A five-sense immersive physical cabin environment adaptive adjustment system

Abstract

The application discloses a five-sense immersive entity cabin environment adaptive adjustment system and belongs to the technical field of intelligent environment control and man-machine interaction. The system comprises a multi-modal sensing layer, which synchronously collects electroencephalogram signals, heart rate variability, temperature and humidity and light intensity; a fusion prediction model adopts a multi-task deep learning network, takes relative power of brain waves, heart rate LF / HF ratio and user subjective comfort evaluation value as joint supervision signals, shares a feature extraction layer and outputs a current comfort demand value; a control execution layer dynamically adjusts light color temperature and illumination, audio content and volume, temperature and fragrance type according to the comfort demand value; and an online preference learning module can realize individualized adaptation. The application takes endogenous physiological signals of users as a control basis, realizes accurate comfort prediction through multi-modal fusion and deep learning, solves the problem that existing environment adjustment depends on threshold triggering and significantly improves the user experience of the immersive cabin.

Description

Technical Field

[0001] This invention relates to the field of intelligent environmental control and human-computer interaction technology, and in particular to a five-sense immersive physical cabin environment adaptive adjustment system. Background Technology

[0002] With increasing work and study pressures in modern society, people have an increasingly urgent need for spaces that provide personalized and comfortable environments. Physical capsule products such as study pods, rest pods, and meditation pods are gradually emerging, their core value being to create a private environment for users, undisturbed by external factors, allowing them to quickly enter a state of focus or relaxation. However, existing capsule environment control technologies have the following significant drawbacks: The adjustment is based on a single and lagging factor: Most cabins use manual control panels or simple timer switches, requiring users to actively adjust the lights, audio, and temperature. This operation is cumbersome and cannot respond to changes in the human body's state in real time. The few systems with automatic adjustment functions make decisions based solely on external environmental parameters such as temperature, humidity, and carbon dioxide concentration, lacking the ability to perceive the human body's internal physiological state (such as mental fatigue, emotional tension, and relaxation level), resulting in a serious disconnect between environmental regulation and the user's actual physical and mental needs.

[0003] The simplistic logic relying on threshold triggers: Existing automatic adjustment systems typically employ rule engines that "execute an action if environmental parameters exceed a preset threshold," such as "turn on cooling if the temperature is above 26°C." This approach cannot handle the nonlinear coupling between multi-dimensional environmental parameters and complex physiological signals, nor can it achieve dynamic prediction based on the user's cognitive state. For example, when a user is focused on studying, even if the ambient temperature is suitable, they may need to appropriately lower the color temperature to alleviate visual strain due to mental fatigue, a situation that threshold triggering systems cannot address.

[0004] Ignoring individual differences and dynamic preferences: Different users experience significant differences in comfort with the same environmental parameters (such as light color temperature and background volume), and the preferences of the same user also change dynamically at different times and under different levels of fatigue. Existing systems lack online learning and personalized adaptive capabilities, and cannot continuously optimize adjustment strategies based on users' manual correction behaviors.

[0005] Insufficient depth of multimodal signal fusion: A few studies have attempted to use brain-computer interfaces or heart rate sensors for emotion recognition, but they usually only use a single signal source (such as EEG or heart rate only) and the signal processing methods are simple (such as extracting the time-domain mean). They have not built a multi-task learning framework with deep feature sharing capabilities, resulting in low prediction accuracy and poor generalization ability.

[0006] Therefore, there is an urgent need for a cabin environment adaptive adjustment system that can integrate multimodal endogenous physiological signals such as EEG and heart rate, combined with deep learning prediction models, to achieve refined, personalized, and seamless cabin environment adaptive adjustment. Summary of the Invention

[0007] Purpose of the Invention: The purpose of this invention is to provide a five-sense immersive physical cabin environment adaptive adjustment system and method to solve the technical problems of existing technologies, such as single-based environmental adjustment, reliance on threshold triggering logic, lack of physiological signal fusion analysis, and individual adaptive capabilities. By collecting the user's endogenous physiological signals (EEG, heart rate variability), a multi-task deep learning model is constructed to predict the user's comfort needs in real time, and to coordinate and regulate multi-dimensional environmental parameters such as light, sound, temperature, and smell inside the cabin, achieving an intelligent immersive experience where "the environment follows the mood."

