Psychological state recognition method based on dynamic optimization support vector data description

Abstract

This invention relates to the field of mental state recognition technology, specifically a mental state recognition method based on dynamically optimized support vector data description. The method includes: collecting physiological and behavioral signal sample data; constructing a dual-boundary SVDD model based on a relaxation threshold and training the model, where relaxation variables are introduced to construct a two-layer decision boundary to distinguish between normal and abnormal mental state samples; determining the hypercenter and radius during model training to form an accurate classification boundary; calculating the spatial and temporal weights of each sample to adjust the model's decision boundary; dynamically updating the model's decision boundary by adjusting sample weights based on newly collected sample data in real time; using the hypercenter trajectory offset index to monitor and predict the changing trend of mental states in real time; and predicting whether an individual's mental state will change significantly by analyzing the positional changes of the hypercenter.

Description

Technical Field

[0001] This invention relates to the field of mental state recognition technology, and more specifically, to a mental state recognition method based on dynamically optimized support vector data description. Background Technology

[0002] In today's rapidly changing society, mental state detection technology is becoming increasingly important, playing an indispensable role in multiple fields. These fields include, but are not limited to, criminal investigation, healthcare, and human-computer interaction, each of which has an urgent need for rapid and accurate identification of mental states.

[0003] In the field of criminal investigation: The application of psychological state detection technology is crucial. During interrogation, investigators need to accurately assess the suspect's psychological state in order to adopt effective interrogation strategies. Accurate assessment of psychological state helps investigators determine whether a suspect is withholding information, is under high stress or panic, and thus more effectively obtains key information. Furthermore, psychological state detection technology can also be used to assess the credibility of witnesses and plays a role in crime prevention and criminal psychology research.

[0004] In the medical field, timely identification of mental state is crucial for developing personalized treatment plans. Changes in mental state are often closely related to various physical illnesses, such as heart disease, hypertension, and sleep disorders. By promptly identifying a patient's mental state, doctors can better understand the patient's overall health condition, thereby developing more precise treatment plans. Furthermore, mental state monitoring technologies play an important role in preventing psychological crises, early intervention, and mental health management, helping to reduce the incidence of mental illness and improve treatment outcomes.

[0005] In the field of human-computer interaction (HCI), understanding users' emotional states is crucial for enhancing user experience. With technological advancements, human-machine interaction is becoming increasingly frequent, from smartphones and smart homes to virtual assistants and robots. Users expect technology to be more intelligent and human-like. Mental state detection technology can help machines better understand users' needs and emotions, thereby providing more personalized and emotional services. For example, intelligent educational software can adjust teaching strategies and improve learning efficiency by recognizing students' emotional states; intelligent customer service systems can provide more considerate services by identifying users' emotions.

[0006] Existing methods for identifying mental states primarily rely on machine learning algorithms, which typically require large amounts of labeled data for training. This data may originate from various physiological signals, such as electroencephalograms (EEG), heart rate (HR), and electrical skin activity (EDA), as well as behavioral signals, such as speech, facial expressions, and posture. By analyzing the characteristics of these signals, machine learning models can learn the mapping relationship between mental states and signal features, thereby enabling the identification of mental states.

[0007] Machine learning algorithms play a central role in the research and application of mental state recognition. These algorithms learn patterns and associations from massive amounts of data, enabling them to identify complex patterns of mental state changes. Physiological signals, as important biomarkers for mental state recognition, provide a wealth of physiological information. For example, electroencephalography (EEG) signals reflect the brain's electrical activity and are an important tool for studying cognitive and emotional states; heart rate (HR) and heart rate variability (HRV) reflect an individual's autonomic nervous system activity and are closely related to emotional states; and electrical skin activity (EDA) reflects an individual's sympathetic nervous system activity and is a sensitive indicator for detecting stress and emotional responses.

[0008] Behavioral signals, such as voice, facial expressions, and posture, are also important sources of psychological state recognition. Changes in tone, rhythm, and intensity in voice signals can reflect an individual's emotional state; facial expressions are a direct manifestation of emotional expression, and by analyzing the movement and changes in facial muscles, various basic emotions can be identified; posture and movements can provide nonverbal information about an individual, reflecting their psychological state and intentions.

