveness and the anti-interference stability at the same time, so that the system can respond quickly to clear turning intentions and resist short-term interference. Through the spatiotemporal context-aware steering pattern recognition framework, the scene adaptive recognition is combined with the riding state information, which solves the problem of inaccurate judgment caused by ignoring the riding context in the prior art, and maintains a high accuracy of steering intention recognition in different riding scenarios. The turn signal control system driven by probability density expresses the turning intention as a continuous probability density function instead of a simple yes / no judgment, breaking through the limitations of the traditional binary logic control mode and providing a more refined and progressive expression of steering intention. Through the dynamic threshold adjustment mechanism driven by decision confidence, the judgment threshold is raised when the uncertainty is high, and the response threshold is lowered when the certainty is high, thereby enhancing the robustness and adaptability of the system in various complex environments.

[0025] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Fix the six-axis attitude sensor at the center of the seat scooter chassis and collect three-axis acceleration data and three-axis angular velocity data; Perform bandpass filtering on the three-axis acceleration data and the three-axis angular velocity data to obtain filtered attitude data, and perform segmentation processing and feature calculation on the filtered attitude data to obtain the target feature space; Based on the target feature space, Gram-Schmidt orthogonalization is performed to obtain the intention-related orthogonal basis and the environment-related orthogonal basis; Based on the intention-related orthogonal basis and the environment-related orthogonal basis, the filtered posture data is orthogonally decomposed and projected to obtain the intention-related subspace data and the environment-related subspace data.

[0026] Specifically, to ensure that the collected data is symmetric and not affected by the chassis structure disturbance, the six-axis attitude sensor is fixed at the center position of the seat scooter chassis, which is conducive to obtaining the central data representing the movement trend of the whole vehicle and minimizing the error diffusion caused by the body structure offset or local vibration. The original data of acceleration and angular velocity are continuously collected at a certain sampling frequency to obtain a six-dimensional attitude vector sequence, where each frame of data is an accurate reflection of the vehicle's attitude state at the current moment. In addition to the attitude changes caused by the driver's active control, the original signal is also mixed with ground vibration, body elastic feedback, electromagnetic interference, as well as the offset and noise generated by the sensor itself. Band-pass filtering is performed on all six-dimensional data to filter out low-frequency drift and high-frequency interference while retaining the frequency band of the human driving signal. The passband of the band-pass filter is designed based on experimental verification results to be between 0.2 Hz and 42 Hz. This frequency band can cover the main energy frequency band exhibited by the user's attitude steering actions (roughly concentrated between 0.3 and 2.5 Hz), and at the same time suppress the high-frequency part (such as frictional vibration, electrical interference) and the low-frequency part (such as gravity slow drift), making the filtered data more real and reliable. After filtering, to extract the stable structural features in the time-series data, the six-dimensional data stream is divided into multiple consecutive time periods. Each segment of data is generated by intercepting through a sliding window. The window length is set to 256 sampling points, and an overlapping mechanism (such as 75%) is introduced to ensure time continuity and the smoothness of state transition. In each segmented window, feature extraction operations are performed. The local feature space of each segment of data is mainly calculated through statistical and structural analysis means, including variance, covariance matrix, main direction vector, and its correlation index. These indicators reflect the trend, amplitude, and change path of the attitude change in the current time period, helping to identify which data main directions are highly correlated with the user's intention and which are more reflected as random noise or environmental disturbances. After obtaining the feature description of each segment, the Gram-Schmidt orthogonalization construction process is performed based on the target feature space. The multi-dimensional vectors in the feature space are converted into a set of mutually independent orthogonal bases through orthogonal projection. The system normalizes the local feature vectors and introduces them in order of importance. Using the step-by-step orthogonal projection and residual minimization strategy, new basis vectors are constructed, thus deconstructing the original coupled multi-dimensional information into a set of orthogonal bases with clear structure and independent directions. During the construction process, the meaning of each orthogonal basis is judged according to the frequency distribution, energy concentration, and time-series stability. The components that are mainly concentrated in the human natural steering frequency range and have time continuity and phase consistency are classified into the intention-related orthogonal basis; while the components with high frequency, fast change, violent fluctuation, or unstable structure are classified into the environment-related orthogonal basis. After completing the construction of the orthogonal bases, orthogonal projection operations are performed on the filtered six-dimensional attitude data based on these two orthogonal basis groups, and the original signal is projected into these two subspaces respectively.The result of projecting onto the intention-related orthogonal basis forms the intention-related subspace data, which mainly contains the control posture signals exerted by the occupant through body movements and is the main basis for subsequent steering intention recognition and behavior modeling. The result of projecting onto the environment-related orthogonal basis constitutes the environment-related subspace data, which contains a large amount of background disturbance information caused by uneven ground, braking impact, acceleration jitter, etc.

