Psychological state evaluation parameter determination method and device, electronic equipment and storage medium

Abstract

This invention relates to the field of non-contact biosignal processing technology, providing a method, device, electronic device, and storage medium for determining psychological state assessment parameters. The method includes: inputting millimeter-wave radar signals into a psychological state assessment model to obtain psychological state assessment parameters; the psychological state assessment model includes a physiological feature extraction module, a micromotor feature extraction module, a multimodal feature fusion module, and a psychological state inference module. This invention reconstructs electrocardiogram waveforms through the physiological feature extraction module to obtain detailed physiological features, extracts micromotor features through the micromotor feature extraction module, and integrates features reflecting internal physiological changes and external behavioral manifestations through the multimodal feature fusion module. This forms a more comprehensive and multidimensional fused feature set for comprehensive analysis by the psychological state inference module, greatly improving the comprehensiveness and accuracy of psychological state assessment in unconstrained daily environments.

Description

Technical Field

[0001] This invention relates to the field of non-contact biosignal processing technology, and in particular to a method, apparatus, electronic device, and storage medium for determining psychological state assessment parameters. Background Technology

[0002] Mental health has become a major public health issue of global concern, and timely and accurate assessment of mental state is crucial for early warning, diagnosis, and intervention. Traditional methods of mental state assessment, such as subjective scales and clinical interviews, while valuable, have limitations including strong subjectivity, recall bias, and low assessment frequency. Therefore, non-contact, objective physiological indicator monitoring technologies have become an important development direction. Among these, millimeter-wave radar technology shows great potential in the field of physiological signal monitoring due to its advantages such as non-contact detection, penetration through clothing, immunity to light, and non-intrusion on visual privacy. Existing technologies have already utilized millimeter-wave radar to monitor physiological parameters such as heart rate and respiratory rate, and some studies have attempted to analyze heart rate variability (HRV) from these parameters.

[0003] However, existing millimeter-wave radar-based psychological state assessment technologies still have the following shortcomings when applied to unconstrained daily environments: First, current technologies lack precision in extracting physiological signals, mostly limiting themselves to extracting relatively macroscopic physiological parameters such as average heart rate and respiratory rate, resulting in insufficient physiological information dimensions for the assessment. Second, in existing technologies, these micromotor signals containing important psychological information are often ignored or filtered out as noise or interference, leading to the loss of effective information. Finally, existing assessment models typically analyze only single-modal physiological rhythm information, lacking a framework that can comprehensively analyze features reflecting internal physiological changes with micromotor features reflecting external behavioral performance, resulting in incomplete and inaccurate assessment results. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for determining parameters for psychological state assessment. It addresses the shortcomings of existing technologies that rely on insufficient precision in extracting physiological signals for psychological state assessment, and that generally treat micro-motion signals containing important information as noise and filter them out, resulting in a lack of effective information. Furthermore, existing solutions often rely on physiological information from a single modality for analysis and lack a comprehensive assessment framework that can synergistically utilize internal physiological and external behavioral characteristics, thus limiting the comprehensiveness and accuracy of the assessment.

[0005] This invention provides a method for determining parameters for psychological state assessment, comprising the following steps:

[0006] Acquire the millimeter-wave radar signal of the target object to be evaluated;

[0007] The millimeter-wave radar signal is input into the psychological state assessment model to obtain the psychological state assessment parameters output by the psychological state assessment model.

[0008] The psychological state assessment model includes a physiological feature extraction module, a micro-motor feature extraction module, a multimodal feature fusion module, and a psychological state inference module.

[0009] The physiological feature extraction module is used to reconstruct the electrocardiogram (ECG) waveform based on the millimeter-wave radar signal and extract physiological features from the ECG waveform; the micro-motion feature extraction module is used to extract micro-motion features from the millimeter-wave radar signal to obtain micro-motion features; the multimodal feature fusion module is used to fuse the physiological features and the micro-motion features to obtain fused features; and the psychological state inference module is used to obtain the psychological state assessment parameters based on the fused features.

[0010] According to a method for determining psychological state assessment parameters provided by the present invention, the training steps of the physiological feature extraction module include:

[0011] Acquire millimeter-wave radar sample signals and reference electrocardiogram activity waveforms acquired synchronously with the millimeter-wave radar sample signals;

[0012] The millimeter-wave radar sample signal is input into the initial physiological feature extraction module to obtain the electrocardiogram activity waveform to be reconstructed output by the initial physiological feature extraction module;

[0013] Based on the clinical diagnostic importance of different waveform segments in the reference ECG waveform, the diagnostic weights corresponding to each waveform segment are determined, and the weighted morphological differences between the ECG waveform to be reconstructed and the reference ECG waveform are calculated based on the diagnostic weights to determine the morphological fidelity loss.

[0014] Based on the morphological fidelity loss, a target loss is determined, and the parameters of the initial physiological feature extraction module are iterated based on the target loss to obtain the physiological feature extraction module.

[0015] According to a method for determining psychological state assessment parameters provided by the present invention, determining the target loss based on the morphological fidelity loss includes:

[0016] The first heart rate variability parameter and the second heart rate variability parameter are extracted from the electrocardiogram waveform to be reconstructed and the reference electrocardiogram waveform, respectively.

[0017] Based on the first heart rate variability parameter and the second heart rate variability parameter, the consistency loss is determined;

[0018] The ECG waveform to be reconstructed and the reference ECG waveform are input into a generative adversarial network (GAN). The discriminator in the GAN determines a first discriminant output for the reference ECG waveform and a second discriminant output for the ECG waveform to be reconstructed. The GAN includes a generator and a discriminator, and the generator is the initial physiological feature extraction module.

[0019] Based on the first discrimination output and the second discrimination output, determine the discriminator loss of the discriminator;

[0020] The generator loss of the generator is determined based on the negative number of the second discrimination output;

[0021] Based on the discriminator loss and the generator loss, the adversarial loss is determined;

[0022] The target loss is determined based on the morphological fidelity loss, the consistency loss, and the adversarial loss.

[0023] According to a method for determining psychological state assessment parameters provided by the present invention, determining the target loss based on the morphological fidelity loss, the consistency loss, and the antagonistic loss includes:

[0024] The ECG waveform to be reconstructed and the millimeter-wave radar sample signal are mapped to a shared feature space to obtain the features of the ECG waveform to be reconstructed and the features of the millimeter-wave radar sample signal. Based on the feature distance between the ECG waveform features to be reconstructed and the millimeter-wave radar sample signal features in the shared feature space, the content loss is determined.

[0025] The target loss is determined based on the morphological fidelity loss, the consistency loss, the adversarial loss, and the content loss.

[0026] According to a method for determining psychological state assessment parameters provided by the present invention, the micro-motion feature extraction module is specifically used for:

[0027] Physiological rhythm filtering is performed on the millimeter-wave radar signal to obtain micro-motion residual signals;

[0028] The residual signal of micro-motion is input into a temporal convolutional network to obtain the initial temporal features output by the temporal convolutional network;

[0029] Temporal attention weights and / or channel attention weights are applied to the initial temporal features to obtain enhanced temporal features;

[0030] Based on the enhanced temporal features, the micro-motion features are determined.

[0031] According to a method for determining psychological state assessment parameters provided by the present invention, the step of acquiring the millimeter-wave radar signal of the target object to be assessed includes:

[0032] Acquire the multi-channel millimeter-wave radar signal of the target object to be evaluated;

[0033] The normalized reflected energy, circular fitting error of the in-phase quadrature component plot, periodic score of the heartbeat signal, harmonic interference ratio of the respiratory signal, and phase standard deviation of each channel signal in the multi-channel millimeter-wave radar signal are determined respectively.

[0034] The comprehensive signal quality index is determined based on at least two of the following: the normalized reflection energy, the circular fitting error of the in-phase orthogonal component plot, the periodic score of the heartbeat signal, the harmonic interference ratio of the respiratory signal, and the phase standard deviation.

[0035] Based on the comprehensive signal quality index, the target channel signal is selected from the multi-channel millimeter-wave radar signals as the millimeter-wave radar signal.

[0036] According to a method for determining psychological state assessment parameters provided by the present invention, the step of selecting a target channel signal as the millimeter-wave radar signal from the multi-channel millimeter-wave radar signal based on the comprehensive signal quality index includes:

[0037] Based on the comprehensive signal quality index, the multi-channel millimeter-wave radar signal is spatially filtered to obtain a spatially filtered signal.

[0038] The spatially filtered signal is subjected to source signal separation processing to obtain the separated radar signal;

[0039] The separated radar signal is subjected to phase unwrapping processing to obtain a continuous phase signal; wherein, the unwrapping path of the phase unwrapping processing is guided by the phase quality map of the separated radar signal;

[0040] The continuous phase signal is subjected to multi-resolution separation to separate the target channel signal related to cardiac activity, and the target channel signal is used as the millimeter-wave radar signal.

[0041] The present invention also provides a device for determining psychological state assessment parameters, comprising the following units:

[0042] The acquisition unit is used to acquire the millimeter-wave radar signal of the target object to be evaluated.

[0043] A psychological state assessment unit is used to input the millimeter-wave radar signal into a psychological state assessment model to obtain psychological state assessment parameters output by the psychological state assessment model.

[0044] The psychological state assessment model includes a physiological feature extraction module, a micro-motor feature extraction module, a multimodal feature fusion module, and a psychological state inference module.

[0045] The physiological feature extraction module is used to reconstruct the electrocardiogram (ECG) waveform based on the millimeter-wave radar signal and extract physiological features from the ECG waveform; the micro-motion feature extraction module is used to extract micro-motion features from the millimeter-wave radar signal to obtain micro-motion features; the multimodal feature fusion module is used to fuse the physiological features and the micro-motion features to obtain fused features; and the psychological state inference module is used to obtain the psychological state assessment parameters based on the fused features.

[0046] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for determining psychological state assessment parameters as described above.

[0047] 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 method for determining psychological state assessment parameters as described above.

[0048] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the method for determining psychological state assessment parameters as described above.