[0008] Technical solution: A five-sense immersive physical cabin environment adaptive adjustment system, comprising: The multimodal sensing layer comprises a group of sensors that synchronously acquire the following signals at a unified timestamp: A non-invasive EEG acquisition headband is used to acquire prefrontal cortex EEG signals, at least extracting... Wave relative power, relative power of the wave and Wave relative power; The millimeter-wave radar heart rate detection unit continuously extracts heart rate variability features in a non-contact manner, including LF low-frequency power, HF high-frequency power, and LF / HF ratio. Temperature and humidity sensors collect data on the ambient temperature and relative humidity inside the cabin. A light sensor collects the ambient light intensity inside the cabin.

[0009] A fusion prediction model, deployed on an edge computing unit or cloud server, is a deep learning-based multi-task regression model. Its input is an aligned multimodal fusion feature vector, and its output is the current comfort requirement value (a continuous value from 0 to 100). The training strategy for this multi-task regression model employs a multi-task learning framework: integrating EEG data... Wave relative power, heart rate variability (LF / HF ratio), and user subjective comfort evaluation value are used as three joint supervision signals, sharing the parameters of the underlying feature extraction network, enabling the model to learn the implicit mapping relationship between physiological signals and subjective comfort.

[0010] The control execution layer includes a main controller and a multi-dimensional environmental actuator connected to it. The main controller receives the comfort requirement value output by the fusion prediction model and dynamically outputs the following adjustment commands according to the preset multi-dimensional control strategy: Lighting adjustment instructions: When the comfort requirement falls into the first preset range (corresponding to focus requirements), drive the adjustable color temperature LED lamp to output cool white light with a color temperature of 5000K~6500K and an illuminance of not less than 300 lux; when it falls into the second preset range (corresponding to relaxation requirements), output warm yellow light with a color temperature of 2700K~3500K and an illuminance of less than 100 lux. Audio adjustment commands: Switch to the corresponding music library. In focus mode, play background sounds that distract attention (such as white noise, rain sounds, and ambient forest sounds). In relaxation mode, play soothing music (such as piano music and ambient electronic music) and automatically adjust the volume to the preset comfortable loudness (35~45dB). Temperature adjustment command: Fine-tune the set temperature of the air conditioner or semiconductor cooling chip in the cabin. The adjustment step shall not exceed ±1℃ / time, and the minimum time interval between two adjustments shall be 3 minutes to avoid over-adjustment of the temperature and causing user discomfort.

[0011] The online preference learning module is configured to record the user's manual correction behavior and the corresponding environmental state after each environmental adjustment. The correction samples are sent as incremental training data to the fusion prediction model to perform online gradient descent to update the bias parameters of the model and gradually converge to the individual optimal environmental setting mapping function.

[0012] Furthermore, the specific network structure of the multi-task regression model is as follows: the input layer receives the concatenated multimodal feature vectors, which are then passed sequentially through three shared convolutional blocks, each containing a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function; subsequently, it splits into three task-specific branches: the first branch uses EEG... With relative wave power as the monitoring target, the output prediction is... The model is designed to improve performance. The first branch is a power-based model; the second branch uses the heart rate variability (LF / HF) ratio as the supervision target and outputs a predicted LF / HF. The third branch uses the user's subjective comfort rating (collected via the built-in touchscreen, where users rate their overall comfort level over the past two minutes from 1 to 5 stars, mapped to a score of 0 to 100) as the supervision target and outputs a predicted comfort value. The total loss function is the sum of the weighted mean square errors of each branch, and the weight coefficients are automatically learned through uncertainty weighting. After the model is trained, only the output of the third branch is retained as the comfort requirement value.

[0013] Furthermore, the process of generating the multimodal fusion feature vector includes: after bandpass filtering and artifact removal of the EEG signal, extracting the feature vector using a fast Fourier transform with a time window of 5 seconds. The relative power of the frequency bands, and the attention index is calculated ( ) and relaxation index ( Peak detection is performed on the I / Q demodulated signal output by the millimeter-wave radar, the RR interval sequence is extracted, and the time-domain indices SDNN and RMSSD and the frequency-domain indices LF, HF, and LF / HF are calculated. The temperature, humidity, and illumination data are filtered by moving average. All the above features are normalized and then concatenated into a feature vector of dimension D, where D ranges from 20 to 50.