[0009] To accurately identify psychological states, effective feature extraction and fusion of these multimodal signals are necessary. Feature extraction is fundamental to training machine learning algorithms, involving extracting key information from raw signals that represents signal characteristics and psychological states. Examples include extracting power spectral density (PSD) features from electroencephalogram (EEG) signals, Mel-frequency cepstral coefficients (MFCC) features from speech signals, and local binary pattern (LBP) features from facial expressions. These features not only capture the static characteristics of the signal but also reflect its dynamic changes.

[0010] Building upon feature extraction, feature fusion is a crucial step in integrating multimodal signals. By fusing features from different modalities, the accuracy and robustness of mental state recognition can be improved. Feature fusion can employ various methods, such as simple feature-level fusion, decision-level fusion, and more complex model-level fusion. Feature-level fusion typically involves concatenating or weighting feature vectors from different modalities; decision-level fusion involves integrating classification results from different modalities during the decision-making stage; and model-level fusion achieves automatic feature fusion and learning by constructing multimodal machine learning models, such as multimodal deep learning networks.

[0011] However, existing methods for identifying mental states face several major problems in practical applications:

[0012] 1. Difficulty in data acquisition: In many real-world scenarios, obtaining a large amount of high-quality training data is extremely difficult. For example, in criminal investigation or medical fields, it may be difficult to obtain a sufficient number of physiological signal data with different psychological states.

[0013] 2. Individual Differences: Under the same psychological state, different individuals may exhibit significant differences in the manifestation of their physiological and behavioral signals. This makes it difficult for models trained on general datasets to adapt to the identification of specific individual psychological states.

[0014] 3. Real-time requirements: In some application scenarios, such as crisis intervention or online education, the model needs to be able to respond to subtle changes in psychological state in real time, while traditional models trained on large amounts of data often cannot meet this requirement.

[0015] 4. Model generalization ability: Existing models often perform well on specific datasets, but their generalization ability is often insufficient when applied to different environments or different groups of people. Summary of the Invention

[0016] To address the aforementioned technical problems and overcome the limitations of existing mental state recognition technologies, this invention proposes an innovative method: a mental state recognition method based on dynamically optimized support vector data description (SVDD). This method is designed to better adapt to real-time, limited, and individually variably distributed sample data, thereby achieving more accurate mental state recognition. The core advantage of this invention lies in its ability to dynamically update the model's decision boundary. This allows the model to not only cope with the dynamic characteristics of data changes, such as data fluctuations caused by individual emotional fluctuations and changes in environmental stimuli at different time points, but also effectively handle the risk of underfitting due to scarce sample size. Specifically, it includes the following steps: constructing an accurate two-layer variable decision classification boundary, and dynamically adaptively adjusting this boundary according to the spatiotemporal characteristics of the samples during the continuous inclusion of new samples. Specifically, it includes...

[0017] Step 1: Collect raw physiological and behavioral signal sample data; the signal sample data includes physiological signals such as electroencephalogram (EEG), heart rate (HR), and electrical skin activity (EDA), as well as speech, facial expression, and posture behavioral signals;

[0018] Step 2: Construct a two-boundary SVDD model based on relaxation thresholds. In this model, relaxation variables are introduced to construct a two-layer decision boundary, which distinguishes signal data samples from samples of normal and abnormal psychological states. During the training process of the model, support vector machine (SVM) or similar algorithms are used to determine the hypersphere center and radius, thereby constructing an accurate two-layer variable decision classification boundary. In the process of continuously incorporating new samples, this boundary will dynamically and adaptively adjust according to the spatiotemporal characteristics of the samples.

[0019] Step 3: Calculate the spatial and temporal weights for each original sample; the spatial weights consider the distribution of the sample in the feature space, while the temporal weights consider the changing trend of the sample over time; these weights will be used to adjust the decision boundary of the model so that it can better adapt to changes in real-time data.

[0020] Step 4: Dynamically update the model's decision boundary based on the new psychological state sample data collected in real time. This step is achieved by adjusting the sample weights, where spatial and temporal weights are used to improve the model's adaptability to new data. The dynamic update mechanism enables the model to respond to changes in psychological states in real time, improving the accuracy and timeliness of recognition.

[0021] Step 5: Utilize the sphere center trajectory offset index to monitor and predict the changing trends of psychological state in real time; by analyzing the positional changes of the supercenter, the model can predict whether an individual's psychological state will undergo significant changes.

