Feature de-entanglement emotion decoding system combining electroencephalogram and electrocardio

Abstract

The invention relates to the technical field of physiological signal processing, in particular to an electroencephalogram and electrocardio combined feature de-entanglement emotion decoding system which comprises a feature de-entanglement module, a three-way dynamic interaction module, a feature semantic mapping module, an abnormal feature verification module and a feature restoration module. According to the method, the electroencephalogram signals and the electrocardiosignals are subjected to multi-stage decomposition to recognize and distinguish public features and invalid private features, synchronous analysis of the signal dimension and the semantic dimension is achieved, feature items inconsistent with task targets are dynamically detected through statistical analysis of feature semantic deviation, stability and abnormal distribution conditions, and the accuracy of the task targets is improved. The semantic state evolution trend of the multi-modal physiological signals is defined, semantic quality control on the feature level is promoted, feature replacement and modeling updating are achieved, and the expression of the multi-modal physiological signals in an emotion recognition task has sensitivity, accuracy and adaptability; and a physiological feature expression system which is clear in structure, definite in semantics and dynamically adjustable is provided for emotional state recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of physiological signal processing, and in particular to an emotion decoding system for feature disentanglement by combining electroencephalogram (EEG) and electrocardiogram (ECG). Background Art

[0002] The technical field of physiological signal processing includes the acquisition, analysis, and feature construction of multi-source physiological signals such as EEG, ECG, and electromyogram (EMG), aiming to explore the internal physiological response laws of individuals in different psychological and physiological states. The core content of this technical field includes the temporal alignment of cross-modal signals, feature attribution modeling, correlation pattern recognition, and trend classification mechanisms. Its systematicness is reflected in the standardization of signal acquisition mechanisms, the analysis of feature structure differences, the collaborative modeling between modalities, and the dynamic adaptation of the signal-emotion mapping mechanism. Physiological signal processing is widely applied in scenarios such as emotion recognition, stress perception, and neurofeedback. By integrating multi-modal data to form a task-oriented physiological feature expression model, it is an important part of cognitive computing and human-computer interaction systems.

[0003] Among them, the emotion decoding system for feature disentanglement by combining EEG and ECG refers to an emotion recognition method based on the joint modeling of EEG and ECG signals, which differentiates different feature dimensions and performs decoupling processing. This patent theme addresses the problems of cross-modal feature aliasing, ambiguous modality attribution, and unclear feature-task association. It covers the alignment of original time-slice data in EEG and ECG signals, the extraction of private and public features and their category attribution recognition, the establishment of an explicit discrimination mechanism for effective and ineffective features in the feature space, the construction of a multi-dimensional interaction path for joint feature reconstruction based on modality specificity, and the generation of a feature semantic contrast mapping relationship through physiological structure attribution rules to construct a multi-dimensional feature analysis scheme for emotion recognition tasks.

[0004] Although the prior art has established a joint modeling framework for multi-source signals such as electroencephalogram (EEG) and electrocardiogram (ECG) in emotion recognition, there are still problems in actual operation, such as inaccurate semantic alignment between signal modalities, fuzzy attribution of feature distributions, and poor adaptability of features to tasks. In conventional multi-modal fusion, features are combined in a splicing or alignment manner, lacking an explicit discrimination mechanism for common and private features, resulting in serious signal aliasing and the masking of common features by private feature noise. During the emotion label mapping process, there is a lack of feature-task semantic adaptation evaluation, and it is impossible to accurately identify key features highly relevant to the task, resulting in weak interpretability of the model decision path and prone to overfitting or transfer failure. There is a lack of means to monitor the temporal stability and semantic deviation of features, making it difficult to detect interfering features in the recognition results, and the recognition accuracy is easily affected by fluctuations, and the system generalization ability is limited. Taking practical applications as an example, in the emotion induction task, although the electrocardiogram waveform changes of the same subject at multiple time periods conform to individual rules, they are treated as the same type of features in the current technology, ignoring their differences from specific emotional states, resulting in a deviation of the emotional attribution mechanism and affecting the reliability of the model. Summary of the Invention

[0005] The object of the present invention is to solve the shortcomings existing in the prior art, and to propose a feature disentanglement emotion decoding system combining electroencephalogram and electrocardiogram.

[0006] To achieve the above object, the present invention adopts the following technical solution. A feature disentanglement emotion decoding system combining electroencephalogram and electrocardiogram includes: The feature disentanglement module is based on electroencephalogram and electrocardiogram signals, performs multi-stage decomposition, extracts common features and invalid private features, and separates the common and private feature attributes through inter-modal semantic association information 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 relationships between common features and the contributions of private features, and generates an emotion feature interaction sequence; The feature semantic mapping module calls the feature items in the emotion feature interaction sequence, analyzes the modality specificity and task attribution mechanism, compares the matching degree between the feature semantics and the task objectives, screens the features with the highest matching degree and analyzes the semantic comparison relationship, and generates 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 deviation degree of the semantic deviation from the task objective, counts the names and frequencies of abnormal features, and generates an abnormal feature list.

[0007] As a further solution of the present invention, the multi-dimensional feature disentanglement table includes a modality correspondence distribution, a feature independence measure, and an attribute separation mapping structure. The emotional feature interaction sequence includes a feature complementary distribution form, a modality interaction structure style, and a semantic connection strength index. The feature semantic mapping table includes a semantic attribution category, a matching priority index, and a semantic comparison structure unit. The abnormal feature list includes a feature name record, an abnormal frequency statistical value, and semantic deviation marked content.

