Seat safety vibration control system and method based on health state dynamic reasoning

Abstract

The invention provides a seat safety vibration control system and method based on health state dynamic reasoning. The method comprises the steps that physiological signals of a user are collected; analyzing according to the collected physiological signals to obtain a user health state grading result and a corresponding seat vibration and environment control execution scheme; according to the seat vibration and environment control execution scheme, the angles and the heights of the seats in all directions are adjusted, and opening and closing and the amplitude of vibrators of all subareas on all the seats are independently controlled. The intelligent health cabin system breaks through the limitation of single function guidance of traditional seat vibration, constructs an intelligent health cabin system from health monitoring to safety control to continuous optimization, has safety, accuracy, compliance and user experience, can be widely adapted to users in different health states and pathological types, and meets the industry development requirements of health cabins.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of vehicle health monitoring, and particularly relates to a seat safety vibration control system and method based on dynamic reasoning of health states. BACKGROUND

[0002] With the popularization of the concept of "healthy cabin", the seat vibration sub-module has become a mainstream configuration for functions such as physiotherapy and massage, but there are three major contradictions:

[0003] First, the vibration function universality conflicts with medical contraindications. For example, low-frequency vibration may aggravate the nerve compression of patients with lumbar disc herniation, and high-frequency vibration is not conducive to patients with ankylosing spondylitis. However, the existing system lacks pathological recognition ability, and there is a medical risk.

[0004] Second, the contradiction between fixed vibration parameters and dynamic health states. Traditional solutions use fixed frequency / amplitude templates, which cannot adapt to real-time physiological changes of users. For example, low-frequency vibration is needed when sitting for a long time causes a decrease in blood flow in the lumbar muscles, but patients with sciatica need to avoid certain areas, and pregnant women and patients with cardiovascular diseases have contraindications for thoracic spine vibration.

[0005] Third, the lack of sign perception and safety intervention boundaries. Existing systems rely on pressure sensors or timers to trigger vibration, but they cannot obtain real-time muscle strain state, skeletal disease history, and dynamic physiological indicators.

[0006] One of the existing technologies collects driver's heart rate, blood pressure and other data through DMS (Driver Monitoring System) camera, combines with real-time vehicle information, uses cloud "thousand faces" model to judge health status, dynamically updates personal benchmark, and executes four-level response strategy according to health status. However, cloud decision-making may be delayed due to large amount of calculation and long reasoning time, affecting the smoothness of experience and reaction in emergency situations. It does not cover special physiological indicators such as blood glucose, and cannot perform effective medical diagnosis. It does not explain the anti-interference ability of the sensor, such as the possible impact of DMS accuracy under strong light.

[0007] The second existing technology uses an Azure Kinect depth camera integrated with a rearview mirror to capture the driver's full-body depth image, extracts skeletal node coordinates through an industrial computer, calculates motor adjustment amount through a mathematical model, controls the motor through a chip, and feeds back the position to achieve millimeter-level skeletal matching according to ergonomics standards, and improve the adaptive ability of seat posture intelligent adjustment. However, this technical solution has high dependence on hardware, high cost, and may have some impact on the privacy of the driver. The integrated depth camera and industrial computer may increase the complexity and failure rate of the system. SUMMARY

[0008] In order to solve the problems of high latency in health detection and lack of precise and safe solutions for seat vibration control in the prior art, the application provides a seat safety vibration control system and method based on health state dynamic reasoning.

[0009] A seat safety vibration control system based on health state dynamic reasoning, which achieves one of the purposes of the application, comprises:

[0010] The perception module is used to collect user physiological signals, and the physiological signals include HRV, heart rate, blood oxygen, blood pressure, respiratory rate, voice tremor, and the like.

[0011] The local decision module is used to analyze the collected user physiological signals to obtain user health state classification results and corresponding seat vibration and environment control execution schemes. The health state classification results include normal state, pre-warning state and emergency state, the heart rate, blood pressure and anxiety index judgment criteria of each state meet the medical health monitoring specifications. The anxiety index is obtained by extracting the voice tremor data in the physiological signals collected by the perception module. The health state classification results are used to determine the vibration area and parameters by the seat multi-modal vibration execution module. The medical health monitoring specifications are, for example, the industry-recognized heart rate and blood pressure normal range standards.

[0012] The seat multi-modal vibration execution module is used to adjust the angles and heights of the seats according to the seat vibration and environment control execution schemes, and to control the opening and closing and amplitude of the vibrators in each partition of each seat independently, so as to realize the closed loop of user physiological condition perception, focus point positioning, vibrator vibration safety control and real-time physiological signal feedback adjustment.

[0013] Further, the perception module is also used to collect user demographic characteristics and real-time environmental data.

[0014] Further, the local decision module further comprises a demand decision model used to obtain user health state classification results. The demand decision model can be a model carried on a vehicle-grade chip, or a model carried on a processor of a device with human body massage function, such as a massage seat.

[0015] Further, a federal learning module is further included, which is used to adopt a local model and a cloud global model based on a FedProx (Federated Proximal) framework, combine user demographic characteristics, physiological signals and real-time environmental data collected by the perception module and cloud multi-user feature label information, and perform local model training, cloud global model parameter aggregation and medical knowledge graph assisted calibration operation to obtain updated parameters of the local decision module; the local model is used to perform localized training based on user demographic characteristics, physiological signals and real-time environmental data collected by the perception module, and generate its own model parameters to be uploaded to the cloud; the local model can be a sub-model deployed on the vehicle end and associated with the demand decision model, and can also be a sub-model deployed on other devices with human massage function and associated with the demand decision model; the cloud global model is an aggregated model deployed on the cloud, which is used to receive and fuse local model parameters uploaded by multiple users, combine a medical knowledge graph to generate optimized local model parameters applicable to each vehicle end, and after updating the local model parameters, the calibrated parameters are transmitted to the demand decision model in the local decision module by the local model, and finally the iterative optimization of the demand decision model parameters is realized; the federal learning module cooperates with the above local model and cloud global model to perform local model training, cloud global model parameter aggregation and medical knowledge graph assisted calibration operation to obtain updated parameters of the local decision module. The updated parameters of the local decision module are transmitted to the demand decision model in the local decision module, which is used to optimize the health state analysis accuracy. The real-time environmental data includes seat pressure, boarding time, location label, outdoor weather and temperature; the demographic characteristics include age, height, weight and past medical history. The beneficial effects include: through the supplement of demographic characteristics and real-time environmental data, the health classification is more comprehensive, and the pathological adaptation loss caused by relying only on physiological signals is avoided; the vehicle-grade chip meets the vehicle-mounted temperature resistance and anti-interference requirements, ensuring the stable operation of the local decision module in complex vehicle-mounted environments and avoiding analysis interruption caused by environmental fluctuations of ordinary chips; the federal learning module updates the model parameters, so that the demand decision model can adapt to the real-time health changes of users and the common characteristics of multiple users, avoiding the defects that traditional fixed parameters cannot adapt to individual differences.

[0016] Further, the cloud global model of the federated learning module is deployed in the cloud, receives local model parameters uploaded by multiple users based on the FedProx framework, is generated by the local model based on user demographic characteristics, physiological signals and real-time environmental data collected by the perception module, and is generated by the cloud global model by aggregating algorithm to fuse the local model parameters uploaded by the multiple users to generate general parameters optimized by the cloud and covering common characteristics of multiple users, and is issued after calibration by the medical knowledge graph. Further, the general parameters are issued to each vehicle end after calibration, and are first used to update the local model of the corresponding vehicle end, and then the updated local model injects optimization parameters into the demand decision model in the local decision module to complete the parameter iteration of the demand decision model. An iterative closed loop is formed from cloud aggregation and calibration to local model update, and then to demand decision model optimization, and the iteration update period is adapted to the user usage frequency. The updated demand decision model is used to improve the accuracy of health status analysis. Further, the local model parameters uploaded by the multiple users are encrypted parameters. The beneficial effects include: avoiding direct transmission of user sensitive data, meeting privacy compliance requirements such as GDPR; the general parameters generated by aggregating multiple user encrypted parameters cover common characteristics of multiple users, are calibrated in combination with the medical knowledge graph, ensure that the parameters meet medical standards, avoid model polarization caused by single user data, and enable each vehicle end local model to obtain safe and adaptive optimization parameters.

