Intelligent teaching system based on AI robot coach

Abstract

The invention provides an intelligent teaching system based on an AI robot coach, and relates to the field of intelligent teaching. Comprising a hardware equipment module, a software platform module, a safety control module, a hierarchical intervention module, a physiological monitoring module and a cognitive load regulation module, the hardware equipment module collects vehicle state, environment and student physiological data in real time through a multi-mode sensor; the software platform module generates a teaching strategy based on a deep reinforcement learning model and analyzes a student behavior mode; the safety control module triggers an automatic braking mechanism through the obstacle distance and the vehicle speed; the grading intervention module is divided into three-level intervention strategies of an L1 unit, an L2 unit and an L3 unit according to the operation error type, and voice prompt and braking intervention are dynamically adjusted; the physiological monitoring module non-inductively evaluates a tension index through micro-expression, grip strength and voice data; the cognitive load adjusting module dynamically adjusts the teaching content difficulty and the information density in combination with the steering frequency, the accelerator and brake alternating times and the tension index.

Description

Technical Field

[0001] The present invention relates to the field of intelligent teaching, and specifically to an intelligent teaching system based on an AI robot coach. Background Art

[0002] With the rapid development of artificial intelligence technology, automated teaching systems have been gradually introduced into the field of intelligent driving education to improve training efficiency; the existing technology mainly combines in-vehicle sensors with preset teaching programs to achieve basic operation guidance and error feedback; traditional driving training relies on the experience judgment of human coaches, resulting in problems such as inconsistent teaching standards and uneven resource allocation, which promotes the transformation of the industry towards intelligence and standardization.

[0003] Most existing driving teaching systems use a single sensor to monitor the vehicle state, such as a steering correction system based on a steering wheel angle sensor or an auxiliary device that identifies lane lines through a camera; some solutions combine a voice prompt function to trigger a warning when crossing the line or speeding; there are also technologies that attempt to evaluate the trainee's tension through a heart rate monitoring bracelet, but the correlation analysis between physiological data and operation behavior is relatively weak; some commercial products use fixed thresholds to judge risk events and lack dynamic adaptability.

[0004] The existing technology generally has problems of insufficient multi-modal data fusion, lack of a collaborative response mechanism between the physiological monitoring and vehicle control modules; the teaching strategy is rigid and cannot adapt to the individual differences of trainees, and error correction only stays in the stage of post-event prompting; safety control relies on manual intervention or simple braking logic, making it difficult to achieve risk grading; the processing delay of sensor data is relatively high, affecting the real-time feedback effect, and the privacy protection mechanism is imperfect, restricting the reliability of the system and the user experience. Summary of the Invention

[0005] (I) Technical Problems to be Solved In view of the deficiencies of the existing technology, the present invention provides an intelligent teaching system based on an AI robot coach to solve the problems of insufficient multi-modal data fusion resulting in uncoordinated system response; lack of a collaborative response mechanism between the physiological monitoring module and the vehicle control module; rigid teaching strategies that cannot adapt to the individual differences of trainees; error correction relying on post-event prompting and lacking dynamic threshold adjustment; and single safety control logic that makes it difficult to achieve risk grading as mentioned in the above background art.

[0006] (II) Technical Solutions To achieve the above objectives, the present invention is realized through the following technical solutions: An intelligent teaching system based on an AI robot coach, comprising: The hardware device module includes an in-vehicle terminal, which is used to perform encrypted communication and real-time preprocessing and instruction distribution of multi-modal sensor data. At the same time, it collects vehicle status data, environmental data, and trainee physiological data in real time through the multi-modal sensors, and conducts data interaction with the software platform module; The software platform module generates teaching strategies based on a deep reinforcement learning model and drives the hardware device module to execute operation instructions. At the same time, it analyzes the trainee's behavior patterns to generate personalized learning paths; The hierarchical intervention module is divided into three-level intervention strategies, namely L1 unit, L2 unit, and L3 unit, based on the error types in the trainee operation data. Among them, the L2 unit dynamically combines or splits the operation step granularity according to the real-time cognitive load score; The safety control module includes a risk grading response unit, which monitors the obstacle distance, vehicle speed, and operation anomalies in real time through radar and divides them into three-level risk levels, working in coordination with the hierarchical intervention module; and triggers an automatic braking mechanism through the obstacle distance and vehicle speed data, generates operation records, and guides the trainee to complete the safety operation review process; The physiological monitoring module constructs trainee psychological state indicators through the trainee physiological data collected in real time, generates a tension index, and inputs it into the cognitive load adjustment module; The cognitive load adjustment module includes an adaptive adjustment unit, which executes a dynamic adjustment strategy according to the trainee's tension index, constructs a cognitive load scoring model based on the vehicle status data, tension index, and historical error rate, generates a cognitive load score, and dynamically adapts the difficulty and information density of the teaching content.

[0007] Preferably, the hardware device module realizes real-time acquisition of vehicle status, environment and trainee physiological data through multi-mode sensors in the data acquisition unit; the vehicle speed sensor measures the wheel speed in the form of pulse signal frequency conversion and transmits it to the embedded system through the CAN bus at a sampling rate of 50 times per second; the steering wheel steering angle sensor uses an optoelectronic encoder to generate digital signals for every 0.1-degree steering angle and transmits them through the SPI protocol; the throttle and brake depth sensor detects the pedal displacement based on the Hall effect with an accuracy of ±1 mm and uploads it after being converted by the ADC as an analog voltage signal; the millimeter-wave radar emits frequency-modulated continuous waves in the 77 GHz frequency band, calculates the distance and relative speed of obstacles through the Doppler effect, with a detection range of 200 meters and an angular resolution of 1 degree; the lidar uses the TOF principle to generate 128-line point cloud data by scanning with a 905 nm wavelength laser, with a frame rate of 10 Hz; the camera is a binocular RGB-IR module with a resolution of 1920×1080, and can identify lane lines and traffic signs in real time through image recognition algorithms; the infrared camera captures the thermal imaging data of the trainee's face at a frame rate of 30 fps and extracts micro-expression features in combination with the OpenFace library; the steering wheel piezoelectric film sensor detects the grip force fluctuation through the piezoresistive effect with a sampling rate of 100 Hz; the noise-canceling microphone uses beamforming technology to filter out environmental noise and extract voice fundamental frequency and tremor parameters; all sensor data is preprocessed by the industrial-grade embedded system of the vehicle-mounted terminal and encrypted and transmitted to the software platform module through the MQTT protocol.

