A Deep Learning-Based Method for Human Orientation Recognition and Adaptive Control of Load-Bearing Surface

Abstract

This invention proposes a deep learning-based method for human orientation recognition and adaptive control of the bearing surface. It captures details of foot pressure distribution using a honeycomb pressure array, combines a dynamic geometric correction algorithm to eliminate interference from shoe shape and standing posture tilt, and utilizes a lightweight spatiotemporal fusion neural network that integrates deep separable convolution, bidirectional GRU, and dual-path attention mechanisms to analyze the true orientation under complex postures. This method overcomes the dependence of traditional solutions on standard standing postures, maintaining high orientation recognition accuracy even in scenarios such as oblique standing and cross-stepping. Furthermore, it achieves intent judgment through pure pressure sensing, offering advantages in both high reliability and low cost.

Description

Technical Field

[0001] This invention relates to the field of smart bathroom products, specifically a method for human body orientation recognition and adaptive control of bearing surface based on deep learning. Background Technology

[0002] With the rapid popularization of smart bathroom appliances, human body sensing technology has become a core element in enhancing user experience. For example, the automatic lid-opening function of smart toilets effectively solves the hygiene hazards associated with traditional manual operation through contactless interaction design. Currently, it mainly relies on two technological routes: electronic nose recognition and camera visual recognition. Its core value lies in achieving basic human body detection at a low cost, triggering mechanical actions through odor recognition or image contour analysis, providing basic intelligent services for homes and public places. However, as users' demands for accuracy and privacy security increase, these two technologies have gradually revealed significant shortcomings in practical deployment.

[0003] While camera-based recognition technology can accurately distinguish between male and female genders, it carries significant privacy risks. Even after edge computing processing of the raw image data captured by the camera, such as using the YOLOv5 model for local inference, the frame buffer in the preprocessing stage can still be compromised by physical bus attacks, potentially leading to the leakage of sensitive user information. Furthermore, this technology performs poorly in adapting to dynamic scenes. In strong backlight conditions, images are prone to overexposure; and in low-light environments at night, even with infrared illumination, pose misjudgments can still occur, making it difficult to meet the reliability requirements of complex home environments.

[0004] Electronic nose odor recognition is based on androsterone-specific detection, using a metal-oxide-semiconductor array to capture steroid molecules volatilized from male sweat. While this technology avoids visual privacy issues, its biomarkers exhibit cross-sensitivity, and individual metabolic differences can easily lead to serious misjudgments. The sensor's response sensitivity to certain common chemicals highly overlaps with the characteristic peaks of androsterone, easily causing false triggers. Moreover, androsterone secretion levels vary significantly across different age groups and among individuals with endocrine disorders, further increasing the possibility of missed detections.

[0005] Deep learning is a type of machine learning technique based on artificial neural networks. By constructing multi-layered neural network models, it enables computers to automatically learn complex patterns and feature representations from large amounts of data. Convolutional Neural Networks (CNNs) excel at spatial feature extraction, gated recurrent units (GRNs) at temporal feature parsing, and attention mechanisms enable adaptive weighted feature fusion. The breakthroughs achieved by these algorithms in multimodal perception have provided new insights into the behavior prediction of smart bathroom devices. Existing research shows that CNNs exhibit excellent performance in pressure distribution pattern classification and graph convolutional networks in dynamic behavior prediction. Deep learning has been widely and effectively applied in many fields, such as accurate lesion identification in medical image diagnosis, efficient quality inspection in industrial automation processes, and optimized traffic flow prediction in intelligent transportation systems. However, in the specific field of smart bathroom devices, various specific deep learning algorithms have not been closely and effectively integrated with the core functional requirements of smart bathroom devices. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a deep learning-based method for human orientation recognition and adaptive control of the bearing surface. By fusing a high-density pressure sensor array with a lightweight spatiotemporal fusion network, it overcomes the technical bottlenecks of traditional solutions, such as high gender misjudgment rates and poor adaptability to posture tilt caused by reliance on single biometric features. This method achieves integrated orientation recognition and adaptive adjustment of the bearing surface while simultaneously controlling the orientation and avoiding infringement on user privacy. The technical solution provided by this invention is as follows:

[0007] A deep learning-based method for human orientation recognition and adaptive control of bearing surfaces includes the following steps:

[0008] Step 1: The infrared module detects the sensing area. When the intensity of the reflected light from a human body reaches a preset intensity and the signal duration is longer than a preset duration, the user is identified as a valid user.