[0008] As a further solution of the present invention, the feature disentanglement module includes: The feature splitting sub-module performs multi-stage decomposition based on electroencephalogram and electrocardiogram signals. After time alignment according to the unified sampling frequency and time window division method, the amplitude sequence and frequency spectrum sequence of the signal segment are extracted. The amplitude difference, frequency component offset value, and perturbation fluctuation degree between adjacent signal segments are compared. According to the amplitude difference and frequency offset, the segment aggregation relationship is judged, and the common features of the signal segments are obtained. The attribute separation sub-module calls the common features of the signal segments, combines the amplitude change range, frequency dense interval, and phase change trend of the electroencephalogram and electrocardiogram signals in the same time period, compares the amplitude change degree and frequency concentration position change of the differential signals, and splits the cross-modal and non-cross-modal parts to obtain the cross-modal attribute difference. The dimension mapping sub-module extracts the value range and change direction of the features under the common attribute set and non-common attribute set according to the cross-modal attribute difference, combines the projection relationship and change rate distribution of the differential features in the coordinate axis mapping, and adjusts the representation form of the original features in multiple dimensions to obtain the multi-dimensional feature disentanglement table.

[0009] As a further solution of the present invention, the three-way dynamic interaction module includes: The feature distribution deconstruction sub-module calls the feature distribution data in the multi-dimensional feature disentanglement table, extracts the multi-modal structure, multi-modal feature dimension, and feature channel type, classifies and divides them according to the multi-modal feature dimension, and calculates the feature frequency value, change value, and feature sparsity value under the multi-modal channel respectively. It judges 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 sub-module extracts the corresponding feature frequency value and change value between the cross-modal channels based on the feature channel screening result, calculates the interaction strength value, judges the complementary or conflict relationship according to the channel type and modality attribution, and sequentially constructs the cross-modal path to generate the cross-channel interaction path map. The emotional feature recognition sub-module extracts the frequency value and interaction intensity value of the path node connection channels according to the cross-channel interaction path graph, combines the modal attribution identifier of the channels, identifies the distribution of public features and private features in the path, statistically calculates the connection frequency and the average value of the interaction intensity of the private feature nodes, arranges and combines them in sequence to form a continuous interaction trajectory, and obtains an emotional feature interaction sequence.

[0010] As a further solution of the present invention, the feature semantic mapping module includes: The feature item parsing sub-module calls the feature items in the emotional feature interaction sequence, performs modal-specific judgment and task attribution mechanism division operations on the feature items in the differential modal data, obtains the modal label and task label value corresponding to each feature, jointly encodes the modal label value and the task label value, and establishes a feature label combination data set; The semantic matching calculation sub-module selects two sets of participating items, namely 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 out the feature combination items that meet the matching requirements, and generates a screened feature semantic list; The semantic comparison sub-module performs semantic difference degree extraction and semantic structure mapping processing according to the feature combination items in the screened feature semantic list, compares the semantic position difference value, semantic pointing value and semantic consistency coefficient of the feature item and the target semantics, and generates a feature semantic mapping table.

[0011] As a further solution of the present invention, the abnormal feature verification module includes: The semantic deviation extraction sub-module calculates the semantic deviation intensity 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 intensity value, and obtains the feature semantic deviation value; The stability index calculation sub-module calls the feature semantic deviation value, calculates the coefficient of variation and the offset jump frequency in the continuous period according to the value fluctuation range and standard deviation of the feature within the time window, and compares with the offset trend distribution to obtain the feature stability index; The abnormal statistical recognition sub-module compares the feature stability index and the feature semantic deviation value with the offset threshold and stability critical value set by the real-time task target, statistically calculates the names of the over-limit features and the cumulative frequency in the distribution, and obtains an abnormal feature list.

[0012] As a further solution of the present invention, the system further includes a feature repair module: The feature repair module calls the abnormal feature list, monitors the real-time semantic state and abnormal times of the corresponding features, identifies the features whose semantic state is abnormal and the abnormal frequency exceeds the tolerance limit, performs feature re-modeling operations and locates alternative features, and obtains an emotional feature repair record; The emotional feature repair record includes the feature status before repair, the alternative feature index, and the feature identifier after reconstruction.

[0013] As a further solution of the present invention, the feature repair module includes: The status monitoring sub-module extracts the real-time semantic status data and the cumulative abnormal record times corresponding to the features in the abnormal feature list, makes an interval judgment on the cumulative abnormal record times and the abnormal frequency tolerance limit value to obtain the number of features with abnormal status; The frequency identification sub-module collects the cumulative abnormal times and the feature change rate per unit time of the feature according to the number of features with abnormal status, cross-compares the cumulative abnormal times and the feature change rate per unit time, and identifies the features whose abnormal fluctuation intensity exceeds the interval range to obtain the abnormal fluctuation rate value; The alternative modeling sub-module 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 and the alternative feature vector value of the feature, reconstructs the modeling parameter combination of the real-time abnormal feature, and records the alternative feature number and the alternative construction value to obtain the emotional feature repair record.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In the present invention, by performing multi-stage decomposition on electroencephalogram and electrocardiogram signals to identify and distinguish common features and invalid private features, and introducing an inter-modal semantic association mechanism to strengthen the feature attribute attribution, synchronous parsing of the signal dimension and the semantic dimension is realized at the feature level. Distributed data is called to construct cross-modal, cross-dimensional, and cross-feature interaction paths, and the attention mechanism is fused to identify complementary relationships and private feature contributions, establishing a more discriminative interaction feature sequence between signal features and emotion labels. Combining the feature semantic parsing mechanism to map and match-screen the task attribution relationship, improving the semantic coupling degree and interpretability between features and emotion classification task objectives. Through statistical analysis of feature semantic deviation, stability, and abnormal distribution, feature items inconsistent with the task objective 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, so that the overall features always maintain the optimal expression structure, effectively avoiding the weakening of the stability and generalization ability of the emotion recognition result by redundant features, making the performance of multi-modal physiological signals in the emotion recognition task sensitive, accurate, and adaptable, and providing a physiological feature expression system with clear structure, clear semantics, and dynamically adjustable for emotion state recognition. Brief Description of the Drawings

[0015] Figure 1 is the system flow chart of the present invention; Figure 2 is the flow chart of the feature disentanglement module in the present invention; Figure 3 This is the flowchart of the three-way dynamic interaction module in the present invention; Figure 4 This is the flowchart of the feature semantic mapping module in the present invention; Figure 5 This is the flowchart of the abnormal feature verification module in the present invention; Figure 6 This is the flowchart of the feature repair module in the present invention. Detailed implementation manners

[0016] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, 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 used to limit the present invention.

