AI-based static intelligent training and evaluation method and device

Abstract

The invention discloses an AI-based intelligent silence training and evaluation method and device, relates to the technical field of psychological training and evaluation, and aims to solve the technical problem that the training effect is difficult to accurately judge in the prior art. The method comprises the following steps: acquiring a house tree person image drawn by a user and a sand table scene image built by the user through a high-speed photographic apparatus; voice descriptions of the user on the room tree person image and the sand table scene image are collected through a microphone, and the voice descriptions are converted into text data; performing feature extraction on the house tree person image to obtain house tree person image features, and generating house tree person psychological feature data in combination with corresponding text data; sand tool identification and spatial relation extraction are carried out on the sand table scene image to obtain sand table scene features, and sand table psychological feature data are generated in combination with corresponding text data; generating a comprehensive psychological assessment result; and matching the corresponding silence training course. The accuracy of an evaluation result can be improved, and a reliable basis is provided for training adjustment.

Description

Technical Field

[0001] This invention relates to the field of intelligent psychological training and assessment, specifically to an AI-based method and device for mindfulness-based intelligence training and assessment, applicable to scenarios of mindfulness-based intelligence enhancement and mental health intervention for individuals or institutions. Background Technology

[0002] With the accelerating pace of modern life and the increasing demand for mental health services, mindfulness training, as an effective intervention for improving emotional regulation and alleviating psychological stress, has been widely adopted in various fields such as healthcare, education, and employee care. Currently, mindfulness-related technologies are developing around two core aspects: training implementation and effectiveness evaluation, resulting in multiple pathways. Traditional mindfulness training techniques are centered on manual guidance and standardized content, such as Mindfulness-Based Stress Reduction (MBSR) courses taught by professional instructors, printed books, and pre-recorded meditation audio, or simple tools with time reminders and breathing light prompts, lacking overall data-driven assessment and adjustment capabilities. While digital mindfulness training technologies provide standardized meditation content through apps or use hardware such as mindfulness meditation pillows and smart bracelets to record training time and heart rate, their assessments largely rely on subjectively completed user-filled Self-Rating Anxiety Scale (SAS) and Mindfulness-Based Attention Scale (MAAS).

[0003] However, existing technologies generally suffer from three major flaws: First, training and evaluation are severely disconnected. They either rely on subjective questionnaires, leading to biased results, or collect only isolated objective data without being able to correlate with key indicators, making it difficult to accurately judge the training effect and provide a basis for subsequent adjustments. Second, the level of intelligence is insufficient. Whether it is traditional courses, digital products, or single hardware, they are all based on standardized content. Third, the functions are fragmented. Training, evaluation, and data viewing are scattered across multiple devices and platforms. Users need to train in an app, evaluate in a separate software, and check data in a wristband app. The operation is cumbersome and the data is fragmented, failing to form a coherent experience of training-evaluation-adjustment, which seriously affects the user's focus and continuity in training.

[0004] Therefore, it is necessary to design a technical solution that can overcome the above-mentioned defects. Summary of the Invention

[0005] One objective of this invention is to provide an AI-based method and apparatus for observing intelligent training and evaluation, which can improve the accuracy of evaluation results and provide a reliable basis for training adjustments.

[0006] To achieve these objectives and other advantages of the present invention, according to one aspect of the present invention, an AI-based method for contemplative intelligent training and evaluation is provided, comprising: S1: acquiring images of the house-tree-person (HTP) drawn by the user and images of a sand table scene built by the user using a high-speed scanner, and acquiring the user's voice description of the HTP images and sand table scene images using a microphone, and converting the voice description into text data; S2: extracting features from the HTP images using a deep learning-based object detection algorithm to obtain HTP image features, and combining them with corresponding text data to generate HTP psychological feature data; S3: using an image segmentation algorithm based on 3D vision technology to perform sand table object recognition and spatial relationship extraction on the sand table scene images to obtain sand table scene features, and combining them with corresponding text data to generate sand table psychological feature data; S4: weightedly fusing the HTP psychological feature data, sand table psychological feature data, physiological data and behavioral data collected during contemplative training to generate a comprehensive psychological evaluation result; S5: matching corresponding contemplative training courses from a training content database based on the comprehensive psychological evaluation result.

[0007] Furthermore, in S1, the human-computer interaction interface provides a dedicated data acquisition entry point; responding to user-triggered operations, it invokes the document scanner and simultaneously displays a dynamic shooting guidance interface on the human-computer interaction interface. The shooting guidance interface includes a visual alignment box and text prompts to guide the user to place the drawn house-tree-person image or the user-built sand table scene image within the optimal shooting area; upon receiving an image capture command, it controls the document scanner to capture the original image; and it provides an image preprocessing toolset, which includes at least image cropping, rotation, brightness adjustment, and contrast adjustment functions, and receives one or more interactive preprocessing operations performed by the user on the original image through the image preprocessing toolset.

[0008] Furthermore, in S2, an object detection algorithm is used to identify and extract the outlines of houses, trees, and people from the house-tree-person image, including the number of doors and windows in the house, the shape of tree branches, and facial expression features of the person. Simultaneously, the color distribution histogram of the house-tree-person image is extracted as color features, and the relative distance between the center point of the person's outline and the center point of the house's outline is calculated as a composition feature. The extracted features are input into a pre-set psychological feature mapping table for matching. The psychological feature mapping table stores the correspondence between features and psychological feature entries, outputting an initial psychological feature vector. Keyword extraction and sentiment analysis are performed on the text data to generate a user subjective feedback vector. The initial psychological feature vector and the user subjective feedback vector are weighted and fused to generate the final house-tree-person psychological feature data. The house-tree-person psychological feature data includes quantified values ​​of family relationship intimacy, self-awareness and confidence, and emotional state calmness tendency.

