A multi-domain feature-based electrodermal activity signal automatic analysis method and system

Abstract

This invention relates to the field of physiological signal analysis technology, specifically to an automatic analysis method and system for electroskin activity (ESA) signals based on multi-domain features. The method includes the following steps: acquiring ESA signals and correcting abnormal fluctuations within a fixed time window to obtain stable and continuous time-series data; filtering and downsampling the corrected signals to characterize the time-domain stability; analyzing signal change trends using moving averages and extracting low-frequency evolution and differential changes to form a multi-component joint characterization; iteratively calculating the multi-component power distribution and spectral centroid features from the frequency domain; statistically judging the frequency domain consistency of the multi-components; and fusing time-domain stability information to form a multi-domain feature analysis result. In this invention, robust statistical correction of abnormal fluctuations and characterization of signal stability ensures that the ESA time series remains consistent under noise drift. The fusion of low-frequency and differential changes constructs a multi-component characterization, and combined with frequency domain consistency analysis, achieves time-frequency verification, improving the stability and comparability of the results.

Description

Technical Field

[0001] This invention relates to the field of physiological signal analysis technology, and in particular to an automatic analysis method and system for skin electrical activity signals based on multi-domain features. Background Technology

[0002] The field of physiological signal analysis technology takes the electrophysiological signals generated by human physiological activities as its research object. It is a technical field that uses the acquisition, recording, quantitative characterization, analysis and modeling of signals to characterize changes in human physiological and psychological states. This field usually covers the acquisition methods of physiological signals, the description of time series characteristics, the construction of statistical features, and the analysis of signal change patterns under different physiological states. The core focus is on the objective characterization and pattern analysis of signals such as cardiac electroencephalogram (CEG), electromyography (EMG), and electrodermal osmosis (EDS). Its technical system emphasizes the consistency of the physiological mechanism of the signal source, the measurability of the signal manifestation, and the repeatability of the analysis process.

[0003] Among them, the traditional automatic analysis method and system for skin electrophysiological activity (SEM) signals refers to the technical solution that processes and analyzes the original continuous signal according to a predetermined process, which is the signal of changes in skin conductance or resistance caused by changes in sweat gland secretion. Typically, SEM signals are collected at a fixed sampling frequency, the signal sequence is segmented, and statistical quantities such as mean, variance, rise time, and fall time are calculated in the time dimension. In the frequency dimension, the power spectrum distribution is obtained through Fourier transform and specific frequency band energy characteristics are extracted. In the amplitude variation dimension, the number of peaks, peak amplitude, and duration of the skin electrophysiological response are identified. Based on manually set discrimination rules or thresholds, the above-mentioned multiple features are combined to complete the automatic analysis of the state of skin electrophysiological activity.

[0004] Traditional automated analysis of electrodermal activity (EDA) signals relies on fixed sampling and segmented statistical methods, directly calculating the mean, variance, and peak characteristics in continuous signals. This approach is sensitive to abnormal fluctuations and transient interference, easily incorporating non-physiological changes into the analysis. Furthermore, the time-domain and frequency-domain features are constructed independently, lacking constraints on the consistency of multi-scale changes. In scenarios involving state transitions or slow signal drift, the analysis results lack stability. The discrimination rules rely on manual experience, limiting the repeatability and objective consistency of results under different experimental conditions, thus affecting the reliability of EDA state assessment. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide an automatic analysis method for skin conductance activity signals based on multi-domain features, comprising the following steps: S1: Collect skin conductance activity signals, calculate the median absolute deviation within a fixed time window and filter out abnormal data, replace them with the median of the corresponding window, and generate corrected skin conductance activity time series data; S2: Based on the corrected skin conductance activity time series data, perform filtering calculations, downsample the filtering results according to a preset sampling ratio, calculate the time domain stability, and generate time domain stable data; S3: Based on the time-domain stable data, perform a centered moving average to calculate the smoothing analysis sequence, and perform low-frequency and differential operations on the smoothing analysis sequence to obtain the low-frequency and differential change sequence. Integrate all sequences to generate multi-component joint analysis results. S4: Based on the results of the multi-component joint analysis, perform spectrum calculation, calculate the total power, spectral centroid frequency, and low-frequency to high-frequency power ratio of the multi-components, compare the consistency of the frequency domain characteristics of the multiple components, and generate frequency domain characteristic analysis results; S5: Based on the frequency domain feature analysis results, extract the total power distribution, the frequency order of the spectral centroid, and the direction of change of the power ratio between low and high frequencies to determine the consistency of direction, and then fuse and weight the results with the time domain stability to generate the multi-domain feature analysis results of skin electrical activity.

[0006] As a further aspect of the present invention, the corrected electrodermal activity time-series data includes a window midpoint sequence, an abnormal substitution location index, and a corrected amplitude time series; the time-domain stable data includes a downsampled amplitude sequence, an adjacent difference mean index, and a stability status indicator; the multi-component joint analysis results include a smoothing analysis sequence, a low-frequency change sequence, and a difference change sequence; the frequency domain feature analysis results include the total power value of the multi-components, the spectral centroid frequency set, and the low-frequency to high-frequency power ratio parameter; and the multi-domain feature analysis results of electrodermal activity include a total power distribution consistency indicator and a power ratio change direction fusion determination result.

[0007] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Collect skin conductance signals, split the signal sequence into fixed time windows, perform sorting operations on the sampled values ​​within multiple windows, select the median value of the sequence as the statistical benchmark, and calculate the absolute value of the difference between the multiple sampled values ​​and the benchmark value to generate the window absolute deviation statistic. S102: Based on the absolute deviation statistic of the window, call the skin conductance activity signal sampling value within the corresponding time window, calculate the absolute value of the difference between the multiple sampling values ​​and the median value, and if the absolute value of the difference is greater than the product of the absolute deviation statistic and the set ratio, it is determined to be abnormal and the abnormal sampling index set is obtained. S103: Based on the abnormal sampling index set, call the original sampling values ​​of the skin conductance activity signal within the corresponding time window, locate the abnormal index and perform a value replacement operation. The replacement value is selected as the median of the non-abnormal sampling sequence within the same window. The replaced sequence is recombined according to time to generate corrected skin conductance activity time series data.

[0008] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Based on the corrected electrodermal activity time series data, perform time-domain smoothing filtering on the continuous sampling point sequence, call the amplitude of adjacent sampling points and perform weighted accumulation for smoothing constraint processing, and rearrange the processed sequence according to the time index to obtain the filtered electrodermal activity time series sequence; S202: Based on the filtered skin electrophysiological time series, call the preset sampling ratio parameter, perform equidistant sampling operation on the time series index, retain the corresponding sampling amplitude, complete the time scale compression process, and generate a downsampled skin electrophysiological time series. S203: For the downsampled electrodermal time series, perform difference calculation on the amplitude of adjacent sampling points, process the absolute value of the difference result, collect all differences and perform mean calculation. If the mean result of the difference does not reach the time domain stability threshold, it is determined to be a stable state value, and time domain stable data is generated.

[0009] As a further aspect of the present invention, the time-domain stability threshold is determined by analyzing the amplitude difference between adjacent sampling points in the downsampled electrodermal time sequence, and by analyzing the standard deviation of the mean of all absolute differences and a preset multiple.

