A method for identifying psychological stress levels based on deep emotional neural networks

Abstract

This invention discloses a method for identifying psychological stress levels based on a deep emotional neural network, belonging to the field of computer technology. It includes information acquisition: acquiring past stress information and real-time multimodal feature information of the user. This invention uses a weighted processing method to derive derived stress information under different weight distributions based on the weight distribution of past stress information, facilitating the establishment of stress recognition models under different weight distributions. A set directory table records the relationship between different weight distributions and the stress recognition models, allowing for the search of corresponding stress recognition models based on weight distributions. The target stress level is obtained by combining the target stress recognition model with the processed real-time multimodal feature information and stress level, thus fully considering individual differences in expression, adapting to different individual expression styles, reducing the probability of impacting individual mental health, and improving user trust and user experience.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a method for identifying psychological stress levels based on deep emotional neural networks. Background Technology

[0002] Psychological stress refers to the psychological and physiological reactions an individual experiences when facing various stressors in life, work, or study. It may manifest as emotional and physical reactions such as anxiety, tension, irritability, and insomnia. Moderate stress can motivate people, but excessive stress can affect physical and mental health and even trigger mental illness.

[0003] A method and apparatus for identifying psychological stress based on a neural network, disclosed in patent publication number CN111063420A, involves acquiring photoplethysmography (PPG) data of a user whose psychological stress is to be identified; segmenting the PPG data to obtain segmented PPG data; equalizing the length of the segmented PPG data to obtain PPG data of equal length; and inputting the equal-length PPG data into a psychological stress identification model to obtain the user's psychological stress information. The psychological stress identification model is obtained by pre-training a preset neural network model using PPG training data labeled with psychological stress information. Based on this scheme, psychological stress can be identified using a neural network.

[0004] Most of the neural networks used to identify psychological stress in the above-mentioned and similar principles are trained under clear signal conditions, lacking robustness to complex noisy environments and failing to fully consider individual differences in expression. This makes it difficult for the model to identify key emotional features representing stress, resulting in poor performance across different populations and an inability to adjust to adapt to different individual expressions. Furthermore, there is a tendency to mistakenly identify users with high stress as having low stress, leading to insufficient psychological intervention and impacting individual mental health. Frequent false alarms may also cause user distrust and negative experiences, hindering the system's promotion and adoption. Therefore, this invention is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying psychological stress levels based on deep emotional neural networks, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying psychological stress levels based on a deep emotional neural network, comprising:

[0007] Information Acquisition: Acquire past stress information and real-time multimodal feature information of users. Past stress information includes past multimodal feature information and stress information corresponding to the past multimodal feature information.

[0008] Information processing: Obtain the target noise by acquiring noise information of the application scenario, and then perform noise reduction on the real-time multimodal feature information based on the target noise using information processing methods to obtain the processed real-time multimodal feature information;

[0009] Its features include:

[0010] Pressure level classification: Pressure levels are obtained by classifying pressure information using a classification method. Each pressure level includes several sub-pressure levels.

[0011] Feature weight distribution: Based on past pressure information, derived pressure information is obtained by weighting the past pressure information according to different feature weights using a weighting method;

[0012] Model Establishment: Based on derived pressure information and past pressure information, pressure recognition models under different feature weight distributions are established, resulting in pressure recognition models and a catalog table;

[0013] Stress Level Identification: Determine the user's expression of stress to obtain change weights. Based on the catalog, select the stress identification model corresponding to the change weights using a selection method to obtain the target stress identification model. Import the processed real-time multimodal feature information into the target stress identification model to obtain the target stress. Based on the target stress and stress level, obtain the target stress level.