[0027] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Perform Gabor transform on the intention-related subspace data to obtain the time-frequency spectrum of the posture change; Use the subspace projection algorithm to perform frequency band separation on the time-frequency spectrum of the posture change to obtain the steering intention signal component; Based on the intention-enhanced adaptive operator, perform signal enhancement on the steering intention signal component to obtain the enhanced steering intention feature signal; Perform inverse Gabor transform processing on the enhanced steering intention feature signal to obtain the steering intention time series feature; Based on the steering intention time series feature, construct a steering intention feature vector including signal energy, main frequency, phase coherence, duration, waveform steepness, energy distribution skewness, kurtosis, and spectral entropy; Perform multi-scale feature analysis and attention mechanism fusion on the steering intention feature vector to obtain the fused feature vector.

[0028] Specifically, the intention-related subspace data is essentially the component of human control intention stripped from the original posture signal. This type of data has characteristics such as low frequency, non-stationarity, and strong persistence, and is suitable for analysis using transformation methods with time-frequency localization characteristics. Therefore, the Gabor transform is selected as the time-frequency mapping means. By constructing a set of composite kernel functions with adjustable center frequencies and window widths, the posture signal is mapped into a two-dimensional time-frequency spectrogram, where the signal structure at each moment is not only expressed as amplitude but also further characterized by its frequency distribution to depict its change trend. The advantage of the Gabor transform lies in its ability to flexibly balance between time accuracy and frequency resolution, enabling the accurate capture of small changes in the signal over different time periods. It is suitable for detecting slow-varying patterns and phase trends in temporal behaviors such as steering behaviors. The subspace projection algorithm is used to perform band separation on the time-frequency spectrum of posture changes. Based on behavioral research experience and statistical analysis of frequency domain characteristics, the intention-related action signals are mainly concentrated between 0.3 Hz and 2.5 Hz. Based on this, a characteristic subspace matching this frequency band is constructed on the entire time-frequency spectrum. Through the method of orthogonal projection, the energy components within this frequency band are extracted, and the main frequency signal segment representing the steering action trend is stripped out to obtain the steering intention signal component. The steering intention signal component is enhanced based on an intention-enhanced adaptive operator. This operator dynamically adjusts the expression intensity of each time-frequency point by simultaneously measuring the consistency of the signal in the time dimension and the energy persistence in the frequency dimension. The consistency index measures the smoothness of the phase and amplitude between adjacent time slices, and the persistence index measures whether the signal in this frequency band exists for a long time in the time dimension. When a certain region has both good phase coherence and energy persistence, a higher weight is assigned to this region, thereby enhancing the expression intensity of this segment of the signal, while short-term abrupt, broken, or isolated interference information is automatically suppressed. The enhanced steering intention feature signal is subjected to inverse Gabor transform processing to restore it from the time-frequency domain to the time-domain form, obtaining a continuous and clear steering intention time-series feature signal. On this basis, high-dimensional feature vectors for quantifying behavioral characteristics are extracted, including statistical and structural description indicators in multiple dimensions. The overall energy of the signal is calculated to measure the intensity of the action; the main frequency is used to express the repeatability and rhythm of the action; at the same time, including phase coherence and action duration, which are used to characterize the structural stability of the action; to describe the asymmetry and mutation degree of the signal, high-order statistics reflecting the signal morphological structure such as waveform steepness, energy distribution skewness, and kurtosis are extracted; and the spectral entropy is used to describe the complexity and information content distribution of the signal. All these indicators together constitute the steering intention feature vector, covering multiple dimensions from macroscopic trends to microscopic perturbations. For feature extraction, to enhance the adaptability of the model to actions of different durations, this feature vector is analyzed on multiple time scales.A short - time window is set to capture sudden reaction behaviors, a medium - time window is used to identify the complete turning process, and a long - time window is used to identify trend behaviors and context associations. The features at each scale are processed through local enhancement and non - linear transformation to ensure their expression intensity and discriminative ability. The attention mechanism is used to weight - fuse the feature vectors at different scales, automatically assigning weights to each feature component according to the stability, discriminability, and temporal consistency of the current data, forming a fused feature vector with global consistency and local sensitivity.

