Anomaly detection method and device for digital psychological sandbox

Abstract

The invention discloses an anomaly detection method and device for a digital psychological sand table, and relates to the technical field of anomaly detection. The invention first extracts process features of each user's sand table game from multiple dimensions according to sand table game records, and extracts image features of each user's sand table game from multiple dimensions according to sand table image data, obtains user data points corresponding to each user, and then cross-combines different evaluation categories under each dimension, thereby obtaining multiple combination layers corresponding to multiple dimensions, and then anomaly detection can be performed based on distance measurement based on each combination layer, reducing the amount of calculation of the distance between user data points, and when anomaly detection is performed based on the distance based on the user data points corresponding to each combination layer, the selection quality of the initial cluster center and the efficiency of sampling classification are significantly improved by combining layered extraction, random extraction and cluster extraction, taking into account the accuracy of anomaly detection of the digital psychological sand table while improving efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of anomaly detection, and in particular to an anomaly detection method and device for a digital psychological sandbox. Background Art

[0002] At present, traditional mental health screening tools include the Symptom Checklist-90 (SCL-90), the Depression Scale (SDS), the Anxiety Scale (SAS), etc. These questionnaires help individuals identify potential psychological problems through standardized self-assessment. This traditional questionnaire method relies on the self-report of the investigator, which may lead to inaccurate results due to the subjective emotions or cognitive bias of the subjects. In addition, some individuals may be unwilling to fill out the questionnaire for various reasons, resulting in distorted screening results. Therefore, the traditional method has the disadvantages of strong subjectivity and easy disguise.

[0003] Psychological sand tray is a common tool for identifying and treating psychological problems. Participants feel less pressure in the game and are more likely to let down their psychological defenses. Psychological counselors can obtain more diagnostic evidence by observing and analyzing behaviors and choices during the process, making hidden psychological problems more visible, which is crucial for mental health education. The emergence of digital psychological sand tray games provides better convenience and accessibility.

[0004] Each sand tool in the sand table has rich symbolic meanings. Although it helps psychologists discover potential mental health problems, it requires a lot of manpower and material resources. Using anomaly detection algorithms in data mining can identify abnormal data patterns that are significantly different from normal behavior from complex data sets. Using anomaly detection algorithms to detect abnormal digital psychological sand tables can greatly reduce labor costs and time.

[0005] Outlier detection is an important task in the field of data mining and machine learning, which aims to identify outliers that are significantly different from other data points in the data set. These outliers may represent abnormal behavior, noise, errors, or important rare events in the data, and have important application value in fraud detection, health monitoring and other fields.

[0006] Existing outlier detection technologies mainly include statistical model-based methods, distance-based methods, clustering-based methods, density-based methods, and deep learning-based methods. However, in the digital psychological sandbox, for the outlier detection method based on deep learning, since the outliers of the digital psychological sandbox are often rare and difficult to label, if it is based on supervised learning, it is usually difficult to train and apply on a large scale; if it is based on unsupervised learning, how to effectively define outliers is a difficult problem to solve.

[0007] Therefore, deep learning methods are rarely used in anomaly detection of digital psychological sand tables, but methods other than deep learning still face some challenges and limitations in practical applications. Specifically, as the digital psychological sand table introduces more psychological indicators or sand table features, the dimension of the user's sand table data increases. When faced with a large number of users' psychological sand table data, the calculation of the distance between the user's sand table data points becomes more complicated and the amount of calculation increases sharply. The existing anomaly detection methods for digital psychological sand tables are less efficient. Summary of the invention

[0008] Based on this, it is necessary to provide an anomaly detection method and device for a digital psychological sandbox to address the above technical problems.

[0009] The present invention adopts the following technical solutions:

[0010] The present invention provides a method for detecting anomalies in a digital psychological sandbox, comprising:

[0011] Encode the sand tools and sand table operations in the sand table respectively, and obtain the sand table game records based on the encoding and the sand table image data after the sand table game of multiple users;

[0012] For each user, process features are extracted from multiple sand tool dimensions according to the sand table game record; and image features are constructed according to multiple basic visual feature dimensions of the sand table image data; the multiple sand tool dimensions include the aggregation degree of the sand tools, the positivity symbolized by the sand tools, the symmetry of the sand tools, and the diversity of the sand tools;

[0013] Cross-combining different evaluation categories of each sand tool dimension in the process feature with different evaluation categories of each basic visual feature dimension in the image feature to obtain multiple combined layers; determining the user data points formed by the process feature and the image feature of the user corresponding to each combined layer;

[0014] Initial candidate points are obtained by random sampling from the user data points corresponding to each combination layer, and the initial candidate points are clustered by cluster sampling to obtain multiple cluster centers as fusion sampling candidate points; all user data points are assigned to the clusters to which the fusion sampling candidate points with the closest distance belong;

[0015] In the cluster to which each fusion sampling candidate point belongs, isolated points in each user data point are determined as outliers according to the distance between each user data point and all other user data points.

[0016] The present invention provides an abnormality detection device for a digital psychological sandbox, comprising:

[0017] An acquisition module is used to encode the sand tools and sand table operations in the sand table, respectively, and obtain sand table game records based on the encoding and sand table image data after the sand table game of multiple users;

[0018] An extraction module is used to extract process features from multiple sand tool dimensions according to the sand table game record for each user; and to construct image features according to multiple basic visual feature dimensions of the sand table image data; the multiple sand tool dimensions include the aggregation degree of the sand tools, the positivity symbolized by the sand tools, the symmetry of the sand tools, and the diversity of the sand tools;

[0019] A hierarchical combination module is used to cross-combine different evaluation categories of each sand tool dimension in the process feature with different evaluation categories of each basic visual feature dimension in the image feature to obtain multiple combined hierarchies; determine the user data points formed by the process feature and image feature of the user corresponding to each combined hierarchy;

[0020] The sampling and classification module is used to randomly sample the user data points corresponding to each combination layer to obtain initial candidate points, cluster the initial candidate points through cluster sampling to obtain multiple cluster centers as fusion sampling candidate points; and assign all user data points to the clusters to which the nearest fusion sampling candidate points belong;

[0021] The anomaly determination module is used to determine, in the cluster to which each fusion sampling candidate point belongs, isolated points in each user data point as an anomaly point according to the distance between each user data point and all other user data points.