[0014] The weight processing method includes: obtaining the basic weight distribution of the previous multimodal feature information in the previous pressure information to obtain the basic weight, the basic weight includes the sub-weight of a single feature information in the previous multimodal feature information, a preset adjustment factor, and obtaining the derived weight by increasing and decreasing the sub-weight in the basic weight in the same way based on the adjustment factor, obtaining the sub-derived feature information based on the derived weight and the previous multimodal feature information, obtaining the pressure of the sub-derived feature information to obtain the sub-derived pressure, establishing the correspondence between the sub-derived feature information and the sub-derived pressure, and integrating the sub-derived feature information and the sub-derived pressure to obtain the derived pressure information;

[0015] The information processing method includes: amplifying and reducing noise based on target noise to obtain derived noise; establishing a noise database to store target noise and derived noise; extracting speech information from multimodal feature information to obtain information to be processed; establishing a timeline; matching the information to be processed with the target noise and derived noise in the noise database based on the timeline to obtain a matching result; extracting the sound information corresponding to the target noise and derived noise in the information to be processed based on the matching result to obtain processed information to be processed; and replacing the information to be processed in the multimodal feature information with the processed information to obtain processed real-time multimodal information.

[0016] Furthermore, the level division method includes: presetting the number of level divisions, the optimal pressure value range, and the decreasing value. The optimal pressure value range is the first sub-pressure level. The optimal pressure value range includes a first boundary value and a second boundary value. The value of the first boundary value is greater than the value of the second boundary value. A third boundary value is obtained based on the second boundary value and the decreasing value. The second boundary value and the third boundary value are integrated to obtain the second optimal pressure value range. The second optimal pressure value range is the second sub-pressure level. A third sub-pressure level is obtained based on the third boundary value and the decreasing value. The first sub-pressure level, the second sub-pressure level, and the third sub-pressure level are integrated to obtain the pressure level.

[0017] Furthermore, the process of establishing a stress identification model is as follows:

[0018] Training data acquisition and processing: The derived pressure information and the previous pressure information are preprocessed to obtain the preprocessed derived pressure information and the preprocessed previous pressure information. Both the preprocessed derived pressure information and the preprocessed previous pressure information are training data.

[0019] Model selection and training: Select a deep learning model as the model base, import the training data into the model base, and train the model base to obtain the initial model;

[0020] Model Adjustment and Output: Preset verification information, including verification feature information and verification pressure. Import the verification feature information into the initial model to obtain the result pressure. Judge the result pressure and the verification pressure to obtain the judgment result. Based on the judgment result, optimize and adjust the initial model to obtain the pressure recognition model.

[0021] Model labeling: Label information is obtained based on the weights of the training data. The association between the label information and the stress recognition model is established to obtain target association. A directory table is established to store the target association.

[0022] Furthermore, the process of determining the user's expression of pressure and obtaining the change weight is as follows: preset basic weights, fixed values, and change thresholds. The basic weights include several sub-basic weights corresponding to individual submodal feature information. Establish time nodes, and obtain sub-change values ​​by monitoring the changes of individual submodal feature information in real-time multimodal feature information based on adjacent time nodes. Determine the relationship between all sub-change values ​​and the change threshold, extract sub-change values ​​that exceed the change threshold to obtain the target change value, add the fixed value to the sub-basic weights of the target change value to obtain the target weight, extract the sub-change values ​​that do not exceed the change threshold to obtain the fixed change value, determine the number of fixed change values ​​to obtain the reduction quantity, average the fixed value based on the reduction quantity to obtain the average value, subtract the average value from the sub-basic weights of the fixed change values ​​to obtain the fixed weight, and integrate all target weights and fixed weights to obtain the change weight.

[0023] Furthermore, the selection method includes: traversing the directory table based on the changing weights to obtain traversal results; when the traversal results indicate that there is a weight in the directory table corresponding to the changing weights, extracting the pressure recognition model corresponding to the changing weights from the directory table to obtain the target pressure recognition model; when the traversal results indicate that there is no weight in the directory table corresponding to the changing weights, extracting the pressure recognition model closest to the changing weights from the directory table to obtain the target pressure recognition model.

