Physiological signal emotion recognition method and equipment based on AI glasses and medium

Abstract

The invention discloses a physiological signal emotion recognition method and device based on AI glasses and a medium, and relates to the technical field of emotion recognition, and the method comprises the steps: carrying out the motion artifact compensation of a photoelectric volume pulse wave signal according to a segmented compensation plan package, obtaining a segmented pure photoelectric volume pulse wave signal and a compensation residual quality index, and carrying out the motion artifact compensation of the photoelectric volume pulse wave signal; integrating into a quality control data packet; based on the quality control data packet, carrying out pulse wave crest positioning, physiological feature calculation and stability index acquisition, and carrying out alignment summarization on the compensation residual quality index to generate a physiological feature stability evaluation packet; and according to the physiological feature stability evaluation packet, combining a rule layer and a machine learning layer, constructing an emotion judgment model, executing emotion recognition and uncertainty control, and generating an emotion result record. According to the invention, high reliability and precision of emotion recognition in a dynamic environment are improved.

Description

Technical Field

[0001] This invention relates to the field of emotion recognition technology, and in particular to a method, device, and medium for physiological signal emotion recognition based on AI glasses. Background Technology

[0002] With the deep integration of artificial intelligence and wearable technology, AI glasses, as a new generation of intelligent terminal devices, have demonstrated significant technological advantages and application prospects in the fields of health monitoring and emotion recognition. In recent years, physiological signal analysis technology based on photoplethysmography (PPG) has made systematic progress. By extracting core physiological features such as heart rate variability and pulse wave transit time with high precision, it has achieved dynamic inference and quantitative assessment of emotional states. At the same time, the mature application of six-axis inertial measurement units has provided reliable technical support for the real-time, high-resolution acquisition of head motion data, promoting the innovative practice of multimodal data fusion strategies in dynamic environments.

[0003] However, existing technologies have two key limitations in dynamic compensation for motion artifacts and real-time quality control of emotion judgment models: First, the motion artifact compensation mechanism lacks dynamic adaptability, compensating only based on global motion intensity, and cannot identify the morphological features of head movements, resulting in inaccurate segmented compensation and residual artifacts interfering with pulse peak localization, thus affecting the computational stability of physiological features such as HRV; Second, the emotion judgment model based on pattern recognition lacks real-time quality control and uncertainty control, and does not integrate signal stability indicators and compensation residual quality, resulting in significant fluctuations in recognition results under motion conditions, and the uncertainty cannot be quantified, making it difficult to meet the reliability requirements of clinical or daily applications. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a physiological signal emotion recognition method based on AI glasses to solve the problems of insufficient dynamic compensation for motion artifacts and lack of real-time quality control of emotion judgment models.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a physiological signal emotion recognition method based on AI glasses, comprising,

[0008] Acquire photoplethysmography (PPG) signals and six-axis inertial data, perform preprocessing, and generate a synchronous preprocessed acquisition dataset.

[0009] Head motion intensity and motion morphology features are extracted from the synchronous preprocessed dataset, and motion categories are identified. Segmented compensation configuration is then performed to generate a segmented compensation plan package.

[0010] Motion artifact compensation is performed on the photoplethysmography (PPG) signal according to the segmented compensation plan package to obtain the segmented pure PPG signal and the quality index of the compensation residual, and integrate them into a quality control data package.

[0011] Based on the quality control data package, pulse wave peak localization is performed, physiological characteristics are calculated, stability indices are obtained, and they are aligned and summarized with the compensation residual quality indices to generate a physiological characteristic stability assessment package.

[0012] Based on the physiological characteristic stability assessment package, and combined with the rule layer and machine learning layer, an emotion judgment model is constructed, and emotion recognition and uncertainty control are performed to generate emotion result records.

[0013] As a preferred embodiment of the physiological signal emotion recognition method based on AI glasses described in this invention, the steps of acquiring photoplethysmography (PPG) signals and six-axis inertial data, and preprocessing them to generate a synchronously preprocessed acquisition dataset are as follows.

[0014] Acquire photoplethysmography (PPG) signals and six-axis inertial data, and add time stamps to generate the original synchronous acquisition dataset;

[0015] Interpolation and gap marking are performed on the original synchronously acquired dataset, and trajectory records are retained and replaced for repeated sampling points to generate a gap-repaired acquisition dataset;

[0016] DC drift removal is performed on the photoplethysmography (PPG) signal in the gap repair acquisition dataset, and gravity component removal is performed on the six-axis inertial data to generate a synchronous preprocessed acquisition dataset.

[0017] As a preferred embodiment of the physiological signal emotion recognition method based on AI glasses described in this invention, the specific steps for extracting head motion intensity and motion morphology features from the synchronously preprocessed acquisition dataset are as follows:

[0018] Based on the synchronous preprocessed acquisition dataset, the dataset is sliced ​​according to the sliding time window, a time window index is constructed, and amplitude clipping and zero bias correction are performed to generate a time window inertial segment dataset.

[0019] Head motion intensity and motion morphology features are extracted from the time window inertial segment dataset and bound according to the time window index to generate a motion feature data table.

[0020] As a preferred embodiment of the physiological signal emotion recognition method based on AI glasses described in this invention, the specific steps for generating the segmented compensation plan package are as follows:

[0021] The motion feature data table is normalized and outlier truncation is performed to obtain standardized motion features. Motion category identification and state switching annotation are then performed to generate a motion category annotation data table.

[0022] Based on the motion category labeling data table, hysteresis determination and de-jitter merging are performed to obtain a segmented index set, and segmented compensation configuration is performed to generate a segmented compensation plan package.

[0023] As a preferred embodiment of the physiological signal emotion recognition method based on AI glasses described in this invention, the steps of performing motion artifact compensation on the photoplethysmography (PPG) signal according to the segmented compensation plan package, obtaining segmented pure PPG signal and compensation residual quality index, and integrating them into a quality control data package are as follows.