[0022] Furthermore, in step 2, a dual-boundary SVDD model based on a relaxation threshold was constructed; wherein the determination of the dual-layer decision boundary is achieved through a spherical boundary offset factor. To make a preliminary division, Defined as:

[0023] ;

[0024] in, This represents the distance from the test sample to the center of the sphere, while setting two threshold values. and ,in , Thus, based on the original SVDD decision boundary, two new layers of variable relaxed decision boundaries are formed, dividing the hypersphere into three regions, each satisfying... , , ;

[0025] For relaxation threshold and The optimal threshold is determined using Otsu's rule. Otsu's method solves for the optimal target threshold. and The calculation formula is as follows:

[0026] ;

[0027] In the formula, This is the average distance level of all samples in the sample set; These represent the probabilities of occurrence for samples of classes A, B, and C, respectively. These are the average distance levels for samples in categories A, B, and C, respectively.

[0028] Furthermore, for step 3, in the field of mental state recognition, spatial weights and temporal weights are introduced to enable the model to more accurately reflect the complexity and dynamic changes of an individual's mental state.

[0029] Spatial weights, by considering the differences in the distribution of samples in the feature space, enhance the model's focus on information-rich regions. Spatial weights are defined as follows:

[0030] ;

[0031] in Representing the Spatial weights of individual data points It is a weighting factor that controls the membership degree and the distance between the local and global centers; It represents The spatial weight is calculated by considering the distance between the local and global center of a data point. A data point with a higher local-to-global center distance will receive a higher weight, meaning it will be assigned a larger spatial weight. Therefore, the introduction of spatial weights will make the identification of support vector samples more accurate, effectively improving the model's accuracy.

[0032] Time weights, by reflecting the changing trend of samples over time, enhance the model's sensitivity to real-time data.

[0033] Sample data of time series Each data point All correspond to timestamps Then the time weight can be defined as:

[0034] ;

[0035] in, It is the latest timestamp; It is the attenuation parameter, which controls the effect of time distance; for each data point Considering the time window arrive Data within the range, This represents the radius of the time window; the significance of doing so is that, in addition to considering the current point... Furthermore, it also considers points that are close in time before and after it. It is a coefficient that adjusts the influence of data points within a time window.

[0036] It is a difference function, used to reflect the differences between data points. With Other points within the range The similarity is specifically represented as:

[0037] ;

[0038] in, Used to control the degree of influence of the difference between two points. When two points are very similar (i.e., their differences are small), Approximately 1; when the difference between the two points is significant, The time weight membership degree not only considers the distance between the data point and the current time point, but also integrates the interaction between data points within the time window, providing a more dynamic and comprehensive time weight allocation method.

[0039] Furthermore, regarding step 4, in reality, the target's psychological state changes due to factors such as environment and time. There is a psychological limit to how much this change in psychological state can be tolerated, and this limit varies from person to person.

[0040] To comprehensively consider the combined impact of time and spatial weights on the model, this invention redefines the joint optimization problem and designs a hierarchical weight adjustment mechanism. Under this mechanism, time weights are used to first adjust the importance of each sample point, adjusting its influence coefficient on the model based on the data point's timestamp, reflecting the timeliness of the data. Subsequently, spatial weights are further adjusted to reflect their spatial distribution characteristics, particularly their sensitivity to boundaries. Data points located on the hypersphere boundary may have a significant impact on the model's decision boundary and updates.

[0041] By introducing new data, the model can efficiently select key support vectors during dynamic updates. As new samples are collected in real time and added to the model optimization, the spatial weight of each sample will be calculated to increase or decrease its impact on the model weights.

[0042] Furthermore, regarding step 5, as the model is dynamically updated, the center of the classification hypersphere exhibits a dynamic displacement over time. The offset distance of the sphere's center after each update represents the degree of abnormal psychological deviation. This deviation may exceed the range of normal psychological state. A deviation index is defined. :

[0043] ;

[0044] in and represent and The center of the ball at a given moment. The acceleration of the ball's center offset before and after the current update can be calculated:

[0045]

[0046] Since psychological states have a tolerance limit before becoming abnormal, the shift acceleration in the early stages of evolution should continuously approach this limit, i.e., continuously decrease. This indicates that the individual's psychological state is rapidly deteriorating, and when it exceeds the tolerance limit and reaches an abnormal state, the shift acceleration should instantaneously deteriorate beyond the relaxation threshold of that instance. It is satisfied at this moment:

[0047] ;

[0048] When the acceleration due to the ball's center of gravity shifts exceeds the relaxation threshold, it indicates that the sample has exceeded its tolerance limit. The psychological state changes from normal to abnormal. This allows us to identify the target's abnormal psychological state.

