Psychological catharsis interaction system based on artificial intelligence

Abstract

The invention relates to the technical field of emotion analysis, and provides a psychological catharsis interaction system based on artificial intelligence, which comprises a data acquisition module, an emotion recognition module, a virtual catharsis scene module and an emotion analysis and guidance module, physiological data and behavior data of students are collected through VR equipment and a sensor, an emotion recognition model is constructed in combination with feature extraction and a multi-modal feature fusion technology, and emotion category recognition is achieved; dynamically generating an adaptive virtual catharsis scene based on an identification result; the emotion analysis and guidance module optimizes hyper-parameters of a TCN-GRU model through an improved grey wolf optimization algorithm, constructs an enhanced TCN-GRU model, extracts long-term and short-term features from virtual catharsis interaction data, analyzes emotion evolution and provides personalized psychological guidance; the system can generate an accurate emotion evolution map and psychological state scores, and provides a scientific and accurate psychological intervention scheme for students.

Description

Technical Field

[0001] The present invention relates to the technical field of emotion analysis, and in particular to a psychological catharsis interaction system based on artificial intelligence. Background Art

[0002] With the rapid development of artificial intelligence technology, more and more mental health intervention methods have begun to introduce artificial intelligence technology to assist in emotion recognition and psychological state analysis, so as to help students relieve psychological pressure more efficiently; however, traditional psychological catharsis interaction systems generally have some shortcomings: first, traditional systems usually rely on only a single data source, ignoring the deep integration of multimodal emotional information, resulting in low accuracy and reliability of emotion recognition; second, traditional systems are insufficient in capturing the dynamic changes of emotions and cannot fully reflect the evolution of user emotions. The results of psychological state analysis are relatively simple and lack personalization; finally, in terms of psychological state assessment, traditional systems usually rely too much on static indicators and ignore the comprehensive extraction of long-term dependencies and short-term dynamic features in time series data, making it difficult to provide accurate emotion recognition and scientific psychological guidance; therefore, there is an urgent need for a psychological catharsis interaction system that can combine multimodal data, dynamically generate virtual catharsis scenes, and has the ability of efficient emotion recognition and accurate psychological state analysis to comprehensively improve the effect of psychological intervention. Summary of the invention

[0003] The present invention provides an artificial intelligence-based psychological catharsis interactive system, which realizes efficient emotion recognition and accurate psychological state analysis through multimodal emotion recognition and dynamic emotion analysis methods, combined with virtual catharsis scene generation and personalized psychological guidance; the system uses a multimodal emotion recognition method, a nested memory feature generation mechanism, a newly added forget gate mechanism and a two-way feature fusion method to extract fused time series features from physiological data and behavioral data, accurately capture emotional changes and output recognition results; at the same time, in emotion analysis and guidance, the present invention combines a dynamic leadership mechanism and a particle swarm optimization mechanism to improve the position update strategy of the gray wolf optimization algorithm, realizes the generation of an emotion evolution map and a refined scoring of the psychological state, and provides targeted psychological guidance for students; the overall system effectively makes up for the problems of weak emotion capture ability, inaccurate analysis results and insufficient personalization in the prior art, and significantly improves the psychological intervention effect.

[0004] The present invention provides a psychological catharsis interaction system based on artificial intelligence, which includes a data acquisition module, an emotion recognition module, a virtual catharsis scene module, and an emotion analysis and guidance module;

[0005] The data acquisition module collects students' physiological data and student behavior data through VR equipment and sensors, and integrates them to obtain original emotional data; students' physiological data include heart rate, skin electrical response, brain waves and breathing rate, and students' behavior data include body movements, voice intonation, eye movement trajectory and facial expressions;

[0006] The emotion recognition module combines feature extraction technology and multimodal feature fusion technology to build a multimodal emotion recognition model, uses the multimodal emotion recognition model to analyze the original emotion data, identifies the student's emotion category, and outputs the emotion recognition result;

[0007] The virtual catharsis scene module generates matching virtual catharsis scenes and virtual catharsis interaction data according to the emotion recognition results; the virtual catharsis scenes include emotion venting scenes, emotion release scenes and confession scenes;

[0008] The emotion analysis and guidance module establishes a TCN-GRU model, improves the position update strategy of the GWO algorithm by introducing a dynamic leadership mechanism and a particle swarm optimization mechanism, constructs an improved GWO algorithm, uses the improved GWO algorithm to optimize the hyperparameters of the TCN-GRU model, and constructs an enhanced TCN-GRU model; uses the enhanced TCN-GRU model to analyze virtual catharsis interaction data and psychological states, obtains an optimized emotion evolution map and optimized psychological state scoring data; and provides personalized psychological guidance based on the obtained optimized emotion evolution map and optimized psychological state scoring data.