[0009] Step 2: Obtain plantar pressure distribution matrix data through a honeycomb pressure sensor array, and quantize the data to the 0-255 range using an analog-to-digital converter;

[0010] Step 3: Eliminate tilt error using a dynamic geometric correction unit;

[0011] Step 4: Human orientation recognition is performed using a spatiotemporal fusion network. This network employs a dual-channel heterogeneous architecture, including a spatial feature extraction branch and a temporal feature extraction branch. The spatial feature extraction branch uses depthwise separable convolution to extract the spatial pattern of pressure distribution channel by channel, capturing the pressure gradient distribution of the forefoot and heel to determine the preliminary features of the main pressure direction of the foot. The temporal feature extraction branch uses compressed bidirectional gated recurrent units to capture the temporal pattern of continuous pressure data, analyzing the foot's center of gravity migration rate and acceleration to determine the user's posture stability and dynamic changes in orientation. A dual-path attention mechanism is used to fuse spatiotemporal features and output the probabilities of urination and defecation intentions.

[0012] Step 5: Start the general adjustment model and calculate the target height / tilt angle parameters of the bearing surface;

[0013] Step 6: Obtain the current height and tilt angle of the bearing surface through the sensor. Calculate the adjustment difference based on the target height / tilt angle parameters of the bearing surface and the current height and tilt angle. If the adjustment difference is less than the preset error range, it is determined that no adjustment is needed; otherwise, drive the motor to rotate and adjust the bearing surface to the height / tilt angle parameters.

[0014] Step 7: When the honeycomb pressure sensor array continues for a preset time and the infrared module does not detect human body reflected light, it is regarded as a user leaving signal, and the system starts the automatic reset program: the drive motor runs at the preset reset speed, and the bearing surface height is adjusted to the initial height through the mechanical transmission mechanism, and the system enters the standby state.

[0015] Preferably, step 1 further filters pet interference and momentary false triggering events using an adaptive biometric discrimination algorithm. The biometric anti-interference condition is as follows:

[0016]

[0017] in, The gradient of light intensity change The equivalent body weight is estimated based on the area of ​​the reflected waveform envelope, and the integral term is the integral of the magnitude of the light intensity gradient from time 0 to time t. This is the threshold for biological characteristics.

[0018] Preferably, the quantification method in step 2 is as follows:

[0019]

[0020]

[0021] in, coordinates The quantized pressure pixel value, Coordinates The raw pressure value collected at the location; the clip function implements hardware clamping, where v is the input value to be clamped, a is the lower limit of the clamping interval, and b is the upper limit of the clamping interval.

[0022] Preferably, the infrared module integrates a 940nm wavelength VCSEL light source array and a photodetector, and uses a ±30° wide-angle lens group to achieve biometric detection within a dynamic distance range of 0.3-1.5m; the honeycomb pressure sensing array includes a central sensing area, an edge support area, and an electromagnetic shielding layer between each sensing unit; the central sensing area uses a substrate with differentiated hardness between the central and edge areas, and the sensing unit density is greater than or equal to a preset density; the edge support area uses a soft composite material to form a hexagonal honeycomb reinforcing rib structure; the electromagnetic shielding layer between each sensing unit is composed of a copper-nickel alloy mesh and a ferrite coating.

[0023] Preferably, the dynamic geometric correction unit is configured with a triple error compensation mechanism: the baseline self-calibration module automatically triggers zero-point calibration of the sensor array under no-load conditions every day to eliminate baseline offset caused by environmental temperature drift; the shoe type matching module constructs a database of material hardness-contact area parameters for various shoe types, extracts real-time pressure distribution features through a convolutional neural network, and dynamically matches the optimal compensation coefficient to achieve contact deformation error correction; the tilt compensation algorithm extracts the maximum eigenvector of the pressure distribution covariance matrix based on principal component analysis, calculates the principal orientation angle of the foot, and establishes a mapping relationship between the principal orientation angle of the pressure distribution and the correction matrix through a formula.

