Psychological stress detection method and device based on transfer learning

Abstract

The present invention provides a method and device for detecting psychological stress based on transfer learning, the method comprising: obtaining and preprocessing a user's multimodal physiological signals; inputting the preprocessed user's multimodal physiological signals into a stress detection migration model, and obtaining the user's stress level output by the stress detection migration model; wherein the stress detection migration model comprises a feature extractor and a stress classifier, the stress detection migration model is obtained by adversarial training with a domain discriminator, the feature extractor is used to extract deep features and fuse them with artificial time-frequency domain features, the stress classifier is used to classify the user's stress level, and the domain discriminator is used to distinguish whether sample data belongs to a source domain or a target domain. The present invention transfers stress detection knowledge applicable to a large population to a small number of target group stress detection tasks, providing a more convenient and accurate solution for establishing personalized models for different groups.

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to a psychological stress detection method and device based on transfer learning. Background Art

[0002] Stress monitoring is one of the key links in the field of healthcare. Stress affects people's cognitive behavior and decision-making, and has a significant impact on people's work efficiency. Long-term excessive stress and negative emotions may even harm people's physical and mental health. In particular, some special occupational groups (such as police, doctors, drivers, etc.) are susceptible to acute stress stimulation in the working environment. Monitoring their response to stress stimulation is conducive to ensuring their physical and mental health and work safety. Physiological signals commonly used for stress detection include electrodermal activity (EDA), photoplethysmography (PPG), skin temperature (ST), respiration and electroencephalogram (EEG). With the development of wearable sensors, the use of wrist-worn devices provides a portable, unobtrusive and non-invasive way to collect physiological signals, enabling the rapid development of real-time stress detection applications in real-life scenarios.

[0003] Conventional methods mostly select a single or multiple physiological signals, calculate their time domain and frequency domain features respectively, and use statistical analysis or traditional machine learning methods such as support vector machines, K nearest neighbor algorithms, and random forests to establish the relationship between these features and stress levels, thereby achieving stress assessment and prediction. In recent years, with the development of deep learning technology, deep networks have gradually been applied to the field of stress detection based on physiological signals. These methods usually directly input physiological signals into deep networks to extract time series features, or extract their spectral features through time-frequency diagrams of physiological signals. Although deep learning-based algorithms have achieved superior performance in stress assessment, they require the models to be trained on larger data sets.

[0004] Although different people have certain similarities in their physiological response trends to stimuli, the intensity of stress induction will be affected by many factors such as gender, nationality, occupation, and collection environment. In particular, people in special occupations will have different reactions to their occupation-related stressors than the general population. Existing technologies often assume that the data distribution of the training data set is consistent with that of the test data set, while ignoring the differences between the model training group and the target group to be tested. Therefore, if the model established on the training data set is directly used to perform stress detection on a specific target population, it is likely to cause a large detection bias due to group differences. If the model is retrained based on the target population data, it will be difficult to collect a sufficient number of target population subjects, and it will take a lot of time and manpower.

[0005] A large number of stress detection studies have shown that some artificial time-frequency domain features calculated manually have a significant correlation with stress and have good interpretability; deep networks have good feature expression and model building capabilities, and have higher accuracy and robustness than traditional algorithms, but network construction relies on a large amount of training data. Existing technologies only analyze artificial time-frequency domain features or deep features, and do not make better use of the advantages of both features. Summary of the invention

[0006] The present invention provides a psychological stress detection method and device based on transfer learning, which is used to solve the defects of the prior art that the difference between the model training group and the target group to be tested is ignored and the two characteristics are not combined for analysis, resulting in a large deviation in stress detection. The present invention can achieve a higher stress detection accuracy, especially in stress detection of special occupational groups, and can obtain better results.