[0022] The present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned digital psychological sandbox anomaly detection method is implemented.

[0023] The present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned anomaly detection method of a digital psychological sandbox when executing the program.

[0024] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0025] The present invention first extracts the process features of each user's sandbox game from multiple dimensions based on the sandbox game records, and extracts the image features of each user's sandbox game from multiple dimensions based on the sandbox image data, to obtain the user data points corresponding to each user, and then cross-combines the different evaluation categories under each dimension, thereby obtaining multiple combination layers corresponding to multiple dimensions, the number of dimensions of each combination layer is the same, and the evaluation category for each dimension is different, so as to determine the user data points corresponding to each combination layer, and then anomaly detection can be performed based on each combination layer based on distance measurement. For the user data points corresponding to each combination layer, the cluster center of each combination layer is first selected by combining layered extraction, random extraction and cluster extraction, and then all user data points corresponding to the combination layer are clustered, and anomaly detection is performed based on the distance between user data points in each cluster obtained by clustering.

[0026] Compared with anomaly detection based on distance metrics according to all user data points, it is often necessary to perform a large amount of calculations to analyze the degree of distance deviation between each user data point and all other user data points in order to determine its abnormal situation; the present application first divides the user data points into multiple combined layers according to different evaluation categories under each dimension, and then selects high-quality cluster centers by combining layered extraction, random extraction and cluster extraction to perform clustering and anomaly detection, which significantly improves the efficiency of sampling classification and the quality of cluster center selection, avoids ignoring boundary points, and only needs to analyze the abnormal points in each cluster in each combined layer during anomaly detection, reducing the complexity of distance metrics, thereby improving the efficiency while taking into account the accuracy of anomaly detection of the digital psychological sandbox. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0028] Figure 1 A schematic flow chart of an abnormality detection method for a digital psychological sandbox provided by the present invention;

[0029] Figure 2 A schematic diagram of an abnormality detection model of a digital psychological sandbox provided by the present invention;

[0030] Figure 3 A schematic diagram of the construction of a sandbox data preprocessing and feature engineering provided by the present invention;

[0031] Figure 4 A schematic diagram of a data fusion sampling process provided by the present invention;

[0032] Figure 5A detailed schematic diagram of distance and outlier detection based on incremental search provided by the present invention;

[0033] Figure 6 A schematic diagram of an anomaly detection device for a digital psychological sandbox provided by the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of 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.

[0035] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0036] Figure 1 The present invention is a schematic flow chart of an abnormality detection method of a digital psychological sandbox, which specifically includes the following steps:

[0037] S101: Encode the sand tools and sandbox operations in the sandbox respectively, and obtain sandbox game records of multiple users based on the encoding and sandbox image data after the sandbox game.

[0038] S102: for each user, extracting process features from multiple sand tool dimensions according to the sand table game record; and constructing image features according to multiple basic visual feature dimensions of the sand table image data; the multiple sand tool dimensions include the aggregation degree of the sand tools, the positivity symbolized by the sand tools, the symmetry of the sand tools, and the diversity of the sand tools.

[0039] S103: Cross-combining different evaluation categories of each sand tool dimension in the process feature with different evaluation categories of each basic visual feature dimension in the image feature to obtain multiple combined layers; determining user data points formed by the process feature and image feature of the user corresponding to each combined layer.

[0040] S104: Randomly sample the user data points corresponding to each combined layer to obtain initial candidate points, cluster the initial candidate points through cluster sampling to obtain multiple cluster centers as fusion sampling candidate points; and assign all user data points to the clusters to which the fusion sampling candidate points closest to each other belong.

[0041] S105: In the cluster to which each fused sampling candidate point belongs, according to the distance between each user data point and all other user data points, determine the isolated points in each user data point as abnormal points.

[0042] For the convenience of explanation, the following description is only based on the server as the execution subject. The server mentioned in the present invention can be a server set up on the business platform, or a device such as a desktop computer, a notebook computer, etc. that can execute the solution of the present invention.

[0043] Outlier detection is an important task in the field of data mining and machine learning, which aims to identify points that are significantly different from other data points in the data set. These outliers may represent abnormal behavior, noise, errors or important rare events in the data, and have important application value in fraud detection, health monitoring and other fields. Existing outlier detection technologies mainly include statistical model-based methods, distance-based methods, clustering-based methods, density-based methods and deep learning-based methods. Although outlier detection has made significant progress, it still faces some challenges and limitations in practical applications. With the increase of data dimensions, the "dimensionality disaster" phenomenon makes outlier detection more difficult. In addition, since outliers are often rare and difficult to label, it is difficult to apply anomaly detection methods based on supervised learning on a large scale. In unsupervised learning, how to effectively define anomalies is also a major challenge. For some scenarios that require real-time detection, the model is required to be able to quickly process large amounts of data. Many existing complex algorithms still need to be improved in processing efficiency.

[0044] In order to solve the above problems, the present invention proposes an innovative multidimensional data feature extraction and clustering-based outlier detection technology based on the study of outlier detection in digital psychological sandbox. The present invention proposes an outlier detection method for digital psychological sandbox, and proposes an outlier detection method based on multidimensional data feature extraction, stratified fusion sampling, incremental distance and clustering. Figure 2 As shown, Figure 2 It is a schematic diagram of an anomaly detection model of a digital psychological sandbox in the present invention.

[0045] First, in the feature engineering stage, the multimodal feature space of user operation behavior, psychological state and image vision was constructed by combining the sand table game records of the digital psychological sand table. The user behavior was quantitatively analyzed from psychological dimensions such as aggregation, positivity, symmetry and diversity, and the low-level visual features of the sand table image data were extracted to achieve multi-level representation of the data, improving the data's expressiveness and effectiveness of expression.

