Optimized stress transfer interfaces enabled wearable nano-electronics for fatigue driving monitoring

Structure and working mechanisms of the interfacial engineering-based triboelectric sensor (IETS)

During ventricular systole, the ejection of blood from the heart into the central and peripheral arteries induces a pulse wave. The real-time monitoring of mechanical information of pulse waves, such as amplitude and frequency, by using wearable sensors can enable applications such as early warning of cardiovascular diseases and assessment of fatigue states. As shown in Fig. 1a and Supplementary Fig. 1, by integrating triboelectric sensors into the wristband of a watch, the mechanical signals of pulse beats can be directly converted into electrical signals for output. After signal processing and Bluetooth module on the front side of the wristband, high-quality pulse wave signals will be sent to a mobile APP for real-time display. Machine learning algorithms in the back-end, by extracting and calculating the characteristic values of the pulse wave signals, can achieve the monitoring of the user’s cardiovascular health and fatigue state. Figure 1b illustrates the structure and equivalent circuit diagram of IETS. The device is composed of a triboelectric sensor with mountain-like triboelectric interface and a piezoelectric sensor with pillar-like sensor-skin interface. By sharing a common electrode to integrate these two kinds of sensors, it realizes aggregating charges on the common electrode from two sensors for amplifying the electrical signal. The working principle of the IETS is depicted in Supplementary Fig. 2. The triboelectric pressure sensor at the lower section is endowed with a specially designed mountain-like microstructures (Fig. 1c), which is constructed by the partial stacking of two conical microstructures with varying heights. The selection criteria for the silicone rubber-silver triboelectric pair are systematically presented in Supplementary Fig. S3. The working mechanism of the triboelectric sensor under different pressures is shown in Supplementary Figs. 4–6. Supplementary Note 1 provides a detailed discussion of the sensing mechanism of the triboelectric sensor under different pressures. The upper part of the IETS features a piezoelectric sensor whose primary component is a polymer-based piezoelectric material (PVDF) with frustums microstructure (Fig. 1d). The lower electrode of the piezoelectric sensor also serves as the upper electrode of the triboelectric sensor and the electrical coupling enhances the output signals. To minimize the impact of external factors, such as body sweat and wear, on the device’s performance, a polymer protective film has been applied to the surface of the piezoelectric sensor’s upper electrode through a spray coating process. The piezoelectric micropillars and mountain-like microstructures are vertically aligned in our design. The vertical alignment of piezoelectric micropillars with peak-like microstructures ensures direct stress transfer to triboelectric zones, avoiding signal loss in non-contact regions.

Figure 1e presents a photograph of a wearable tribo-electronic system designed for driver fatigue and health monitoring. The exterior of the wristband is equipped with a flexible circuit board for signal processing and data transmission, while the interior of the wristband is secured with IETS. From a micro perspective, the human wrist is not an absolute plane. When the sensor contacts with the wrist, the traditional sensor with a smooth outer surface encounters recessed regions, preventing it from closely adhering to the skin, as depicted in left panel of Fig. 1f. The weak pulsation cannot effectively transfer mechanical energy to the device at these recessed interfaces, resulting in a low signal-to-noise ratio for the sensor. Additionally, the triboelectric interface with a regular structure tends to quickly reach its deformation limit under the action of external forces, leading to a rapid decline in the sensitivity of the triboelectric sensor in high pressure range. This work optimizes the transfer and distribution of stress through interface microstructures (right panel of Fig. 1f), achieving precise detection of pulse waves by IETS even under a pre-stress. The incorporation of the piezoelectric sensor not only enhances the output signal but also makes a significant positive contribution in terms of mechanical properties. Pillar-like microstructures, fabricated by a laser demolding technique, can closely adhere to the skin under the pre-tension of the wristband (Supplementary Figs. 7 and 8). These microstructures efficiently transfer the faint mechanical energy from pulse beats to the main body of IETS, generating a distinct electrical signal. Moreover, the piezoelectric microstructure themselves generate charge through the piezoelectric effect, serving dual roles in stress transfer and electrical signal enhancement. Figure 1g clearly illustrates the stress distribution of the device with and without the piezo-frustums microstructures. In areas where the wrist is recessed, stress cannot be effectively transmitted to the triboelectric sensor. However, with the presence of piezo-frustums microstructures, stress is efficiently conveyed to the sensor. For the triboelectric sensor, the larger the effective contact area, the higher for the output signal. Depressions at the skin-device interface leads to a reduction in the effective working area of the triboelectric sensor. The introduction of piezo-frustums microstructures allows for the effective transmission of pulse-induced stress at recessed areas, thereby increasing the effective working area of the triboelectric sensor. The triboelectric interface with mountain-like microstructures endows the sensor with high sensitivity under high pressure through multiple stress concentration points. Under a low pressure, the deformation is predominantly localized in the taller peak, which ensures the device’s high sensitivity for tiny pressure. With the application of a preload, the deformation capacity of the taller peaks is constrained after compression. The tips of the secondary peaks can provide the necessary deformation under this condition, ensuring the device’s high sensitivity under a greater pressure (Fig. S7, Supporting Information).