[0010] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Based on the time-domain stable data, call the centered sliding window for the values ​​corresponding to the continuous time index, perform weighted summation on the values ​​in the window and divide by the window length to obtain the values ​​of the smoothed sequence, and rearrange them according to the original time index order to generate a smoothed analysis sequence. S302: Based on the smoothing analysis sequence, perform low-frequency component extraction operation on the overall numerical distribution of the sequence, perform difference calculation on the corresponding values ​​of adjacent time indices, and maintain the correspondence of time indices to generate a low-frequency analysis sequence and a difference change sequence; S303: Call the smoothing analysis sequence, the low-frequency analysis sequence, and the differential change sequence, perform alignment mapping on the three types of sequences based on time index consistency, and aggregate the corresponding index position values ​​to generate multi-component joint analysis results.

[0011] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Based on the results of the multi-component joint analysis, obtain the numerical sequence of the time index of the multi-component, perform discrete spectrum transformation on the time series data of each component, map the time domain values ​​to the frequency axis and form an amplitude distribution, retain the correspondence between frequency resolution and index, and generate the component spectrum sequence. S402: Based on the component spectrum sequence, perform integration and weighting on the spectrum amplitude and frequency, calculate the total power as the sum of the squared amplitudes of the entire spectrum, calculate the weighted average of frequency and amplitude as the centroid frequency of the spectrum, and divide the low and high frequency bands according to the preset frequency boundary value and calculate the ratio to obtain the component frequency domain parameter set; S403: Call the component frequency domain parameter set, perform item-by-item difference and relative deviation calculation on the parameter vectors corresponding to multiple components, compare the consistency of the total power, spectral centroid frequency and power ratio results, and generate frequency domain feature analysis results.

[0012] As a further aspect of the present invention, the frequency boundary value is statistically analyzed based on the frequency axis range corresponding to the component spectrum sequence. The amplitude square distribution corresponding to multiple frequency points in the spectrum is accumulated and sorted to form a cumulative power distribution. 50% of the total power value is calculated as the energy median threshold. The frequency point in the cumulative power distribution that reaches the energy median threshold is located and determined as the frequency boundary value.

[0013] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Based on the frequency domain feature analysis results, extract the total power numerical sequence corresponding to multiple components, perform normalization and rearrangement of the total power according to the component index, calculate the difference in total power of adjacent components and form a distribution vector, perform concentration judgment on the distribution vector and record it, and generate a total power distribution statistical vector. S502: Based on the total power distribution statistical vector and calling the spectral centroid frequency data in the frequency domain feature analysis results, perform sorting operation on the spectral centroid frequencies according to the component order, calculate the frequency offset direction of adjacent sorting positions and encode the change symbol to obtain the spectral centroid frequency sorting consistency identifier. S503: Call the spectral centroid frequency sorting consistency identifier and obtain the low-frequency and high-frequency power ratio change direction sequence. Combine the time-domain stability sequence of the time-domain stable data, perform a weighted summation of the power ratio direction sequence and the time-domain stability sequence to generate the multi-domain feature analysis result of skin electric field activity.

[0014] An automatic skin conductance signal analysis system based on multi-domain features includes: The data correction module collects skin conductance activity signals, calculates the median absolute deviation within a fixed time window, filters out abnormal data, replaces it with the median of the corresponding window, generates corrected skin conductance activity time series data, and transmits it to the time domain stabilization module. The temporal stabilization module performs filtering calculations based on the corrected skin conductance activity time series data, downsamples the filtering results according to a preset sampling ratio, calculates the temporal stability, generates temporal stable data, and transmits it to the multi-sequence fusion module. The multi-sequence fusion module calculates a smoothed analysis sequence based on the time-domain stable data by performing a centered moving average, and performs low-frequency and differential operations on the smoothed analysis sequence to obtain low-frequency and differential change sequences. It then integrates all sequences to generate multi-component joint analysis results and transmits them to the frequency domain feature module. The frequency domain feature module performs spectrum calculation based on the multi-component joint analysis results, calculates the total power, spectral centroid frequency, and low-frequency to high-frequency power ratio of the multi-components, compares the consistency of the frequency domain features of the multiple components, generates frequency domain feature analysis results, and transmits them to the multi-domain fusion module. The multi-domain fusion module, based on the frequency domain feature analysis results, extracts the total power distribution, the frequency order of the spectral centroid, and the direction of change of the power ratio between low and high frequencies to determine the consistency of direction, and then fuses and weights the results with the time domain stability to generate multi-domain feature analysis results of skin conductance activity.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, robust statistical constraints are introduced within a time window to correct abnormal fluctuations, and the stability of the signal is characterized during filtering and downsampling. This ensures that the temporal sequence of electrodermal activity remains consistent under noise and drift conditions. A multi-component joint characterization is constructed by combining the low-frequency evolution and differential changes of the smoothed sequence. Then, consistency analysis is performed on the power distribution and spectral centroid of each component in the frequency domain. This achieves mutual verification of time-domain and frequency-domain information, reduces dependence on empirical thresholds, and improves the stability and comparability of analysis results under different physiological states. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0019] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0020] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0021] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0022] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0023] Please see Figure 1 This invention provides an automatic analysis method for skin conductance activity signals based on multi-domain features, comprising the following steps: S1: Collect skin conductance activity signals, calculate the median absolute deviation within a fixed time window and filter out abnormal data, replace them with the median of the corresponding window, and generate corrected skin conductance activity time series data; S2: Based on the corrected skin conductance activity time series data, filter calculation is performed, the filter result is downsampled according to the preset sampling ratio, and the time domain stability is calculated to generate time domain stable data; S3: Based on the time-domain stable data, the smoothing analysis sequence is calculated by the center moving average, and the smoothing analysis sequence is subjected to low-frequency and differential operations to obtain the low-frequency and differential change sequence. All sequences are integrated to generate multi-component joint analysis results. S4: Perform spectrum calculation based on the results of multi-component joint analysis, calculate the total power, spectral centroid frequency, and low-frequency to high-frequency power ratio of the multi-components, compare the consistency of frequency domain characteristics of multiple components, and generate frequency domain characteristic analysis results; S5: Based on the frequency domain feature analysis results, extract the total power distribution, the frequency order of the spectral centroid, and the direction of change of the power ratio between low and high frequencies to determine the consistency of direction, and then fuse and weight the results with the time domain stability to generate the multi-domain feature analysis results of skin electrical activity.

[0024] The corrected electrodermal activity (EDA) time-series data includes the window midpoint sequence, abnormal substitution location index, and corrected amplitude time series. The time-domain stable data includes the downsampled amplitude sequence, the mean index of adjacent differences, and the stability status indicator. The multi-component joint analysis results include the smoothing analysis sequence, the low-frequency change sequence, and the differential change sequence. The frequency domain feature analysis results include the total power value of the multi-components, the spectral centroid frequency set, and the low-frequency to high-frequency power ratio parameter. The multi-domain feature analysis results of EDA include the consistency indicator of the total power distribution and the fusion judgment result of the power ratio change direction.