[0014] Furthermore, the control execution layer also includes an olfactory actuator, which is a replaceable fragrance capsule release device that selects the corresponding fragrance type according to the comfort requirement value: in focus mode, it releases rosemary or lemon to refresh the mind; in relaxation mode, it releases lavender or sandalwood to calm the mind, thereby achieving immersive regulation of the five senses (sight, hearing, touch / temperature, and smell).

[0015] Furthermore, the multi-dimensional control strategy preset in the main controller is generated using a deep Q-network pre-training to overcome the discontinuity of simple threshold triggering. The comfort requirement value E and the current environmental state vector S (current temperature, humidity, light intensity, and volume) are used as inputs, and a pre-trained deep Q-network outputs the optimal control action. The training environment of this network is a digital twin environment simulating the user's physiological response.

[0016] This invention also provides a method for adaptive adjustment of the five-sense immersive physical cabin environment based on the above system, comprising the following steps: S1. Multimodal signal synchronous acquisition: Start all sensors with a unified system timestamp (millisecond-level synchronization) to acquire signal segments with a duration of T seconds, where T≥5, to obtain raw EEG time-series data, heart rate pulse sequence, instantaneous temperature and humidity values, and light intensity values.

[0017] S2, Multimodal Fusion Feature Extraction: Preprocessing and feature engineering are performed on the signals acquired in S1 to generate fusion feature vectors; among which, EEG feature extraction includes independent component analysis for artifact removal and wavelet packet decomposition; heart rate feature extraction includes adaptive threshold peak detection and abnormal RR interval correction; environmental parameters are taken as the mean within the window.

[0018] S3. Comfort Need Prediction: Input the fused feature vector obtained in S2 into the trained multi-task regression model, and perform forward inference to obtain the current comfort need value E (0~100). If E≥70, it is determined to be a focus need; if E≤40, it is determined to be a relaxation need; if it is in between, it is a neutral maintenance mode.

[0019] S4. Multi-dimensional environmental coordinated control: Based on the E value and the current environmental state, the main controller retrieves or calculates in real time the light color temperature C, light illuminance L, audio track ID, audio volume V, temperature setpoint T_set, and fragrance type F from the preset control strategy library. After receiving the command, the actuator completes the adjustment within a limited time (light response < 0.5 seconds, audio < 0.2 seconds, temperature change rate ≤ 0.5℃ / minute).

[0020] S5. Online Preference Learning and Model Fine-tuning: Monitors whether the user manually corrects the settings via the in-cabin control panel within a preset time (e.g., 3 minutes). If a correction is detected, the current environmental state and the corrected control action are stored as preference samples. An online gradient descent update is triggered every N samples (N≥10), updating only the weight parameters of the last two layers of the model while keeping the underlying shared feature extractor fixed to achieve rapid personalized adaptation.

[0021] Beneficial effects: Endogenous physiological regulation: using EEG With endogenous signals such as wave relative power and heart rate LF / HF ratio as core inputs, it directly reflects the user's focus / relaxation state, realizing proactive predictive adjustment of "environment following mood". It can optimize the environment before the user feels fatigued, thereby improving focus duration and rest efficiency.

[0022] Multimodal fusion with high precision: It adopts a multi-task learning framework, which forces the sharing of deep features of EEG, heart rate and subjective rating, which is significantly better than the single signal source model. It has the ability to handle complex nonlinear mappings, which is different from the simple rules of traditional threshold triggering.

[0023] Gradual and seamless adjustment: gradual changes in light brightness, limited temperature increments, and fade-in / fade-out audio ensure that users are almost unaware of the environmental parameter switching process, maintaining the continuity of the immersive experience.

[0024] Online personalized adaptation: It records each manual correction behavior of the user and updates the model bias parameters in real time through online gradient descent. It can converge to the individual optimal preference in about 10 interactions, which greatly improves user satisfaction. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention.

[0026] Figure 2 This is a schematic diagram of the structure of the multi-task deep learning model of the present invention.

[0027] Figure 3 This is a flowchart illustrating the environmental adaptive adjustment method of the present invention. Detailed Implementation

[0028] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] Example 1: Application of the Learning Focus Pod This embodiment provides a five-sense immersive learning and focus cabin. The cabin's external dimensions are: 1.2m long × 1.0m wide × 2.0m high. The interior features an ergonomic seat, a foldable table, a surround LED light strip, a roof-mounted Bluetooth speaker, a small semiconductor-cooled air conditioner, and a replaceable fragrance box. The overall system architecture is as follows: Figure 1 As shown.