Mechanism of the laser-based microstructure fabrication method

In this work, we propose a method for preparing surface microstructures on the interface of polymer materials by using a CO₂ laser to pattern etch PMMA substrate. Herein, the formation mechanism of the microstructures is further explained. The microscopic process of the CO₂ laser acting on the PMMA substrate is depicted in Fig. 2a. According to the keyhole model, when the energy of the laser is absorbed by PMMA, a high temperature is instantly generated internally, as illustrated in Supplementary Note 2. When the temperature is higher than the vaporization point of PMMA, the intermediate PMMA vaporizes and evaporates, leaving cavities in the substrate^36,37,38. Around the cavity, since the temperature is lower than the vaporization point of PMMA, the PMMA becomes a liquid molten pool. When the laser is turned off or moves away, the molten pool cools and becomes solid PMMA, leaving the middle hole structure (Supplementary Movie 1). The energy of the light beam emitted by the laser generally obeys a Gaussian distribution; thus the energy absorbed by PMMA also shows a Gaussian distribution in the irradiated area by the laser. Figure 2b shows the simulation and experimental results of the cavity depth in the PMMA substrate under different laser power conditions. When the distance between the laser head and the PMMA is fixed, the higher the laser power is, the deeper the cavity depth in the substrate. Figure 2c depicts the simulated temperature distribution in the substrate due to the Gaussian distribution of laser energy, resulting in a conical cavity morphology in PMMA. This enables indirect control of cone microstructure sizes by adjusting laser power to manipulate the temperature field. Figure 2d presents the relationship between the laser power and height of the conical microstructures. The higher the laser power is, the higher the height of the conical microstructures. Additionally, the conical microstructures created by laser etching are highly consistent due to the constancy of the laser power, as shown in Supplementary Fig. 9. Herein, a triboelectric layer with cone microstructures (width of 500 μm and height of 600 μm) is chosen as an example to analyze the deformation process under pressure. Supplementary Fig. 10 shows the deformation process of microstructures under certain pressures through simulation and experimental pressing tests. As the pressure increases, the vertical compression deformation of the cone microstructures increases. Due to the specific stress concentrations of the microstructures, the device deforms greatly, even under a low pressure. Different microstructures possess distinct stress concentration effects, thereby endowing the sensor with diverse sensing capabilities.

By designing various etching patterns, microstructures with different morphologies can be obtained. As depicted in Fig. 2e and Supplementary Fig. 11, when the laser etching pattern consists of two intersecting circles with one circle etched at a low power, mountain-like cavities with two apexes can be attained in the PMMA mold. In the subsequent replica molding process, the cavities can be filled with silicone rubber, thereby obtaining a mountain-shaped surface microstructure on the cured silicone rubber layer. By adjusting the laser power and the distance between the laser head and substrate, the mountain-like microstructure can exhibit the same height as the conical microstructure (Supplementary Fig. 12). Figure 2f presents a comparison of the displacement of the two types of microstructures under the same pressure through simulation. The mountain-like microstructure exhibits ~150 μm deformation under the same pressure, which is greater than that of the conical microstructure with same bottom diameter and same height. The outcome is mainly from the smaller volume of the mountain-like microstructure comparing to the conical microstructures of the same size to resist the applied external force. In addition, it has two peak structures at the top, which has a stronger stress concentration effect than the conical structure. This characteristic allows the microstructure to keep a large deformation capacity under high pressure without quickly reaching deformation saturation. Having greater deformation under high pressure implies that sensors with mountain-like microstructures maintain high sensitivity even under a pre-load stress. Figure 2g shows the potential for the large area fabrication of the mountain-like microstructures based on laser-etching processing, which also exhibits the capability for mass production of the IETS (Supplementary Figs. 13 and 14). Figure 2h shows the stress‒strain curves of different microstructures, further illustrating that the mountain-like microstructure can experience relatively great deformation under a large pressure. This result provides a mechanical basis for improving the sensitivities of triboelectric sensors under a prestress. To investigate the contribution of various interface structures to the output signal, the charge output of three kinds of devices is shown in Fig. 2i. A smooth acrylic plate was selected as the contact interface to avoid measurement errors caused by surface roughness. The test results indicate that the output signal is primarily provided by the TENG in the lower half of IETS. The contribution of the piezoelectric micropillars to the sensor mainly comes from their ability to enhance the stress transfer efficiency at the interface between the sensor and external objects. Concurrently, the piezoelectric micro-columns in the upper half also offer a portion of the electrical output, which is coupled with the signal of the TENG to enhance the signal-to-noise ratio.