[0046] When using clustering methods for outlier detection, in order to solve the limitations of traditional clustering methods in dealing with complex high-dimensional data and data with uneven density distribution, the present invention proposes a strategy of stratified fusion sampling. According to the global distribution characteristics of the data, the present invention improves the problem that traditional random sampling easily ignores boundary points and outliers, and the high computational complexity of K-means++ sampling by combining stratified extraction, random extraction and K-means++ extraction, significantly improves the selection quality of the initial cluster center and improves the sampling efficiency. The candidate points obtained by fusion sampling are processed using the K-means method, and cluster allocation of the entire data set is performed on this basis, which improves the clustering effect on unevenly distributed data. Distance and metric methods are introduced to better characterize the relationship between sample points and cluster centers, which can accurately identify outliers and combine with incremental search strategies to effectively improve the efficiency of processing large-scale data.

[0047] In order to achieve the above object, the present invention adopts the following technical solution:

[0048] The present invention proposes an abnormal digital psychological sandbox detection method based on multidimensional feature extraction and hierarchical distance and clustering, and the method mainly comprises the following steps:

[0049] 1. Multi-dimensional feature extraction: This invention innovates the extraction of data from the digital psychological sandbox. During data preparation and preprocessing, it not only integrates process data and image data, but also designs noise removal rules to ensure the validity and consistency of the data. This fusion of multi-dimensional data provides a more comprehensive framework for evaluating the user's psychological state by associating operational behavior with the visual information of the image. In addition, multi-scale edge detection is used to separate the background and foreground of the image, laying the foundation for the accuracy of feature extraction. When constructing feature engineering, a series of feature extraction methods for user operational behaviors are proposed from the four psychological dimensions of aggregation, positivity, symmetry, and diversity. Combined with psychological theories, complex behavioral patterns are converted into easy-to-analyze features. At the same time, by extracting features from multiple basic visual feature dimensions of sandbox image data, a multi-level image description method is formed, which improves the accuracy and interpretability of psychological sandbox data analysis. Figure 3 As shown, Figure 3 The figure is a schematic diagram of the construction of a sandbox data preprocessing and feature engineering in the present invention.

[0050] 2. Clustering-based outlier detection: This paper proposes an innovative clustering method for efficient outlier detection, which has significant technical breakthroughs. First, the method combines stratified sampling, random sampling and K-means++ sampling, balancing sample representativeness and computational efficiency. Through stratified sampling to comprehensively cover different feature categories, random sampling to quickly reduce the data size, and combined with K-means++ to achieve optimal selection of cluster centers, it successfully solves the problems of insufficient random sampling representativeness and high initial computational complexity of K-means++. Figure 4 As shown, Figure 4 The figure is a schematic diagram of a data fusion sampling process in the present invention. Secondly, the K-means algorithm is used to efficiently cluster the candidate points obtained by fusion sampling to ensure accurate processing of large-scale data. In response to the needs of outlier detection, a distance and isolated point detection method based on distance and distance is proposed, and an incremental search strategy is introduced for optimization, which greatly reduces the computational complexity and improves the detection accuracy and efficiency. This hierarchical distance and distance clustering method provides a new solution for outlier detection in complex data environments. Figure 5 As shown, Figure 5 A schematic diagram of the details of distance and outlier detection based on incremental search in the present invention.

[0051] First, in one or more embodiments of the present invention, when a user performs a sandbox game on a digital psychological sandbox, the user's sandbox data may be collected, which may include procedural sandbox game records and image-based sandbox image data.

[0052] For process data preparation and preprocessing, in one or more embodiments of the present invention, the present invention designs detailed recording rules for digital psychological sand table process data, and ensures the independence and traceability of data by assigning a unique operation ID to each step of operation; recording the operation time accurately captures the dynamic trajectory of each user interaction; the association between the operation type and the operation object further reveals the characteristics of user behavior and supports in-depth analysis of the behavior pattern; and the operation position fully reflects the dynamic distribution of the sand tool through three-dimensional spatial coordinates. The innovation of this data design is reflected in the all-round capture and in-depth characterization of the operation behavior, which provides a basic basis for the subsequent noise removal rules, such as combining the operation frequency, time logic and multi-level verification of sand tool matching to ensure high-quality data input and accurate identification of abnormal points.

[0053] At the same time, in order to further improve the consistency and accuracy of the data, a data dictionary was constructed based on the digital psychological sandbox process data (as shown in Tables 1 and 2).

[0054] Table 1 Dictionary of sand tools

[0055]

[0056] Table 2 Sandbox operation dictionary

[0057]

[0058] In one or more embodiments of the present invention, the present invention also proposes a set of systematic elimination rules for the noise data that may exist in the psychological sand table data, aiming to improve the data quality and the accuracy of analysis. The rules cover multiple dimensions: first, by eliminating the records with less than or equal to 20 operations and a total duration of no more than 20 seconds, invalid samples with insufficient data or atypical interactions are effectively eliminated; second, in view of the problem of non-monotonic time records, the sand table game records with non-monotonic user operation time during the recording process can be eliminated to clear inconsistent data caused by input errors or system failures; finally, the sand table game records can be matched with the codes of sand tools and sand table operations in the sand table to eliminate sand table game records involving invalid sand tools. By strictly matching the sand tool dictionary, abnormal records involving invalid sand tools are deleted. This series of elimination rules, combined with the characteristics of psychological sand table data, comprehensively covers the key quality indicators of operation behavior and data records, ensures the validity, consistency and representativeness of the input data, and provides a reliable data basis for subsequent analysis.

[0059] Regarding the preparation and preprocessing of sand table image data, in one or more embodiments of the present invention, the image data of the sand table obtained after the digital sand table game is completed can obtain details such as the sand table background and color distribution that are not displayed in the process data, which is crucial for a deep understanding of the emotional characteristics of the creator.