[0049] The present invention provides a method, apparatus, electronic device, and storage medium for determining psychological state assessment parameters. The method involves acquiring millimeter-wave radar signals from the target object to be assessed; inputting the millimeter-wave radar signals into a psychological state assessment model to obtain psychological state assessment parameters; wherein the psychological state assessment model includes a physiological feature extraction module for reconstructing electrocardiogram waveforms and extracting physiological features based on millimeter-wave radar signals, a micro-motion feature extraction module for extracting micro-motion features from millimeter-wave radar signals, a multi-modal feature fusion module for fusing physiological features and micro-motion features to obtain fused features, and a psychological state inference module for obtaining psychological state assessment parameters based on the fused features. This invention reconstructs electrocardiogram waveforms through a physiological feature extraction module to obtain detailed physiological features, while simultaneously extracting micromotor features through a micromotor feature extraction module. Furthermore, a multimodal feature fusion module integrates features reflecting internal physiological changes and external behavioral manifestations, forming a more comprehensive and multidimensional fused feature for the psychological state inference module to perform comprehensive analysis. This solves the problems of incomplete and inaccurate assessment results caused by insufficient precision in physiological information extraction, neglect of micromotor information, and single assessment modality in existing technologies, greatly improving the comprehensiveness and accuracy of psychological state assessment in unconstrained daily environments. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0051] Figure 1 This is one of the flowcharts illustrating the method for determining psychological state assessment parameters provided by this invention.

[0052] Figure 2 This is a schematic diagram of multiple losses in determining the target loss provided by the present invention.

[0053] Figure 3 This is a schematic diagram of the micro-motion feature extraction module provided by the present invention.

[0054] Figure 4 This is a schematic diagram of the selection of millimeter-wave radar signals provided by the present invention.

[0055] Figure 5 This is the second flowchart of the method for determining psychological state assessment parameters provided by the present invention.

[0056] Figure 6 This is a schematic diagram of the structure of the device for determining psychological state assessment parameters provided by the present invention.

[0057] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0059] The terms "first," "second," etc., used in this invention are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and that the objects distinguished by "first," "second," etc., are generally of the same class.

[0060] This invention provides a method for determining parameters for psychological state assessment. This method aims to solve the problems of strong subjectivity, susceptibility to motion artifacts, neglect of micro-motion information, and single assessment dimension in existing psychological state assessment methods. Thus, it provides a technical solution that can comprehensively and accurately assess an individual's psychological state in a non-contact setting. Figure 1 This is one of the flowcharts illustrating the method for determining psychological state assessment parameters provided by this invention, such as... Figure 1 As shown, the method includes steps 110 and 120.

[0061] Step 110: Obtain the millimeter-wave radar signal of the target object to be evaluated.

[0062] Specifically, firstly, millimeter-wave radar signals from the target object can be acquired. The target object can be any living organism requiring psychological state assessment, typically referring to a person in a natural state. For example, an office worker, a driver resting, or an individual in a home environment. The non-contact nature of this method allows the assessment process to be conducted without disturbing the target object's normal activities or infringing on their visual privacy.

[0063] Here, acquiring the millimeter-wave radar signal of the target object to be evaluated can be achieved through a millimeter-wave radar device. The millimeter-wave radar device is directed towards the target object, for example, aimed at the upper part of the target object's torso, and emits a millimeter-wave beam, receiving the echo signal reflected from the target object's body surface. Because physiological activities such as breathing and heartbeat cause minute displacements (typically at the sub-millimeter to millimeter level) in the chest and abdomen, these minute displacements modulate the phase and amplitude of the reflected echo. Simultaneously, unconscious micro-movements caused by changes in an individual's psychological state, such as tension, anxiety, irritability, or cognitive activity—e.g., slight swaying of the body, muscle tremors, subtle adjustments of posture—are also captured by the radar signal.

[0064] Therefore, millimeter-wave radar signals are time-series signals used to reflect minute displacement information of the body surface of an object being evaluated over time. Millimeter-wave radar signals are essentially composite signals, carrying information caused by macroscopic physiological rhythms such as heartbeat and respiration, as well as more subtle, usually non-periodic, motion information caused by micro-movements of the body. In specific implementations, millimeter-wave radar signals can be digital intermediate frequency signals or I / Q (In-phase / Quadrature) signals obtained from the millimeter-wave radar front-end after mixing, filtering, and analog-to-digital conversion.

[0065] Here, the acquisition of the millimeter-wave radar signal of the target object to be evaluated can be done through single-channel acquisition or through multi-channel acquisition using a multiple-input multiple-output (MIMO) antenna array to obtain richer spatial information. This embodiment of the invention does not specifically limit this.

[0066] Step 120: Input the millimeter-wave radar signal into the psychological state assessment model to obtain the psychological state assessment parameters output by the psychological state assessment model;

[0067] The psychological state assessment model includes a physiological feature extraction module, a micro-motor feature extraction module, a multimodal feature fusion module, and a psychological state inference module.

[0068] The physiological feature extraction module is used to reconstruct the electrocardiogram (ECG) waveform based on the millimeter-wave radar signal and extract physiological features from the ECG waveform; the micro-motion feature extraction module is used to extract micro-motion features from the millimeter-wave radar signal to obtain micro-motion features; the multimodal feature fusion module is used to fuse the physiological features and the micro-motion features to obtain fused features; and the psychological state inference module is used to obtain the psychological state assessment parameters based on the fused features.

[0069] Specifically, after obtaining the millimeter-wave radar signal, the millimeter-wave radar signal can be input into the psychological state assessment model to obtain the psychological state assessment parameters output by the psychological state assessment model.

[0070] Here, psychological state assessment parameters are used to reflect the mental health or emotional state of the target subject at the assessment time point. Psychological state assessment parameters can be discrete classification labels, such as calm, mild anxiety, moderate tension, high cognitive load, etc., or they can be one or more continuous numerical scores, such as stress index (0-100 points), anxiety score (0-10 points), etc., thereby achieving a quantitative assessment of psychological state. This embodiment of the invention does not specifically limit these parameters.

[0071] To achieve the conversion from millimeter-wave radar signals to the final psychological assessment results, the psychological state assessment model can include a physiological feature extraction module, a micro-motion feature extraction module, a multimodal feature fusion module, and a psychological state inference module.

[0072] The physiological feature extraction module is used to reconstruct electrocardiogram (ECG) waveforms based on millimeter-wave radar signals and extract physiological features from these waveforms. The primary task of this module is to separate and reconstruct signal components strongly correlated with cardiac activity from the millimeter-wave radar signals, which carry complex motion information.

[0073] Electrocardiographic activity (ECG) waveforms refer to a time series reconstructed from millimeter-wave radar signals that highly simulates the waveform morphology and rhythm of a real ECG signal. ECG waveforms reflect the fine patterns of mechanical vibrations of the chest wall caused by each heartbeat, and these patterns closely correspond to the heart's electrical activity processes, such as atrial depolarization, ventricular depolarization, and repolarization. The reconstruction of ECG waveforms can be achieved using advanced signal processing algorithms or deep learning-based generative models, with the goal of generating a waveform that includes morphological details such as P waves, QRS complexes, and T waves.

[0074] Here, physiological characteristics are quantitative features used to reflect cardiac function and the activity state of the autonomic nervous system. Physiological characteristics may include basic parameters such as heart rate (HR) and respiratory rate; more specifically, they may include a series of refined heart rate variability parameters, obtained through statistical analysis of successive heartbeat interval time series, which are recognized as powerful indicators for assessing psychological stress, anxiety, and autonomic nervous system balance. For example, physiological characteristics may include time-domain parameters such as SDNN (Standard Deviation of Normal-to-Normal RR Intervals) and RMSSD (Root Mean Square of Successive Differences between Normal RR Intervals), and frequency-domain parameters such as LF (Low Frequency Power), HF (High Frequency Power), and LF / HF (Ratio of Low Frequency to High Frequency Power). This embodiment of the invention does not specifically limit these parameters. In addition, parameters extracted from the morphology of the reconstructed electrocardiogram waveforms, such as T-wave amplitude and QT-interval, can also be used as part of the physiological characteristics.

[0075] Here, the MicroMotionNet module is used to extract micromotion features from millimeter-wave radar signals, obtaining micromotion characteristics. Unlike the physiological feature extraction module, which focuses on the periodic and rhythmic physiological components in the signal, the micromotion feature extraction module focuses on the non-physiological rhythmic and weaker body motion information in the signal.

[0076] One approach to extracting micro-motion features from millimeter-wave radar signals can be to first preprocess the signals, for example, by using filter banks or adaptive filtering algorithms to suppress or filter out the main components of the signal caused by heartbeat and respiration, thus obtaining residual signals that primarily reflect body micro-motion. Then, feature engineering is performed on this residual signal to obtain the micro-motion features.

[0077] Here, micromotor features are characteristics used to quantify and describe subtle body movement patterns. Since micromotor activity is highly correlated with an individual's psychological state, such as anxiety, tension, and restlessness, micromotor features can correspondingly include the energy or amplitude of the micromotor, its dominant frequency, occurrence frequency, duration, or entropy values ​​describing the complexity or irregularity of its movement patterns. This embodiment of the invention does not specifically limit these features.

[0078] In more complex implementations, the micromotor feature extraction module can be a classifier used to identify specific micromotor patterns, such as anxiety tremor, restlessness, etc., and output the classification results or confidence scores of these patterns as micromotor features.

[0079] Here, the multimodal feature fusion module is used to fuse physiological features and micromotor features to obtain fused features. The multimodal feature fusion module receives outputs from the two modules mentioned above: physiological features representing internal physiological states and micromotor features representing external behavioral performance. The purpose of fusion is to effectively integrate information from these two different modalities, both related to psychological states, to form a more comprehensive and robust feature representation than any single modality.

[0080] Here, the fusion of physiological features and micromotor features can be achieved by splicing the physiological features and micromotor features together, or by using an attention mechanism to weight the physiological features and micromotor features before splicing them together. This embodiment of the invention does not specifically limit this method.

[0081] Here, the mental state inference module is used to obtain mental state assessment parameters based on fusion features.