[0017] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. In addition, in the description of the present invention, the meaning of "a plurality" is two or more, unless otherwise specifically defined.

[0018] Please refer to Figure 1 , a feature disentanglement emotion decoding system combining electroencephalogram and electrocardiogram includes: The feature disentanglement module performs multi-stage decomposition based on electroencephalogram and electrocardiogram signals, extracts common features and invalid private features, and separates the common and private feature attributes through inter-modal semantic association information 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 common features and the contribution of private features, and generates an emotion feature interaction sequence; The feature semantic mapping module calls the feature items in the emotion feature interaction sequence, analyzes the modality specificity and task attribution mechanism, compares the matching degree between the feature semantics and the task objective, screens the features with the highest matching degree and analyzes the semantic comparison relationship, and generates 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 deviation degree of the semantic deviation from the task objective, counts the abnormal feature names and frequencies, and generates an abnormal feature list; The feature repair module calls the abnormal feature list, monitors the real-time semantic status and the number of abnormal occurrences of the corresponding features, identifies the features whose semantic status is abnormal and the abnormal frequency exceeds the tolerance limit, performs the feature re-modeling operation and locates the alternative features, and obtains the sentiment feature repair record; The multi-dimensional feature disentanglement table includes the modal correspondence distribution, the feature independence measure, and the attribute separation mapping structure. The sentiment feature interaction sequence includes the feature complementary distribution form, the modal interaction structure style, and the semantic connection strength index. The feature semantic mapping table includes the semantic attribution category, the matching priority index, and the semantic comparison structure unit. The abnormal feature list includes the feature name record, the abnormal frequency statistical value, and the semantic shift marker content. The sentiment feature repair record includes the feature status before repair, the alternative feature index, and the feature identifier after reconstruction.

[0019] Please refer to Figure 2 , the feature disentanglement module includes: Based on the electroencephalogram (EEG) and electrocardiogram (ECG) signals, the feature splitting sub-module performs multi-stage decomposition. After time alignment according to the unified sampling frequency and time window division method, it extracts the amplitude sequence and frequency spectrum sequence of the signal segment, compares the amplitude difference, frequency component offset value, and perturbation fluctuation degree between adjacent signal segments, and determines the segment aggregation relationship based on the amplitude difference and frequency offset. Then it obtains the common features of the signal segments; 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 Lead II and V5 are selected. The sampling frequency is uniformly set to 256 Hz, the time window is divided every 5 seconds, and each window contains 1280 sampling points. Continuous data segments are extracted by the sliding window method. The amplitude sequence of each segment of the signal is directly extracted from the voltage sampling value and recorded as a one-dimensional vector. At the same time, the frequency spectrum sequence is obtained through a frequency spectrum calculation tool such as the fft module in Matlab, and the main frequency distribution and energy concentration interval are recorded. After the extraction is completed, adjacent signal segments are compared pairwise to extract the difference range and fluctuation state between the amplitudes. Assuming segment A and segment B are in the same channel, calculate their maximum amplitude difference, frequency component offset value, and the stability degree of their fluctuation amplitude. In this process, a certain range of difference fluctuations is used as the segment aggregation judgment standard. In actual judgment, if the maximum amplitude difference between two segments is within 0.5 μV, the main frequency component offset is within 3 Hz, and there is no obvious mutation in the fluctuation amplitude, it is determined that the two belong to the same feature class. Further merge their frequency spectrum structures and take the common feature range. After aggregating all the segments that meet the aggregation relationship, extract their average frequency energy, main frequency band position, amplitude upper bound, and lower bound as the representative information of the current class, and obtain the common features of the signal segments.

[0020] The attribute separation sub-module calls the common features of the 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 the change of frequency concentration position between different signals, splits the cross and non-cross parts between modalities, and obtains the cross-modal attribute differences. On the basis of time synchronization, extract the original amplitude data, frequency distribution structure, and phase change trajectory of EEG and ECG signals in the current time period. The amplitude variation range consists of the upper and lower intervals formed by the maximum and minimum values of each signal segment. The frequency concentration interval delimits the upper and lower boundaries by observing the main frequency band where the energy is concentrated in the spectrogram. The phase change trend requires applying the Hilbert transform to the signal to extract the instantaneous phase and performing time series analysis. Classify the EEG and ECG signals respectively, and judge whether there are similar or significantly different change trends in a specific time window. If the frequency of the ECG signal is stably concentrated in the 6Hz - 10Hz interval, while the frequency of the EEG signal is concentrated in the 4Hz - 8Hz interval in the same time window, it is considered that there is a frequency band cross structure. By extracting the difference in amplitude variation degree, the difference in the main frequency segment change, and the coincidence degree of the phase time change trend, a splitting and judgment criterion for the two-modal signal structure is formed. The parts in the common features that meet the conditions of consistent amplitude trend, close frequency distribution concentration interval, and simultaneous phase change are classified as the cross region, and the rest are classified as the non-cross attribute region. Record the corresponding feature classification identifier and modal attribute identifier, delimit the cross-modal attribute boundary line, and obtain the cross-modal attribute differences.