[0060] In addition, the sand table image data can be preprocessed. For example, the Canny multi-scale edge detection can be used to process the sand table image data. Not only can the fine and coarse edges be captured through different thresholds, but also edge expansion and closing operations are used to ensure the continuity and integrity of the edges. This method significantly improves the separation effect of the foreground of the sand table image data, laying a solid foundation for subsequent feature extraction.

[0061] Furthermore, a higher threshold can be set to capture smaller and more subtle edges. The formula is as follows: .in, is the image after Gaussian blur processing. and They are The lower and upper thresholds for detecting subtle edges, For passing The edge detection result image obtained after algorithm processing contains small and subtle edges.

[0062] You can also set a lower threshold to detect rough edges in the image and help identify the main structure of the image. The formula is as follows: .

[0063] In the formula, This is the result image used to detect the rough edges of the image after being processed by the Canny algorithm. It is the high threshold value when the Canny algorithm detects rough edges. The low threshold for the Canny algorithm to detect rough edges.

[0064] By performing clear masking operations on the foreground and background, the present invention proposes an efficient background removal method, which only retains the image part related to the user operation. This precise foreground extraction method reduces unnecessary information interference and improves the accuracy of image data processing. At the same time, a foreground pixel ratio screening method is set: , the sandbox with foreground pixel ratio lower than the set threshold is regarded as a noise sandbox and removed. is the pixel ratio threshold set, a is the number of foreground pixels of the sandbox image data, b It is the total number of pixels of the sandbox image data.

[0065] Afterwards, from psychology to data science, the pre-processed sandbox game records and sandbox image data can be subjected to feature extraction from multiple dimensions to apply subsequent anomaly detection.

[0066] 1. Extraction of process data features:

[0067] Based on psychological theory, a comprehensive feature extraction framework was designed on the basis of preliminary data preparation to link the user's sandbox game records with emotional characteristics. The user's operation behavior was quantified through four dimensions: aggregation, positivity, symmetry, and diversity, reflecting the multi-dimensional characteristics of the psychological state.

[0068] 1. Clustering dimension: The relevant feature is used to measure the spatial density of objects in the sandbox. High clustering may reflect the maker's emotional closure or need for security, which is manifested as a tendency to psychologically protect or resist external interference, while low clustering may indicate openness and extroversion, reflecting the maker's social inclination and acceptance of the outside world.

[0069] The server can first determine the total number of sand tools after the sand table game, the area occupied by all sand tools, and the spatial distribution density of the sand tools based on the sand table game records; then divide the sand table into grids to determine the number of sand tools in the grid with the most sand tools; determine the grid aggregation degree based on the total number of sand tools and the total number of grids; and determine the number of grids occupied by all sand tools.

[0070] 1) Total number of sand tools in the sandbox , reflecting the complexity of the sandbox.

[0071] 2) The area occupied by all sand tools , take the square convex hull of the occupied area to reflect the distribution of sand tools in the sandbox: .

[0072] In the formula, and It's sand tools. The maximum and minimum coordinates in the direction, and It's sand tools. The maximum and minimum coordinates in a direction.

[0073] 3) Spatial distribution density of sand tools , reflecting the spatial distribution characteristics of sand tools: In the formula, It is the size of a sandbox.

[0074] 4) Mesh features , divide the sandbox into Calculate the number and proportion of sand tools in each grid to reflect the degree of aggregation of sand tools in different areas. Calculate the following aspects respectively:

[0075] i. The number of sand tools in the grid with the most sand tools , reflecting the highest density area of ​​sand tools in the sandbox: .

[0076] In the formula, Indicates The number of sand tools in a grid, Indicates The number of sand tools in a grid.

[0077] ii. Grid aggregation : , and the number of grids occupied by the sand tools : .

[0078] Based on the above grid characteristic indicators, the grid Gini coefficient is calculated by the following formula to reflect the imbalance of the distribution of sand tools in the grid: .

[0079] In the formula, is the number of grids.

[0080] Based on the above constructed aggregation components, the final aggregation score is obtained by weighting : .

[0081] In the formula, … The weights corresponding to the aggregation components can be adjusted according to experiments and expert experience to ensure that the influence of each indicator in the total score is reasonable.

[0082] 2. Positiveness dimension: measures the degree of positivity and negativity of the use of sand tools in the sand tray, thereby reflecting the user's emotional bias. In the digital psychological sand tray, each sand tool has its symbolic meaning, and its emotional value score can be obtained according to the SentimentIntensityAnalyzer in the NLTK library. According to the emotional value of a single sand tool, the positivity score of the entire sand tray can be obtained.

[0083] The server can first determine the proportion of negative emotion sand tools, neutral emotion sand tools and positive emotion sand tools in all the sand tools in the sand tray according to the emotion values ​​symbolized by the sand tools in the sand tray game records; and then determine the total emotion score of all the sand tools in the sand tray, the average emotion value of the sand tools in the sand tray and the median emotion value of the sand tools in the sand tray.

[0084] 1) The distribution ratio of the emotional range of the sand table. The proportion of negative emotional sand tools (The emotion value range of sand tools is [−1.0, −0.6]): .

[0085] Percentage of neutral emotional sand toys (The emotional value range of the sand tool is [−0.2, 0.2]): .

[0086] Proportion of positive emotion sand toys (The emotional value range of sand tools is [0.6, 1.0]): .

[0087] In the formula, Indicates i The emotional score of each sand tool.

[0088] 2) Total emotional score of the sand tools in the sand tray , reflecting the overall situation of evaluating sentiment distribution: .

[0089] 3) The average emotional value of the sand tools in the sand tray and median : .

[0090] Based on the above constructed positivity components, the final positivity score is obtained by weighting : .

[0091] In the formula, … They correspond to the weights of the positivity components respectively.

[0092] 3. Symmetry dimension: reflects whether the arrangement of sand table elements follows a certain symmetry pattern, reflects the design style and layout pattern of the sand table, and reduces noise and errors in clustering. Considering the characteristics of different sand tables, two different symmetry scores are constructed.