[0008] Preferably, the software platform module constructs a deep reinforcement learning model, and the input layer includes trainee operation data, physiological monitoring data and cognitive load scores; the deep reinforcement learning model analyzes the temporal correlation between the steering wheel steering angle and the throttle and brake through the Q-learning algorithm and outputs teaching strategy instructions; the deep reinforcement learning model performs clustering analysis on the operation records of historical trainees to generate a typical error pattern library; the generation logic of the personalized learning path is as follows: when the trainee conducts the first training, a basic teaching template is matched according to the initial ability test results; during the training process, the weighted score of the trainee's operation accuracy and tension index is calculated every 5 minutes. If the score is lower than the threshold, the decomposed teaching video in the special training content library is called; the dynamic adjustment of the teaching strategy is specifically manifested as: when the trainee does not press the line for 5 consecutive times in the reverse parking project, the training difficulty is automatically increased to the side parking project, and a curve simulation scenario is added.

[0009] Preferably, the risk grading and response unit of the safety control module perceives the distance of obstacles through the fusion of millimeter-wave radar and camera, and uses the Kalman filter algorithm to predict the collision time; when the obstacle distance is less than 3 meters and the relative speed is greater than 20 km / h, a slow braking mechanism is triggered, and the braking deceleration is 2 m / s 2; The risk level is divided into three levels: the first-level risk is that the distance to the obstacle is less than 1 meter, directly triggering the emergency braking of the L3 unit; the second-level risk is that the distance to the obstacle is 1 - 3 meters and the trainee does not perform the braking operation, triggering the intervention strategy of the L2 unit; the third-level risk is that the seat belt is not fastened or the accelerator is misstepped, triggering a voice alarm and limiting the vehicle speed to 10 km / h; the cooperation logic between the module and the hierarchical intervention module is: when the cognitive load score ≥ 75 points, even if the physical risk does not reach the threshold, the L2 unit is still forced to activate to lock the teaching item; the slow braking command is sent to the electronic stability control system through the CAN bus, first reducing the engine torque and then gradually applying braking force to avoid vehicle out of control; during the braking process, the on-vehicle screen displays the dynamic curve of the obstacle distance and the braking progress bar in real time.

[0010] Preferably, the triggering condition of the L1 unit is that the steering wheel response delay exceeds 0.5 seconds or the deviation from the line is less than 20 cm; when the condition is detected, a voice prompt is generated through the TTS engine, and the content is "Please adjust the steering wheel to the left dotted line"; at the same time, a semi-transparent green virtual marking is superimposed on the in-vehicle screen, and the marking width is 1.2 times that of the real lane line, and the ideal trajectory is highlighted and flashed in red dotted line; during the autonomous adjustment of the trainee, the training process continues to run. If the error is not corrected within 5 seconds, the voice prompt is repeated but no new marking is added; the dynamic threshold is generated through a preset weighted calculation model, and its input parameters include the real-time scoring data output by the cognitive load adjustment module, the trainee's historical error rate data, and the difficulty coefficient of the current training stage; the weighted calculation model is: dynamic threshold = α × real-time scoring + β × historical error rate + γ × difficulty coefficient, where α, β, and γ are preset weight coefficients, and α + β + γ = 1. For example, when the trainee's operation accuracy rate is increased by 10%, the delay threshold is relaxed to 0.8 seconds, and the line-pressing threshold is expanded to 25 cm. When the cognitive load score ≥ 70 points, the voice prompt density is reduced from 3 times per minute to 2 times, and the command is simplified to a single-parameter command such as "Turn left 15 degrees". If the error is not corrected within 5 seconds, the system enters multi-level intervention: when the preset duration of the multi-level progressive intervention stage reaches 5 - 10 seconds, the direction-specific touch of the steering wheel and the micro-path animation rendered by AI in real time are superimposed. After more than 10 seconds, the steering angle of the steering wheel is forcibly restricted and the speed is reduced to the safety threshold; all intervention data is encrypted and stored in the local SQLite database, and a daily report containing the error distribution and scoring curve is generated. In the abnormal state, that is, when the cognitive load score > 75 points or the stress index ≥ 65 points, non-critical visual elements are automatically hidden, the tactile intensity is attenuated, and the deep breathing rhythm voice guidance synchronized with the heart rate is activated. The non-critical visual elements refer to the auxiliary visual content on the in-vehicle screen except for the core driving guidance information, such as lane lines, obstacle warnings, and vehicle trajectories, including but not limited to background grids, secondary icons, and decorative elements; the attenuation of the tactile intensity refers to the vibration feedback intensity of the piezoelectric actuator built into the steering wheel being dynamically reduced according to a preset ratio. For example, in the normal state, the tactile feedback intensity is 100%; when the stress index ≥ 65 points, the intensity is attenuated to 50% to avoid the high-intensity vibration from exacerbating the trainee's anxiety.

[0011] Preferably, the initial value of the preset duration of the multi-level progressive intervention stage can be set according to the vehicle type, training scenario, or trainee group differences. In the system initialization or safety-critical scenarios, the default durations are set as follows: the first preset duration is 5 seconds, the second preset duration is 10 seconds, and the third preset duration is 15 seconds. Subsequently, it is gradually optimized according to the output of the deep Q-network model. For example, when the trainee conducts the first training, a fixed duration is used as the baseline; when the model accumulates enough data, it switches to the dynamic adjustment mode.