Sensing performance of IETS

In order to authentically simulate the device’s response to pressure in a wearable scenario, the surface of a silicone rubber pad is etched using a laser to mimic the irregular curvature and roughness characteristic of the skin surface. For the testing procedure, the rough side of the silicone rubber pad is brought into close contact with the sensor, while a pressure gauge is utilized to exert force on the smooth side of the silicone rubber pad. As shown in Fig. 3a, the sensitivities of sensor with piezo-frustums microstructures are higher than those of devices with mountain-like and cone microstructures but without piezo-frustums. This is because the outer surface of the sensor without piezo-frustums is a smooth plane, when in contact with the roughness silicone pad, stress can be only transmitted through points of partial contact. This results in a smaller effective contact area at the triboelectric interface, thereby reducing the sensitivity of the triboelectric sensor. For IETS with both piezo-frustums and mountain-like microstructure, it exhibits the highest sensitivity of 4.28 V/kPa within the pressure range of 0-12 kPa and retains a high sensitivity of 0.18 V/kPa even under the pressure over 100 kPa. To further demonstrate the device’s pressure sensing capability across a wide pressure range, the sensor’s output voltages were measured under varying pressure conditions (50 kPa, 70 kPa, 90 kPa and 110 kPa), demonstrating maintained pressure resolution even under high-pressure conditions. Additionally, by placing the sensor on a volunteer’s heel and monitoring the voltage output during slow walking, the study provided further evidence of the sensor’s ability to detect pressure variations (Supplementary Fig. 15). The high sensitivity of IETS can be attributed to two aspects: the optimization of interfacial stress and the enhanced output signals by the coupling of piezoelectric and triboelectric effects. Interestingly, among devices with piezo-frustums, those with a mountain-like microstructure at the triboelectric interface exhibit higher sensitivity and a broader detection range under low pressure compared to those with conical structures. This is due to the stress concentration effect of the peak-like microstructures, which can produce large deformation, resulting in significant changes in contact area and gap distance. Overall, IETS with piezo-frustums and mountain-like microstructures have higher sensitivity and wider detection range, making it more suitable in the application that pre-load stress is required.

To investigate the influence of multi-peak microstructure shapes and peak number on sensor performance, finite element simulation and experiments are conducted as shown in Supplementary Fig. 16 and 17. For a given base diameter, pyramidal multi-peak microstructures undergo greater deformation than conical multi-peak microstructures under the same pressure, theoretically resulting in higher sensitivity. The number of peaks in multi-peak microstructures also significantly impacts sensor performance. Specifically, with a fixed base diameter, increasing the number of peaks enhances sensor sensitivity but correspondingly narrows the detection range.