[0017] Further, the demand decision model in the local decision module is a health analysis model trained based on a deep learning architecture, and the training data covers multiple user health samples.

[0018] Further, the seat multi-modal vibration execution module includes a targeted vibrator sub-module, which independently controls the on-off of the partition vibrator including the headrest and / or the waist and / or the hips and / or the Achilles tendon, and adjusts the frequency and amplitude parameters. For different pathological users, according to the past medical history in the demographic characteristics collected by the perception module, the vibration contraindicated area adapted to the pathological state is turned off, and vibration is output to the non-vibration contraindicated area. The frequency and amplitude parameters of the vibration are determined based on the seat vibration execution scheme output by the local decision module, which corresponds to the user health status classification result. The corresponding principles include: the local decision module matches the preset vibration parameter interval and the vibrator opening range according to the risk degree of the health status classification result. The higher the risk degree, the lower the vibration frequency and amplitude, and the smaller the vibrator opening range. The beneficial effects include: for pathological users, the contraindicated area is turned off to avoid aggravating the condition, and the adaptive vibrator is enabled to achieve a balance between safe vibration and soothing effect; the vibration frequency / amplitude is determined based on the health classification execution scheme to avoid the problem of mismatch between traditional fixed parameters and real-time health status of the user, and to improve the accuracy and safety of vibration control.

[0019] Further, a data preprocessing module is further included, which is configured to clean sensitive information from physiological signals collected by the perception module, to obtain user feature label information, and transmit the user feature label information to the local decision module as input data for analyzing the user's health status; the sensitive information is information that may be associated with the user's identity, reveal individual health privacy, or affect data security. The beneficial effects include: generating structured user feature labels (such as “30 years old / no medical history / normal HRV”) by cleaning sensitive information reduces the interference of invalid data on the local decision module, and improves the efficiency and accuracy of health analysis; cleaning sensitive information in advance reduces the risk of privacy leakage in the data processing process from the source, and provides a pre- guarantee for subsequent privacy encryption.

[0020] Further, the data preprocessing module includes a denoising submodule configured to filter physiological signals collected by the perception module using Kalman filtering; further configured to filter seat pressure signals in real-time environmental data collected by the perception module using wavelet transform, to filter out noise caused by road bumps, electromagnetic interference, and seat fabric friction, and improve the signal-to-noise ratio of the data. Kalman filtering is suitable for the low-frequency noise characteristics of physiological signals (such as heart rate data fluctuations caused by road bumps), and wavelet transform is suitable for the 5-50Hz frequency band requirement of seat pressure signals (such as pressure noise caused by fabric friction). After filtering, the input data for subsequent feature extraction and health analysis is accurate, and the health state misjudgment caused by noise (such as misjudging heart rate fluctuations caused by bumps as an early warning) is avoided.

[0021] Further, the data preprocessing module includes a feature extraction submodule configured to extract time domain features, frequency domain features, and EMG features from the denoised data; the denoised data is the data processed by the denoising submodule from the physiological signals collected by the perception module; the time domain features include SDNN and RMSSD, the frequency domain features include LF / HF, and the EMG features include integrated electromyography and average power frequency; the extracted features are transmitted to the local decision module as input data for analyzing the user's health status. The beneficial effects include: by extracting multi-dimensional features, fine-grained judgment of the health status is achieved, such as time domain feature SDNN reflecting the overall situation of heart rate variability, which can be used to evaluate autonomic nervous function; EMG feature integrated electromyography reflecting muscle strain intensity, which can be used to judge lumbar muscle fatigue caused by long sitting; frequency domain feature LF / HF reflecting sympathetic / parasympathetic nerve balance, which can be used to evaluate anxiety state, replacing traditional coarse analysis relying only on heart rate / blood pressure, making the health classification more in line with the user's real health status.

[0022] Further, the data preprocessing module comprises a privacy encryption submodule, which performs anonymization processing and transmission encryption operation on past medical history in user demographic characteristics, HRV and heart rate data in physiological signals collected by the perception module, so as to meet the privacy compliance requirements, avoid leakage of user sensitive information, and use the encrypted data for subsequent model training of the federated learning module.

[0023] A seat safety vibration control method based on health state dynamic reasoning is implemented to achieve the second purpose of the application, comprising:

[0024] Collecting physiological signals of a user;

[0025] Analyzing the collected physiological signals to obtain a user health state classification result and a corresponding seat vibration and environment control execution scheme;

[0026] Adjusting the angles and heights of the seats according to the seat vibration and environment control execution scheme, and independently controlling the opening and closing and amplitude of the vibrators in each partition of each seat.

[0027] A non-transitory computer readable storage medium is implemented to achieve the third purpose of the application, which stores a computer program, and the computer program is executed by a processor to implement the steps of the seat safety vibration control method based on health state dynamic reasoning.

[0028] A computer program product is implemented to achieve the fourth purpose of the application, which comprises a computer program / instruction, and the computer program / instruction is executed by a processor to implement the steps of the seat safety vibration control method based on health state dynamic reasoning.

[0029] The beneficial effects of the application include:

[0030] 1. The application combines demographic characteristics (medical history and medical knowledge graph calibration) to accurately identify vibration contraindications and avoid secondary damage to pathological users caused by vibration; the local decision module analyzes physiological signals in real time, quickly triggers the closing of the vibrator + air conditioning ventilation in emergency state, adjusts the vibration parameters in the warning state, and forms a safe closed loop from real-time monitoring to classification intervention.

[0031] 2. Multi-source data and multi-dimensional feature extraction make health state judgment more comprehensive, such as adjusting vibration and fragrance in combination with muscle strain and anxiety index;

[0032] 3. The federated learning module implements the whole process of local training, cloud aggregation and parameter distribution, so that the model parameters continuously adapt to individual changes of users and commonalities of multiple users, avoid the extensive control of traditional fixed parameters, and realize vibration adaptation for thousands of people.

[0033] 4. Full-link privacy protection ensures that sensitive data such as user medical history and HRV are available and unrecognizable, in line with international privacy standards. Federated learning not only optimizes the model with multi-user data, but also avoids the risk of data leakage caused by centralized storage, balancing privacy protection and technical iteration.

[0034] 5. Car-grade chips support local decision-making, solving the delay problem of cloud decision-making, ensuring that vibration control and health status respond synchronously; deep learning models and multi-user samples improve the accuracy of health classification, denoising processing improves data precision, reduces misjudgment and invalid control, and combined with the cooperative control of seat angle / vibrator / environment (air conditioner / fragrance), the cabin health experience can be significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a schematic diagram of the system described in the present application;

[0036] Figure 2 is a schematic diagram of the business process;

[0037] Figure 3 is a front view of the vibrator arrangement structure of the seat;

[0038] Figure 4 is a side view of the vibrator arrangement structure of the seat;

[0039] Among them, 1: air sound exciter, 2: somatosensory exciter, 3: mixed exciter. DETAILED DESCRIPTION

[0040] The following detailed description is used to explain the technical solutions of the present application, so that those skilled in the art can understand the present application. The protection scope of the present application is not limited to the following specific implementation structure. The technical solutions of the present application are also included in the protection scope of the present application, which are different from the following specific embodiments.

[0041] A seat safety vibration control system based on dynamic reasoning of health status, comprising a perception module, a local decision-making module, a federated learning module and a seat multi-modal vibration execution module.