[0024] Based on the segmented compensation scheme package, the six-axis inertial data and photoplethysmography pulse wave signal in the synchronous preprocessed acquisition dataset are rearranged into a noisy reference sequence, and then sliced ​​and paired to generate a segmented aligned input set;

[0025] The protection time window is extracted from the segmented aligned input set, amplitude limiting and abrupt segment shielding are performed, and normalization and direction locking are performed on the non-protection time window to generate a segmented compensation execution set;

[0026] Based on the segmented compensation execution set, motion artifact compensation is performed on the photoplethysmography signal to obtain segmented pure photoplethysmography signal and compensation residual quality index.

[0027] Time-stamped splicing is performed on segmented pure photoplethysmography (PPG) signals, and weighted fusion is performed at segment boundaries. Simultaneously, the data is aligned with the quality index of the compensation residual to generate a quality control data package.

[0028] As a preferred embodiment of the physiological signal emotion recognition method based on AI glasses described in this invention, the steps include: localizing pulse peaks based on quality control data packets, calculating physiological characteristics, obtaining stability indices, and aligning and summarizing these with compensation residual quality indices to generate a physiological characteristic stability evaluation package. The specific steps are as follows:

[0029] Based on the quality control data package, pulse wave peak localization is performed, and local extreme peak candidates are extracted. At the same time, minimum interval constraints and period consistency constraints are applied for screening to generate peak event sequences.

[0030] The time intervals between adjacent peaks are extracted from the peak event sequence and organized into a pulse interval sequence. At the same time, abnormal interval removal, interpolation annotation and physiological feature window clipping are performed to calculate physiological features.

[0031] A cross-window fluctuation assessment is performed on physiological characteristics to obtain stability indices, which are then aligned with the quality indices of the compensated residuals to generate a physiological characteristic stability assessment package.

[0032] As a preferred embodiment of the physiological signal emotion recognition method based on AI glasses described in this invention, the specific steps for constructing the emotion judgment model are as follows:

[0033] The rule layer is used to receive the physiological feature stability assessment package, perform stability gating screening, obtain a set of stable window segments, perform physiological feature binning mapping, and generate rule feature vectors.

[0034] The machine learning layer is used to input the regular feature vectors into the lightweight neural network, perform physiological feature weight reduction labeling and physiological feature enhancement labeling, and output the emotion category and confidence level;

[0035] By jointly assembling the rule layer and the machine learning layer, a sentiment judgment model is constructed.

[0036] As a preferred embodiment of the physiological signal emotion recognition method based on AI glasses described in this invention, the specific steps for generating the emotion result record are as follows:

[0037] Based on the emotion judgment model, emotion recognition is performed on the physiological characteristic stability assessment package to obtain candidate emotion conflicts, and priority locking is performed to generate candidate emotion records.

[0038] Based on candidate emotion records, uncertainty control and hierarchical conservative output are implemented to generate emotion result records.

[0039] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the physiological signal emotion recognition method based on AI glasses as described in the first aspect of the present invention.

[0040] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the physiological signal emotion recognition method based on AI glasses as described in the first aspect of the present invention.

[0041] The beneficial effects of this invention are as follows: by generating a segmented compensation plan package, accurate compensation for dynamic motion artifacts is achieved, which improves the purity of photoplethysmography pulse wave signals and the stability of physiological feature calculation; by constructing a two-layer emotion judgment model through pattern recognition of the rule layer and machine learning layer, real-time quality control and uncertainty quantification of emotion recognition are achieved, which improves the high reliability and accuracy of emotion recognition in dynamic environments. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0043] Figure 1 This is a flowchart of a physiological signal emotion recognition method based on AI glasses.

[0044] Figure 2 The flowchart for generating a motion feature data table.

[0045] Figure 3 This is a flowchart for motion artifact compensation.

[0046] Figure 4 A flowchart for constructing an emotion judgment model. Detailed Implementation

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0048] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0049] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0050] Reference Figures 1-4 This is one embodiment of the present invention, which provides a physiological signal emotion recognition method based on AI glasses, including the following steps:

[0051] S1: Acquire photoplethysmography (PPG) signals and six-axis inertial data, perform preprocessing, and generate a synchronous preprocessed acquisition dataset;

[0052] S1.1: Acquire photoplethysmography (PPG) signals and six-axis inertial data, add time stamps, and generate the original synchronous acquisition dataset;

[0053] Specifically, in the nose pad area of ​​the AI ​​glasses, a photoplethysmography (PPG) sensor continuously collects PPG signals at a specific sampling rate (e.g., 100 Hz). Simultaneously, a six-axis inertial measurement unit integrated in the temples synchronously collects six-axis inertial data containing three-axis acceleration and three-axis angular velocity at a specific sampling rate (e.g., 50 Hz). At the start of the acquisition, the PPG signals and six-axis inertial data need to be synchronized in time. A time stamp accurate to the millisecond level is added to each sampling point, and the time-stamped PPG signals and six-axis inertial data are merged in chronological order to generate a raw synchronous acquisition dataset containing timestamps, PPG signal data, and six-axis inertial data.

[0054] S1.2: Perform interpolation and gap marking on the original synchronous acquisition dataset, and retain and replace trajectory records for repeated sampling points to generate a gap repair acquisition dataset;

[0055] Furthermore, based on a preset time interval threshold, the intervals between adjacent timestamps are checked and the timestamp gap positions and lengths are registered. For each registered timestamp gap, the starting and ending sampling points of the gap are locked and a gap boundary index is formed. Based on the gap boundary index, the timestamp sequence within the gap interval is interpolated and filled, and gap markers are written for the filled sampling points. At the same time, duplicate sampling points with the same timestamp or falling within the duplicate judgment tolerance range (e.g., 1 millisecond) are detected and summarized into a duplicate sampling point list. In the duplicate sampling point list, a retention marker is written according to the retention judgment (retaining the sampling point closest to the expected timestamp), and a replacement trajectory record is written for the unretained sampling points and associated with the sampling point identifier corresponding to the retention marker. The interpolated and filled sampling points, gap markers, retention markers and replacement trajectory records are merged, checked and rearranged according to timestamps to generate a gap repair acquisition dataset.