[0007] The present invention provides a psychological stress detection method based on transfer learning, comprising:

[0008] Acquire and preprocess the user's multimodal physiological signals;

[0009] Inputting the preprocessed multimodal physiological signal of the user into the stress detection migration model to obtain the stress level of the user output by the stress detection migration model;

[0010] Among them, the stress detection migration model includes a feature extractor and a stress classifier. The stress detection migration model is obtained by adversarial training with a domain discriminator. The feature extractor is used to extract deep features and fuse them with artificial time-frequency domain features. The stress classifier is used to classify the user's stress level. The domain discriminator is used to distinguish whether the sample data belongs to a source domain or a target domain. The source domain and the target domain are two groups with different stress-inducing intensities.

[0011] According to a psychological stress detection method based on transfer learning provided by the present invention, the feature extractor is also used for:

[0012] The deep features are fused and dimensionally reduced and then cascaded with artificial time-frequency domain features, and the cascaded features are fused through a fully connected network.

[0013] According to a psychological stress detection method based on transfer learning provided by the present invention, the preprocessing includes: motion segment removal, filtering and denoising, and outlier inspection and removal.

[0014] According to a psychological stress detection method based on transfer learning provided by the present invention, the stress classifier is used to classify the stress level of the user, including: the stress classifier classifies the stress level of the user according to the input artificial time-frequency domain features and the depth features, and the stress level includes high stress and low stress;

[0015] The domain discriminator is used to distinguish whether the sample data belongs to a source domain or a target domain, including: the domain discriminator is used to be trained by domain labels of source domain and target domain sample data, and is used to distinguish whether the sample data belongs to the source domain or the target domain according to the input artificial time-frequency domain features and the deep features.

[0016] The present invention also provides a psychological stress detection device based on transfer learning, comprising:

[0017] Acquisition module, which acquires and preprocesses the user's multimodal physiological signals;

[0018] A detection module, used for inputting the pre-processed multimodal physiological signal of the user into a stress detection migration model, and obtaining the stress level of the user output by the stress detection migration model;

[0019] Among them, the stress detection migration model includes a feature extractor and a stress classifier. The stress detection migration model is obtained by adversarial training with a domain discriminator. The feature extractor is used to extract deep features and fuse them with artificial time-frequency domain features. The stress classifier is used to classify the user's stress level. The domain discriminator is used to distinguish whether the sample data belongs to a source domain or a target domain. The source domain and the target domain are two groups with different stress-inducing intensities.

[0020] According to a psychological stress detection device based on transfer learning provided by the present invention, the feature extractor is also used for:

[0021] The deep features are fused and dimensionally reduced and then cascaded with artificial time-frequency domain features, and the cascaded features are fused through a fully connected network.

[0022] According to a psychological stress detection device based on transfer learning provided by the present invention, the preprocessing includes: motion segment removal, filtering and denoising, and outlier inspection and removal.

[0023] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any of the above-mentioned psychological stress detection methods based on transfer learning are implemented.

[0024] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-mentioned psychological stress detection methods based on transfer learning.

[0025] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of any of the above-mentioned psychological stress detection methods based on transfer learning.

[0026] The present invention provides a psychological stress detection method and device based on transfer learning, and proposes an improved scheme for the situation where the training group is different from the target group to be tested for stress. A personalized group stress detection migration model can be trained by adding only a small number of labeled samples of the target group. A transfer learning algorithm is introduced to make the feature distribution of the two domains consistent by mining the similarities between the target domain and the source domain, and the source domain stress detection knowledge is transferred to the target domain to achieve stress detection in the target domain. Specifically, in addition to the feature extractor and the stress classifier, a domain discriminator is also added to distinguish whether the training data comes from the target domain or the source domain. The optimization goals of the feature extractor include two items: extracting feature vectors that are more conducive to stress detection, and generating feature vectors that are difficult to be successfully distinguished by the domain discriminator. Through this adversarial idea, feature vectors that are insensitive to domain differences can be obtained in the process of iterative optimization, so that the detection model can also obtain a higher stress detection accuracy on the target group.