[0029] In a specific embodiment, the process of performing multi - scale feature analysis and attention - mechanism fusion on the steering - intention feature vector to obtain the fused feature vector may specifically include the following steps: By setting a 0.5 - second short - time window, a 2 - second medium - time window, and a 5 - second long - time window, a sliding analysis is performed on the steering - intention feature vector to respectively obtain an initial short - time feature vector, an initial medium - time feature vector, and an initial long - time feature vector; Perform local feature enhancement on the initial short - time feature vector, the initial medium - time feature vector, and the initial long - time feature vector to obtain an enhanced short - time feature vector, an enhanced medium - time feature vector, and an enhanced long - time feature vector; Perform non - linear transformation processing on the enhanced short - time feature vector, the enhanced medium - time feature vector, and the enhanced long - time feature vector to obtain a transformed short - time feature vector, a transformed medium - time feature vector, and a transformed long - time feature vector; Based on the soft - attention model, perform dynamic weight - coefficient calculation and weighted fusion on the transformed short - time feature vector, the transformed medium - time feature vector, and the transformed long - time feature vector to obtain a weighted - fusion vector; Calculate the cross - correlation coefficient matrix between the different - time - scale features in the initial short - time feature vector, the initial medium - time feature vector, and the initial long - time feature vector, and perform feature - consistency verification on the weighted - fusion vector based on the cross - correlation coefficient matrix to obtain the fused feature vector.

[0030] Specifically, to capture the change patterns of steering intentions in different time dimensions, three sliding analysis windows with representative time lengths are set, namely a short-term window of 0.5 seconds, a medium-term window of 2 seconds, and a long-term window of 5 seconds. Each window uses an equally spaced sliding method to traverse and analyze the real-time generated steering intention feature vectors, ensuring continuous tracking and time coverage of the behavior evolution process. In the short-term window, the system pays more attention to capturing rapid dynamic responses such as a sudden increase in steering intention, a drastic change in phase, or a steep increase in waveform slope; in the medium-term window, the system extracts the energy trend, structural continuity, and rhythm pattern of the complete steering process; while in the long-term window, the system analyzes the global fluctuation trend of steering intention, the state retention characteristics, and the mapping relationship between the historical behavior trajectory and the current state, thereby obtaining the initial short-term feature vector, the initial medium-term feature vector, and the initial long-term feature vector respectively. Local feature enhancement processing is performed on the initial feature vectors at these three scales to highlight the most discriminative feature dimensions at each time scale. The enhancement methods mainly include amplitude amplification of key dimensions, frequency response enhancement, and correlation amplification between adjacent features. Through convolution or filter bank methods, higher expression weights are assigned to feature segments with high energy concentration, strong phase continuity, or stable spectra, thereby enhancing the dominance of these high-confidence regions over the overall structure. After the enhancement processing, the enhanced short-term feature vector, the enhanced medium-term feature vector, and the enhanced long-term feature vector are obtained respectively. Nonlinear transformation operations are sequentially performed on the three types of enhanced feature vectors. Nonlinear mapping can expand the expression ability of the original features and improve the separability between features. This transformation is achieved through activation functions. For example, by using nonlinear functions with saturation segments such as LeakyReLU or Swish, the system can enhance the boundary contrast between features, amplify the response differences caused by small perturbations, and compress redundant or repeated signal regions, obtaining the transformed short-term feature vector, the transformed medium-term feature vector, and the transformed long-term feature vector respectively. Based on the above nonlinear processing results, a soft attention mechanism is introduced to fuse and model the transformed feature vectors at three time scales. The soft attention model dynamically calculates the weight coefficients of each scale feature vector by comparing the salience of features at each scale in terms of expression intensity, structural clarity, and time stability, and uses them as the weight basis for that scale in the final fusion expression. The attention mechanism models the importance of features at each scale by introducing a learnable parameter matrix. The system uses these dynamic weights to perform weighted fusion on the three types of feature vectors to obtain a weighted fusion vector. To ensure the stability of the fusion result in the time dimension and cross-scale consistency, cross-correlation analysis is performed on the correlation relationships between the initial short-term, medium-term, and long-term feature vectors. By calculating the cross-correlation coefficient matrix among the three, the feature coordination degree, change coupling strength, and phase synchronization degree between each time scale are quantified.If a time-scale feature is strongly consistent with other scales in structure, its contribution will be retained and enhanced in the final fusion vector; if there is a scale feature with significantly low correlation with other scales, the interference effect of this redundant or abnormal feature will be suppressed by reducing its attention weight. Feature consistency verification is performed on the weighted fusion vector based on the cross-correlation coefficient matrix, and after the verification passes, it is output as the final fusion feature vector.

[0031] In a specific embodiment, the process of executing step S3 may specifically include the following steps: Collect the speed data and vibration characteristic data of the seat scooter, and perform multi-feature analysis on the speed data and vibration characteristic data through the DBSCAN clustering algorithm to obtain the current riding scenario category; Perform joint representation processing on the fusion feature vector and the environment-related subspace data to obtain a comprehensive feature representation; Perform manifold space mapping on the comprehensive feature representation and the current riding scenario category to obtain a low-dimensional steering intention representation; According to the current riding scenario category, select the corresponding pre-trained deep neural network classifier to classify the low-dimensional steering intention representation to obtain an initial steering probability vector; Calculate the cosine similarity of the initial steering probability vector between consecutive frames. When the similarity is greater than the preset threshold and the duration exceeds the set threshold, it is confirmed as a valid steering intention to obtain a time persistence index.

