A feature-deconvoluted emotion decoding system combining EEG and ECG

Abstract

The present invention relates to the field of physiological signal processing technology, specifically a feature disentanglement emotion decoding system combining EEG and ECG, the system including a feature disentanglement module, a three-way dynamic interaction module, a feature semantic mapping module, an abnormal feature verification module, and a feature repair module. In the present invention, by performing multi-stage decomposition of EEG and ECG signals to identify and distinguish public features from invalid private features, the synchronous analysis of signal dimension and semantic dimension is achieved, and through statistical analysis of feature semantic deviation, stability and abnormal distribution, feature items inconsistent with task objectives are dynamically detected, and their semantic state evolution trend is clarified, which promotes semantic quality control at the feature level, realizes feature replacement and modeling update, and makes the performance of multimodal physiological signals in emotion recognition tasks have sensitivity, accuracy and adaptability, and provides a physiological feature expression system with clear structure, clear semantics and dynamic adjustment for emotional state recognition.

Description

Technical Field

[0001] The present invention relates to the field of physiological signal processing technology, and in particular to a feature-deconvoluted emotion decoding system combining electroencephalogram (EEG) and electrocardiogram (ECG). Background Art

[0002] The field of physiological signal processing technology involves the collection, analysis, and feature construction of multi-source physiological signals such as EEG, ECG, and EMG, aiming to explore the inherent physiological response patterns of individuals under different psychological and physiological states. The core content of this technical field includes time series alignment of cross-modal signals, feature attribution modeling, correlation pattern recognition, and trend classification mechanisms. Its systematic nature is reflected in the standardization of signal acquisition mechanisms, feature structure difference analysis, inter-modal collaborative modeling, and dynamic adaptation of signal-emotion mapping mechanisms. Physiological signal processing is widely used in scenarios such as emotion recognition, stress perception, and neural feedback. It integrates multimodal data to form task-oriented physiological feature expression models and is an important component of cognitive computing and human-computer interaction systems.

[0003] Among them, the feature disentanglement emotion decoding system combining EEG and ECG refers to an emotion recognition method based on joint modeling of EEG and ECG signals, distinguishing different feature dimensions and performing decoupling processing. This patent subject addresses the problems of cross-modal feature aliasing, modality attribution ambiguity, and unclear association between features and tasks. It covers the alignment of raw time slice data in EEG and ECG signals, extraction of private and public features and category attribution identification, establishment of an explicit discrimination mechanism for valid and invalid features in the feature space, construction of multi-dimensional interaction paths for joint feature reconstruction based on modality specificity, and generation of feature semantic comparison mapping relationships through physiological structure attribution rules, thus constructing a multi-dimensional feature analysis solution for emotion recognition tasks.

[0004] While existing technologies have established joint modeling frameworks for emotion recognition using multiple signal sources, such as EEG and ECG, practical applications still face challenges such as inaccurate semantic alignment between signal modalities, ambiguous feature distribution attribution, and poor feature-task adaptability. In conventional multimodal fusion, features are combined through splicing or alignment, lacking an explicit mechanism for distinguishing between public and private features. This leads to severe signal aliasing and the masking of public features by private feature noise. During the emotion label mapping process, the lack of feature-task semantic adaptation assessment prevents the accurate identification of key features highly relevant to the task, resulting in weak interpretability of the model's decision path and prone to overfitting or transfer failure. The lack of monitoring methods for feature temporal stability and semantic bias makes it difficult to detect interfering features in the recognition results, making recognition accuracy susceptible to fluctuations and limiting the system's generalization capabilities. For example, in an emotion induction task, ECG waveform changes for the same subject at multiple time intervals, while consistent with individual patterns, are currently treated as equivalent features, ignoring their differences from specific emotional states. This can lead to biased emotion attribution mechanisms and compromise model reliability. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the existing technology and propose a feature de-entanglement emotion decoding system that combines EEG and ECG.

[0006] To achieve the above objectives, the present invention adopts the following technical solutions: a feature-deconvoluted emotion decoding system combining EEG and ECG, comprising:

[0007] The feature disentanglement module performs multi-stage decomposition based on EEG and ECG signals to extract public features and invalid private features. It also separates public and private feature attributes through semantic association information between modalities to generate a multi-dimensional feature disentanglement table.

[0008] The three-way dynamic interaction module calls the feature distribution data in the multi-dimensional feature disentanglement table, constructs cross-modal, cross-dimensional and cross-feature interaction paths, identifies the complementary relationship between public features and the contribution of private features, and generates an emotional feature interaction sequence;

[0009] The feature semantic mapping module calls the feature items in the emotional feature interaction sequence, analyzes the modality specificity and task attribution mechanism, compares the matching degree between the feature semantics and the task goal, selects the features with priority matching degree and analyzes the semantic comparison relationship to generate a feature semantic mapping table;

[0010] The abnormal feature verification module extracts the feature semantic deviation value, feature stability index and abnormal identifier according to the feature distribution in the feature semantic mapping table, compares the degree of deviation between the semantic deviation and the task target, counts the abnormal feature names and frequencies, and generates an abnormal feature list.

[0011] As a further solution of the present invention, the multidimensional feature disentanglement table includes modal correspondence distribution, feature independence measurement, and attribute separation mapping structure; the emotional feature interaction sequence includes feature complementary distribution form, modal interaction structure style, and semantic connection strength index; the feature semantic mapping table includes semantic attribution category, matching priority index, and semantic comparison structure unit; the abnormal feature list includes feature name record, abnormal frequency statistics, and semantic offset mark content.

[0012] As a further solution of the present invention, the feature detangling module includes:

[0013] The feature decomposition submodule performs multi-stage decomposition based on EEG and ECG signals. After time alignment according to a unified sampling frequency and time window division method, it extracts the amplitude sequence and spectrum sequence of the signal segments, compares the amplitude difference, frequency component offset value and disturbance fluctuation degree between adjacent signal segments, and determines the segment aggregation relationship based on the amplitude difference and frequency offset to obtain the common features of the signal segments.