[0030] 1. Multimodal sensing layer configuration EEG acquisition headband: A custom-designed frontal single-lead dry electrode headband using the NeuroSkyThinkGear ASIC module, with a sampling rate of 512Hz, and a reference electrode placed behind the ear. Real-time extraction. relative power of wave (4~8Hz), relative power of wave (8~13Hz), Wavelength (13~30Hz) relative power. The headband transmits EEG signature packets to the cabin control host via Bluetooth 5.0.

[0031] Millimeter-wave radar heart rate detection unit: Utilizing a Texas Instruments IWR684 360GHz radar module, it is installed 1 meter in front of the seat, level with the user's chest. The radar transmits frequency-modulated continuous waves, and after receiving the reflected signal, it extracts the chest cavity micro-displacement signal through range FFT and Doppler FFT. Heart rate pulses are extracted using an adaptive moving average filter, and the RR interval sequence is calculated. Every 5 seconds, it outputs a set of HRV features: MeanHR, SDNN, RMSSD, LF (0.04~0.15Hz), HF (0.15~0.40Hz), and LF / HF.

[0032] Temperature and humidity sensor: The SHT30 digital sensor is used, with an accuracy of ±0.2℃ and ±2%RH. It is installed on the rear wall of the cabin at a height of 1.5m above the ground, with a sampling interval of 1 second.

[0033] Light sensor: A BH1750 light intensity sensor, with a range of 0~65535 lux, is installed on the cabin ceiling to measure the current ambient illuminance. Color temperature is indirectly obtained through feedback values ​​from the LED driver.

[0034] All sensors are connected to the main controller via an RS485 bus. The main controller has a built-in high-precision real-time clock (RTC) that aligns data with millisecond-level timestamps. Each acquisition window is 10 seconds long, with a sliding step of 2 seconds (overlap rate 80%).

[0035] 2. Construction and Training of Multi-Task Deep Learning Models Model structure as follows Figure 2 As shown. The input feature vector has a dimension D=36, and its specific components are: EEG features ( relative power and The original features are: (ratio) × 4 + HRV features (SDNN, RMSSD, LF, HF, LF / HF) × 5 + environmental parameters (temperature, humidity, illuminance, color temperature estimation) × 4, for a total of 13 original features. The first-order difference (change in the window before and after) and the second-order difference are additionally concatenated for each feature, resulting in 39 dimensions (36 dimensions are actually taken according to the design).

[0036] The shared feature extraction part consists of three one-dimensional convolutional blocks: Conv1: Convolution kernel size 3, stride 1, output channels 32, followed by BatchNorm and ReLU; Conv2: kernel size 5, stride 2, output channels 64; Conv3: Kernel size 3, stride 1, output channels 128; Then flatten it and connect it to a fully connected layer (output 128 dimensions).

[0037] All three task branches use a two-layer fully connected network (128→64→1) to output the predicted θ power, predicted LF / HF, and predicted comfort value, respectively.

[0038] Training data was collected from 50 volunteers (aged 22-45, half male and half female). Each volunteer used the cabin 6 times, each session consisting of a 20-minute focus task (reading technical documents) and a 20-minute relaxation task (closed-eye meditation). During the tasks, a touchscreen popped up every 2 minutes asking, "Please rate your overall comfort level over the past 2 minutes" (1-5 stars, mapped to 0-100 points). Approximately 6000 labeled samples were obtained. The Adam optimizer was used with an initial learning rate of 0.001, a batch size of 64, 50 training epochs, and an early stopping mechanism (the training stopped if the validation set loss did not decrease for 5 consecutive epochs). The final mean squared error (MSE) of the comfort prediction branch on the validation set was 68.3 (corresponding to a difference of approximately 8.3 points between the predicted and actual values). R... 2 =0.87.

[0039] 3. Environmental regulation strategy library The control strategy employs a combination of deep Q-networks and fuzzy rules. First, a deep Q-network is trained using historical data, with a state space of (comfort prediction value E, current temperature, current illuminance, current volume, current color temperature) and an action space of: Light color temperature: Discretized into 5 levels (2700K, 3500K, 4500K, 5500K, 6500K); Light intensity: continuously adjustable (50~500 lux), with an actual step size of 50 lux. Audio library: Focus library (numbers A001~A050), Relaxation library (numbers B001~B050); Volume: 35dB, 40dB, and 45dB (three levels); Temperature setting: 20℃~26℃, in increments of 0.5℃; Fragrance: Rosemary (focus), Lavender (relaxation), Unscented (neutral).