[0009] Furthermore, the initial psychological feature vector and the user's subjective feedback vector are input into an attention mechanism network. The attention mechanism network is trained to output dynamic weights corresponding to the consistency between the objective features of the image and the user's subjective description. When the emotional polarity value in the user's subjective feedback vector is highly consistent with the predicted emotional tendency in the initial psychological feature vector, a higher first weight is assigned to the user's subjective feedback vector, and a lower second weight is assigned to the initial psychological feature vector. When there is a significant inconsistency between the two, the first weight is reduced and the second weight is increased. Based on a pre-set psychological rule base, feature conflicts between the initial psychological feature vector and the user's subjective feedback vector are detected. If a conflict is detected, a conflict resolution procedure is initiated: similar conflict cases in historical data and their final evaluation results confirmed by experts are called, and the bias weight of the current conflict is calculated through similarity matching, and the dynamic weight is corrected accordingly. Using the corrected or uncorrected dynamic weights, the initial psychological feature vector and the user's subjective feedback vector are weighted and summed, and the summation result is mapped onto a predefined psychological scale to generate House-Tree-Person (HTP) psychological feature data containing quantitative values ​​of family relationship intimacy, self-awareness and self-confidence, and emotional state calmness tendency.

[0010] Further, in S3, depth images of the sand table scene are acquired using a depth camera based on structured light principles. Combined with RGB images, an image segmentation algorithm is used to identify and classify sand table objects, including human figures, animals, and buildings. The three-dimensional spatial distances between sand table objects and the sequence of lines connecting their center points are calculated based on the depth images. Each identified sand table object type and its three-dimensional spatial relationship are matched with a pre-set database of symbolic meanings of sand table objects, and an initial symbolic weight is assigned to each sand table object and its spatial relationship. Semantic analysis is performed on the text data to extract the user's subjective interpretation of specific sand table objects and their placement. Based on this subjective interpretation, the initial symbolic weights of the corresponding sand table objects are dynamically adjusted to generate a weighted symbolic feature vector. The weighted symbolic feature vector is input into a pre-trained long short-term memory network model. The long short-term memory network model is trained using a dataset of real sand table cases and corresponding professional psychological assessment results. The output sand table psychological feature data includes quantitative values ​​of emotional need intensity, psychological stress level, and subconscious conflict index.

[0011] Furthermore, based on the temporal data of sand table placement recorded by a depth camera during user sand table construction, the weighted symbolic feature vectors are constructed into a spatiotemporal input sequence according to time order. Each time step in the sequence contains the type of sand table placement, its symbolic weight, and its three-dimensional coordinates within the sand table. The sequence, ordered by placement time, is input into a first long short-term memory network model, which specifically learns the dynamic evolution pattern of user construction behavior and outputs a temporal feature vector. The final sand table layout is then subjected to cluster analysis to identify core and peripheral areas, generating a spatial structure sequence from core to edge. This spatial structure sequence is input into a second long short-term memory network model, which learns the static geometry and symbolic structure of the overall sand table layout and outputs a spatial feature vector. The temporal feature vector and the spatial feature vector are concatenated and input into a fully connected neural network, outputting quantitative values ​​for emotional need intensity, psychological stress level, and subconscious conflict indicators.

[0012] Furthermore, a multi-source data vector is constructed, including House-Tree-Person (HTP) psychological characteristic data, sandplay psychological characteristic data, physiological data collected by sensors during mindfulness training, and behavioral data recorded through a human-computer interaction interface. This multi-source data vector is then input into a pre-trained decision tree model. The decision tree model is trained based on the user's historical assessment data and expert annotation results, and is used to dynamically allocate the first weight value for the HTP psychological characteristic data, the second weight value for the sandplay psychological characteristic data, the third weight value for the physiological data, and the fourth weight value for the behavioral data. Based on the first, second, third, and fourth weight values ​​output by the decision tree model, a weighted sum is calculated on the multi-source data vector to generate the comprehensive psychological assessment result.

[0013] Furthermore, the confidence level of image data is calculated based on the clarity and integrity of house-tree-person images and sand table scene images; the confidence level of text data is calculated based on the emotional word density and syntactic complexity of the speech description text data; the confidence level of physiological data is calculated based on the signal stability of physiological data within the acquisition time window; and the confidence level of behavioral data is calculated based on the event record integrity of behavioral data. The consistency of the first evaluation index of emotional state in the house-tree-person psychological feature data and the sand table psychological feature data is compared; the consistency of the second evaluation index representing emotional arousal in the physiological data is compared with the first evaluation index; and the consistency comparison results are verified based on abnormal interaction events recorded in the behavioral data. In response to the consistency comparison results, the first weight value is dynamically adjusted to the fourth weight value: when the evaluation indicators of two or more data sources are consistent, the weight of the corresponding weight value of the data source is increased; when the evaluation indicators of a certain data source significantly conflict with other data sources, and its data quality confidence level is lower than a preset threshold, the weight of the corresponding weight value of that data source is decreased.

[0014] Furthermore, it also includes: periodically aggregating anonymized multi-source data vectors from multiple users and corresponding expert-confirmed evaluation results on the cloud platform to form an incremental training dataset; using the incremental training dataset to incrementally train the decision tree model and generate an updated decision tree model; based on the historical multi-source data vector sequence of a specific user, performing personalized fine-tuning on the updated decision tree model to generate a personalized decision tree model for the corresponding user; the personalized fine-tuning process uses temporal cross-validation to optimize the model's adaptability to the evolution trend of the user's psychological state; and deploying the personalized decision tree model or the updated decision tree model for subsequent dynamic allocation of weight values.