[0012] Preferably, the preset duration of the multi-level progressive intervention stage is dynamically optimized based on a deep Q-network model; the state space of the deep Q-network model includes real-time cognitive load scores, stress indices, historical operation correction efficiencies, and the cumulative uncorrected duration of the current intervention stage; the action space is defined as adjustment strategies for shortening, maintaining, or extending the preset duration, with an amplitude between 10% and 30%; the reward function comprehensively considers the correction success rate, the increase in cognitive load scores, and the decrease in stress indices, and quantifies the intervention effect through fixed weight coefficients, including focusing on operation effect α = 0.5, suppressing cognitive overload β = 0.3, and relieving psychological stress γ = 0.2. The specific formula is: Reward function = α·correction success rate - β·increase in cognitive load score + γ·decrease in stress index, where the correction success rate is the ratio of the learner's independent correction operations in the current stage; the increase in cognitive load score is the increase in the score during the intervention stage, and if the score decreases, it is recorded as a negative value; the decrease in stress index is the decrease in the stress index during the intervention stage; the neural network uses a 4D input layer, two fully connected hidden layers, and a 3D output layer; during training, data is sampled in batches through an experience replay pool, the Q-value prediction is optimized with a mean squared error loss function, and the target network parameters are updated synchronously. The loss function formula is , where E is the expectation operator, representing the average of the squared errors of all data points in the batch sample, and Q target is the target Q-value, and the calculation formula is , where R is the immediate reward obtained by the current action; γ is the discount factor, where 0 < γ < 1, which is used to balance the importance of the current reward and future rewards; is the next state is the maximum Q-value of all possible actions in the next state, calculated by the target network, that is, the non-real-time updated network, and Q predicted is the Q-value prediction of the main network, that is, the real-time updated network, for the current state-action pair; among them, offline training triggers model weight updates when the learner successfully corrects independently twice under high-pressure conditions; after the update, a personalized stage duration threshold is generated. For example, in practical applications, the model can dynamically adjust the intervention duration according to the learner's state. For example, in a high-pressure state: shorten the duration by 30% at a cognitive load of 85 points and a stress index of 75 points, and adjust the current stage duration from 10 seconds to 7 seconds; the learner receives tactile feedback and voice prompts in a shorter time, reducing cognitive stress, or in a stable state: extend the duration by 20% at a cognitive load of 60 points and a stress index of 50 points, and extend the stage duration from 10 seconds to 12 seconds; reduce frequent interventions, increase the learner's independent operation time, and improve the skill proficiency by 15% to enhance the independent training space. The experience replay pool stores the interaction data of the learner's intervention stage, including states, actions, rewards, and the next state in the experience pool; the batch sampling randomly extracts 32 groups of data from the experience pool each time for training; the update of the target network parameters synchronizes the main network parameters to the target network every 100 steps.

[0013] Preferably, the triggering condition of the L2 unit is that the number of consecutive occurrences of the same error type reaches the preset threshold, and the preset threshold is 5 consecutive occurrences within 10 minutes or more than the maximum allowable value of 10 times accumulated in a single day; after triggering, extract the error types within the TOP3 of this trainee from the historical error database and generate a special module containing 5 progressive training scenarios; for example, for the error of "excessive turning radius", the first scenario requires the trainee to drive on a curve with a radius of 8 meters, and after success, it is reduced to 6 meters; the implementation method of locking the teaching project is: mark the current project as "uncompleted" in the teaching schedule of the in-vehicle terminal and disable the selection button for the next project; when the teaching video is forced to play, the trainee must watch at least 90% of the duration and pass the in-class test to unlock; if the cognitive load score continues to be higher than 65 points, each operation step is split into three sub-steps of "steering wheel adjustment - throttle control - trajectory correction", and the score data is collected in real time after each sub-step is completed; if the sub-step score is lower than 60 points, repeat the step and superimpose auxiliary virtual markings; all special training data is encrypted by HMAC-SHA256 and uploaded to the cloud.

[0014] Preferably, in the hierarchical intervention module, when the cognitive load score of the trainee exceeds the preset high load threshold of 75 points for 10 consecutive minutes, the special training module is split into independent steps to be executed sequentially according to the operation logic. For example, the parallel parking project is split into four steps: adjusting the initial parking space distance, turning the steering wheel fully to the left, straightening the vehicle body, and fine-tuning the position. The target score threshold for each step is set at 65 points. The steering deviation value is collected with a precision of 0.1 degrees by the steering wheel angle sensor, the alternating frequency data is recorded at a sampling rate of 100 Hz by the accelerator and brake depth sensor, and the stress index updated every 2 seconds by the physiological monitoring module. Based on the weighted scoring model, the step execution score is calculated. The weight distribution of the weighted scoring model is: steering deviation score accounts for 40%, accelerator and brake alternating frequency score accounts for 30%, stress index score accounts for 20%, and historical error rate score accounts for 10%. If the single-step score is lower than the lower limit of the safe load range of 65 points, the system triggers the closed-loop feedback mechanism: a red auxiliary virtual marking with a width 1.5 times that of the actual marking is superimposed on the in-vehicle screen, and a breakdown action video with a duration of 15 seconds is played, forcing the trainee to repeat the current step at least 3 times until the score reaches or exceeds 65 points for two consecutive times. If the score meets the standard, an unlocking instruction is sent to the terminal through the AES-256 encryption protocol, the auxiliary virtual marking is cleared, and the next step is entered. When the trainee's historical error rate drops by 10%, the target score threshold is automatically increased to 70 points. If the cognitive load score continues to be higher than 75 points, the step granularity will be further refined. For example, the fine-tuning position will be split into two sub-steps: horizontal fine-tuning and vertical fine-tuning. When the step execution score is lower than the lower limit of the safe load range for two consecutive times, the adaptive adjustment unit triggers the interface information simplification operation, closes the non-critical display information, and extends the voice prompt interval. At the same time, the L2 unit dynamically combines or splits the operation step granularity according to the real-time cognitive load score, forming a joint decompression closed loop from the operation layer to the information layer. The effective range of the safe load range is 60-75 points of the cognitive load score. Below this range, the step repetition mechanism is triggered, and above this range, the information density is forced to be simplified.