[0056] It should be noted that the preset time interval thresholds include the photoplethysmography (PPG) signal time interval threshold and the six-axis inertial data time interval threshold. The PPG signal time interval threshold is based on the expected sampling period of the PPG signal and combined with the time synchronization jitter tolerance definition. The six-axis inertial data time interval threshold is based on the expected sampling period of the six-axis inertial data and combined with the time synchronization jitter tolerance definition. For example, when the expected sampling period of the PPG signal is 10 milliseconds, the example range of the PPG signal time interval threshold is 12 milliseconds to 18 milliseconds. When the expected sampling period of the six-axis inertial data is 20 milliseconds, the example range of the six-axis inertial data time interval threshold is 24 milliseconds to 36 milliseconds.

[0057] S1.3: Perform DC drift removal on the photoplethysmography pulse wave signal in the gap repair acquisition dataset and remove the gravity component from the six-axis inertial data to generate a synchronous preprocessed acquisition dataset.

[0058] Specifically, a DC baseline reference sequence is extracted from the photoplethysmography (PPG) signal using a sliding time window, and this sequence is bound to a timestamp to form a baseline index. After aligning the PPG signal with the baseline index according to the timestamp, the DC baseline reference sequence is subtracted from each sample point to obtain the de-DC drifted PPG signal. The amplitude fluctuation of the six-axis inertial data within the sliding time window is checked. If the amplitude fluctuation is less than a preset fluctuation threshold (based on the upper limit of amplitude fluctuation of the pre-acquired and labeled six-axis inertial data within the sliding time window, with an example range of 0.02 to 0.10), the sliding time window is... The time window is registered as a stationary time window and used to extract the gravity component reference sequence and bind it with the timestamp to form a gravity index. The three-axis acceleration in the six-axis inertial data is aligned with the gravity index according to the timestamp and the gravity component reference sequence is subtracted point by point to obtain the gravity component removed three-axis acceleration. The three-axis angular velocity in the six-axis inertial data is retained according to the timestamp and merged with the gravity component removed three-axis acceleration at the same timestamp. The photoplethysmography pulse wave signal after DC drift removal is checked against the gravity component removed six-axis inertial data according to the timestamp consistency, and the notch mark and replacement trajectory record are retained and then merged and written to generate a synchronous preprocessed acquisition dataset.

[0059] S2: Extract head motion intensity and motion morphology features from the synchronous preprocessed acquisition dataset, identify the motion category, configure segmented compensation, and generate a segmented compensation plan package;

[0060] S2.1: Based on the synchronous preprocessed acquisition dataset, slice it according to the sliding time window, construct the time window index, and perform amplitude clipping and zero bias correction to generate the time window inertial segment dataset;

[0061] Specifically, based on the synchronously preprocessed acquisition dataset, six-axis inertial data is read in timestamp order, and a time window boundary sequence is generated by selecting the sliding time window length and sliding step size (e.g., the sliding time window length is 1 to 3 seconds, and the sliding step size is 0.2 to 1 second). The time window boundary sequence is then written with the time window number, time window start timestamp, and time window end timestamp to form a time window index. The time window index is then used to retrieve six-axis inertial data falling within the range of the time window start timestamp and time window end timestamp (determined by the sliding time window length). The data is encapsulated into time window segment entries. Simultaneously, amplitude clipping is performed on each sampling point of the time window segment entries. Six-axis inertial data values ​​exceeding the amplitude clipping upper and lower limits are replaced with the amplitude clipping upper and lower limits and clipping markers are registered (for example, the amplitude clipping upper and lower limits are 0.80 to 0.98 of the device range). Zero-bias reference segments are extracted from the amplitude-clipped time window segment entries and zero-bias reference values ​​are written. Zero-bias correction is completed by subtracting the zero-bias reference values ​​from each sampling point of the time window segment entries on each axis and zero-bias correction markers are registered, generating a time window inertial segment dataset.

[0062] S2.2: Extract head motion intensity and motion morphology features from the time window inertial segment dataset, and bind them according to the time window index to generate a motion feature data table;

[0063] Furthermore, the time window number, start timestamp, end timestamp, and corresponding time window segment entries are read one by one from the time window inertial segment dataset according to the time window index. The six-axis inertial data in the time window segment entries are organized into amplitude sequences and direction sequences according to the timestamp order. In the process of extracting head motion intensity and motion morphology features, the amplitude sequences are summarized into intervals and peak-valley segments are registered within the same time window segment entries to form a head motion intensity feature set. The direction sequences are registered into direction switching segments, rhythm segments, and morphological segments to form a head motion morphology feature set. The clipping mark and zero bias correction mark are retrieved synchronously from the time window segment entries and written into the feature source mark. The head motion intensity feature set, head motion morphology feature set, feature source mark, and time window index are aligned and solidified into a record row using the time window number as the binding key. All record rows are summarized in order of time window number to generate a motion feature data table.