[0042] The perception module is arranged at the vehicle end, is used for collecting demographic characteristics including age, height, weight, past medical history, etc. through the occupant monitoring system (DMS / OMS) and wearable devices and mobile phone health apps; is also used for collecting physiological signals including HRV, heart rate, blood oxygen, blood pressure, respiratory rate, voiceprint tremor (anxiety index and cough sound recognition), etc. through the capacitive film steering wheel MIT flexible electronic sensor, the DMS system millimeter wave radar micro-motion detection algorithm and the microphone array; and is also used for collecting real-time environmental data including seat pressure, boarding time, location label, outdoor weather and temperature conditions through the seat sensor and other system information.

[0043] In an embodiment, the blood pressure detection accuracy of the capacitive film steering wheel MIT flexible electronic sensor is ±3 mmHg, the HRV detection error is ≤5 ms, and the response time is ≤30 ms; in an embodiment, the sampling rate of the millimeter wave radar is 20 Hz, the respiratory rate is calculated through the peak value interval of the chest and abdominal micro-motion signals (the peak value corresponds to one breath), and the frequency detection error after filtering is ≤1 times / min; in an embodiment, the detection accuracy of the seat pressure sensor is ±5 kPa, the sampling rate is 10 Hz, and the lumbar pressure center point can be accurately positioned (error ≤1 cm).

[0044] In an embodiment, the system further includes a data preprocessing module, which is used for cleaning sensitive information of the user data collected by the perception module to obtain feature label information of the user; specifically including:

[0045] The de-noising module is configured to filter the collected physiological signals by using Kalman filtering, and configured to filter the collected pressure signals by using wavelet transform method to filter out the noise caused by road bumps and electromagnetic interference, and improve the signal-to-noise ratio of the data. In an embodiment, the process noise covariance Q of Kalman filtering is in the range of 1e-4 to 1e-3, which is suitable for the low-frequency noise characteristics of the physiological signals. In another embodiment, the measurement noise covariance R is in the range of 1e-2 to 1e-1, which is suitable for the high-frequency noise characteristics of the road bumps. In another embodiment, the wavelet transform selects db4 wavelet basis, which is suitable for the frequency band requirement of 5-50 Hz of the seat pressure signal, and the signal-to-noise ratio of the filtered pressure signal is improved by more than 15 dB. The wavelet basis is the core basic function used for signal decomposition and reconstruction in wavelet transform, and the structural characteristics such as support length, orthogonality and time-frequency localization characteristics determine the accuracy of signal filtering and noise filtering effect. Different wavelet bases adapt to the processing requirements of different types of signals. The db4 wavelet basis is a 4th order Daubechies wavelet basis, which is a typical type of Daubechies wavelet basis. The time energy of the wavelet basis is concentrated, which can reduce the edge distortion of the seat pressure signal. The orthogonality of the wavelet basis can make the signal decomposition without redundant information, and avoid data redundancy after filtering. The good time-frequency localization characteristics of the wavelet basis can accurately capture the low-frequency interference such as road bumps and high-frequency noise such as seat fabric friction in the seat pressure signal, which is suitable for filtering multiple frequency band noise and retaining the real data of the seat pressure. It avoids the problem of incomplete noise filtering caused by the too strong symmetry of other sym series wavelet bases.

[0046] The feature extraction submodule is configured to extract the time domain features, frequency domain features and EMG features of the de-noised collected data. The time domain features include SDNN and RMSSD. The frequency domain features include LF / HF. The EMG features include integrated electromyography and average power frequency.

[0047] The pre-processed physiological signals (including HRV signals, EMG signals) are subjected to index analysis and calculation operations to obtain time-domain indicators of HRV (SDNN, RMSSD), frequency-domain indicators of HRV (LF (Low Frequency), HF (High Frequency)), and EMG features (integrated electromyography, mean power frequency), which provide core feature data for subsequent health status judgment and decision-making. The HRV (Heart Rate Variability) signal is used to indicate heart rate variability, that is, the difference between consecutive heartbeats. EMG (Electromyography) is a muscle electromyogram that reflects the muscle function state by detecting muscle electrical activity. SDNN (Standard Deviation of NN Intervals) is the standard deviation of normal heartbeat intervals, which is one of the time-domain indicators of HRV, reflecting the overall variability of the heartbeat interval in a period of time. RMSSD (Root Mean Square of Successive Differences) is the root mean square of consecutive heartbeat intervals, which is one of the time-domain indicators of HRV, reflecting the variability of adjacent heartbeat intervals. The low-frequency component LF in the frequency-domain indicator of HRV has a frequency range of 0.04-0.15 Hz, mainly reflecting the combined activity of the sympathetic and parasympathetic nerves. The high-frequency component HF has a frequency range of 0.15-0.4 Hz, mainly reflecting the activity of the parasympathetic nerve. Integrated electromyography and mean power frequency are both EMG features. Integrated electromyography refers to the integral value of electromyography signals within a certain time, reflecting the total intensity of muscle activity. Mean power frequency refers to the frequency corresponding to the average power of the power spectrum of electromyography signals, reflecting the fatigue degree and other states of muscle contraction.

[0048] In an embodiment, the data preprocessing module further includes a privacy encryption submodule for anonymizing and transmission encryption of sensitive data according to the privacy protection requirements of user sensitive health data (medical history, heart rate variability, etc.), to obtain encrypted data that meets privacy compliance requirements and avoid leakage of user sensitive information. In an embodiment, differential privacy method is used to anonymize user sensitive health data (such as medical history, heart rate variability, etc.). By adding specific noise to the data, the data can be used for statistical analysis while being unable to be accurately associated with a specific individual, thereby realizing data usability and unrecognizability and preventing individual information leakage. In an embodiment, the noise strength of differential privacy is set to ε = 0.1, and test verification shows that the data statistical error is ≤8% under this value, while meeting the compliance requirements of GDPR for user data anonymization. In an embodiment, homomorphic encryption (BFV algorithm) is used to ensure the security of sensitive data transmission and cloud processing. It allows direct addition, multiplication and other operations on encrypted data streams, without the need for decryption to complete data processing, thereby avoiding the risk of leakage due to decryption during data transmission or cloud processing. In an embodiment, the key length of homomorphic encryption is set to 2048 bits, and the encryption / decryption time on the vehicle-grade NVIDIA Orin chip is ≤50 ms, which does not affect the latency requirement of ≤500 ms for local decision-making. In an embodiment, all encryption operations of the privacy encryption submodule need to meet the requirements of GDPR (General Data Protection Regulation) for user data privacy protection, to ensure that the system complies with international privacy standards when processing user health data and avoid legal risks.

[0049] The local decision-making module is used to obtain user health status classification results and corresponding seat vibration and environmental control preliminary execution scheme based on the collected user physiological characteristic data or preprocessed user physiological characteristic data (heart rate, blood pressure, muscle electrical characteristic, etc.). In an embodiment, the demand decision-making model can be loaded on a vehicle-grade chip on a vehicle to perform preliminary rapid analysis, and the demand decision-making model can also be loaded on a massage seat processing chip to perform preliminary rapid analysis, mainly including seat posture, cabin air and odor health. The vehicle-grade chip (Automotive-Grade Chip) is a chip that meets the requirements of vehicle environment (temperature resistance, anti-interference, etc.), such as NVIDIA Orin chip, which can provide hardware support for local rapid calculation.

[0050] In an embodiment, the demand decision model is a health analysis model trained based on a ResNet50 architecture, and the training data includes 50,000 user health samples. The health analysis model can output a health status (normal, warning, emergency) according to input physiological characteristic data. The health status classification includes: normal state, warning state, and emergency state. In the normal state, the heart rate is 60-100 beats per minute, and the blood pressure is 90 / 60-140 / 90 mmHg. In the warning state, the heart rate is >100 beats per minute or <60 beats per minute. In the emergency state, the heart rate is >120 beats per minute or <50 beats per minute.