[0014] The attribute separation submodule calls the common features of the signal segments, combines the amplitude variation range, frequency concentration interval and phase variation trend of the EEG and ECG signals in the same time period, compares the amplitude variation degree and frequency concentration position change between the differentiated signals, and separates the cross-section and non-cross-section between the modes to obtain the modal cross-attribute difference;

[0015] The dimensional mapping submodule extracts the value range and change direction of the features under the common attribute set and the non-common attribute set based on the modal cross-attribute difference, combines the projection relationship and change rate distribution of the differentiated features in the coordinate axis mapping, adjusts the expression of the original features in multiple dimensions, and obtains the multi-dimensional feature de-entanglement table.

[0016] As a further solution of the present invention, the three-way dynamic interaction module includes:

[0017] The feature distribution deconstruction submodule calls the feature distribution data in the multidimensional feature deentanglement table, extracts the multimodal structure, multimodal feature dimension and feature channel type, classifies and divides them according to the multimodal feature dimension, and calculates the feature frequency value, change value and feature sparsity value under the multimodal channel respectively, determines whether the feature sparsity value is less than the set feature sparsity threshold, filters the feature channels that meet the threshold, and obtains the feature channel screening result;

[0018] The interaction path construction submodule extracts the characteristic frequency values ​​and change values ​​corresponding to the cross-modal channels based on the characteristic channel screening results, calculates the interaction strength value, determines the complementary or conflicting relationship according to the channel type and modality attribution, sequentially constructs cross-modal paths, and generates a cross-channel interaction path map;

[0019] The emotional feature recognition submodule extracts the frequency value and interaction strength value of the path node connection channel according to the cross-channel interaction path map, combines the modal attribution identifier of the channel, identifies the distribution of public features and private features in the path, counts the connection frequency and interaction strength mean of the private feature nodes, arranges and combines them in sequence to form a continuous interaction trajectory, and obtains the emotional feature interaction sequence.

[0020] As a further solution of the present invention, the feature semantic mapping module includes:

[0021] The feature item parsing submodule calls the feature items in the emotional feature interaction sequence, performs modality-specific judgment and task attribution mechanism division operations on the feature items in the differentiated modal data, obtains the modality label and task label value corresponding to each feature, jointly encodes the modality label value and the task label value, and establishes a feature label combination dataset;

[0022] The semantic matching calculation submodule selects two groups of participating items, the feature semantic vector and the task target semantic vector, based on the feature label combination data set, calculates the semantic deviation value, compares the semantic deviation value with the matching threshold, screens the feature combination items that meet the matching requirements, and generates a screened feature semantic list;

[0023] The semantic comparison submodule performs semantic difference extraction and semantic structure mapping processing based on the feature combination items in the filtered feature semantic list, compares the semantic position difference value, semantic orientation value and semantic consistency coefficient of the feature items with the target semantics, and generates a feature semantic mapping table.

[0024] As a further solution of the present invention, the abnormal feature verification module includes:

[0025] The semantic deviation extraction submodule calculates the semantic deviation strength value of the real-time feature based on the feature distribution in the feature semantic mapping table, identifies the offset trend distribution corresponding to multiple features according to the deviation strength value, and obtains the feature semantic deviation value;

[0026] The stability index calculation submodule calls the feature semantic deviation value, calculates the coefficient of variation and offset jump frequency in consecutive cycles based on the feature value fluctuation range and standard deviation in the time window, and compares them with the offset trend distribution to obtain the feature stability index;

[0027] The abnormal statistics identification submodule compares the offset threshold and stability critical value set by the real-time task target according to the feature stability index and feature semantic deviation value, counts the names of the out-of-limit features and the cumulative frequency in the distribution, and obtains a list of abnormal features.

[0028] As a further solution of the present invention, the system further includes a feature repair module:

[0029] The feature repair module calls the abnormal feature list, monitors the real-time semantic state and abnormal number of corresponding features, identifies features whose semantic state is abnormal and whose abnormal frequency exceeds the tolerance limit, performs feature remodeling operations and locates replacement features, and obtains emotional feature repair records;

[0030] The emotional feature restoration record includes the feature state before restoration, the replacement feature index, and the feature identifier after reconstruction.

[0031] As a further solution of the present invention, the feature repair module includes:

[0032] The state monitoring submodule extracts the real-time semantic state data and the cumulative number of abnormal records corresponding to the features in the abnormal feature list, performs interval judgment on the cumulative number of abnormal records and the abnormal frequency tolerance limit, and obtains the number of state abnormal features;

[0033] The frequency identification submodule collects the cumulative number of abnormal features and the rate of change of the features per unit time according to the number of abnormal state features, cross-compares the cumulative number of abnormal features with the rate of change of the features per unit time, identifies the features whose abnormal fluctuation intensity exceeds the interval range, and obtains the abnormal fluctuation rate value;

[0034] The alternative modeling submodule calls the abnormal fluctuation rate value, extracts the construction model parameter value and the alternative feature vector value of the corresponding feature, matches and selects the construction model parameter value of the feature with the alternative feature vector value, reconstructs the modeling parameter combination of the real-time abnormal feature, and records the replaced feature number and the alternative construction value to obtain the emotional feature repair record.

[0035] Compared with the prior art, the advantages and positive effects of the present invention are:

[0036] In this invention, EEG and ECG signals are decomposed in multiple stages to identify and distinguish public features from invalid private features. At the same time, an inter-modal semantic association mechanism is introduced to strengthen the attribution of feature attributes. At the feature level, synchronous analysis of signal dimensions and semantic dimensions is achieved. Distributed data is called to construct cross-modal, cross-dimensional, and cross-feature interaction paths. The attention mechanism is integrated to identify complementary relationships and private feature contributions. A more discriminatory interactive feature sequence is established between signal features and emotion labels. The feature semantic parsing mechanism is combined to map and match task attribution relationships, thereby improving the semantic coupling and interpretability between features and emotion classification task objectives. Through statistical analysis of feature semantic deviation, stability, and abnormal distribution, feature items that are inconsistent with task objectives are dynamically detected, and their semantic state evolution trends are clarified, promoting semantic quality control at the feature level. Feature replacement and modeling update are realized to ensure that the overall features are always maintained in the optimal expression structure, effectively avoiding the weakening of the stability and generalization ability of emotion recognition results by redundant features, so that the performance of multimodal physiological signals in emotion recognition tasks is sensitive, accurate and adaptable, and a physiological feature expression system with clear structure, clear semantics and dynamic adjustment is provided for emotional state recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a system flow chart of the present invention;

[0038] Figure 2 This is a flow chart of the characteristic de-entanglement module in the present invention;

[0039] Figure 3 This is a flow chart of the three-way dynamic interaction module in the present invention;

[0040] Figure 4 This is a flow chart of the feature semantic mapping module in the present invention;

[0041] Figure 5 This is a flow chart of the abnormal feature verification module in the present invention;

[0042] Figure 6 This is a flow chart of the feature repair module in the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0044] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0045] See also Figure 1 , a feature-detangling emotion decoding system combining EEG and ECG includes:

[0046] The feature disentanglement module performs multi-stage decomposition based on EEG and ECG signals to extract public features and invalid private features. It also separates public and private feature attributes through semantic association information between modalities to generate a multi-dimensional feature disentanglement table.