[0024] Preferably, in step 5, the collected biometric data, such as weight, height, and shoe size, are input into a pre-trained identity feature database for matching. If historical registered user feature values ​​exist, the preset bearing surface parameters corresponding to that user are directly retrieved.

[0025] Preferably, in step 6, during the adjustment process, pressure sensor array data is collected in real time to monitor the uniformity of pressure distribution on the bearing surface: if the pressure difference in the edge area exceeds the preset balance threshold, a fine-tuning mechanism is triggered to correct the height with a small amplitude until the pressure distribution is restored to balance; if the pressure distribution is uniform, it is continuously adjusted at a preset speed until the target height is reached.

[0026] Preferably, step 6 also includes a ceramic heating module, which uses gradient power control technology to raise the contact temperature of the bearing surface from room temperature to a constant temperature of 36±1℃ within 10 seconds.

[0027] Preferably, step 7 also includes an ultraviolet disinfection device that is automatically activated after the user leaves to perform a 5-minute sterilization program.

[0028] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: Within a certain range of user standing posture tilt angle, the method of this invention achieves a human orientation recognition accuracy of ≥92% based on dynamic geometric correction and spatiotemporal feature fusion. Furthermore, under non-standard standing postures such as oblique standing or cross-stepping, the simulated recognition accuracy decreases by no more than 5%, significantly outperforming traditional visual / infrared solutions. Through the hardware reuse design of a cellular array and the compression of a depthwise separable convolutional network, the core sensing and processing modules are designed at low cost, resulting in low daily power consumption. This invention completely abandons biometric technologies such as face and fingerprint recognition, relying solely on the spatiotemporal pattern of foot pressure distribution to achieve anonymized intent judgment, thus avoiding the risk of user privacy leakage from the data collection source. Simultaneously, through dual verification of infrared biometrics and pressure distribution, it filters out misidentification caused by pets and transient interference, resulting in a high anti-interference rate. Attached Figure Description

[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0030] Figure 1 This is the main flowchart of the present invention;

[0031] Figure 2 This is a simulation comparison diagram showing the difference before and after correcting the oblique standing posture of sports shoes according to the present invention;

[0032] Figure 3 This is a timing simulation in a specific scenario of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] To make the above-mentioned objectives, features and effects of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] Example 1: A deep learning-based method for human orientation recognition and adaptive control of bearing surface, such as... Figure 1 As shown, it includes the following steps:

[0036] Step 1: The infrared module detects the sensing area. When the intensity of reflected light from a human body reaches a preset intensity and the signal duration exceeds a preset duration, a valid user identification flag is triggered. The combined triggering condition of light intensity and time is as follows:

[0037]

[0038] Where L(t) is the infrared light intensity reflected by the object in the sensing area collected by the infrared module at time t, a is the preset light intensity threshold, and b is the preset continuous duration for the reflected light intensity to be greater than the threshold. This is an indicator function; it takes the value 1 if the condition within the parentheses is met, and 0 otherwise. In this embodiment, a is set to 200 lux and b is set to 2 seconds.

[0039] Furthermore, an adaptive biometric discrimination algorithm is used to filter out pet interference and momentary false triggering events. The biometric anti-interference condition is as follows:

[0040]

[0041] in, The gradient of light intensity change The equivalent body weight is estimated based on the area of ​​the reflected waveform envelope, and the integral term is the integral of the magnitude of the light intensity gradient from time 0 to time t. The biometric threshold is preferably set at 15 kg.

[0042] In summary, the comprehensive discriminant is:

[0043]

[0044] in, To determine the overall confidence level for effective human identification, condition 1 is the indicator function based on the fundamental triggering condition (light intensity + duration), and sigmoid is the S-shaped activation function. For the adaptive calibration coefficient of the equivalent weight term, This represents the adaptive calibration coefficient for the light intensity gradient term. When... Time-triggered effective identification is achieved, and light intensity conditions are ensured to be continuously triggered through time integration. Highly reliable human body identification is achieved through a dual verification mechanism (physical light intensity + biometric features).