[0051] In an embodiment, the local decision module further includes a health demand analysis submodule. Based on the health status classification result output by the local decision module, the health demand analysis submodule further converts the health demand into a specific demand, such as adjusting the seat posture to a comfortable sitting posture, and passing the demand to the execution layer. The data input shares the preprocessed physiological characteristic data with the local decision module, and the output result is directly transmitted to the seat multi-modal vibration execution module, which cooperates with the output of the local decision module to guide the execution operation.

[0052] The federated learning module is used to train the local model based on the FedProx framework, aggregate the parameters of the global model in the cloud, and perform medical knowledge graph assisted calibration operations (such as CDC disease database calibration of abnormal threshold) to obtain an iteratively updated demand decision model. The demand decision model is used to predict and adapt to the health demand of the user in advance, and the update period is shortened to ≤3 driving times. While protecting the privacy of the user data (data does not move the model moves), the generalization and adaptation ability of the model to different user health states and multiple environments is improved.

[0053] In an embodiment, the federated learning module further includes a deep decision submodule. The deep decision submodule is used to fuse the parameters aggregated by the global model in the cloud and the rules of the medical knowledge graph, and convert the parameters into specific vibration control rules, such as limiting the parameters of the lumbar vibration sub-module of the user with L5-S1 disc herniation. The deep decision submodule does not independently participate in data collection or model training, and only serves as a parameter conversion bridge between the global model in the cloud and the local model.

[0054] In an embodiment, the federated learning module further comprises a medical knowledge graph submodule, configured to construct a "pathology-vibration parameter" mapping rule library covering spine / skeletal diseases (such as L5-S1 disc herniation, ankylosing spondylitis, osteoporosis) according to the accessed CDC disease database and clinical medical data, and to calibrate the health abnormal threshold output by the deep decision-making module to obtain vibration contraindications (such as lumbar-sacral region > 10 Hz vibration contraindication, sacral vibration contraindication) and parameter suggestions conforming to medical standards, thereby providing authoritative medical basis for the local decision-making module and the federated learning module and ensuring the safety and pathology adaptation of seat vibration control. Only the parameter calibration link of the cloud global model is involved, and the local model parameters uploaded by multiple users are corrected to obtain general parameters, and the corrected general parameters are then distributed to the local model, without directly participating in the training of the local model.

[0055] In an embodiment, the local model is deployed in a vehicle-grade chip at the vehicle end, and is a specific implementation of the federated learning framework at the user side. Based on the full data of the user (including historical health records, long-term feedback, and pathological changes), the local model is periodically trained (e.g., updated once every 3 times of driving) by the FedProx framework to continuously optimize the pathological recognition threshold, such as the vibration contraindication parameter of L5-S1 protrusion, the user adaptability coefficient such as the tolerance threshold of the elderly to vibration intensity, and other underlying parameters, and finally outputs an optimized model parameter set. It can be understood analogously that the algorithm parameter library of a mobile phone camera is used to continuously optimize the exposure and focusing parameters according to the user's shooting habits, and the mobile phone camera itself does not directly take photos, but only provides better parameter support for the shooting function.

[0056] The optimized parameter set generated by the local model through federated learning iteration is periodically injected into the demand decision-making model, so that the decision basis is upgraded from the initial general parameters to precise parameters adapted to the individual user, for example, the heart rate warning threshold of an ordinary user is 100 times per minute, and after optimization by the local model, the threshold of a user with heart disease may be adjusted to 90 times per minute. The demand decision-making model only calls parameters and does not participate in parameter training; the local model only optimizes parameters and does not participate in real-time decision-making.

[0057] In an embodiment, the local model completes the model training process based on the FedProx framework. Each user's vehicle end has an independent local model that is trained based on the user's local data (physiological characteristics, feedback scores, etc.) to adapt to individual health characteristics, such as the heart rate fluctuation pattern of a certain user or pathological contraindications. Each user's vehicle end device uses the locally collected health data (such as physiological characteristics and feedback scores) to independently train the local demand decision model. After 10 iterations, the model loss rate is reduced to less than 5%, and a single iteration takes ≤10s on an NVIDIA Orin chip, without occupying too much vehicle-mounted computing power; the batchsize is set to 32, which is suitable for the amount of physiological characteristic data collected by the vehicle end at a time, and the data amount is 32-dimensional features per time, ensuring a balance between training efficiency and data utilization. During training, the loss function correction term of the FedProx framework ensures that the local model parameters not only adapt to individual characteristics such as the heart rate threshold difference between older and younger people, but also do not deviate from the common rules of the cloud global model, i.e., avoid parameter extremeization. After training, only the encrypted model parameters are uploaded to the cloud for aggregation.

[0058] The loss function correction term is a constraint term used to alleviate the problem of unstable model training caused by too large differences in data distribution among different users. During local model training, the model is required to not only fit its own data, but also to not deviate too much from the current cloud global model parameters. This allows each user's local model to adapt to its own data while considering the global common rules, avoiding the training of extreme parameters due to special own data. Specifically, it can be understood as follows: the health data of different users has significant differences. This large data distribution difference can lead to the following problem: if the local model is only trained based on its own data, the parameters of the local models of different users will deviate too far from each other, for example, the model of user A may consider that a heart rate >100 beats per minute needs to be warned, and the model of user B may consider that a heart rate >110 beats per minute needs to be warned, which ultimately leads to the inability to form a stable global model when aggregating in the cloud.

[0059] In an embodiment, the calculation formula of the loss function correction term includes:

[0060] (λ×(local model parameters-current cloud global model parameters) 2 )

[0061] λ is a weight coefficient used to control the constraint strength; in an embodiment, λ is set to 0.01. Tests have verified that this value can control the deviation of the parameters of the local models of different users within a set proportion of 15%, ensuring the stability of the global model after aggregation in the cloud.

[0062] Through this modification, the parameters of each local model can maintain a certain compatibility even if the data of different users differs greatly, ensuring that the cloud global model generated after aggregation can both adapt to the common characteristics of most users and retain a reasonable space for individual differences, solving the problem of poor model generalization ability caused by data heterogeneity.

[0063] In an embodiment, the cloud global model is deployed in the cloud and completes parameter aggregation and optimization based on the same FedProx framework. After receiving the encrypted local model parameters uploaded by multiple users, the cloud global model fuses the parameters through the aggregation algorithm (such as weighted average) of the FedProx framework to generate general parameters covering the common characteristics of multiple users; at the same time, the general parameters are calibrated in combination with a medical knowledge graph such as the CDC disease database pathology-vibration rule to ensure that the model output meets the medical standards. The calibrated general parameters are then distributed to each user's vehicle end to update the local model, forming an iterative closed loop of "local training→ parameter uploading→ cloud aggregation and calibration→ parameter distribution and updating".

[0064] The seat multi-modal vibration execution module is used to adjust the angles and heights of the seat in all directions to adaptively match the targeted needs of healthy users and sub-healthy users using vibration functions in static sitting / lying, dynamic riding, and dynamic / static states. By individually controlling the opening and closing of each vibrator and its amplitude, a closed loop is realized from user physiological condition sensing to focal point positioning, to vibrator vibration safety control, to real-time physiological signal feedback adjustment of the user.

[0065] In an embodiment, the seat multi-modal vibration execution module includes a seat adjustment motor, a targeted vibrator sub-module, and an environment control sub-module.