As shown in Fig. 3b, as the pressure increases, the output voltage of the IETS also increases, showing a good correlation. Supplementary Fig. 18 shows the output response of device with mountain-like microstructures but without piezo-frustums under an increasing pressure. It also proves the ability of triboelectric sensor to distinguish pressure amplitude. To further demonstrate the potential of the sensors in pressure distribution monitoring, a 30 × 20 sensor array has been fabricated as shown in Supplementary Fig. 19. The results show that the sensor array could accurately detect both the position and intensity of finger presses. When a roll of tape was placed on the sensor array, the output pressure distribution image clearly reflected the outer contour of the tape, confirming its promising application in pressure distribution detection. To test the limit of detection (LOD) of the IETS at low pressures, sand paper with different sizes are used as pressure sources on the surface of the sensor. As shown in Fig. 3c, the test results show that when the sand paper weight is reduced to 4.19 mg, the IETS reaches its response limit, corresponding to an LOD of 2 Pa. To further test the device response to low pressures, different numbers of water drops (around 50 mg of each drop) are continuously dripped on a weight surface with 5 kPa at the same height. Figure 3d shows that the IETS is highly sensitive to low pressures even under a pre-load stress of 5 kPa and it can clearly detect the dripping and accumulation of water drops on the weight surface with a fast response time of 70 ms (Supplementary Fig. 20). In addition, the durability of the sensor is an important indicator affecting its actual service life. As shown in Supplementary Fig. 21, when a cyclic pressure of 2 kPa is applied to the surface of the sensor, the device does not show obvious signal attenuation after 5000 cycles, which indicates that the device has excellent durability. Moreover, the IETS can monitor static pressure of 2 kPa in real time and the static drift within 45 min is <15%, as shown in Supplementary Fig. 22. Under normal operating environments below 75 °C and 80% RH, the IETS demonstrates excellent stability (Supplementary Fig. 23). Upon comprehensive analysis, the interfacial stress engineering at the sensor-skin interface and the triboelectric interface significantly enhances sensing performance. Figure 3e presents a performance comparison between IETS and recently reported triboelectric and piezoelectric sensors in terms of sensitivity and detection range^{39,40,41,42,43}. The reported IETS in this work possesses both high sensitivity and a broad detection range within a pressure range of <12 kPa. Although some previously reported triboelectric sensors have achieved higher sensitivity through stress concentration strategies, their high-sensitivity detection range is limited to below 7 kPa. By designing a mountain-like microstructure, the sensitivity and detection range of the sensor in this work are effectively ensured.

The pulse wave signal is an important indicator for cardiovascular health status and driver fatigue monitoring. The signal quality of the pulse wave directly determines the accuracy of the diagnosis. As shown in Fig. 3f, real-time monitoring characteristics of pulse waves from the same volunteer under a preload of 10 kPa during the same period by sensors with different microstructure types are compared. The results show that the IETS responds sensitively to pulse waves, and its output pulse wave signal has three obvious characteristic peaks. In contrast, the sensor with mountain-like microstructures but without piezo-frustums can just detect two characteristic peaks of the pulse wave. The device with 600 μm conical microstructures can just detect the main peak of the pulse wave, while the device without microstructures cannot effectively detect the pulse wave signal. The results for pulse wave monitoring demonstrate that the piezo-frustums and mountain-shaped microstructures can effectively improve the sensitivity of the triboelectric sensor and thus the IETS with piezo-frustums and mountain-shaped microstructures are selected in subsequent actual applications.

Applications of IETS in monitoring driver physiological signals and behaviors

To achieve real-time monitoring of driver’s health and fatigue state, the system has to obtain the pulse wave parameter of the driver in real time. Herein, the IETS is integrated into the strap of the smart watch and connected to the driver mobile terminal via a Bluetooth module to acquire and process pulse wave data. As shown in Fig. 4a, the sensor is integrated into a specific part of the strap so that the sensor is in close contact with the wrist artery when the watch is properly worn. When blood flow causes pulse beats in the artery, the highly sensitive sensor can capture these low-pressure signals in real time and convert them into electrical signals. The edge data processing module and Bluetooth module can process and send pulse wave data to the user mobile phone in real time. Figure 4b illustrates the schematic diagram of the real-time wireless pulse wave monitoring system based on IETS. Due to the low amplitude and frequency (<5 Hz) values of pulsation signals, this system is designed with unique amplification and filtering circuits. The pulse wave signal with a high signal-to-noise ratio is subsequently collected by an analog-to-digital converter (ADC) for initial shaping and feature extraction. The signal is then transmitted via Bluetooth to a mobile device for further analysis. Figure 4c presents the physical components of the hardware section of the system. Electronic components are integrated on a flexible circuit board, enabling the hardware system to bend and accommodate the users with various wrist shapes. When the complete hardware system is affixed to the surface of a cylindrical object with a radius of 2.5 cm, as shown in Fig. 4d, the system can continue to function as intended. Figure 4e displays the waveform of the pulse wave signal detected by IETS after undergoing analog signal processing. Evidently, the signal exhibits a high signal-to-noise ratio and can distinctly reflect information from multiple characteristic peaks. Figure 4f shows the Bluetooth connection interface on the mobile phone to prove the feasibility of the communication module. After the mobile phone and the sensor module successfully connect via Bluetooth, the sensor module can send preprocessed data to the mobile phone in real time and display the pulse waveform on the APP interface. This process provides a solid foundation for implementing subsequent eigenvalue extraction, frequency domain signal conversion and algorithms process. Figure 4g depicts the physical layout of the hardware system after the wristwatch strap is removed. Apparently, the sensor and flexible circuit board are installed on opposite sides of the strap that are interconnected by wires. Once the sensor and flexible circuit board are integrated with the wristwatch strap, the real-time collection of the driver pulse wave signal can be achieved, as shown in Fig. 4h. Assessing the driver fatigue level based on the pulse wave signal typically relies on indicators of heart rate variability (HRV). Figure 4i shows a conceptual diagram for calculating HRV. The core of HRV calculation is to obtain the time difference between two consecutive pulse beats. In practical measurement scenarios, the real-time collected pulse wave signal is a temporal signal. To effectively extract HRV parameters, the pulse wave signal is generally transformed into the frequency domain through fast Fourier transformation (FFT), as shown in Fig. 4j. In the frequency domain signal, the system can acquire parameters for each characteristic frequency and determine the volunteer fatigue state based on the proportion of these parameters^3,44,45. As depicted in Fig. S18, noticeable distinctions in characteristic frequency parameters are observed for the same volunteer between wakeful morning and drowsy states. Additionally, based on the feature values in the temporal signal of the pulse wave, the system can assess the user cardiovascular health status. On this basis, we develop a driver fatigue and health status monitoring system, as demonstrated in Fig. 4k and Supplementary Fig. 24. By processing and analyzing the temporal and frequency domain signals of raw data, this system enables real-time monitoring and early warning of driver fatigue and physical health status (Supplementary Movie 2).