[0093] 1) The central symmetry point is the center of all sand tools. Then, the center of all the sand tools is taken as the first central symmetry point, and the symmetry positions of all the sand tools are generated for the horizontal, vertical, positive diagonal, and negative diagonal lines determined based on the first central symmetry point. The distances between each sand tool and the corresponding symmetry position are accumulated to obtain the four symmetry component scores corresponding to the sand tools: ; ; ; .

[0094] In the formula, Indicates The location of the sand tools, Represents the total number of sand tools. Generate symmetric point sets of the original data point sets of the sand tool positions in the sand table process record for the horizontal, vertical, positive diagonal, and negative diagonal lines, and find the points closest to the original data points from the symmetric point sets and sum them up. Among them, Indicates finding the concentrated distance of symmetric points The nearest symmetry point. For the four symmetry component scores corresponding to the sand tool, take the minimum value as the final sand tool center symmetry score .

[0095] 2) The central symmetry point is the center of the sand table. , similar to taking the symmetry center of the sand table, take the center of the sand table as the second central symmetry point, generate the symmetry positions of all the sand tables for the horizontal, vertical, positive diagonal, and negative diagonal lines determined based on the second central symmetry point, respectively, accumulate the distances between each sand table and the corresponding symmetry position, and obtain the four symmetry component scores corresponding to the sand table; take the minimum value from the four symmetry component scores corresponding to the sand table as the final sand table central symmetry score .

[0096] Based on the symmetry components constructed above and , and the final symmetry score is obtained by weighting: In the formula, and They correspond to the weights of the symmetry components respectively.

[0097] 4. Diversity dimension: reflects the diversity of the sandbox, and then reflects the individual's psychological state. Diversity includes individual diversity, that is, the number of types of sand tools; type diversity, that is, the number of different types of sand tools used in the current sandbox; the ratio of individual diversity relative to the number of types; the ratio of individual diversity relative to the number of individuals; the maximum number of repetitions, that is, the maximum number of times a single sand tool is used in a sandbox. Finally, these components are weighted and summed to obtain the final diversity score.

[0098] 5. Process characteristics processing:

[0099] 1) By using RobustScaler to perform robust scaling on feature data and introducing methods for processing the median and interquartile range, the negative impact of noise data on model training is avoided. Compared with traditional scaling methods based on mean and standard deviation, this robust scaling method is more robust in maintaining data structure characteristics: .in, , , are the median, the first and third quartiles, X is the original process characteristic data, It is the process characteristic data after robust scaling.

[0100] 2) To address the long-tail distribution problem of data, logarithmic transformation is used to make skewed data closer to normal distribution, reducing the impact of extreme values ​​on the model: . Where N is the parameter of distribution transformation, is the process characteristic data after logarithmic transformation.

[0101] 3) Use The function performs nonlinear processing on the transformed scores and normalizes them again to unify the scale of the features to improve the discrimination of the data: .

[0102] In the formula, For passing The data after nonlinear processing by the function and normalization again, The score data is logarithmically transformed.

[0103] Use the custom amplification function Amplify to asymmetrically scale the data with the median as the center to amplify the feature differences of different data segments: .

[0104] In the formula, is the center point of the characteristic data of a sand tool dimension process. To adjust the degree of zoom.

[0105] This method of data processing can greatly improve the model's ability to distinguish different clusters and further optimize the interpretability and separability of features (that is, after data processing, data of different categories or clusters can be more clearly distinguished).

[0106] 2. Feature extraction of image data:

[0107] 1. Image feature construction: By focusing on the basic visual elements of the image, it can reflect information such as color and brightness, which helps to express emotional characteristics. It is mainly extracted through the basic visual characteristics of the image such as color, brightness, saturation, etc., and the distribution vector of the current dimension in the entire image range is obtained in each dimension: .

[0108] in, To represent hue, brightness or saturation in different dimensions, For the The corresponding value of pixels in the current dimension, is the histogram vector , is the total number of pixels.

[0109] Afterwards, from feature stratification to clustering optimization, stratified distance and clustering outlier detection can be carried out from large-scale to refined, achieving efficient data point selection based on fusion sampling.

[0110] In the case of uneven data distribution, in order to better represent different parts of the data and improve the processing efficiency of the data set, the present invention proposes a sampling method that combines stratified sampling, random sampling, and K-means++ sampling. This method combines the advantages of the three, namely, low distance calculation overhead in large-scale data sets, and uniform sampling coverage and strong representativeness. The main idea is to first stratify the data set according to certain features, and then use random sampling and K-means++ sampling in each layer.

[0111] Step 1: Based on the feature components constructed in feature engineering, first select the basic visual feature dimensions in the sand tool dimensions and image features as the basis for stratification:

[0112] Process data characteristics: aggregation, positivity, symmetry, and diversity.

[0113] Image data features: hue, brightness, and saturation.

[0114] Of course, in one or more embodiments of the present invention, psychological state dimensions may be added as a stratification basis, that is, three dimensional features may be used as a stratification basis:

[0115] Process data characteristics: aggregation, positivity, symmetry, and diversity.

[0116] Image data features: hue, brightness, and saturation.

[0117] Psychological state: emotional score.

[0118] Step 2: Group different evaluation categories of each feature according to the calculated value of each feature component in feature engineering.

[0119] Process data characteristics: aggregation (low aggregation, medium aggregation, high aggregation), positivity (negative, neutral, positive), symmetry (asymmetric, symmetric), diversity (single, diverse).

[0120] Image data features: hue (cold, neutral, warm), brightness (bright, dark), saturation (low, medium, high).

[0121] Sentiment score: (low, medium, high).

[0122] Step 3: Cross-group categories on each feature dimension to form a combination layer. For example:

[0123] High concentration + positive + symmetry + diversity + warm colors + bright + high saturation + high emotional score, etc.

[0124] Step 4: Randomly select a portion of data points in each layer as initial candidate points. According to experimental results, the number of initial candidate points is generally or Random sampling can efficiently extract candidate points from large-scale data sets, significantly reducing the size of the original data set. Although random sampling has low computational cost, it is less representative and may miss key clustering structures.