[0082] The mental state inference module can be a machine learning model trained under supervised learning. If the mental state evaluation parameters are discrete categories, the module can be a classifier, such as a Support Vector Machine (SVM) or a Deep Neural Network (DNN) classifier. If the mental state evaluation parameters are continuous values, the module can be a regression model to learn a complex nonlinear mapping from fused features to a specific mental state.

[0083] There are two training schemes for the mental state inference module. Scheme 1 is to fine-tune the pre-trained physiological signal base model. First, a model that has been pre-trained on a large-scale multimodal physiological signal dataset and performed well is selected, and then fine-tuned on the dataset annotated by psychiatrists, which is unique to this invention. Full model fine-tuning or parameter effectiveness fine-tuning methods such as LoRA (Low-Rank Adaptation) can be used. Scheme 2 is to design a dedicated Transformer-based classification / regression network if a suitable pre-trained model is not available. Its training depends on the aforementioned annotated dataset. At the same time, in terms of the reintegration of psychiatric professional knowledge, a post-processing calibration layer can be designed to calibrate the original output of the model. It is also possible to integrate technologies such as SHAP (SHapley Additive Explanations) or LIME (Local Interpretable Model-agnostic Explanations) to provide an interpretability interface. Finally, the system provides a user-friendly graphical interface to display the assessment results, including a comprehensive assessment of the current psychological state, visualization of key physiological indicators and micromotor activity levels, and a historical trend chart of psychological state changes. With user authorization, it can also provide detailed professional reports and raw feature data summaries to professionals, and has user privacy settings and data management functions.

[0084] The method provided in this invention acquires millimeter-wave radar signals of the target object to be evaluated; inputs the millimeter-wave radar signals into a psychological state assessment model to obtain psychological state assessment parameters; wherein, the psychological state assessment model includes a physiological feature extraction module for reconstructing electrocardiogram waveforms and extracting physiological features based on millimeter-wave radar signals, a micro-motion feature extraction module for extracting micro-motion features from millimeter-wave radar signals, a multi-modal feature fusion module for fusing physiological features and micro-motion features to obtain fused features, and a psychological state inference module for obtaining psychological state assessment parameters based on the fused features. This invention reconstructs electrocardiogram waveforms through the physiological feature extraction module to obtain refined physiological features, extracts micro-motion features through the micro-motion feature extraction module, and then integrates features reflecting internal physiological changes and external behavioral manifestations through the multi-modal feature fusion module, thereby forming a more comprehensive and dimensional fused feature for comprehensive analysis by the psychological state inference module. This solves the problems of incomplete and inaccurate assessment results caused by insufficient refinement of physiological information extraction, neglect of micro-motion information, and single assessment modality in existing technologies, greatly improving the comprehensiveness and accuracy of psychological state assessment in unconstrained daily environments.

[0085] Based on the above embodiments, the training steps of the physiological feature extraction module include:

[0086] Step 21: Acquire millimeter-wave radar sample signals and reference electrocardiogram activity waveforms acquired synchronously with the millimeter-wave radar sample signals;

[0087] Step 22: Input the millimeter-wave radar sample signal into the initial physiological feature extraction module to obtain the electrocardiogram activity waveform to be reconstructed output by the initial physiological feature extraction module;

[0088] Step 23: Determine the diagnostic weight of each waveform segment according to the clinical diagnostic importance of different waveform segments in the reference ECG waveform, and calculate the weighted morphological difference between the ECG waveform to be reconstructed and the reference ECG waveform based on the diagnostic weight to determine the morphological fidelity loss.

[0089] Step 24: Based on the morphological fidelity loss, determine the target loss, and perform parameter iteration on the initial physiological feature extraction module based on the target loss to obtain the physiological feature extraction module.

[0090] Specifically, firstly, millimeter-wave radar sample signals and reference electrocardiogram waveforms acquired synchronously with the millimeter-wave radar sample signals can be obtained. The millimeter-wave radar sample signals consist of numerous radar signal fragments acquired using millimeter-wave radar equipment under various scenarios. These scenarios can include different environments, different target objects to be evaluated, and different psychological state induction conditions.

[0091] Here, the reference ECG waveform serves as a training standard signal. This reference ECG waveform can be acquired synchronously with radar signal acquisition using medical-grade ECG equipment, such as a multi-lead ECG machine. It should be understood that the reference ECG waveform is clear, accurate, and can be annotated and screened by a professional psychiatrist to ensure it accurately reflects the electrophysiological manifestations under specific psychological states.

[0092] Then, the millimeter-wave radar sample signal can be input into the initial physiological feature extraction module to obtain the electrocardiogram waveform to be reconstructed output by the initial physiological feature extraction module. The initial physiological feature extraction module can be a deep learning network structure with preset parameters or randomly initialized parameters, such as the generator in a Generative Adversarial Network (GAN) consisting of an encoder and a decoder.

[0093] Furthermore, the diagnostic weights of each waveform segment can be determined based on the clinical diagnostic importance of different waveform segments in the reference ECG waveform, and the weighted morphological differences between the ECG waveform to be reconstructed and the reference ECG waveform can be calculated based on the diagnostic weights to determine the morphological fidelity loss.

[0094] Here, diagnostic weights are assigned numerical values ​​to each time point or waveform segment on the electrocardiogram based on the aforementioned clinical diagnostic importance. Waveform segments with greater diagnostic significance are assigned higher diagnostic weights; conversely, waveform segments with less diagnostic significance are assigned lower weights.

[0095] Here, the weighted morphological difference can be calculated by multiplying the absolute value (L1 distance) or square value (L2 distance) of the difference between the ECG waveform to be reconstructed and the reference ECG waveform at each time point by the diagnostic weight corresponding to that time point, and then summing or averaging over the entire waveform length.

[0096] Here, the formula for morphological fidelity loss is as follows:

[0097] ;

[0098] in, Indicates morphological fidelity loss. This indicates the waveform of the electrocardiogram to be reconstructed. Indicates the reference electrocardiogram waveform. This indicates the diagnostic weight corresponding to each waveform segment.

[0099] Understandably, the greater the morphological difference between the waveform of the ECG activity to be reconstructed and the reference ECG activity waveform, the greater the loss of morphological fidelity; conversely, the smaller the morphological difference between the waveform of the ECG activity to be reconstructed and the reference ECG activity waveform, the smaller the loss of morphological fidelity.

[0100] Clinical diagnostic importance refers to the degree of significance of different components of the electrocardiogram waveform, such as the P wave, QRS complex, T wave, ST segment, and QT interval, in diagnosing specific psychological states, such as anxiety, depression, and stress. For example, psychiatrists point out that in an anxious state, changes in the morphology of the T wave (such as flattening or inversion) and changes in the QT interval are more diagnostically valuable than the morphology of the QRS complex; therefore, the corresponding waveform segments have different diagnostic weights. A higher value will be assigned. The morphology of the QRS complex is crucial for the accurate characterization of ventricular depolarization and is therefore given a high weight. This knowledge is incorporated to ensure that the reconstructed ECG is not only mathematically similar but also clinically meaningful.

[0101] Finally, the target loss can be determined based on the morphological fidelity loss, and the parameters of the initial physiological feature extraction module can be iterated based on the target loss. The initial physiological feature extraction module after parameter iteration is then used as the physiological feature extraction module.

[0102] The method provided in this invention introduces diagnostic weights and morphological fidelity loss corresponding to each waveform segment guided by psychiatric expertise. This allows the physiological feature extraction module to focus on waveform features that are most clinically significant for psychological state assessment during training. This ensures that the reconstructed electrocardiogram waveform not only closely approximates the real signal mathematically but also maintains a high degree of consistency in clinical interpretation, greatly improving the effectiveness of subsequent physiological feature extraction and the accuracy of the final psychological state assessment parameters.

[0103] Based on the above embodiments, step 24, determining the target loss based on the morphological fidelity loss, includes:

[0104] Step 241: Extract the first heart rate variability parameter and the second heart rate variability parameter from the ECG activity waveform to be reconstructed and the reference ECG activity waveform, respectively.

[0105] Step 242: Determine the consistency loss based on the first heart rate variability parameter and the second heart rate variability parameter;

[0106] Step 243: The ECG waveform to be reconstructed and the reference ECG waveform are input into a generative adversarial network (GAN). The discriminator in the GAN determines the first discriminant output of the reference ECG waveform and the second discriminant output of the ECG waveform to be reconstructed. The GAN includes a generator and a discriminator, and the generator is the initial physiological feature extraction module.

[0107] Step 244: Based on the first discrimination output and the second discrimination output, determine the discriminator loss of the discriminator;

[0108] Step 245: Determine the generator loss of the generator based on the negative number of the second discrimination output;

[0109] Step 246: Determine the adversarial loss based on the discriminator loss and the generator loss;

[0110] Step 247: Determine the target loss based on the morphological fidelity loss, the consistency loss, and the adversarial loss.

[0111] Specifically, firstly, the first heart rate variability parameter and the second heart rate variability parameter can be extracted from the ECG waveform to be reconstructed and the reference ECG waveform, respectively.

[0112] The first and second heart rate variability parameters refer to a set of statistical indicators used to describe heart rate variability, such as RMSSD, SDNN, pNN50 (Percentage of Successive RR Intervals Differentying by More Than 50 Milliseconds), LF / HF ratio, etc. The embodiments of the present invention do not specifically limit these parameters.

[0113] The extraction method involves first detecting the R-peak position on the ECG waveform to be reconstructed and the reference ECG waveform, calculating the RR interval sequence, and then calculating the first and second heart rate variability parameters based on the RR interval sequence.

[0114] In one specific embodiment, firstly, a robust R-peak detection algorithm (e.g., an improved algorithm based on the Pan-Tompkins algorithm combined with wavelet transform denoising and adaptive thresholding) is applied to accurately locate the R-peak on the electrocardiogram waveform. Then, based on the detected RR interval sequence, a complete set of heart rate variability parameters is calculated strictly according to international standards (such as the Task Force standard). These heart rate variability parameters constitute the first / second heart rate variability parameters, which are a comprehensive set of indicators that fully reflect autonomic nervous function. Specifically, they may include: time-domain indicators (such as SDNN, RMSSD, NN50, pNN50, etc.), frequency-domain indicators (such as the absolute power of VLF, LF, HF, and the LF / HF ratio, etc.), and nonlinear HRV indicators, such as Poincaré plot parameters SD1, SD2, approximate entropy (ApEn), sample entropy (SampEn), etc.