[0027] The present invention constructs a stress classifier for giving stress detection results, and constructs a domain discriminator for distinguishing whether the input samples are from the target domain or the source domain during the training process. The partial network for feature extraction and fusion, the stress classifier, and the domain discriminator are jointly trained using a transfer learning algorithm based on adversarial thinking to extract robust features that are insensitive to group differences and implement stress detection migration model training suitable for the target group. The present invention can transfer stress detection knowledge suitable for a large population to the target group stress detection task while only obtaining a small number of target group samples, providing a more convenient and accurate solution for establishing personalized models for different groups. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0029] Figure 1 is a flow chart of a psychological stress detection method based on transfer learning provided by an embodiment of the present invention;

[0030] Figure 2 is a network structure diagram of a psychological stress detection method based on transfer learning provided by an embodiment of the present invention;

[0031] Figure 3 is a data preprocessing flow chart of a psychological stress detection method based on transfer learning provided by an embodiment of the present invention;

[0032] Figure 4 This is a data collection experimental process of a psychological stress detection method based on transfer learning provided by an embodiment of the present invention;

[0033] Figure 5 It is a structural schematic diagram of a psychological stress detection device based on transfer learning provided by the present invention;

[0034] Figure 6 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] Combine the following Figure 1 The present invention describes a psychological stress detection method based on transfer learning, comprising:

[0037] S101, obtaining and preprocessing multimodal physiological signals of a user;

[0038] S102, inputting the pre-processed multimodal physiological signal of the user into a stress detection migration model, and obtaining the stress level of the user output by the stress detection migration model;

[0039] Among them, the stress detection migration model includes a feature extractor and a stress classifier. The stress detection migration model is obtained by adversarial training with a domain discriminator. The feature extractor is used to extract deep features and fuse them with artificial time-frequency domain features. The stress classifier is used to classify the user's stress level. The domain discriminator is used to distinguish whether the sample data belongs to a source domain or a target domain. The source domain and the target domain are two groups with different stress-inducing intensities.

[0040] The present invention proposes an improved scheme for the situation where the training group is different from the target group to be tested for stress. A personalized group stress detection migration model can be trained by adding only a small number of labeled samples of the target group. To this end, a transfer learning algorithm is introduced to make the feature distribution of the two domains tend to be consistent by mining the similarities between the target group (target domain) and the training group (source domain), and to migrate the source domain stress detection knowledge to the target domain to achieve stress detection in the target domain. Specifically, in addition to the feature extractor and the stress classifier, a domain discriminator is added to distinguish whether the training data comes from the target domain or the source domain. The optimization goals of the feature extractor include two items: extracting feature vectors that are more conducive to stress detection, and generating feature vectors that are difficult to be successfully distinguished by the domain discriminator. Through this adversarial idea, feature vectors that are insensitive to domain differences can be obtained in the process of iterative optimization, so that the detection model can also obtain a higher stress detection accuracy on the target group.

[0041] During the training phase, the feature extractor, stress classifier, and domain discriminator are jointly trained adversarially. The feature extractor is used to learn deep features that are insensitive to differences in user groups and fuse them with artificial time-frequency domain features. The stress classifier is used to classify the user's stress level. The domain discriminator is used to distinguish whether the sample data belongs to the source domain or the target domain. The source domain and the target domain are two groups with different stress-induced intensities. After training, the stress migration detection model only includes the feature extractor and the stress classifier.

[0042] It should be noted that the multimodal physiological signals include: skin electrical activity signals, photoplethysmographic signals and skin temperature signals. The psychological stress detection method based on transfer learning of the embodiment of the present invention uses a set of wristband wearable devices for data collection. The device can collect multimodal physiological signals including EDA, PPG, ST and accelerometer data. It is connected to a smartphone via Bluetooth and can display multi-channel signal waveforms in real time in the supporting APP.

[0043] The embodiment of the present invention takes the target group as police academy students and selects video stimulation to induce stress as an example. This method is also applicable to other target groups and stress induction methods. Participants belonging to two groups were recruited: police academy students (target domain) and ordinary college students (source domain). The police academy student group included 32 healthy participants (24 males and 8 females) with an average age of 20.9 years old. The ordinary college student group included 264 healthy participants (151 males and 113 females) with an average age of 22.8 years old.