[0125] Step 5: In order to make up for the shortcomings of random sampling mentioned above, Based on the initial candidate points, the present invention uses K-means++ cluster sampling and then performs fine clustering on these initial candidate points to obtain Initial cluster centers, where .

[0126] The K-means++ method improves the clustering effect and reduces the risk of falling into local optimality by optimizing the selection of cluster centers. It can better capture the distribution characteristics of the data, that is, the initial centers are distributed in more dispersed positions in the data set, and the initial range of each cluster is more reasonable, which can more accurately represent the overall structure of its cluster.

[0127] In the fusion sampling Based on the initial cluster centers, the present invention can also use the K-means method to Initial cluster centers are clustered to obtain The groups corresponding to the fusion sampling candidate points. Represents the data selected from the original data set by fusion sampling Initial cluster centers, . Indicates the expected number of groups ( ). Represents the final grouping obtained by the K-means algorithm, , . By p arrive K Then k On the one hand, it reduces the distance calculation amount when classifying all subsequent user data points. On the other hand, by gradually converging to the desired number of groups, the selection of fusion sampling candidate points is more stable, more representative, and has better effect.

[0128] After the initial clustering centers complete clustering, all remaining user data points are assigned to the cluster to which the nearest representative point belongs based on their distance from the fusion sampling candidate points: .

[0129] In the formula, Represents a data point in the original dataset. Indicates that clustering is complete There are fused sampling candidate points, and each fused sampling candidate point corresponds to a cluster. Represents user data points and fusion sampling candidate points The distance between. Represents user data points This step enables all user data points corresponding to the original data set to be efficiently assigned to appropriate clusters without the need for complex calculations on the entire data set.

[0130] A new model for isolation assessment can then be built, based on incremental search optimization of distances and outlier identification.

[0131] In each cluster obtained using the K-means method, the present invention uses an innovative distance sum-based outlier detection algorithm. That is, the sum of the distances between each data point and all other data points is calculated to evaluate the isolation of the data point. The distance between each data point and other data points is calculated using the Euclidean distance, and the distance between all data points is calculated: , generating a distance matrix , each element in the matrix Represents data points , The distance between The first i The user data point and j The distance between user data points, The first i User data points, For user data points No. k The characteristics of the dimension, is the total number of feature dimensions of user data points.

[0132] For each data point , calculate the sum of the distances between it and all other data points: , is the sum of the distances between the i-th user data point and all other user data points in the cluster. Distance and Reflects the data points The greater the degree of deviation from other data points, the more likely it is an isolated point.

[0133] In order to make the method consistent for clusters with different numbers of data points and different data sets, the distance sum results are normalized, that is, the distance sum of all data points is normalized to Range: . is the sum of the distances between the cluster and all other user data points and the distance corresponding to the largest user data point, is the sum of the distances between the cluster and all other user data points and the distance corresponding to the user data point with the smallest distance, For the i The isolation of a user's data point. According to the normalized distance sum, the data point with a larger distance sum is an isolated point with a higher degree of deviation.

[0134] The distance sum based isolated point detection algorithm needs to calculate the distance matrix in each cluster. Although the number of data points in each cluster has been greatly reduced in the initial grouping process, it is still necessary to calculate the distance between a point and all other points in the cluster. Based on the distance sum, the present invention proposes an improved distance sum isolated point detection algorithm using incremental search. From the perspective of density, this method observes the "density" around the data point by gradually expanding the radius. As the radius continues to increase, the accumulated distance sum can reflect the distribution density of the data points around the point. If the distance sum exceeds the preset distance threshold after a round of radius expansion, it means that the distribution of data points around the point is relatively sparse, and it can be identified as an isolated point. This density-based isolated point detection method can more effectively capture data points that deviate from the main distribution in the anomaly detection scenario of the digital psychological sandbox. In terms of the application of the distance sum, the algorithm does not simply judge isolated points based on the size of the distance, but comprehensively considers the accumulated distance sum and combined with the radius change. Through this accumulated distance sum method, more abundant information is provided for the judgment of isolated points, which helps to more accurately identify isolated points in complex data distribution.

[0135] Incrementally update the distance sum in the local area of ​​the cluster to which each fusion sampling candidate point belongs, that is, gradually expand the calculated user data points. After each new user data point is introduced, only the distance from the newly added user data point to the current user data point is calculated. The distance is calculated and the result is added to the total distance sum.

[0136] Initially, preset an initial radius (like ), based on each user data point and the preset initial radius , only the initial radius is calculated The distances of the interior points are: Then the radius is expanded according to the preset initial radius and the preset radius increment. For example, the radius is expanded to ,calculate , Between the newly added user data point and the current user data point and add its increment to the distance sum: Through multiple rounds of radius expansion search, the sum of the distances between other user data points within the radius and the user data point is updated during the search process; if the current sum of distances is greater than the preset distance threshold, the search is stopped and the user data point is determined to be an isolated point.

[0137] This method of accumulating distance sums provides richer information for judging isolated points, which helps to more accurately identify isolated points in complex data distribution.

[0138] based on Figure 1The anomaly detection method of the digital psychological sand table shown in the figure, the present invention first extracts the process characteristics of each user's sand table game from multiple dimensions according to the sand table game record, and extracts the image characteristics of each user's sand table game from multiple dimensions according to the sand table image data, to obtain the user data point corresponding to each user, and then cross-combines the different evaluation categories under each dimension, so as to obtain multiple combination layers corresponding to multiple dimensions, the number of dimensions of each combination layer is the same, and the evaluation category for each dimension is different, so as to determine the user data point corresponding to each combination layer, and then anomaly detection can be performed based on the distance metric based on each combination layer. For the user data points corresponding to each combination layer, the cluster center of each combination layer is first selected by combining layered extraction, random extraction and cluster extraction, and then all the user data points corresponding to the combination layer are clustered, and anomaly detection is performed based on the distance between the user data points in each cluster obtained by clustering.