[0015] Preferably, the L2 unit dynamic step splitting of the hierarchical intervention module, the closed-loop control unit, and the adaptive adjustment unit achieve deep cooperation through heterogeneous data mirroring feedback and dynamic priority nesting. When the dynamic step splitting unit splits the special training into an independent operation step sequence, if the step execution score of a certain step is lower than the lower limit of the safety load range for two consecutive times, the adaptive adjustment unit turns off the non-critical display information of the in-vehicle screen and extends the voice prompt interval to twice the original interval, reducing the cognitive load of the trainee in reverse by simplifying the interface; when the step splitting granularity and the information density optimization threshold are dynamically bound through the adaptive adjustment unit, when the number of independent steps after splitting exceeds the preset threshold, the adaptive adjustment unit automatically increases the lower limit of the safety range of the cognitive load score by 10%, forcing subsequent steps to be merged into a coarse-grained operation unit. For example, "steering wheel adjustment - throttle control - trajectory correction" is merged into a single instruction "steering wheel adjustment"; if the adaptive adjustment unit triggers the breathing rhythm guidance function animation due to the tension index exceeding the limit, a deep breathing instruction synchronized with the real-time heart rate is output through the voice module, and the dynamic step splitting unit freezes the execution state of the current step. After the animation ends, the step execution score is recalculated based on the steering wheel steering angle deviation value and the throttle and brake alternation frequency at the freezing moment. If the score still does not meet the standard, it jumps to a finer-grained sub-step splitting mode, such as splitting "throttle control" into a two-step "depth calibration - rhythm matching".

[0016] Preferably, the L3 unit identifies abnormal operation modes of the throttle or brake by real-time monitoring of the trainee's operation behavior and vehicle dynamic parameters. When it detects that the operation behavior causes the vehicle acceleration to exceed the safety threshold, the system immediately executes a slow braking instruction, gradually reducing the vehicle speed to the safe range and locking the vehicle control authority; during the braking process, the system packages and sends the operation state data, including the throttle and brake dynamic curves, vehicle body attitude, and environmental perception data, to the coach end through an encrypted communication channel; the coach needs to complete biometric verification at the in-vehicle terminal to restore vehicle control authority, and then the system automatically starts the review process, restoring the operation scenario through a visual interface and marking the key error nodes; in addition, the trigger condition of the L3 unit is dynamically matched with the tension index output by the physiological monitoring module and the cognitive load score. When the tension index or cognitive load exceeds the preset threshold, the system preferentially activates the L3 unit intervention strategy to ensure the timeliness and safety of risk response.

[0017] Preferably, the trigger condition of the L3 unit is that the throttle depth increases from 0% to 80% within 0.5 seconds and the vehicle acceleration exceeds 4m / s 2 , or the deceleration caused by emergency braking exceeds 6m / s 2When the conditions are met, apply a gentle brake through the electronic hydraulic braking system to reduce the vehicle speed to 5 km / h within 2 seconds; at the same time, the steering wheel motor locks the steering angle and disables the throttle pedal signal input; and the data packet sent to the coach terminal includes the throttle depth curve, the Euler angle data of the vehicle body attitude, and the distance to the nearest obstacle of the millimeter-wave radar; the coach needs to continuously press the fingerprint recognition module on the in-vehicle terminal for 3 seconds to complete the identity verification. After unlocking, the operation review animation will be automatically played. The animation restores the accident scene with a 3D model and marks the points where the trainee made mistakes. The coach can provide remote guidance on relevant experience; the priority logic of the L3 unit is: when the tension index ≥ 65 points and the cognitive load score ≥ 70 points, even if the acceleration does not exceed the preset threshold, the braking intervention will still be triggered; the braking records are saved as non-erasable read-only files for accident liability tracing.

[0018] Preferably, the physiological monitoring module constructs a trainee's psychological state index to generate a tension index based on the physiological data of the trainee collected in real time; the physiological data includes facial micro-expressions, hand grip force fluctuations, and voice signals. Among them, the infrared camera is used to capture the trainee's facial images at 30 fps, and micro-expression features such as the degree of mouth corner drooping and the degree of eyebrow raising are extracted through the ResNet-18 model; the piezoelectric film sensor on the steering wheel detects the grip force fluctuations, and the FFT algorithm is used to calculate the energy ratio in the frequency band of 0.5 - 4 Hz as the grip force stability index; after the noise-canceling microphone collects the voice signal, the standard deviation of the fundamental frequency is extracted through the Praat tool to measure the tremor degree; the edge computing module organizes the collected physiological data into micro-expression scores, grip force stability scores, and voice tremor scores and normalizes them to 0 - 100 points, where the micro-expression weight accounts for 40%, the grip force fluctuation accounts for 30%, and the voice tremor accounts for 30%; the specific formula is: tension index = 0.4×S 微表情 + 0.3×S 握力 + 0.3×S 语音 where S 微表情 is the micro-expression score; S 握力 is the grip force stability score; S 语音 is the voice tremor score; for example, if the degree of mouth corner drooping of the trainee exceeds the threshold, the micro-expression item will be deducted 15 points; the tension index is updated every 2 seconds and is pushed to the dashboard on the in-vehicle screen in real time through the WebSocket protocol; the trainee can turn off the infrared camera monitoring through the dashboard, but the data collection functions of the steering wheel and the microphone will be forcibly retained to ensure basic safety.