[0009] Furthermore, the process of the emotion recognition module outputting the emotion recognition result specifically includes the following steps:

[0010] Step S1: Physiological feature extraction: using wavelet transform to extract the time-frequency domain features of students' physiological data, and performing dimensionality reduction and optimization through sparse autoencoder to obtain students' physiological feature data;

[0011] Step S2: Behavior feature extraction: extracting feature data of student behavior data through eGeMAPS feature set, convolutional neural network and gated recurrent unit respectively to obtain student behavior feature data;

[0012] Step S3: Preliminary fusion: Preliminary fusion feature data is obtained by preliminarily fusion of the student's physiological feature data and the student's behavioral feature data;

[0013] Step S4: Advanced feature fusion: A nested He-BiLSTM model is constructed through a nested memory feature generation mechanism, a newly added forget gate mechanism, and bidirectional feature fusion. The preliminary fused feature data is input into the nested He-BiLSTM model to capture the bidirectional time dependency in the feature data and obtain fused time series feature data. The nested He-BiLSTM model includes a forward I-LSTM and a backward I-LSTM.

[0014] Step S5: Attention weighting: The time series feature data is processed and fused through the attention mechanism, and importance weights are assigned to obtain weighted time feature data;

[0015] Step S6: layer normalization: performing layer normalization processing on the weighted time feature data to reduce the influence of feature scale differences and obtain normalized time feature data;

[0016] Step S7: nonlinear activation: input the normalized time feature data into the ReLU activation function, perform nonlinear mapping, capture the nonlinear relationship between features, and obtain nonlinear time feature data;

[0017] Step S8: Regularization processing: input the nonlinear time feature data into the Dropout layer in the multimodal emotion recognition model, randomly discard some neurons to prevent overfitting, and obtain regularized time feature data;

[0018] Step S9: residual connection: performing residual connection on the preliminary fusion feature data and the regularized time feature data to obtain advanced fusion feature data;

[0019] Step S10: Classification output: Input the high-level fusion feature data into the fully connected layer in the multimodal emotion recognition model to obtain the emotion recognition result.

[0020] Furthermore, the step S4 specifically includes the following steps:

[0021] Step S41: Receive the preliminary fusion feature data and input it into the nested He-BiLSTM model. Through the forward propagation forget gate mechanism, preliminarily screen the long-term memory state in the preliminary fusion feature data. Combined with the reverse memory state of the current time step in the nested He-BiLSTM model, generate the nested forward memory feature data through dynamic nesting operation. The formula used is as follows:

[0022] Forget gate calculation formula:

[0023] ;

[0024] in, Represents the time step Indicates the hidden state, represents the forget gate, represents the output value of the forward forget gate, represents the activation function, Represents the weight matrix from the input feature to the forward forget gate; represents the input features of the current time step, Represents the weight matrix from the previous hidden state to the forward forget gate; represents the forward hidden state of the previous time step; Represents the bias term of the forget gate;

[0025] Nested memory feature generation formula:

[0026] ;

[0027] in, Represents the nested forward memory feature data; The weight matrix of the nested mechanism, represents the reverse memory state of the current time step, Represents the forward memory state of the previous time step, that is, the long-term memory state;

[0028] Step S42: The forward newly added forget gate mechanism is used to filter noise and compress information redundancy on the nested forward memory feature data to generate forward high-quality long-term memory feature data. The formula used is as follows:

[0029] ;

[0030] in, express The transpose of express The correlation matrix of

[0031] ;

[0032] in, represents the redundant compression weight matrix, Represents the compressed forward memory feature data

[0033] ;

[0034] in, Indicates the memory status. represents the gate mechanism, Represents the output of the newly added forget gate, Represents the weight matrix of the newly added forget gate;

[0035] ;

[0036] in, Represents forward high-quality long-term memory feature data;

[0037] Step S43: Processing the forward high-quality long-term memory feature data through the output gate mechanism to generate forward hidden state feature data;

[0038] Step S44: Preliminarily screen the future long-term memory state in the preliminary fusion feature data through the backward propagation forget gate mechanism, combine the forward memory state of the current time step in the nested He-BiLSTM model, and generate nested backward memory feature data through dynamic nesting operation;

[0039] Step S45: performing noise filtering and information redundancy compression on the nested backward memory feature data to generate backward high-quality long-term memory feature data;

[0040] Step S46: Processing the backward high-quality long-term memory feature data through the output gate mechanism to generate backward hidden state feature data;

[0041] Step S47: fusing the forward hidden state feature data and the backward hidden state feature data to generate bidirectional time step feature data;

[0042] Step S48: Input the bidirectional time step feature data into the fully connected layer in the nested He-BiLSTM model to generate fused time series feature data.