[0115] Then, the consistency loss can be determined based on the first and second heart rate variability parameters, as shown in the following formula:

[0116] ;

[0117] in, Indicates consistency loss. This represents the first heart rate variability parameter. This represents the second heart rate variability parameter.

[0118] Understandably, the smaller the difference between the first and second heart rate variability parameters, the closer the reconstructed waveform reflects the autonomic nervous system regulation information to the actual situation.

[0119] It should be understood that consistency loss is used to ensure that the deep physiological rhythm information contained in the reconstructed waveform is consistent with the actual situation.

[0120] In this invention, the specific network responsible for the physiological feature extraction module is called Adv-ECG-RecNet (Advanced ECG Reconstruction Network), which is a Conditional Generative Adversarial Network (CGAN) designed specifically for high-fidelity reconstruction of ECG waveforms and extraction of fine cardiac activity features. The training process of this network is achieved by optimizing the composite target loss defined in this embodiment.

[0121] Then, the ECG waveform to be reconstructed and the reference ECG waveform can be input into a generative adversarial network (GAN). The discriminator in the GAN determines the first discriminant output of the reference ECG waveform and the second discriminant output of the ECG waveform to be reconstructed. The GAN consists of a generator and a discriminator, with the generator being the initial physiological feature extraction module.

[0122] The generator's input is a preprocessed millimeter-wave radar signal segment primarily containing information about cardiac mechanical vibrations. For example, it could be a 5-second time series with 2500 data points sampled at 500Hz. The generator comprises an encoder and a decoder, with the encoder employing an innovative 1D ResNet-Transformer hybrid architecture. The signal is first passed through multiple 1D residual convolutional blocks (ResNet blocks) to effectively extract local temporal features and patterns across different receptive fields of the cardiac mechanical motion signal. These local feature sequences (typically downsampled) are then fed into a multi-layer Transformer encoder. Leveraging the Transformer's powerful self-attention mechanism, the model captures long-range dependencies and global contextual information crucial for understanding the complete cardiac cycle.

[0123] The decoder part of the generator uses a series of 1D transposed convolutional layers and combines them with skip connections to introduce low-level features from different levels in the encoder, gradually upsampling the abstract latent representation of the encoder output, and finally reconstructing a high-resolution, high-fidelity electrocardiogram waveform corresponding to the input radar segment.

[0124] Here, the discriminator's input includes two types of data: one is the ECG waveform segment to be reconstructed generated by generator G, and the other is a real reference ECG waveform segment. To ensure training quality, the reference ECG waveform segment comes from a large-scale, high-quality synchronously acquired dataset. This dataset not only includes millimeter-wave radar sample signals, but also standard medical-grade multi-lead ECG signals, as well as detailed psychological state assessment data of individuals conducted by psychiatrists using standardized assessment tools. More importantly, psychiatrists are deeply involved in data screening and annotation, selecting ECG segments with typical psychophysiological characteristics (such as ECG manifestations under specific stress conditions) as high-quality training samples.

[0125] Here, standardized assessment tools may include at least one of MINI (Mini-International Neuropsychiatric Interview), SCID (Structured Clinical Interview for DSM Disorders), HAM-A (Hamilton Anxiety Rating Scale), and PSS (Perceived Stress Scale), and this embodiment of the invention does not specifically limit them.

[0126] Here, the discriminator's network structure can adopt the PatchGAN architecture. Unlike traditional discriminators that output a single scalar ("true" or "false"), PatchGAN segments the input ECG waveform into multiple overlapping local segments and independently judges the authenticity of each segment, ultimately outputting a discrimination matrix. This mechanism encourages the network to pay more attention to the local details and high-frequency components of the ECG waveform, thereby significantly improving the overall realism and detail fidelity of the reconstructed waveform.

[0127] Here, the discriminator outputs a first discriminant output (expected to be close to 1) for the reference ECG waveform and a second discriminant output (expected to be close to 0) for the ECG waveform to be reconstructed. The discriminator loss can be determined based on the first and second discriminant outputs; specifically, it can be determined based on the difference between the first and second discriminant outputs and their respective expected targets (1 and 0). This discriminator loss is used to train the discriminator, making its discriminative ability increasingly stronger.

[0128] Furthermore, the generator loss of the generator can be determined based on the negative of the second discriminant output. This generator loss is used to train the generator, i.e., the physiological feature extraction module, with the goal of making its generated waveforms able to deceive the discriminator, i.e., making the second discriminant output as close to 1 as possible.

[0129] Finally, the adversarial loss can be determined based on the discriminator loss and the generator loss, as shown in the following formula:

[0130] ;

[0131] ;

[0132] ;

[0133] in, Indicates adversarial loss. Indicates discriminator loss. Indicates generator loss. This indicates the output of the second discrimination. Represents the gradient penalty coefficient. This indicates the distribution of generated data produced by the generator. This represents the average score (expected value) of the real data in the discriminator. Specifically, the symbol... The reference ECG activity waveform is derived from the distribution of real data. The real samples obtained from sampling This is the score given by the discriminator to the real sample. (Symbol) This represents the mathematical expectation, which is the average score calculated over a large number of such real samples. During the training of a generative adversarial network (GAN), the discriminator aims to maximize this value. It needs to learn to assign high scores to real data, thereby effectively distinguishing it from the fake data generated by the generator. Indicates the distribution of real data With the distribution of generated data Random uniform interpolation along the connection line, Indicates the distribution from All ECG activity waveforms to be reconstructed from the mid-sampled data Calculate the mathematical expectation.

[0134] in, The expression represents the gradient (i.e., rate of change) of the discriminator function with respect to its input samples. Where, the symbols... Used to calculate the derivative vector The term refers to the ECG waveform to be reconstructed, which is an interpolated sample (i.e., points obtained by random interpolation between real and generated samples). It is about reconstructing the waveform of cardiac activity. Differentiating, This is the output of the discriminator to the reconstructed ECG waveform. This term is mainly used in the gradient penalty mechanism, and its core purpose is to constrain the norm of the gradient. The value is close to 1. By restricting drastic changes in gradients, this term forces the discriminator to satisfy the Lipschitz continuity constraint, thereby effectively preventing gradient vanishing or exploding problems during training and ensuring the stability of model learning. Finally, the target loss is jointly determined based on morphological fidelity loss, consistency loss, and adversarial loss.

[0135] The method provided in this invention, compared to using only morphological fidelity loss, constructs a comprehensive target loss function with multiple tasks and perspectives by introducing consistency loss and adversarial loss. Adversarial loss ensures the overall realism and high-frequency details of the reconstructed waveform, consistency loss guarantees the accuracy of core physiological rhythm information, and morphological fidelity loss injects clinical expertise, ensuring the fidelity of key diagnostic features. The combination of these three factors guides the physiological feature extraction module to generate electrocardiogram waveforms that highly approximate real electrocardiographic activity in terms of morphology, rhythm, and clinical significance.

[0136] Based on the above embodiments, step 247 includes:

[0137] Step 2471: Map the ECG activity waveform to be reconstructed and the millimeter-wave radar sample signal to a shared feature space to obtain the ECG activity waveform features to be reconstructed and the millimeter-wave radar sample signal features, and determine the content loss based on the feature distance between the ECG activity waveform features to be reconstructed and the millimeter-wave radar sample signal features in the shared feature space.

[0138] Step 2472: Determine the target loss based on the morphological fidelity loss, the consistency loss, the adversarial loss, and the content loss.

[0139] Specifically, Figure 2 This is a schematic diagram of multiple losses in determining the target loss provided by the present invention, such as... Figure 2 As shown, firstly, the ECG activity waveform to be reconstructed and the millimeter-wave radar sample signal are mapped to a shared feature space to obtain the features of the ECG activity waveform to be reconstructed and the features of the millimeter-wave radar sample signal. Then, based on the feature distance between the ECG activity waveform features to be reconstructed and the millimeter-wave radar sample signal features in the shared feature space, the content loss is determined.

[0140] The shared feature space is a high-dimensional, abstract representation space. Mapping signals of different modalities (radar signals and electrocardiogram waveforms) to this space allows for comparison of their inherent information content on a unified dimension. This mapping process can be implemented using a fixed feature extractor, such as a pre-trained neural network or mathematical transformations like wavelet transform.

[0141] The feature distance can be Euclidean distance, cosine distance, etc., used to measure the similarity between two feature vectors. This embodiment of the invention does not impose specific limitations on this.

[0142] Here, the formula for content loss is as follows:

[0143]

[0144] in, Indicates content loss. This indicates the waveform characteristics of the electrocardiogram activity to be reconstructed. This represents the characteristics of millimeter-wave radar sample signals. and This represents the feature extraction function.

[0145] It is understandable that the greater the feature distance between the waveform features of the ECG activity to be reconstructed and the features of the millimeter-wave radar sample signal, the greater the content loss; conversely, the smaller the feature distance between the waveform features of the ECG activity to be reconstructed and the features of the millimeter-wave radar sample signal, the smaller the content loss.

[0146] Finally, the target loss can be determined by combining morphological fidelity loss, consistency loss, adversarial loss, and content loss. The formula for the target loss is as follows:

[0147]

[0148] in, Indicates target loss. Indicates adversarial loss. Indicates consistency loss. Indicates morphological fidelity loss. Indicates content loss. The weights representing adversarial losses The weights representing the loss of morphological fidelity The weights representing the consistency loss The weight representing the content loss.

[0149] Here, The optimal value is determined through experimentation and validation (in conjunction with an assessment by a psychiatrist).

[0150] Based on the above embodiments, the micro-motion feature extraction module is specifically used for:

[0151] Step 310: Filter the millimeter-wave radar signal by physiological rhythm to obtain residual micro-motion signals;

[0152] Step 320: Input the micro-motion residual signal into a temporal convolutional network to obtain the initial temporal features output by the temporal convolutional network;

[0153] Step 330: Apply temporal attention weights and / or channel attention weights to the initial temporal features to obtain enhanced temporal features;

[0154] Step 340: Determine the micro-motion features based on the enhanced temporal features.