[0066] The seat adjustment motor is used to drive and adjust the seat backrest angle (adjustment range 0-120°) and the seat cushion height (adjustment range 400-550 mm) with millimeter-level precision (adjustment accuracy ±0.1 mm) according to user demographic characteristics (height, weight), seat pressure distribution data (such as lumbar pressure value, ischial tuberosity pressure value), and health needs (such as lumbar disc herniation, lumbar spondylolisthesis user sitting posture adaptation needs), with a response time ≤100 ms, thereby realizing precise adaptation of the seat static sitting / lying, dynamic riding posture, and user physiological characteristics and pathological state, relieving local pressure concentration (such as adjusting the backrest angle to reduce pressure when the lumbar pressure is >80 kPa), and ensuring the health of the sitting posture and the comfort of the ride.

[0067] The target vibrator sub-module is used to independently control and adjust the frequency and amplitude parameters of the headrest, waist, back, hip and other partition vibrators according to the user's pathological information (such as L5-S1 disc herniation, ankylosing spondylitis, osteoporosis, etc.), real-time physiological signals (such as muscle strain state) and seat pressure distribution data, to close the vibration contraindication area adapted to the user's pathological state, and output precise vibration that meets the user's health needs (such as muscle relaxation and avoidance of bone conduction damage) to the non-vibration contraindication area, so as to ensure the safety and adaptation to individual pathological differences of the seat vibration function, while achieving targeted vibration relaxation or auxiliary treatment effect. Among them, the Achilles tendon vibrator enabled for osteoporosis users is installed at the rear end of the seat cushion ankle support area, adopts a fit design, the vibrator size is 15mmx8mm, and the vibration transmission medium is medical silicone with a hardness of 50 Shore A, which ensures that the vibration energy is effectively transmitted to the Achilles tendon without discomfort; the vibration parameter is set to a frequency of 3Hz and an amplitude of 0.3mm, which is based on the bone conduction vibration safety threshold of osteoporosis patients in the "Orthopedic Rehabilitation Medicine Guide", to avoid secondary damage to the skeleton caused by vibration.

[0068] The environmental control sub-module is used to cooperatively control the cabin air conditioning system (temperature regulation accuracy ±0.5℃, fresh air volume regulation range 50-100m 3 / h) and the fragrance system (supporting 2 kinds of essential oils of lavender and mint, concentration regulation range 0.05-0.2mg / m 3 ) according to the user's physiological signals (anxiety index, breathing rate), external environment data (outdoor temperature, weather conditions) and health needs (such as the relaxation needs of anxious users and the air improvement needs of abnormal breathing users), to realize the adaptation of cabin temperature, air quality (such as CO2 concentration <1000ppm), fragrance type and concentration to the user's health state, to assist in relieving user anxiety (such as enabling mint fragrance when the anxiety index is ≥7 points) and improving the breathing environment, and to cooperate with the seat vibration function to improve the cabin health experience.

[0069] The business process of the method embodiment based on the above system includes five stages of data acquisition, preprocessing, decision making, execution and feedback, and the timing and operation details of each stage are as follows:

[0070] Step 1, multi-source data acquisition based on the perception module, specifically including:

[0071] Demographic collection: through the linkage of vehicle-mounted system and mobile phone APP, automatically synchronize user's age, height, weight, medical history, such as 30 years old, 175 cm tall, 70 kg, no past medical history. Manually supplement when first used, and automatically update subsequently;

[0072] Physiological signal collection: use DMS camera to collect heart rate and blood oxygen through facial photoplethysmography (PPG); such as heart rate 70 beats per minute, blood oxygen 98%; use flexible steering wheel sensor to collect blood pressure and HRV through fingertip pulse wave; the blood pressure detection accuracy of flexible electronic sensor is ±3 mmHg. Use millimeter wave radar to calculate respiratory rate by detecting chest and abdominal micro-movement; use microphone array to collect voiceprint, extract anxiety index (0-10 points, ≥7 points for high anxiety) and identify cough sound through Mel-frequency cepstral coefficient (MFCC);

[0073] Real-time data collection: collect pressure distribution through seat pressure sensor at a sampling rate of 10 Hz, and locate the lumbar pressure center point; obtain the location (city / highway), outdoor temperature and weather (sunny / rain / snow) through vehicle-mounted GPS and weather interface; record the boarding time (accurate to the minute) through vehicle-mounted clock, and determine whether it is a rush hour (such as 7:00-9:00, 17:00-19:00). Collect data such as: seat pressure distribution is uniform, boarding time is 8:00 am, location is urban road, outdoor temperature is 25℃.

[0074] Step 2, data preprocessing

[0075] 1. Noise reduction processing: for physiological signals, use Kalman filter to filter out HRV noise caused by road bumps to reduce the SDNN error after filtering; for pressure signals, use wavelet transform (such as db4 wavelet basis) to filter out noise caused by seat fabric friction to improve pressure detection accuracy;

[0076] 2. Feature extraction: for time domain features, extract SDNN (24-hour heart rate variability) and RMSSD (standard deviation of adjacent heart rate intervals) of HRV; IEMG (integral electromyogram, reflecting muscle activity intensity) of EMG; for frequency domain features, extract LF (low frequency component, 0.04-0.15 Hz, reflecting sympathetic nerve activity) and HF (high frequency component, 0.15-0.4 Hz, reflecting parasympathetic nerve activity) of HRV;

[0077] 3. Privacy encryption: add differential privacy noise (ε=0.1, meet GDPR privacy requirements) to sensitive data such as user's past medical history and HRV; use BFV homomorphic encryption for data transmitted to the cloud, with key length of 2048 bits.

[0078] Step 3, local decision

[0079] The demand decision model inputs the pre-processed features and outputs the preliminary health status:

[0080] Normal state: heart rate 60-100 beats / min, blood pressure 90 / 60-140 / 90 mmHg, anxiety index ≤5 points;

[0081] Warning state: heart rate > 100 beats / min or < 60 beats / min, blood pressure > 140 / 90 mmHg or < 90 / 60 mmHg, anxiety index 6-8 points;

[0082] Emergency state: heart rate > 120 beats / min or < 50 beats / min, blood pressure > 160 / 100 mmHg or < 80 / 50 mmHg, anxiety index ≥ 9 points;

[0083] Match the preliminary implementation plan: if normal state → turn on the waist back vibrator (frequency 5 Hz, amplitude 1 mm), warning state → reduce the amplitude to 0.5 mm, emergency state → turn off all vibrators and trigger the air conditioning ventilation. After 1000 user tests, the recognition accuracy of the local decision model for normal / warning / emergency state is 98%, 95%, and 92% respectively, meeting the health risk judgment needs.

[0084] Step 4, cloud deep decision

[0085] Cloud deep decision and local decision are parallel, and the decision result is used for iterative model:

[0086] Local encrypted features are uploaded to the cloud, and the cloud global model (FedProx framework) aggregates multi-user data and calibrates parameters combined with medical knowledge graph;

[0087] For example, if the user's medical history is L5-S1 disc herniation, the knowledge graph calls the rule: lumbosacral region (frequency of the two vibrators below the waist back ≤8 Hz, amplitude ≤0.5 mm;

[0088] For example, if the user's medical history is osteoporosis, the rule is: disable the sacrum (middle vibrator in the buttocks), ischial tuberosity (vibrators on both sides of the buttocks), and enable the Achilles tendon vibrator (new, frequency 3 Hz, amplitude 0.3 mm);

[0089] The cloud will issue the updated "pathology-vibration" rule to the vehicle end and iterate the local model (update cycle ≤ 3 times of driving).

[0090] The health demand analysis submodule performs preliminary rapid analysis based on the above collected data and obtains the following seat health needs of the corresponding user:

[0091] Seat health needs: seat posture adjustment to comfortable sitting position, fresh air in the cabin, no odor.

[0092] Based on the above cabin health needs, the seat is automatically adjusted to the preset comfortable sitting posture, and the air conditioning system starts the air purification mode.