[0044] During the preparation of the video stimulus material, the blood cortisol concentrations of another 10 subjects were collected before and after watching the video. The experimental results prove that the videos selected in this study can effectively induce stress response. The collection experiment plan includes 4 stages: relaxation 1, video 1, relaxation 2, video 2. Participants are required to wear wearable devices and sit in front of the screen for about 10 minutes, and sit quietly or watch the video according to the prompts. After each stage, the subjects were asked to verbally report their current self-assessed level of stress. The data collection experiment process is designed as follows Figure 4 As shown, including:

[0045] Relaxation 1: Sit quietly for 2 minutes.

[0046] Video 1: Watch a clip from the movie "Final Destination 5", which is 3 minutes and 25 seconds long.

[0047] Relaxation 2: Sit quietly for 2 minutes.

[0048] Video 2: Watching a documentary clip about the riots, duration 2 minutes 56 seconds. This stimulus is closely related to the future working environment of police academy students.

[0049] In the embodiment of the present invention, the training data can use two stress level labeling methods to indicate the stress level obtained at the end of each stage: stress level labeling based on self-report and stress level labeling based on the experimental induction stage.

[0050] Stress level annotation based on self-report: During the data collection process, the self-assessed stress levels of the participants at each stage were obtained and annotated with 0 and 1 as stress category training labels, with 0 representing low stress and 1 representing high stress.

[0051] Stress level annotation based on the experimental induction stage: The data were divided into two categories according to the experimental stage of data collection, namely low stress (corresponding to relaxation 1 and relaxation 2 stages) and high stress (corresponding to video 1 and video 2 stages).

[0052] like Figure 3 As shown, the present invention performs data preprocessing on the collected signal to remove noise and abnormal data points on the multimode signal. The preprocessing includes:

[0053] S201, motion segment removal;

[0054] S202, filtering and denoising;

[0055] S203: Check and remove outliers.

[0056] It should be noted that the motion segment removal includes screening signal segments based on hand motion, estimating hand motion according to acceleration sensor signals, and removing low-quality data segments caused by violent motion.

[0057] The filtering and denoising includes filtering the PPG signal with a fourth-order Butterworth bandpass filter with a cutoff frequency of 0.5-3 Hz; filtering the EDA signal with a fourth-order Butterworth low-pass filter with a cutoff frequency of 1 Hz; and filtering the ST signal with a moving average filter.

[0058] The outlier detection and removal includes performing singular value detection on the signal based on wavelet transform, and determining and eliminating outliers by detecting the modulus maximum point of the wavelet transform coefficient.

[0059] It should be noted that the time-frequency domain features are extracted by the following method:

[0060] In an embodiment of the present invention, when performing time-frequency domain feature extraction, the EDA signal is decomposed into tonic components and phasic components related to the skin conductance level (SCL) and the skin conductance response (SCR), respectively. Based on these components, features such as the mean, standard deviation, number of fluctuations, and rise time are calculated. For the PPG signal, the peak detection algorithm is first executed to extract the RR interval for calculating the heart rate variability (HRV). The extracted PPG features include time domain features such as beats per minute (BPM) and beat interval (IBI) and frequency domain features such as low frequency (LF) power and high frequency (HF) power. For the ST signal, its mean and standard deviation are calculated as features. The selected artificial time-frequency domain features and their descriptions are shown in Table 1.

[0061] Table 1 Artificial time-frequency domain features and their descriptions

[0062]

[0063]

[0064] It should be noted that the deep features are extracted by the following method: The embodiment of the present invention uses a short-time Fourier transform with a window length of 1s to obtain the frequency components of the PPG and EDA signals in different time periods, and the time series data is converted into an image representation based on the frequency domain. Then, the spectrogram is input into the pre-trained VGG16 network for deep feature extraction. Among them, the parameters of the last two layers of the VGG16 network participate in the joint training of the model, and the parameters of other layers are fixed. The ST signal is input into a two-layer long short-term memory network (LSTM) for temporal feature extraction.