[0019] Preferably, the physiological monitoring module captures real-time facial micro-expressions and pupil change data of the trainee through the infrared camera collected by the multi-mode sensor, continuously monitors the handgrip force fluctuation frequency of the trainee through the piezoelectric film sensor on the steering wheel, and analyzes the voice tremor data of the trainee after collecting through the noise-canceling microphone; the multi-mode sensor synchronously collects the steering wheel steering frequency data output by the steering wheel angle sensor and the number of alternations of the accelerator and brake output by the accelerator and brake depth sensor; the data collected by the multi-mode sensor is only stored locally in the vehicle terminal and supports selectively turning off single monitoring functions through the supporting mobile application; based on the vehicle state data, that is, the steering frequency data collected by the steering wheel angle sensor and the number of alternations data collected by the accelerator and brake depth sensor, plus the stress index data and historical error rate data of the physiological monitoring module as input parameters of the cognitive load adjustment module, which are used to construct a cognitive load scoring model.

[0020] Preferably, the input parameters of the cognitive load scoring model include the steering wheel steering frequency, the number of accelerator and brake alternations, the stress index, and the historical error rate; the cognitive load scoring model uses a multiple linear regression algorithm, and the formula is: score = 0.3×steering frequency + 0.4×number of alternations + 0.2×stress index + 0.1×historical error rate; the scoring result is updated every 30 seconds. If the score is ≥70 points for 3 consecutive times, the information density optimization strategy is triggered, which specifically includes: reducing the parking space width for reverse parking from the standard 2.5 meters to 2.3 meters, and adding an S-curve to the simulated exam route; at the same time, the in-vehicle screen turns off non-critical information such as the remaining fuel and gear status, and only retains the vehicle trajectory line and obstacle warning frame; the voice prompt interval is extended from 10 seconds to 15 seconds, and only core instructions such as "turn left" or "brake" are broadcast; if the score drops below 60 points, the basic training mode is automatically restored and positive voice feedback for successful operations is increased.

[0021] Preferably, the cognitive load adjustment module includes an adaptive adjustment unit that executes a dynamic adjustment strategy according to the trainee's stress index. When the stress index reaches the first preset threshold, the in-vehicle screen is automatically switched to the breathing rhythm guidance animation interface, the voice prompt frequency is reduced to once per minute, and the training difficulty parameters are dynamically adjusted. If the stress index continues to be higher than the second preset threshold, the current training process is paused and a remote intervention request is sent to the coach terminal. At the same time, in combination with the cognitive load score, when the cognitive load score reaches or exceeds the specified threshold, an information density optimization strategy is executed, non-critical display information is turned off, only the vehicle's real-time trajectory and key ground markings are retained, and the voice prompt interval is extended to 1.5 times the original interval, thereby reducing the trainee's cognitive load and improving the training effect. The first preset threshold is that the stress index reaches 70 points, and the breathing guidance animation interface is triggered when the stress index rises to 70 points. The second preset threshold is that the stress index reaches 70 points and continues not to fall below 60 points. For example, if the stress index does not fall below 60 points within 2 minutes, the training is paused and a request for the coach to intervene remotely is made.

[0022] Preferably, when the millimeter-wave radar detects that the distance to an obstacle enters the 5-meter range, the risk classification response unit first checks the current vehicle speed: if the vehicle speed > 30 km / h, the L3 unit is directly triggered; if the vehicle speed ≤ 30 km / h, the type of the obstacle is judged in combination with the target recognition result of the camera, including pedestrians, vehicles, and static objects. For pedestrian targets, the safety threshold is reduced to 3 meters. At the same time, the unit receives the cognitive load score in real time. If the score ≥ 75 points, the L2 unit intervention strategy is forcibly activated regardless of whether the physical risk exceeds the limit. For example, in the case of high cognitive load, even if the obstacle distance is 4 meters and the vehicle speed is 20 km / h, the training item will be locked and special exercises will be pushed. The collaborative logic is implemented through a message subscription mechanism, and status data is transmitted between modules in JSON format to ensure that the response delay is less than 100 milliseconds.

[0023] Preferably, when the vehicle is ignited, the industrial-grade embedded system executes a startup self-check process: first, the GPS / Beidou dual-mode signal strength of the positioning module is detected through the PING command. If the signal is lost for more than 10 seconds, an alarm is triggered. Then, the calibration error detection of the millimeter-wave radar and the lidar is carried out, and the maximum allowable angle deviation is ±0.5 degrees. For the brake system pressure test, the ESP module is required to establish a braking pressure of 10 MPa within 500 ms. If the standard is not met, a fault code P0x12 is generated. After the self-check passes, the lightweight Linux kernel is loaded, and the real-time data processing thread is started: the steering wheel and throttle data are directly written into the shared memory through the DMA channel for the cognitive load adjustment module to call.

[0024] (III) Beneficial effects The present invention provides an intelligent teaching system based on an AI robot coach, having the following beneficial effects: 1. The present invention collects vehicle status and environmental data in real time through a multi-mode sensor, and realizes accurate measurement of the distance to obstacles by combining the fusion perception of millimeter-wave radar and camera, effectively reducing the collision risk; the hierarchical intervention module divides the L1 unit, L2 unit, and L3 unit three-level intervention strategies based on dynamic thresholds, and responds step by step through voice prompts, special training, and emergency braking when the trainee makes an operation error, significantly improving the safety of driving teaching; the cognitive load adjustment module constructs a scoring model through vehicle status data, stress index, and historical error rate, dynamically adapts the difficulty and information density of teaching content, and ensures that the training intensity of the trainee is always in the optimal load range, improving the efficiency of skill mastery.