[0045] The infrared module integrates a 940nm wavelength VCSEL light source array and a photodetector, employing a ±30° wide-angle lens group to achieve biometric detection within a dynamic distance range of 0.3-1.5m. This module achieves ultra-low power operation through adaptive pulse drive technology, specifically employing a timing control strategy that alternates between a 10ms working cycle and a 990ms sleep cycle. During the working cycle, target detection, distance calculation, and signal verification are completed, while maintaining a power consumption level of <0.1mW in standby mode. When a valid biological target is detected, the module outputs a wake-up signal to trigger the pressure sensing array to enter working mode.

[0046] Step 2: Acquire plantar pressure distribution matrix data using a honeycomb pressure sensor array. The data is quantized to the 0-255 range using an analog-to-digital converter. In this embodiment, the honeycomb pressure sensor array initiates dynamic scanning at a 20Hz sampling rate, continuously capturing 5 frames of 64×64 pixel pressure matrix data, with a total acquisition time of 250ms. Hardware-level clamping processing is applied to pressure data exceeding the 100kPa range. The quantization method is as follows:

[0047]

[0048]

[0049] in, coordinates The quantized pressure pixel value, Coordinates The raw pressure value collected at the location; the clip function implements hardware clamping, where v is the input value to be clamped, a is the lower limit of the clamping interval, and b is the upper limit of the clamping interval.

[0050] The honeycomb pressure sensing array includes a central sensing area, an edge support area, and an electromagnetic shielding layer between each sensing unit. The central sensing area uses a substrate with a different hardness between the central and edge areas (such as an elastomer material with a hardness of 70 Shore A), and the sensing unit density is greater than or equal to a preset density. The edge support area uses a TPU composite material with a hardness of Shore 60A to form a hexagonal honeycomb reinforcing rib structure. The electromagnetic shielding layer between each sensing unit is composed of a copper-nickel alloy mesh and a ferrite coating.

[0051] Step 3: Eliminate tilt error using a dynamic geometric correction unit. Call the pre-stored shoe feature database, containing material hardness and contact area parameters for 10 shoe types including athletic shoes, leather shoes, and slippers. Match the deformation compensation coefficient corresponding to the current shoe type using a convolutional neural network. Then, use an improved cosine compensation algorithm to eliminate tilt angle interference. The formula is:

[0052]

[0053] in, For the corrected pressure, This is the baseline pressure under no-load conditions, used to eliminate zero-point drift. The deformation compensation coefficient for the shoe shape is obtained from a shoe shape feature database. This represents the real-time tilt angle between the sensor array plane and the horizontal plane.

[0054] A simulation comparison was conducted to examine the differences in posture before and after correction for an inclined stance in athletic shoes. The experimental results are as follows: Figure 2The left image shows the original pressure data when a user stands at an angle wearing hard-soled shoes, showing localized high pressure concentration (bright areas); the right image reconstructs the pressure distribution through a dynamic geometric correction algorithm (shoe shape compensation and tilt error elimination), making it more in line with the real foot shape and significantly improving the accuracy of recognizing posture orientation and behavioral intention.

[0055] The dynamic geometric correction unit is equipped with a triple error compensation mechanism: the baseline self-calibration module automatically triggers zero-point calibration of the sensor array under no-load conditions at 3:00 AM every day to eliminate baseline offset caused by environmental temperature drift; the shoe matching module constructs a database of material hardness-contact area parameters for 10 types of shoes, including sports shoes, leather shoes, and slippers, and extracts real-time pressure distribution features through a convolutional neural network to dynamically match the optimal compensation coefficient to achieve contact deformation error correction; the tilt compensation algorithm extracts the maximum eigenvector of the pressure distribution covariance matrix based on principal component analysis, calculates the principal direction angle of the foot, and establishes a mapping relationship between the principal direction angle of the pressure distribution and the correction matrix through a formula to achieve pressure field reconstruction within a tilt angle range of ±25°.

[0056] Step 4: Human orientation recognition is performed through a spatiotemporal fusion network. The spatiotemporal fusion network adopts a dual-channel heterogeneous architecture design, including a spatial feature extraction branch and a temporal feature extraction branch.