[0047] The three-way dynamic interaction module calls the feature distribution data in the multi-dimensional feature disentanglement table to construct cross-modal, cross-dimensional, and cross-feature interaction paths, identify the complementary relationship between public features and the contribution of private features, and generate emotional feature interaction sequences;

[0048] The feature semantic mapping module calls the feature items in the emotional feature interaction sequence, analyzes the modality specificity and task attribution mechanism, compares the matching degree between feature semantics and task objectives, selects features with priority matching degree, analyzes the semantic comparison relationship, and generates a feature semantic mapping table;

[0049] The abnormal feature verification module extracts the feature semantic deviation value, feature stability index and abnormal identifier according to the feature distribution in the feature semantic mapping table, compares the degree of deviation between the semantic deviation and the task target, counts the abnormal feature names and frequencies, and generates an abnormal feature list;

[0050] The feature repair module calls the abnormal feature list, monitors the real-time semantic state and abnormality count of the corresponding features, identifies features whose semantic state is abnormal and whose abnormality frequency exceeds the tolerance limit, performs feature remodeling operations and locates replacement features to obtain emotional feature repair records;

[0051] The multidimensional feature disentanglement table includes the modal correspondence distribution, feature independence measurement, and attribute separation mapping structure. The emotional feature interaction sequence includes the feature complementary distribution form, modal interaction structure style, and semantic connection strength index. The feature semantic mapping table includes semantic attribution category, matching priority index, and semantic comparison structure unit. The abnormal feature list includes feature name record, abnormal frequency statistics, and semantic offset mark content. The emotional feature repair record includes the feature status before repair, the replacement feature index, and the feature identifier after reconstruction.

[0052] See also Figure 2 , the feature disentanglement module includes:

[0053] The feature decomposition submodule performs multi-stage decomposition based on EEG and ECG signals. After time alignment according to a unified sampling frequency and time window division method, it extracts the amplitude sequence and spectrum sequence of the signal segments, compares the amplitude difference, frequency component offset value and disturbance fluctuation degree between adjacent signal segments, and determines the segment aggregation relationship based on the amplitude difference and frequency offset to obtain the common features of the signal segments.

[0054] After time alignment according to the unified sampling frequency and time window division method, signal segments are extracted. In actual implementation, EEG channels such as Fp1 and F3 and ECG signal channels such as II lead and V5 lead need to be selected. The sampling frequency is uniformly set to 256Hz, and the time window is divided into 5 seconds. Each window contains 1280 sampling points. Continuous data segments are extracted by sliding window method, and the amplitude sequence of each segment is directly extracted from the voltage sampling value and recorded as a one-dimensional vector. At the same time, the spectrum sequence is obtained through spectrum calculation tools such as the fft module in Matlab, and the main frequency distribution and energy concentration interval are recorded. After the extraction is completed, the adjacent signal segments are compared pairwise to extract the difference between the amplitudes. Range and fluctuation state: Set fragment A and fragment B in the same channel, calculate their maximum amplitude difference, frequency component offset value, and the stability of their fluctuation amplitude. In this process, the difference fluctuation must be less than a certain range as the fragment aggregation judgment standard. In actual judgment, if the maximum amplitude difference between the two fragments is within 0.5μV, the frequency principal component offset is within 3Hz, and there is no obvious mutation in the fluctuation amplitude, then the two are judged to belong to the same feature class. Further merge their spectral structures and take their common feature range. After aggregating all fragments that meet the aggregation relationship, extract their average frequency energy, main frequency band position, upper and lower amplitude limits as representative information of the current class to obtain the common features of the signal fragments.

[0055] The attribute separation submodule uses the common features of signal segments, combines the amplitude variation range, frequency concentration interval and phase change trend of EEG and ECG signals in the same time period, compares the amplitude variation degree and frequency concentration position change between differentiated signals, and separates the cross-section and non-cross-section between modes to obtain the cross-modal attribute difference;

[0056] On the basis of time synchronization, the original amplitude data, frequency distribution structure and phase change trajectory of the EEG and ECG signals in the current time period are extracted. The amplitude variation range is composed of the upper and lower intervals of the maximum and minimum values ​​of each signal segment. The frequency-intensive interval is demarcated by observing the upper and lower boundaries of the main frequency band where the energy is concentrated in the spectrum diagram. The phase change trend needs to be extracted by applying Hilbert transform to the signal and performing time series analysis. The EEG and ECG signals are classified separately to determine whether there is a change trend with similar structure or significant difference in a specific time window. If the frequency of the ECG signal is stable and concentrated in the range of 6Hz-10H z interval, while the EEG signal frequency is concentrated in 4Hz-8Hz in the same time window, it is considered that there is a frequency band crossover structure. By extracting the difference in the degree of amplitude change and the difference in the frequency main segment change, as well as the degree of overlap of the phase time change trend, a judgment standard for splitting the two-modal signal structure is formed. The common features that meet the same amplitude trend, close frequency distribution concentration interval, and phase change trend are divided into crossover areas, and the rest are divided into non-crossover attribute areas. The corresponding feature classification identifiers and modal attribute identifiers are recorded, the cross-modal attribute dividing line is drawn, and the modal cross-attribute difference is obtained.

[0057] The dimension mapping submodule extracts the value range and change direction of features in the common attribute set and the non-common attribute set based on the difference in modal cross-attributes. It combines the projection relationship and change rate distribution of the differentiated features in the coordinate axis mapping to adjust the representation of the original features in multiple dimensions and obtain a multi-dimensional feature disentanglement table.