[0032] Specifically, speed data and vibration characteristic data during the operation of the seat scooter are collected in real time. These two types of information together constitute physical indicators related to ground conditions, acceleration state, and operation stability during riding. The speed data is obtained by a wheel speed sensor or an encoder, reflecting the current driving state and dynamic change trend of the vehicle; while the vibration characteristic data comes from accelerometers installed on the chassis or seat structure, and the high-frequency vibration components collected by them can better reveal road surface flatness, obstacle response, and vibration feedback caused by acceleration and deceleration. To make full use of these data, a multi-dimensional feature set is constructed, integrating multiple dimensional indicators including instantaneous speed, acceleration change rate, vibration intensity, vibration frequency components, kurtosis, and skewness, etc., to characterize the current riding state. After obtaining the above features, the DBSCAN clustering algorithm is used to perform unsupervised clustering on all feature samples. This algorithm identifies high-density aggregation points and eliminates outliers through density neighborhood analysis, and does not depend on the preset number of categories, so it is suitable for actual scenarios with non-linear and non-uniform distributions. During the DBSCAN clustering process, appropriate neighborhood distance thresholds and minimum point number parameters are set to enable it to effectively identify typical scenario categories during riding, such as stationary, starting, accelerating, straight-line uniform speed, decelerating, driving on a bumpy road, high-speed turning, emergency obstacle avoidance, etc. Each category reflects different vehicle states and driver operation characteristics. After identifying the current riding scenario, the previously extracted fused feature vector and environment-related subspace data are jointly represented. The fused feature vector contains the behavioral structure of the steering intention feature on multiple time scales, while the environment-related subspace data records the vibrations, interferences, or background dynamics caused by non-intentional factors in the current path. The two together constitute a complete context behavior representation. These two vectors are synthesized into a unified comprehensive feature representation through concatenation, weighting, or embedding methods. The joint manifold mapping process is performed on the comprehensive feature representation and the current scenario category to embed the high-dimensional complex features into a low-dimensional structure space to extract the essential behavior representation structure. For this purpose, a manifold learning method under Lie group algebra is introduced, and the transformation between the tangent space and the manifold space is realized through logarithmic mapping and exponential mapping respectively, so that the original feature structure reduces the overall dimension while maintaining the local relative geometric relationship. And the non-linear boundary expression ability is enhanced through the kernel function-assisted mapping method to obtain a low-dimensional steering intention representation with a compact structure and strong discriminability. According to the currently identified riding scenario category, a pre-trained deep neural network classifier corresponding to it is selected from a preset multi-classifier model library to classify and judge this intention representation. Each classifier is trained for a specific scenario, with a unified structure but different weight parameters, and has the ability to adapt to signal feature differences in different riding environments. After the classifier inputs the low-dimensional intention representation, it outputs a three-dimensional probability vector, corresponding to the prediction results of the three behavior possibilities of turning left, going straight, and turning right respectively. This vector is the initial steering probability vector.To improve the system's judgment robustness and anti-jitter ability in the time dimension, a cosine similarity calculation mechanism for continuous frame probability vectors is introduced to analyze the behavior continuity between different time points. In this process, the turning probability vector of the current frame is evaluated for similarity with the vector of the previous frame, and the cosine similarity is used to measure their consistency in the direction space. When the similarity exceeds the system-set threshold and this high-similarity state persists in time for more than a set number of frames or time period threshold, the system confirms that this turning intention has a stable structure and is a real and continuous operation behavior, rather than a short-term anomaly caused by random fluctuations or external disturbances. The output of this judgment is the time persistence index.

[0033] In a specific embodiment, the process of executing step S4 may specifically include the following steps: Construct a time series conditional probability model of the turning intention based on the initial turning probability vector and the time persistence index, and use a mixture Gaussian process to model the time series conditional probability model to obtain the conditional probability distribution of the turning intention in the continuous angle space; Perform kernel function enhancement processing on the conditional probability distribution of the turning intention in the continuous angle space to obtain the target feature mapping relationship; Perform probability density inference based on the target feature mapping relationship to obtain the probability density function of the continuous turning angle; Discretize the probability density function of the continuous turning angle in the range of [-π / 2, π / 2] to obtain the turning intention density vector, and calculate the left-turn probability, straight-ahead probability, and right-turn probability through interval integration; Calculate the probability distribution entropy value of the turning intention density vector to evaluate the uncertainty degree of the turning decision and obtain the decision confidence; Determine the optimal turn signal control strategy based on the left-turn probability, straight-ahead probability, right-turn probability, and decision confidence, and send a turn signal control instruction to the control unit of the seat scooter.