[0043] Furthermore, the emotion analysis and guidance module obtains a process of optimizing the emotion evolution map and optimizing the psychological state scoring data, which specifically includes the following steps:

[0044] Step C1: Modeling and initialization: Initialize the hyperparameters of the TCN-GRU model, which includes the TCN model and the GRU model;

[0045] Step C2: Data cleaning: Clean the virtual catharsis interaction data, process missing values ​​and outliers, and obtain cleaned interaction data;

[0046] Step C3: Data normalization: Min-Max normalization is performed on the cleaned interaction data, and the cleaned interaction data is scaled to the range of [0, 1] to reduce the feature scale difference and obtain normalized interaction data;

[0047] Step C4: Data alignment: perform time step alignment on the normalized interaction data to obtain aligned time series data;

[0048] Step C5: Clustering processing: Determine the dense area of ​​the aligned time series data through the mean shift algorithm, cluster the aligned time series data through fuzzy C-means, and obtain a similar emotion data set;

[0049] Step C6: Temporal convolutional network processing: Input the similar emotion dataset into the TCN model, use the dilated convolution and causal convolution of the TCN model to extract the long-term dependency of the time series in the similar emotion dataset, capture the long-term characteristics of emotion evolution, and obtain long-term temporal feature data;

[0050] Step C7: Gated recurrent unit processing: input the long-term temporal feature data into the GRU model, use the update gate and reset gate in the GRU model to extract the short-term dynamic features in the long-term temporal feature data, capture short-term emotional fluctuations, and obtain short-term dynamic feature data;

[0051] Step C8: Emotion classification and psychological state assessment: Input the short-term dynamic feature data into the classification layer of the TCN-GRU model, analyze the psychological state, and obtain the emotion evolution map and psychological state score data;

[0052] Step C9: Gray Wolf Optimization: Use the improved version of the GWO algorithm to optimize the hyperparameters of the TCN-GRU model to obtain the optimal hyperparameter combination. According to the optimal hyperparameter combination, the output of the TCN-GRU model is optimized to obtain the optimized emotion evolution map and optimized psychological state score data.

[0053] Furthermore, the step C9 specifically includes the following steps:

[0054] Step C91: Define the optimization objective function: Use the root mean square error as the fitness function to measure the prediction performance of the TCN-GRU model;

[0055] Step C92: Initialize the gray wolf population: randomly generate gray wolf individuals in the gray wolf population, and the gray wolf individuals represent the hyperparameter combination of the TCN-GRU model;

[0056] Step C93: Calculate the fitness of gray wolves: Input the gray wolf individuals into the TCN-GRU model for training, calculate the fitness value of the hyperparameter combination of each gray wolf individual according to the fitness function, and sort the fitness values ​​to obtain the fitness ranking of the gray wolf population;

[0057] Step C94: Update the positions of α, β and δ gray wolves: Select the best α wolf, the second best β wolf and the third best δ wolf as the leader wolf according to the fitness ranking of the gray wolf population, introduce a dynamic leadership mechanism, and allow β wolf and δ wolf to replace α wolf; and dynamically update the positions of other gray wolves through the gray wolf position update formula of particle swarm optimization, and get closer to the leader wolf to optimize the hyperparameter combination. The formula used is as follows:

[0058] Dynamic Leadership Mechanism Formula:

[0059] ;

[0060] in, represents the probability of wolf α being replaced, represents the weight of replacing α wolf with β wolf, represents the weight of replacing α wolf with δ wolf; represents the fitness ranking of wolf α, represents the fitness ranking of β wolf, represents the fitness ranking of δwolves; represents the sensitivity control parameter, represents the natural exponential function;

[0061] The gray wolf position update formula of particle swarm optimization is:

[0062] ;

[0063] in, represents the speed of the individual gray wolf after a new round of iteration, represents the speed of the gray wolf in the last round, represents the inertia weight, represents the cognitive coefficient, represents the social coefficient, and Represents a random number, Indicates the current location of the gray wolf individual, represents the best historical position of the gray wolf individual, represents the global optimal position of the gray wolf population;

[0064] ;

[0065] in, represents the position of the individual gray wolf after a new round of iteration, Indicates the position of the individual gray wolf in the previous round;

[0066] Step C95: Convergence judgment: set the maximum number of iterations, iterate steps C33-C94 until the maximum number of iterations is met, and output the optimal hyperparameter combination.