[0058] Analyze the value differences and trend directions of each feature in the public attribute set and the non-public attribute set. Specifically, extract the value range of each feature in each set and record its upward or downward trend. Classify and organize multiple features based on the trend direction. Merge and classify features with similar change directions and similar value ranges. Use them as the input basis for multidimensional mapping. Model the spatial mapping of features. Select feature combinations in typical change intervals to establish coordinate axes. Rely on three-dimensional space modeling tools such as Origin and matplotlib in Python to construct feature point distribution maps. By observing the distribution structure and projection direction, adjust the original feature expression. When constructing coordinates, use the amplitude value range, frequency main peak position and phase fluctuation trend as references in the x, y and z directions respectively. Form a feature dimension mapping structure based on modal structure differences. Redefine the original feature content in the new coordinate system and classify it as a multidimensional coordinate record. The position of the feature points in this space will be used as the basis for feature comparison, modal identification and interaction analysis to obtain a multidimensional feature disentanglement table.

[0059] See also Figure 3 , the three-way dynamic interaction module includes:

[0060] The feature distribution deconstruction submodule calls the feature distribution data in the multidimensional feature disentanglement table, extracts the multimodal structure, multimodal feature dimension and feature channel type, classifies and divides them according to the multimodal feature dimension, and calculates the feature frequency value, change value and feature sparsity value under the multimodal channel respectively, determines whether the feature sparsity value is less than the set feature sparsity threshold, and selects the feature channels that meet the threshold to obtain the feature channel screening results;

[0061] Clarify the target modality and channel type, set extraction rules based on the data field, and load the multimodal input data stored locally or on the data server into the cache. In a typical example, the visual modality channel contains 64 channels, each representing the response feature of a convolution kernel output, the text modality contains 32 channels, and the audio modality contains 16 channels. After the channels are dimensionally unified and normalized, they need to be divided into channel groups based on the modality label. The feature frequency value is calculated. The frequency value is defined as the sum of the number of times the response value of a channel in the sample set is greater than the activation threshold (such as 0.8). Assume that there are 100 frames of images in the dataset, and the response value of visual channel 3 exceeds 0.8 in 28 frames, then the frequency value of this channel is 28. Calculate the channel change value, which is the average of the change amplitude of the feature output of adjacent time slices. Assume that the response value of visual channel 5 in 10 frames changes to [0.9, 1.1, 1.2, 0.7, 0.8, 1.0, 0.9, 1.3, 0.6, 1.1]. The change value is the average of the adjacent differences: ; For the calculation of sparsity, the sparsity formula is used ,in The number of channels in the current channel group whose response values ​​are greater than the threshold and whose times are not 0, is the total number of channels. If there are 32 channels in text mode and only 10 channels are activated in more than 5 samples, then , compared with the set sparse threshold of 0.7. Since 0.6875 is less than 0.7, the mode can be retained. After screening, the channel number and corresponding frequency value that meet the conditions in each mode are recorded, and the results of feature channel screening for subsequent use are summarized.

[0062] The interactive path construction submodule extracts the corresponding characteristic frequency values ​​and change values ​​between cross-modal channels based on the characteristic channel screening results, using the formula:

[0063] ;

[0064] Calculate the interaction strength value, determine the complementary or conflicting relationship based on the channel type and modality, construct cross-modal paths in sequence, and generate a cross-channel interaction path map;

[0065] in, is the interaction strength value, Indicates channel The characteristic frequency value of Indicates channel The characteristic frequency value of 、 Channel and In the The characteristic change value of the dimension, 、 Indicates channel and The range of the median change value, Represents the interaction gain between features in the path, is the number of dimensions;

[0066] Table 1 Interaction path sample data table:

[0067] ;

[0068] Table 1 lists the parameter values ​​involved in typical interaction paths and their calculated interaction strength values ​​for subsequent path judgment and classification;

[0069] The benefit of this formula is that by simultaneously introducing frequency difference, variation difference, and channel range terms, and jointly calculating the interaction strength between features, it avoids the problem of a single feature dimension dominating path construction and improves the multidimensional robustness of path determination.

[0070] Extract the characteristic frequency value between each pair of modal channels 、 The corresponding change value vector 、 Taking the visual-text modality channel as an example, let the visual channel frequency be 18 and the text channel frequency be 13, and their respective change value vectors are and , then the sum of the differences can be calculated as ;

[0071] Further calculate the range of each channel change value, the visual channel range , the text channel is extremely poor , interaction gain It is defined as the sum of the variances of the cross terms of the two channel change values ​​and is set to 1.2;

[0072] Then substitute into the formula:

[0073] ;

[0074] The results show that the interaction strength value is 1.96. In the interaction strength value reference table, if the value is between 1.5 and 2.5, it is marked as a complementary feature channel. A complementary path is established between the channels. Subsequently, the channel frequency is used as the path weight, and the path network is constructed through depth-first method. For example, when the visual channel is the root node, the one with the highest frequency value in the complementary path is preferentially connected. After the construction is completed, the connection relationship is recorded according to the path order to generate a cross-channel interaction path map.

[0075] The emotional feature recognition submodule extracts the frequency and interaction strength values ​​of the path nodes connecting the channels based on the cross-channel interaction path map. Combined with the modal attribution of the channels, it identifies the distribution of public and private features in the path, counts the connection frequency and interaction strength mean of the private feature nodes, and sequentially arranges and combines them to form a continuous interaction trajectory to obtain the emotional feature interaction sequence.

[0076] Extract the channel pairs composed of path nodes in the atlas and their corresponding interaction attributes. For example, path 1 connects visual channel a and text channel b, with frequency values ​​of 22 and 16 respectively, and an interaction strength value of 1.83. The channel type field is recorded as visual-shared and text-private. The path needs to be classified as "public feature path" or "private feature path" based on the channel type field. If a path contains at least one private modal channel, it is classified as a private path. Count the number of times each private feature appears in the atlas as a frequency indicator. If visual channel a appears in 7 paths and text channel b appears in 5 paths, the corresponding frequencies are 7 and 5; count the average interaction strength value corresponding to the feature in the path. The strength value is the average value of the channel corresponding to the feature in all paths in the path. Set the interaction strength value of visual channel a to [1.2, 1.3, 1.4, 1.1, 1.3, 1.2, 1.4], then the average is ; According to the order of channel appearance in the atlas, the channel connection paths are arranged into a sequence, such as channel a→b→c→e, and the interaction trajectory is constructed. Each pair of connecting channels in the path must meet the interaction intensity not less than the set path intensity threshold, such as 1.0. The feature ID of each path node and its channel frequency and average interaction intensity value are output in sequence and recorded in the structured interaction sequence to obtain the emotional feature interaction sequence.