[0024] Furthermore, user information is obtained by integrating real-time multimodal feature information, change weights, target stress levels, and target stress recognition models. A repository is established to store user information, and recognition requirements are obtained. Based on the recognition requirements, the user information corresponding to the recognition requirements is searched in the repository to obtain search results.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] This psychological stress level recognition method based on deep emotional neural networks derives derived stress information under different weight distributions by setting a weight processing method based on the weight distribution of previous stress information. This facilitates the establishment of stress recognition models under different weight distributions. Through a set catalog table, the relationship between different weight distributions and stress recognition models can be recorded, so as to find the corresponding stress recognition model according to the weight distribution. The target stress level is obtained by combining the target stress recognition model with the processed real-time multimodal feature information and stress level. This method fully considers the differences in individual expression, adapts to different individual expression styles, reduces the probability of affecting individual mental health, and improves user trust and user experience.

[0027] Simultaneously, by using the set information processing method to denoise the sound information in the real-time multimodal feature information, the processed real-time multimodal feature information is obtained. Through denoising, the accuracy of the multimodal feature information can be improved, thereby improving the accuracy of subsequent identification of psychological stress levels and facilitating use. By using the set level division method, stress levels can be set according to the distribution of stress information and usage needs, so that the stress levels meet the needs of administrators. By using the set weight processing method, derived stress information under different weight distributions can be obtained based on the weight distribution of previous stress information, so as to establish stress recognition models under different weight distributions.

[0028] Simultaneously, a directory table is established to store target relationships, facilitating the subsequent generation of corresponding stress recognition models based on their respective weights. Training a single stress recognition model using training data under a single weight distribution improves the training efficiency of the stress recognition model. Furthermore, separating stress recognition models under different weight distributions allows for simultaneous processing of multimodal feature information under corresponding weight distributions, thereby enhancing stress recognition efficiency. By matching user recognition needs with user information in the repository to obtain search results, it is possible to determine whether a user has previously used this method for psychological stress recognition and obtain the result of the user's last psychological stress level recognition. This allows for further assessment of changes in the user's psychological state, improving the user experience. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the main process structure of the present invention;

[0030] Figure 2 This is a schematic diagram of the structure for obtaining the target pressure level according to the present invention;

[0031] Figure 3 This is a schematic diagram of the characteristic numerical change structure of the present invention;

[0032] Figure 4 This is a schematic diagram of the catalog structure of the present invention;

[0033] Figure 5 This is a schematic diagram of the derived weight structure of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Identifying psychological stress levels can help detect an individual's high-pressure state early, enabling scientific assessment and timely intervention, thereby preventing mental health problems and promoting physical and mental health development. The psychological stress level identification method based on deep emotional neural networks uses deep learning technology to automatically determine the user's psychological stress level based on the user's emotional expression (such as voice, text, and physiological signals).

[0036] like Figures 1-5 As shown, the present invention provides a technical solution: a method for identifying psychological stress levels based on deep emotional neural networks, comprising:

[0037] Information Acquisition: Acquire past stress information and real-time multimodal feature information of users. Past stress information includes past multimodal feature information and stress information corresponding to the past multimodal feature information.

[0038] It should be noted that real-time multimodal feature information and past multimodal feature information include speech, facial expressions, physiological signals and behavioral actions. Past stress information can be obtained from the operation logs of the corresponding model used for previous psychological stress level recognition, while real-time multimodal feature information is obtained by installing corresponding sensors and camera components.

[0039] Information processing: Obtain the target noise by acquiring noise information of the application scenario, and then perform noise reduction on the real-time multimodal feature information based on the target noise using information processing methods to obtain the processed real-time multimodal feature information;

[0040] It is important to note that the target noise is the noise information of the application scenario. Specifically, it can be obtained by installing a noise sensor in the application scenario. The sound information in the real-time multimodal feature information is denoised using the set information processing method to obtain the processed real-time multimodal feature information. By denoising, the accuracy of the multimodal feature information can be improved, thereby improving the accuracy of subsequent identification of psychological stress levels and making it easier to use.