[0064] S2.3: Perform feature normalization and outlier truncation on the motion feature data table to obtain standardized motion features, and perform motion category identification and state switching annotation to generate a motion category annotation data table;

[0065] Specifically, the motion feature data table is read row by row according to the time window number, and the values ​​of head motion intensity and motion morphology features corresponding to the time window index are read one by one. The values ​​of head motion intensity and motion morphology features are scaled and mapped according to a unified standardized range to obtain standardized motion features (e.g., 0 to 1). The upper and lower limits of the standardized motion features are checked column by column (based on the quantile range definition of the standardized motion feature values ​​in the motion feature data table, the example range is 0.02 to 0.98). Values ​​exceeding the upper limit are replaced with the upper limit, and values ​​below the lower limit are replaced with the lower limit to complete the outlier truncation. After the standardized motion features are written back to the corresponding time window number record row of the motion feature data table, the motion category judgment conditions of each row of standardized motion features are checked one by one and the motion category label is written. The motion category labels of adjacent time window number record rows are checked for continuity and the positions where changes are written into the state switching label. The time window index, standardized motion features, motion category label and state switching label are merged and written to generate the motion category label data table.

[0066] S2.4: Based on the motion category labeling data table, perform hysteresis judgment and de-jitter merging, obtain the segmented index set, configure segmented compensation, and generate a segmented compensation plan package.

[0067] Specifically, the state transition marker is registered as a candidate transition point while maintaining the motion category label before the transition. If the new motion category label is maintained for no less than the number of confirmation windows (e.g., 2 to 5 confirmation windows) and the confidence flag is not lower than the confidence threshold within the continuous time window after the candidate transition point, the candidate transition point is registered as a confirmed transition point. If the conditions are not met, the candidate transition point is revoked and the motion category label before the transition is restored. Confirmed transition points are aggregated according to the time window index to form segment entries and de-jitter merging is performed. Segment entries with a duration less than the shortest segment duration are merged into adjacent segment entries and the segment start and end information is updated (e.g., the shortest segment duration is 0.5 seconds to 2 seconds). The merged segment entries are written with the segment number and segment start and end information to form a segment index set. The segment index set is written to the segment compensation configuration one by one according to the motion category label and bound to the segment number to generate a segment compensation plan package.

[0068] It should be noted that the confidence threshold is based on the criteria defined in the sports category labeling data table for distinguishing between stable and unstable sports category labels, with an example range of 0.60 to 0.90.

[0069] S3: Perform motion artifact compensation on the photoplethysmography signal according to the segmented compensation plan package, obtain the segmented pure photoplethysmography signal and the quality index of the compensation residual, and integrate them into a quality control data package;

[0070] S3.1: Based on the segmented compensation plan package, the six-axis inertial data and photoplethysmography pulse wave signal in the synchronous preprocessed acquisition dataset are rearranged into a noisy reference sequence, and sliced ​​and paired to generate a segmented aligned input set;

[0071] Specifically, in the synchronous preprocessing acquisition dataset, six-axis inertial data and photoplethysmography (PPG) signals are extracted according to the segment start and end timestamp ranges. The six-axis inertial data is rearranged by timestamp within the segment start and end timestamp ranges and concatenated according to a predetermined dimensional order (meaning the six-axis inertial data is concatenated into a noise reference sequence according to a uniform dimensional arrangement to ensure different segments use the same arrangement). Simultaneously, a unified timestamp sequence is established, with the interval between adjacent timestamps determined by the PPG signal sampling period or the six-axis inertial data sampling period. Missing sampling points in the unified timestamp sequence are filled using linear interpolation, spline interpolation, or hold-at-the-moment interpolation and marked with the interpolation method. When multiple sampling points exist at the same timestamp position in a unified timestamp sequence for the volumetric pulse wave signal, they are written in segments. The noise reference sequence and the photoplethysmography (PPG) signal correspond one-to-one at the same timestamp position in the unified timestamp sequence. Synchronous slicing is performed on the noise reference sequence and the PPG signal according to a fixed segment length (meaning that the unified timestamp sequence is divided into segments with a fixed time span so that the noise reference sequence slices and the PPG signal slices have the same start and end timestamp range, such as 0.5 seconds to 2 seconds). The noise reference sequence slices and the PPG signal slices within the same segment start and end timestamp are paired. At the same time, the segment number, segment start timestamp, and segment end timestamp are bound to the paired slices and written. All paired slices are summarized to form a segment-aligned input set.

[0072] S3.2: Extract the protection time window from the segmented aligned input set, perform amplitude limiting and abrupt segment shielding, and perform normalization and direction locking on the non-protection time window to generate a segmented compensation execution set;

[0073] Furthermore, from the segmented aligned input set, segment start timestamps, segment end timestamps, noise reference sequence slices, and photoplethysmography (PPG) signal slices are read one by one according to segment number. Based on the amplitude fluctuations and direction reversal density of the noise reference sequence slices and the waveform continuity verification results of the PPG signal slices, a protection time window is registered, and the start and end timestamps of the protection time window are written (e.g., the protection time window length is 0.1 seconds to 0.5 seconds). Within the protection time window, amplitude limiting is performed on the noise reference sequence slices and PPG signal slices, and a limiting mark is registered. Sampling points exceeding the limiting boundary are replaced with the limiting boundary (e.g., the limiting boundary is 0.05 to 0.5 seconds). .95), within the protection time window, the jump amplitude of adjacent sampling points is checked and the interval exceeding the mutation constraint is registered as the mutation segment shielding interval. Sampling points within the mutation segment shielding interval are replaced with nearby stable sampling points and the mutation segment shielding mark is registered (e.g., the mutation constraint is 0.10 to 0.30). Within the non-protection time window, the noise reference sequence slice is normalized and the normalization mark is registered. At the same time, the direction lock is performed on the noise reference sequence slice and the direction lock mark is registered. The processed noise reference sequence slice and the processed photoplethysmography pulse wave signal slice, together with the start and end timestamps of the protection time window and each mark, are bound and summarized according to the start and end timestamps of the segments to generate a segmented compensation execution set.

[0074] S3.3: Based on the segmented compensation execution set, motion artifact compensation is performed on the photoplethysmography (PPG) signal to obtain segmented pure PPG signals and compensation residual quality indicators.