[0021] The dimension mapping sub-module extracts the value intervals and change directions of the features under the common attribute set and non-common attribute set according to the cross-modal attribute differences, combines the projection relationship and change rate distribution of the different features in the coordinate axis mapping, adjusts the manifestation form of the original features in multiple dimensions, and obtains the multi-dimensional feature disentanglement table. Analyze the value differences and trend directions presented by current features in the common attribute set and non-common attribute set. Specifically, by extracting the numerical range of each feature's change in each set and recording its upward or downward trend, classify and organize multiple features in combination with the trend direction. Merge and classify features with similar change directions and similar value ranges, and use them as the input basis for multi-dimensional mapping. Build a spatial mapping model for features, select a combination of features in the typical change interval to establish coordinate axes, and rely on 3D spatial modeling tools such as Origin and the matplotlib tool in Python to construct a feature point distribution map. Adjust the original feature representation form by observing the distribution structure and projection direction. When constructing the coordinates, use the amplitude value range, the position of the main frequency peak, and the phase fluctuation trend as the reference for the x, y, and z directions respectively to form a feature dimension mapping structure based on the modal structure difference. Redefine the original feature content in the new coordinate system and classify it in the form of multi-dimensional coordinates for recording. The position of the feature points in this space will be used as the basic basis for feature comparison, modal recognition, and interaction analysis to obtain a multi-dimensional feature disentanglement table.

[0022] Please refer to Figure 3 , the three-way dynamic interaction module includes: The feature distribution deconstruction sub-module calls the feature distribution data in the multi-dimensional feature disentanglement table, extracts the multi-modal structure, multi-modal feature dimensions, and feature channel types, classifies and divides them according to the multi-modal feature dimensions, and calculates the feature frequency value, change value, and feature sparsity value under the multi-modal channels respectively. Judge whether the feature sparsity value is less than the set feature sparsity threshold, and screen the feature channels that meet the threshold to obtain the feature channel screening result; Define the target modality and channel type, and according to the data field setting extraction rules, load the multi-modal input data stored locally or in the data server into the buffer area. In a typical example, the visual modality channel has 64 channels, each channel represents the response feature output by a convolution kernel, the text modality has 32 channels, and the audio modality has 16 channels. After dimension unification and normalization processing of the channels, it is necessary to divide them into channel groups according to the modality label and perform the calculation of the feature frequency value. The frequency value is defined as the sum of the number of times the response value of a certain channel in the sample set is greater than the activation threshold (such as 0.8). Suppose there are 100 frames of images in the dataset, and 28 frames of the response value of visual channel 3 exceed 0.8, then the frequency value of this channel is 28; calculate the channel change value, and the change value is the average value of the change amplitude of the feature output in adjacent time slices. Set the response value change of visual channel 5 in 10 frames to [0.9, 1.1, 1.2, 0.7, 0.8, 1.0, 0.9, 1.3, 0.6, 1.1], then the change value is the average of the adjacent differences: ; For the calculation of sparsity, use the sparsity formula , where is the number of channels in the current channel group whose response value is greater than the threshold and the number of times is not zero. is the total number of channels. It is set that there are 32 channels in the text modality, and only 10 channels are activated in more than 5 samples. Then , compare with the set sparse threshold of 0.7. Since 0.6875 is less than 0.7, this modality can be retained. After screening, record the eligible channel numbers and corresponding frequency values in each modality, and summarize them to form the feature channel screening results for subsequent use.

[0023] Based on the feature channel screening results, the interaction path construction sub-module extracts the corresponding feature frequency values and variation values between cross-modal channels, and uses the formula: ; Calculate the interaction intensity value, judge the complementary or conflict relationship according to the channel type and modality attribution, and sequentially construct the cross-modal path to generate the cross-channel interaction path map; Among them, is the interaction intensity value, represents the feature frequency value of channel , represents the feature frequency value of channel , , are respectively the feature variation values of channels and in the dimension, , represent the range of variation values in channels and , represents the interaction gain degree between features in the path, is the number of dimensions; Table 1 Sample data table of interaction paths: ; Table 1 lists the parameter values involved in the typical interaction path and the calculated interaction intensity values for subsequent path judgment and classification; The advantage of the formula is that by simultaneously introducing the frequency difference, variation difference and channel range terms, the interaction intensity between features is jointly calculated, avoiding the problem of a single feature dimension dominating the path construction and improving the multi-dimensional robustness of path determination; Extract the feature frequency values , and the corresponding variation value vectors , between each pair of modality channels. Taking the visual-text modality channels as an example, assume that the frequency of the visual channel is 18 and the frequency of the text channel is 13, and their respective variation value vectors are and , the total difference can be calculated as ; Further calculate the range of the change values of each channel. The range of the visual channel , the range of the text channel , and the interaction gain is defined as the sum of the cross-term variances of the change values of the two channels, and is set to 1.2; Then substitute into the formula: ; The results show that the interaction intensity value is 1.96. In the interaction intensity value reference table, if this value is between 1.5 and 2.5, it is marked as a complementary feature channel, and a complementary path is established between the channels. Subsequently, the channel frequency is used as the path weight, and a path network is constructed through depth-first search. 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 in the order of the path, and a cross-channel interaction path map is generated.

[0024] The emotional feature recognition sub-module extracts the frequency value and interaction intensity value of the channels connected by the path nodes in 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 the average value of the interaction intensity of the private feature nodes, arranges them in combination in sequence to form a continuous interaction trajectory, and obtains an emotional feature interaction sequence; Extract the channel pairs composed of path nodes in the map 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 the interaction intensity value is 1.83. The channel type field is recorded as visual - shared, text - private. It is necessary to classify the path as a "public feature path" or a "private feature path" according to 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 map as the frequency index. 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 value of the interaction intensity values corresponding to this feature in the path. The intensity value is the average value of the channels corresponding to this feature in all paths in the path. Set the interaction intensity value of visual channel a to [1.2, 1.3, 1.4, 1.1, 1.3, 1.2, 1.4], then the average value is ; According to the order of the channels appearing in the map, organize the connection paths of each channel into a sequence, such as channel a → b → c → e, to construct an interaction trajectory. Each pair of connected channels in the path needs to satisfy that the interaction intensity is not lower than the set path intensity threshold, such as 1.0. Output the feature ID of each path node, its channel frequency, and the average interaction intensity value in order, and record them in the structured interaction sequence to obtain the emotional feature interaction sequence.