[0093] The depth decision sub-module adopts a local learning and cloud aggregation mode. Through local learning, it is found that the user's recent heart rate fluctuation is large, which may be in the cold period. Through cloud aggregation, the global model and medical knowledge graph in the cloud are combined to update the abnormal threshold.

[0094] Model update: the demand decision model iterates according to new data.

[0095] Step 5, atomization execution

[0096] The seat multi-modal vibration execution module performs the following operations:

[0097] Seat adjustment: according to the user's health needs, the seat angle is automatically adjusted, and the height is adapted. For example, according to the pressure distribution, the backrest angle is adjusted from 90° to 100° when the lumbar pressure is >80kPa; when the user's height is 175cm, the seat cushion height is adapted to 450mm; it is found through actual measurement that this adjustment can reduce the lumbar pressure by 20%-30%, and the user's sitting posture comfort score is increased by ≥40%, which is based on a 5-point scale.

[0098] Vibration control: independent on-off control and frequency, amplitude parameter adjustment are performed on the partition vibrator containing the headrest and / or waist and / or hip and / or Achilles tendon; for different pathological users, according to the past medical history in the demographic characteristics collected by the perception module, the vibration contraindication area suitable for the pathological state is closed, and vibration is output to the non-vibration contraindication area. For example: normal user: turn on 4 waist and back vibrators, frequency 5-8Hz, amplitude 0.8-1mm, 3 hip vibrators, frequency 3-5Hz, amplitude 0.5-0.8mm; L5-S1 intervertebral disc herniation user: turn off 2 lower waist and back vibrators (lumbosacral region), turn on 2 upper waist and back vibrators, frequency 6Hz, amplitude 0.5mm; ankylosing spondylitis user: turn off all waist and back vibrators, only turn on headrest vibrator, frequency 2Hz, amplitude 0.3mm; for 50 cases of L5-S1 intervertebral disc herniation users, the VAS pain score of the user is reduced from 4-5 to 2-3 after using this vibration scheme, and the pain relief rate reaches 40%-50%.

[0099] Environmental collaborative control: when the anxiety index ≥7: turn on the mint fragrance, the concentration is 0.1mg / m 3 , the air conditioner is adjusted to 24℃, the wind speed is level 2; when the respiratory rate >20 times / min: turn on the air conditioner external circulation, the fresh air volume is 100m 3 / h, reduce the CO2 concentration to <1000ppm. It is found through testing that the HRV high frequency component (HF) of the user with anxiety index ≥7 is increased by ≥20% after using the mint fragrance, and the anxiety state relief effect is significant.

[0100] Step 6, Real-time Feedback: Continuous monitoring of user physiological signals, and dynamic adjustment of vibration function based on feedback.

[0101] Real-time feedback is started after a set time (e.g. 600ms) after the first control operation of the seat multi-modal vibration execution module. Thereafter, it is continuously performed during the user's ride, and the core ensures that the seat vibration and cabin environment control always adapt to the user's real-time health status through the closed-loop process of effect evaluation, parameter adjustment, and model iteration. The specific process is as follows:

[0102] 1. Effect Evaluation

[0103] It includes subjective evaluation and objective evaluation to verify the adaptability of the control scheme to user health needs: subjective evaluation is achieved by popping up a visual analog scale pain score on the vehicle screen, with a score range of 0 to 10 points. The user completes the selection through touch operation, and the evaluation frequency is once every 5 minutes, so as to quantify the user's subjective feeling of seat vibration and environmental control. Objective evaluation is achieved by continuously monitoring electromyography and heart rate variability indicators. In the electromyography indicator, the integral electromyography value decreases by no less than a set percentage of 15%, representing that the muscle strain state is relieved. In the heart rate variability indicator, the high-frequency component increases by no less than a set percentage of 20%, representing that the user's anxiety state is relieved. Both indicators are objective basis for the effectiveness of the control scheme. After continuous monitoring of 200 users, the consistency of subjective evaluation and objective evaluation results is more than 85%, which can effectively verify the effect of the control scheme.

[0104] 2. Parameter Adjustment

[0105] Based on the results of effect evaluation, dynamic optimization is performed for seat vibration parameters and cabin environment parameters, including:

[0106] If the visual analog scale pain score exceeds 5 points, or the integral electromyography value does not decrease, the system will reduce the seat vibrator frequency by 1 Hz each time, and reduce the vibrator amplitude by a set amplitude value of 0.2 mm each time.

[0107] If the high-frequency component of the heart rate variability indicator does not increase, the system will increase the concentration of mint fragrance by 0.05 mg / m3 each time, and increase the air conditioner wind speed level by 1 level each time, to assist in relieving the user's anxiety state through coordinated optimization of environmental parameters. After testing, the user's VAS score is reduced by an average of 1-2 points after parameter adjustment, and the HF index is increased by 15%-25%, with significant adjustment effect.

[0108] 3. Model Iteration

[0109] Model iteration incorporates effect evaluation data and parameter adjustment records into the training system of the federal learning framework:

[0110] The system adds the subjective score of this feedback, the change of the objective index, and the adjusted parameter to the training data set of the local model at the vehicle end, supplements the adaptation relationship between the real-time health status of the user and the control scheme; when the next global model at the cloud end starts aggregation update, the local model at the vehicle end will generate new parameters based on the supplemented training data set, upload to the cloud end to participate in general parameter fusion, and finally through the calibration of the global model at the cloud end, the new general parameters will be issued to the vehicle end, completing the parameter update of the local model and the demand decision model, and ensuring that the subsequent decision and control scheme continue to adapt to the change of the health status of the user. After 500 user model iteration tests, the pathological recognition accuracy of the iteration model is improved from 90% to 95%, and the adaptability is significantly improved.

[0111] In addition to the above user examples, through the above system or method, when the user is identified as having "L5-S1 intervertebral disc herniation", the lumbar sacral region > 10 Hz vibration is turned off; the vibration function of the locked spine region of the ankylosing spondylitis patient is locked; the vibration region of the sacrum and the ischial tuberosity of the osteoporosis user is disabled, and is replaced by the Achilles tendon micro-vibration to avoid bone conduction; the L1-L5 vertebral body corresponding region of the lumbar spondylolisthesis user is disabled, and is replaced by the biceps femoris low-frequency vibration to promote compensation, etc. Intelligent adjustment. For the above 4 types of pathological users, each 30 cases were tested, and the vibration contraindication region recognition accuracy of the system was ≥95%, and there was no false triggering of the contraindication region vibration, and the safety met the design requirements.

[0112] It should be understood that the size of the serial number of each step in the above embodiment does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0113] The embodiment of the present application also provides a seat safety vibration control system based on dynamic reasoning of health status, comprising:

[0114] The perception module is used for collecting physiological signals of the user, demographic characteristics including age, height, weight, and past medical history, physiological signals including HRV, heart rate, blood oxygen, blood pressure, respiratory rate, voiceprint tremor, and real-time environmental data including seat pressure, boarding time, location label, outdoor weather and temperature;

[0115] The local decision module is a demand decision model relying on a vehicle-level chip, which is used for analyzing the collected physiological signals to obtain user health status classification results and corresponding seat vibration and environmental control execution schemes;

[0116] The seat multi-modal vibration execution module is used for adjusting the angles and heights of the seats in each direction according to the seat vibration, environmental control execution scheme, and realizing the closed loop of user physiological condition sensing, focus point positioning, vibrator vibration safety control and real-time physiological signal feedback adjustment by independently controlling the opening and closing and amplitude of the vibrators in each partition on each seat.