[0041] Its features include:

[0042] Pressure level classification: Pressure levels are obtained by classifying pressure information using a classification method. Each pressure level includes several sub-pressure levels.

[0043] It is important to note that the pressure levels should be set according to the distribution of pressure information and usage requirements through the set level classification method, so that the pressure levels meet the administrator's needs.

[0044] Feature weight distribution: Based on past pressure information, derived pressure information is obtained by weighting the past pressure information according to different feature weights using a weighting method;

[0045] It is important to note that by using the weighting processing method, derived pressure information under different weight distributions can be obtained based on the weight distribution of previous pressure information, so as to establish pressure recognition models under different weight distributions.

[0046] Model Establishment: Based on derived pressure information and past pressure information, pressure recognition models under different feature weight distributions are established, resulting in pressure recognition models and a catalog table;

[0047] It is important to note that stress recognition models are built based on derived stress information and past stress information. The number of stress recognition models is consistent with the number of different weight distributions. By setting up a directory table, the relationship between different weight distributions and stress recognition models can be recorded, so as to find the corresponding stress recognition model according to the weight distribution. This is to fully consider the differences in individual performance, adapt to different individual expression styles, reduce the probability of affecting individual mental health, and improve user trust and user experience.

[0048] Stress Level Identification: Determine the user's expression of stress to obtain change weights. Based on the catalog table, select the stress identification model corresponding to the change weights to obtain the target stress identification model. Import the processed real-time multimodal feature information into the target stress identification model to obtain the target stress. Based on the target stress and stress level, obtain the target stress level.

[0049] It is important to note that the process of determining the user's expression of stress and obtaining the change weights involves several steps. First, based on the user's expression, a weight distribution matching the user's individual characteristics is determined. Then, a stress recognition model corresponding to the change weights is selected using a selection method based on a directory table. Next, a stress recognition model relevant to the user is found based on the weight distribution matching the user's individual characteristics to obtain the target stress recognition model. Finally, the processed real-time multimodal feature information is imported into the target stress recognition model to obtain the target stress. Based on the obtained target stress and stress levels, the distribution of the target stress within each stress level is determined to obtain the target stress level.

[0050] In practical use, by obtaining the address information of the administrator or user, the generated target pressure level is fed back to the administrator or user based on the address information, thus completing the presentation of the pressure identification results to the administrator or user.

[0051] like Figure 1 As shown, the pressure level classification method includes: a preset number of pressure levels, an optimal pressure value range, and a decreasing value. The optimal pressure value range is the first sub-pressure level. The optimal pressure value range includes a first boundary value and a second boundary value. The value of the first boundary value is greater than the value of the second boundary value. Based on the second boundary value and the decreasing value, a third boundary value is obtained. The second boundary value and the third boundary value are integrated to obtain the second optimal pressure value range. The second optimal pressure value range is the second sub-pressure level. Based on the third boundary value and the decreasing value, the third boundary value is used to obtain the third sub-pressure level. The first sub-pressure level, the second sub-pressure level, and the third sub-pressure level are integrated to obtain the pressure level.

[0052] It is important to note that the number of pressure levels refers to the specific number of pressure levels. The optimal pressure value range is the range of pressure values ​​with the lowest possible response pressure. Decreasing values ​​are the values ​​that are reduced. The second sub-pressure level is obtained by selecting the smaller boundary of the optimal pressure value range and combining it with the decreasing value. This process is repeated to obtain several sub-pressure levels. The pressure level is obtained by integrating all sub-pressure levels. By setting the pressure level classification method, administrators can customize the corresponding pressure levels according to their own needs, thereby improving the user experience.