[0057] The spatial feature extraction branch employs depthwise separable convolution, using a 3×3 convolution kernel to extract the spatial pattern of pressure distribution channel by channel, focusing on capturing the pressure gradient distribution of the forefoot and heel to determine the preliminary features of the main pressure direction of the foot; where the size of the l-th layer convolution kernel is... The parameter compression amount meets the requirements. .

[0058] Spatial branching extracts forefoot pressure gradient distribution features and identifies abrupt changes in heel contact area. The calculation formula is as follows:

[0059]

[0060] The intensity of the pressure variation in the left-right (horizontal) direction. This represents the intensity of pressure changes in the forward and backward (vertical) directions. The changes in both directions are combined, and the absolute values ​​of the horizontal and vertical changes are added together, including both "sudden increases" and "sudden decreases" in the statistics.

[0061] Temporal feature extraction branch: This branch utilizes compressed bidirectional gated recurrent units to capture the temporal patterns of 5 frames of continuous pressure data, analyzing the foot center of gravity migration rate and acceleration to determine the user's posture stability and dynamic changes in orientation. The number of hidden layer units is compressed to 32 dimensions, with residual connections added every 3 time steps, and residual coefficients... Based on decay, acceleration is calculated using the following formula:

[0062]

[0063] Where a is the resultant acceleration of the center of mass of the foot. Let t be the position vector of the plantar pressure centroid at time t, and 0.5 be the centroid acceleration trigger threshold.

[0064] Dynamic feature fusion: Spatiotemporal features are fused through a dual-path attention mechanism while satisfying the following constraints:

[0065]

[0066] in, For spatial feature channels, attention weights Temporal attention weights are used to represent time features; the Softmax function outputs two types of behavior probabilities: urination intention probability P1 (triggered by threshold > 85%) and defecation intention probability P2 (triggered by threshold > 75%). Combining features from different dimensions, the probability is displayed as a percentage.

[0067]

[0068]

[0069] in, Let denot be the probability of the intention to urinate, and sigmoid be the sigmoid activation function. These are the characteristic values ​​of the pressure gradient. The characteristic value of the center of mass acceleration, To solve for the probability of behavioral intent. This represents the characteristic value of the effective contact area of ​​the sole of the foot. The characteristic value for the duration of stable standing.

[0070] Step 5: Activate the general adjustment model and calculate the target bearing surface parameters using the ergonomic mapping function. Further, input the collected biometric data, such as weight, height, and shoe size, into a pre-trained identity feature database for matching. If historical registered user feature values ​​exist, directly retrieve the preset bearing surface parameters corresponding to that user.

[0071] Step 6: Obtain the current real-time height / tilt angle of the target bearing surface (such as a smart seat, adaptive control panel, etc.) using an angle sensor or displacement sensor, and calculate the adjustment difference based on the target bearing surface parameters and the real-time status. When the adjustment difference is less than the preset error range, it is determined that no adjustment is needed; otherwise, drive the motor to rotate and adjust the bearing surface to the target height / tilt angle.

[0072] Furthermore, during the adjustment process, the system collects pressure sensor array data in real time to monitor the uniformity of pressure distribution on the bearing surface: if the pressure difference in the edge area exceeds the preset balance threshold, a fine-tuning mechanism is triggered to correct the height with a small amplitude until the pressure distribution is restored to balance; if the pressure distribution is uniform, it continues to adjust at a preset speed until the target height is reached.

[0073] Furthermore, the present invention can also be configured with a ceramic heating module, which adopts gradient power control technology to raise the contact temperature of the bearing surface from room temperature to a constant temperature of 36±1℃ within 10 seconds.

[0074] Step 7: When the pressure sensor array has no signal for a preset duration and the infrared sensor does not detect human body reflected light, it is regarded as a user leaving signal. The system starts the automatic reset program: the drive motor runs at the preset reset speed, and the bearing surface height is adjusted to the initial height through the mechanical transmission mechanism. The system enters the standby state.

[0075] Furthermore, the present invention can also be configured with an ultraviolet disinfection device (wavelength 275nm), which is automatically activated after the user leaves and performs a 5-minute sterilization program (applicable to public service surfaces).