[0049] The effects of the key technologies employed in this invention are as follows:

[0050] 1) Relaxed Threshold Two-Boundary SVDD Model: In traditional SVDD models, data points are mapped to a high-dimensional space, and a hypersphere is used to distinguish between normal and abnormal samples. However, this method may not be flexible enough when dealing with boundary samples. To improve the robustness of the model, this invention introduces relaxation variables to construct a two-layer decision boundary. This two-layer boundary consists of an inner layer and an outer layer. The inner layer represents core normal psychological state samples, while the outer layer contains more boundary samples. In this way, the model can more effectively distinguish between normal and abnormal psychological state samples, while reducing the risk of misclassification and missed classification. The introduction of relaxation variables makes the model more flexible and accurate when dealing with samples that are difficult to classify.

[0051] 2) Dynamic Update Mechanism for Spatiotemporal Weights: Psychological state data typically exhibits both time-series and spatial distribution characteristics, meaning the importance of data points may change over time and space. To fully utilize these characteristics, this invention proposes a dynamic update mechanism combining spatial and temporal weights. Spatial weights are calculated based on the distribution of data points in the feature space, considering the proximity between data points and their distance from the decision boundary. Temporal weights are calculated based on the timestamps of the data points, considering their trends over time and their impact on the current psychological state. Through this dynamic update mechanism, the model can adjust its decision boundary in real time to adapt to changes in new data, thereby improving the model's adaptability and classification accuracy.

[0052] 3) Center-of-Sphere Trajectory Shift Index: The hypercenter of the SVDD model represents the central position of the data, and its changes can reflect the overall trend of the data distribution. This invention proposes a center-of-sphere trajectory shift index for real-time monitoring and prediction of changes in psychological states. This index is defined by calculating the positional changes of the hypercenter at consecutive time points. It can sensitively capture subtle changes in data distribution, thus providing an effective tool for real-time monitoring and prediction of psychological states. By analyzing the trend and speed of center-of-sphere trajectory shift, the model can predict whether an individual's psychological state will undergo significant changes, thus providing the possibility for timely intervention. Attached Figure Description

[0053] Figure 1 A flowchart for the SVDD abnormal psychological state detection system;

[0054] Figure 2 A schematic diagram of a hypersphere for classifying double-boundary SVDDs;

[0055] Figure 3 The local-global center distance of the model samples;

[0056] Figure 4 SVDD model prediction: Schematic diagram of SVDD sphere center offset. Detailed Implementation

[0057] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0058] refer to Figure 1 This invention provides a method for identifying mental states in SVDD anomaly detection, the implementation steps of which are as follows:

[0059] Step 1: In the first step of this method, physiological and behavioral signal data are collected from the target individual. This data includes physiological signals such as electroencephalogram (EEG), heart rate (HR), and electrical skin activity (EDA), as well as behavioral signals such as speech, facial expressions, and posture.

[0060] Step 2: Based on feature extraction, a double-boundary SVDD model based on a relaxed threshold was constructed.

[0061] This method constructs a two-layer decision boundary by introducing slack variables, which effectively distinguishes samples of normal and abnormal psychological states. During the training process of the model, support vector machine (SVM) or similar algorithms will be used to determine the hypersphere center and radius, thereby forming an accurate classification boundary.

[0062] like Figure 2 As shown. Define the sphere boundary offset factor. , ,in This represents the distance from the test sample to the center of the sphere, while setting two threshold values. and ,in , This results in two new, variable, relaxed decision boundaries based on the original SVDD decision boundaries, dividing the hypersphere into three regions, A, B, and C, each satisfying... , , threshold and This corresponds to the two-layer relaxed boundary effect in the SVDD model. The relaxed boundary reflects the model's tolerance for anomalous mental state points, weakening the influence of normal samples (far from the center) and anomalous samples (closer to the center) that are far from the boundary. After training, the threshold... and The SVDD boundary is adjusted so that most samples that are difficult to identify directly and accurately enter region B, in order to balance the misclassification ratio of the SVDD boundary.