[0075] Specifically, the process involves reading the segment start timestamp, segment end timestamp, processed noise reference sequence slice, processed photoplethysmography (PPG) signal slice, and start and end timestamps of the protection time window according to segment number. The processed noise reference sequence slice is then aligned with the processed PPG signal slice by timestamp to form a compensation sequence. This compensation sequence is generated through a linear mapping at each sampling point. The three-axis acceleration and three-axis angular velocity channels in the noise reference sequence slice are read point by point and concatenated in a fixed channel order to form a noise vector. Based on the amplitude change direction of the noise vector and the processed PPG signal slice at the same timestamp, a sign consistency check is performed and a direction consistency marker is written. At the sampling point corresponding to the direction consistency marker, the amplitude of each channel of the noise vector is weighted and summed to obtain the compensation sequence sampling point value. At sampling points where the weights do not meet the amplitude constraints... The compensation sequence sampling point values ​​are replaced with adjacent stable sampling point values ​​and written into the replacement mark. Within the protection time window, the compensation sequence is subjected to amplitude limiting replacement and shielding replacement according to the amplitude limiting mark and the abrupt segment shielding mark. Within the non-protection time window, the compensation sequence is subjected to interval scaling and direction consistency according to the normalization mark and the direction locking mark. The processed photoplethysmography (PPG) signal slices are obtained by subtracting the compensation sequence from each sampling point. The difference between the processed PPG signal slices and the segmented pure PPG signal slices is written into the compensation residual sequence. The residual amplitude ratio and residual abrupt change ratio are checked according to the segment range and written into the compensation residual quality index (e.g., 0.0 to 1.0). The segmented pure PPG signal slices and the compensation residual quality index are summarized to obtain the segmented pure PPG signal and compensation residual quality index.

[0076] S3.4: Perform time-stamped splicing on the segmented pure photoplethysmography (PPG) signals, perform weighted fusion at the segment boundaries, and align with the compensation residual quality index to generate a quality control data package.

[0077] Specifically, time stamps are added to each sampling point in the sequence, and the sequences are spliced ​​together in ascending order of timestamps. The spliced ​​sequences are checked for time continuity at the end timestamps of adjacent segments and the start timestamps of adjacent segments, and gap or overlap marks are registered. At the segment boundaries, fusion intervals are defined according to the length of the fusion interval (e.g., 0.05 seconds to 0.20 seconds). Within the fusion interval, sampling points of pure photoplethysmography (PPG) signals from the left and right segments are taken simultaneously at the same timestamp. The left segment's weight coefficient is written in descending order, and the right segment's weight coefficient is written in ascending order, and the fused sampling points are synthesized and the fusion mark is registered. The compensation residual quality index is aligned to the segment start timestamp and segment end timestamp according to the segment number and written into the corresponding timestamp interval. Within the fusion interval, the compensation residual quality index is synthesized into a fusion quality mark with the same weight coefficient and the fusion mark is registered. The spliced ​​sequence, gap or overlap marks, fusion marks, and compensation residual quality index are merged and written to generate a quality control data package.

[0078] It should be noted that the weighting coefficient refers to the coefficient used to allocate the contribution ratio of the left segment pure photoplethysmography (PPG) signal to the right segment pure PPG signal within the fusion interval. The weighting coefficient is determined by the relative position of each timestamp within the fusion interval between the start and end timestamps of the fusion interval in a linearly varying manner.

[0079] S4: Based on the quality control data package, perform pulse wave peak localization, calculate physiological characteristics, obtain stability indicators, and align and summarize them with the compensation residual quality indicators to generate a physiological characteristic stability assessment package.

[0080] S4.1: Based on the quality control data package, perform pulse wave peak localization and extract local extreme peak candidates. At the same time, perform minimum interval constraint and period consistency constraint screening to generate peak event sequence;

[0081] Specifically, based on the spliced ​​segmented pure photoplethysmography (PPG) signals in the quality control data package, the amplitude change direction of adjacent sampling points is compared within the time interval where the compensation residual quality index meets the quality threshold. The position where the amplitude change direction changes from rising to falling is identified as a local extreme peak candidate, and the peak candidate timestamp and peak candidate amplitude are retained. After the local extreme peak candidates are sorted by timestamp, a minimum interval constraint is applied to the difference between the timestamps of adjacent peak candidates. Peak candidates with timestamp differences less than the minimum interval threshold are merged into a single peak candidate, and the peak candidate with the more prominent amplitude is retained as the retained peak. After the retained peaks form a candidate peak sequence in timetamp order, the period consistency screening is performed on the adjacent peak intervals of the candidate peak sequence. The candidate peaks that cause significant deviations in the peak interval are replaced with candidate peaks with more prominent amplitudes within the adjacent timetamp range to maintain rhythm continuity. The selected candidate peaks are organized into a peak event sequence in timetamp order.

[0082] It should be noted that the quality threshold is based on the criteria for distinguishing between available and unavailable time intervals using the compensated residual quality index, with an example range of 0.60 to 0.90.

[0083] The minimum interval threshold is based on the shortest allowable time interval between adjacent true peaks under normal pulse rhythm and combined with the peak candidate false detection tolerance definition, with an example range of 0.25 seconds to 0.60 seconds.

[0084] The adjacent timestamp range refers to the range of retrieval time windows that extend forward and backward centered on the timestamp of the replaced candidate peak.

[0085] S4.2: Extract the time interval between adjacent peaks from the peak event sequence and organize it into a pulse interval sequence. At the same time, perform abnormal interval removal, interpolation annotation and physiological feature window clipping, and calculate physiological features.