In addition, when a driver is tired, they may display irregular behaviors such as frequent abrupt braking and yawning due to distractions. The wearable sensors can be utilized to monitor these driver behaviors, providing an alternative method for assessing fatigue. On this basis, IETSs are wore on the driver face or attached on driving cab components, such as the accelerator and brake pedals, for the real-time collection of driver behaviors, as shown in Supplementary Fig. 25. Due to the high sensitivity of the IETS in a wide pressure range, the sensor can accurately detect both low pressure signals, such as eye movements, and large pressure signals, such as stepping on the brake pedals and the driver leaving the seat. As shown in Supplementary Fig. 26a, when a IETS is attached to the corner of the eye, blinking causes slight pressure changes due to contraction of the eye muscles. The sensor will clearly capture the driver blinking action. In a normal state, the frequency of driver blinking is maintained near a fixed value. In a state of fatigue, the driver usually experiences a period of first unblinking and then rapid blinking. The IETS can clearly monitor the driver blinking signal for the evaluation of the driver fatigue state. Similarly, the sensor can be attached to the driver’s mouth corner to monitor yawning. As shown in Supplementary Fig. 26b, when the driver is in a state of fatigue, yawning causes the mouth corner muscles to contract for usually 2–5 s, which is significantly different from normal talking and chewing actions. Therefore, monitoring driver yawning behaviors through the sensors can assist in evaluating the state of driver fatigue. In addition, installing sensors on the accelerator and brake pedals can also monitor driver actions in real time. As shown in Supplementary Fig. 26c, the driver’s behavioral data acquired by the IETS installed on the accelerator pedal show that when the vehicle starts, the output voltage of the sensor first slowly increases and then remains stable. When overtaking is needed, the force applied by the driver foot increases, the output voltage of the sensor increases as well. Similarly, the IETS attached to the brake pedal can provide real-time feedback on the changes in the force applied by the driver foot when braking. When the driver is in a state of fatigue, emergencies cause the driver to brake frequently and suddenly, and the sensor outputs sudden braking signals, as shown in Supplementary Fig. 26d.

Regarding to driver safety, IETSs can be integrated into seat belt buckles or cushions to determine whether the driver has fastened the seat belt or left the seat, respectively, by detecting changes in pressure (Supplementary Movie 3). As shown in Supplementary Fig. 26e, when the seat belt is fastened, the sensor in the buckle receives compression and outputs a continuous output signal. When the seat belt is released, the signal returns to the initial state. Supplementary Fig. S26f shows the output signal of the sensor under the seat cushion after being pressed by the driver weight. Upon the parallel integration of 12 IETSs to form an array device, the quantity of sensors under compression can be detected by assessing the magnitude of the output signal. This method allows for determination of the driver’s seating posture (Supplementary Fig. 27). The IETS can accurately detect whether the driver has seat or left the seat. These applications show that the sensor has an ultrawide detection range from low pressures, such as pulse waves, to large pressures, such as body weight.