[0015] According to another aspect of the present invention, an AI-based static intelligent training and evaluation device is also provided, comprising: a data acquisition module, configured with a document scanner and a microphone, for acquiring user-drawn house-tree-person images, user-built sand table scene images, and user's voice descriptions of the house-tree-person images and sand table scene images; a human-computer interaction module, for providing a data acquisition entry point and displaying a dynamic shooting guidance interface; a speech-to-text module, connected to the data acquisition module, for converting the voice descriptions acquired by the microphone into text data; and a house-tree-person feature analysis module, connected to the image preprocessing module and the speech-to-text module, for using a deep learning-based object detection algorithm to extract features from the standardized house-tree-person images, obtaining house-tree-person image features, and combining them with the corresponding text data to generate house-tree-person psychological feature data. The system includes: a sand table feature analysis module, connected to the image preprocessing module and the speech-to-text module; a sand table feature analysis module, connected to the image preprocessing module and the speech-to-text module; a multi-source data fusion and evaluation module, connected to the house-tree-person feature analysis module and the sand table feature analysis module; a multi-source data fusion and evaluation module, connected to the house-tree-person feature analysis module and the sand table feature analysis module; a multi-source data fusion and evaluation module, connected to the multi-source data fusion and evaluation module; and a training course matching module, connected to the multi-source data fusion and evaluation module. The training course matching module matches the corresponding mindfulness training course from the training content database based on the comprehensive psychological evaluation result.

[0016] The present invention has at least the following beneficial effects: This invention utilizes a high-speed scanner and microphone to collect multimodal images of the house-tree-person (HTP), sand table scene images, and voice descriptions. It combines deep learning and 3D vision technology to extract features and integrate text data, breaking through the limitations of existing technologies that rely on subjective or isolated data evaluation. This achieves a deep fusion of objective features and subjective feedback, resulting in more comprehensive psychological feature data and more accurate overall evaluation results, providing a reliable basis for training adjustments. Based on the comprehensive psychological evaluation results, mindfulness training courses are matched, changing the traditional standardized training model and enabling personalized course adaptation. This accurately meets the psychological needs of different users, improving the targeting and effectiveness of training. Integrating the entire process of data collection, feature analysis, evaluation fusion, and course matching avoids fragmented operation across multiple devices, forming a closed loop of collection-evaluation-training. This simplifies user operation, ensures a focused and immersive experience in mindfulness training, and promotes the development of mindfulness training towards higher efficiency and intelligence.

[0017] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0018] Figure 1 This is a flowchart of one embodiment of this application. Detailed Implementation

[0019] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0020] It should be understood that terms such as "having," "comprising," and "including" used in the embodiments of this application do not exclude the presence or addition of one or more other elements or combinations thereof. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationship and movement of components in a specific posture. If the specific posture changes, the directional indication will also change accordingly. When an element is referred to as "fixed to" or "set on" another element, it can be directly on the other element or may have an intervening element present. When an element is referred to as "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element through an intervening element. Descriptions involving "first," "second," etc., in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features.

[0021] It should be noted that the technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0022] like Figure 1 As shown, embodiments of this application provide an AI-based mindfulness training and evaluation method, including: S1: acquiring user-drawn house-tree-person (HTP) images and user-built sand table scene images using a high-speed scanner, and acquiring user's voice descriptions of the HTP images and sand table scene images using a microphone, converting the voice descriptions into text data; S2: extracting features from the HTP images using a deep learning-based object detection algorithm, obtaining HTP image features, and combining them with corresponding text data to generate HTP psychological feature data; S3: using an image segmentation algorithm based on 3D vision technology to perform sand table object recognition and spatial relationship extraction on the sand table scene images, obtaining sand table scene features, and combining them with corresponding text data to generate sand table psychological feature data; S4: weightedly fusing the HTP psychological feature data, sand table psychological feature data, physiological data and behavioral data collected during mindfulness training to generate a comprehensive psychological evaluation result; S5: matching corresponding mindfulness training courses from the training content database based on the comprehensive psychological evaluation result.

[0023] For example, in step S1, the document scanner can be the Chengzhe Technology Elf EA3 MAX model, which can be placed 15-20 cm above the table to capture the house-tree-person image drawn by the user on A4-sized drawing paper, as well as the sand table scene image built by the user in a standard sand table with a length of 60 cm, a width of 40 cm, and a height of 8 cm; the microphone can be the Logitech Blue Yeti Nano model, which can be placed 30-50 cm away from the user's mouth on the table to capture the user's voice description of the house-tree-person image (such as describing the number of doors and windows of the house, the shape of tree branches and leaves, etc.) and the sand table scene image (such as describing the placement of sand table objects, their meanings, etc.); the process of converting speech to text data can be implemented through Baidu AI Cloud's speech recognition API, which can convert the captured speech signal (sampling rate 16000Hz or 48000Hz) into UTF-8 encoded text data. In step S2, the deep learning-based object detection algorithm can be YOLOv8. This algorithm performs pixel-by-pixel analysis on the house-tree-person image (resolution set to 1920×1080 pixels) to extract house outlines (such as rectangular or irregular polygonal outlines), tree outlines (such as branching outlines), person outlines (such as upright human figure outlines), the number of doors and windows in the house (such as 2 doors or 4 windows), the shape of tree branches (such as thick branches or thin branches), and facial expression features of the person (such as smiling or calm expression), thereby obtaining house-tree-person image features. The corresponding text data is the text converted in S1. The image features and text data are combined by vector concatenation. For example, the image feature vector (512 dimensions) and the text data vector (256 dimensions) are concatenated into a 768-dimensional fusion vector, which is then input into the pre-trained classification model to generate house-tree-person psychological feature data. In step S3, the image segmentation algorithm based on 3D vision technology can be the Mask R-CNN algorithm. This algorithm combines RGB images (1280×720 pixels resolution) and depth images (depth range 0.5-3 meters) of the sand table scene acquired by the Kinect Azure depth camera (which can be placed 80-100 cm directly above the sand table) to identify the sand table objects. The sand table objects can include human figures (such as a 5 cm tall human-shaped doll), animal figures (such as an 8 cm long bear doll), and building figures (such as a 10 cm tall house model). At the same time, the three-dimensional spatial distance between the sand table objects (such as a 15 cm distance between the center points of two sand table objects) and the sequence of lines connecting the center points (such as a broken line sequence formed according to the order in which the sand table objects are placed) are calculated through the depth images to obtain the sand table scene features. Then, these features are combined with the text data of the corresponding sand table scene in S1. For example, an attention mechanism is used to assign higher weights to the sand table object features mentioned in the text to generate sand table psychological feature data.In step S4, the House-Tree-Person (HTP) psychological characteristic data may include quantitative indicators such as family relationship intimacy and self-awareness confidence (range 0-100 points). The sandplay psychological characteristic data may include quantitative indicators such as emotional need intensity and psychological stress level (range 0-100 points). Physiological data can be collected by a wristband (worn on the user's wrist) with heart rate (normal range 60-100 beats / minute) and blood oxygen saturation (normal range 95%-100%). Behavioral data can be collected by the user's training duration (e.g., 30 minutes) and operation frequency (e.g., 5 times / minute) recorded by the computer-side interaction log. During the weighted fusion process, initial weights can be assigned to the four types of data (HTP psychological characteristic data weight 0.3, sandplay psychological characteristic data weight 0.3, physiological data weight 0.2, and behavioral data weight 0.2). Then, the quantitative value of each type of data is multiplied by the corresponding weight and summed to obtain the comprehensive psychological assessment result (range 0-100 points). In step S5, the training content database can be a database stored on a local server, containing 100-200 mindfulness training courses. Each course corresponds to a specific range of comprehensive psychological assessment results (e.g., a comprehensive score of 60-70 points corresponds to a basic stress reduction course). After the comprehensive psychological assessment results are generated, the course corresponding to the range of the result is matched through a database query statement. For example, a comprehensive score of 55 points matches a beginner emotion regulation course.