[0155] Specifically, Figure 3 This is a schematic diagram of the micro-motion feature extraction module provided by the present invention, as shown below. Figure 3 As shown, the micro-motion feature extraction module is specifically used for:

[0156] First, the millimeter-wave radar signal is filtered for physiological rhythms to obtain residual micromotor signals. Since the signal components caused by heartbeat and respiration in millimeter-wave radar signals are typically high-energy and periodic, they can mask weaker micromotor information. Therefore, the first step is to remove these strong physiological rhythm signals. Methods for physiological rhythm filtering can include using notch filters to filter out the dominant frequencies and harmonics of heartbeat and respiration, or employing more advanced signal separation techniques, such as Independent Component Analysis (ICA) or Singular Value Decomposition (SVD), to separate and remove the physiological rhythm components. The resulting residual micromotor signal is a time series primarily containing information about body micromotor functions.

[0157] Then, the residual micro-motion signals are input into a Temporal Convolutional Network (TCN) to obtain the initial temporal features output by the TCN. The TCN is a deep learning network particularly well-suited for processing time-series data. By employing causal convolution, dilated convolution, and residual connections, the TCN can efficiently learn long-term dependencies in a sequence, making it ideal for capturing complex temporal dynamics such as micro-motion patterns that may span several seconds.

[0158] Furthermore, the initial temporal features can be input into the Channel-Temporal Attention Module (CTAM), which applies temporal attention weights and / or channel attention weights to the initial temporal features to obtain enhanced temporal features. This is to allow the micro-motion feature extraction module to adaptively focus on the most informative parts. Specifically, the temporal attention weights are used to identify which time points or time periods in the micro-motion sequence are most critical for judging the current psychological state (e.g., the start of a sudden tremor or a slight postural adjustment), and assign higher weights to the features of these critical moments.

[0159] Here, channel attention weights are used to select the channels most relevant to the target task from the multiple feature channels output by the temporal convolutional network (each channel may represent an abstract motion pattern) and assign them higher weights. The temporal attention weights and channel attention weights can work in parallel to jointly weight the initial temporal features, thereby obtaining enhanced temporal features that highlight key points and concentrate information.

[0160] Finally, based on the enhanced temporal features, micromotor features are determined. This is typically achieved through one or more fully connected layers, mapping the enhanced temporal features to the final output space. A classification layer follows one or more fully connected layers, and the resulting micromotor features can be classification probabilities for specific micromotor patterns, such as anxiety tremor, involuntary trunk swaying, or restlessness patterns, or they can be regression values ​​for parameters that quantify micromotor features, such as dominant frequency, average amplitude, energy, occurrence density, and rhythmic disorder.

[0161] In one embodiment, collaboration with psychiatrists can be used to define a set of micromotor pattern prototypes or feature descriptions that are detectable by millimeter-wave radar and are associated with common psychological states (particularly anxiety disorders, stress responses, certain manifestations of Attention Deficit Hyperactivity Disorder (ADHD), or involuntary movements under high cognitive load), based on clinical observations and relevant literature. For example:

[0162] Anxiety-related tremor: low-amplitude tremors in the extremities or trunk within a specific frequency range (e.g., 4-12 Hz); restless pattern: characterized by frequent, small-scale, irregular shifts in the body's center of gravity while seated, or repetitive small movements of the lower limbs; head micro-posture adjustments under cognitive load: small, frequent adjustments of the head in a specific pattern during the performance of challenging cognitive tasks (which can be combined with radar data from head tracking). These definitions will guide the design of the output labels for the MicroMotionNet module and the annotation of training data.

[0163] Based on the above embodiments, step 110 includes:

[0164] Step 111: Acquire the multi-channel millimeter-wave radar signal of the target object to be evaluated;

[0165] Step 112: Determine the normalized reflection energy, circular fitting error of the in-phase quadrature component map, periodic score of the heartbeat signal, harmonic interference ratio of the respiratory signal, and phase standard deviation of each channel signal in the multi-channel millimeter-wave radar signal.

[0166] Step 113: Determine the comprehensive signal quality index based on at least two of the following: the normalized reflection energy, the circular fitting error of the in-phase orthogonal component plot, the periodic score of the heartbeat signal, the harmonic interference ratio of the respiratory signal, and the phase standard deviation.

[0167] Step 114: Based on the comprehensive signal quality index, select the target channel signal from the multi-channel millimeter-wave radar signals as the millimeter-wave radar signal.

[0168] Specifically, Figure 4 This is a schematic diagram of the selection of millimeter-wave radar signals provided by the present invention, as shown below. Figure 4 As shown, in practical applications, the signal quality received by different channels of a multi-channel radar changes dynamically due to the changing attitude, position, and relative orientation of the target object to be evaluated. Automatically selecting the channel with the best quality for subsequent analysis is a crucial prerequisite for ensuring the stability and accuracy of the evaluation system.

[0169] Accordingly, firstly, multi-channel millimeter-wave radar signals of the target object to be evaluated can be acquired. This is typically achieved by a MIMO radar system, which combines multiple transmit antennas (Tx) and receive antennas (Rx) into multiple virtual receive channels.

[0170] In a specific implementation, the device used to acquire multi-channel millimeter-wave radar signals is a parameter-optimized millimeter-wave radar front-end module. To maximize the sensitivity to weak physiological signals and micro-motion patterns, the preferred parameter configuration for this millimeter-wave radar front-end module is as follows:

[0171] To achieve highly sensitive capture of weak psychophysiological signals (especially submillimeter-level chest cavity displacement and micro-motions related to heartbeat), the following carefully optimized combination of millimeter-wave radar parameters is used in this embodiment of the invention:

[0172] The 79 GHz band offers a good trade-off between wavelength and atmospheric attenuation. Its shorter wavelength (approximately 3.8 mm) is more sensitive to sub-millimeter displacements, overcoming the limitations of traditional radar, which is crucial for accurately detecting chest surface vibrations caused by heartbeats and subtle bodily tremors. Simultaneously, compared to higher frequency bands (e.g., >100 GHz), its atmospheric attenuation is relatively small, ensuring effective detection over a certain distance. This band also boasts a large usable bandwidth. Accordingly, the operating frequency is chosen to be a 79 GHz center frequency (e.g., operating in the 77-81 GHz band).

[0173] According to the distance resolution formula R = c / (2B), where c is the speed of light and B is the bandwidth. A bandwidth of 4 GHz can achieve a distance resolution of approximately 3.75 cm. Such a high distance resolution is crucial for distinguishing the faint physiological vibrations of specific parts of the chest cavity from the movements of other body parts or internal organs, and for accurately locating the reflected signals from the chest cavity surface when covered by clothing or thin blankets.

[0174] Accordingly, the bandwidth uses a 4GHz sweep bandwidth, for example, 77GHz to 81GHz.

[0175] The antenna configuration employs a 4-transmit (Tx) and 8-receive (Rx) multiple-input multiple-output (MIMO) antenna array. The 4Tx / 8Rx MIMO configuration allows for the formation of a large-aperture virtual antenna array with 32 virtual channels through signal processing. This not only significantly improves angular resolution, helping to spatially distinguish desired signals from the chest cavity from interference signals from other body parts, but more importantly, it provides the necessary hardware foundation for subsequent use of advanced digital beamforming techniques and spatial filtering algorithms. This enables the radar energy to be focused on specific areas of the human chest and abdomen, effectively suppressing motion interference and clutter signals from other directions or other parts of the body.

[0176] The sampling rate / frame rate of the ADC (Analog-to-Digital Converter) uses a high-speed ADC to ensure that the sampling rate of the original intermediate frequency (IF) signal is much higher than the Nyquist frequency (e.g., up to tens of MHz) to retain all the information in the signal without distortion.

[0177] The signal processing frame rate (Chirp sequence repetition frequency) is preferably set to 500 Hz. A high frame rate of 500 Hz allows for high-density sampling of cardiac mechanical motion. This is crucial for accurately capturing rapidly changing components (such as QRS complexes) and high-frequency details (such as the fine morphology of P and T waves) in the electrocardiogram waveform, as well as analyzing the high-frequency components of heart rate variability. Compared to the 200 Hz frame rate mentioned in the prior art, a higher frame rate helps improve the ability to capture the more subtle morphological features and rapid changes of interest in this invention. According to psychiatric expertise, these subtle features of electrocardiogram activity and high-frequency changes in HRV are closely related to the functional state of the autonomic nervous system and are key to assessing psychological states such as anxiety and stress. Simultaneously, a high frame rate is also beneficial for tracking rapid micro-movements of the body.

[0178] Transmit Power: While strictly adhering to international safety standards, the Effective Isotropic Radiated Power (EIRP) should be controlled within a safe threshold, for example, an average power density below 10 dBm, to ensure human safety as the primary prerequisite. Based on this, the transmit power should be appropriately optimized, combined with a high-gain antenna and a low-noise receiver, to obtain a sufficient signal-to-noise ratio (SNR), especially when monitoring requires penetration through clothing or bedding.

[0179] The noise figure is selected from receiver front-ends with low noise figures (e.g., <12dB), such as LNAs (Low Noise Amplifiers) and mixers. A low noise figure means less noise introduced by the receiver itself, thus improving the system's sensitivity to weak physiological signals, especially submillimeter-level chest vibrations.

[0180] Phase noise was achieved using a local oscillator (LO) source with extremely low phase noise. Phase information is crucial for extracting minute displacements caused by physiological activities such as heartbeat and respiration. Low phase noise directly affects the accuracy and stability of phase measurements, which is essential for accurately extracting inter-beat interval (IBI) sequences, calculating HRV parameters, and identifying micro-motion patterns.

[0181] The radar front-end module that implements the above configuration may include an FMCW (Frequency-Modulated Continuous Wave) signal generator, a power distribution network, four transmit antennas, eight receive antennas, an array of eight parallel low-noise amplifiers (LNAs), a mixer array, a baseband filter array, a high-speed multi-channel synchronous analog-to-digital converter (ADC), and a field-programmable gate array (FPGA) or a dedicated digital signal processor (DSP) for radar parameter configuration, chirp sequence generation and timing control, and initial buffering and formatting of raw data.