[0053] In the specific implementation process, the number of pressure levels can be set to four, namely the first sub-pressure level, the second sub-pressure level, the third sub-pressure level, and the fourth sub-pressure level. The optimal pressure value range is a pressure value of 75-100. At this point, the lower the pressure value, the greater the pressure. The decreasing value is 25. At this point, the pressure value range of the first pressure level is 75-100, the pressure value range of the second pressure level is 50-75, the pressure value range of the third pressure level is 25-50, and the pressure value range of the fourth pressure level is 0-25.

[0054] like Figure 1 and Figure 5 As shown, the weight processing method includes: obtaining the basic weight distribution of the previous multimodal feature information in the previous pressure information to obtain the basic weight; the basic weight includes the sub-weight of the corresponding single feature information in the previous multimodal feature information; a preset adjustment factor; the sub-weights in the basic weight are increased and decreased in the same way based on the adjustment factor to obtain the derived weight; the derived weight is combined with the previous multimodal feature information to obtain the sub-derived feature information; the pressure of the sub-derived feature information is obtained to obtain the sub-derived pressure; the correspondence between the sub-derived feature information and the sub-derived pressure is established; and the sub-derived feature information and the sub-derived pressure are integrated to obtain the derived pressure information.

[0055] It is important to note that the process of obtaining the basic weights from the basic weight distribution of previous multimodal feature information in the previous stress information involves obtaining the weight distribution or attention of the previous multimodal feature information from the production logs of the model used in the past to obtain the basic weights. The adjustment factor is a specific value. The sub-weights in the basic weights are increased or decreased accordingly by adjusting the adjustment factor to obtain the derived weights. The process of obtaining the sub-derived pressure from the pressure of the sub-derived feature information can be obtained by directly obtaining the pressure generated by multimodal feature information under the same weight distribution, or by obtaining it from the stress recognition model under the corresponding weight distribution. By integrating the sub-derived feature information and the sub-derived pressure to obtain the derived pressure information, the training data for subsequent training of stress recognition models under different weight distributions can be obtained.

[0056] In specific implementation, feature contribution or attention weights are extracted from existing stress recognition models (such as multimodal fusion networks, attention mechanisms, and feature importance analysis) to obtain basic weights. Based on the derived weights of the derived stress information, existing stress recognition models corresponding to the derived weights are searched. Sub-derived feature information is imported into the stress recognition model to obtain sub-derived stress. Integrating the sub-derived stress with the sub-derived feature information yields the derived stress information. Figure 5 As shown, the basic weight, derived weight 1, and derived weight 2 include five features. The area size of the sub-weights in the basic weight, derived weight 1, and derived weight 2 represents the corresponding weight. In the overall weight, shrinking the area is consistent with increasing the area.

[0057] like Figure 1 and Figure 4 As shown, the process of establishing the pressure recognition model is as follows:

[0058] Training data acquisition and processing: The derived pressure information and the previous pressure information are preprocessed to obtain the preprocessed derived pressure information and the preprocessed previous pressure information. Both the preprocessed derived pressure information and the preprocessed previous pressure information are training data.

[0059] It should be noted that preprocessing the derived stress information and the previous stress information can improve their accuracy. The preprocessed derived stress information and the preprocessed previous stress information are both training data. That is, the preprocessed derived stress information is a single set of training data, and the preprocessed previous stress information is also a single set of training data.

[0060] Model selection and training: Select a deep learning model as the model base, import the training data into the model base, and train the model base to obtain the initial model;

[0061] It is important to note that deep learning models can be convolutional neural networks, etc., and the initial model is obtained by importing training data into the model base for training.

[0062] Model Adjustment and Output: Preset verification information, including verification feature information and verification pressure. Import the verification feature information into the initial model to obtain the result pressure. Judge the result pressure and the verification pressure to obtain the judgment result. Based on the judgment result, optimize and adjust the initial model to obtain the pressure recognition model.