[0025] Please refer to Figure 4, the feature semantic mapping module includes: The feature item parsing sub-module calls the feature items in the emotion feature interaction sequence, performs modal-specific judgment and task attribution mechanism division operations on the feature items in the differential modal data, obtains the modal label and task label value corresponding to each feature, jointly encodes the modal label value and task label value, and establishes a feature label combination data set; Relying on the time series arrangement of the emotion data, the emotion expression vectors associated with the feature interaction points are extracted one by one, and their source identifiers in different modal structures are recorded. The image modality can be set to , the text modality is , the audio modality is , so as to distinguish the annotation sources of emotion features in the multi-modal corpus. An initial feature table is constructed by comparing the feature sequence with the modal identifier. Further, for the modal-specific judgment link, the appearance frequency and context relevance of the feature in each modality are used as the judgment benchmark. The appearance frequency of the feature "corner of the eye bent" in the image modality is set to 0.87, and the word frequency of "happy" in the text is 0.32. Then the modal attribution degree of this feature in the image modality is stronger. The attribution benchmark is set as the frequency difference ≥0.4, belonging to the stronger modal characteristics. In the task attribution mechanism division, according to the mapping performance of the current feature in specific emotion recognition tasks and emotion intensity judgment tasks, such as the recognition accuracy of a certain feature in the "pleasure / depression" task is 89%, while the recognition accuracy in the "excited / calm" task is 63%, then it is preferentially attributed to the former task dimension. The value-taking method is the task directivity value T = max(task performance ratio). The modal label and task label are respectively uniformly mapped and encoded. Set to be 1, the task T is 2, then the corresponding feature code is (1, 2), forming a feature type combination array such as [(1, 2), (2, 1), (3, 2)], etc., which is used as the feature source characteristic set and input into the subsequent comparison system to establish a feature label combination data set.

[0026] Based on the feature label combination data set, the semantic matching calculation sub-module selects two participating items, the feature semantic vector and the task target semantic vector, and uses the formula: ; Calculate the semantic deviation value, compare the semantic deviation value with the matching threshold, screen out the feature combination items that meet the matching requirements, and generate a screened feature semantic list; Among them, represents the semantic deviation value, represents the th feature contribution amount, represents the th target expectation item, represents the th modal interference degree, Represents the semantic vector value of the th feature, represents the semantic vector value of the th project slogan, represents the error coefficient between the th semantic item and the task item; Calculation logic of the formula: used to calculate the semantic matching deviation value between feature semantic vectors. Its core idea is to judge the semantic matching degree by measuring the deviation degree between the current task target semantic vector and the reference feature semantic vector. The numerator part calculates the semantic differences of each item , where is the feature contribution weight, is the target item expected value, is the current mode interference degree; the denominator part is the error normalization factor, which accumulates the semantic error ratio through to play a balancing role. Through this value, the interference information of multiple features can be fused, the most semantic-demanding feature combination can be screened out, and accurate matching and task evaluation can be achieved; The semantic deviation value is a quantitative indicator for measuring the semantic gap between the feature semantic vector and the task target semantic vector. Considering the feature contribution degree, the target and interference differences, and the semantic error comprehensively, the smaller the deviation value, the higher the matching degree, and it can be preferentially used as a candidate feature; The benefit of the formula is that through the fusion of the semantic matching vector and multiple task interference parameters, a multi-angle comprehensive scoring mechanism is formed, effectively stripping the interference of single semantic similarity, and integrating the feature contribution deviation into the structure calculation system to enhance the recognition accuracy of effective feature matching, as shown in Table 2; Table 2 Feature Semantic Matching Analysis Data Table: ; Extract the semantic expression vector corresponding to the encoded feature item and the semantic vector corresponding to the task target from Table 2, perform Euclidean distance and dot product calculations on each pair of feature vectors to form a semantic basic fitness score, and then introduce the corresponding three participation items of the feature contribution amount, the target expected item, and the modal interference degree. Specifically, analyze and calculate with the data in Table 2 as an example; Parameter meaning and calculation process: Represents the th feature contribution amount, and its value comes from the proportion of the feature's participation in the correct prediction in the task; is the th project target expected item, which is calculated by the ideal expression intensity set by the target task for this semantic item; is the modal interference degree, which is converted by the misrecognition frequency of this feature in the modal context; and are the th feature and task semantic vector, respectively, quantified from the fitting degree of sentiment vocabulary and image feature dictionary; is the semantic error coefficient, estimated by the average difference between the feature prediction value and the target value; In the formula, the numerator part is the square root of the product of the contribution of each feature item and the difference between the target expectation item and the modal interference degree to calculate its difference influence score. The denominator part is the sum of the absolute values of the dot product of the semantic vectors divided by the corresponding error term, and the specific calculation is carried out through the formula; Calculation of the numerator part: ; ; Calculation of the denominator part: ; Substitute into the formula for calculation: ; The result shows that the semantic deviation value is , if the matching threshold is set to 1.1, then this feature combination does not meet the screening requirements. The screening threshold needs to be determined through experimental tests to find its optimal matching effect interval. The experimental setting matching threshold interval is [0.9, 1.3], and the matching accuracy rate is increased to 92.3%. Through this value and the matching threshold for screening, the feature combinations that meet the conditions are extracted to generate a screened feature semantic list.