[0025] Please see Figure 2 The specific steps of S1 are as follows: S101: Collect skin conductance signals, split the signal sequence into fixed time windows, perform sorting operations on the sampled values ​​within multiple windows, select the median value of the sequence as the statistical benchmark, and calculate the absolute value of the difference between the multiple sampled values ​​and the benchmark value to generate the window absolute deviation statistic. In the practical application scenario of pilot simulated flight stress monitoring, a high-precision skin conductance sensor using silver / silver chloride electrodes (Ag / AgCl) is first activated, with a sampling frequency set to 1000 Hz to continuously collect skin conductance activity signals from the pilot's fingertips. The analog electrical signals are converted into digital sequences and stored in a high-speed buffer in real time. A fixed time window segmentation operation is performed, with the window length set to 200 milliseconds. This window length is selected based on the physiological time constant of human skin conductance response (typically 1 to 3 seconds) and the frequency characteristics of transient noise (such as millisecond-level spikes generated by electrode micro-movements), aiming to ensure the local stationarity of the signal statistical characteristics within the window. For continuous time-series data in the buffer, it is divided into a series of non-overlapping subsequences, each containing 200 sampling points according to the 1000 Hz sampling rate. For any given time window to be processed, all sampled values ​​within that window are read, and a quicksort algorithm is executed to arrange the 200 voltage values ​​in ascending order. After sorting, the center of the sorted sequence is located. Since the sequence length is even, the arithmetic mean of the 100th and 101st digits is selected and established as the statistical benchmark (i.e., median) for this window. Then, a difference operation is performed, subtracting the statistical benchmark from each original sampled value within the window and taking the absolute value of the subtraction result to generate an absolute deviation sequence containing 200 non-negative values. To quantify the dispersion of the data within this window and avoid interference from extreme outliers, the absolute deviation sequence is sorted again, and the median value after sorting is selected as the absolute deviation statistic (i.e., MAD value) for this window. For example, if a sampled value is 3.50 microsiemens and the median calculated for this window is 3.60 microsiemens, then the absolute deviation corresponding to that point is 0.10 microsiemens. All split time windows are traversed, and the absolute deviation statistic for each window is calculated and stored sequentially as a dynamic benchmark for subsequent anomaly detection. This process does not rely on the normal distribution setting of the data and can effectively deal with common non-Gaussian noise and asymmetric motion artifact interference in electrodermal signals.

[0026] S102: Based on the absolute deviation statistic of the window, call the skin conductance activity signal sampling value within the corresponding time window, calculate the absolute value of the difference between the multiple sampling values ​​and the median value. If the absolute value of the difference is greater than the product of the absolute deviation statistic and the set ratio, it is determined to be abnormal, and the abnormal sampling index set is obtained. The calculated absolute deviation statistic of the window is used, combined with a preset anomaly multiplier coefficient, to calculate the dynamic anomaly judgment threshold for the current time window. In practice, the anomaly multiplier coefficient is set to 5. This coefficient setting has undergone rigorous experimental verification: in a comparative experiment of static benchmark testing and simulated vigorous hand movement testing, skin conductance data from 50 different subjects in static and dynamic states were collected. Statistical results showed that the signal fluctuation amplitude caused by physiological skin conductance responses (such as a surge in skin conductance caused by fright) is usually distributed within the range of the median of the benchmark value plus or minus 3 times the absolute deviation statistic, while the amplitude deviation of instantaneous spike noise caused by poor electrode contact or mechanical pulling generally exceeds 5 times the absolute deviation statistic. Therefore, setting the coefficient to 5 can effectively distinguish between physiological signal changes and mechanical noise. A multiplication operation is performed, multiplying the absolute deviation statistic of the current window by 5 to obtain the deviation tolerance value. Subsequently, threshold boundary definition is performed to construct the judgment range of the closed interval, with the lower limit being the statistical benchmark value minus the deviation tolerance value, and the upper limit being the statistical benchmark value plus the deviation tolerance value. Each raw EEG sample value within the time window is iterated and compared to the judgment range. If a sample value is less than the lower limit or greater than the upper limit, its index position in the original time series is immediately locked, and it is marked as an anomaly. Continuing with the previous example, the median of this window is set to 3.60 microsiemens, the calculated absolute deviation statistic is 0.10 microsiemens, the deviation tolerance is 0.50 microsiemens, and the judgment range is 3.10 microsiemens to 4.10 microsiemens. If the value of the 4th sample point in this window is 12.50 microsiemens (caused by momentary contact jitter), since 12.50 is greater than 4.10, index 4 is added to the anomaly sampling index set. This logic is executed for all windows, ultimately generating an anomaly sampling index set containing the locations of all detected noise points. This set precisely points to the time coordinates that require numerical correction, providing a precise address basis for subsequent data cleaning.

[0027] S103: Based on the abnormal sampling index set, call the original sampled values ​​of skin electrophysiological activity signals within the corresponding time window, locate the abnormal index and perform a value replacement operation. The replacement value is selected as the median of the non-abnormal sampling sequence within the same window. The replaced sequence is recombined according to time to generate corrected skin electrophysiological activity time series data. Based on the output abnormal sampling index set, the data correction process is initiated. First, a new data buffer is allocated in memory to construct the corrected EEG activity time-series data. For each abnormal location recorded in the index set, the time window to which the abnormal point belongs is traced back. To ensure the physiological rationality of the replacement values ​​and maintain the local trend of the signal, all sampling points within the window that are not marked as abnormal are selected, forming a non-abnormal sampling sequence. Subsequently, the non-abnormal sampling sequence is sorted and its median is calculated to obtain the net median of the window. The median of the non-abnormal sequence is chosen here instead of directly using the median of the entire window to completely eliminate the slight influence of extreme outliers on the replacement benchmark. A numerical replacement operation is performed, directly overwriting the values ​​at the abnormal index positions in the original data sequence (such as the aforementioned 12.50 microSiemens) with the calculated net median. For index positions not marked as abnormal, their original sampling values ​​are directly retained. After completing all replacement operations within a single window, the processed window data are seamlessly spliced ​​together in chronological order to reconstruct continuous corrected EEG activity time-series data. Taking a specific embodiment as an example, for a 200-millisecond window containing artificially introduced noise, the original sampled data includes sequence segments with values ​​of 3.50, 3.60, 3.50, 12.50, and 3.70 microsiemens. First, the median of the entire window is calculated to be 3.60 microsiemens, and the absolute deviation statistic is 0.10 microsiemens, thus determining the upper limit for anomaly judgment as 4.10 microsiemens. Since the 12.50 microsiemens in the sequence is greater than the upper limit of 4.10 microsiemens, this point is judged as an anomaly. Subsequently, non-abnormal data (i.e., 3.50, 3.60, 3.50, 3.70, etc.) within the same window are extracted, and the non-abnormal sequences are sorted (sorted to 3.50, 3.50, 3.60, 3.70), and their median is calculated as the average of 3.50 and 3.60, which is 3.55 microsiemens. Finally, the outlier of 12.50 microSiemens was replaced with 3.55 microSiemens to generate corrected time-series data. This result demonstrates that a robust statistical method based on absolute deviation statistics can accurately locate and repair outliers in high-sampling-rate data, thereby eliminating the potential interference of high-amplitude noise on subsequent power spectral density analysis while preserving the overall baseline level of the signal.