[0040] The reward function is defined as follows: if the user does not manually correct the settings within 3 minutes after adjustment, a positive reward of +1 is given; if the user manually corrects to a better setting (judged by subsequent improvement in comfort score), a negative reward of -0.5 is given; if the user directly shuts down the system, a negative reward of -2 is given. After training convergence, the optimal action lookup table output by the Q network is fixed to the controller as the initial policy.

[0041] 4. Actual Operation Process One afternoon, User A entered the learning focus pod, put on the EEG headband, and the system automatically started after the pod door was closed.

[0042] t=0~10 seconds: The system acquires baseline data. Initial EEG theta wave relative power is 0.28 (low), alpha wave 0.35, beta wave 0.37, LF / HF=1.2 (normal range). Temperature and humidity sensor reading: 25.3℃, 48%RH. Light sensor reading: 180 lux (from the cabin background light).

[0043] t=10 seconds: The model is input with fused feature vectors, and the output comfort requirement value E=78 (≥70) is determined to be a focus requirement.

[0044] t=10~15 seconds: The controller executes focus mode: the LED light strip gradually changes from the current 180 lux to 350 lux, and the color temperature gradually changes from 4000K to 5500K (rate 200 lux / second); the audio switches to music library A, playing rain sound white noise at a volume of 40dB; the air conditioner is reduced from 25.3℃ to 24.8℃ (in 0.5℃ increments, in compliance with regulations); the fragrance releases rosemary scent at a low concentration.

[0045] t = 15 seconds ~ 20 minutes: User A begins reading a document. During this period, a sliding window predicts the E value every 2 seconds. The system records the EEG at the 8th minute. The relative power of the wave increased to 0.42 (indicating signs of fatigue), and the LF / HF ratio increased to 1.8 (indicating high sympathetic nerve excitation). The model output E value decreased from 78 to 65 (approaching neutral). The controller did not switch modes, but slowly reduced the light illuminance to 280 lux and the volume to 38 dB to reduce visual and auditory load.

[0046] At t=20 minutes: User A completes the learning task and provides feedback of "overall comfort level 4 stars" via the in-cabin touchscreen. The system records a subjective score of 85 points, which deviates from the model's predicted value (average score of 72 points during the period). This sample is then entered into the online learning cache.

[0047] At t=21 minutes: The user manually lowered the light color temperature from 5500K to 4800K. The system detected the manual correction and stored the state-action pair before and after the correction in the preference cache. After accumulating 10 samples, an online update was triggered: the Adam optimizer was used to perform a single-step update on the last two fully connected layers of the model with a learning rate of 0.0001 and the loss function was mean squared error. After the update, the model's bias in predicting user A's comfort level was adjusted, and the default color temperature in focus mode gradually converged to approximately 4900K during subsequent use.

[0048] 5. Compare the test results To verify the effectiveness of this invention, a control group was set up: the same cabin was used only with threshold triggering logic (e.g., "warm light + soothing music is activated when EEG theta wave > 0.4 and LF / HF > 1.5"). Thirty subjects alternately used the system of this invention and the control group system, and completed a subjective satisfaction questionnaire (1-10 points) after each use. Results showed: The average satisfaction score of the system of this invention was 8.7 points, while that of the control group was 6.2 points (p<0.001). In this invention system, the average number of manual corrections by the user is 0.8 times / hour, compared to 3.2 times / hour in the control group. The system of this invention successfully identified the transition from "focus to relaxation" with an accuracy rate of 91% (based on user-declared gold standard), compared to 67% in the control group. Regarding model inference latency, the edge computing unit (NVIDIA Jetson Nano) takes 12ms for a single forward inference, which meets the real-time requirements.

[0049] 6. Other application scenarios This system is also applicable to scenarios such as meditation cabins, sleep cabins, and virtual reality experience cabins. Only the comfort requirement range and actuator control parameters need to be adjusted (e.g., turning off all lights, raising the temperature to 26°C, and playing delta-wave induced music in sleep mode). The system supports cloud-based model updates, allowing manufacturers to periodically aggregate anonymized preference correction data, retrain the base model, and push the data to all cabins via OTA, achieving continuous evolution of collective intelligence.