[0034] Specifically, the initial turning probability vector and the time persistence index within the corresponding time frame are combined into multi-moment input samples to construct a time series conditional probability model for turning intention. The initial turning probability vector consists of instantaneous probability values in three directions: left turn, straight ahead, and right turn. The time persistence index reflects whether the current probability trend is continuous and stable. Therefore, this combination captures the behavioral tendency at a certain moment and introduces a reference for cross-time behavioral consistency. When constructing the time series structure, the probability-stability combinations within several consecutive historical frames are regarded as time-related observation sequences, and based on this, a dynamic conditional distribution is constructed to predict the specific turning trend probability in the continuous angular space at the current moment. To perform high-precision modeling of the time series conditional probability model, a mixture Gaussian process is used as the probability regression framework. While fully utilizing the mapping relationship between the input sequence and the target angle response in the training samples, a kernel function structure is introduced to model non-linear changes. The mixture Gaussian process contains the definitions of the mean function and covariance function in the traditional Gaussian process and allows the system to construct a linear combination structure of local Gaussian kernels on each input sequence segment, thus more flexibly fitting the distribution pattern of turning intention in the continuous angular space. In this process, the angular space is regarded as a continuous interval from -90 degrees to +90 degrees, and the model performs probability prediction on the intention response of each angle point within this interval through the Gaussian process, forming a conditional probability distribution surface. To improve the identification accuracy of the model in the boundary response, angle-dense regions, and noise sections, kernel function enhancement processing is performed on the above conditional probability distribution. This processing maps the data in the original input space to a high-dimensional non-linear feature space by introducing a deep feature mapping mechanism, and a new kernel function is constructed in this space to capture more complex similarities between inputs. A residual neural network structure is used to perform end-to-end encoding on the input sequence, making the generated feature space have stronger expressiveness and discriminative ability. Furthermore, a new non-linear expression dimension is introduced into the Gaussian kernel structure to obtain an enhanced target feature mapping relationship. Using the target feature mapping relationship, probability density inference is performed on the continuous angular space. The inference result is a continuous probability density function defined within the interval [-π / 2, π / 2], which gives the probability value that the user may turn to this angle at each angle point. This density function is more continuous and interpretable than the traditional three-class probability and is suitable for generating high-precision control strategies. To make this function applicable to subsequent discrete decision-making processing, uniform discretization processing is performed on it in the angular space. The angular space is divided into several equally spaced angular segments, and the probability density integral within each segment interval is calculated to form a turning intention density vector. This density vector is a one-dimensional array of fixed length, where each item represents the probability mass within a certain angular segment. According to the set angular threshold, the entire density vector is divided into three interval segments: left turn, straight ahead, and right turn. Numerical integration operations are performed on each segment respectively to obtain three types of values: left turn probability, straight ahead probability, and right turn probability.This operation is achieved by the cumulative integration of the probability density function over different angular intervals, such that the probability of each type of behavior direction is determined not only by a single point value estimate, but jointly by the distribution trend of the entire angular interval, enhancing the stability and overall reliability of the probability expression. Meanwhile, the system calculates the probability distribution entropy of the entire density vector to quantify the degree of uncertainty in the current steering intention expression. The higher the probability distribution entropy value, the closer the density function is to a uniform distribution, indicating that the system has a high degree of uncertainty about the user's intention; conversely, when the entropy value is low, the density function has an obvious concentration in a certain direction, indicating that the current prediction result has a high confidence level. This entropy value will be used as a confidence index in the final control decision-making process to dynamically adjust the strategy trigger threshold and the behavior confirmation mechanism. The calculated left-turn probability, straight-ahead probability, right-turn probability, and decision confidence are jointly input into the control strategy decision-making module. Based on the current probability magnitude ranking and combined with the entropy value, this module determines whether the decision condition for performing the turn signal switching operation is met. If the confidence is high, the probability distribution is concentrated, and the current judgment is inconsistent with the previous control state, the system determines the optimal turn signal control strategy according to the maximum probability direction and sends the corresponding control instruction to the control unit of the seat scooter via the CAN bus to drive the left-turn light on, the right-turn light on, or keep the lights off.

[0035] In a specific embodiment, the process of determining the optimal turn signal control strategy based on the left-turn probability, straight-ahead probability, right-turn probability, and decision confidence and sending the turn signal control instruction to the control unit of the seat scooter may specifically include the following steps: Construct a turn signal control state transition function, where the turn signal control state transition function includes the state space of three states: left-turn light on, lights off, and right-turn light on, and the action space of three operations: activate the left-turn light, turn off the turn signal, and activate the right-turn light; Construct a composite reward function based on the left-turn probability, straight-ahead probability, and right-turn probability, where the composite reward function includes a steering intention matching degree reward component, an energy consumption penalty component, and a state switching frequency penalty component; Dynamically adjust the steering angle threshold according to the decision confidence to obtain a dynamic steering threshold; Based on the turn signal control state transition function, the composite reward function, and the dynamic steering threshold, solve the optimal turn signal control strategy through the Q-learning algorithm; Determine the turn signal control action to be executed according to the optimal turn signal control strategy, and send the turn signal control instruction corresponding to the turn signal control action to the control unit of the seat scooter via the CAN bus.