[0028] Please see Figure 3 The specific steps of S2 are as follows: S201: Based on the corrected electrodermal activity time series data, perform time-domain smoothing filtering on the continuous sampling point sequence, call the amplitude of adjacent sampling points and perform weighted accumulation for smoothing constraint processing, and rearrange the processed sequence according to the time index to obtain the filtered electrodermal activity time series sequence; The system receives the generated corrected electrodermal activity (EDA) time-series data and initiates frequency domain suppression. This operation aims to separate effective physiological components and filter out high-frequency noise, using a fourth-order Butterworth low-pass filter algorithm. The filter cutoff frequency is set to 1 Hz, and the sampling frequency is set to 1000 Hz. Filter coefficients are pre-calculated using the bilinear transform method, generating a feedforward coefficient array and a feedback coefficient array containing five elements. During execution, a recursive difference equation operation is performed on the corrected time-series data: for the current output value, the input amplitudes of the current time and the past four time points are multiplied by the corresponding elements in the feedforward coefficient array with weights; simultaneously, the output amplitudes of the past four time points are multiplied by the corresponding elements in the feedback coefficient array with weights; and the two sets of weighted results are summed to obtain the preliminary filtered value for the current time. Subsequently, a zero-phase filtering operation is performed. First, the preliminary data sequence obtained through the aforementioned recursive operation is rearranged in reverse order by time index, placing the last data point at the beginning and the first data point at the end. Then, the same recursive filtering operation is performed again on the reversed sequence. Finally, the results of the secondary filtering are reversed and restored to their original time order. Taking a small high-frequency jump in the input sequence as an example, if the original data oscillates repeatedly between 3.550 μSiemens and 3.560 μSiemens at a frequency of 50 Hz, after bidirectional filtering with a cutoff frequency of 1 Hz, the corresponding values ​​in the output sequence are smoothly constrained into a gentle curve close to the mean, stabilizing around 3.555 μSiemens. The final output is a filtered electrodermal time series. The amplitude-frequency characteristic of this series attenuates by 3 dB at 1 Hz, and due to the bidirectional filtering, the signal phase does not shift.

[0029] S202: Based on the filtered EEG time series, call the preset sampling ratio parameter, perform equidistant sampling operation on the time index, retain the corresponding sampling amplitude, complete the time scale compression process, and generate a downsampled EEG time series. The filtered EEG time series output is retrieved, currently maintaining a sampling rate of 1000 Hz. A preset sampling ratio parameter is set to 10, meaning the goal is to convert the sampling rate to 100 Hz. An equidistant decimation operation is performed, first initializing the storage space of the downsampled EEG time series. Using the original time index as a reference, a decimation step size of 10 is set. Starting from index 0, the sampling amplitudes at multiples of 10 (index 0, index 10, index 20, etc.) are extracted sequentially and mapped to the new index positions in the downsampled sequence. During this process, data points located in the middle positions of indices 1-9 and 11-19 in the original sequence are discarded and do not proceed to the next stage of processing. For example, a 1-second data segment containing 1000 data points in the original sequence, after decimation at a ratio of 10, results in a sequence containing 100 data points. Specifically, if the amplitude at time point 0.010 (corresponding to original index 10) in the original sequence is 3.551 microsiemens, this value is directly retained and written to index 1 of the downsampled sequence; while the amplitude of 3.548 microsiemens at time point 0.009 (corresponding to original index 9) is discarded. This step completes the time-scale compression process, generating a downsampled electrodermal time series sequence. The time resolution of this sequence is transformed to 10 milliseconds, and the data volume is reduced to one-tenth of the original. Table 1 details the numerical changes of data at some key time points during the filtering and downsampling process, with the "Retention Mark" column clearly indicating the selected data points.

[0030] S203: For downsampled electrodermal skin time series, perform difference calculation on the amplitude of adjacent sampling points, process the absolute value of the difference result, collect all differences and perform mean calculation. If the mean result of the difference does not reach the time domain stability threshold, it is determined to be a stable state value and generate time domain stable data. For the generated downsampled EEG time series, temporal stability calculation and data generation are performed. A sliding window length of 50 sampling points, corresponding to a duration of 0.5 seconds, is set and slid across the sequence point by point. For each sliding window, the amplitudes of 50 consecutive sampling points within the window are first extracted. The difference between adjacent sampling points is calculated, i.e., the value of the later sampling point minus the value of the previous sampling point, and the absolute value of all calculated differences is taken. Subsequently, the arithmetic mean of these 49 absolute differences is calculated and defined as the window mean variability. A preset temporal stability threshold is applied, set to 0.01 microsiemens. This threshold is set based on the fact that the baseline drift velocity of EEG in a non-stressed state is usually extremely low; if the average change between adjacent points exceeds this value, it often indicates significant motion artifacts or sudden physiological excitation. The calculated window mean variability is compared with the temporal stability threshold: if the window mean variability is less than 0.01 microsiemens, the signal within that window is considered to be in a stable state. After determining a stable state, the arithmetic mean of the 50 sampled amplitudes within the window is calculated to generate the time-domain stable data corresponding to that moment. Continuing with the aforementioned data, if the downsampled data within a window slowly varies between 3.550 μSiemens and 3.552 μSiemens, the calculated average absolute difference between adjacent points is 0.002 μSiemens. This result of 0.002 μSiemens is compared with the threshold of 0.01 μSiemens. Since 0.002 is less than 0.01, the stability condition is met, and the mean of the data within that window (e.g., 3.551 μSiemens) is calculated and used as the final output time-domain stable data. Table 1 lists the specific numerical changes of the data during the filtering, downsampling, and stability determination processes.

[0031] Table 1 Data on Electrodermal Signal Filtering, Downsampling, and Temporal Stabilization ; As shown in Table 1, the correction data output in step S103 exhibits slight fluctuations within the time interval from 0.000 seconds to 0.020 seconds; for example, the value is 3.560 microsiemens at 0.010 seconds. After the 1 Hz low-pass filtering in step S201, the value at this point is smoothed and constrained to 3.551 microsiemens, eliminating high-frequency jitter. In step S202, based on a sampling ratio of 10, only the data at indices 0, 10, and 20 are retained, generating a downsampling sequence with a frequency of 100 Hz. Finally, in step S203, based on the average variability within the window, the average absolute difference calculated in this example is approximately 0.001 microsiemens. This value is less than the preset threshold of 0.01 microsiemens, therefore the signal is determined to be stable, and the window mean of 3.551 microsiemens is output as the time-domain stable data. This result indicates that the data after filtering and stability verification accurately reflects the baseline level of the electrodermal signal and removes noise interference.

[0032] Please see Figure 4 The specific steps of S3 are as follows: S301: Based on time-domain stable data, a centered sliding window is called for the values ​​corresponding to the continuous time index. The values ​​within the window are weighted and summed and divided by the window length to obtain the values ​​of the smoothed sequence. The values ​​are then rearranged according to the original time index order to generate a smoothed analysis sequence. The system receives the output EEG signal sequence containing time-domain stable data and initiates the centered sliding window smoothing logic. First, a floating-point array of the same length as the input sequence is allocated in memory to store the smoothed analysis sequence. The sliding window length is set to 5. This value of 5 corresponds to a 0.05-second time span at a 100Hz sampling rate; choosing an odd length ensures the window has a unique center time index, allowing the smoothed values ​​to be precisely aligned to the original time point and preventing phase shift. The weighting coefficient array is defined as an array of all 1s, employing an equal-weight strategy to perform unbiased mean smoothing. During execution, a loop variable is initialized and iterates from the second index position to the second-to-last index position in the sequence. For each specific time index, five consecutive sampling points in the input sequence, centered at that index and extending two positions before and after it, are locked. The floating-point values ​​at these five memory addresses are read sequentially and accumulated. Taking the data point corresponding to time index 0.030 seconds as an example, the numerical sets of this moment and the two sampling points before and after it are extracted, and the values ​​of this local sequence are set as 3.550 microsiemens, 3.551 microsiemens, 3.552 microsiemens, 3.553 microsiemens, and 3.554 microsiemens, respectively. The arithmetic logic unit performs an addition operation, that is, adds the above five values ​​together to obtain 17.760 microsiemens. Then, the window length parameter 5 is called to perform a division operation on the cumulative sum, and 3.552 microsiemens is calculated. This calculation result is immediately written to the storage unit corresponding to the current index in the smoothing analysis sequence. For the boundary points at both ends of the sequence, a boundary copying and padding strategy is adopted, that is, the closest effective calculated value is used to fill in the boundary points to keep the sequence length unchanged. Finally, the complete smoothing analysis sequence is output in the original time index order. This sequence retains the peak shape of the skin conductance response while filtering out residual small quantization noise.