Deep learning-enabled cardiovascular monitoring and driver fatigue monitoring

Biological signs can reliably indicate early fatigue and prevent accidents from occurring. Cardiovascular signals, such as electrocardiography (ECG) and photoplethysmography (PPG), can accurately monitor fatigue status. However, due to the limited viability of invasive sensors for real-time driver wear and the low sensitivity of noninvasive sensors influenced by humans or the environment, normal driver behavior and fatigue monitoring accuracy are compromised. Due to the high sensitivity in detecting imperceptible low pressures, the IETS can be used to monitor weak pulse waves to continuously assess cardiovascular and fatigue conditions. Moreover, a combination of artificial intelligence (AI), such as deep learning, can improve the accuracy and actionability of new sensing devices, ultimately facilitating real-time personal identification and driver fatigue (Fig. 5a). By comparing the voltage signal obtained by IETS with the typical arterial pulse wave in Fig. 5b, we find that the voltage signal matches the typical arterial pulse waveform and can detect low intra-arterial blood pressure oscillations with systolic, reflected and diastolic peaks (P₁, P₂ and P₃). These peaks can also be used to further calculated the detail feature parameters for artery evaluating. Parameters such as diastolic blood pressure, systolic blood pressure and blood flow velocity shown in the diagram can be used to accurately predict the user’s cardiovascular status through cloud-based algorithms, enabling early warning of sudden cardiovascular disease. Therefore, the cardiovascular condition and degree of arteriosclerosis can be assessed by comparing the measured results with references.

Figure 5c and Supplementary Fig. 28 show the reflected wave transit time (RWTT), systolic–diastolic time (PPT), upstroke time (UT) and left ventricular ejection time (LVET) from the continuous 20 voltage signal cycles, all of which are correlated with the reported reference values of a healthy individual. Cardiovascular conditions are related to heart rate (HR) and HRV. Figure 5 d reflects a mapped scatter points plot of pulse interval (Ti) to reflect the HRV when the subject is nonfatigued (HRV = 8.9) and fatigued (HRV = 5.4); HRV is a valuable predictor of sudden cardiac death and arrhythmic events. The lower the HRV value is, the more likely the subject is to suffer from acute myocardial infarction and arrhythmias, and the stronger the need to relax and rest in a timely manner. Moreover, the frequency–time distribution chart (Fig. 5e) illustrates that when the subject is fatigued, the frequency change over time is unstable, aiding in demonstrating the instability of heart rate variability during fatigue. Heart rate decreases significantly in the fatigued state of the subject, as shown in Fig. 5f–j. The heart rate and HRV index in pulse wave signals are crucial physiological indicators for evaluating fatigue. However, this method requires long-term pulse wave data, which cannot monitor and assess the driver fatigue status and alert the driver in real time. Thus, in this study, the pulse wave data are divided into time lengths of 750 ms per sample, and the short-term signals are classified by a one-dimensional (1 d) convolutional neural network (CNN)-based method for recognizing driver behavior and fatigue. To reduce the effects of noise, baseline drift and environment, the collected data are preprocessed (wavelet noise reduction, R-peak splitting and normalization, respectively). The detailed framework and parameters used to construct the CNN model are labeled in Supplementary Table 1. Supplementary Fig. 29 supplies the typical pulse signals of the 5 different subjects. Through data collection and signal preprocessing, 625 sample data points with each 750 data points constitute dataset 1 (80% training set and 20% test set). The average recognition accuracy is 94% (Supplementary Fig. 30a), providing great potential for high-accuracy behavioral recognition based on deep learning (DL) prediction. After training in the 1d-CNN model with 80 training epochs, the maximum accuracy is achieved, and the dropout layer avoids overfitting, as shown in Supplementary Fig. 30b. In contrast to unprocessed data and support vector machine (SVM) models, which require pre-extracted features, the end-to-end CNN model exhibits increased accuracy and reduced overall complexity in Supplementary Fig. 30c.

Pulse signal-based identification technology avoids the drawbacks of traditional identity authentication and is less susceptible to copying and counterfeiting. Combining it with deep learning networks increases the convenience and effectiveness of identification. As depicted in Fig. 5h, k, the morphology of the pulse changes after fatigue, with reflected and diastolic peaks decreasing relative to the systolic peaks. By merging the data of the different states for several days, dataset 2 is trained and validated by the deep learning network described above. The average accuracy reaches 98% for one subject, as shown in Fig. 5 l. Fatigue classification based on short-time pulse signals can achieve the real-time and accurate detection of driver fatigue, providing a certain guarantee for safe driving.