[0036] Specifically, an executable turn signal control state transition function is constructed at the behavior modeling layer to describe the state evolution process of the turn signal under different operations. This state transition function defines a clear state space and action space. The state space consists of three discrete states, namely "left turn signal on", "lights off", and "right turn signal on", which correspond to the turn signals currently displayed by the vehicle; the action space also includes three operation options, namely "activate left turn signal", "turn off turn signal", and "activate right turn signal". The system can only be in one of these three states at any given moment and performs a state jump by executing the corresponding action after receiving the policy decision result. The state transition function clarifies the next state that may be reached after executing a specific action in the current state and provides structural constraints for subsequent policy optimization. On this basis, a composite reward function is constructed as the driving basis for controlling policy optimization. This reward function reflects the rationality of control behavior, energy consumption performance, and behavior smoothness, and is therefore designed to include three core components. First is the "turn intention matching reward component", which measures the consistency between the current light control state and the user's actual turn intention. For example, when the system determines that the current probability of turning left is the highest and the control state is the left turn signal on, the reward value of this item is set to a positive value; conversely, if there is a deviation, a negative reward is given to correct the mismatch between behavior and intention. Secondly, there is an "energy consumption penalty component", which mainly punishes the ineffective energy consumption caused by frequent turning on and off of the lights, especially significantly reducing the reward value in unnecessary flashing or short-cycle operations to encourage the system to tend to a stable state rather than high-frequency changes. The "state switching frequency penalty component" is introduced, which penalizes according to the frequency of state switching in the recent several time steps. The more frequent the switching, the higher the penalty, so as to suppress irregular switching behavior while maintaining response sensitivity, and further improve the consistency of the user experience and the clarity of the lighting logic. To enhance the adaptability of the control policy to uncertain environments, a dynamic steering threshold adjustment mechanism based on decision confidence is introduced. This mechanism adjusts the steering angle judgment threshold in real time according to the entropy value or confidence index calculated in the probability density evaluation result. If the current decision confidence is high, the system appropriately reduces the steering judgment threshold to enable the policy to quickly respond to the behavior intention; conversely, when the confidence is low, the system increases the angle threshold to suppress the wrong judgment and premature control behavior caused by uncertainty. This mechanism improves the agility of the policy in high-confidence situations and enhances the fault tolerance ability in low-confidence states, constructing a dynamically adjustable and highly adaptable policy perception gating mechanism. Based on the above state definition, reward construction, and confidence gating mechanism, the Q-learning algorithm is used as the basis to perform the step-by-step solution process of the optimal policy. In this reinforcement learning process, the state-action pair is used as the learning unit to construct a Q-value table, which is used to represent the cumulative expected return that can be obtained by executing a certain action in the current state.At each moment, according to the current state and the predicted result of the steering intention, calculate the Q-values of all executable actions, select the action with the maximum Q-value as the current execution strategy, record the actual obtained reward feedback at the same time, and iteratively update the Q-values according to the update formula of Q-learning. As the number of learning times increases, the system continuously approaches the optimal strategy, enabling it to always select the optimal turn signal control operation when facing different scenarios and user intentions. After the policy learning stage is completed, the system can determine the turn signal control action to be executed according to the current state and the learned optimal strategy in each decision-making cycle. For example, when the current state is "lights off", and it is determined that the left turn threshold has been reached based on the left turn probability and confidence, and it is found through policy evaluation that the expected benefit of "activating the left turn signal" is the highest, the system then executes this action to switch to the state of "left turn signal on". After completing the action decision, the system maps this control action into a standardized turn signal control instruction and sends this instruction to the underlying control unit of the seat scooter in real time through the CAN bus to drive the corresponding turn signal module to enter the target state and achieve the response output at the hardware level.

[0037] The automatic control method of the turn signal in the embodiment of the present invention has been described above. Next, the automatic control system of the turn signal in the embodiment of the present invention will be described. Please refer to Figure 2 , an embodiment of the automatic control system of the turn signal in the embodiment of the present invention includes: A decomposition module, configured to perform orthogonal basis decomposition on the three-axis acceleration data and three-axis angular velocity data of the seat scooter to obtain intention-related subspace data and environment-related subspace data; A fusion module, configured to perform harmonic attitude decomposition and attention mechanism fusion on the intention-related subspace data to obtain a fusion feature vector; An identification module, configured to perform spatio-temporal context steering pattern identification on the fusion feature vector in combination with the environment-related subspace data to obtain an initial turn probability vector and a time persistence index; A calculation module, configured to calculate the left turn probability, straight-ahead probability, right turn probability and decision confidence according to the initial turn probability vector and the time persistence index, and send a turn signal control instruction to the control unit of the seat scooter.