[0033] S302: Based on the smoothing analysis sequence, perform low-frequency component extraction operation on the overall numerical distribution of the sequence, perform difference calculation on the corresponding values ​​of adjacent time indices, and maintain the correspondence of time indices to generate low-frequency analysis sequence and difference change sequence; The generated smoothing analysis sequence is invoked, and low-frequency baseline extraction and high-frequency rate of change calculation are performed respectively to separate the steady-state and transient components in the electrodermal signal. First, a second-order Butterworth low-pass filter is configured for acquiring the low-frequency analysis sequence. The cutoff frequency parameter is set to 0.05 Hz, which is based on the fact that the physiological drift cycle of human electrodermal levels is usually relatively long. Based on a sampling rate of 100 Hz, the recursive coefficients of the filter are pre-calculated through bilinear transformation. A recursive operation pipeline is established to perform time-domain difference equation operations on the smoothing analysis sequence: for the current output value, the current and past input values ​​are multiplied with the feedforward coefficients, and the past output values ​​are multiplied with the feedback coefficients. These products are then algebraically summed. The filter is set to be initialized and in the steady-state tracking phase, and the current input smoothing value is 3.552 microsiemens. Due to the extremely low cutoff frequency, the baseline value of the filter output will closely follow the long-term mean of the signal. If the signal has recently fluctuated slightly around 3.550 microsiemens, then the low-frequency component value calculated at the current moment is approximately 3.550 microsiemens, which is stored in the low-frequency analysis sequence. Next, for the acquisition of the differential change sequence, a first-order backward differential operation is performed. A traversal pointer is set, starting from the first index of the sequence. For each index, the value of the current index and the value of the previous index are read simultaneously from the smoothing analysis sequence. A subtraction instruction is executed, that is, the value at the current moment is subtracted from the value at the previous moment. Continuing with the aforementioned data, if the smoothing value at 0.030 seconds is 3.552 microsiemens, and the smoothing value at 0.020 seconds is 3.551 microsiemens, then the difference is calculated to be 0.001 microsiemens. This positive value indicates that the signal is at a rising edge, corresponding to an increase in sweat gland secretion activity. The calculated 0.001 microsiemens is written to the corresponding position in the differential change sequence. By traversing the entire sequence, we finally obtained the low-frequency analysis sequence reflecting baseline drift and the differential change sequence reflecting instantaneous excitation rate.

[0034] S303: Call the smoothing analysis sequence, low-frequency analysis sequence and differential change sequence, perform alignment mapping on the three types of sequences based on time index consistency, aggregate the corresponding index position values ​​in parallel, and generate multi-component joint analysis results; The output smoothing analysis sequence, low-frequency analysis sequence, and difference change sequence are invoked to initiate the multidimensional feature alignment and aggregation program. First, the time length and index integrity of the three sequences are verified to ensure no data is missing. Then, a multi-column data container is initialized to store the multi-component joint analysis results. The primary key index is set to the timestamp, and data aggregation is performed row by row. For any identical time index, the corresponding values ​​are addressed and extracted from the memory blocks of the three source sequences: the total signal amplitude is extracted from the smoothing analysis sequence, the baseline amplitude from the low-frequency analysis sequence, and the rate of change amplitude from the difference change sequence. These three independent scalar values ​​are packaged into a feature vector and mapped to the corresponding row in the result container. As shown in Table 2, taking time point 0.030 seconds as an example, the read smoothing value is 3.552 microsiemens, the calculated low-frequency baseline value is 3.550 microsiemens, and the calculated difference value is 0.001 microsiemens. These three values ​​are aggregated side-by-side to form the joint analysis record for that moment. This aggregation operation not only integrates the time-domain form and frequency-domain components of the signal, but also directly correlates the signal's absolute level and trend. By traversing all time points, it completes the structured reorganization of data across all time periods, generating the final multi-component joint analysis results, and providing standardized multi-dimensional input data for subsequent decoding of psychophysiological states.

[0035] Table 2. Example of data from multi-component joint analysis. ; Table 2 shows the joint analysis data for specific time segments. At time 0.030, the smoothing analysis sequence value is 3.552 microsiemens, representing the actual skin conductance level after denoising; the low-frequency analysis sequence value is 3.550 microsiemens, representing the physiological baseline at this time, i.e., the steady-state component, whose numerical change is lagging and gradual, consistent with low-pass filtering characteristics; the differential change sequence value is 0.001 microsiemens, representing the increment of the signal at this time relative to the previous time, i.e., 0.020 seconds, indicating the rising phase of the transient skin conductance response, i.e., the transient component. This result demonstrates that a single skin conductance signal was successfully decomposed into multidimensional numerical features describing different physiological mechanisms, and that each feature is strictly aligned on the time axis.

[0036] Please see Figure 5 The specific steps of S4 are as follows: S401: Based on the results of multi-component joint analysis, obtain the numerical sequence of the corresponding time index of the multi-component, perform discrete spectrum transformation on the time series data of each component, map the time domain values ​​to the frequency axis and form an amplitude distribution, retain the correspondence between frequency resolution and index, and generate the component spectrum sequence. The generated multi-component joint analysis results are retrieved, and three sets of numerical arrays—a smoothed analysis sequence aligned by time index, a low-frequency analysis sequence, and a differential variation sequence—are extracted and subjected to discrete spectrum transformation. To capture the frequency domain characteristics of short-term physiological signals, the window length parameter for spectrum analysis is set to 1024 sampling points, corresponding to a duration of 10.24 seconds at a sampling rate of 100 Hz. A sliding strategy with a 50% overlap rate is used to truncate data segments, and a Hamming window function is applied to each truncated time-domain data segment to suppress spectral leakage. The Hamming window coefficients are calculated by subtracting the product of 0.46 and a cosine term from a constant of 0.54; the independent variable of the cosine term depends on the relative position of the sampling points within the window. Subsequently, a fast Fourier transform is performed on the windowed sequence to map the time-domain amplitude to the frequency axis. Based on a sampling rate of 100 Hz and a transform length of 1024 points, the generated frequency resolution is approximately 0.0976 Hz. The magnitude of the Fast Fourier Transform (FFT) result is calculated, generating amplitude spectrum sequences for each component, where the frequency index covers the Nyquist frequency range from 0 Hz to 50 Hz. Taking the differential transformation sequence as an example, within a 10.24-second window, this sequence contains a fluctuating signal with a frequency of 0.2 Hz and an amplitude of 0.05 microsiemens. After transformation, the spectral amplitude near 0.195 Hz corresponding to the frequency index will exhibit a significant peak. Finally, three independent component spectrum sequences are output, representing the energy distribution in the frequency domain of the original signal's overall trend, baseline drift, and instantaneous rate of change.