[0067] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows:

[0068] The present invention realizes the deep fusion and precise analysis of multimodal emotion data by combining the multimodal emotion recognition method with the nested memory feature generation mechanism, the newly added forget gate mechanism and the two-way feature fusion technology. The system can not only effectively extract the temporal characteristics of physiological data and behavioral data, but also comprehensively capture the dynamic changes of emotions, thus improving the accuracy and sensitivity of emotion recognition, thereby providing a more reliable basis for emotion category judgment for the psychological catharsis interaction system. This improvement overcomes the shortcomings of the traditional system, which is highly dependent on a single data source and has low emotion recognition accuracy, significantly improves the performance of the emotion recognition module, and lays a solid foundation for subsequent psychological state analysis and personalized psychological guidance.

[0069] The present invention introduces a dynamic leadership mechanism and a particle swarm optimization mechanism through an emotion analysis and guidance module, improves the position update strategy of the gray wolf optimization algorithm, and optimizes the parameters of the TCN-GRU model, thereby realizing the system's in-depth analysis of emotion evolution and accurate assessment of psychological state; providing students with targeted psychological guidance suggestions; this improvement solves the problem of insufficient comprehensiveness and personalization of psychological state analysis in traditional systems, and enhances the scientific nature and intervention effect of psychological guidance;

[0070] In addition, the virtual catharsis scene module of the present invention can dynamically generate matching virtual scenes and interactive content according to the emotion recognition results, so as to meet the students' psychological catharsis needs in different emotional states; through this module, the present invention realizes the closed-loop design of the psychological catharsis interaction system from emotion recognition to psychological intervention, solves the problems of single scene and rigid interaction mode in traditional psychological catharsis system, and significantly enhances the practicality and scope of application of the system; the present invention comprehensively utilizes multimodal emotion recognition, emotion dynamic analysis and scene generation technology, which improves the efficiency and effect of psychological catharsis as a whole, and has important practical application value in the field of mental health intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A schematic diagram of a module of a psychological catharsis interaction system based on artificial intelligence proposed by the present invention;

[0072] Figure 2 This is a schematic diagram of the structure of the nested He-BiLSTM model in step S4 proposed by the present invention;

[0073] Figure 3 This is a flow chart of the enhanced TCN-GRU model proposed in the present invention;

[0074] Figure 4 This is a flow chart of the improved GWO algorithm proposed in Example 6. DETAILED DESCRIPTION

[0075] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0076] Embodiment 1, according to Figure 1 ,The present invention provides a psychological catharsis interaction system based on artificial intelligence, which includes a data acquisition module, an emotion recognition module, a virtual catharsis scene module, and an emotion analysis and guidance module;

[0077] The data acquisition module collects students' physiological data and student behavior data through VR equipment and sensors, and integrates them to obtain original emotional data; students' physiological data include heart rate, skin electrical response, brain waves and breathing rate, and students' behavior data include body movements, voice intonation, eye movement trajectory and facial expressions;

[0078] The emotion recognition module combines feature extraction technology and multimodal feature fusion technology to build a multimodal emotion recognition model, uses the multimodal emotion recognition model to analyze the original emotion data, identifies the student's emotion category, and outputs the emotion recognition result;

[0079] The virtual catharsis scene module generates matching virtual catharsis scenes and virtual catharsis interaction data according to the emotion recognition results; the virtual catharsis scenes include emotion venting scenes, emotion release scenes and confession scenes;

[0080] The emotion analysis and guidance module establishes a TCN-GRU model, improves the position update strategy of the GWO algorithm by introducing a dynamic leadership mechanism and a particle swarm optimization mechanism, constructs an improved GWO algorithm, uses the improved GWO algorithm to optimize the hyperparameters of the TCN-GRU model, and constructs an enhanced TCN-GRU model; uses the enhanced TCN-GRU model to analyze virtual catharsis interaction data and psychological states, obtains an optimized emotion evolution map and optimized psychological state scoring data; and provides personalized psychological guidance based on the obtained optimized emotion evolution map and optimized psychological state scoring data.