[0038] Through the synergy of the above components, the harmonic attitude decomposition technology and the biorthogonal basis mapping algorithm are used to achieve accurate separation of the turning intention signal and the environmental interference signal, effectively solving the technical problem that the traditional method cannot accurately distinguish between intentional turning and road vibration interference, and significantly reducing the false trigger rate. Based on the multi-level dynamic time window feature fusion technology, the attitude change patterns of short-term windows, medium-term windows and long-term windows are analyzed simultaneously, overcoming the technical contradiction that a single time scale cannot take into account both instantaneous responsiveness and anti-interference stability at the same time, so that the system can respond quickly to clear turning intentions and resist short-term interference. Through the spatiotemporal context-aware steering pattern recognition framework, the scene adaptive recognition is combined with the riding state information to solve the problem of inaccurate judgment caused by ignoring the riding context in the existing technology, and maintain a high accuracy of steering intention recognition in different riding scenarios. The turn signal control system driven by probability density expresses the turning intention as a continuous probability density function instead of a simple yes / no judgment, breaking through the limitations of the traditional binary logic control mode and providing a more refined and progressive expression of steering intention. Through the dynamic threshold adjustment mechanism driven by decision confidence, the judgment threshold is raised when the uncertainty is high, and the response threshold is lowered when the certainty is high, thereby enhancing the robustness and adaptability of the system in various complex environments.

[0039] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0040] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0041] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An automatic control method for a turn signal, characterized in that, Including: Performing orthogonal basis decomposition on the three-axis acceleration data and three-axis angular velocity data of the seat scooter to obtain intention-related subspace data and environment-related subspace data; Performing harmonic attitude decomposition and attention mechanism fusion on the intention-related subspace data to obtain a fused feature vector; Combining the environment-related subspace data to perform spatio-temporal context turning pattern recognition on the fused feature vector to obtain an initial turning probability vector and a time persistence index; Calculating the left-turn probability, straight-ahead probability, right-turn probability and decision confidence according to the initial turning probability vector and the time persistence index, and sending a turn signal control instruction to the control unit of the seat scooter.

2. The automatic control method of the turn signal according to claim 1, characterized in that The performing orthogonal basis decomposition on the three-axis acceleration data and three-axis angular velocity data of the seat scooter to obtain intention-related subspace data and environment-related subspace data includes: Fixing a six-axis attitude sensor at the center position of the seat scooter chassis and collecting three-axis acceleration data and three-axis angular velocity data; Performing band-pass filtering on the three-axis acceleration data and the three-axis angular velocity data to obtain filtered attitude data, and performing segmentation processing and feature calculation on the filtered attitude data to obtain a target feature space; Based on the target feature space, performing Gram-Schmidt orthogonalization construction to obtain an intention-related orthogonal basis and an environment-related orthogonal basis; Based on the intention-related orthogonal basis and the environment-related orthogonal basis, performing orthogonal decomposition projection on the filtered attitude data to obtain intention-related subspace data and environment-related subspace data.

3. The automatic control method of the turn signal according to claim 1, wherein The performing harmonic attitude decomposition and attention mechanism fusion on the intention-related subspace data to obtain a fused feature vector includes: Performing Gabor transform on the intention-related subspace data to obtain a time-frequency spectrum of attitude changes; Using a subspace projection algorithm to perform frequency band separation on the time-frequency spectrum of attitude changes to obtain a steering intention signal component; Based on an intention enhancement adaptive operator, performing signal enhancement on the steering intention signal component to obtain an enhanced steering intention feature signal; Performing inverse Gabor transform processing on the enhanced steering intention feature signal to obtain a steering intention time series feature; Based on the steering intention time series feature, constructing a steering intention feature vector including signal energy, main frequency, phase coherence, duration, waveform steepness, energy distribution skewness, kurtosis and spectral entropy; Performing multi-scale feature analysis and attention mechanism fusion on the steering intention feature vector to obtain a fused feature vector.

4. The automatic control method of the turn signal according to claim 3, characterized in that, The performing multi-scale feature analysis and attention mechanism fusion on the steering intention feature vector to obtain a fused feature vector includes: By setting a 0.5-second short-time window, a 2-second medium-time window and a 5-second long-time window, performing sliding analysis on the steering intention feature vector to respectively obtain an initial short-time feature vector, an initial medium-time feature vector and an initial long-time feature vector; Performing local feature enhancement on the initial short-time feature vector, the initial medium-time feature vector and the initial long-time feature vector to obtain an enhanced short-time feature vector, an enhanced medium-time feature vector and an enhanced long-time feature vector; Perform a non - linear transformation on the enhanced short - term feature vector, the enhanced medium - term feature vector, and the enhanced long - term feature vector to obtain a transformed short - term feature vector, a transformed medium - term feature vector, and a transformed long - term feature vector; Based on the soft attention model, calculate the dynamic weight coefficients and perform weighted fusion on the transformed short - term feature vector, the transformed medium - term feature vector, and the transformed long - term feature vector to obtain a weighted fusion vector; Calculate the cross - correlation coefficient matrix between the different time - scale features in the initial short - term feature vector, the initial medium - term feature vector, and the initial long - term feature vector, and verify the feature consistency of the weighted fusion vector based on the cross - correlation coefficient matrix to obtain a fused feature vector.