[0024] In existing technologies, common mindfulness training and assessment methods provide standardized meditation audio through mobile apps, while users fill out paper-based Self-Rating Anxiety Scales (SAS) to provide subjective feedback. Physiological data is collected via separate smart bracelets, and there is no correlation analysis between different data types. Fixed courses can only be recommended based on the user's subjective questionnaire results. This embodiment uses multimodal data collection via devices such as document scanners and microphones, combines deep learning and 3D vision technology to achieve precise feature extraction, then weights and fuses multi-source data to generate assessment results, and finally matches corresponding courses. This solves the problems of single data collection, one-sided assessment, and untargeted course recommendations in existing technologies. It can more comprehensively reflect the user's psychological state, provide users with training content that better meets their needs, and all data processing and course matching processes are seamless, eliminating the need for switching between multiple devices.

[0025] In another embodiment, in S1, the human-computer interaction interface provides a dedicated data acquisition entry point; responding to the user's trigger operation, it calls the high-speed document scanner and simultaneously displays a dynamic shooting guidance interface on the human-computer interaction interface. The shooting guidance interface includes a visual alignment box and text prompts to guide the user to place the drawn house-tree-person image or the sand table scene image built by the user in the optimal shooting area; after receiving the image capture command, it controls the high-speed document scanner to take a picture to obtain the original image; it provides an image preprocessing toolset, which includes at least image cropping, rotation, brightness adjustment, and contrast adjustment functions, and receives one or more interactive preprocessing operations performed by the user on the original image through the image preprocessing toolset.

[0026] For example, in S1, the human-computer interaction interface can be a software interface running on a computer. This interface has a dedicated data acquisition entry point on the left side, presented as a "Start Acquisition" button. The button is 3 cm × 1.5 cm in size and light blue in color. When the user clicks the "Start Acquisition" button (i.e., triggers the operation), the software calls the document scanner (such as the Chengzhe Elf EA3 MAX) via USB interface. Simultaneously, a dynamic shooting guidance interface is displayed in the center of the human-computer interaction interface. The visual alignment frame of this guidance interface is a black rectangle, with the dimensions within the frame matching the optimal shooting range of the document scanner (e.g., 21 cm × 29.7 cm, corresponding to A4 paper size). Text prompts are displayed in white font below the alignment frame, stating "Please place the house-tree-person image or sand table scene within the frame, ensuring the image edges are unobstructed," with a font size of 12. Once the user confirms that the image has been placed within the alignment frame and clicks the "Shoot" button on the interface (i.e., the image capture command), the software sends a control signal to the document scanner to control the lens (with an adjustable focal length of 50mm) to focus and take a picture, acquiring an original image with a resolution of 28 million pixels (5152×5152 pixels) in JPEG format. Meanwhile, the right side of the human-computer interaction interface provides an image preprocessing toolset, arranged in the form of icons, each corresponding to a function: the image cropping function allows users to select the cropping area by dragging the mouse, and the minimum resolution of the cropped image is no less than 1000×1000 pixels; the rotation function supports rotating the image in units of 90 degrees, 180 degrees, or 270 degrees; the brightness adjustment function is controlled by a slider, with an adjustment range of -50 to +50 (initial value is 0), and each unit adjustment corresponds to a 1% change in brightness value; the contrast adjustment function is also controlled by a slider, with an adjustment range of -30 to +30 (initial value is 0), and each unit adjustment corresponds to a 1% change in contrast value; users can select one or more functions by clicking the corresponding icon, for example, first rotating the image by 90 degrees, and then adjusting the brightness to +10. The software receives the user's operation commands in real time and processes the original image, and the processed image is displayed in real time in the interface preview area.

[0027] In existing technologies, the human-computer interaction interface lacks dedicated guidance during data acquisition, requiring users to place images based on experience. This can easily lead to image misalignment or edge loss. Furthermore, the raw images lack built-in preprocessing tools, necessitating export to third-party software like Photoshop for processing, which is cumbersome. This embodiment, through a dedicated data acquisition entry point and dynamic guidance interface, helps users quickly place images in the optimal shooting area, reducing shooting errors. Simultaneously, the built-in preprocessing toolset allows users to adjust images without switching software, simplifying the operation process and ensuring more stable image quality during subsequent feature extraction, preventing image issues from affecting the accuracy of evaluation results.