[0081] Embodiment 2, according to Figure 2 Based on the first embodiment, the process of the emotion recognition module outputting the emotion recognition result specifically includes the following steps:

[0082] Step S1: Physiological feature extraction: using wavelet transform to extract the time-frequency domain features of students' physiological data, and performing dimensionality reduction and optimization through sparse autoencoder to obtain students' physiological feature data;

[0083] Step S2: Behavior feature extraction: extracting feature data of student behavior data through eGeMAPS feature set, convolutional neural network and gated recurrent unit respectively to obtain student behavior feature data;

[0084] Step S3: Preliminary fusion: Preliminary fusion feature data is obtained by preliminarily fusion of the student's physiological feature data and the student's behavioral feature data;

[0085] Step S4: Advanced feature fusion: A nested He-BiLSTM model is constructed through a nested memory feature generation mechanism, a newly added forget gate mechanism, and bidirectional feature fusion. The preliminary fused feature data is input into the nested He-BiLSTM model to capture the bidirectional time dependency in the feature data and obtain fused time series feature data. The nested He-BiLSTM model includes a forward I-LSTM and a backward I-LSTM.

[0086] Step S5: Attention weighting: The time series feature data is processed and fused through the attention mechanism, and importance weights are assigned to obtain weighted time feature data;

[0087] Step S6: layer normalization: performing layer normalization processing on the weighted time feature data to reduce the influence of feature scale differences and obtain normalized time feature data;

[0088] Step S7: nonlinear activation: input the normalized time feature data into the ReLU activation function, perform nonlinear mapping, capture the nonlinear relationship between features, and obtain nonlinear time feature data;

[0089] Step S8: Regularization processing: input the nonlinear time feature data into the Dropout layer in the multimodal emotion recognition model, randomly discard some neurons to prevent overfitting, and obtain regularized time feature data;

[0090] Step S9: residual connection: performing residual connection on the preliminary fusion feature data and the regularized time feature data to obtain advanced fusion feature data;

[0091] Step S10: Classification output: Input the high-level fusion feature data into the fully connected layer in the multimodal emotion recognition model to obtain the emotion recognition result.

[0092] Embodiment 3, based on embodiment 2, the step S4 specifically includes the following steps:

[0093] Step S41: Receive the preliminary fusion feature data and input it into the nested He-BiLSTM model. Through the forward propagation forget gate mechanism, preliminarily screen the long-term memory state in the preliminary fusion feature data. Combined with the reverse memory state of the current time step in the nested He-BiLSTM model, generate the nested forward memory feature data through dynamic nesting operation. The formula used is as follows:

[0094] Forget gate calculation formula:

[0095] ;

[0096] in, represents the time step, Indicates the hidden state, represents the forget gate, represents the output value of the forward forget gate, represents the activation function, Represents the weight matrix from the input feature to the forward forget gate; represents the input features of the current time step, Represents the weight matrix from the previous hidden state to the forward forget gate; represents the forward hidden state of the previous time step; Represents the bias term of the forget gate;

[0097] Nested memory feature generation formula:

[0098] ;

[0099] in, Represents the nested forward memory feature data; The weight matrix of the nested mechanism, represents the reverse memory state of the current time step, Represents the forward memory state of the previous time step, that is, the long-term memory state;

[0100] Step S42: The forward newly added forget gate mechanism is used to filter noise and compress information redundancy on the nested forward memory feature data to generate forward high-quality long-term memory feature data. The formula used is as follows:

[0101] ;

[0102] in, express The transpose of express The correlation matrix of

[0103] ;

[0104] in, represents the redundant compression weight matrix, Represents the compressed forward memory feature data

[0105] ;

[0106] in, Indicates the memory status. represents the gate mechanism, Represents the output of the newly added forget gate, Represents the weight matrix of the newly added forget gate;

[0107] ;

[0108] in, Represents forward high-quality long-term memory feature data;

[0109] Step S43: Processing the forward high-quality long-term memory feature data through the output gate mechanism to generate forward hidden state feature data;

[0110] Step S44: Preliminarily screen the future long-term memory state in the preliminary fusion feature data through the backward propagation forget gate mechanism, combine the forward memory state of the current time step in the nested He-BiLSTM model, and generate nested backward memory feature data through dynamic nesting operation;

[0111] Step S45: performing noise filtering and information redundancy compression on the nested backward memory feature data to generate backward high-quality long-term memory feature data;

[0112] Step S46: Processing the backward high-quality long-term memory feature data through the output gate mechanism to generate backward hidden state feature data;

[0113] Step S47: fusing the forward hidden state feature data and the backward hidden state feature data to generate bidirectional time step feature data;

[0114] Step S48: Input the bidirectional time step feature data into the fully connected layer in the nested He-BiLSTM model to generate fused time series feature data.