[0037] S402: Based on the component spectrum sequence, perform integration and weighting on the spectrum amplitude and frequency, calculate the total power as the sum of the squared amplitudes of the entire spectrum, calculate the weighted average of frequency and amplitude as the centroid frequency of the spectrum, and divide the low and high frequency bands according to the preset frequency boundary value and calculate the ratio to obtain the component frequency domain parameter set; Based on the output component spectrum sequence, frequency domain characteristic parameters are calculated for each component to construct a component frequency domain parameter set. Three core characteristic indicators are defined: total power, spectral centroid frequency, and high-low frequency power ratio. First, the total power is calculated by summing the squares of the amplitudes across the entire frequency band (0 to 50 Hz), i.e., calculating and summing the squares of the amplitudes at each frequency point. Second, the spectral centroid frequency is calculated, reflecting the center position of the spectral energy. The calculation process involves multiplying the frequency value of each frequency point by its corresponding amplitude and summing the results, then dividing the sum by the total amplitude of all frequency points. Finally, a preset frequency boundary value of 0.2 Hz is applied, based on the physiological characteristic that the main energy of the skin conductance response is concentrated in the 0.05 to 0.5 Hz frequency band. The cumulative power values ​​in the low-frequency band (0 to 0.2 Hz) and the high-frequency band (0.2 to 5 Hz) are calculated separately, and the ratio is obtained by dividing the high-frequency power by the low-frequency power. Taking the spectral data of a differential variation sequence as an example, if the calculated total power is 0.0045 square microsiemens, and the energy is mainly concentrated around 0.3 Hz, resulting in a calculated spectral centroid frequency of 0.28 Hz, with a high-frequency power of 0.0030 and a low-frequency power of 0.0015, then the high-low frequency power ratio is 2.0. This calculation process quantifies the frequency domain energy structure of the signal components. Table 3 lists the calculated results of the frequency domain parameters for each component.

[0038] Table 3. Calculation results of component frequency domain parameters ; As shown in Table 3, the smoothed analysis sequence, due to the inclusion of a DC baseline, has an extremely high total power of 12.8500 and an extremely low centroid frequency of 0.015 Hz, indicating that the energy is mainly concentrated in the region close to 0 Hz. The low-frequency analysis sequence, after being low-pass filtered at 0.05 Hz, has a high-to-low frequency power ratio of 0.000, verifying that the high-frequency components were effectively filtered out. While the differential variation sequence has a small total power of 0.0045, its spectral centroid frequency is 0.280 Hz, and the high-to-low frequency power ratio is as high as 2.000, clearly revealing that this component is mainly composed of rapidly changing signals above 0.2 Hz, consistent with the frequency domain characteristics of transient skin conductance responses.

[0039] S403: Call the component frequency domain parameter set, perform item-by-item difference and relative deviation calculation on the parameter vectors corresponding to multiple components, compare the consistency of the total power, spectral centroid frequency and power ratio results, and generate frequency domain feature analysis results; The generated component frequency domain parameter set is invoked, and the parameter vectors corresponding to the smoothed analysis sequence and the low-frequency analysis sequence are subjected to term-by-term difference and relative deviation calculations to quantify the frequency domain contribution of transient components to the overall signal, generating frequency domain feature analysis results. First, two sets of parameter vectors are extracted: the smoothed sequence parameter vector and the low-frequency sequence parameter vector. Each set of vectors contains three components: total power, spectral centroid frequency, and high-low frequency power ratio. The total power difference is calculated by subtracting the total power of the low-frequency sequence from the total power of the smoothed sequence. This difference physically represents the pure transient response power after baseline removal. Next, the relative power deviation rate is calculated by dividing the calculated total power difference by the total power of the low-frequency sequence and converting the result into a percentage to assess the proportion of stress response intensity relative to the baseline level. Simultaneously, the spectral centroid frequency shift is calculated by subtracting the spectral centroid frequency of the low-frequency sequence from the spectral centroid frequency of the smoothed sequence. A positive frequency shift indicates the introduction of high-frequency stress components into the signal. Substituting the actual data from Table 3, the total power of the smoothed sequence is 12.8500, and the total power of the low-frequency sequence is 12.8000. The power difference is calculated to be 12.8500 - 12.8000 = 0.0500 square microsiemens. The relative power deviation rate is calculated as 0.0500 / 12.8000 * 100% = 0.39%. The spectral centroid frequency shift is calculated as 0.015 - 0.005 = 0.010 Hz. This series of calculations shows that although the transient response accounts for a small proportion of the total energy, only 0.39%, it introduces a significant rightward shift in the spectrum, i.e., an increase in the centroid frequency. This small variation in the frequency domain structure is accurately captured and serves as a key frequency domain characteristic for determining whether the driver experiences latent psychological stress.

[0040] Please see Figure 6 The specific steps of S5 are as follows: S501: Based on the frequency domain feature analysis results, extract the total power numerical sequence corresponding to multiple components, perform normalization and rearrangement of the total power according to the component index, calculate the difference of total power of adjacent components and form a distribution vector, perform concentration judgment on the distribution vector and record it, and generate a total power distribution statistical vector. The generated frequency domain feature analysis results are invoked. First, the total power storage area in memory is located, and the total power values ​​corresponding to the smoothed analysis sequence, low-frequency analysis sequence, and differential change sequence are extracted to construct the original power vector. Based on the calculation data from the previous steps, the actual values ​​are substituted: smoothed sequence power 12.8500, low-frequency sequence power 12.8000, and differential sequence power 0.0045. To eliminate dimensional differences and assess the relative proportion of each component in the total energy, the normalization operation logic is initiated. First, the total power is calculated by adding the above three values, i.e., 12.8500 + 12.8000 + 0.0045 = 25.6545 square microsiemens. Then, element-wise division is performed, i.e., the original power value of each component is divided by the total power to obtain the normalized power vector. The specific calculations are as follows: the proportion of the smooth sequence is 12.8500 / 25.6545 = 0.5009; the proportion of the low-frequency sequence is 12.8000 / 25.6545 = 0.4989; and the proportion of the difference sequence is 0.0045 / 25.6545 = 0.0002. Next, the normalized vector is sorted in descending order. The quicksort algorithm is called to rearrange the values ​​from largest to smallest, generating an ordered sequence of 0.5009, 0.4989, and 0.0002, corresponding to the smooth component, low-frequency component, and difference component. Based on this, the total power difference between adjacent components is calculated. The first step is performed: 0.5009 - 0.4989 = 0.0020; the second step is performed: 0.4989 - 0.0002 = 0.4987. These two differences constitute the distribution vector. Finally, a concentration assessment is performed, quantifying the non-uniformity of energy distribution by calculating the standard deviation of the distribution vector. First, the mean of the distribution vector is calculated: (0.0020 + 0.4987) / 2 = 0.25035. Then, the variance is calculated by squareding the difference between each value and the mean, then taking the average, and finally taking the square root of this average, yielding a standard deviation of approximately 0.2484. This standard deviation of 0.2484 is recorded as the core indicator of the total power distribution statistical vector, indicating that energy is highly concentrated in the first two components—the smooth and low-frequency components—while the differential component, representing transient changes, has a very low energy proportion but exhibits significant discontinuities with other components. This distribution characteristic is typical of high signal-to-noise ratio physiological signals.