[0063] It is important to note that validation information can be obtained by selecting one or more pieces of information from the training data, and renaming the multimodal feature information and stress in the information to obtain validation feature information and validation stress. By importing the validation feature information into the initial model, the result stress under the validation feature information can be obtained. The stress recognition model can be obtained by optimizing and adjusting the initial model according to the difference between the result stress and the validation stress.

[0064] Model labeling: Label information is obtained based on the weights of the training data. The association between the label information and the stress recognition model is established to obtain target association. A directory table is established to store the target association.

[0065] It's important to note that the model labeling process involves separating the stress recognition models trained on training data under different weight distributions and creating a directory table to store target relationships. This allows for the subsequent generation of corresponding stress recognition models based on their respective weights. Training a single stress recognition model using training data under a single weight distribution improves the training efficiency of the stress recognition model. Furthermore, separating stress recognition models under different weight distributions allows for simultaneous processing of multimodal feature information under their respective weight distributions, thereby enhancing the efficiency of stress recognition. Figure 4 As shown, the catalog includes several stress recognition models and the weight distributions corresponding to the training data.

[0066] like Figure 1 As shown, the process of determining the user's expression of pressure and obtaining the change weight is as follows: Preset a base weight, a fixed value, and a change threshold. The base weight includes several sub-base weights corresponding to individual sub-modal feature information. Establish time nodes. Based on adjacent time nodes, monitor the changes in individual sub-modal feature information in real-time multimodal feature information to obtain sub-change values. Determine the relationship between all sub-change values ​​and the change threshold. Extract sub-change values ​​exceeding the change threshold to obtain the target change value. Add the fixed value to the base weight of the target change value to obtain the target weight. Extract sub-change values ​​that do not exceed the change threshold to obtain the fixed change value. Determine the number of fixed change values ​​to obtain the reduction quantity. Average the fixed value based on the reduction quantity to obtain the average value. Subtract the average value from the base weight of the fixed change value to obtain the fixed weight. Integrate all target weights and fixed weights to obtain the change weight.

[0067] It is important to note that the base weights, fixed values, and change thresholds are determined based on actual usage, as are the time points, which are multiple time points within the same time interval. The target weight is obtained by adding the fixed value to the sub-base weights of the target change value, and the fixed weight is obtained by subtracting the average value from the sub-base weights of the fixed change value. This ensures that the overall weights remain stable while adjusting the weights according to changes in the user's own characteristics. This helps determine the corresponding weights that conform to multimodal feature information, ensuring that the subsequent stress recognition results match the user's individual characteristics.

[0068] In practical application, when the multimodal feature information contains five features, the sub-basic weights can be 0.15, 0.2, 0.3, 0.2, and 0.15, corresponding to the five features: Feature 1, Feature 2, Feature 3, Feature 4, and Feature 5. When the sub-change value of Feature 4 exceeds the change threshold, a fixed value is added to the sub-basic weight of Feature 4, and this fixed value is divided into four equal parts to reduce the sub-basic weights of Features 1, Feature 2, Feature 3, and Feature 5. The increased and reduced sub-basic weights are then integrated to obtain the change weight. The specific determination of whether a sub-change value exceeds the change threshold can be as follows: Figure 3 As shown, by determining the slope of the change threshold, and then comparing the slope of the change trend of the segment feature values ​​at adjacent time nodes with the slope of the change threshold, it is possible to determine whether the sub-change value exceeds the change threshold.

[0069] like Figure 2 As shown, the information processing method includes: amplifying and reducing the target noise to obtain derived noise; establishing a noise database to store the target noise and derived noise; extracting speech information from the multimodal feature information to obtain the information to be processed; establishing a timeline; matching the information to be processed with the target noise and derived noise in the noise database based on the timeline to obtain a matching result; extracting the sound information corresponding to the target noise and derived noise in the information to be processed based on the matching result to obtain the processed information to be processed; and replacing the information to be processed in the multimodal feature information with the processed information to obtain the processed real-time multimodal information.