[0065] In at least one embodiment of the present invention, the feature extractor is further used for:

[0066] The deep features are fused and dimensionally reduced and then cascaded with artificial time-frequency domain features, and the cascaded features are fused through a fully connected network.

[0067] The embodiment of the present invention combines the advantages of artificial time-frequency domain features and deep features, extracts two features of multimodal signals respectively and fuses them by adding a fully connected layer. This feature extraction and fusion method is closely integrated with the training and optimization phase of the above-mentioned migration algorithm, which helps to enhance the robustness of the extracted feature vector.

[0068] In at least one embodiment of the present invention, the stress classifier is used to classify the stress level of the user, including: the stress classifier classifies the stress level of the user according to the input artificial time-frequency domain features and the depth features, and the stress level includes high stress and low stress;

[0069] The domain discriminator is used to distinguish whether the sample data belongs to a source domain or a target domain, including: the domain discriminator is used to be trained by domain labels of source domain and target domain sample data, and is used to distinguish whether the sample data belongs to the source domain or the target domain according to the input artificial time-frequency domain features and the deep features.

[0070] It should be noted that, in the embodiment of the present invention, the data in the source domain can be represented as S s =*X s ,Y s ,G s}, where X s represents the intermediate result of feature extraction in the source domain that is input into the trainable network, Y s Indicates the corresponding pressure label, G s = *0,0,…,0} is the domain label of the source domain. Similarly, the data of the target domain can be represented as S t =*X t ,Y t ,G t}, where G t=*1,1,…,1} is the domain label of the target domain. The ultimate goal is to build a discriminative model that can predict the stress level y given x and is insensitive to domain differences.

[0071] like Figure 2 As shown, the network structure diagram of the psychological stress detection method based on transfer learning provided by an embodiment of the present invention includes a feature extractor F (only the variable parameter part and the feature fusion part participate in the optimization calculation of the algorithm), a stress classifier C for giving stress assessment results, and a discriminator D for determining which field the sample belongs to. The goal of D is to distinguish the source domain and the target domain based on the features generated by F, while the goal of F is to generate domain-independent features as much as possible. The two modules can make the feature distributions of the source domain and the target domain after transformation consistent through mutual game. At the same time, the training of C ensures that F can extract features that are helpful for stress detection, so that the generated model can be applied to stress assessment in both the source domain and the target domain.

[0072] It should be noted that in the embodiment of the present invention, both D and C are composed of multiple fully connected layers and a softmax layer. The two fully connected layers that fuse deep features and artificial time-frequency domain features have 50 units each. The pressure classifier contains two fully connected layers, with the number of units set to 50 and 20; the discriminator contains two fully connected layers, with the number of units set to 30 and 10.

[0073] For a multimodal signal x input by the user, the output of F is expressed as F(x), the output of D is expressed as D(F(x)), and the output of C is expressed as C(F(x)). Then the classification losses of C and D can be expressed as:

[0074]

[0075]

[0076]

[0077] in represents the mathematical expectation, K is the number of pressure categories, in the embodiment of the present invention, K=2, C k and D i denote the components corresponding to the kth and ith categories in the output of C and D, respectively. The low-pressure and high-pressure categories are denoted as k=0 and k=1, respectively. The source domain and target domain categories are denoted as i=0 and i=1, respectively. Based on the above analysis of the purposes of the three modules, C minimizes L cls To optimize, D is minimized by L adv To optimize, F is minimized by L cls and maximize L advOptimize. The optimization formulas of C, D and F can be expressed as:

[0078]

[0079]

[0080]

[0081] During the network training process, the parameters of the three modules are updated alternately according to the above three formulas.