[0025] 2. The present invention uses an infrared camera and a piezoelectric film sensor in the physiological monitoring module to non-intrusively capture the micro-expression and grip force fluctuation data of the trainee, and generates a stress index in real time through edge computing to accurately evaluate the mental state; the software platform module analyzes the trainee's historical operations and real-time data based on deep learning algorithms, generates a personalized learning path, and dynamically adjusts the voice prompt frequency and training scenario complexity; the safety control module and the hierarchical intervention module cooperate across layers, and when the cognitive load score is abnormal, a targeted intervention strategy is forcibly triggered to achieve a double balance of teaching safety and mental load, comprehensively improving the intelligence and humanization level of driving training. Specific embodiments

[0026] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] An embodiment of the present invention provides an intelligent teaching system based on an AI robot coach. When a trainee starts the vehicle for reverse parking training, the system first executes a self-check process. The industrial-grade embedded system detects the signal strength of the GPS / Beidou dual-mode positioning module through the PING command. If the signal is lost for more than 10 seconds, a red warning box will pop up on the in-vehicle screen and the training start will be prohibited. At the same time, the millimeter-wave radar and the lidar perform calibration error detection, calculate the angle deviation using a checkerboard calibration board, and the maximum allowable deviation is ±0.5 degrees. If the threshold is exceeded, the preset compensation parameters are called to correct the detection data. The brake system pressure test establishes a braking pressure of 10 MPa within 500 ms through the ESP module. If the standard is not met, a fault code P0x12 is generated and the brake system abnormality is announced by voice. After the self-check passes, the lightweight Linux kernel is loaded, and the real-time data processing thread is started. The steering wheel angle sensor samples data at a rate of 0.1 degree per time through the SPI protocol, and the throttle and brake depth sensors collect the pedal displacement at a sampling rate of 100 Hz. All data is written into the shared memory through the DMA channel for subsequent module calls. After the trainee fastens the seat belt and confirms the training item, the multi-mode sensors of the hardware device module run at full speed. The millimeter-wave radar emits a frequency-modulated continuous wave in the 77 GHz band to detect the distance of obstacles within 5 meters behind the vehicle tail, with an accuracy of ±5 cm. The lidar generates 128-line point cloud data to construct a three-dimensional map, which is fused with the image collected by the camera with a resolution of 1920×1080, and a virtual-real combined parking space line frame is superimposed on the in-vehicle screen. The infrared camera captures the facial thermal image of the trainee at 30 fps, and extracts the features of the degree of drooping corners of the mouth and the degree of raising eyebrows through the ResNet-18 model. The piezoelectric film sensor of the steering wheel monitors the grip force fluctuation frequency, and the noise-canceling microphone collects the voice signal and calculates the standard deviation of the fundamental frequency through the Praat tool. The edge computing module normalizes the micro-expression, grip force stability, and voice tremor data into a percentage-based tension index, which is updated and pushed to the dashboard every 2 seconds. When the reverse operation starts, the software platform module calls the trainee's historical data. If it is a new trainee, the basic teaching template is matched, and a green virtual marking line and an ideal trajectory arrow are displayed on the in-vehicle screen. During the process of the trainee manipulating the steering wheel to reverse, the vehicle speed sensor transmits the wheel speed data in real time. When the vehicle speed exceeds 3 km / h, the system issues a voice prompt through the TTS engine to control the vehicle speed. The risk grading response unit of the safety control module continuously analyzes the millimeter-wave radar data. If the distance between the vehicle tail and the obstacle is less than 1.5 meters and the relative speed is greater than 0.5 m / s, the L1 unit intervention strategy is triggered, a red warning box is superimposed on the screen, and a pulse vibration reminder is generated. When the trainee repeatedly experiences steering wheel steering delay, the L1 unit of the hierarchical intervention module activates dynamic intervention; if the steering wheel response delay exceeds 0.5 seconds or the deviation from the line reaches 20 cm, the system gives a voice prompt to adjust 15 degrees to the right and superimposes a semi-transparent green correction marking on the screen; during the trainee's autonomous adjustment, the training process continues. If the error is not corrected within 5 seconds, the voice prompt is repeated but no new markings are added; the dynamic threshold is automatically adjusted according to the trainee's progress speed. When the operation accuracy rate increases by 10%, the delay threshold is relaxed to 0.8 seconds, and the deviation threshold from the line is increased to 25 cm; if the cognitive load score ≥ 70 points, the system reduces the voice prompt density from 3 times per minute to 2 times and uses a single-parameter instruction to reduce the information load; When the trainee fails to align with the center line of the parking space three times in a row, the L2 unit is activated; the system extracts the under-steering label from the historical error library, generates a special training module and locks the current project; the in-vehicle terminal forcibly plays a pre-recorded teaching video, and the trainee needs to watch the complete video and pass a 5-question multiple-choice test to unlock the next training; if the cognitive load score remains higher than 65 points, the system splits the reverse operation into three sub-steps: initial steering wheel adjustment, fine adjustment, and vehicle body alignment. After each sub-step is completed, the scoring data is collected in real time; if the score of a certain sub-step is lower than 60 points, the step is repeated and an auxiliary virtual marking is superimposed on the screen until the score reaches the standard; During the training process, when it is detected that the trainee's tension index rises to 70 points, due to the energy ratio of the grip force fluctuation frequency band exceeding 30% and the standard deviation of the voice tremor reaching 25 Hz, the emergency strategy of the adaptive adjustment unit is triggered; the in-vehicle screen switches to the breathing rhythm guidance animation interface, the voice prompt frequency drops to 1 time per minute, and the width of the reverse parking space is temporarily adjusted from the standard 2.5 meters to 2.7 meters to reduce the operation difficulty; if the tension index does not drop below 60 points within 2 minutes, the system pauses the training and sends a remote intervention request to the coach terminal. After the coach unlocks the control right through biometric identification, the coach can manually adjust the training parameters or start the voice soothing mode; When the trainee accidentally steps on the accelerator to 80% depth, the L3 unit triggers a response within 0.2 seconds; the electro-hydraulic braking system issues a slow braking command to reduce the vehicle speed to a standstill within 1.5 seconds and locks the drive control; the steering wheel motor fixes the current steering angle to prevent the vehicle body from shifting; the encrypted communication module sends an operation data packet to the coach terminal through the TLS1.3 protocol, including the accelerator depth curve, vehicle body pitch angle data, and obstacle distance; after the coach completes fingerprint verification, the system plays a 3D accident review animation to mark the moment of the wrong step on the accelerator and the recommended braking timing; the event record is written into a read-only file and synchronized to the cloud for liability tracing and teaching analysis; After the training is completed, the comprehensive operation data is used to update the cognitive load score; the training log is uploaded to the cloud through the differential upgrade technology, and only 15% of the newly added data volume is transmitted to optimize the network load; the coach-side APP generates a daily report including the error type distribution, the stress index curve, and the cognitive load score, and recommends relaxation training courses to help the trainees relieve psychological pressure; the entire implementation process realizes module coordination through the message subscription mechanism of the ROS system, and each unit exchanges data in JSON format to ensure that the response delay is less than 100 milliseconds. From the vehicle start-up self-check to the training end feedback, while ensuring safety, personalized teaching is realized, and the technical closed-loop and logical self-consistency of intelligent driving training are fully presented.