[0139] Compared with anomaly detection based on distance metrics according to all user data points, it is often necessary to perform a large amount of calculations to analyze the degree of distance deviation between each user data point and all other user data points in order to determine its abnormal situation; the present application first divides the user data points into multiple combined layers according to different evaluation categories under each dimension, and then selects high-quality cluster centers by combining layered extraction, random extraction and cluster extraction to perform clustering and anomaly detection, which significantly improves the efficiency of sampling classification and the quality of cluster center selection, avoids ignoring boundary points, and only needs to analyze the abnormal points in each cluster in each combined layer during anomaly detection, reducing the complexity of distance metrics, thereby improving the efficiency while taking into account the accuracy of anomaly detection of the digital psychological sandbox.

[0140] When applying the anomaly detection method of the digital psychological sandbox provided by the present invention, it is not necessary to Figure 1 The steps are executed in the order shown. The specific execution order of the steps can be determined according to needs, and the present invention does not limit this.

[0141] The above is an abnormality detection method for a digital psychological sandbox provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding abnormality detection device for a digital psychological sandbox, such as Figure 6 shown.

[0142] Figure 6 A schematic diagram of an abnormality detection device for a digital psychological sandbox provided by the present invention includes:

[0143] An acquisition module 201 is used to encode the sand tools and sandbox operations in the sandbox, respectively, and obtain sandbox game records of multiple users based on the encoding and sandbox image data after the sandbox game;

[0144] The extraction module 202 is used to extract process features from multiple sand tool dimensions according to the sand table game record for each user; and construct image features according to multiple basic visual feature dimensions of the sand table image data; the multiple sand tool dimensions include the aggregation degree of the sand tools, the positivity symbolized by the sand tools, the symmetry of the sand tools, and the diversity of the sand tools;

[0145] The hierarchical combination module 203 is used to cross-combine different evaluation categories of each sand tool dimension in the process feature with different evaluation categories of each basic visual feature dimension in the image feature to obtain multiple combined hierarchies; determine the user data points formed by the process feature and the image feature of the user corresponding to each combined hierarchy;

[0146] The sampling classification module 204 is used to randomly sample the user data points corresponding to each combination layer to obtain initial candidate points, cluster the initial candidate points through cluster sampling to obtain multiple cluster centers as fusion sampling candidate points; and assign all user data points to the clusters to which the fusion sampling candidate points with the closest distance belong;

[0147] The abnormality determination module 205 is used to determine, in the cluster to which each fused sampling candidate point belongs, isolated points in each user data point as abnormal points according to the distance between each user data point and all other user data points.

[0148] For the specific definition of the anomaly detection device of the digital psychological sandbox, please refer to the definition of the anomaly detection method of the digital psychological sandbox above, which will not be repeated here. Each module in the above-mentioned anomaly detection device of the digital psychological sandbox can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0149] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 Anomaly detection method for digital psychological sandbox is provided.

[0150] The present invention also provides a schematic diagram of the structure of a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 Anomaly detection method for digital psychological sandbox is provided.

[0151] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0152] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. A method for detecting anomalies in a digital psychological sandbox, characterized in that: include: Encode the sand tools and sand table operations in the sand table respectively, and obtain the sand table game records based on the encoding and the sand table image data after the sand table game of multiple users; For each user, process features are extracted from multiple sand tool dimensions based on sand table game records; and constructing image features according to multiple basic visual feature dimensions of the sand table image data; the multiple sand tool dimensions include the degree of aggregation of the sand tools, the positivity symbolized by the sand tools, the symmetry of the sand tools, and the diversity of the sand tools; Cross-combining different evaluation categories of each sand tool dimension in process features with different evaluation categories of each basic visual feature dimension in image features, to obtain multiple combined layers; Determine user data points formed by process features and image features of the user corresponding to each combined layer; Initial candidate points are obtained by random sampling from the user data points corresponding to each combination layer, and the initial candidate points are clustered by cluster sampling to obtain multiple cluster centers as fusion sampling candidate points; All user data points are assigned to the cluster to which the nearest fusion sampling candidate points belong; In the cluster to which each fusion sampling candidate point belongs, isolated points in each user data point are determined as outliers according to the distance between each user data point and all other user data points.

2. The method for detecting anomalies in a digital psychological sandbox according to claim 1, wherein: According to the sand table game records, the process characteristics are extracted from the aggregation of sand tools, including: According to the sand table game records, determine the total number of sand tools after the sand table game, the area occupied by all sand tools, and the spatial distribution density of sand tools; Divide the sand table into grids, determine the number of sand tools in the grid with the most sand tools; determine the grid aggregation degree based on the total number of sand tools and the total number of grids; determine the number of grids occupied by all sand tools; The grid Gini coefficient is determined by the following formula: ; The total number of sand tools, the spatial distribution density of sand tools, the number of sand tools in the grid with the most sand tools, the grid aggregation, the number of grids occupied by all sand tools, and the grid Gini coefficient are weighted and added to determine the final aggregation score of the sand table game record; in, is the grid Gini coefficient, is the number of grids, and Respectively represent and The number of sand tools in a grid, is the average number of sand tools in each grid.

3. The method for detecting anomalies in a digital psychological sandbox according to claim 1, wherein: According to the sand table game records, the process characteristics are extracted from the positivity symbolized by the sand tools, including: According to the emotional values ​​symbolized by the sand tools in the sand table game records, determine the proportion of negative emotional sand tools, neutral emotional sand tools, and positive emotional sand tools in all the sand tools in the sand table; Determine the total emotion score of all the sand toys in the sand tray, the mean emotion value of the sand toys in the sand tray, and the median emotion value of the sand toys in the sand tray; The proportion of negative emotion sand tools, neutral emotion sand tools, and positive emotion sand tools in all sand tools, the total emotion score of all sand tools, the average emotion value of the sand tools, and the median emotion value of the sand tools in the sand tray are weighted and added together to determine the final positivity score of the sand tray game record.