[0028] In another embodiment, in S2, an object detection algorithm is used to identify and extract the outlines of houses, trees, and people from the house-tree-person image, including the number of doors and windows in the house, the shape of tree branches, and facial expression features of the person. Simultaneously, the color distribution histogram of the house-tree-person image is extracted as a color feature, and the relative distance between the center point of the person's outline and the center point of the house's outline is calculated as a composition feature. The extracted features are input into a pre-set psychological feature mapping table for matching. The psychological feature mapping table stores the correspondence between features and psychological feature entries, outputting an initial psychological feature vector. Keyword extraction and sentiment analysis are performed on the text data to generate a user subjective feedback vector. The initial psychological feature vector and the user subjective feedback vector are weighted and fused to generate the final house-tree-person psychological feature data. The house-tree-person psychological feature data includes quantified values ​​of family relationship intimacy, self-awareness and confidence, and emotional state calmness tendency.

[0029] For example, in S2, the target detection algorithm used can be the YOLOv8 algorithm, which runs on a server configured with an NVIDIA RTX 4090 graphics card to process the house-tree-person image (resolution adjusted to 640×640 pixels): The algorithm's feature extraction network (backbone is CSPDarknet) identifies the outlines of houses (e.g., rectangular, with an edge error of less than 2 pixels), tree outlines (e.g., branched, with an error of less than 1 branch), and human outlines (e.g., upright human figure, with over 95% completeness of body parts). It also counts the number of doors and windows in the houses (e.g., identifying 2 doors and 4 windows), determines the shape of tree branches (e.g., thick branches with a diameter greater than 5 pixels), and analyzes the facial features of the person. The system identifies facial features (such as calm expressions, with an accuracy rate of over 90% in recognizing facial muscle states); it also extracts the color distribution histogram of the house-tree-person image using the OpenCV library, selecting RGB as the color space and setting the number of bins per channel to 256, resulting in 3×256 dimensional color features; when calculating the relative distance between the center point of the person's outline and the center point of the house's outline, the system first marks the coordinates of the center points of the two outlines using an algorithm (e.g., the center point coordinates of the person (320, 400) and the center point coordinates of the house (320, 200)), and then calculates the relative distance using the Euclidean distance formula, finding it to be 200 pixels. The extracted contour features, number of doors and windows, branch morphology, facial expression features, color features, and composition features are integrated into a 1024-dimensional feature vector. This vector is then input into a pre-built psychological feature mapping table in a local server database (such as a MySQL database). This mapping table stores the correspondence between features and psychological feature entries. For example, "number of doors and windows ≥ 3" corresponds to the psychological feature entry of a family with a high degree of openness. By looking up the table, an initial 128-dimensional psychological feature vector is output (each dimension ranges from 0 to 1). When extracting keywords from the text data (i.e., the UTF-8 encoded text converted in S1), the TF-IDF algorithm is used to extract keywords such as "warmth" and "safety" (5-10 keywords are extracted). Sentiment analysis uses the sentiment analysis API of Baidu AI Open Platform to output the sentiment polarity value of the text (ranges from -1 to 1, where -1 represents negative and 1 represents positive). The keyword vector (512 dimensions) is combined with the sentiment polarity value to generate a 256-dimensional user subjective feedback vector (each dimension ranges from 0 to 1).When weighted and fused the initial psychological feature vector and the user subjective feedback vector, weights are first assigned to the two vectors (0.6 for the initial psychological feature vector and 0.4 for the user subjective feedback vector). Then, the values ​​of the corresponding dimensions are multiplied and summed to obtain a 128-dimensional fused vector. Finally, the fused vector is mapped to a quantization range of 0-100 points to generate the final House-Tree-Person psychological feature data. The quantization values ​​for family relationship intimacy (e.g., 75 points), self-awareness and self-confidence (e.g., 80 points), and emotional state calmness tendency (e.g., 65 points) are all within the range of 0-100 points.

[0030] In existing technologies, house-tree-person image feature extraction relies on manual observation, which is inefficient and prone to subjective errors. Text data is merely recorded without sentiment analysis, and psychological feature data is mostly qualitative description without quantitative values, failing to accurately reflect the user's psychological state. This embodiment automatically extracts multi-dimensional image features using object detection algorithms, combines these with algorithms for keyword extraction and sentiment analysis of text data, and then weights and fuses the two types of features to generate quantified psychological feature data. This not only improves the efficiency and accuracy of feature extraction but also achieves a quantitative expression of psychological features, providing more accurate data support for subsequent comprehensive evaluation.

[0031] In another embodiment, the AI-based static intelligence training and evaluation method described above is characterized by inputting an initial psychological feature vector and a user subjective feedback vector into an attention mechanism network. The attention mechanism network is trained to output dynamic weights corresponding to the consistency between objective image features and the user's subjective description. When the emotional polarity value in the user's subjective feedback vector is highly consistent with the predicted emotional tendency in the initial psychological feature vector, a higher first weight is assigned to the user's subjective feedback vector, and a lower second weight is assigned to the initial psychological feature vector. When there is a significant inconsistency between the two, the first weight is reduced and the second weight is increased. Based on the prediction... The system uses a pre-defined psychological rule base to detect feature conflicts between the initial psychological feature vector and the user's subjective feedback vector. If a conflict is detected, a conflict resolution procedure is initiated: similar conflict cases from historical data and their final evaluation results confirmed by experts are retrieved, and the bias weight of the current conflict is calculated through similarity matching, and the dynamic weight is adjusted accordingly. Using the adjusted or unadjusted dynamic weight, the initial psychological feature vector and the user's subjective feedback vector are weighted and summed, and the summation result is mapped onto a predefined psychological scale to generate House-Tree-Person (HTP) psychological feature data that includes quantitative values ​​for family relationship intimacy, self-awareness and self-confidence, and emotional calmness tendencies.