[0027] The semantic comparison sub-module performs semantic difference extraction and semantic structure mapping processing according to the feature combination items in the screened feature semantic list, compares the semantic position difference value, semantic pointing value and semantic consistency coefficient between the feature item and the target semantics, and generates a feature semantic mapping table; Extract the semantic expression structure in the modal corpus and its matching semantic representation structure in the task objective. During the comparison process, it is necessary to analyze the semantic bit difference, that is, the vector distance of the corresponding feature semantic center in the embedding space. For example, the vector corresponding to "happy" in the vector space is [0.74, 0.65], and the task objective "pleased" is [0.82, 0.61]. Then the bit difference is the Euclidean distance √[(0.74 - 0.82)² + (0.65 - 0.61)²] ≈ 0.089. Further, in combination with the semantic pointing value, by comparing the semantic tendency encoding of the feature item with the target semantic encoding. For example, the feature encoding is (1, 0, 1), and the target is (1, 1, 0). The Hamming distance between the two is 2, then the semantic pointing value is 2 / 3 = 0.667. The semantic consistency coefficient is obtained by converting the matching ratio. For example, if 2 out of 3 dimensions of the vector are completely matched, the consistency coefficient is 2 / 3 = 0.667. Construct the three numerical values into a joint comparison structure as the basis for semantic mapping determination, forming a one-to-one 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, which is output as a feature semantic mapping table.

[0028] Please refer to Figure 5 , the abnormal feature verification module includes: The semantic deviation extraction sub-module is based on the feature distribution in the feature semantic mapping table and uses the formula: ; Calculate the semantic deviation intensity value of the real-time feature, identify the offset trend distribution corresponding to multiple features according to the deviation intensity value, and obtain the feature semantic deviation value; Among them, represents the semantic deviation intensity value, represents the real-time th feature value, is the mean value of the th feature, represents the average state under normal conditions, is the th range of the feature, is the th feature value, is the feature total number, is the index of the feature; The formula is used to calculate the semantic deviation intensity value of features. Its rationality lies in taking into account both the deviation degree of the feature itself and the stability of the overall feature distribution. The numerator part of the formula is the absolute value of the difference between the current value of a certain feature and its average value, which intuitively reflects the degree of difference between the current state of the feature and the normal state. The denominator part introduces the total deviation of the feature (including the range of the feature and the sum of the deviations 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 is as follows: if the overall fluctuation is large, the outlier of an individual feature is part of the overall trend change, and the semantic deviation intensity value should be relatively reduced; on the contrary, if the overall is relatively stable, then the deviation of a single feature is more worthy of attention, and the semantic deviation intensity value should increase. This formula eliminates the interference of overall fluctuations on anomaly judgment through relative standardization, making the calculation result more robust and distinguishable, and can more accurately identify the features that are truly deviated from the normal state. Introducing the range as a reference value for intra-feature fluctuation also improves local sensitivity and can effectively identify extreme deviations. The design fully integrates the standardization idea in statistics and the relative anomaly detection requirements in actual business, and has good theoretical basis and practical applicability; Table 3 Feature Semantic Deviation Calculation Table: ; Extract the mean, range, and current sampling value of each feature under each category. When setting the processing of five features numbered F01 to F05, the corresponding current values extracted 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 representation point, which is user behavior, monitoring value, environmental parameter. By comparing it with the mean and range of the feature, it is judged whether it significantly deviates from the normal range, which is the core reference index for anomaly detection. Obtain the difference value by subtracting the mean value of each feature from its current value. For example, the difference value of F01 is 7.5. Then compare this difference value with its range for normalization, calculate the offset amplitude, and then calculate the average offset degree after statistically calculating the absolute value of the difference of the feature. 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, and the sum is 21.5, and the average is 4.3. Substitute it into the formula for deviation intensity calculation; ; Set F01 substituted to get: ; As shown in Table 3, the semantic deviation intensity value of each feature is calculated through a formula. Then, the deviation values of multiple features are plotted according to the time series to obtain the offset trend distribution, and the feature semantic deviation value of each feature is obtained. This value can be used as an important input parameter for subsequent stability analysis and anomaly judgment. The benefit of the formula is that it corrects the limitations of single-feature fluctuations by introducing a global average offset term, enhancing the overall adaptability of the deviation measure.

[0029] The stability index calculation sub-module calls the feature semantic deviation value, calculates the coefficient of variation and the offset jump frequency within a continuous period according to the value fluctuation range and standard deviation of the feature within the time window, and compares with the offset trend distribution to obtain the feature stability index; Call the feature semantic deviation value as the state quantity input of the current feature, extract the value sequence of the corresponding feature within multiple time period windows. Set the values of F01 in the past 5 time segments to 90.1, 91.3, 92.5, 91.7, 92.2 respectively. Then calculate the standard deviation of this sequence as 0.79 and the mean as 91.56. Calculate the coefficient of variation CV = standard deviation / mean = 0.79 / 91.56 ≈ 0.0086. Then judge its fluctuation intensity, and at the same time identify whether there is a mutation behavior. Set whether the difference between any two consecutive periods exceeds 3%. If it exceeds, it is regarded as a jump. Set the offset jump threshold to 2.7%. In the above sequence, the change range from 91.3 to 92.5 is 1.2 / 91.3 ≈ 1.31%, which does not constitute a jump. Then count the jump frequency as 0. Combining with the fluctuation curve of F01 in the deviation trend graph, the stability score can be obtained as the weighted average of the reciprocal of the offset value and the jump frequency. The feature stability index of F01 is 0.72.