[0028] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent teaching system based on AI robot coach, characterized in that: include: The hardware device module includes an on-board terminal, which is used to perform encrypted communication and real-time preprocessing of multi-mode sensor data and command distribution, and collect vehicle status data, environmental data and trainee physiological data in real time through the multi-mode sensor, and interact with the software platform module for data; The software platform module generates teaching strategies based on the deep reinforcement learning model and drives the hardware device module to execute operation instructions, while analyzing the student behavior patterns to generate personalized learning paths; The graded intervention module is divided into three levels of intervention strategies: L1 unit, L2 unit and L3 unit based on the error types in the trainees’ operation data; wherein the L2 unit dynamically merges or splits the operation step granularity according to the real-time cognitive load score; The safety control module includes a risk classification response unit, which uses radar to monitor obstacle distance, vehicle speed and operation abnormalities in real time and divides them into three risk levels, and works in conjunction with the graded intervention module; and triggers the automatic braking mechanism through obstacle distance and vehicle speed data, generates operation records and guides trainees to complete the safe operation review process; The physiological monitoring module constructs the trainee's psychological state index through the real-time collected trainee's physiological data, generates a stress index and inputs it into the cognitive load adjustment module; The cognitive load adjustment module includes an adaptive adjustment unit, which executes dynamic adjustment strategies according to the student's stress index, and builds a cognitive load scoring model based on vehicle status data, stress index and historical error rate to generate a cognitive load score and dynamically adapt the difficulty and information density of the teaching content.

2. The intelligent teaching system based on AI robot coach according to claim 1, characterized in that: The multi-mode sensors in the hardware device module include: a vehicle speed sensor, a steering wheel steering angle sensor, and an accelerator and brake depth sensor for collecting vehicle status data; millimeter wave radar, lidar, and camera for environmental perception; an infrared camera, a steering wheel piezoelectric film sensor, and a noise reduction microphone for monitoring trainee physiological data. The camera and lidar are used to detect the vehicle's lane deviation in real time based on the position relationship between the vehicle and the lane line.

3. The intelligent teaching system based on AI robot coach according to claim 2, characterized in that: The operation data includes steering wheel steering angle deviation, accelerator and brake depth abnormal values, and operation timing logic errors. In the L1 unit, when a steering wheel response delay or line pressure deviation is detected to be less than a dynamic threshold preset in the graded intervention module, the trainee's line of sight focus position is tracked by an infrared camera, and the vehicle screen is driven to superimpose dynamic enhanced virtual markings, and the transparency and color saturation of the markings are adjusted based on the line of sight focus coordinates to match the attention area; The piezoelectric actuator built into the steering wheel is activated synchronously to generate direction-specific tactile feedback. Left deviation triggers high-frequency vibration on the left side, and right deviation triggers high-frequency vibration on the right side. The density of voice content and tone intensity are dynamically compressed in combination with the tension index output by the physiological monitoring module. When the tension index exceeds the preset threshold, the complex command is simplified into a single action keyword and the volume is lowered. The L1 unit divides the duration of the uncorrected operation into preset durations of multiple progressive intervention stages: when the steering wheel response delay or the line pressure deviation is lower than the dynamic threshold and the uncorrected operation continues for a first preset duration, the dynamic enhanced virtual marking line and the single voice command are activated; if the uncorrected operation continues to accumulate to a second preset duration, the direction-specific tactile feedback of the steering wheel is superimposed and a miniature animation is played; when the uncorrected operation continues for more than a third preset duration, the steering angle of the steering wheel is limited to a preset safety range, and the vehicle speed is forced to be reduced to a preset safety threshold through the safety control module; The preset duration of the multi-level progressive intervention stage is dynamically adjusted through the deep Q network model. The input parameters include the real-time cognitive load score, the tension index and the historical operation correction efficiency. The output layer generates a stage duration threshold that matches the current state of the trainee. If the trainee succeeds in self-correction twice in a row under high-pressure conditions, the deep Q network model triggers the offline training process to form a personalized progressive intervention strategy. When the cognitive load score exceeds the preset high load threshold, the virtual reticle is dynamically enhanced and non-critical visual elements are automatically hidden; If the tension index exceeds the associated threshold simultaneously, the tactile feedback intensity will be attenuated according to the preset ratio, and the breathing rhythm guidance function will be activated, and deep breathing instructions synchronized with the real-time heart rate will be output through the voice module.