[0086] Furthermore, the timestamp differences between adjacent peak events are extracted from the peak event sequence in timestamp order and organized into records of adjacent peak time intervals. These records are then sorted by their sequence number. Each adjacent peak time interval record is compared with the physiologically acceptable range. Records exceeding the physiologically acceptable range are registered as abnormal intervals and removed (e.g., 0.30 seconds to 2.00 seconds). The time axis gaps in the peak event sequence corresponding to the abnormal intervals are interpolated using the valid adjacent peak time interval records at both ends of the gap, and interpolation labels are registered. The valid adjacent peak time interval records are then segmented according to a fixed window duration to form... Physiological feature windows are generated and window pruning is performed. Window pruning ensures that the proportion of valid adjacent peak time interval records within the physiological feature window is not less than the proportion threshold and the duration of the physiological feature window meets the window duration requirement (e.g., window duration of 30 to 120 seconds). Within each physiological feature window, the number of valid adjacent peak time interval records is counted, and physiological features are calculated based on the correspondence between adjacent peak time interval records, the sum of the mean, and the next adjacent peak time interval record. Physiological features include the average heart rate feature value, the dispersion feature value of adjacent peak time intervals, and the short-term fluctuation feature value of adjacent peak time intervals.

[0087] The formula for calculating physiological characteristics is:

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] in, Indicates physiological characteristics, This represents the characteristic value of average heart rate. This represents the characteristic value of the dispersion of the time interval between adjacent wave peaks. This represents the short-term fluctuation characteristic value of the time interval between adjacent peaks. This indicates the number of records of valid adjacent peak time intervals within the physiological characteristic window. The window representing the physiological characteristics of the first Records of time intervals between adjacent peaks. This represents the index number used to represent the sum of the means within the physiological characteristic window. The window representing the physiological characteristics of the first Record the time interval between adjacent peaks.

[0093] It should be noted that the percentage threshold is defined based on the minimum acceptable ratio of the number of valid adjacent peak time interval records within the physiological characteristic window to the total number of adjacent peak time interval records within the window, with an example range of 0.70 to 0.90.

[0094] S4.3: Perform cross-window fluctuation assessment on physiological characteristics, obtain stability indices, align them with compensated residual quality indices, and generate a physiological characteristic stability assessment package.

[0095] Specifically, physiological features are arranged in chronological order according to their windows, and the start and end timestamps of each window are locked. Values ​​of the same physiological feature in adjacent windows are compared item by item and summarized into cross-window fluctuation records. These records are then used to determine stability based on a preset fluctuation threshold. Physiological feature windows with cross-window fluctuations consistently below the threshold are marked as stable windows and a stability index is generated. Conversely, physiological feature windows with cross-window fluctuations exceeding the threshold are marked as fluctuating windows and a stability index is generated. Compensation residual quality indices are aligned to the start and end timestamps of each physiological feature window. Values ​​of the compensation residual quality indices are extracted within the window range and summarized into window quality markers. Stability indices and window quality markers are bound and compiled according to the physiological feature window number and the start and end timestamps. Finally, physiological features, stability indices, window quality markers, and the start and end timestamps are summarized into a physiological feature stability assessment package.

[0096] S5: Based on the physiological characteristic stability assessment package, combined with the rule layer and machine learning layer, construct an emotion judgment model, perform emotion recognition and uncertainty control, and generate emotion result records.

[0097] S5.1: The rule layer is used to receive the physiological feature stability assessment package, perform stability gating screening, obtain a set of stable window segments, perform physiological feature binning mapping, and generate rule feature vectors;

[0098] Specifically, after receiving the physiological characteristic stability assessment package, the rule layer expands the average heart rate characteristic value, the dispersion characteristic value of the time interval between adjacent peaks, the short-term fluctuation characteristic value of the time interval between adjacent peaks, the stability index, and the compensation residual quality index in the physiological characteristic stability assessment package according to the physiological characteristic window index. Stability gating screening is performed on the physiological characteristic window index. The stability gating screening performs stability index threshold verification on the stability index (comparing the stability index corresponding to the physiological characteristic window index with the stability index threshold and writing a pass or fail mark), and performs compensation residual quality threshold verification on the compensation residual quality index. The compensating residual quality index corresponding to the physiological feature window index is compared with the compensating residual quality threshold and written into a pass or fail flag. The duration of the physiological feature window and the percentage threshold of the effective adjacent peak time interval are verified simultaneously. The physiological feature window indexes that meet the gating conditions are registered and summarized into a stable window segment set. The rule layer extracts the average heart rate feature value, the dispersion feature value of the adjacent peak time interval and the short-term fluctuation feature value of the adjacent peak time interval from the stable window segment set, performs physiological feature binning mapping and writes the binning number. The rule layer encodes the binning number in a fixed field order and concatenates it into a rule feature vector.

[0099] It should be noted that the stability index threshold is based on the criteria for distinguishing stable and unstable window segments in the physiological characteristic stability assessment package, with an example range of 0.60 to 0.90.

[0100] The quality threshold for compensated residuals is defined in the physiological characteristics stability assessment package as a criterion for distinguishing between acceptable and unacceptable compensated residuals. The example range is 0.60 to 0.90.

[0101] Meeting the gating conditions means that the stability index meets the stability index threshold, the compensation residual quality index meets the compensation residual quality threshold, the physiological characteristic window duration meets the window duration requirement, and the proportion of records with effective adjacent peak time intervals meets the proportion threshold.