[0082] Combine the following Figure 5 The psychological stress detection device based on transfer learning provided by the embodiment of the present invention is described. The psychological stress detection device based on transfer learning described below and the psychological stress detection method based on transfer learning described above can be referred to each other. A psychological stress detection device based on transfer learning includes:

[0083] Acquisition module 501, used to acquire and pre-process the multimodal physiological signals of the user;

[0084] A detection module 502 is used to input the pre-processed multimodal physiological signal of the user into a stress detection migration model to obtain the stress level of the user output by the stress detection migration model;

[0085] Among them, the stress detection migration model includes a feature extractor and a stress classifier. The stress detection migration model is obtained by adversarial training with a domain discriminator. The feature extractor is used to extract deep features and fuse them with artificial time-frequency domain features. The stress classifier is used to classify the user's stress level. The domain discriminator is used to distinguish whether the sample data belongs to a source domain or a target domain. The source domain and the target domain are two groups with different stress-inducing intensities.

[0086] The transfer learning-based psychological stress detection device of the embodiment of the present invention constructs a stress classifier to give stress detection results, and constructs a domain discriminator to distinguish whether the input samples come from the target domain or the source domain during the training process. The partial network for feature extraction and fusion, the stress classifier, and the domain discriminator are jointly trained using a transfer learning algorithm based on adversarial thinking to extract robust features that are insensitive to group differences and implement stress detection migration model training suitable for the target group. The present invention can transfer stress detection knowledge suitable for a large population to the target group stress detection task when only a small number of target group samples are obtained, providing a more convenient and accurate solution for establishing personalized models for different groups.

[0087] In at least one embodiment of the present invention, the feature extractor is further used for:

[0088] The deep features are fused and dimensionally reduced and then cascaded with artificial time-frequency domain features, and the cascaded features are fused through a fully connected network.

[0089] In at least one embodiment of the present invention, the preprocessing includes: motion segment removal, filtering and denoising, and outlier detection and removal.

[0090] The recognition accuracy experiment of the embodiment of the present invention is as follows, comparing the method in the embodiment of the present invention with the following baseline methods: a) Support vector machine trained using only artificial time-frequency domain features of the general population; b) Support vector machine trained using only artificial time-frequency domain features of the target population; c) Method with the same structure as the present invention but using only artificial time-frequency domain features; d) Method with the same structure as the present invention but using only deep features. The recognition accuracy is shown in the following table:

[0091] Table 2 Comparison of recognition accuracy of different methods

[0092]

[0093]

[0094] Figure 6 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 6 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630 and a communication bus 640, wherein the processor 610, the communication interface 620 and the memory 630 communicate with each other through the communication bus 640. The processor 610 may call the logic instructions in the memory 630 to execute the psychological stress detection method based on transfer learning, and the method includes:

[0095] S101, obtaining and preprocessing multimodal physiological signals of a user;

[0096] S102, inputting the pre-processed multimodal physiological signal of the user into a stress detection migration model, and obtaining the stress level of the user output by the stress detection migration model;

[0097] Among them, the stress detection migration model includes a feature extractor and a stress classifier. The stress detection migration model is obtained by adversarial training with a domain discriminator. The feature extractor is used to extract deep features and fuse them with artificial time-frequency domain features. The stress classifier is used to classify the user's stress level. The domain discriminator is used to distinguish whether the sample data belongs to a source domain or a target domain. The source domain and the target domain are two groups with different stress-inducing intensities.

[0098] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0099] On the other hand, the present invention further provides a computer program product, the computer program product includes a computer program, the computer program can be stored in a non-transitory computer-readable storage medium, when the computer program is executed by a processor, the computer can execute the psychological stress detection method based on transfer learning provided by the above methods, the method includes:

[0100] S101, obtaining and preprocessing multimodal physiological signals of a user;

[0101] S102, inputting the pre-processed multimodal physiological signal of the user into a stress detection migration model, and obtaining the stress level of the user output by the stress detection migration model;

[0102] Among them, the stress detection migration model includes a feature extractor and a stress classifier. The stress detection migration model is obtained by adversarial training with a domain discriminator. The feature extractor is used to extract deep features and fuse them with artificial time-frequency domain features. The stress classifier is used to classify the user's stress level. The domain discriminator is used to distinguish whether the sample data belongs to a source domain or a target domain. The source domain and the target domain are two groups with different stress-inducing intensities.