[0063] For relaxation threshold and The optimal value is determined as follows: Otsu's rule is used to determine the optimal value. Thresholding. Otsu's method, also known as the maximum inter-class variance method, has high segmentation accuracy and adaptability. The core idea of ​​this method is to select the optimal threshold that maximizes the separation performance between different classes, and to use the Otsu method to select the optimal double boundary. and The training samples and newly added samples are combined to form a sample set to determine the optimal threshold. The distances from the center R of the hypersphere obtained after processing the combined sample set using the SVDD algorithm to each sample in the high-dimensional feature space are divided into levels from 0 to m, where level 0 corresponds to the center R, and level m corresponds to the sample with the largest sample distance. Assume that distance levels exist. , Divide all samples in the sample set into three categories: , , Otsu then solves for the optimal target threshold. and The formula for calculation is:

[0064] ;

[0065] In the formula, This is the average distance level of all samples in the sample set; These represent the probabilities of occurrence for samples of classes A, B, and C, respectively. These represent the average distance levels for samples in classes A, B, and C, respectively. Based on this, the optimal threshold conforming to the model is determined, yielding the optimal double-boundary SVDD, achieving a more reasonable division of boundary samples and improving the model's detection accuracy.

[0066] Step 3: Spatiotemporal weight calculation. In the field of mental state recognition, due to the randomness and nonlinearity of the target's mental state fluctuations, the contributions of collected samples to the classification hypersphere vary. Therefore, it is necessary to assign different spatial weights to samples at different distribution locations.

[0067] When considering relevant methods, the following problems arise from focusing solely on the density of data points.

[0068] (1) When the support vector set has a low density, using the fuzzy SVDD-based method will reduce the contribution of the support vector set;

[0069] (2) In practical applications, if the initial sample data is small or the number of samples acquired in real time is insufficient, the above method cannot obtain accurate density. These two problems are frequently encountered in mental state recognition, where the target sample data is usually insufficient. This will lead to a decrease in the performance of the decision boundary.

[0070] Based on the above problems, this method proposes a local-global center distance:

[0071] ;

[0072] ;

[0073] ;

[0074] The local-global center distance in the model is shown in the appendix. Figure 3 As shown.

[0075] in, represent The The nearest neighbor, It can be regarded as the center of the training data points, that is, the global center. express and its own first Because it is a neighbor, it can be called a local center. The value is chosen empirically as the square root of the number of training data samples. It represents The distance between the local center and the global center. For outliers, the distance between the sample and the global center should not be used as the sole basis for fuzzy construction; otherwise, outliers will receive excessive weight. They should be treated as outliers to reduce their impact on the decision boundary. To relatively weaken outliers Weight values. Outliers. local center With global center distance Less than Therefore, the weight is relatively low for points near the support vector set. Local center near Therefore, sample points The weights do not change significantly and remain high. Therefore, using the local-global center distance instead of the density distribution to consider global information yields better results, focusing more attention on the support vector set. Based on this, the spatial weights of the samples are defined as:

[0076] ;

[0077] in, Representing the Spatial weights of individual data points It is a weighting factor that controls the membership degree and the distance between the local and global centers. If a data point has a high local-to-global center distance, it will also receive a high weight, that is, be assigned a larger spatial weight.

[0078] Psychological state data exhibits significant time-series characteristics, which change continuously with different scenarios and mindsets, significantly impacting the detection of anomalies. Therefore, it is crucial to fully consider these time-series characteristics to avoid accuracy degradation caused by static data distribution. This method proposes time weights to reflect the dynamic timeliness of the data. Based on the time-series characteristics, each sample is assigned a unique time weight to adjust its impact on the model, eliminating samples that are not significant for the decision boundary.

[0079] Sample data of time series Each data point All correspond to timestamps Then the time weight can be defined as:

[0080] ;

[0081] in, It is the latest timestamp; It is the attenuation parameter, which controls the effect of time distance; for each data point Considering the time window arrive Data within the range, This represents the radius of the time window; the significance of doing so is that, in addition to considering the current point... Furthermore, it also considers points that are close in time before and after it. It is a coefficient that adjusts the influence of data points within a time window. It is a difference function, used to reflect the differences between data points. With Other points within the range The similarity is specifically represented as:

[0082] ;

[0083] in, Used to control the degree of influence of the difference between two points. When two points are very similar (i.e., their differences are small), Approximately 1; when the difference between the two points is significant, The time weight membership degree not only considers the distance between the data point and the current time point, but also integrates the interaction between data points within the time window, providing a more dynamic and comprehensive time weight allocation method.

[0084] Step 4: Dynamic Model Update. Based on newly acquired sample data in real time, this method dynamically updates the model's decision boundary. This is achieved by adjusting sample weights, where spatial and temporal weights are used to improve the model's adaptability to new data.