[0041] S502: Based on the total power distribution statistical vector and calling the spectral centroid frequency data in the frequency domain feature analysis results, sort the spectral centroid frequencies according to the component order, calculate the frequency offset direction of adjacent sorting positions and encode the change sign to obtain the spectral centroid frequency sorting consistency identifier. Based on the determined component sorting order (smooth component, low-frequency component, differential component), the corresponding spectral centroid frequency data are retrieved from the frequency domain feature analysis results, and frequency-level logical verification is performed. First, an array of centroid frequencies to be sorted is established, substituting the values ​​calculated in previous step S402: smooth component frequency 0.015 Hz, low-frequency component frequency 0.005 Hz, and differential component frequency 0.280 Hz. An ascending sort operation is performed on this array, resulting in a rearranged frequency sequence of 0.005 Hz, 0.015 Hz, and 0.280 Hz, with the corresponding component physical meaning order changed to low-frequency component, smooth component, and differential component. Subsequently, the frequency offset direction of adjacent sorting positions is calculated. The first comparison step: subtracting the first position's value of 0.005 from the second position's value of 0.015 yields a positive 0.010, with a positive encoding sign. The second comparison step: subtracting 0.015 from the third position's value of 0.280 yields a positive 0.265, with a positive encoding sign. Combining these two codes generates a spectral centroid frequency ordering consistency identifier, which consists of two consecutive positive signs. This all-positive sequence has a clear physical meaning: it verifies that the frequency domain hierarchy of the signal decomposition conforms to the physiological law that the baseline drift frequency is less than the average frequency of the mixed signal, which is less than the transient change frequency. To quantify this consistency, a consistency scoring function is defined: if all the identifiers are positive, the score is 1.0; if an inversion (negative sign) occurs, 0.5 is deducted for each occurrence. In this embodiment, because the sequence is strictly monotonically increasing, the calculated consistency score is 1.0. This result confirms that the multi-component data within the current analysis window is a valid physiological signal without aliasing in the frequency domain structure, ruling out the anomaly of low-frequency and high-frequency component inversion caused by motion artifacts.

[0042] S503: Call the spectral centroid frequency sorting consistency identifier and obtain the low-frequency and high-frequency power ratio change direction sequence. Combine the time-domain stability sequence of the time-domain stable data, perform a weighted summation of the power ratio direction sequence and the time-domain stability sequence, and generate the multi-domain feature analysis results of skin electric field activity. The generated spectral centroid frequency sorting consistency score (1.0) is used, and the change direction sequence of the low-frequency and high-frequency power ratio is obtained. The power ratio of the low-frequency sequence is known to be 0.000, and the power ratio of the difference sequence is 2.000, with the change direction being 2.000 - 0.000 = 2.000, indicating a significant increase in the proportion of high-frequency energy along the component evolution direction. Simultaneously, the temporal stability value generated in step S203 is read. According to the record, the temporal stability output by step S203 for the current time window is 3.551. At this point, a multi-domain feature fusion vector is constructed, containing the consistency score 1.0, the power ratio increment 2.0, and the temporal stability 3.551. A consistency fusion judgment is performed, setting a judgment threshold set: the consistency score must be greater than 0.8, the power ratio increment must be greater than 1.5, and the stability must be greater than 0.9. The logical operation unit performs comparisons: 1.0 greater than 0.8 satisfies the conditions, 2.0 greater than 1.5 satisfies the conditions, and 3.551 greater than 0.9 satisfies the conditions. If all three conditions are met simultaneously, the logic output is true. Finally, the multi-domain feature analysis results of skin conductance activity are generated through weighted aggregation. The formula is: 0.4 multiplied by the consistency score, plus 0.3 multiplied by the normalized power ratio increment (the normalization factor is set to 2.0 here), plus 0.3 multiplied by the normalized temporal stability (to eliminate the influence of physical dimensions, the temporal stability normalization factor is set to 4.0 here). The specific calculation is: 0.4×1.0 +0.3×(2.0 / 2.0)+0.3×(3.551 / 4.0)=0.4+0.3+0.3×0.88775=0.4+0.3+0.2663, and the final result is 0.966 with three decimal places. Table 4 lists the final multi-domain feature fusion analysis results. The normalized high confidence score (close to 1.0) indicates that the driver's psychological stress response at approximately 0.03 seconds was accurately identified.

[0043] Table 4. Results of Multi-Domain Feature Fusion Analysis ; As shown in Table 4, by fusing the frequency domain structural consistency value of 1.000, the energy distribution significance value of 2.000, and the time domain signal quality value of 3.551, and after uniform dimensional normalization, the final output score is 0.966. This result verifies the reliability and accuracy of the multi-domain feature fusion method in recognizing complex physiological signals.

[0044] Please see Figure 7 An automatic skin conductance signal analysis system based on multi-domain features, comprising: The data correction module collects skin conductance activity signals, calculates the median absolute deviation within a fixed time window, filters out abnormal data, replaces it with the median of the corresponding window, generates corrected skin conductance activity time series data, and transmits it to the time domain stabilization module. The temporal stabilization module performs filtering calculations based on the corrected skin conductance activity time series data, downsamples the filtering results according to the preset sampling ratio, calculates the temporal stability, generates temporal stable data, and transmits it to the multi-sequence fusion module. The multi-sequence fusion module calculates the smoothing analysis sequence based on the central moving average of the time-domain stable data, performs low-frequency and differential operations on the smoothing analysis sequence to obtain the low-frequency and differential change sequence, integrates all sequences, generates multi-component joint analysis results, and transmits them to the frequency domain feature module. The frequency domain feature module performs spectrum calculations based on the joint analysis results of multiple components, calculates the total power, spectral centroid frequency, and low-frequency to high-frequency power ratio of multiple components, compares the consistency of frequency domain features of multiple components, generates frequency domain feature analysis results, and transmits them to the multi-domain fusion module. The multi-domain fusion module, based on the frequency domain feature analysis results, extracts the total power distribution, the frequency order of the spectral centroid, and the direction of change of the power ratio between low and high frequencies to determine the consistency of direction. It then fuses and weights the results with the time domain stability to generate the multi-domain feature analysis results of skin conductance activity.

[0045] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An automatic analysis method for skin conductance activity signals based on multi-domain features, characterized in that, Includes the following steps: S1: Collect skin conductance activity signals, calculate the median absolute deviation within a fixed time window and filter out abnormal data, replace them with the median of the corresponding window, and generate corrected skin conductance activity time series data; S2: Based on the corrected skin conductance activity time series data, perform filtering calculations, downsample the filtering results according to a preset sampling ratio, calculate the time domain stability, and generate time domain stable data; S3: Based on the time-domain stable data, perform a centered moving average to calculate the smoothing analysis sequence, and perform low-frequency and differential operations on the smoothing analysis sequence to obtain the low-frequency and differential change sequence. Integrate all sequences to generate multi-component joint analysis results. S4: Based on the results of the multi-component joint analysis, perform spectrum calculation, calculate the total power, spectral centroid frequency, and low-frequency to high-frequency power ratio of the multi-components, compare the consistency of the frequency domain characteristics of the multiple components, and generate frequency domain characteristic analysis results; S5: Based on the frequency domain feature analysis results, extract the total power distribution, the frequency order of the spectral centroid, and the direction of change of the power ratio between low and high frequencies to determine the consistency of direction, and then fuse and weight the results with the time domain stability to generate the multi-domain feature analysis results of skin electrical activity.

2. The automatic analysis method for skin conductance activity signals based on multi-domain features according to claim 1, characterized in that, The corrected electrodermal activity time-series data includes the window midpoint sequence, the abnormal substitution location index, and the corrected amplitude time series. The time-domain stable data includes the downsampled amplitude sequence, the mean value index of adjacent differences, and the stability status indicator. The multi-component joint analysis results include the smoothing analysis sequence, the low-frequency change sequence, and the difference change sequence. The frequency domain feature analysis results include the total power value of the multi-components, the spectral centroid frequency set, and the low-frequency to high-frequency power ratio parameter. The multi-domain feature analysis results of electrodermal activity include the total power distribution consistency indicator and the fusion judgment result of the power ratio change direction.