[0182] Next, the normalized reflected energy, IQ plot circular fitting error, cardiac periodicity score, respiratory harmonic interference ratio, and phase standard deviation of each channel in the multi-channel millimeter-wave radar signal were determined. The normalized reflected energy reflects the strength of the channel signal; higher energy generally indicates a better signal-to-noise ratio. The IQ plot circular fitting error: Under ideal stationary conditions, the IQ plot of a single target reflection should be circular. Motion artifacts distort the circle; therefore, the IQ plot circular fitting error reflects the degree of motion interference, and a smaller error indicates better quality.

[0183] Here, the periodicity score of the heartbeat signal is used to assess the periodicity intensity and stability of the signal within the heartbeat frequency band. A higher periodicity score indicates a clearer heartbeat signal. Specifically, the calculation process for the periodicity score is as follows: First, preliminary bandpass filtering is performed on the raw multi-channel millimeter-wave radar signal, with the filtering frequency band set to 0.8-2.5Hz to accurately extract the heartbeat frequency band signal. Then, an autocorrelation function is performed on the filtered signal, and the score is quantified by analyzing the main peak characteristics of the autocorrelation results. The significance of the main peak is characterized by the ratio of the peak amplitude to the sidelobe amplitude; a higher ratio indicates a more prominent periodicity. Peak position stability is assessed by the degree of fit between the actual peak frequency and the clinically recognized normal heart rate range (usually 60-100 beats / min, corresponding to a frequency of 1-1.67Hz). Finally, the periodicity score, formed by combining these two indicators, indicates that a higher value represents a clearer periodicity and better stability of the heartbeat signal, providing a reliable basis for subsequent heartbeat signal quality assessment and physiological state monitoring.

[0184] The respiratory signal harmonic interference ratio (HRI) is used to assess the degree of interference of respiratory signal harmonics (0.1-0.5 Hz) on the heartbeat signal frequency band. The lower the HRI value, the less interference and the better the signal quality.

[0185] The phase standard deviation is used to characterize signal stability. The smaller the phase fluctuation after unwrapping within a short time window, the less the signal is affected by noise and random motion interference.

[0186] Here, the formula for the overall signal quality index is as follows:

[0187] ;

[0188] in, This represents the overall signal quality index. This represents the normalized reflected energy. This indicates the circular fitting error of the in-phase orthogonal component plot. This represents the periodic score of the heartbeat signal. Indicates the harmonic interference ratio of the respiratory signal. Indicates the phase standard deviation. The weights representing the normalized reflected energy. The weights representing the circular fitting error of the in-phase orthogonal component plot. The weights representing the periodic scores of the heartbeat signal. The weights representing the harmonic interference ratio of the respiratory signal. The weights representing the phase standard deviation This represents each channel signal in a multi-channel millimeter-wave radar signal.

[0189] Finally, based on the comprehensive signal quality index, the target channel signal is selected from the multi-channel millimeter-wave radar signals as the millimeter-wave radar signal. The most direct way is to select the signal from the channel with the highest comprehensive signal quality index score as the input for subsequent processing (such as input to the physiological feature extraction module and the micro-motion feature extraction module).

[0190] The method provided in this invention proposes a multi-factor, data-driven adaptive channel selection method. By evaluating the signal quality of each channel in real time and selecting the optimal channel, this method can effectively cope with signal quality fluctuations caused by changes in target posture and random motion, ensuring that the signal input to the downstream analysis module is always the best available signal source, thereby greatly enhancing the robustness and reliability of the entire psychological state assessment system.

[0191] Based on the above embodiments, step 114 includes:

[0192] Step 1141: Based on the comprehensive signal quality index, perform spatial filtering on the multi-channel millimeter-wave radar signal to obtain a spatially filtered signal;

[0193] Step 1142: Perform source signal separation processing on the spatial filtered signal to obtain the separated radar signal;

[0194] Step 1143: Perform phase unwrapping processing on the separated radar signal to obtain a continuous phase signal; wherein, the unwrapping path of the phase unwrapping processing is guided by the phase quality map of the separated radar signal;

[0195] Step 1144: Perform multi-resolution separation on the continuous phase signal to separate the target channel signal related to cardiac activity, and use the target channel signal as the millimeter-wave radar signal.

[0196] Specifically, firstly, spatial filtering can be performed on multi-channel millimeter-wave radar signals based on the comprehensive signal quality index to obtain a spatially filtered signal. The comprehensive signal quality index can be used not only for channel selection but also to guide spatial filtering. For example, adaptive digital beamforming (ADBF), such as MVDR (Minimum Variance Distortionless Response), can be used to take the channel signal with a high comprehensive signal quality index as the desired signal and the channel signal with a low comprehensive signal quality index as the interference reference, thereby forming a virtual beam pointing towards the chest cavity of the target object. This spatially suppresses interference from other directions (such as a swinging arm) and yields a spatially filtered signal with a higher signal-to-noise ratio.

[0197] Then, the spatially filtered signal can be processed to separate the source signals, resulting in a separated radar signal. Even after spatial filtering, the signal may still contain residual artifacts such as heartbeats, respiration, and micro-motion artifacts. Source signal separation can be achieved using Constrained Independent Component Analysis (cICA), a blind separation technique that utilizes the statistical properties of source signals (such as the quasi-periodicity of heartbeats) to decompose the mixed signal into multiple independent source signal components, thus obtaining a purer separated radar signal.

[0198] Furthermore, the separated radar signals can be phase-unwrapped to obtain continuous phase signals. Since the original phase measured by the radar is limited to the [-π, π] interval, unwrapping is necessary to reflect the true displacement. A key point of this embodiment is that the unwrapping path is guided by the phase quality map of the separated radar signals. This phase quality map can be generated from some comprehensive signal quality indicators, such as phase standard deviation. The phase quality map can identify the quality status of different spatiotemporal points in the separated radar signals, clearly indicating which spatiotemporal points have higher quality and which have lower quality. The unwrapping algorithm will prioritize along the high-quality path, avoiding starting unwrapping from low-quality regions, which would lead to error accumulation and propagation, thus significantly improving the accuracy of unwrapping.

[0199] In the phase unwrapping process, this embodiment employs a dynamic threshold hierarchical discrimination mechanism to improve unwrapping accuracy, specifically for the phase jump correction stage. Specifically, when the phase difference between adjacent sampling points exceeds a preset wide dynamic threshold based on the dynamic characteristics of physiological signals (this threshold is set according to the maximum possible phase change amount of physiological activities such as heartbeat), the system first determines it as a 2π integer multiple jump and performs preliminary correction. To further suppress noise interference and preserve the rapid change characteristics of the true physiological signal, a narrow dynamic threshold is simultaneously introduced as a secondary criterion: if the phase difference after correction still continues to exceed the narrow dynamic threshold range, joint discrimination is performed through time-series signal trend analysis, and secondary correction is implemented by combining the phase change characteristics of the preceding and following times.

[0200] This hierarchical discrimination mechanism effectively solves the problem of noise sensitivity in traditional path integral methods by dynamically adjusting threshold parameters. While maintaining the ability to capture rapidly changing features of real physiological signals, it significantly improves the noise resistance of phase unwrapping. Its technical principle absorbs the essence of classic algorithms such as Differentiated and Cross-Multiplied Demodulation (DACM) in phase ambiguity processing, and achieves methodological innovation through dynamic threshold adaptive adjustment and a two-level verification mechanism, forming a phase unwrapping optimization scheme that combines theoretical rigor with engineering practicality.

[0201] Finally, multi-resolution separation can be performed on the continuous phase signal to isolate the target channel signal related to cardiac activity, which is then used as the final millimeter-wave radar signal for analysis. Multi-resolution separation, such as Multi-resolution Singular Value Decomposition (MSVD), is an advanced band separation technique. It decomposes the signal into different wavelet scales (resolutions) and then uses SVD to accurately separate the cardiac activity signal (typically with dominant energy in the 0.8–4.0 Hz band) and the respiratory signal (typically in the 0.1–0.8 Hz band) at specific scales, while further filtering out out-of-band noise.

[0202] Understandably, compared to traditional bandpass filtering, MSVD has better adaptability and the ability to process non-stationary signals; compared to VMD (Variational Mode Decomposition) or EMD (Empirical Mode Decomposition), it is less sensitive to parameter selection and the decomposition is more stable.

[0203] The method provided in this invention, through a series of precise processes including spatial filtering, source signal separation, high-quality guided phase unwrapping, and multi-resolution separation, can suppress various interferences and noises to the greatest extent, and extract cardiac activity-related signals with extremely high signal-to-noise ratio and extremely pure morphology from multi-channel signals. This provides the most ideal input for the subsequent physiological feature extraction module to reconstruct high-fidelity cardiac activity waveforms, and is an important guarantee for achieving high-precision psychological state assessment.

[0204] Based on any of the above embodiments Figure 5 This is the second flowchart illustrating the method for determining psychological state assessment parameters provided by this invention, as shown below. Figure 5 As shown, the first step is to configure the radar parameters. This step aims to optimize the radar system's operating parameters to enhance its sensitivity to weak physiological and micro-motion signals. After configuration, the multi-channel radar signal acquisition step is performed, using a multi-input multi-output radar system to acquire multi-channel reflected signals from the target object to be evaluated.

[0205] Subsequently, the system performs real-time signal quality assessment and optimal channel selection / fusion on the acquired signals. This step calculates the comprehensive quality index of each channel signal, selects the single channel with the best signal quality, or performs weighted fusion of multiple channels to obtain a high-quality signal source for subsequent processing.

[0206] Next, a motion artifact intelligent recognition and adaptive compensation step is performed on the high-quality signal. This step employs advanced signal processing algorithms to identify, eliminate, or suppress signal interference caused by the macroscopic motion of the target object's body.

[0207] After removing major motion artifacts, the signal undergoes phase unwrapping, converting the radar-measured phase, which is confined to the range of [-π, π], into a continuous phase signal that directly reflects minute displacements on the body surface. Then, preliminary cardiopulmonary signal separation is performed on the continuous phase signal. Using filtering or source separation techniques, the signal is roughly separated into physiological rhythm components related to heartbeat and respiration, as well as other residual signal components.