[0032] For example, an initial psychological feature vector (128 dimensions, value range 0-1) and a user subjective feedback vector (256 dimensions, value range 0-1) are input into an attention mechanism network. This network can be an attention model based on the Transformer architecture. The network is pre-trained on training data labeled with the correspondence between feature consistency and dynamic weights, and can output dynamic weights in the 0-1 range. For example, the attention mechanism network adopts a 2-layer Transformer encoder structure, with the hidden layer dimension set to 256, configured with 4 attention heads, and ReLU activation function. It calculates the association weight between the objective feature vector of the image and the user subjective feedback vector through a self-attention mechanism, and outputs dynamic weights. The sentiment polarity value is calculated by the sentiment analysis module from the user's subjective feedback vector (range -1 to 1). The predicted sentiment tendency in the initial psychological feature vector is calculated by the vector mean (mean > 0.6 is considered positive, < 0.4 is considered negative, and 0.4-0.6 is considered neutral). When the sentiment polarity value > 0.6 and the initial sentiment tendency is positive (highly consistent), the first weight (user subjective feedback vector weight) can be assigned 0.7, and the second weight (initial psychological feature vector weight) can be assigned 0.3. When the sentiment polarity value < 0.4 but the initial sentiment tendency is positive (significantly inconsistent, difference > 0.2), the first weight is reduced to 0.3, and the second weight is increased to 0.7. A pre-built psychological rule base can be stored in a PostgreSQL database. This database contains over 200 conflict rules, such as image features indicating distant family relationships but subjective descriptions of family harmony. A rule-matching algorithm detects conflicts between two types of vectors. For example, if the initial psychological feature vector has a family relationship intimacy metric score <40, but the user's subjective feedback vector shows a family harmony keyword percentage >60%, it is considered a conflict. When the conflict resolution program starts, it calls a historical conflict case database (containing over 50,000 cases) stored locally on the server. A cosine similarity algorithm is used to calculate the similarity between the current conflict vector and historical case vectors (the similarity threshold is set to 0.8; values ​​above the threshold are considered similar cases). For example, if the current conflict similarity is 0.85 with a historical case, the result of the expert-confirmed bias towards the initial psychological feature vector weight of 0.6 in that case is used, and the current dynamic weights are adjusted to a first weight of 0.4 and a second weight of 0.6. When performing weighted summation, if the initial weights are 0.7 and 0.3, the initial value of a certain dimension of the psychological feature vector is 0.6, and the corresponding dimension value of the user's subjective feedback vector is 0.8, then the weighted sum of that dimension is 0.6 × 0.3 + 0.8 × 0.7 = 0.74. Then, the weighted sum of all dimensions is mapped to a psychological scale of 0-100 points (such as the sub-dimensional scoring criteria of the SCL-90 scale). For example, 0.74 corresponds to 74 points, and finally, the House-Tree-Person psychological feature data are generated, including quantitative values ​​for family relationship intimacy (such as 74 points), self-awareness and self-confidence (such as 78 points), and emotional state calmness tendency (such as 68 points).

[0033] This embodiment dynamically allocates weights through an attention mechanism network, combines a psychological rule base to detect conflicts and call historical cases to resolve them, and can achieve intelligent weight adjustment without manual intervention, ensuring that the final psychological feature data is more in line with the user's real state and reducing evaluation bias caused by data conflicts.

[0034] In another embodiment, in S3, a depth image of the sand table scene is acquired using a depth camera based on structured light principles. Combined with the RGB image, an image segmentation algorithm is used to identify and classify the sand table objects, including human figures, animals, and buildings. The three-dimensional spatial distances between the sand table objects and the sequence of lines connecting their center points are calculated based on the depth image. Each identified sand table object type and its three-dimensional spatial relationship are matched with a pre-set database of symbolic meanings of sand table objects, and an initial symbolic weight is assigned to each sand table object and its spatial relationship. Semantic analysis is performed on the text data to extract the user's subjective interpretation of specific sand table objects and their placement. Based on this subjective interpretation, the initial symbolic weights of the corresponding sand table objects are dynamically adjusted to generate a weighted symbolic feature vector. The weighted symbolic feature vector is input into a pre-trained long short-term memory network model. The long short-term memory network model is trained using a dataset of real sand table cases and corresponding professional psychological assessment results. The model outputs sand table psychological feature data, which includes quantitative values ​​of emotional need intensity, psychological stress level, and subconscious conflict index.

[0035] For example, in S3, the structured light-based depth camera can be an Intel RealSense D455, which can be fixed 100-120 cm directly above the sand table to acquire depth images (1280×720 pixels, depth accuracy ±2%) and RGB images (1920×1080 pixels) of the sand table scene at a frame rate of 30 frames per second. The image segmentation algorithm used can be Mask R-CNN, which runs on a workstation equipped with an NVIDIA A100 graphics card (such as a Dell Precision 7920). By performing pixel-level segmentation on the RGB images, it identifies and classifies the sand table objects: human figures can include plastic dolls 3-8 cm tall (such as Lego minifigures), animal figures can include resin animal models 5-10 cm long (such as bear and deer models), and architectural figures can include wooden architectural models 6-15 cm high (such as small houses and bridge models). When calculating the 3D spatial distance between sandplay figures based on depth images, the 3D coordinates of each sandplay figure are first extracted using an algorithm (e.g., sandplay figure A has coordinates of (20, 30, 5) cm, sandplay figure B has coordinates of (40, 30, 5) cm, and the Z-axis represents height). Then, the distance between the two sandplay figures is calculated using the 3D distance formula, resulting in a distance of 20 cm. The sequence of lines connecting the center points is generated according to the order in which the sandplay figures were placed, such as sandplay figure A-sandplay figure B-sandplay figure C, forming a broken line sequence. For example, a pre-set database of symbolic meanings of sandplay figures can be stored in a MongoDB database, constructed based on Jungian psychological theory. For instance, human-shaped sandplay figures correspond to the symbolic meaning of self-identity, with an initial symbolic weight of 0.6; animal-shaped sandplay figures (such as bears) correspond to the symbolic meaning of security, with an initial symbolic weight of 0.5; and a distance of less than 10 cm between sandplay figures corresponds to the symbolic meaning of close relationship, with an initial symbolic weight of 0.7. When performing semantic analysis on text data (speech text converted in S1), a BERT pre-trained model is used to extract subjective interpretations such as "This little house represents my home, and I want to be closer to it." If the user emphasizes the importance of "home," the initial symbolic weight of the corresponding building-type sand figure (little house) is adjusted from 0.6 to 0.8, generating a 256-dimensional weighted symbolic feature vector (each dimension corresponds to the weight of a type of sand figure or spatial relationship, with a value range of 0-1). The pre-trained Long Short-Term Memory (LSTM) network model uses a training dataset containing real sandplay cases (including images of sandplay object layouts and text descriptions) and corresponding professional psychological assessment results (annotated by three Level 2 psychological counselors). The model's input layer has a dimension of 256, the hidden layer has 128 nodes, and the output layer has a dimension of 3. After inputting the weighted symbolic feature vector into the model, it outputs quantitative values ​​for emotional need intensity (e.g., 65 points), psychological stress level (e.g., 55 points), and subconscious conflict index (e.g., 40 points). All quantitative values ​​range from 0 to 100.