[0077] See also Figure 4 , the feature semantic mapping module includes:

[0078] The feature item parsing submodule calls the feature items in the emotional feature interaction sequence, performs modality-specific judgment and task attribution mechanism division operations on the feature items in the differentiated modal data, obtains the modality label and task label value corresponding to each feature, jointly encodes the modality label value and the task label value, and establishes a feature label combination dataset;

[0079] Relying on the temporal arrangement of emotional data, the emotional expression vectors associated with the feature interaction points are extracted one by one, and their source identifications in different modal structures are recorded. The image modality can be set to , the text modal is , audio mode is , in order to distinguish the annotation sources of emotional features in multimodal corpus, construct the initial feature table by comparing the feature sequence with the modal identifier, and further target the modal specificity judgment link, take the frequency of occurrence of features in each modality and the context relevance as the judgment benchmark, set the frequency of occurrence of the feature "curved corners of the eyes" in the image modality to 0.87, and the word frequency of "happy" in the text to 0.32, then the modal attribution of this feature in the image modality is stronger, set the attribution benchmark to frequency difference ≥ 0.4, and attribute it to a stronger modal characteristic. In the task attribution mechanism division, according to the mapping performance of the current feature in the specific emotion recognition task and the emotion intensity judgment task, if a feature has an identification accuracy of 89% in the "pleasure / depression" task and an identification accuracy of 63% in the "excitement / calmness" task, it will be preferentially attributed to the previous task dimension, and the value is taken as the task directionality value T = max (task performance ratio). The modal label and the task label are uniformly mapped and encoded respectively, and set If the feature type is 1 and the task T is 2, the corresponding feature code is (1, 2), forming a feature type combination array such as [(1, 2), (2, 1), (3, 2)], etc., which is input into the subsequent comparison system as a feature source feature set to establish a feature label combination dataset.

[0080] The semantic matching calculation submodule selects two groups of participants, the feature semantic vector and the task target semantic vector, based on the feature label combination dataset, and uses the formula:

[0081] ;

[0082] Calculate the semantic deviation value, compare the semantic deviation value with the matching threshold, filter the feature combination items that meet the matching requirements, and generate a filter feature semantic list;

[0083] in, Represents the semantic deviation value, Representative The contribution of the feature, Representative Project expectations, Representative The degree of modal interference, Representative Item feature semantic vector value, Representative The semantic vector value of the project target, Representative The error coefficient between the semantic item and the task item, represents the number of features;

[0084] The calculation logic of the formula is used to calculate the semantic matching deviation value between feature semantic vectors. The core idea is to judge the degree of semantic matching by measuring the degree of deviation between the current task target semantic vector and the reference feature semantic vector, and calculate the semantic differences in the numerator. ,in is the feature contribution weight, is the expected value of the target item, is the current mode interference level; the denominator is the error normalization factor, which is obtained by The accumulated semantic error ratio plays a balancing role. This value can be used to integrate the interference information of multiple features, screen out the feature combination that best meets the semantic requirements, and achieve accurate matching and task evaluation;

[0085] The semantic deviation value is a quantitative indicator that measures the semantic gap between the feature semantic vector and the task target semantic vector. It comprehensively considers the feature contribution, the difference between the target and the interference, and the semantic error. The smaller the deviation value, the higher the matching degree, and the feature can be prioritized as a candidate feature.

[0086] The benefit of the formula is that it forms a multi-angle comprehensive scoring mechanism by fusing the semantic matching vector with multiple task interference parameters, effectively removing the interference of single semantic similarity and integrating feature contribution deviation into the structural calculation system, thus enhancing the recognition accuracy of effective feature matching, as shown in Table 2.

[0087] Table 2 Feature semantic matching analysis data table:

[0088] ;

[0089] Extract the semantic expression vectors corresponding to the encoded feature items and the semantic vectors corresponding to the task goals from Table 2, perform Euclidean distance and dot product calculations on each pair of feature vectors to form a semantic basic fit score, and then introduce the corresponding feature contribution, target expectation item, and modal interference degree. Specifically, the data in Table 2 is used as an example for analysis and calculation;

[0090] Parameter meaning and calculation process:

[0091] Representative The contribution of a feature is derived from the proportion of the feature in the correct prediction of the task;

[0092] For the The target expectation item is calculated by the ideal expression strength set by the target task for the semantic item;

[0093] is the degree of modal interference, which is converted by the frequency at which the feature is easily misidentified in the modal context;

[0094] and Respectively Item features and task semantic vectors, quantified from the fit of sentiment vocabulary and image feature dictionary;

[0095] is the semantic error coefficient, which is estimated by the average difference between the feature prediction value and the target value;

[0096] The numerator of the formula is the product of the difference between the contribution of each feature item and the target expectation item minus the modal interference degree, and then the square root is taken to calculate its difference impact score. The denominator is the dot product of the semantic vector divided by the corresponding error term and the absolute value summed. The specific calculation is performed through the formula;

[0097] Molecular calculation:

[0098] ;

[0099] ;

[0100] Denominator calculation:

[0101] ;

[0102] Substitute into the formula for calculation:

[0103] ;

[0104] The results show that the semantic bias value If the matching threshold is set to 1.1, the feature combination does not meet the screening requirements. The screening threshold needs to be tested experimentally to determine its optimal matching effect range. The experimental setting matching threshold range is set to [0.9, 1.3], and the matching accuracy is increased to 92.3%. This value and the matching threshold are used for screening to extract the feature combinations that meet the conditions and generate a screening feature semantic list.