[0208] After that, the signal processing flow is divided into two parallel branches: one part of the signal is sent to the physiological feature extraction module, which is used to reconstruct high-fidelity electrocardiogram waveforms and extract fine physiological features such as heart rate variability from them; the other part of the signal is sent to the micromotor feature extraction module, which focuses on identifying and quantifying the body micromotor pattern features related to specific psychological states from the signal.

[0209] Next, the physiological features and micro-motion features extracted by the two modules are simultaneously input into the multimodal feature fusion module. This module is responsible for effectively integrating the features from these two different modalities to generate a fused feature with richer information dimensions and stronger discriminative ability.

[0210] Subsequently, this fused feature is fed as input to the mental state inference module. This module is typically a pre-trained deep learning model that performs comprehensive analysis and inference based on the input fused feature. Finally, the mental state inference module outputs an evaluation conclusion, namely the mental state evaluation parameters, thereby completing an objective and quantitative assessment of the target object's current mental state.

[0211] In summary, the method provided by this invention, through optimized radar parameters, a hierarchical adaptive motion artifact suppression strategy, and intelligent signal quality assessment, exhibits extremely high motion robustness and signal fidelity. It effectively removes motion interference during everyday minor activities while preserving weak physiological signals and micro-motion information related to psychological states. The physiological feature extraction module, combined with morphological fidelity loss guided by psychiatric knowledge, achieves unprecedented precision in physiological information, enabling the reconstruction of high-fidelity ECG waveforms and the calculation of accurate HRV parameters. The micro-motion feature extraction module innovatively conducts micro-motion psychological correlation analysis, intelligently identifying and quantifying subtle body movement patterns related to specific psychological states. Simultaneously, the deep integration of psychiatric expertise into the entire technical process significantly enhances clinical applicability and interpretability. By comprehensively analyzing high-precision physiological and micro-motion behavior data using large-scale AI models or dedicated Transformer networks, a comprehensive, objective, and dynamic psychological state profile can be provided. Ultimately, this technology provides a key breakthrough for non-contact psychological health monitoring, early warning, and assisted diagnosis in real-world scenarios, possessing significant social and clinical application value.

[0212] The following describes the apparatus for determining psychological state assessment parameters provided by the present invention. The apparatus for determining psychological state assessment parameters described below can be referred to in correspondence with the method for determining psychological state assessment parameters described above.

[0213] Based on any of the above embodiments, the present invention provides a device for determining psychological state assessment parameters. Figure 6 This is a schematic diagram of the structure of the psychological state assessment parameter determination device provided by the present invention, as shown below. Figure 6 As shown, the device includes:

[0214] Acquisition unit 610 is used to acquire millimeter-wave radar signals of the target object to be evaluated;

[0215] The psychological state assessment unit 620 is used to input the millimeter-wave radar signal into the psychological state assessment model to obtain the psychological state assessment parameters output by the psychological state assessment model.

[0216] The psychological state assessment model includes a physiological feature extraction module, a micro-motor feature extraction module, a multimodal feature fusion module, and a psychological state inference module.

[0217] The physiological feature extraction module is used to reconstruct the electrocardiogram (ECG) waveform based on the millimeter-wave radar signal and extract physiological features from the ECG waveform; the micro-motion feature extraction module is used to extract micro-motion features from the millimeter-wave radar signal to obtain micro-motion features; the multimodal feature fusion module is used to fuse the physiological features and the micro-motion features to obtain fused features; and the psychological state inference module is used to obtain the psychological state assessment parameters based on the fused features.

[0218] The apparatus provided in this invention acquires millimeter-wave radar signals of the target object to be evaluated; inputs the millimeter-wave radar signals into a psychological state assessment model to obtain psychological state assessment parameters; wherein, the psychological state assessment model includes a physiological feature extraction module for reconstructing electrocardiogram waveforms and extracting physiological features based on millimeter-wave radar signals, a micro-motion feature extraction module for extracting micro-motion features from millimeter-wave radar signals, a multimodal feature fusion module for fusing physiological features and micro-motion features to obtain fused features, and a psychological state inference module for obtaining psychological state assessment parameters based on the fused features. This invention reconstructs electrocardiogram waveforms through the physiological feature extraction module to obtain refined physiological features, extracts micro-motion features through the micro-motion feature extraction module, and then integrates features reflecting internal physiological changes and external behavioral manifestations through the multimodal feature fusion module, thereby forming a more comprehensive and dimensionally rich fused feature for comprehensive analysis by the psychological state inference module. This solves the problems of incomplete and inaccurate assessment results caused by insufficient refinement of physiological information extraction, neglect of micro-motion information, and single assessment modality in existing technologies, greatly improving the comprehensiveness and accuracy of psychological state assessment in unconstrained daily environments.

[0219] Based on any of the above embodiments, a training unit is further included, wherein the training unit specifically includes:

[0220] A sample acquisition unit is used to acquire millimeter-wave radar sample signals and reference electrocardiogram activity waveforms that are synchronously acquired with the millimeter-wave radar sample signals.

[0221] The feature extraction unit is used to input the millimeter-wave radar sample signal into the initial physiological feature extraction module to obtain the electrocardiogram activity waveform to be reconstructed output by the initial physiological feature extraction module;

[0222] A morphological fidelity loss determination unit is used to determine the diagnostic weight corresponding to each waveform segment based on the clinical diagnostic importance of different waveform segments in the reference electrocardiogram waveform, and to calculate the weighted morphological difference between the electrocardiogram waveform to be reconstructed and the reference electrocardiogram waveform based on the diagnostic weight, thereby determining the morphological fidelity loss.

[0223] The parameter iteration unit is used to determine the target loss based on the morphological fidelity loss, and to perform parameter iteration on the initial physiological feature extraction module based on the target loss to obtain the physiological feature extraction module.

[0224] Based on any of the above embodiments, the parameter iteration unit specifically includes:

[0225] The parameter extraction unit is used to extract a first heart rate variability parameter and a second heart rate variability parameter from the electrocardiogram waveform to be reconstructed and the reference electrocardiogram waveform, respectively.

[0226] A consistency loss determination unit is used to determine the consistency loss based on the first heart rate variability parameter and the second heart rate variability parameter;

[0227] The discriminant output unit is used to input the ECG activity waveform to be reconstructed and the reference ECG activity waveform into the generative adversarial network (GAN). The discriminator in the GAN determines a first discriminant output for the reference ECG activity waveform and a second discriminant output for the ECG activity waveform to be reconstructed. The GAN includes a generator and a discriminator, and the generator is the initial physiological feature extraction module.

[0228] A discriminator loss determination unit is configured to determine the discriminator loss of the discriminator based on the first discrimination output and the second discrimination output;

[0229] A generator loss determination unit is used to determine the generator loss of the generator based on the negative number of the second discrimination output;

[0230] An adversarial loss determination unit is used to determine the adversarial loss based on the discriminator loss and the generator loss;

[0231] A target loss unit is defined to determine the target loss based on the morphological fidelity loss, the consistency loss, and the adversarial loss.

[0232] Based on any of the above embodiments, determining the target loss is specifically used for:

[0233] The ECG waveform to be reconstructed and the millimeter-wave radar sample signal are mapped to a shared feature space to obtain the features of the ECG waveform to be reconstructed and the features of the millimeter-wave radar sample signal. Based on the feature distance between the ECG waveform features to be reconstructed and the millimeter-wave radar sample signal features in the shared feature space, the content loss is determined.

[0234] The target loss is determined based on the morphological fidelity loss, the consistency loss, the adversarial loss, and the content loss.

[0235] Based on any of the above embodiments, the micro-motion feature extraction module is specifically used for:

[0236] Physiological rhythm filtering is performed on the millimeter-wave radar signal to obtain micro-motion residual signals;

[0237] The residual signal of micro-motion is input into a temporal convolutional network to obtain the initial temporal features output by the temporal convolutional network;

[0238] Temporal attention weights and / or channel attention weights are applied to the initial temporal features to obtain enhanced temporal features;

[0239] Based on the enhanced temporal features, the micro-motion features are determined.

[0240] Based on any of the above embodiments, the acquisition unit 610 specifically includes:

[0241] A signal acquisition unit is used to acquire the multi-channel millimeter-wave radar signal of the target object to be evaluated.

[0242] The indicator unit is used to determine the normalized reflection energy, circular fitting error of the in-phase quadrature component plot, periodic score of the heartbeat signal, harmonic interference ratio of the respiratory signal, and phase standard deviation of each channel signal in the multi-channel millimeter-wave radar signal.

[0243] A comprehensive signal quality index unit is defined to determine the comprehensive signal quality index based on at least two of the following: the normalized reflection energy, the circular fitting error of the in-phase orthogonal component plot, the periodic score of the heartbeat signal, the harmonic interference ratio of the respiratory signal, and the phase standard deviation.

[0244] The selection unit is used to select a target channel signal as the millimeter-wave radar signal from the multi-channel millimeter-wave radar signals based on the comprehensive signal quality index.

[0245] Based on any of the above embodiments, the selection unit is specifically used for:

[0246] Based on the comprehensive signal quality index, the multi-channel millimeter-wave radar signal is spatially filtered to obtain a spatially filtered signal.

[0247] The spatially filtered signal is subjected to source signal separation processing to obtain the separated radar signal;

[0248] The separated radar signal is subjected to phase unwrapping processing to obtain a continuous phase signal; wherein, the unwrapping path of the phase unwrapping processing is guided by the phase quality map of the separated radar signal;

[0249] The continuous phase signal is subjected to multi-resolution separation to separate the target channel signal related to cardiac activity, and the target channel signal is used as the millimeter-wave radar signal.