[0085] To comprehensively consider the combined effects of time and space weights on the model, the joint optimization problem is redefined.

[0086] Traditional SVDD aims to improve classification performance by finding the SV of the target data point to provide the best data description. Assume we have a dataset... ,in The target number of data points. The problem is described as follows:

[0087] ;

[0088] In this function, Represents the radius of the sphere. It is a punishment factor. This represents a slack variable. This represents a nonlinear mapping that maps the original input data to a new high-dimensional feature space.

[0089] This method designs a hierarchical weight adjustment mechanism. Time weight. This can be used to adjust the importance of each sample point first, and then adjust its coefficient of influence on the model based on the timestamp of the data point, reflecting the timeliness of the data. Spatial weights can then be applied. Further adjustments are needed to reflect their spatial distribution characteristics, particularly their sensitivity to boundaries. Data points located on the boundaries of the hypersphere may have a significant impact on the model's decision boundaries and updates.

[0090] Using local-global center distance To relatively weaken outliers Membership degree. For example... Figure 3 As shown, outliers local center With global center distance Less than This means that the weight is relatively low. For points near SV... Local center near The membership degree of all samples will not change significantly and will remain large.

[0091] Therefore, the optimization objective of the above joint problem can be improved based on spatial weights:

[0092] ;

[0093] ;

[0094] By using spatial weights Directly applied to constraints and slack variables, in the objective function The influence of samples with high membership degrees has been increased. (This is related to) the constraints. This ensures that samples with high contribution are more closely packed around the center of the sphere, while samples with low contribution are allowed to deviate more.

[0095] By introducing new data, the model can efficiently select key support vectors during dynamic updates. As new samples are collected in real time and added to the model optimization, the spatial weight of each sample will be calculated to increase or decrease its impact on the model weights.

[0096] Step 5, Abnormal State Detection: The center of the classified hypersphere exhibits a dynamic displacement that changes over time. The offset distance of the center after each update represents the degree of abnormal psychological deviation. This deviation may exceed the normal range of psychological state. A deviation index is defined. : ;

[0097] in and represent and The center of the ball at a given moment. The acceleration of the ball's center offset before and after the current update can be calculated:

[0098] ;

[0099] Since psychological states have a tolerance limit before becoming abnormal, the shift acceleration in the early stages of evolution should continuously approach this limit, i.e., continuously decrease. This indicates that the individual's psychological state is rapidly deteriorating, and when it exceeds the tolerance limit and reaches an abnormal state, the shift acceleration should instantaneously deteriorate beyond the relaxation threshold of that instance. It is satisfied at this moment:

[0100] ;

[0101] When the acceleration due to the ball's center of gravity shifts exceeds the relaxation threshold, it indicates that the sample has exceeded its tolerance limit. The psychological state changes from normal to abnormal. This allows us to identify the target's abnormal psychological state.

[0102] This method employs data validation in its experiments. Over time, the continuously collected EEG data exhibits a state shift trend. Without any specific abnormalities, the test subject's mental state will eventually manifest as an abnormal state. Accurate prediction of the test subject's mental state not only facilitates further mental state detection but also helps determine the model's update boundaries, reducing time and computational complexity and improving model performance. To clearly demonstrate the shift process of the SVDD sphere center, the prediction results for 1 minute of EEG data are shown below. Figure 4 As shown.

[0103] During the online real-time updates, the center of the SVDD hypersphere shifted, and this shift precisely reflected the trend of psychological state changes caused by EEG data. When the SVDD hypersphere boundary was initially established, the test subject's psychological state was at its most stable. As the model dynamically updates, the center of the hypersphere continuously shifts from the boundary. When the model's center shifts back to the initial outer boundary of the hypersphere, the model reaches its termination point.

[0104] The above are all the steps of one embodiment of the present invention, including data acquisition, model building, calculation of spatiotemporal weights, dynamic model updating, and abnormal psychological state detection. Finally, a fast and accurate method for detecting abnormal psychological states is realized, which can be widely applied to the field of intelligent recognition of psychological states.