[0115] Embodiment 4, based on embodiment 2, the step S4 specifically includes the following steps:

[0116] Step S41: Establish a He-BiLSTM model, receive preliminary fusion feature data, input it into the He-BiLSTM model, perform preliminary screening of the long-term memory state in the preliminary fusion feature data through the forward propagation forget gate mechanism, and combine the reverse memory state of the current time step in the He-BiLSTM model to generate preliminary forward memory feature data;

[0117] ;

[0118] in, represents preliminary forward memory feature data;

[0119] Step S42: Processing the preliminary forward memory feature data through the output gate mechanism to generate forward hidden state feature data;

[0120] Step S43: Preliminarily screen the future long-term memory state in the preliminary fusion feature data through the backward propagation forget gate mechanism, and generate preliminary backward memory feature data in combination with the forward memory state of the current time step in the He-BiLSTM model;

[0121] Step S44: Processing the preliminary backward memory feature data through the output gate mechanism to generate backward hidden state feature data;

[0122] Step S45: fusing the forward hidden state feature data and the backward hidden state feature data to generate bidirectional time step feature data;

[0123] Step S46: Input the bidirectional time step feature data into the fully connected layer in the He-BiLSTM model to generate fused time series feature data.

[0124] Embodiment 5, according to Figure 3 Based on the third embodiment, the emotion analysis and guidance module obtains the process of optimizing the emotion evolution map and optimizing the psychological state scoring data, which specifically includes the following steps:

[0125] Step C1: Modeling and initialization: Initialize the hyperparameters of the TCN-GRU model, which includes the TCN model and the GRU model;

[0126] Step C2: Data cleaning: Clean the virtual catharsis interaction data, process missing values ​​and outliers, and obtain cleaned interaction data;

[0127] Step C3: Data normalization: Min-Max normalization is performed on the cleaned interaction data, and the cleaned interaction data is scaled to the range of [0, 1] to reduce the feature scale difference and obtain normalized interaction data;

[0128] Step C4: Data alignment: perform time step alignment on the normalized interaction data to obtain aligned time series data;

[0129] Step C5: Clustering processing: Determine the dense area of ​​the aligned time series data through the mean shift algorithm, cluster the aligned time series data through fuzzy C-means, and obtain a similar emotion data set;

[0130] Step C6: Temporal convolutional network processing: Input the similar emotion dataset into the TCN model, use the dilated convolution and causal convolution of the TCN model to extract the long-term dependency of the time series in the similar emotion dataset, capture the long-term characteristics of emotion evolution, and obtain long-term temporal feature data;

[0131] Step C7: Gated recurrent unit processing: input the long-term temporal feature data into the GRU model, use the update gate and reset gate in the GRU model to extract the short-term dynamic features in the long-term temporal feature data, capture short-term emotional fluctuations, and obtain short-term dynamic feature data;

[0132] Step C8: Emotion classification and psychological state assessment: Input the short-term dynamic feature data into the classification layer of the TCN-GRU model, analyze the psychological state, and obtain the emotion evolution map and psychological state score data;

[0133] Step C9: Gray Wolf Optimization: Use the improved GWO algorithm to optimize the hyperparameters of the TCN-GRU model to obtain the optimal hyperparameter combination. According to the optimal hyperparameter combination, the output of the TCN-GRU model is optimized to obtain the optimized emotion evolution map and optimized psychological state score data.

[0134] Embodiment 6, according to Figure 4 Based on the fifth embodiment, the step C9 specifically includes the following steps:

[0135] Step C91: Define the optimization objective function: Use the root mean square error as the fitness function to measure the prediction performance of the TCN-GRU model;

[0136] Step C92: Initialize the gray wolf population: randomly generate gray wolf individuals in the gray wolf population, and the gray wolf individuals represent the hyperparameter combination of the TCN-GRU model;

[0137] Step C93: Calculate the fitness of gray wolves: Input the gray wolf individuals into the TCN-GRU model for training, calculate the fitness value of the hyperparameter combination of each gray wolf individual according to the fitness function, and sort the fitness values ​​to obtain the fitness ranking of the gray wolf population;

[0138] Step C94: Update the positions of α, β and δ gray wolves: Select the best α wolf, the second best β wolf and the third best δ wolf as the leader wolf according to the fitness ranking of the gray wolf population, introduce a dynamic leadership mechanism, and allow β wolf and δ wolf to replace α wolf; and dynamically update the positions of other gray wolves through the gray wolf position update formula of particle swarm optimization, and get closer to the leader wolf to optimize the hyperparameter combination. The formula used is as follows:

[0139] Dynamic Leadership Mechanism Formula:

[0140] ;

[0141] in, represents the probability of wolf α being replaced, represents the weight of replacing α wolf with β wolf, represents the weight of replacing α wolf with δ wolf; represents the fitness ranking of wolf α, represents the fitness ranking of β wolf, represents the fitness ranking of δwolves; represents the sensitivity control parameter, represents the natural exponential function;

[0142] The gray wolf position update formula of particle swarm optimization is:

[0143] ;

[0144] in, represents the speed of the individual gray wolf after a new round of iteration, represents the speed of the gray wolf in the last round, represents the inertia weight, represents the cognitive coefficient, represents the social coefficient, and Represents a random number, Indicates the current location of the gray wolf individual, represents the best historical position of the gray wolf individual, represents the global optimal position of the gray wolf population;

[0145] ;

[0146] in, represents the position of the individual gray wolf after a new round of iteration, Indicates the position of the individual gray wolf in the previous round;

[0147] Step C95: Convergence determination: set the maximum number of iterations, iterate step C93-step C94 until the maximum number of iterations is met, and output the optimal hyperparameter combination;

[0148] Set the maximum number of iterations: 150.

[0149] Embodiment 7, based on embodiment 5, the step C9 specifically includes the following steps:

[0150] Step C91: Define the optimization objective function: Use the root mean square error as the fitness function to measure the prediction performance of the TCN-GRU model;

[0151] Step C92: Initialize the gray wolf population: randomly generate gray wolf individuals in the gray wolf population, and the gray wolf individuals represent the hyperparameter combination of the TCN-GRU model;

[0152] Step C93: Calculate the fitness of gray wolves: Input the gray wolf individuals into the TCN-GRU model for training, calculate the fitness value of the hyperparameter combination of each gray wolf individual according to the fitness function, and sort the fitness values ​​to obtain the fitness ranking of the gray wolf population;

[0153] Step C94: Update the positions of α, β and δ gray wolves: Select the best α wolf, the second best β wolf and the third best δ wolf as the leader wolf according to the fitness ranking of the gray wolf population; and dynamically update the positions of other gray wolves through the gray wolf position update formula to get closer to the leader wolf to optimize the hyperparameter combination. The formula used is as follows:

[0154] Gray wolf position update formula:

[0155] ;

[0156] in, represents a multidimensional random vector, represents the distance vector between the current position of the gray wolf and the leading wolf, represents the convergence factor, Indicates the current iteration number, represents the exponential decay term;

[0157] Step C95: Convergence determination: set the maximum number of iterations, iterate steps C93-C84 until the maximum number of iterations is met, and output the optimal hyperparameter combination;

[0158] Set the maximum number of iterations: 150.

[0159] The present invention and its embodiments are described above, which is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual structure is not limited to this. In short, if ordinary technicians in this field are inspired by it and do not deviate from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.

Claims

1. A psychological catharsis interactive system based on artificial intelligence, comprising a data acquisition module, wherein the data acquisition module collects students' physiological data and students' behavioral data, and integrates them to obtain original emotional data, characterized in that: The system also includes an emotion recognition module, a virtual catharsis scene module, and an emotion analysis and guidance module; The emotion recognition module combines feature extraction technology and multimodal feature fusion technology to build a multimodal emotion recognition model, uses the multimodal emotion recognition model to analyze the original emotion data, and outputs the emotion recognition result; The virtual catharsis scene module generates a matching virtual catharsis scene and virtual catharsis interaction data according to the emotion recognition result; The emotion analysis and guidance module establishes a TCN-GRU model, improves the position update strategy of the GWO algorithm by introducing a dynamic leadership mechanism and a particle swarm optimization mechanism, constructs an improved GWO algorithm, uses the improved GWO algorithm to optimize the hyperparameters of the TCN-GRU model, and constructs an enhanced TCN-GRU model; uses the enhanced TCN-GRU model to analyze virtual catharsis interaction data to obtain an optimized emotion evolution map and optimized psychological state score data.