[0102] S5.2: The machine learning layer is used to input the regular feature vector into the lightweight neural network, perform physiological feature weight reduction labeling and physiological feature enhancement labeling, and output the emotion category and confidence level;

[0103] Specifically, after receiving the regular feature vectors, the machine learning layer performs field integrity checks and fixed field order checks. The machine learning layer groups the regular feature vectors according to the physiological feature window index and converts them into an input sequence that can be accepted by the lightweight neural network. Placeholders and alignment markers are written to the missing bin number positions in the input sequence to ensure that the length of the input sequence is consistent. The lightweight neural network performs one forward inference on the input sequence and uses dual output heads to synchronously output the annotation strength corresponding one-to-one with the bin number positions, as well as the emotion category and confidence bound to the physiological feature window index. The annotation strength is used to write physiological feature weight reduction annotations or physiological feature enhancement annotations at the corresponding positions of the regular feature vectors and retains the positions that have not been marked as unmarked positions. The emotion category and confidence are bound to the physiological feature window index for output (e.g., confidence is 0.60 to 0.95).

[0104] S5.3: Combine the rule layer and the machine learning layer to build a sentiment judgment model;

[0105] Furthermore, the emotion judgment model receives a physiological feature stability assessment package and uses the physiological feature window index as the binding key. The emotion judgment model executes the rule layer in a fixed calling order to generate rule feature vectors and retains the physiological feature window indexes corresponding to the stable window fragment set. Then, the rule feature vectors are passed to the machine learning layer according to the physiological feature window index to generate emotion categories and confidence scores, and retain physiological feature deweighting labels and physiological feature enhancement labels. The emotion judgment model aligns and verifies the output of the rule layer and the output of the machine learning layer one by one according to the physiological feature window index and verifies the consistency of field completeness and fixed field order. The emotion judgment model solidifies the aligned output relationship into a unified assembly link and registers the mapping relationship from input to output, thus constructing the emotion judgment model.

[0106] S5.4: Based on the emotion judgment model, perform emotion recognition on the physiological characteristic stability assessment package, obtain candidate emotion conflicts, perform priority locking, and generate candidate emotion records;

[0107] Specifically, based on the emotion judgment model, the physiological feature stability assessment package performs emotion recognition item by item according to the physiological feature window index. The consistency of emotion categories in adjacent physiological feature window indices is checked, and adjacent window pairs whose emotion categories have changed and whose confidence differences meet the conflict judgment conditions are registered as candidate emotion conflicts. At the same time, emotion category candidates with multiple emotion category candidates in the same physiological feature window index and similar confidence are registered as candidate emotion conflicts. Priority locking sorts each group of emotion category candidates in the candidate emotion conflict according to priority rules and selects the emotion category candidate with higher priority as the locked emotion category. At the same time, the suppressed emotion category candidates and their corresponding confidence are retained as conflict traces. The locked emotion category, locked confidence, candidate emotion conflict constituent entries, physiological feature window index and conflict traces are merged and written to generate candidate emotion records.

[0108] It should be noted that the conflict determination condition refers to the fact that the emotion categories corresponding to adjacent physiological feature window indices are inconsistent, and the difference in confidence between the two emotion category candidates falls within the conflict difference range and simultaneously meets the lower confidence limit requirement (for example, the conflict difference range is 0.05 to 0.20, and the lower confidence limit is 0.60 to 0.90).

[0109] The priority rule refers to sorting candidate emotional conflicts from high to low confidence to lock in the emotional category. When the confidence difference falls within the range of the conflict difference, it is sorted from high to low stability index. When the stability index is consistent, the stability window takes priority, followed by the fluctuation window. When the quality index of the compensation residual is consistent, it is sorted from long to short physiological characteristic window duration.

[0110] S5.5: Based on the candidate emotion records, perform uncertainty control and hierarchical conservative output to generate emotion result records.

[0111] Specifically, after sorting candidate emotion records by physiological feature window index, each record is retrieved for locking the emotion category, locking confidence level, and candidate emotion conflict traces. Uncertainty control compares the locking confidence level with the uncertainty control threshold, and combines the number of conflict entries and the number of conflict duration windows in the candidate emotion conflict traces to complete the classification. Records with locking confidence levels higher than the uncertainty control threshold and small numbers of conflict entries and conflict duration windows are registered as low uncertainty; records with locking confidence levels close to the uncertainty control threshold or with the number of conflict entries and conflict duration windows at an intermediate level are registered as medium uncertainty; records with locking confidence levels lower than the uncertainty control threshold are registered as low uncertainty. Records with large thresholds or a large number of conflict entries and conflict duration windows are registered as high uncertainty. For tiered conservative output, the emotion category and confidence level are locked when there is low uncertainty; the emotion category is locked and a conservative label is registered when there is medium uncertainty, and the output confidence level is adjusted to a conservative confidence level; when there is high uncertainty, the emotion category corresponding to the previous physiological feature window index is locked and a conservative label is registered, and the output confidence level is adjusted to a conservative confidence level (e.g., a conservative confidence level of 0.50 to 0.80). The physiological feature window index, output emotion category, output confidence level, uncertainty label level, and conservative label are merged and written to generate an emotion result record.

[0112] It should be noted that the uncertainty control threshold is defined based on the locked confidence level in the candidate emotion record, which is used to distinguish between emotion categories that can be directly output and those that require graded conservative output. The example range is 0.60 to 0.90.

[0113] This embodiment also provides a computer device applicable to the physiological signal emotion recognition method based on AI glasses, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the physiological signal emotion recognition method based on AI glasses as proposed in the above embodiment.

[0114] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0115] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the physiological signal emotion recognition method based on AI glasses as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0116] In summary, this invention achieves precise compensation for dynamic motion artifacts by generating a segmented compensation plan package, thereby improving the purity of photoplethysmography pulse wave signals and the stability of physiological feature calculations. Furthermore, it constructs a two-layer emotion judgment model through pattern recognition at the rule layer and machine learning layer, enabling real-time quality control and uncertainty quantification in emotion recognition, thus enhancing the reliability and accuracy of emotion recognition in dynamic environments.