[0250] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 7 As shown, the electronic device may include: a processor 710, a communications interface 720, a memory 730, and a communications bus 740, wherein the processor 710, the communications interface 720, and the memory 730 communicate with each other through the communications bus 740. The processor 710 can call logical instructions in the memory 730 to execute a method for determining psychological state assessment parameters. This method includes: acquiring millimeter-wave radar signals of the target object to be assessed; inputting the millimeter-wave radar signals into a psychological state assessment model to obtain psychological state assessment parameters output by the model; wherein the psychological state assessment model includes a physiological feature extraction module, a micro-motion feature extraction module, a multimodal feature fusion module, and a psychological state inference module; the physiological feature extraction module is used to reconstruct electrocardiogram (ECG) waveforms based on the millimeter-wave radar signals and extract physiological features from the ECG waveforms; the micro-motion feature extraction module is used to extract micro-motion features from the millimeter-wave radar signals to obtain micro-motion features; the multimodal feature fusion module is used to fuse the physiological features and the micro-motion features to obtain fused features; and the psychological state inference module is used to obtain the psychological state assessment parameters based on the fused features.

[0251] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0252] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the psychological state assessment parameter determination method provided by the above methods. The method includes: acquiring millimeter-wave radar signals of the target object to be assessed; inputting the millimeter-wave radar signals into a psychological state assessment model to obtain psychological state assessment parameters output by the psychological state assessment model; wherein, the psychological state assessment model includes a physiological feature extraction module, a micro-motion feature extraction module, a multimodal feature fusion module, and a psychological state inference module; the physiological feature extraction module is used to reconstruct electrocardiogram waveforms based on the millimeter-wave radar signals and extract physiological features from the electrocardiogram waveforms; the micro-motion feature extraction module is used to extract micro-motion features from the millimeter-wave radar signals to obtain micro-motion features; the multimodal feature fusion module is used to fuse the physiological features and the micro-motion features to obtain fused features; and the psychological state inference module is used to obtain the psychological state assessment parameters based on the fused features.

[0253] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for determining psychological state assessment parameters provided by the methods described above. This method includes: acquiring millimeter-wave radar signals of a target object to be assessed; inputting the millimeter-wave radar signals into a psychological state assessment model to obtain psychological state assessment parameters output by the psychological state assessment model; wherein the psychological state assessment model includes a physiological feature extraction module, a micro-motion feature extraction module, a multimodal feature fusion module, and a psychological state inference module; the physiological feature extraction module is used to reconstruct an electrocardiogram waveform based on the millimeter-wave radar signals and extract physiological features from the electrocardiogram waveform; the micro-motion feature extraction module is used to extract micro-motion features from the millimeter-wave radar signals to obtain micro-motion features; the multimodal feature fusion module is used to fuse the physiological features and the micro-motion features to obtain fused features; and the psychological state inference module is used to obtain the psychological state assessment parameters based on the fused features.

[0254] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0255] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0256] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for determining parameters for psychological state assessment, characterized in that, include: Acquire the millimeter-wave radar signal of the target object to be evaluated; The millimeter-wave radar signal is input into the psychological state assessment model to obtain the psychological state assessment parameters output by the psychological state assessment model. The psychological state assessment model includes a physiological feature extraction module, a micro-motor feature extraction module, a multimodal feature fusion module, and a psychological state inference module. The physiological feature extraction module is used to reconstruct the electrocardiogram (ECG) waveform based on the millimeter-wave radar signal and extract physiological features from the ECG waveform; the micro-motion feature extraction module is used to extract micro-motion features from the millimeter-wave radar signal to obtain micro-motion features; the multimodal feature fusion module is used to fuse the physiological features and the micro-motion features to obtain fused features; the psychological state inference module is used to obtain the psychological state assessment parameters based on the fused features; the ECG waveform refers to a time series reconstructed from the millimeter-wave radar signal that highly simulates the real ECG signal in waveform morphology and rhythm; The training steps of the physiological feature extraction module include: Acquire millimeter-wave radar sample signals and reference electrocardiogram activity waveforms acquired synchronously with the millimeter-wave radar sample signals; The millimeter-wave radar sample signal is input into the initial physiological feature extraction module to obtain the electrocardiogram activity waveform to be reconstructed output by the initial physiological feature extraction module; Based on the clinical diagnostic importance of different waveform segments in the reference ECG waveform, the diagnostic weights corresponding to each waveform segment are determined, and the weighted morphological differences between the ECG waveform to be reconstructed and the reference ECG waveform are calculated based on the diagnostic weights to determine the morphological fidelity loss. Based on the morphological fidelity loss, a target loss is determined, and the parameters of the initial physiological feature extraction module are iterated based on the target loss to obtain the physiological feature extraction module.

2. The method for determining psychological state assessment parameters according to claim 1, characterized in that, The determination of the target loss based on the morphological fidelity loss includes: The first heart rate variability parameter and the second heart rate variability parameter are extracted from the electrocardiogram waveform to be reconstructed and the reference electrocardiogram waveform, respectively. Based on the first heart rate variability parameter and the second heart rate variability parameter, the consistency loss is determined; The ECG waveform to be reconstructed and the reference ECG waveform are input into a generative adversarial network (GAN). The discriminator in the GAN determines a first discriminant output for the reference ECG waveform and a second discriminant output for the ECG waveform to be reconstructed. The GAN includes a generator and a discriminator, and the generator is the initial physiological feature extraction module. Based on the first discrimination output and the second discrimination output, determine the discriminator loss of the discriminator; The generator loss of the generator is determined based on the negative number of the second discrimination output; Based on the discriminator loss and the generator loss, the adversarial loss is determined; The target loss is determined based on the morphological fidelity loss, the consistency loss, and the adversarial loss.

3. The method for determining psychological state assessment parameters according to claim 2, characterized in that, The determination of the target loss based on the morphological fidelity loss, the consistency loss, and the adversarial loss includes: The ECG waveform to be reconstructed and the millimeter-wave radar sample signal are mapped to a shared feature space to obtain the features of the ECG waveform to be reconstructed and the features of the millimeter-wave radar sample signal. Based on the feature distance between the ECG waveform features to be reconstructed and the millimeter-wave radar sample signal features in the shared feature space, the content loss is determined. The target loss is determined based on the morphological fidelity loss, the consistency loss, the adversarial loss, and the content loss.

4. The method for determining psychological state assessment parameters according to any one of claims 1 to 3, characterized in that, The micro-motion feature extraction module is specifically used for: Physiological rhythm filtering is performed on the millimeter-wave radar signal to obtain micro-motion residual signals; The residual signal of micro-motion is input into a temporal convolutional network to obtain the initial temporal features output by the temporal convolutional network; Temporal attention weights and / or channel attention weights are applied to the initial temporal features to obtain enhanced temporal features; Based on the enhanced temporal features, the micro-motion features are determined.

5. The method for determining psychological state assessment parameters according to any one of claims 1 to 3, characterized in that, The acquisition of the millimeter-wave radar signal of the target object to be evaluated includes: Acquire the multi-channel millimeter-wave radar signal of the target object to be evaluated; The normalized reflected energy, circular fitting error of the in-phase quadrature component plot, periodic score of the heartbeat signal, harmonic interference ratio of the respiratory signal, and phase standard deviation of each channel signal in the multi-channel millimeter-wave radar signal are determined respectively. The comprehensive signal quality index is determined based on at least two of the following: the normalized reflection energy, the circular fitting error of the in-phase orthogonal component plot, the periodic score of the heartbeat signal, the harmonic interference ratio of the respiratory signal, and the phase standard deviation. Based on the comprehensive signal quality index, the target channel signal is selected from the multi-channel millimeter-wave radar signals as the millimeter-wave radar signal.

6. The method for determining psychological state assessment parameters according to claim 5, characterized in that, The step of selecting the target channel signal as the millimeter-wave radar signal from the multi-channel millimeter-wave radar signals based on the comprehensive signal quality index includes: Based on the comprehensive signal quality index, the multi-channel millimeter-wave radar signal is subjected to spatial filtering to obtain a spatially filtered signal. The spatially filtered signal is subjected to source signal separation processing to obtain the separated radar signal; The separated radar signal is subjected to phase unwrapping processing to obtain a continuous phase signal; wherein, the unwrapping path of the phase unwrapping processing is guided by the phase quality map of the separated radar signal; The continuous phase signal is subjected to multi-resolution separation to separate the target channel signal related to cardiac activity, and the target channel signal is used as the millimeter-wave radar signal.

7. A device for determining parameters for psychological state assessment, characterized in that, include: The acquisition unit is used to acquire the millimeter-wave radar signal of the target object to be evaluated. A psychological state assessment unit is used to input the millimeter-wave radar signal into a psychological state assessment model to obtain psychological state assessment parameters output by the psychological state assessment model. The psychological state assessment model includes a physiological feature extraction module, a micro-motor feature extraction module, a multimodal feature fusion module, and a psychological state inference module. The physiological feature extraction module is used to reconstruct the electrocardiogram (ECG) waveform based on the millimeter-wave radar signal and extract physiological features from the ECG waveform; the micro-motion feature extraction module is used to extract micro-motion features from the millimeter-wave radar signal to obtain micro-motion features; the multimodal feature fusion module is used to fuse the physiological features and the micro-motion features to obtain fused features; the psychological state inference module is used to obtain the psychological state assessment parameters based on the fused features; the ECG waveform refers to a time series reconstructed from the millimeter-wave radar signal that highly simulates the real ECG signal in waveform morphology and rhythm; It also includes a training unit, which is specifically used for: Acquire millimeter-wave radar sample signals and reference electrocardiogram activity waveforms acquired synchronously with the millimeter-wave radar sample signals; The millimeter-wave radar sample signal is input into the initial physiological feature extraction module to obtain the electrocardiogram activity waveform to be reconstructed output by the initial physiological feature extraction module; Based on the clinical diagnostic importance of different waveform segments in the reference ECG waveform, the diagnostic weights corresponding to each waveform segment are determined, and the weighted morphological differences between the ECG waveform to be reconstructed and the reference ECG waveform are calculated based on the diagnostic weights to determine the morphological fidelity loss. Based on the morphological fidelity loss, a target loss is determined, and the parameters of the initial physiological feature extraction module are iterated based on the target loss to obtain the physiological feature extraction module.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method for determining psychological state assessment parameters as described in any one of claims 1 to 6.

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, it implements the method for determining psychological state assessment parameters as described in any one of claims 1 to 6.

Images