[0050] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A five-sense immersive physical cabin environment adaptive adjustment system, characterized in that, include: The multimodal sensing layer comprises a sensor array that synchronously acquires the following signals at a unified timestamp: A non-invasive EEG acquisition headband is used to acquire prefrontal cortex EEG signals, at least extracting... Wave relative power, relative power of the wave and Wave relative power; The millimeter-wave radar heart rate detection unit continuously extracts heart rate variability features in a non-contact manner, including LF low-frequency power, HF high-frequency power, and LF / HF ratio. Temperature and humidity sensors collect data on the ambient temperature and relative humidity inside the cabin. Light sensor to collect ambient light intensity inside the cabin; The fusion prediction model is deployed on an edge computing unit or a cloud server. The model is a deep learning-based multi-task regression model. Its input is an aligned multimodal fusion feature vector, and its output is the current comfort requirement value (a continuous value from 0 to 100). The training strategy of the multi-task regression model adopts a multi-task learning framework: the relative power of the EEG theta wave, the LF / HF ratio of heart rate variability, and the user's subjective comfort evaluation value are used as three joint supervision signals, sharing the parameters of the underlying feature extraction network, so that the model learns the implicit mapping relationship between physiological signals and subjective comfort. The control execution layer includes a main controller and a multi-dimensional environmental actuator connected to it. The main controller receives the comfort requirement value output by the fusion prediction model and dynamically outputs the following adjustment commands according to the preset multi-dimensional control strategy: Lighting adjustment command: When the comfort requirement value falls within the first preset range, drive the adjustable color temperature LED lamp to output cool white light with a color temperature of 5000K~6500K and an illuminance of not less than 300 lux; when it falls within the second preset range, output warm yellow light with a color temperature of 2700K~3500K and an illuminance of less than 100 lux. Audio adjustment commands: Switch to the corresponding music library; in Focus mode, play background music that distracts attention; in Relax mode, play soothing music. Temperature adjustment command: Fine-tune the set temperature of the air conditioner or semiconductor cooling chip in the cabin, with an adjustment step of no more than ±1℃ / time, and a minimum time interval of 3 minutes between two adjustments; The online preference learning module is configured to record the user's manual correction behavior and the corresponding environmental state after each environmental adjustment. The correction samples are sent as incremental training data to the fusion prediction model to perform online gradient descent to update the bias parameters of the model and gradually converge to the individual optimal environmental setting mapping function.

2. The five-sense immersive physical cabin environment adaptive adjustment system according to claim 1, characterized in that, The specific network structure of the multi-task regression model is as follows: the input layer receives the concatenated multimodal feature vectors, which are then passed through three shared convolutional blocks, each containing a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function; subsequently, it splits into three task-specific branches: the first branch uses EEG... With relative wave power as the monitoring target, the output prediction is... The first branch has a power rating; the second branch uses the heart rate variability (LF / HF) ratio as the supervision target and outputs the predicted LF / HF; the third branch uses the user's subjective comfort evaluation value as the supervision target and outputs the predicted comfort value; the total loss function is the sum of the weighted mean square errors of each branch, and the weight coefficients are automatically learned through uncertainty weighting; after the model is trained, only the output of the third branch is retained as the comfort requirement value.

3. The five-sense immersive physical cabin environment adaptive adjustment system according to claim 1, characterized in that, The process of generating the multimodal fusion feature vector includes: bandpass filtering and artifact removal of the EEG signal, followed by extraction using a fast Fourier transform with a 5-second time window. The relative power of the frequency bands, and the attention index is calculated ( ) and relaxation index ( Peak detection is performed on the I / Q demodulated signal output by the millimeter-wave radar, the RR interval sequence is extracted, and the time-domain indices SDNN and RMSSD and the frequency-domain indices LF, HF, and LF / HF are calculated. The temperature, humidity, and illumination data are filtered by moving average. All the above features are normalized and then concatenated into a feature vector of dimension D, where D ranges from 20 to 50.