[0036] In existing technologies, sand table object recognition relies on manual classification, spatial relationships are judged solely by visual inspection, symbolic weights lack dynamic adjustment mechanisms, and psychological feature data is mostly qualitative description. This embodiment achieves automatic object recognition and spatial distance calculation through a depth camera and image segmentation algorithm, adjusts symbolic weights by combining semantic analysis, and then outputs quantitative psychological feature data through a pre-trained LSTM model, significantly improving the efficiency and accuracy of sand table analysis and reducing subjective errors caused by manual intervention.

[0037] In another embodiment, based on the temporal data of sand table placement recorded by a depth camera during user sand table construction, a spatiotemporal input sequence is constructed from weighted symbolic feature vectors in chronological order. Each time step in the sequence includes the type of sand table placement, its symbolic weight, and its three-dimensional coordinates within the sand table. The sequence, ordered by placement time, is input into a first long short-term memory network model, which specifically learns the dynamic evolution pattern of user construction behavior and outputs a temporal feature vector. The final sand table layout is then subjected to cluster analysis to identify core and peripheral regions, generating a spatial structure sequence from core to edge. This spatial structure sequence is input into a second long short-term memory network model, which learns the static geometry and symbolic structure of the overall sand table layout and outputs a spatial feature vector. The temporal feature vector and the spatial feature vector are concatenated and input into a fully connected neural network, which outputs quantitative values ​​for emotional need intensity, psychological stress level, and subconscious conflict indicators.

[0038] For example, based on the temporal data of sand table placement recorded by a depth camera (a timestamp is recorded for each sand table placed, with millisecond precision), the weighted symbolic feature vector (256 dimensions) is used to construct a spatiotemporal input sequence in chronological order. The sequence length is determined by the number of sand tables (e.g., if 8 sand tables are placed, the sequence length is 8). Each time step data includes: sand table type, symbolic weight (e.g., 0.7), and three-dimensional coordinates (e.g., (15, 25, 6) cm). The first Long Short-Term Memory (LSTM) network model can be built based on the TensorFlow framework. The input sequence dimension is (8, 256) (sequence length 8, feature dimension 256 for each time step), with 2 hidden layers, 128 nodes per layer, and a dropout probability of 0.2. After pre-training with 10,000 sets of sand table temporal data, it specifically learns the dynamic evolution pattern of user building behavior (e.g., the behavior pattern of placing building-type sand tables first and then placing human-type sand tables), and outputs a 128-dimensional temporal feature vector (value range 0-1). When performing cluster analysis on the final sand table layout, the K-means clustering algorithm was used (K value set to 2, corresponding to the core region and the edge region). The two-dimensional coordinates (X, Y axes) of all sand table objects were used as clustering data, with a cluster radius threshold of 15 cm. Sand table objects closer to the cluster center (<15 cm) constituted the core region, and those farther away (≥15 cm) constituted the edge region. The weighted symbolic feature vectors of the sand table objects within each region were arranged in order from the core to the edge, generating a spatial structure sequence of length 8 (consistent with the number of sand table objects). The second LSTM model had the same structure as the first LSTM model (2 hidden layers, 128 nodes per layer). After pre-training with 10,000 sets of sand table spatial layout data, it learned the static geometry and symbolic structure of the sand table (e.g., the layout structure of the core region, which mainly consists of human-shaped sand table objects), outputting a 128-dimensional spatial feature vector (value range 0-1). The temporal feature vector (128-dimensional) and the spatial feature vector (128-dimensional) are concatenated into a 256-dimensional fusion vector, which is then input into a fully connected neural network (256-dimensional input layer, 128-dimensional hidden layer, 3-dimensional output layer, with activation functions ReLU, ReLU, and Sigmoid, respectively). After training on 3000 sets of labeled data, the network outputs quantitative values ​​of emotional need intensity (e.g., 68 points), psychological stress level (e.g., 52 points), and subconscious conflict index (e.g., 38 points). The quantitative values ​​are then mapped to the 0-100 score range after being output through the Sigmoid function.

[0039] In existing technologies, sand table analysis only focuses on the static features of the final layout, ignoring the dynamic information of the placement sequence, resulting in an incomplete interpretation of the user's psychological state. This embodiment uses two LSTM models to learn the temporal dynamic pattern and spatial static structure respectively, and then fuses the results through a fully connected network. This takes into account both the behavioral evolution of the user during the construction process and the spatial features of the final layout, making the evaluation of the sand table's psychological feature data more comprehensive and more in line with the user's psychological formation process.

[0040] In another embodiment, a multi-source data vector is constructed, including House-Tree-Person (HTP) psychological characteristic data, sandplay psychological characteristic data, physiological data collected by sensors during mindfulness training, and behavioral data recorded through a human-computer interaction interface. The multi-source data vector is input into a pre-trained decision tree model, which is generated based on the user's historical assessment data and expert annotation results. This decision tree model dynamically allocates a first weight value for the HTP psychological characteristic data, a second weight value for the sandplay psychological characteristic data, a third weight value for the physiological data, and a fourth weight value for the behavioral data. Based on the first, second, third, and fourth weight values ​​output by the decision tree model, a weighted summation is performed on the multi-source data vector to generate the comprehensive psychological assessment result.