[0117] As shown in Figure 3 and 4 , in the seat multi-modal vibration execution module, the vibrators in each partition on the seat include: a somatosensory exciter 2 arranged in the seat, and a mixed exciter 3 for receiving an electrical signal and capable of converting the electrical signal into a mechanical vibration and an audible sound signal. The vibration driving signal is transmitted to the somatosensory exciter 2 and / or the mixed exciter 3 for vibration. In an embodiment, the seat further includes an air sound exciter 1, an acoustic driving digital signal is routed to the air sound exciter 1 to sound, and the user's health condition and other information related to the user can be played through the air sound exciter 1; it is arranged on the side of the headrest and the cushion; it is close to the human ear, can reach a higher sound pressure level under a smaller power, more efficiently transmits the sound to the user's ear, and improves the clarity and sound quality of the sound. And arranged on the side of the cushion, it can cover the lower half of the human body with sound, so that the sound field in the car is more uniform, and the overall auditory experience is improved.

[0118] The arrangement of the air sound exciter 1, the somatosensory exciter 2 and the mixed exciter 3 in the seat satisfies the target function relationship between the exciter position and the human body perception intensity, and the target function expression of the exciter position and the human body perception intensity is:

[0119] In order to maximize the comprehensive perception of the human body to the sound and the somatosensory, β audio (U) and β haptic (U) are adaptive weights combined with the user feature vector U, respectively 0.6+δ u and 0.4-δ u , δ u ∈[-0.1,0.1];

[0120] In order to minimize the crosstalk between the exciters, the crosstalk physical model is calculated, wherein k is the coupling coefficient, f i is the exciter related parameter, d ij is the distance between the exciter i and the exciter j, γ is the penalty coefficient, γ∈[0.2,0.5], the mechanical crosstalk hard constraint is d ij ≥80+∈, ∈≥0, and the acoustic sensitive area constraint is wherein x iis the spatial position vector of the ith exciter ear is the ear position, k is the attenuation coefficient, and the somatosensory sensitive area constraint is S body,i = a T(x i ), a T(x i ) > T min , T(x i ) is the muscle thickness, T min is the muscle thickness threshold, a is the correlation coefficient, and the neural dense area avoidance constraint is ||x i -x nerve || > R safe , x nerve is the position vector of the neural dense area, and R safe is the safety distance of the exciter from the neural dense area.

[0121] The constraint processing rule is: if d ij < 80, the solution is invalid, and if there are multiple solutions that satisfy the constraint, the solutions are sorted in descending order of the objective function value, and the solution with the larger objective function value is selected first.

[0122] The minimum setting distance and the maximum setting number between the exciters can be obtained through the above formula, and the arrangement and verification of the exciters can be performed according to actual needs in the later stage.

[0123] In an embodiment, the mixed exciter 3 is arranged at a position at 50%-60% of the height of the backrest of the seat.

[0124] In an embodiment, the distance between the mixed exciter 3 and the pinna of the 50th percentile human body is not more than 150 mm, and the distance between the mixed exciter 3 and the scapula of the 50th percentile human body is not more than 40 mm.

[0125] In an embodiment, the installation depth of the mixed exciter 3 in the backrest of the seat is 20-30 mm.

[0126] In an embodiment, the somatosensory exciter 2 is arranged at a position below the backrest, the cushion, the thigh, the leg support, etc., and can perform low-frequency vibration feedback on the back of the human body; and the mixed exciter 3 is arranged at a position on the backrest, etc.

[0127] Table 1 below is a specific parameter and arrangement table of the air sound exciter, the somatosensory exciter, and the mixed exciter in an embodiment:

[0128] Table 1

[0129]

[0130] The seat multi-modal vibration execution module further comprises a feedback collection module for collecting actual mechanical vibration felt by the user, comparing the actual mechanical vibration felt by the user with the expected mechanical vibration, and judging whether the vibration is synchronized. The processing module of the seat can receive feedback information of the feedback collection module, and optimize parameters of the vibration signal to make the expected mechanical vibration consistent with the actual mechanical vibration.

[0131] In an embodiment, the perception module is further configured to collect user demographic characteristics and real-time environmental data, and the local decision module further comprises a demand decision model carried by a vehicle-grade chip and configured to obtain a user health state classification result. The system further comprises a federated learning module configured to use a local model and a cloud global model based on a FedProx framework, combine user demographic characteristics, physiological signals and real-time environmental data collected by the perception module stored at the vehicle end and cloud multi-user feature label information, perform local model training, cloud global model parameter aggregation and medical knowledge graph assisted calibration operations, and obtain updated parameters of the local decision module. The updated parameters are transmitted to the demand decision model in the local decision module to optimize the health state analysis accuracy.

[0132] In an embodiment, the local model parameters uploaded by the multiple users are encrypted parameters.

[0133] In an embodiment, the cloud global model of the federated learning module is deployed in the cloud, receives local model parameters uploaded by multiple users based on the FedProx framework, is trained and generated by the local model based on user demographic characteristics, physiological signals and real-time environmental data collected by the perception module, and generates general parameters by aggregating the local model parameters uploaded by the multiple users. The cloud global model is updated by combining the general parameters with the medical knowledge graph and then distributed to each vehicle end to update the model parameters, forming an iterative closed loop and adapting the iteration update period to the user usage frequency. The updated local model is used to optimize the analysis logic of the demand decision model.

[0134] In an embodiment, the demand decision model in the local decision module is a health analysis model trained based on a deep learning architecture, and the training data covers multiple user health samples. The multiple user health samples include user demographic characteristics, physiological signals and real-time environmental data. The health state classification result includes normal state, warning state and emergency state, and the heart rate, blood pressure and anxiety index judgment criteria of each state meet the medical health monitoring specifications. The anxiety index is extracted from the voiceprint tremor data in the physiological signals collected by the perception module. The health state classification result is used by the seat multi-modal vibration execution module to determine the vibration area and parameters. The medical health monitoring specifications are, for example, the industry-recognized heart rate and blood pressure normal range standards.

[0135] In an embodiment, the seat multi-modal vibration execution module comprises a targeted vibrator sub-module, which independently controls the on-off of the zoned vibrator including the headrest and / or the waist and / or the hip and / or the Achilles tendon and adjusts the frequency and amplitude parameters; for different pathological users, according to the past medical history in the demographic characteristics collected by the perception module, the vibration taboo area adapted to the pathological state is closed, and vibration is output to the non-vibration taboo area, and the frequency and amplitude parameters of the vibration are determined based on the seat vibration execution scheme output by the local decision module, which corresponds to the user health state classification result. In an embodiment, the corresponding principles include: the local decision module matches the preset vibration parameter interval and the vibrator opening range according to the risk level of the health state classification result, the higher the risk level, the lower the vibration frequency and amplitude, and the smaller the vibrator opening range. The health state classification result includes normal state, warning state and emergency state, and one of the embodiments of the corresponding execution scheme is: when the classification result is normal state (such as heart rate 60-100 times / min, blood pressure 90 / 60-140 / 90 mmHg), the execution scheme is to open the waist + Achilles tendon vibrator, the vibration frequency is 30-50 Hz, and the amplitude is 0.8-1.2 mm; the second embodiment of the corresponding execution scheme is: when the classification result is the warning state (heart rate 101-120 times / min or blood pressure 141 / 91-160 / 100 mmHg), the execution scheme is to only open the Achilles tendon vibrator, the vibration frequency is 20-30 Hz, and the amplitude is 0.4-0.8 mm; the third embodiment of the corresponding execution scheme is: when the classification result is the emergency state (heart rate > 120 times / min or blood pressure > 160 / 100 mmHg), the execution scheme is to close all vibrators, and further, the seat ventilation function can also be opened.

[0136] In an embodiment, a data preprocessing module is further included, which is used for cleaning sensitive information of the user physiological signals collected by the perception module to obtain user feature label information, and the user feature label information is transmitted to the local decision module as input data for analyzing the user health state; the sensitive information is information that may be associated with the user's identity, reveal individual health privacy or affect data security.