4. The five-sense immersive physical cabin environment adaptive adjustment system according to claim 1, characterized in that, The control execution layer also includes an olfactory actuator, which is a replaceable fragrance capsule release device that selects the corresponding fragrance type according to the comfort requirement: in focus mode, it releases a refreshing scent of rosemary or lemon; in relaxation mode, it releases a calming scent of lavender or sandalwood.

5. The five-sense immersive physical cabin environment adaptive adjustment system according to claim 1, characterized in that, The multi-dimensional control strategy preset in the main controller is generated by pre-training a deep Q-network to overcome the discontinuity of simple threshold triggering. Specifically, the comfort requirement value E and the current environmental state vector S are taken as inputs, and the optimal control action is output through a pre-trained deep Q-network. The training environment of this network is a digital twin environment that simulates the user's physiological response.

6. The five-sense immersive physical cabin environment adaptive adjustment system according to claim 1, characterized in that, The millimeter-wave radar heart rate detection unit uses a 60GHz frequency-modulated continuous wave radar module, which is installed inside the cabin at a distance of 0.5~1.5m from the user's chest. It extracts the chest cavity micro-displacement signal through distance FFT and Doppler FFT, and extracts the heart rate pulse sequence through an adaptive moving average filter.

7. The five-sense immersive physical cabin environment adaptive adjustment system according to claim 1, characterized in that, When the control execution layer executes the light adjustment command, the illuminance change rate does not exceed 100 lux / second; when executing the audio adjustment command, a fade-in / fade-out switching method is used; when executing the temperature adjustment command, the temperature change rate does not exceed 0.5℃ / minute.

8. A method for adaptive adjustment of the environment in a five-sense immersive physical cabin, characterized in that, This method is implemented based on the system described in any one of claims 1 to 7, and specifically includes the following steps: S1. Multimodal signal synchronous acquisition: Start all sensors with a unified system timestamp, acquire signal segments with a duration of T seconds, T≥5, and obtain raw EEG time-series data, heart rate pulse sequence, instantaneous temperature and humidity values, and light intensity values. S2, Multimodal Fusion Feature Extraction: Preprocessing and feature engineering are performed on the signals acquired in S1 to generate fusion feature vectors; among which, EEG feature extraction includes independent component analysis for artifact removal and wavelet packet decomposition; heart rate feature extraction includes adaptive threshold peak detection and abnormal RR interval correction; environmental parameters are taken as the mean within the window; S3. Comfort Need Prediction: Input the fused feature vector obtained in S2 into the trained multi-task regression model, and forward inference to obtain the current comfort need value E, 0~100; if E≥70, it is determined to be a focus need; if E≤40, it is determined to be a relaxation need; if it is in between, it is a neutral maintenance mode. S4. Multi-dimensional environmental coordinated control: The main controller retrieves or calculates in real time the light color temperature C, light illuminance L, audio track ID, audio volume V, temperature setpoint T_set, and fragrance type F from the preset control strategy library based on the E value and the current environmental status; the actuator completes the adjustment within a limited time after receiving the instruction. S5. Online Preference Learning and Model Fine-tuning: Monitor whether the user has made manual corrections through the cabin control panel within the preset time after adjustment; if a correction action is detected, the current environmental state and the corrected control action are stored as preference samples. Every N accumulated samples trigger an online gradient descent update, updating only the weight parameters of the last two layers of the model, keeping the underlying shared feature extractor fixed.

9. The five-sense immersive physical cabin environment adaptive adjustment method according to claim 8, characterized in that, The fused feature vector described in step S2 also includes the first-order and second-order differences of each original feature to capture the dynamic change trend of physiological signals.

10. The five-sense immersive physical cabin environment adaptive adjustment method according to claim 8, characterized in that, The training data for the multi-task regression model described in step S3 comes from multimodal physiological signals collected from at least 50 volunteers during focus and relaxation tasks, as well as subjective comfort scores actively labeled every 2 minutes. The training uses the Adam optimizer and sets an early stop mechanism.

11. The five-sense immersive physical cabin environment adaptive adjustment method according to claim 8, characterized in that, In step S5, the learning rate for online gradient descent updates is set to 0.0001, and N is between 10 and 20. When a manual correction action is detected, if the correction is consistent with the current model output direction, the reward coefficient is positive; otherwise, a negative reward is applied to accelerate personalized convergence.

12. A physical cabin, characterized in that, The system includes the five-sense immersive physical cabin environment adaptive adjustment system as described in any one of claims 1 to 7.

Images