[0105] The semantic comparison submodule performs semantic difference extraction and semantic structure mapping based on the feature combination items in the filtered feature semantic list, compares the semantic position difference value, semantic orientation value and semantic consistency coefficient of the feature items with the target semantics, and generates a feature semantic mapping table;

[0106] Extract the semantic expression structure in the modal corpus and its matching semantic expression structure in the task target. In the comparison process, it is necessary to analyze the semantic position difference value, that is, the vector distance between the corresponding feature semantic center in the embedding space. For example, the corresponding vector of "happy" in the vector space is [0.74, 0.65], and the task target "joy" is [0.82, 0.61]. The position difference is the Euclidean distance √[(0.74-0.82)²+(0.65-0.61)²]≈0.089. Further combined with the semantic pointing value, the semantic tendency coding of the feature item is compared with the target semantic coding, such as the feature The feature code is (1, 0, 1), the target is (1, 1, 0), and the Hamming distance between the two is 2. The semantic pointing value is 2 / 3=0.667. The semantic consistency coefficient is converted by the matching ratio. If there is a 2-dimensional complete match in the 3-dimensional vector, the consistency coefficient is 2 / 3=0.667. The three numerical values ​​are constructed into a joint comparison structure as the basis for semantic mapping judgment, forming a one-to-one corresponding mapping list item. The structure of each mapping item is {feature item ID, semantic bit difference value, semantic pointing value, consistency coefficient}. Multiple mapping items form an overall comparison table, and the output is a feature semantic mapping table.

[0107] See also Figure 5 , the abnormal feature verification module includes:

[0108] The semantic deviation extraction submodule is based on the feature distribution in the feature semantic mapping table and uses the formula:

[0109] ;

[0110] Calculate the semantic deviation strength value of the real-time feature, identify the offset trend distribution corresponding to multiple features based on the deviation strength value, and obtain the feature semantic deviation value;

[0111] in, represents the semantic bias strength value, Represents real-time eigenvalues, For the The mean of the features, represents the average state under normal conditions, For the The extreme difference of a characteristic, For the eigenvalues, Characterized by The total number of is the index of the feature;

[0112] The formula is used to calculate the semantic deviation strength value of a feature. Its rationality is reflected in its consideration of both the degree of deviation of the feature itself and the stability of the overall feature distribution. The numerator of the formula is the absolute value of the difference between the current value of a feature and its mean, which intuitively reflects the degree of difference between the current state of the feature and the normal state. The denominator introduces the total deviation of the feature (including the sum of the range of the feature and the deviation of the current values ​​of the remaining features from the mean). After normalization, it reflects the fluctuation level of the overall feature. The rationality of this normalization operation lies in: if the overall fluctuation is large, the outliers of individual features are part of the overall trend change, and the semantic deviation strength value should be relatively reduced. Conversely, if the overall situation is relatively stable, the deviation of a single feature is more worthy of attention, and the semantic deviation strength value should increase. This formula eliminates the interference of overall fluctuations on anomaly judgment through relative standardization, making the calculation result more robust and recognizable, and can more accurately identify features that truly deviate from the normal state. Introducing the range as a reference value for intra-feature fluctuations also improves local sensitivity and can effectively identify extreme deviations. The design fully integrates the standardization concept in statistics with the relative anomaly detection requirements in actual business, and has a good theoretical basis and practical applicability.

[0113] Table 3 Feature semantic deviation calculation table:

[0114] ;

[0115] Extract the mean, range and current sampling value of each feature in each category. When the five features numbered F01 to F05 are processed, the corresponding extracted current values ​​are 92.5, 84.1, 97.3, 76.2, and 88.4, respectively. The corresponding means are 85.0, 82.0, 90.0, 75.0, and 85.0, and the ranges are 10.0, 12.0, 15.0, 8.0, and 11.0. The feature value is a quantifiable data performance point, which is user behavior, monitoring value, and environmental parameter. Comparing the mean and range of each feature to determine whether it deviates significantly from the normal range is the core reference indicator for anomaly detection. The difference value is obtained by subtracting the mean from the current value of each feature. For example, the difference of F01 is 7.5. The difference is then normalized by comparing it with the range, and the deviation amplitude is calculated. The absolute value of the difference of the features is then counted to calculate the average deviation degree. For example, the absolute values ​​of the differences of F01 to F05 are 7.5, 2.1, 7.3, 1.2, and 3.4, respectively. The sum is 21.5, and the average is 4.3. Substitute this into the formula to calculate the deviation intensity.

[0116] ;

[0117] Set F01 and substitute:

[0118] ;

[0119] As shown in Table 3, the semantic deviation strength value of each feature is calculated using the formula. The deviation values ​​of multiple features are then plotted in a time series to show the offset trend distribution. This yields the feature semantic deviation value of each feature. This value serves as an important input parameter for subsequent stability analysis and anomaly judgment. The formula is beneficial in that it introduces a global average offset term to correct for the limitations of single-feature fluctuations and enhance the overall adaptability of the deviation measure.

[0120] The stability index calculation submodule calls the feature semantic deviation value, calculates the coefficient of variation and offset jump frequency in consecutive cycles based on the feature value fluctuation range and standard deviation within the time window, and compares them with the offset trend distribution to obtain the feature stability index;

[0121] The feature semantic deviation value is used as the state input of the current feature. The value sequence of the corresponding feature in multiple time period windows is extracted. The values ​​of F01 in the past five time segments are set to 90.1, 91.3, 92.5, 91.7, and 92.2, respectively. The standard deviation of the sequence is calculated to be 0.79, and the mean is 91.56. The coefficient of variation (CV) is calculated as CV = standard deviation / mean = 0.79 / 91.56 ≈ 0.0086. The intensity of its fluctuation is then determined, and whether there is sudden change behavior is identified. The difference between any two consecutive periods is set to exceed 3%. If it exceeds, it is considered a jump. The offset jump threshold is set to 2.7%. In the above sequence, the change from 91.3 to 92.5 is 1.2 / 91.3 ≈ 1.31%, which does not constitute a jump. The jump frequency is statistically calculated as 0. Combined with the fluctuation curve of F01 in the deviation trend chart, its stability score can be obtained as the weighted average of the inverse of the offset value and the jump frequency, resulting in a feature stability index of 0.72 for F01.