[0117] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for emotion recognition based on physiological signals using AI glasses, characterized in that: include, Acquire photoplethysmography (PPG) signals and six-axis inertial data, perform preprocessing, and generate a synchronous preprocessed acquisition dataset. Head motion intensity and motion morphology features are extracted from the synchronous preprocessed dataset, and motion categories are identified. Segmented compensation configuration is then performed to generate a segmented compensation plan package. Motion artifact compensation is performed on the photoplethysmography (PPG) signal according to the segmented compensation plan package to obtain the segmented pure PPG signal and the quality index of the compensation residual, and integrate them into a quality control data package. Based on the quality control data package, pulse wave peak localization is performed, physiological characteristics are calculated, stability indices are obtained, and they are aligned and summarized with the compensation residual quality indices to generate a physiological characteristic stability assessment package. Based on the physiological characteristic stability assessment package, and combined with the rule layer and machine learning layer, an emotion judgment model is constructed, and emotion recognition and uncertainty control are performed to generate emotion result records.

2. The physiological signal emotion recognition method based on AI glasses as described in claim 1, characterized in that: The acquisition of photoplethysmography (PPG) signals and six-axis inertial data, followed by preprocessing to generate a synchronously preprocessed acquisition dataset, is detailed in the following steps. Acquire photoplethysmography (PPG) signals and six-axis inertial data, and add time stamps to generate the original synchronous acquisition dataset; Interpolation and gap marking are performed on the original synchronously acquired dataset, and trajectory records are retained and replaced for repeated sampling points to generate a gap-repaired acquisition dataset; DC drift removal is performed on the photoplethysmography (PPG) signal in the gap repair acquisition dataset, and gravity component removal is performed on the six-axis inertial data to generate a synchronous preprocessed acquisition dataset.

3. The physiological signal emotion recognition method based on AI glasses as described in claim 2, characterized in that: The specific steps for extracting head motion intensity and motion morphology features from the synchronously preprocessed dataset are as follows. Based on the synchronous preprocessed acquisition dataset, the dataset is sliced ​​according to the sliding time window, a time window index is constructed, and amplitude clipping and zero bias correction are performed to generate a time window inertial segment dataset. Head motion intensity and motion morphology features are extracted from the time window inertial segment dataset and bound according to the time window index to generate a motion feature data table.

4. The physiological signal emotion recognition method based on AI glasses as described in claim 3, characterized in that: The specific steps for generating the segmented compensation plan package are as follows: The motion feature data table is normalized and outlier truncation is performed to obtain standardized motion features. Motion category identification and state switching annotation are then performed to generate a motion category annotation data table. Based on the motion category labeling data table, hysteresis determination and de-jitter merging are performed to obtain a segmented index set, and segmented compensation configuration is performed to generate a segmented compensation plan package.

5. The physiological signal emotion recognition method based on AI glasses as described in claim 4, characterized in that: The process involves performing motion artifact compensation on the photoplethysmography (PPG) signal according to the segmented compensation plan package, obtaining segmented clean PPG signals and compensation residual quality indicators, and integrating them into a quality control data package. The specific steps are as follows: Based on the segmented compensation scheme package, the six-axis inertial data and photoplethysmography pulse wave signal in the synchronous preprocessed acquisition dataset are rearranged into a noisy reference sequence, and then sliced ​​and paired to generate a segmented aligned input set; The protection time window is extracted from the segmented aligned input set, amplitude limiting and abrupt segment shielding are performed, and normalization and direction locking are performed on the non-protection time window to generate a segmented compensation execution set; Based on the segmented compensation execution set, motion artifact compensation is performed on the photoplethysmography signal to obtain segmented pure photoplethysmography signal and compensation residual quality index. Time-stamped splicing is performed on segmented pure photoplethysmography (PPG) signals, and weighted fusion is performed at segment boundaries. Simultaneously, the data is aligned with the quality index of the compensation residual to generate a quality control data package.

6. The physiological signal emotion recognition method based on AI glasses as described in claim 5, characterized in that: The process involves locating the pulse wave peak based on the quality control data package, calculating physiological characteristics, obtaining stability indices, and aligning and summarizing these with the compensated residual quality indices to generate a physiological characteristic stability assessment package. The specific steps are as follows: Based on the quality control data package, pulse wave peak localization is performed, and local extreme peak candidates are extracted. At the same time, minimum interval constraints and period consistency constraints are applied for screening to generate peak event sequences. The time intervals between adjacent peaks are extracted from the peak event sequence and organized into a pulse interval sequence. At the same time, abnormal interval removal, interpolation annotation and physiological feature window clipping are performed to calculate physiological features. A cross-window fluctuation assessment is performed on physiological characteristics to obtain stability indices, which are then aligned with the quality indices of the compensated residuals to generate a physiological characteristic stability assessment package.

7. The physiological signal emotion recognition method based on AI glasses as described in claim 6, characterized in that: The specific steps for constructing the emotion judgment model are as follows. The rule layer is used to receive the physiological feature stability assessment package, perform stability gating screening, obtain a set of stable window segments, perform physiological feature binning mapping, and generate rule feature vectors. The machine learning layer is used to input the regular feature vectors into the lightweight neural network, perform physiological feature weight reduction labeling and physiological feature enhancement labeling, and output the emotion category and confidence level; By jointly assembling the rule layer and the machine learning layer, a sentiment judgment model is constructed.

8. The physiological signal emotion recognition method based on AI glasses as described in claim 7, characterized in that: The specific steps for generating the emotion result record are as follows: Based on the emotion judgment model, emotion recognition is performed on the physiological characteristic stability assessment package to obtain candidate emotion conflicts, and priority locking is performed to generate candidate emotion records. Based on candidate emotion records, uncertainty control and hierarchical conservative output are implemented to generate emotion result records.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the physiological signal emotion recognition method based on AI glasses as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the physiological signal emotion recognition method based on AI glasses as described in any one of claims 1 to 8.

Images