4. The method for detecting anomalies in a digital psychological sandbox according to claim 1, wherein: According to the sand table game records, the process characteristics are extracted from the symmetry of the sand tools, including: Taking the center of all the sand tools as the first central symmetry point, respectively generate the symmetric positions of all the sand tools for the horizontal, vertical, positive diagonal, and negative diagonal determined based on the first central symmetry point, respectively accumulate the distance between each sand tool and the corresponding symmetry position, and obtain the four symmetry component scores corresponding to the sand tools; take the minimum value from the four symmetry component scores corresponding to the sand tools as the final central symmetry score of the sand tools; Taking the center of the sand table as the second central symmetry point, generate the symmetric positions of all sand tools for the horizontal, vertical, positive diagonal, and negative diagonal lines determined based on the second central symmetry point, respectively, and accumulate the distances between each sand tool and the corresponding symmetry position to obtain the four symmetry component scores corresponding to the sand table; take the minimum value from the four symmetry component scores corresponding to the sand table as the final sand table central symmetry score; The center symmetry score of the sand tool and the center symmetry score of the sand table are weighted and added together to determine the final symmetry score of the sand table game record.

5. The method for detecting anomalies in a digital psychological sandbox according to claim 1, wherein: The method of randomly sampling the user data points corresponding to each combined layer to obtain initial candidate points, and clustering the initial candidate points through cluster sampling to obtain multiple cluster centers as fusion sampling candidate points specifically includes: The amount of data randomly sampled from the user data points corresponding to each combined layer is determined according to the following formula: or ; According to the determined data amount, random sampling is performed from the user data points corresponding to each combination layer to obtain initial candidate points; The initial candidate points are finely clustered by K-means++ clustering to obtain multiple initial cluster centers, and the initial cluster centers are clustered by K-means clustering to obtain multiple final cluster centers as fusion sampling candidate points; in, The amount of data randomly sampled for each combination, The total amount of user data points corresponding to each combination stratum.

6. The method for detecting anomalies in a digital psychological sandbox according to claim 1, wherein: The step of determining, in the cluster to which each fused sampling candidate point belongs, isolated points in each user data point according to the distance between each user data point and all other user data points specifically includes: In the cluster to which each fusion sampling candidate point belongs, the distance between each user data point and any other user data point is determined by the following formula: ； The isolation of each user data point is determined by the sum of the distances between each user data point and all other user data points using the following formula: ， ； Determine isolated points among the user data points according to the isolation of each user data point and a preset isolation threshold; in, The first i User data points, For user data points No. k The characteristics of the dimension, is the total number of feature dimensions of user data points, For the i The isolation of each user data point, is the sum of the distances between the cluster and all other user data points and the distance corresponding to the user data point with the smallest distance, is the sum of the distances between the cluster and all other user data points and the distance corresponding to the largest user data point, The first i The user data point and j The distance between user data points, The first i The sum of the distances between a user data point and all other user data points.

7. The method for detecting anomalies in a digital psychological sandbox according to claim 1, wherein: The step of determining, in the cluster to which each fused sampling candidate point belongs, isolated points in each user data point according to the distance between each user data point and all other user data points specifically includes: In the cluster to which each fusion sampling candidate point belongs, according to each user data point and the preset initial radius, the sum of the distances between the user data point and other user data points within the preset radius of the user data point is determined by the following formula: ; The radius is expanded according to the preset initial radius and the preset radius increment. The sum of the distances between the newly added user data point and the user data point after the radius expansion is determined by the following formula, and the sum of the distances between the user data point and other user data points within the preset radius of the user data point is incrementally updated: ; Through multiple rounds of radius expansion search, the sum of distances between other user data points within the radius and the user data point is updated during the search process; if the current sum of distances is greater than a preset distance threshold, the search is stopped and the user data point is determined to be an isolated point; in, For user data points Preset Radius Other user data points and user data points in range The distance between For user data points With user data points The distance between The radius from Expand to Other user data points and user data points within the radius The distance between and.

8. The method for detecting anomalies in a digital psychological sandbox according to claim 1, wherein: The sandbox game record includes: the user's operation time, the sand tool object operated by the user, the operation type of the sand tool object, and the operation position of the sand tool object.

9. The method for detecting anomalies in a digital psychological sandbox according to claim 1, wherein: Before extracting process features from multiple sand tool dimensions according to the sand table game record, the method further includes: Eliminate sandbox game records with less than or equal to 20 operations and a total duration of no more than 20 seconds; Eliminate sandbox game records where the user's operation time is non-monotonic during the recording process; Match the sandbox game records with the codes of the sand tools and sandbox operations in the sandbox, and eliminate the sandbox game records involving invalid sand tools.

10. An abnormality detection device for a digital psychological sandbox, characterized in that: include: An acquisition module is used to encode the sand tools and sand table operations in the sand table, respectively, and obtain sand table game records based on the encoding and sand table image data after the sand table game of multiple users; An extraction module is used to extract process features from multiple sand tool dimensions according to the sand table game records for each user; and constructing image features according to multiple basic visual feature dimensions of the sand table image data; the multiple sand tool dimensions include the degree of aggregation of the sand tools, the positivity symbolized by the sand tools, the symmetry of the sand tools, and the diversity of the sand tools; A hierarchical combination module is used to cross-combine different evaluation categories of each sand tool dimension in the process feature with different evaluation categories of each basic visual feature dimension in the image feature to obtain multiple combined hierarchies; Determine user data points formed by process features and image features of the user corresponding to each combined layer; The sampling and classification module is used to randomly sample the user data points corresponding to each combination layer to obtain initial candidate points, and cluster the initial candidate points through cluster sampling to obtain multiple cluster centers as fusion sampling candidate points; All user data points are assigned to the cluster to which the nearest fusion sampling candidate points belong; The anomaly determination module is used to determine, in the cluster to which each fusion sampling candidate point belongs, isolated points in each user data point as an anomaly point according to the distance between each user data point and all other user data points.

Images