2. The artificial intelligence-based psychological catharsis interactive system according to claim 1, characterized in that: The process of outputting the emotion recognition result by the emotion recognition module specifically includes the following steps: Step S1: extracting characteristic data of student physiological data to obtain student physiological characteristic data; Step S2: extracting characteristic data of student behavior data to obtain student behavior characteristic data; Step S3: Preliminary fusion feature data is obtained by preliminarily fusing the student's physiological feature data and the student's behavioral feature data; Step S4: construct a nested He-BiLSTM model through a nested memory feature generation mechanism, a newly added forget gate mechanism, and bidirectional feature fusion, input the preliminary fused feature data into the nested He-BiLSTM model, and obtain fused time series feature data; Step S5: attention weighting; Step S6: layer normalization; Step S7: nonlinear activation; Step S8: regularization processing; Step S9: residual connection; Step S10: Classification output: The fused time series feature data is subjected to attention weighting, layer normalization, nonlinear activation, regularization processing and residual connection to obtain the emotion recognition result.

3. The artificial intelligence-based psychological catharsis interactive system according to claim 2, characterized in that: The step S4 specifically comprises the following steps: Step S41: receiving preliminary fusion feature data, inputting it into the nested He-BiLSTM model, preliminarily screening the long-term memory state in the preliminary fusion feature data through the forward propagation forget gate mechanism, combining the reverse memory state of the current time step in the nested He-BiLSTM model, and generating nested forward memory feature data through dynamic nesting operation; Step S42: noise filtering and information redundancy compression are performed on the nested forward memory feature data through a forward newly added forget gate mechanism to generate forward high-quality long-term memory feature data; Step S43: Processing the forward high-quality long-term memory feature data through the output gate mechanism to generate forward hidden state feature data; Step S44: Preliminarily screen the future long-term memory state in the preliminary fusion feature data through the backward propagation forget gate mechanism, combine the forward memory state of the current time step in the nested He-BiLSTM model, and generate nested backward memory feature data through dynamic nesting operation; Step S45: performing noise filtering and information redundancy compression on the nested backward memory feature data to generate backward high-quality long-term memory feature data; Step S46: Processing the backward high-quality long-term memory feature data through the output gate mechanism to generate backward hidden state feature data; Step S47: fusing the forward hidden state feature data and the backward hidden state feature data to generate bidirectional time step feature data; Step S48: Input the bidirectional time step feature data into the fully connected layer in the nested He-BiLSTM model to generate fused time series feature data.

4. The artificial intelligence-based psychological catharsis interactive system according to claim 1, characterized in that: The emotion analysis and guidance module obtains the process of optimizing the emotion evolution map and optimizing the psychological state scoring data, which specifically includes the following steps: Step C1: Initialize the hyperparameters of the TCN-GRU model, which includes the TCN model and the GRU model; Step C2: data cleaning; Step C3: data normalization; Step C4: Data alignment: Perform data cleaning, data normalization and data alignment on the virtual catharsis interaction data to obtain aligned time series data; Step C5: Determine the dense area of ​​the aligned time series data by using the mean shift algorithm, cluster the aligned time series data by using fuzzy C-means to obtain a similar emotion data set; Step C6: Input similar emotion datasets into the TCN model to capture the long-term characteristics of emotion evolution and obtain long-term temporal feature data; Step C7: Input the long-term temporal feature data into the GRU model to capture short-term emotional fluctuations and obtain short-term dynamic feature data; Step C8: Input the short-term dynamic feature data into the classification layer of the TCN-GRU model, and output the emotion evolution map and psychological state score data; Step C9: Use the improved GWO algorithm to optimize the hyperparameters of the TCN-GRU model to obtain the optimal hyperparameter combination, optimize the output of the TCN-GRU model based on the optimal hyperparameter combination, and obtain the optimized emotion evolution map and optimized psychological state score data.

5. The artificial intelligence-based psychological catharsis interactive system according to claim 4, characterized in that: The step C9 specifically comprises the following steps: Step C91: using the root mean square error as the fitness function; Step C92: randomly generate gray wolf individuals from the gray wolf population, where the gray wolf individuals represent the hyperparameter combination of the TCN-GRU model; Step C93: input the gray wolf individuals into the TCN-GRU model for training, calculate the fitness value of the hyperparameter combination of each gray wolf individual according to the fitness function, and sort the fitness values ​​to obtain the fitness ranking of the gray wolf population; Step C94: Select the best α wolf, the second best β wolf and the third best δ wolf as the leader wolves according to the fitness ranking of the gray wolf population, introduce a dynamic leadership mechanism, and allow β wolf and δ wolf to replace α wolf; and dynamically update the positions of other gray wolves through the gray wolf position update rule of particle swarm optimization, so as to get closer to the leader wolf and optimize the hyperparameter combination; Step C95: Set the maximum number of iterations, iterate steps C93-C94 until the maximum number of iterations is met, and output the optimal hyperparameter combination.

Images