[0030] The anomaly statistics and identification sub-module compares the offset threshold and stability critical value set by the real-time task target according to the feature stability index and the feature semantic deviation value, and counts the names of the over-limit features and their cumulative frequencies in the distribution to obtain the anomaly feature list; According to the obtained feature semantic deviation value and feature stability index, compare them with two preset critical benchmarks in the task target. Set the offset threshold to 1.5 and the stability lower limit value to 0.6. For F01, its offset intensity is 1.98, higher than the threshold 1.5, and the stability index is 0.72, higher than the critical value, meeting the primary condition for anomaly determination. Then record F01 as an anomaly item once, and continue to traverse features F02 to F05. Among them, the deviation intensity of F03 is 2.0 and the stability is 0.58, not meeting the stability index, and it is also recorded as an anomaly feature. Count the anomaly frequency, name each type of anomaly feature and establish a frequency record table. For example, F01 appears 3 times as an anomaly and F03 appears 2 times, to obtain the anomaly feature list.

[0031] Please refer to Figure 6 , the feature repair module includes: The status monitoring sub-module extracts the real-time semantic status data and the cumulative abnormal record times corresponding to the features in the abnormal feature list, makes an interval judgment on the cumulative abnormal record times and the abnormal frequency tolerance limit value, and obtains the number of status abnormal features; Call the electroencephalogram and electrocardiogram synchronous data, such as P300, alpha wave power, R-R interval, QRS waveform variation and other data, perform redundancy removal processing and generate a preliminary screening list of abnormal features. Each feature in this list needs to correspond to the semantic status data sampled in the current time window, set emotional recognition classification results such as semantic labels like "nervous" and "angry", and at the same time count the number of times the feature appears abnormally. For example, if a user's alpha wave power is lower than 10 μV² and has been lower than the set threshold three times in a row, the system will record its cumulative abnormal times as 3. At the same time, set the abnormal frequency tolerance limit value of this feature to 2 to 4 times, that is, this value should be less than or equal to 4 times. If the cumulative abnormal times fall within this range, it is judged that the feature is in an abnormal state within the frequency tolerance interval. If it exceeds the interval, it is determined to be a persistent abnormality. If a user's R-R interval fluctuation range is 0.2 s during quiet reading, and if it increases to 0.8 s and appears frequently 5 times within a period of time, the comparison between the cumulative times and the tolerance limit (set to 3 times) is exceeded and is included in the number of status abnormal features.

[0032] The frequency identification sub-module collects the cumulative abnormal times and the feature change rate per unit time of the feature according to the number of status abnormal features, cross-compares the cumulative abnormal times and the feature change rate per unit time, and identifies the features whose abnormal fluctuation intensity exceeds the interval range to obtain the abnormal fluctuation rate value; Deeply analyze the change trend and stability of each abnormal feature. The specific process includes extracting the cumulative abnormal times of each feature during the monitoring period, and combining the change of this feature per unit time to judge its fluctuation intensity. For vibration features, it will collect its change value per minute and calculate the difference change between adjacent sampling points to obtain the change rate per unit time. By matching and analyzing the cumulative abnormal times and the change rate, set multiple evaluation intervals. Set the abnormal times of a certain feature to 7 times and the change rate per unit time to 40%, then it falls into the high-frequency and high-change interval, identify the features with both high abnormal times and high change rates, confirm that some features show a strong abnormal trend, and generate an abnormal fluctuation rate value for each qualified feature according to the information.

[0033] The surrogate modeling sub-module calls the abnormal fluctuation rate value, extracts the construction model parameter values and alternative feature vector values corresponding to the features, matches and selects the construction model parameter values and alternative feature vector values of the features, reconstructs the modeling parameter combination of the real-time abnormal features, and records the surrogate feature number and surrogate construction value to obtain the emotional feature repair record; After identifying features with an obviously abnormal fluctuation rate, the model repair mechanism will be activated to construct alternative parameters for the abnormal features. It will judge whether the abnormal fluctuation rate value exceeds the established tolerance threshold, and set the features exceeding a certain value to be marked as requiring model repair. The parameter configuration used by the feature in the modeling will be retrieved. The parameter configuration contains the values used for model construction, such as setting the linear trend factor, cycle adjustment parameter, time window length, etc. At the same time, the alternative feature vector values related to the abnormal feature will be screened from the current operating state of the device. The values are derived from the remaining features that have a strong correlation with the abnormal feature. The similarity between the parameters will be evaluated through the set matching rules. If the gap between the parameters in the current model and the alternative parameters is small, the alternative value can be used for substitution, and the modeling combination of the feature will be reconstructed. During the parameter substitution process, the number of the replaced feature will be recorded, and the specific value of the alternative parameter will be indicated, listing in detail the time of repair, the substitution process, the source of the adopted parameters and the reasons for selection, so as to facilitate subsequent review and continuous optimization of the model for use. This process can ensure rapid substitution when the model fails and maintain the continuity of data modeling, and obtain the emotional feature repair record.

[0034] The above are only the preferred embodiments of the present invention, and do not limit the present invention in other forms. Any person skilled in the art may use the technical content disclosed above to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A feature disentanglement emotion decoding system combining electroencephalogram and electrocardiogram, characterized in that The system includes: Based on electroencephalogram (EEG) and electrocardiogram (ECG) signals, the feature disentanglement module performs multi-stage decomposition, extracts common features and invalid private features, and separates the common and private feature attributes through inter-modal semantic association information 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 interaction paths across modalities, dimensions, and features, identifies the complementary relationships between common features and the contributions 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 objectives, screens the features with higher matching priority 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 deviation degree of the semantic deviation from the task objective, counts the names and frequencies of abnormal features, and generates an abnormal feature list.