5. The automatic control method of the turn signal according to claim 1, characterized in that, Combining the environmental - related subspace data to perform spatio - temporal context turning pattern recognition on the fused feature vector to obtain an initial turning probability vector and a time persistence index, including: Collect the speed data and vibration characteristic data of the seat scooter, and perform multi - feature analysis on the speed data and the vibration characteristic data through the DBSCAN clustering algorithm to obtain the current riding scene category; Perform joint representation processing on the fused feature vector and the environmental - related subspace data to obtain a comprehensive feature representation; Perform manifold space mapping on the comprehensive feature representation and the current riding scene category to obtain a low - dimensional turning intention representation; According to the current riding scene category, select the corresponding pre - trained deep neural network classifier to classify the low - dimensional turning intention representation to obtain an initial turning probability vector; Calculate the cosine similarity of the initial turning probability vector between consecutive frames. When the similarity is greater than a preset threshold and the duration exceeds a set threshold, it is recognized as a valid turning intention to obtain a time persistence index.

6. The automatic control method of the turn signal according to claim 1, wherein Calculating the left - turn probability, straight - ahead probability, right - turn probability, and decision confidence according to the initial turning probability vector and the time persistence index, and sending a turn - signal control instruction to the control unit of the seat scooter, including: Construct a time - series conditional probability model of the turning intention based on the initial turning probability vector and the time persistence index, and use a mixture Gaussian process to model the time - series conditional probability model to obtain the conditional probability distribution of the turning intention in the continuous angle space; Perform kernel function enhancement processing on the conditional probability distribution of the turning intention in the continuous angle space to obtain a target feature mapping relationship; Perform probability density inference based on the target feature mapping relationship to obtain the probability density function of the continuous turning angle; Perform discretization processing on the probability density function of the continuous turning angle within the range of [-π / 2, π / 2] to obtain a turning intention density vector, and calculate the left - turn probability, straight - ahead probability, and right - turn probability through interval integration; Calculate the probability distribution entropy value of the turning intention density vector to evaluate the uncertainty degree of the turning decision to obtain the decision confidence; Determine the optimal turn - signal control strategy according to the left - turn probability, the straight - ahead probability, the right - turn probability, and the decision confidence, and send a turn - signal control instruction to the control unit of the seat scooter.

7. The automatic control method of the turn signal according to claim 6, wherein Determining an optimal turn signal control strategy based on the left turn probability, the straight-ahead probability, the right turn probability, and the decision confidence, and sending a turn signal control instruction to the control unit of the seat scooter, includes: Constructing a turn signal control state transition function, where the turn signal control state transition function includes a state space of three states: left turn light on, light off, and right turn light on, and an action space of three operations: activating the left turn light, turning off the turn signal, and activating the right turn light; Constructing a composite reward function based on the left turn probability, the straight-ahead probability, and the right turn probability, where the composite reward function includes a steering intention matching degree reward component, an energy consumption penalty component, and a state switching frequency penalty component; Dynamically adjusting the steering angle threshold according to the decision confidence to obtain a dynamic steering threshold; Based on the turn signal control state transition function, the composite reward function, and the dynamic steering threshold, solving for the optimal turn signal control strategy through the Q-learning algorithm; Determining the turn signal control action to be executed according to the optimal turn signal control strategy, and sending the turn signal control instruction corresponding to the turn signal control action to the control unit of the seat scooter through the CAN bus.

8. An automatic control system for a turn signal, characterized in that, For executing the automatic control method of the turn signal according to any one of claims 1-7, the automatic control system of the turn signal includes: A decomposition module for performing orthogonal basis decomposition on the three-axis acceleration data and three-axis angular velocity data of the seat scooter to obtain intention-related subspace data and environment-related subspace data; A fusion module for performing harmonic attitude decomposition and attention mechanism fusion on the intention-related subspace data to obtain a fused feature vector; An identification module for performing spatio-temporal context steering pattern identification on the fused feature vector in combination with the environment-related subspace data to obtain an initial steering probability vector and a time persistence index; A calculation module for calculating the left turn probability, the straight-ahead probability, the right turn probability, and the decision confidence according to the initial steering probability vector and the time persistence index, and sending a turn signal control instruction to the control unit of the seat scooter.

Images