4. The intelligent teaching system based on AI robot coach according to claim 3 is characterized by: The L2 unit includes: when the trainee's consecutive number of errors of the same type reaches a preset threshold or exceeds the maximum allowed value in a single day, it automatically generates dynamically optimized special training content and locks the current teaching project, prohibiting the trainee from entering the next stage; forcibly pushes pre-recorded targeted teaching videos through the vehicle terminal, requiring the trainee to complete the special training module; if the cognitive load score continues to exceed the high load threshold and the intervention of the L1 unit does not achieve the expected effect, the special training is split into a sequence of independent operation steps executed in sequence, and the student's cognitive load score is collected in real time after each step is executed; when the score drops to the safe load range, the subsequent operation steps are unlocked, and if the score still exceeds the safe range, the current operation step is repeated until the standard is met, forming a closed-loop feedback control mechanism; during the execution of the steps, the training difficulty and information density are dynamically adjusted according to the real-time score.

5. The intelligent teaching system based on AI robot coach according to claim 4 is characterized in that: In the graded intervention module, when the cognitive load adjustment module determines that the cognitive load score of the trainee is in a preset high load interval for a continuous time period, the following operations are performed: The current special training content is divided into a sequence of independent operation steps to be executed in sequence according to the operation logic. Each step corresponds to a single action instruction and sets a target score threshold. The steering wheel steering angle sensor, accelerator and brake depth sensor and physiological monitoring module are used to collect the steering wheel steering angle deviation value, accelerator and brake alternation frequency and tension index data of the trainee when performing each independent operation step in real time; After the trainee completes each step, the cognitive load scoring model is called to perform scoring by combining sensor data with historical error rate calculation steps; If the step execution score is lower than the preset lower limit threshold of the safe load range, the step repetition mechanism is triggered: the auxiliary virtual markings and dynamic correction arrows of the current step are superimposed on the vehicle screen, and a micro-teaching video clip containing a decomposed action demonstration is generated; the trainee is forced to repeat the current step at least 3 times, and the step execution score is recalculated after each execution; When the step execution score reaches the safe load range twice in a row, the next operation step is unlocked; If the step execution score is in the safe load range, then if the score is in the safe range, an unlocking instruction is sent to the vehicle terminal to enter the next step, and the auxiliary virtual marking line of the current step is cleared; When the step execution score is lower than the lower limit of the safe load range for two consecutive times, the adaptive adjustment unit triggers the interface information simplification operation, turns off non-critical display information and extends the voice prompt interval; at the same time, the L2 unit dynamically merges or splits the operation step granularity according to the real-time cognitive load score.

6. The intelligent teaching system based on AI robot coach according to claim 5, characterized in that: In the L3 unit, when it is detected that the trainee's misoperation causes the vehicle acceleration to exceed the preset safety range, the L3 unit triggers the slow braking mechanism and locks the vehicle driving control right, and at the same time sends a data packet containing the operation record to the coach end through an encrypted communication protocol; The coach needs to unlock the control right through biometric verification, and guide the trainee to complete the safe operation review process based on the operation record; the trigger logic of the L3 unit is dynamically associated with the trainee's psychological state index output by the physiological monitoring module and the cognitive load score of the cognitive load adjustment module. When the trainee's psychological load reaches the preset threshold, the braking intervention of the L3 unit is triggered first. The operation record guides the trainee to complete the safe operation review process through a visual interface. The review process includes error node marking and operation scene restoration.

7. The intelligent teaching system based on AI robot coach according to claim 6, characterized in that: The adaptive adjustment unit executes a dynamic adjustment strategy according to the trainee's tension index: when the tension index reaches a first preset threshold, the vehicle screen is automatically switched to a breathing rhythm guidance animation interface, the voice prompt frequency is reduced to once per minute, and the training difficulty parameter is dynamically adjusted; if the tension index continues to be higher than a second preset threshold, the current training process is paused and a remote intervention request is sent to the coach; at the same time, combined with the cognitive load score, when the cognitive load score reaches or exceeds a specified threshold, an information density optimization strategy is executed, non-critical display information is turned off, only the vehicle's real-time trajectory and key ground markings are retained, and the voice prompt interval is extended to 1.5 times the original interval, thereby reducing the trainee's cognitive load and improving the training effect.

8. The intelligent teaching system based on AI robot coach according to claim 7, characterized in that: The risk grading response unit monitors the obstacle distance, vehicle speed and abnormal operation in real time through radar, and divides the risk into three levels: the first level risk is that the obstacle distance is less than the preset safety threshold and the relative speed exceeds the limit, triggering the L3 unit emergency braking; the second level risk is that the obstacle distance is within the warning range and the trainee does not perform the braking operation, triggering the L2 unit intervention strategy; The third level of risk is when the seat belt is not fastened or the accelerator is accidentally stepped on, which triggers a voice alarm and limits the vehicle speed; the risk classification response unit also receives the scoring data of the cognitive load adjustment module. When the cognitive load score is ≥75 points, even if the physical risk threshold is not reached, the L2 unit intervention strategy is still forcibly triggered to lock the current teaching project and generate a special training module.

9. The intelligent teaching system based on AI robot coach according to claim 8, characterized in that: The hardware device module and the software platform module are connected through an encrypted communication protocol, wherein the vehicle-mounted terminal also has a built-in industrial-grade embedded system, which realizes high-precision positioning through real-time correction of multi-satellite positioning signals, and synchronizes the vehicle status based on the millisecond-level response of sensor data; during the ignition self-starting stage, the embedded system performs integrity verification on the positioning signal, sensor equipment and safety devices, including positioning signal delay detection, sensor calibration error verification and brake system pressure testing. If the verification fails, the training process is prohibited from starting and a fault code is generated.