[0070] It is important to note that by amplifying and reducing the target noise to obtain derived noise, the number of samples without target noise can be increased, thus reducing the amount of data for subsequent target noise processing. The specific amplification and reduction factors can be determined based on the actual usage. The timeline represents the time process, i.e., the time process of the generation of multimodal feature information. Based on the matching results, the sound information corresponding to the target noise and derived noise in the information to be processed is extracted to obtain the processed information. In other words, the target noise and derived noise are similarly matched in the information to be processed, and the information corresponding to the target noise and derived noise is removed from the information to be processed to obtain the processed information.

[0071] like Figure 2 As shown, the selection method includes: traversing the directory table based on the changing weights to obtain the traversal results; when the traversal results indicate that there is a weight in the directory table corresponding to the changing weights, extracting the pressure recognition model corresponding to the changing weights from the directory table to obtain the target pressure recognition model; when the traversal results indicate that there is no weight in the directory table corresponding to the changing weights, extracting the pressure recognition model closest to the changing weights from the directory table to obtain the target pressure recognition model.

[0072] It should be noted that by setting the selection method, the target pressure identification model can be obtained by selecting a pressure identification model that corresponds to or is close to the change weight in the catalog table according to the change weight, thus achieving the function of screening pressure identification models that meet the change weight.

[0073] User information is obtained by integrating real-time multimodal feature information, change weights, target pressure levels, and target pressure recognition models. A repository is established to store user information, and recognition requirements are obtained. Based on the recognition requirements, the user information corresponding to the recognition requirements is searched in the repository to obtain the search results.

[0074] It is important to note that user information is obtained by recording and integrating real-time multimodal feature information, change weights, target stress levels, and target stress recognition models. This user information can be understood as a user profile. A repository is established to store user profiles. Users whose needs are to perform psychological stress recognition are identified. By matching the user's recognition needs with the user information in the repository, search results are obtained. Based on the search results, it can be determined whether the user has used this method for psychological stress recognition and the result of the user's last psychological stress level recognition can be obtained, so as to further judge the changes in the user's psychological state and improve the user experience.