[0105] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying mental states based on dynamically optimized support vector data description, characterized in that, Includes the following steps, Step 1: Collect raw physiological and behavioral signal sample data; Step 2: Construct a two-boundary SVDD model based on relaxation threshold and train the model. The model introduces relaxation variables to construct a two-layer decision boundary to distinguish between normal and abnormal psychological state samples of signal data samples. During the training process of the model, the two-layer variable decision classification boundary is constructed by determining the hypersphere center and radius. Step 3: Calculate the spatial and temporal weights for each original sample to adjust the decision boundary of the model; the spatial weights consider the distribution of the samples in the feature space, and the temporal weights consider the changing trend of the samples over time. Step 4: Based on the new sample data of real-time collected psychological state data, dynamically update the decision boundary of the model by adjusting the sample weights; Step 5: Use the ball center trajectory offset index to monitor and predict the changing trend of psychological state in real time; by analyzing the positional changes of the super-center, the model can predict whether the individual's psychological state will change significantly.

2. The mental state recognition method based on dynamically optimized support vector data description as described in claim 1, characterized in that, In step 1, the physiological signal sample data includes electroencephalography (EEG), heart rate (HR), and electrical skin activity (EDA); the behavioral signals include speech, facial expressions, and posture.

3. The mental state recognition method based on dynamically optimized support vector data description as described in claim 1, characterized in that, In step 2, the determination of the two-layer decision boundary is achieved through the sphere boundary offset factor. To make a preliminary division, Defined as: ； in, This represents the distance from the test sample to the center of the sphere, while setting two threshold values. and ,in , Thus, based on the original SVDD decision boundary, two new layers of variable relaxed decision boundaries are formed, dividing the hypersphere into three regions, each satisfying... , , ; For relaxation threshold and The optimal threshold is determined using Otsu's rule. Otsu's method solves for the optimal target threshold. and The calculation formula is as follows: ； In the formula, This is the average distance level of all samples in the sample set; These represent the probabilities of occurrence for samples of classes A, B, and C, respectively. These are the average distance levels for samples in categories A, B, and C, respectively.

4. The mental state recognition method based on dynamically optimized support vector data description as described in claim 1, characterized in that, In step 3, the spatial weights are defined by considering the differences in the distribution of samples in the feature space: ; in Representing the Spatial weights of individual data points It is a weighting factor that controls the membership degree and the distance between the local and global centers; It represents The distance between the local center and the global center; Sample data of time series Each data point All correspond to timestamps Then the time weight can be defined as: ； in, It is the latest timestamp; It is the attenuation parameter, which controls the effect of time distance; for each data point Considering the time window arrive Data within the range, Indicates the radius of the time window; It is a coefficient that adjusts the influence of data points within a time window; It is a difference function, used to reflect the differences between data points. With Other points within the range The similarity is specifically represented as: ； in, Used to control the degree of influence of differences between two points; when two points are very similar. Approximately 1; when the difference between the two points is significant, It approaches 0.

5. The mental state recognition method based on dynamically optimized support vector data description as described in claim 1, characterized in that, For step 4, considering the combined influence of time and spatial weights on the model, a joint optimization problem is defined, and a hierarchical weight adjustment mechanism is designed. Under this weight adjustment mechanism, the time weight is used to adjust the importance of each sample point first, and the coefficient of its influence on the model is adjusted according to the timestamp of the data point. Then, the spatial weights are further adjusted to reflect their spatial distribution characteristics, especially their sensitivity to the boundary. By introducing new data, it is ensured that the model can select key support vectors during the dynamic update process. As new samples are collected in real time and added to the model optimization, the spatial weight of the sample will be calculated to increase or decrease its influence on the model weights.

6. The mental state recognition method based on dynamically optimized support vector data description as described in claim 1, characterized in that, For step 5, as the model is dynamically updated, the center of the classification hypersphere exhibits a dynamic displacement that changes over time. The offset distance of the sphere's center after each update represents the degree of abnormal psychological deviation, and a deviation index is defined. : ; in, and represent and The center of the ball at time t; calculate the acceleration of the ball's center offset before and after the current update: ； In the early stages of evolution, the offset acceleration continuously approaches its limit, i.e., it continuously decreases; this indicates that the individual's psychological state is rapidly deteriorating. When the tolerance limit is exceeded and an abnormal state is reached, the offset acceleration should instantly deteriorate beyond the relaxation threshold of that instance. At this moment, the following conditions are met: ； When the acceleration of the ball's center of gravity deviates from the relaxation threshold, it means that the sample has exceeded the tolerance limit; the psychological state changes from normal to abnormal; thus, the abnormal psychological state of the target can be identified.

Images