[0076] Example 2: Taking a smart toilet as a preferred embodiment, this invention demonstrates that the method is applicable to the intelligent transformation of bathrooms in aging societies and public places. The technical solutions in this embodiment are described in detail below in conjunction with actual scenarios.

[0077] A male urinating scene (standing at a 35° angle). First, pressure detection revealed that the user was wearing size 42 athletic shoes. When standing with the left foot tilted forward at 35°, the honeycomb pressure array detected a concentrated load in the forefoot area, with a peak pressure of 85 kPa (coordinates [28,41]) and a contact area of ​​approximately 120 cm². Dynamic correction of the standing position was then performed: the dynamic geometric correction unit called upon the athletic shoe material parameters (Shore A hardness 65±3, elastic modulus 1.8 MPa), generated a compensation coefficient of 0.87 through a CNN, and calculated the actual tilt angle θ = 33.7° (error -1.3°). Next, user intent was determined: the spatiotemporal fusion network analyzed the forefoot pressure gradient to be 12.3 kPa / cm and the center of gravity shift acceleration to 0.63 m / s², outputting a urination intent probability P1 = 88.5%. Finally, the electromechanical response was achieved: a dual-path brushless motor synchronously drove the seat ring and cover at 1500 rpm, raising the opening angle to 75° within 1.5 seconds, with a power consumption of 0.12 Wh.

[0078] Female defecation scenario. Pressure sensor performs posture detection: When the user turns 55° to the right to prepare to sit down, the pressure distribution shows that the load in the heel area accounts for 68% (coordinates [15-20, 45-50]), and the contact area reaches 210cm²; Use mode judgment: The system recognizes the sudden increase in contact area and directly triggers the defecation intention judgment (P2=91.4%); After the judgment, temperature control is executed: The single motor drives the cover to open to 60°, the ceramic heating element heats up at a rate of 3℃ / s, and the seat temperature reaches 36.1℃ within 8 seconds, and the surface temperature-sensing ink changes from blue to orange; After use, the response when leaving the seat: After the weight sensor detects that the load has dropped to 0.5kg, the cover automatically resets and starts the centrifugal exhaust device, and the wind speed is maintained at 0.5m³ / min for 30 seconds; Safety monitoring: The pressure sampling rate drops to 1Hz 10 seconds after the action is completed, the infrared module (102) switches to pulse detection mode, and the overall power consumption of the system drops to 0.9W.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based method for human orientation recognition and adaptive control of bearing surface, characterized in that, Includes the following steps: Step 1: The infrared module detects the sensing area. When the intensity of the reflected light from a human body reaches a preset intensity and the signal duration is longer than a preset duration, the user is identified as a valid user. Step 2: Obtain plantar pressure distribution matrix data through a honeycomb pressure sensor array, and quantize the data to the 0-255 range using an analog-to-digital converter; Step 3: Eliminate tilt error using a dynamic geometric correction unit; Step 4: Human orientation recognition is performed through a spatiotemporal fusion network. The spatiotemporal fusion network adopts a dual-channel heterogeneous architecture design, including a spatial feature extraction branch and a temporal feature extraction branch. The spatial feature extraction branch uses depthwise separable convolution to extract the spatial pattern of pressure distribution channel by channel, capture the pressure gradient distribution of the forefoot and heel, and determine the preliminary features of the main direction of foot pressure. The temporal feature extraction branch uses a compressed bidirectional gated loop unit to capture the temporal pattern of continuous pressure data, analyzes the foot center of gravity migration rate and acceleration, and determines the user's standing posture stability and dynamic changes in orientation; it also uses a dual-path attention mechanism to fuse spatiotemporal features and output the probabilities of urination and defecation intentions. Step 5: Start the general adjustment model and calculate the target height / tilt angle parameters of the bearing surface; Step 6: Obtain the current height and tilt angle of the bearing surface through the sensor. Calculate the adjustment difference based on the target height / tilt angle parameters of the bearing surface and the current height and tilt angle. If the adjustment difference is less than the preset error range, it is determined that no adjustment is needed; otherwise, drive the motor to rotate and adjust the bearing surface to the height / tilt angle parameters. Step 7: When the honeycomb pressure sensor array continues for a preset time and the infrared module does not detect human body reflected light, it is regarded as a user leaving signal, and the system starts the automatic reset program: the drive motor runs at the preset reset speed, and the bearing surface height is adjusted to the initial height through the mechanical transmission mechanism, and the system enters the standby state.