2. The feature disentanglement emotion decoding system combining electroencephalogram and electrocardiogram according to claim 1, characterized in that, The multi-dimensional feature disentanglement table includes the modality correspondence distribution, feature independence measurement, and attribute separation mapping structure. The emotional feature interaction sequence includes the feature complementary distribution form, modality interaction structure style, and semantic connection strength index. The feature semantic mapping table includes the semantic attribution category, matching priority index, and semantic comparison structure unit. The abnormal feature list includes the feature name record, abnormal frequency statistical value, and semantic deviation mark content.

3. The feature disentanglement emotion decoding system combining electroencephalogram and electrocardiogram according to claim 1, characterized in that, The feature disentanglement module includes: The feature splitting sub-module performs multi-stage decomposition based on EEG and ECG signals. After time alignment according to the unified sampling frequency and time window division method, it extracts the amplitude sequence and frequency spectrum sequence of the signal segment, compares the amplitude difference, frequency component offset value, and perturbation 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 segment; The attribute separation sub-module calls the common features of the signal segment, combines the amplitude change range, frequency dense interval, and phase change trend of EEG and ECG signals in the same time period, compares the amplitude change degree and frequency concentration position change of different signals, splits the cross and non-cross parts between modalities, and obtains the cross-modal attribute difference; The dimension mapping sub-module extracts the value interval and change direction of the feature under the common attribute set and non-common attribute set according to the cross-modal attribute difference, combines the projection relationship and change rate distribution of the different features in the coordinate axis mapping, adjusts the representation form of the original feature in multiple dimensions, and obtains a multi-dimensional feature disentanglement table.

4. The feature disentangled emotion decoding system combining electroencephalogram and electrocardiogram according to claim 3, characterized in that, The three-way dynamic interaction module includes: The feature distribution deconstruction sub-module calls the feature distribution data in the multi-dimensional feature disentanglement table, extracts the multi-modal structure, multi-modal feature dimensions, and feature channel types, classifies and divides them according to the multi-modal feature dimensions, and calculates the feature frequency value, change value, and feature sparsity value under the multi-modal channels respectively. It judges whether the feature sparsity value is less than the set feature sparsity threshold, screens the feature channels that meet the threshold, and obtains the feature channel screening result; The interaction path construction sub-module extracts the corresponding feature frequency values and variation values between cross-modal channels based on the feature channel screening results, calculates the interaction intensity value, judges the complementary or conflict relationship according to the channel type and modal attribution, and sequentially constructs cross-modal paths to generate a cross-channel interaction path map; The emotional feature recognition sub-module extracts the frequency value and interaction intensity value of the path node connection channels according to the cross-channel interaction path map, combines the modal attribution identifiers of the channels, identifies the distribution of public features and private features in the path, counts the connection frequency and the average interaction intensity of the private feature nodes, and sequentially arranges and combines them to form a continuous interaction trajectory to obtain an emotional feature interaction sequence.

5. The feature disentanglement emotion decoding system combining electroencephalogram and electrocardiogram according to claim 4, characterized in that, The feature semantic mapping module includes: The feature item parsing sub-module calls the feature items in the emotional feature interaction sequence, performs modal-specific judgment and task attribution mechanism division operations on the feature items in the differential modal data, obtains the modal label and task label value corresponding to each feature, jointly encodes the modal label value and the task label value, and establishes a feature label combination data set; The semantic matching calculation sub-module selects two sets of participating items, namely 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 sub-module performs semantic difference extraction and semantic structure mapping processing according to the feature combination items in the screened feature semantic list, compares the semantic position difference value, semantic pointing value and semantic consistency coefficient of the feature item and the target semantics, and generates a feature semantic mapping table.

6. The feature disentanglement emotion decoding system combining electroencephalogram and electrocardiogram according to claim 5, characterized in that The abnormal feature verification module includes: The semantic deviation extraction sub-module calculates the semantic deviation intensity 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 intensity value, and obtains the feature semantic deviation value; The stability index calculation sub-module calls the feature semantic deviation value, calculates the coefficient of variation and offset jump frequency in a continuous period according to the value fluctuation range and standard deviation of the feature within the time window, and compares with the offset trend distribution to obtain the feature stability index; The abnormal statistics recognition sub-module compares the offset threshold and stability critical value set by the real-time task target according to the feature stability index and the feature semantic deviation value, counts the names of the over-limit features and their cumulative frequencies in the distribution, and obtains a list of abnormal features.

7. The feature disentanglement emotion decoding system combining electroencephalogram and electrocardiogram according to claim 1, characterized in that, The system further includes a feature repair module: The feature repair module calls the list of abnormal features, monitors the real-time semantic state and the number of abnormalities of the corresponding features, identifies the features whose semantic state is abnormal and the number of abnormalities exceeds the tolerance limit, performs feature re-modeling operations and locates alternative features, and obtains an emotional feature repair record; The emotional feature repair record includes the feature state before repair, the alternative feature index, and the feature identifier after reconstruction.

8. The feature disentangled emotion decoding system combining electroencephalogram and electrocardiogram according to claim 7, characterized in that, The feature repair module includes: The status monitoring sub-module extracts the real-time semantic status data and the cumulative abnormal record times corresponding to the features in the abnormal feature list, makes an interval judgment on the cumulative abnormal record times and the abnormal frequency tolerance limit value, and obtains the number of status abnormal features; The frequency identification sub-module collects the cumulative abnormal times and the feature change rate per unit time of the features according to the number of status abnormal features, makes a cross comparison between the cumulative abnormal times and the feature change rate per unit time, identifies the features whose abnormal fluctuation intensity exceeds the interval range, and obtains the abnormal fluctuation rate value; The alternative modeling sub-module calls the abnormal fluctuation rate value, extracts the construction model parameter value and the alternative feature vector value of the corresponding feature, makes a matching selection between the construction model parameter value and the alternative feature vector value of the feature, reconstructs the modeling parameter combination of the real-time abnormal feature, and records the replaced feature number and the replaced construction value to obtain the emotional feature repair record.

Images