[0122] The abnormal statistical identification submodule compares the feature stability index and feature semantic deviation value with the offset threshold and stability critical value set by the real-time task target, counts the names of the out-of-limit features and the cumulative frequency in the distribution, and obtains a list of abnormal features;

[0123] According to the acquired feature semantic deviation value and feature stability index, they are compared with the two critical benchmarks preset in the task objective, and the offset threshold is set to 1.5 and the stability lower limit is 0.6. For F01, its offset strength is 1.98, which is higher than the threshold 1.5, and the stability index is 0.72, which is higher than the critical value. It meets the primary condition for abnormal judgment, so F01 is recorded as an abnormal item once, and the features F02 to F05 are traversed. Among them, F03 has a deviation strength of 2.0 and a stability of 0.58, which does not meet the stability index and is also recorded as an abnormal feature. The abnormal frequency is counted, and each type of abnormal feature is named and a frequency record table is established. For example, F01 appears abnormally 3 times and F03 appears twice, and a list of abnormal features is obtained.

[0124] See also Figure 6 , the feature repair module includes:

[0125] The state monitoring submodule extracts the real-time semantic state data and the cumulative number of abnormal records corresponding to the features in the abnormal feature list, performs interval judgment on the cumulative number of abnormal records and the abnormal frequency tolerance limit, and obtains the number of state abnormal features;

[0126] Synchronized EEG and ECG data is used to extract features such as P300, alpha wave power, RR interval, and QRS waveform variation. Redundancy is removed and a preliminary screening list of abnormal features is generated. Each feature in this list corresponds to the semantic state data sampled in the current time window. Semantic labels such as "nervous" and "angry" are assigned to the emotion recognition classification results, and the number of abnormal occurrences of each feature is counted. For example, if a user's alpha wave power is below 10μV² and has fallen below the set threshold three times in a row, the system will record the cumulative number of abnormalities as 3. The abnormal frequency tolerance limit for this feature is set to 2 to 4 times, meaning it should be less than or equal to 4 times. If the cumulative number of abnormalities falls within this range, the feature is considered to be abnormal within the frequency tolerance interval. If it exceeds the range, it is considered to be persistently abnormal. For example, if the RR interval fluctuates within 0.2s while the user is reading quietly, and if the fluctuation increases to 0.8s over a period of time and occurs frequently for a total of 5 times, the cumulative number of abnormalities exceeds the tolerance limit (set to 3 times) and is counted as an abnormal feature.

[0127] The frequency identification submodule collects the cumulative number of abnormal features and the rate of change of features per unit time according to the number of abnormal state features, cross-compares the cumulative number of abnormal features with the rate of change of features per unit time, identifies the features whose abnormal fluctuation intensity exceeds the interval range, and obtains the abnormal fluctuation rate value;

[0128] In-depth analysis of the changing trend and stability of each abnormal feature. The specific process includes extracting the cumulative number of abnormalities of each feature during the monitoring period, and judging the severity of its fluctuations based on the changes in the feature within unit time. For vibration features, the per-minute change value is collected and the difference between adjacent sampling points is calculated to obtain the rate of change within unit time. By matching and analyzing the cumulative number of abnormalities with the rate of change, multiple evaluation intervals are set. If the number of abnormalities of a certain feature is set to 7 times and the rate of change per unit time is 40%, it falls into the high-frequency and high-variability interval. Features with both high abnormality times and high change rates are identified, and some features are confirmed to show a severe abnormal trend. Based on the information, an abnormal fluctuation rate value is generated for each qualified feature.

[0129] The alternative modeling submodule calls the abnormal fluctuation rate value, extracts the construction model parameter value and the alternative feature vector value of the corresponding feature, matches and selects the construction model parameter value of the feature with the alternative feature vector value, reconstructs the modeling parameter combination of the real-time abnormal feature, and records the replaced feature number and the replacement construction value to obtain the emotional feature repair record;

[0130] After identifying a feature with a significantly abnormal rate of fluctuation, the model repair mechanism is activated to construct alternative parameters for the abnormal feature. The abnormal rate of fluctuation is determined to determine whether it exceeds a predetermined tolerance threshold. Features exceeding a certain value are marked as requiring model repair. The parameter configuration used in the modeling of the feature is retrieved. This parameter configuration includes values ​​used for model construction, setting linear trend factors, period adjustment parameters, time window length, and more. At the same time, alternative feature vector values ​​related to the abnormal feature are selected from the current operating status of the device. These values ​​are derived from remaining features that have a strong correlation with the abnormal feature. The similarity between the parameters is evaluated using predefined matching rules. If the difference between the parameters in the current model and the alternative parameters is small, the alternative values ​​are used to replace them, and the modeling combination for the feature is rebuilt. During the parameter replacement process, the replaced feature number is recorded, along with the specific values ​​of the replacement parameters. The time of the repair, the replacement process, the source of the parameters used, and the reasons for their selection are detailed to facilitate subsequent review and continuous model optimization. This process ensures rapid replacement when a model fails, maintains data modeling continuity, and obtains a record of emotional feature repair.

[0131] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A feature-based emotion decoding system combining EEG and ECG, characterized by: The system includes a feature disentanglement module, a three-way dynamic interaction module, a feature semantic mapping module, and an abnormal feature verification module; The feature disentanglement module performs multi-stage decomposition based on EEG and ECG signals to extract public features and invalid private features. It also separates public and private feature attributes through semantic association information between modalities to generate a multi-dimensional feature disentanglement table. The three-way dynamic interaction module calls the feature distribution data in the multi-dimensional feature disentanglement table, constructs cross-modal, cross-dimensional and cross-feature interaction paths, identifies the complementary relationship between public features and the contribution of private features, and generates an emotional feature interaction sequence; The feature semantic mapping module calls the feature items in the emotional feature interaction sequence, analyzes the modality specificity and task attribution mechanism, compares the matching degree between the feature semantics and the task goal, selects the features with priority matching degree and analyzes the semantic comparison relationship to generate a feature semantic mapping table; The abnormal feature verification module extracts the feature semantic deviation value, feature stability index and abnormal identifier according to the feature distribution in the feature semantic mapping table, compares the degree of deviation between the semantic deviation and the task target, counts the abnormal feature names and frequencies, and generates an abnormal feature list; The three-way dynamic interaction module includes a feature distribution deconstruction submodule, an interaction path construction submodule, and an emotion feature recognition submodule; The feature distribution deconstruction submodule calls the feature distribution data in the multidimensional feature deentanglement table, extracts the multimodal structure, multimodal feature dimension and feature channel type, classifies and divides them according to the multimodal feature dimension, and calculates the feature frequency value, change value and feature sparsity value under the multimodal channel respectively, determines whether the feature sparsity value is less than the set feature sparsity threshold, filters the feature channels that meet the threshold, and obtains the feature channel screening result; The interaction path construction submodule extracts the characteristic frequency values ​​and change values ​​corresponding to the cross-modal channels based on the characteristic channel screening results, calculates the interaction strength value, determines the complementary or conflicting relationship according to the channel type and modality attribution, sequentially constructs cross-modal paths, and generates a cross-channel interaction path map; The emotional feature recognition submodule extracts the frequency value and interaction strength value of the path node connection channel according to the cross-channel interaction path map, combines the modal attribution identifier of the channel, identifies the distribution of public features and private features in the path, counts the connection frequency and interaction strength mean of the private feature nodes, arranges and combines them in sequence to form a continuous interaction trajectory, and obtains the emotional feature interaction sequence.