[0103] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to perform the psychological stress detection method based on transfer learning provided by the above methods, the method comprising:

[0104] S101, obtaining and preprocessing multimodal physiological signals of a user;

[0105] S102, inputting the pre-processed multimodal physiological signal of the user into a stress detection migration model, and obtaining the stress level of the user output by the stress detection migration model;

[0106] Among them, the stress detection migration model includes a feature extractor and a stress classifier. The stress detection migration model is obtained by adversarial training with a domain discriminator. The feature extractor is used to extract deep features and fuse them with artificial time-frequency domain features. The stress classifier is used to classify the user's stress level. The domain discriminator is used to distinguish whether the sample data belongs to a source domain or a target domain. The source domain and the target domain are two groups with different stress-inducing intensities.

[0107] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the embodiments of the present invention. Those of ordinary skill in the art may understand and implement them without creative effort.

[0108] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A psychological stress detection method based on transfer learning, characterized in that: include: Acquire and preprocess the user's multimodal physiological signals; Inputting the preprocessed multimodal physiological signal of the user into the stress detection migration model to obtain the stress level of the user output by the stress detection migration model; Among them, the stress detection migration model includes a feature extractor and a stress classifier. The stress detection migration model is obtained by adversarial training with a domain discriminator. The feature extractor is used to extract deep features and fuse them with artificial time-frequency domain features. The stress classifier is used to classify the user's stress level. The domain discriminator is used to distinguish whether the sample data belongs to a source domain or a target domain. The source domain and the target domain are two groups with different stress-inducing intensities.

2. The psychological stress detection method based on transfer learning according to claim 1 is characterized in that: The feature extractor is also used to: The deep features are fused and dimensionally reduced and then cascaded with artificial time-frequency domain features, and the cascaded features are fused through a fully connected network.

3. The psychological stress detection method based on transfer learning according to claim 1 is characterized in that: The preprocessing includes: motion segment removal, filtering and denoising, and outlier inspection and removal.

4. The psychological stress detection method based on transfer learning according to claim 1 is characterized in that: The stress classifier is used to classify the stress level of the user, including: the stress classifier classifies the stress level of the user according to the input artificial time-frequency domain features and the depth features, and the stress level includes high stress and low stress; The domain discriminator is used to distinguish whether the sample data belongs to a source domain or a target domain, including: the domain discriminator is used to be trained by domain labels of source domain and target domain sample data, and is used to distinguish whether the sample data belongs to the source domain or the target domain according to the input artificial time-frequency domain features and the deep features.

5. A psychological stress detection device based on transfer learning, characterized in that: include: Acquisition module, which acquires and preprocesses the user's multimodal physiological signals; A detection module, used for inputting the pre-processed multimodal physiological signal of the user into a stress detection migration model, and obtaining the stress level of the user output by the stress detection migration model; Among them, the stress detection migration model includes a feature extractor and a stress classifier. The stress detection migration model is obtained by adversarial training with a domain discriminator. The feature extractor is used to extract deep features and fuse them with artificial time-frequency domain features. The stress classifier is used to classify the user's stress level. The domain discriminator is used to distinguish whether the sample data belongs to a source domain or a target domain. The source domain and the target domain are two groups with different stress-inducing intensities.

6. The psychological stress detection device based on transfer learning according to claim 5, characterized in that: The feature extractor is also used to: The deep features are fused and dimensionally reduced and then cascaded with artificial time-frequency domain features, and the cascaded features are fused through a fully connected network.

7. The psychological stress detection device based on transfer learning according to claim 5, characterized in that: The preprocessing includes: motion segment removal, filtering and denoising, and outlier inspection and removal.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the psychological stress detection method based on transfer learning as described in any one of claims 1 to 4 are implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the psychological stress detection method based on transfer learning as claimed in any one of claims 1 to 4 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the psychological stress detection method based on transfer learning as claimed in any one of claims 1 to 4 are implemented.

Images