[0075] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A psychological stress level recognition method based on a deep emotional neural network, comprising: information acquisition: acquiring past stress information and real-time multi-modal feature information of a user, the past stress information including past multi-modal feature information and stress information corresponding to the past multi-modal feature information; information processing: obtaining target noise by acquiring application scene noise information, and performing denoising processing on the real-time multi-modal feature information based on the target noise to obtain processed real-time multi-modal feature information; characterized in that it comprises: stress level division: dividing the stress level based on the stress information by a level division method to obtain a stress level, the stress level including a plurality of sub-stress levels; feature weight distribution: deriving the past stress information according to different feature weights based on the past stress information by a weight processing method to obtain derived stress information; model establishment: establishing a stress recognition model under different feature weight distributions based on the derived stress information and the past stress information to obtain a stress recognition model and a directory table; stress level recognition: determining the expression stress mode of the user to obtain a change weight, selecting a stress recognition model corresponding to the change weight based on the directory table by a selection method to obtain a target stress recognition model, inputting the processed real-time multi-modal feature information into the target stress recognition model to obtain a target stress, and obtaining a target stress level based on the target stress and the stress level; the weight processing method comprises: obtaining a basic weight distribution of the past multi-modal feature information in the past stress information to obtain a basic weight, the basic weight including a sub-weight corresponding to a single feature information in the past multi-modal feature information, a preset adjustment factor, increasing and reducing the sub-weights in the basic weight based on the adjustment factor to obtain a derived weight, obtaining a sub-derived feature information based on the derived weight and the past multi-modal feature information, obtaining a sub-derived stress by obtaining a stress of the sub-derived feature information, establishing a corresponding relationship between the sub-derived feature information and the sub-derived stress, and integrating the sub-derived feature information and the sub-derived stress to obtain the derived stress information; the information processing method comprises: obtaining derived noise by noise amplification and noise reduction based on the target noise, establishing a noise library for storing the target noise and the derived noise, extracting voice information in the multi-modal feature information to obtain to-be-processed information, establishing a timeline, matching the to-be-processed information with the target noise and the derived noise in the noise library based on the timeline to obtain a matching result, extracting sound information corresponding to the target noise and the derived noise in the to-be-processed information based on the matching result to obtain processed to-be-processed information, and replacing the to-be-processed information in the multi-modal feature information with the processed to-be-processed information to obtain processed real-time multi-modal information. 2.The method of claim 1, wherein the method further comprises: The grading method comprises: presetting a grading number, an optimal pressure value interval and a decrement value, the optimal pressure value interval being a first sub-pressure grade, the optimal pressure value interval comprising a first boundary value and a second boundary value, the first boundary value being greater than the second boundary value, a third boundary value being obtained based on the second boundary value and the decrement value, a sub-optimal pressure value interval being obtained by integrating the second boundary value and the third boundary value, the sub-optimal pressure value interval being a second sub-pressure grade, a third sub-pressure grade being obtained based on the third boundary value and the decrement value, and a pressure grade being obtained by integrating the first sub-pressure grade, the second sub-pressure grade and the third sub-pressure grade. 3.The method of claim 1, wherein the method further comprises: determining a stress level of the user based on the stress level of the user and the stress level of the user in the past. The process of establishing the pressure identification model comprises: Training data acquisition and processing: pre-processing derivative pressure information and past pressure information to obtain pre-processed derivative pressure information and pre-processed past pressure information, both of which are training data; Model selection and training: selecting a deep learning model as a model base, importing the training data into the model base for training to obtain an initial model; Model adjustment and output: presetting verification information, which comprises verification feature information and verification pressure, importing the verification feature information into the initial model to obtain a result pressure, judging the result pressure and the verification pressure to obtain a judgment result, optimizing and adjusting the initial model based on the judgment result to obtain a pressure identification model; Model labeling: obtaining labeling information based on the weight of the training data, establishing an association between the labeling information and the pressure identification model to obtain a target association, and establishing a directory table for storing the target association.

4. The method of claim 1, wherein the method is based on a deep affective neural network. The process of judging the expression pressure mode of the user to obtain the change weight comprises: presetting a basic weight, a fixed value and a change threshold, the basic weight comprising a plurality of sub-basic weights corresponding to a single sub-modal feature information, establishing a time node, obtaining a sub-change value based on the change of a single sub-modal feature information in real-time multi-modal feature information between adjacent time nodes, judging the relationship between all sub-change values and the change threshold, extracting a target change value from the sub-change value exceeding the change threshold, adding a fixed value to the sub-basic weight of the target change value to obtain a target weight, extracting a fixed change value from the sub-change value not exceeding the change threshold, judging the number of fixed change values to obtain a reduction number, obtaining an average value based on the average fixed value of the reduction number, and subtracting the average value from the sub-basic weight of the fixed change value to obtain a fixed weight, and integrating all target weights and fixed weights to obtain a change weight.

5. The method of claim 1, wherein the method is based on a deep affective neural network. The selection method comprises: traversing the directory table based on the change weight to obtain a traversal result, when the traversal result feedback is that there is a weight corresponding to the change weight in the directory table, extracting the pressure identification model corresponding to the weight from the directory table to obtain a target pressure identification model, and when the traversal result feedback is that there is no weight corresponding to the change weight in the directory table, extracting the pressure identification model corresponding to the weight closest to the change weight from the directory table to obtain a target pressure identification model.

6. The method of claim 1, wherein the method is based on a deep affective neural network. The user information is obtained by integrating real-time multi-modal feature information, change weight, target pressure level and target pressure recognition model, a storage library is established to store the user information, a recognition demand is acquired, and the storage library is searched based on the recognition demand to obtain a search result corresponding to the recognition demand.

Images