2. The deep learning-based human orientation recognition and adaptive control method for bearing surface as described in claim 1, characterized in that, Step 1 also filters pet interference and momentary false triggering events using an adaptive biometric discrimination algorithm. The biometric anti-interference condition is as follows: ； in, The gradient of light intensity change The equivalent body weight is estimated based on the area of ​​the reflected waveform envelope, and the integral term is the integral of the magnitude of the light intensity gradient from time 0 to time t. This is the threshold for biological characteristics.

3. The deep learning-based human orientation recognition and adaptive control method for bearing surface as described in claim 2, characterized in that, The quantification method for step 2 is as follows: ； ； in, coordinates The quantized pressure pixel value, Coordinates The raw pressure value collected at the location; the clip function implements hardware clamping, where v is the input value to be clamped, a is the lower limit of the clamping interval, and b is the upper limit of the clamping interval.

4. The deep learning-based human orientation recognition and adaptive control method for bearing surface as described in claim 3, characterized in that, The infrared module integrates a 940nm wavelength VCSEL light source array and a photodetector, and uses a ±30° wide-angle lens group to achieve biometric detection within a dynamic distance range of 0.3-1.5m; the honeycomb pressure sensing array includes a central sensing area, an edge support area, and an electromagnetic shielding layer between each sensing unit; the central sensing area uses a substrate with different hardness between the central and edge areas, and the sensing unit density is greater than or equal to a preset density; the edge support area uses a soft composite material to form a hexagonal honeycomb reinforcing rib structure; the electromagnetic shielding layer between each sensing unit is composed of a copper-nickel alloy mesh and a ferrite coating.

5. The deep learning-based human orientation recognition and adaptive control method for bearing surface as described in claim 3, characterized in that, The dynamic geometric correction unit is equipped with a triple error compensation mechanism: the baseline self-calibration module automatically triggers zero-point calibration of the sensor array under no-load conditions every day to eliminate baseline offset caused by environmental temperature drift; the shoe type matching module constructs a database of material hardness-contact area parameters for various shoe types, extracts real-time pressure distribution features through a convolutional neural network, and dynamically matches the optimal compensation coefficient to achieve contact deformation error correction; the tilt compensation algorithm extracts the maximum eigenvector of the pressure distribution covariance matrix based on principal component analysis, calculates the principal orientation angle of the foot, and establishes a mapping relationship between the principal orientation angle of the pressure distribution and the correction matrix through a formula.

6. The deep learning-based human orientation recognition and adaptive control method for bearing surface as described in claim 1, characterized in that, Step 5 involves inputting the collected biometric data, such as weight, height, and shoe size, into a pre-trained identity feature database for matching. If historical registered user feature values ​​exist, the preset bearing surface parameters corresponding to that user are directly retrieved.

7. The deep learning-based human orientation recognition and adaptive control method for bearing surface according to claim 6, characterized in that, Step 6: During the adjustment process, pressure sensor array data is collected in real time to monitor the uniformity of pressure distribution on the bearing surface. If the pressure difference in the edge area exceeds the preset balance threshold, a fine-tuning mechanism is triggered to correct the height with a small amplitude until the pressure distribution is restored to balance. If the pressure distribution is uniform, the adjustment continues at a preset speed until the target height is reached.

8. The deep learning-based human orientation recognition and adaptive control method for bearing surface according to claim 6, characterized in that, Step 6 also includes a ceramic heating module, which uses gradient power control technology to raise the contact temperature of the bearing surface from room temperature to a constant temperature of 36±1℃ within 10 seconds.

9. The deep learning-based human orientation recognition and adaptive control method for bearing surface according to claim 6, characterized in that, Step 7 also includes an ultraviolet disinfection device that automatically activates after the user leaves, performing a 5-minute sterilization process.

Images