3. The automatic analysis method for skin conductance activity signals based on multi-domain features according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Collect skin conductance signals, split the signal sequence into fixed time windows, perform sorting operations on the sampled values ​​within multiple windows, select the median value of the sequence as the statistical benchmark, and calculate the absolute value of the difference between the multiple sampled values ​​and the benchmark value to generate the window absolute deviation statistic. S102: Based on the absolute deviation statistic of the window, call the skin conductance activity signal sampling value within the corresponding time window, calculate the absolute value of the difference between the multiple sampling values ​​and the median value, and if the absolute value of the difference is greater than the product of the absolute deviation statistic and the set ratio, it is determined to be abnormal and the abnormal sampling index set is obtained. S103: Based on the abnormal sampling index set, call the original sampling values ​​of the skin conductance activity signal within the corresponding time window, locate the abnormal index and perform a value replacement operation. The replacement value is selected as the median of the non-abnormal sampling sequence within the same window. The replaced sequence is recombined according to time to generate corrected skin conductance activity time series data.

4. The automatic analysis method for skin conductance activity signals based on multi-domain features according to claim 1, characterized in that, The specific steps of S2 are as follows: S201: Based on the corrected electrodermal activity time series data, perform time-domain smoothing filtering on the continuous sampling point sequence, call the amplitude of adjacent sampling points and perform weighted accumulation for smoothing constraint processing, and rearrange the processed sequence according to the time index to obtain the filtered electrodermal activity time series sequence; S202: Based on the filtered skin electrophysiological time series, call the preset sampling ratio parameter, perform equidistant sampling operation on the time series index, retain the corresponding sampling amplitude, complete the time scale compression process, and generate a downsampled skin electrophysiological time series. S203: For the downsampled electrodermal time series, perform difference calculation on the amplitude of adjacent sampling points, process the absolute value of the difference result, collect all differences and perform mean calculation. If the mean result of the difference does not reach the time domain stability threshold, it is determined to be a stable state value, and time domain stable data is generated.

5. The automatic analysis method for skin conductance activity signals based on multi-domain features according to claim 4, characterized in that, The time-domain stability threshold is determined by analyzing the amplitude difference between adjacent sampling points in the downsampled electrodermal time sequence, and by analyzing the standard deviation of the mean of all absolute differences and a preset multiple.

6. The automatic analysis method for skin conductance activity signals based on multi-domain features according to claim 1, characterized in that, The specific steps for S3 are as follows: S301: Based on the time-domain stable data, call the centered sliding window for the values ​​corresponding to the continuous time index, perform weighted summation on the values ​​in the window and divide by the window length to obtain the values ​​of the smoothed sequence, and rearrange them according to the original time index order to generate a smoothed analysis sequence. S302: Based on the smoothing analysis sequence, perform low-frequency component extraction operation on the overall numerical distribution of the sequence, perform difference calculation on the corresponding values ​​of adjacent time indices, and maintain the correspondence of time indices to generate a low-frequency analysis sequence and a difference change sequence; S303: Call the smoothing analysis sequence, the low-frequency analysis sequence, and the differential change sequence, perform alignment mapping on the three types of sequences based on time index consistency, and aggregate the corresponding index position values ​​to generate multi-component joint analysis results.

7. The automatic analysis method for skin conductance activity signals based on multi-domain features according to claim 1, characterized in that, The specific steps of S4 are as follows: S401: Based on the results of the multi-component joint analysis, obtain the numerical sequence of the time index of the multi-component, perform discrete spectrum transformation on the time series data of each component, map the time domain values ​​to the frequency axis and form an amplitude distribution, retain the correspondence between frequency resolution and index, and generate the component spectrum sequence. S402: Based on the component spectrum sequence, perform integration and weighting on the spectrum amplitude and frequency, calculate the total power as the sum of the squared amplitudes of the entire spectrum, calculate the weighted average of frequency and amplitude as the centroid frequency of the spectrum, and divide the low and high frequency bands according to the preset frequency boundary value and calculate the ratio to obtain the component frequency domain parameter set; S403: Call the component frequency domain parameter set, perform item-by-item difference and relative deviation calculation on the parameter vectors corresponding to multiple components, compare the consistency of the total power, spectral centroid frequency and power ratio results, and generate frequency domain feature analysis results.

8. The automatic skin conductance signal analysis method based on multi-domain features according to claim 7, wherein the frequency boundary value is statistically analyzed based on the frequency axis range corresponding to the component spectrum sequence, the amplitude square distribution corresponding to multiple frequency points in the spectrum is accumulated and sorted to form a cumulative power distribution, 50% of the total power value is calculated as the energy median threshold, the frequency point in the cumulative power distribution that reaches the energy median threshold is located, and it is determined as the frequency boundary value.

9. The automatic analysis method for skin conductance signals based on multi-domain features according to claim 1, characterized in that, The specific steps of S5 are as follows: S501: Based on the frequency domain feature analysis results, extract the total power numerical sequence corresponding to multiple components, perform normalization and rearrangement of the total power according to the component index, calculate the difference in total power of adjacent components and form a distribution vector, perform concentration judgment on the distribution vector and record it, and generate a total power distribution statistical vector. S502: Based on the total power distribution statistical vector and calling the spectral centroid frequency data in the frequency domain feature analysis results, perform sorting operation on the spectral centroid frequencies according to the component order, calculate the frequency offset direction of adjacent sorting positions and encode the change symbol to obtain the spectral centroid frequency sorting consistency identifier. S503: Call the spectral centroid frequency sorting consistency identifier and obtain the low-frequency and high-frequency power ratio change direction sequence. Combine the time-domain stability sequence of the time-domain stable data, perform a weighted summation of the power ratio direction sequence and the time-domain stability sequence to generate the multi-domain feature analysis result of skin electric field activity.

10. An automatic skin conductance signal analysis system based on multi-domain features, characterized in that, The system is used to implement the automatic skin conductance signal analysis method based on multi-domain features as described in any one of claims 1-9, the system comprising: The data correction module collects skin conductance activity signals, calculates the median absolute deviation within a fixed time window, filters out abnormal data, replaces it with the median of the corresponding window, generates corrected skin conductance activity time series data, and transmits it to the time domain stabilization module. The temporal stabilization module performs filtering calculations based on the corrected skin conductance activity time series data, downsamples the filtering results according to a preset sampling ratio, calculates the temporal stability, generates temporal stable data, and transmits it to the multi-sequence fusion module. The multi-sequence fusion module calculates a smoothed analysis sequence based on the time-domain stable data by performing a centered moving average, and performs low-frequency and differential operations on the smoothed analysis sequence to obtain low-frequency and differential change sequences. It then integrates all sequences to generate multi-component joint analysis results and transmits them to the frequency domain feature module. The frequency domain feature module performs spectrum calculation based on the multi-component joint analysis results, calculates the total power, spectral centroid frequency, and low-frequency to high-frequency power ratio of the multi-components, compares the consistency of the frequency domain features of the multiple components, generates frequency domain feature analysis results, and transmits them to the multi-domain fusion module. The multi-domain fusion module, based on the frequency domain feature analysis results, extracts the total power distribution, the frequency order of the spectral centroid, and the direction of change of the power ratio between low and high frequencies to determine the consistency of direction, and then fuses and weights the results with the time domain stability to generate multi-domain feature analysis results of skin conductance activity.

Images