[0137] In an embodiment, the data preprocessing module comprises a denoising sub-module, which is used for filtering the physiological signals collected by the perception module by using Kalman filtering and filtering the seat pressure signals in the real-time environmental data collected by the perception module by using wavelet transform, so as to filter out the noise caused by road bumps, electromagnetic interference and seat fabric friction and improve the signal-to-noise ratio of the data.

[0138] In an embodiment, the data preprocessing module comprises a feature extraction submodule for extracting time domain features, frequency domain features and EMG features of the denoised data; the denoised data is the physiological signal collected by the perception module and processed by the denoising submodule; the time domain features include SDNN and RMSSD, the frequency domain features include LF / HF, and the EMG features include integral electromyography and average power frequency; the extracted features are transmitted to the local decision module as input data for analyzing the user's health status.

[0139] In an embodiment, the data preprocessing module comprises a privacy encryption submodule, which anonymizes and encrypts the past medical history in the user's demographic characteristics, the HRV and heart rate data in the physiological signal collected by the perception module, to meet the privacy compliance requirements and avoid leakage of sensitive user information; the encrypted data is used for subsequent model training of the federated learning module.

[0140] The embodiment of the present application also provides a seat safety vibration control method based on dynamic reasoning of health status, comprising:

[0141] Collecting physiological signals of a user;

[0142] Based on a vehicle demand decision model carried by a vehicle-grade chip, analyzing the collected physiological signals to obtain a user health status classification result and a corresponding seat vibration and environment control execution scheme;

[0143] Adjusting the angles and heights of the seat in each direction according to the seat vibration and environment control execution scheme, and independently controlling the opening and closing and amplitude of the vibrators in each partition of each seat.

[0144] The embodiment of the present application also provides a computer program product comprising computer programs / instructions, which, when executed by a processor, implement the steps of the seat safety vibration control method based on dynamic reasoning of health status.

[0145] The embodiment of the present application also provides a non-transitory computer readable storage medium storing a computer program, which comprises program instructions that, when executed by a processor, implement the steps of the method of the present application, which will not be described here.

[0146] The computer readable storage medium can be the internal storage unit of the data transmission device or the computer device, such as the hard disk or the memory of the computer device. The computer readable storage medium can also be the external storage device of the computer device, such as the plug-in hard disk, the smart media card (SMC), the secure digital (SD) card, the flash card and the like.

[0147] Further, the computer readable storage medium can include both the internal storage unit and the external storage device of the computer device. The computer readable storage medium is used to store the computer program and other programs and data required by the computer device. The computer readable storage medium can also be used to temporarily store the data to be output or the data that has been output.

[0148] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer-usable program code.

[0149] The present application is described with reference to the flowcharts and / or block diagrams of the method, device (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of the flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device implemented in the flowcharts and / or block diagrams. Figure 1 The function specified in one flow or multiple flows and / or blocks Figure 1 The function specified in one flow or multiple flows and / or blocks

[0150] These computer program instructions can also be stored in a computer readable storage medium capable of directing the computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable storage medium produce a manufactured product including instruction devices, which implement the functions specified in the flowcharts and / or block diagrams. Figure 1 The function specified in one flow or multiple flows and / or blocks Figure 1 The function specified in one flow or multiple flows and / or blocks

[0151] These computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable data processing devices to generate a computer-implemented process, so that the instructions executed on the computer or other programmable data processing devices provide steps for implementing the function specified in the flowchart Figure 1 one or more flows and / or blocks Figure 1 one or more blocks or steps of the function specified in the flowchart

[0152] The content of the present specification not described in detail belongs to the prior art known to those skilled in the art.

Claims

1. A seat safety vibration control system based on health state dynamic inference, characterized in that, include: Sensing module: Used to collect user physiological signals; Local decision module: used to analyze the collected user physiological signals to obtain the user's health status classification results and corresponding seat vibration and environmental control execution schemes; Seat multimodal vibration execution module: used to adjust the angle and height of the seat in all directions according to the seat vibration and environmental control execution scheme, and to independently control the opening and closing of the oscillators and amplitude of each zone on each seat.

2. The health status dynamic inference based seat safety vibration control system of claim 1, wherein, The local decision-making module also includes a demand decision model mounted on an automotive-grade chip for obtaining user health status classification results.

3. The health status dynamic inference based seat safety vibration control system of claim 2, wherein, The sensing module is also used to collect user demographic characteristics and real-time environmental data.

4. The health status dynamic inference based seat safety vibration control system according to any one of claims 1-3, wherein, It also includes a federated learning module, which uses a local model based on the FedProx framework and a cloud-based global model. It combines user demographic features, physiological signals, and real-time environmental data collected by the perception module with feature label information from multiple users in the cloud to perform local model training, cloud-based global model parameter aggregation, and medical knowledge graph-assisted calibration operations to obtain updated parameters for the local decision module.

5. The health status dynamic inference based seat safety vibration control system of claim 4, wherein, The cloud-based global model is deployed in the cloud and receives local model parameters uploaded by multiple users based on the FedProx framework. The local model is trained and generated based on user demographic features, physiological signals and real-time environmental data collected by the perception module. The local model parameters uploaded by multiple users are fused through an aggregation algorithm to generate general parameters that are optimized in the cloud and cover common features of multiple users. The general parameters are calibrated in conjunction with a medical knowledge graph before being distributed.

6. The health status-based dynamic inference enabled seat safety vibration control system of claim 5, wherein, The local model parameters uploaded by multiple users are encrypted.

7. The health status dynamic inference based seat safety vibration control system of claim 2, wherein, The demand decision model in the local decision module is a health analysis model trained based on a deep learning architecture, and the training data covers health samples from multiple users.

8. The health status dynamic inference based seat safety vibration control system according to any one of claims 1-3, wherein, The multimodal vibration execution module of the seat includes a targeted oscillator submodule, which independently controls the on / off state and adjusts the frequency and amplitude parameters of the zoned oscillators, including the headrest and / or lumbar and / or hip and / or Achilles tendon. For different pathological users, based on the past medical history in the demographic characteristics collected by the sensing module, the vibration prohibition area adapted to the pathological state is closed, and vibration is output to the non-vibration prohibition area.

9. The health status dynamic inference based seat safety vibration control system according to any one of claims 1 to 3, wherein, It also includes a data preprocessing module, which is used to clean the data collected by the perception module to obtain user feature tag information.

10. The health status-based dynamic inference enabled seat safety vibration control system of claim 9, wherein, The data preprocessing module includes a denoising submodule, which uses Kalman filtering to filter the physiological signals acquired by the sensing module.

11. The health status dynamic inference based seat safety vibration control system of claim 10, wherein, The denoising submodule is also used to filter the seat pressure signal in the real-time environmental data collected by the sensing module using wavelet transform.

12. The seat safety vibration control system based on dynamic reasoning of health status as described in claim 10 or 11, characterized in that, The data preprocessing module includes a feature extraction submodule, which is used to extract the time-domain features, frequency-domain features and EMG features of the denoised data.

13. The seat safety vibration control system based on dynamic reasoning of health status as described in claim 12, characterized in that, The data preprocessing module includes a privacy encryption submodule, which anonymizes and encrypts the past medical history, HRV and heart rate data in the user's demographic characteristics collected by the sensing module.

14. A seat safety vibration control method based on dynamic reasoning of health status in the system as described in claim 1, characterized in that, include: Collect user physiological signals; Based on the analysis of the collected physiological signals, the results of the user's health status classification and the corresponding seat vibration and environmental control execution schemes are obtained; The seat's angle and height are adjusted in all directions according to the seat vibration and environmental control execution scheme, and the opening and closing of the oscillators and the amplitude of each zone on each seat are controlled independently.

15. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the seat safety vibration control method based on dynamic reasoning of health status as described in claim 14.

16. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the seat safety vibration control method based on dynamic reasoning of health status as described in claim 14.

Images