2. The feature-based de-entanglement emotion decoding system combining EEG and ECG according to claim 1 is characterized in that: The multidimensional feature disentanglement table includes modal correspondence distribution, feature independence measurement, and attribute separation mapping structure; the emotional feature interaction sequence includes feature complementary distribution form, modal interaction structure style, and semantic connection strength index; the feature semantic mapping table includes semantic attribution category, matching priority index, and semantic comparison structure unit; the abnormal feature list includes feature name record, abnormal frequency statistics, and semantic offset mark content.

3. The feature-based de-entanglement emotion decoding system combining EEG and ECG according to claim 1 is characterized in that: The feature disentanglement module includes a feature splitting submodule, an attribute separation submodule, and a dimension mapping submodule; The feature decomposition submodule performs multi-stage decomposition based on EEG and ECG signals. After time alignment according to a unified sampling frequency and time window division method, it extracts the amplitude sequence and spectrum sequence of the signal segments, compares the amplitude difference, frequency component offset value and disturbance fluctuation degree between adjacent signal segments, and determines the segment aggregation relationship based on the amplitude difference and frequency offset to obtain the common features of the signal segments. The attribute separation submodule calls the common features of the signal segments, combines the amplitude variation range, frequency concentration interval and phase variation trend of the EEG and ECG signals in the same time period, compares the amplitude variation degree and frequency concentration position change between the differentiated signals, and separates the cross-section and non-cross-section between the modes to obtain the modal cross-attribute difference; The dimensional mapping submodule extracts the value range and change direction of the features under the common attribute set and the non-common attribute set based on the modal cross-attribute difference, combines the projection relationship and change rate distribution of the differentiated features in the coordinate axis mapping, adjusts the expression of the original features in multiple dimensions, and obtains the multi-dimensional feature de-entanglement table.

4. The feature-based de-entanglement emotion decoding system combining EEG and ECG according to claim 3 is characterized in that: The feature semantic mapping module includes a feature item parsing submodule, a semantic matching calculation submodule, and a semantic comparison submodule; The feature item parsing submodule calls the feature items in the emotional feature interaction sequence, performs modality-specific judgment and task attribution mechanism division operations on the feature items in the differentiated modal data, obtains the modality label and task label value corresponding to each feature, jointly encodes the modality label value and the task label value, and establishes a feature label combination dataset; The semantic matching calculation submodule selects two groups of participating items, the feature semantic vector and the task target semantic vector, based on the feature label combination data set, calculates the semantic deviation value, compares the semantic deviation value with the matching threshold, screens the feature combination items that meet the matching requirements, and generates a screened feature semantic list; The semantic comparison submodule performs semantic difference extraction and semantic structure mapping processing based on the feature combination items in the filtered feature semantic list, compares the semantic position difference value, semantic orientation value and semantic consistency coefficient of the feature items with the target semantics, and generates a feature semantic mapping table.

5. The feature-based de-entanglement emotion decoding system combining EEG and ECG according to claim 4 is characterized in that: The abnormal feature verification module includes a semantic deviation extraction submodule, a stability index calculation submodule, and an abnormal statistics identification submodule; The semantic deviation extraction submodule calculates the semantic deviation strength value of the real-time feature based on the feature distribution in the feature semantic mapping table, identifies the offset trend distribution corresponding to multiple features according to the deviation strength value, and obtains the feature semantic deviation value; The stability index calculation submodule calls the feature semantic deviation value, calculates the coefficient of variation and offset jump frequency in consecutive cycles based on the feature value fluctuation range and standard deviation in the time window, and compares them with the offset trend distribution to obtain the feature stability index; The abnormal statistics identification submodule compares the offset threshold and stability critical value set by the real-time task target according to the feature stability index and feature semantic deviation value, counts the names of the out-of-limit features and the cumulative frequency in the distribution, and obtains a list of abnormal features.

6. The feature-based de-entanglement emotion decoding system combining EEG and ECG according to claim 1 is characterized in that: The system also includes a feature repair module: The feature repair module calls the abnormal feature list, monitors the real-time semantic state and abnormal number of corresponding features, identifies features whose semantic state is abnormal and whose abnormal frequency exceeds the tolerance limit, performs feature remodeling operations and locates replacement features, and obtains emotional feature repair records; The emotional feature restoration record includes the feature state before restoration, the replacement feature index, and the feature identifier after reconstruction.

7. The feature-based de-entanglement emotion decoding system combining EEG and ECG according to claim 6 is characterized in that: The feature repair module includes a state monitoring submodule, a frequency identification submodule, and an alternative modeling submodule; The state monitoring submodule extracts the real-time semantic state data and the cumulative number of abnormal records corresponding to the features in the abnormal feature list, performs interval judgment on the cumulative number of abnormal records and the abnormal frequency tolerance limit, and obtains the number of state abnormal features; The frequency identification submodule collects the cumulative number of abnormal features and the rate of change of the features per unit time according to the number of abnormal state features, cross-compares the cumulative number of abnormal features with the rate of change of the features per unit time, identifies the features whose abnormal fluctuation intensity exceeds the interval range, and obtains the abnormal fluctuation rate value; The alternative modeling submodule calls the abnormal fluctuation rate value, extracts the construction model parameter value and the alternative feature vector value of the corresponding feature, matches and selects the construction model parameter value of the feature with the alternative feature vector value, reconstructs the modeling parameter combination of the real-time abnormal feature, and records the replaced feature number and the alternative construction value to obtain the emotional feature repair record.

Images