[0041] For example, when constructing the multi-source data vector, the House-Tree-Person psychological characteristic data is 3-dimensional (family relationship intimacy, self-awareness and self-confidence, and emotional state calmness tendency, all from 0 to 100 points), and the sandplay psychological characteristic data is 3-dimensional (emotional need intensity, psychological stress level, and subconscious conflict index, all from 0 to 100 points). Physiological data is collected through sensors: a wristband (worn on the user's left wrist) can be selected to collect heart rate (60-100 beats / minute), blood oxygen saturation (95%-100%), and skin conductance (0.1-10μS), totaling 3 dimensions. Behavioral data is recorded through the human-computer interaction interface (the software interface of the Lenovo ThinkPad X1 Carbon laptop): training duration (10-60 minutes), operation frequency (1-10 times / minute), and task completion rate (0%-100%), totaling 3 dimensions. The total dimensions of the multi-source data vector are 3+3+3+3=12 dimensions, and all data are normalized to the 0-1 range. The pre-trained decision tree model can be an XGBoost model. The training dataset includes users' historical evaluation data (multi-source data vectors) and the weight allocation results labeled by experts (determined jointly by 5 psychology experts). After optimization by 10-fold cross-validation, the model is used to dynamically allocate weights: for example, when a user's house-tree-person psychological characteristic data standard deviation <10 (high data stability) and physiological data fluctuation <5% (stable signal), the first weight value is 0.3, the second weight value is 0.25, the third weight value is 0.25, and the fourth weight value is 0.2 (the weight sum is 1). When calculating the weighted sum, if the house-tree-person psychological characteristic data vector is (75, 80, 65) (normalized to (0.75, 0.8, 0.65)), multiplying it by the first weight of 0.3 yields (0.225, 0.24, 0.195); and the sandplay psychological characteristic data vector is (65, 55, 40) (normalized to (0.65, 0.55, 0.4)), multiplying it by the second weight of 0.25 yields (0.1625, 0.1375, 0.1); The rational data vector (70, 98, 5) (normalized to (0.7, 0.98, 0.05)) is multiplied by the third weight of 0.25 to obtain (0.175, 0.245, 0.0125); the behavioral data vector (40, 5, 80) (normalized to (0.4, 0.05, 0.8)) is multiplied by the fourth weight of 0.2 to obtain (0.08, 0.01, 0.16); the summation of the corresponding dimensions of the four types of data yields the fusion vector (0.6425, 0.6325, 0.4675), which is then mapped to a score of 0-100 to obtain the comprehensive psychological assessment result (64.25, 63.25, 46.75), rounded to the nearest integer as (64, 63, 47). During the initial assessment, since the user had not yet started mindfulness training, the physiological / behavioral data collected during the mindfulness training were replaced with the user's baseline physiological data in a seated state during the initial assessment phase and the initial behavioral data during the assessment process.Initial behavioral data is automatically generated by recording operation logs through the human-computer interaction module. Both types of data serve as a benchmark for subsequent static training data. After the user conducts the first training session, real-time physiological / behavioral data from the training process will be added and integrated.

[0042] In existing technologies, multi-source psychological data often uses fixed-weight fusion, failing to consider data stability and individual user differences, resulting in low accuracy of the comprehensive evaluation results. This embodiment dynamically allocates weights through a pre-trained decision tree model, adjusting the weight ratio of each data source based on data quality and historical annotation results, and then generates a comprehensive evaluation result through weighted summation, making the evaluation results more accurate and better reflecting the user's multi-dimensional psychological state.

[0043] In another embodiment, image data confidence is calculated based on the clarity and integrity of the house-tree-person image and the sand table scene image; text data confidence is calculated based on the emotional word density and syntactic complexity of the speech description text data; physiological data confidence is calculated based on the signal stability of physiological data within the acquisition time window; and behavioral data confidence is calculated based on the event record integrity of behavioral data. The consistency of the first evaluation index regarding emotional state in the house-tree-person psychological feature data and the sand table psychological feature data is compared; the consistency of the second evaluation index representing emotional arousal in the physiological data is compared with the first evaluation index; and the consistency comparison results are verified based on abnormal interaction events recorded in the behavioral data. In response to the consistency comparison results, the first weight value is dynamically adjusted to the fourth weight value: when the evaluation indicators of two or more data sources are consistent, the weight of the corresponding weight value of the data source is increased; when the evaluation indicators of a certain data source significantly conflict with other data sources, and its data quality confidence is lower than a preset threshold, the weight of the corresponding weight value of that data source is decreased.

[0044] For example, when calculating the confidence score of image data, the sharpness of the house-tree-person image and the sand table scene image is calculated using the peak signal-to-noise ratio (PSNR). A PSNR value ≥ 35dB is considered sharp, and the sharpness score is calculated as (PSNR-20) / 15 (value range 0-1). Integrity is determined by the percentage of unobstructed areas (house-tree-person / sand table) in the image. An unobstructed percentage ≥ 90% earns 1 point; otherwise, the percentage is taken as (0-1). The image data confidence score is the average of the sharpness score and the integrity score (value range 0-1). When calculating the confidence score of text data, the sentiment word density is the ratio of the number of sentiment words (such as "happy" and "anxious") in the text to the total number of words. A ratio ≥ 15% earns 1 point; otherwise, the ratio is taken as (0-1). Syntactic complexity is calculated using the average sentence length (number of words). An average length ≥ 8 words earns 1 point; otherwise, the average length is taken as (0-1). The text data confidence score is the average of the two (value range 0-1). When calculating the confidence level of physiological data, the acquisition time window is set to 5 minutes. Signal stability is determined by the coefficient of variation (standard deviation / mean) of heart rate data. A coefficient of variation ≤ 10% earns 1 point; otherwise, it is calculated as (10% - coefficient of variation) / 10% (0-1). The confidence level of physiological data is the signal stability score (value 0-1). When calculating the confidence level of behavioral data, event recording completeness is the ratio of the number of recorded events (such as training start / pause, operation trigger) to the number of events that should be recorded. A ratio ≥ 95% earns 1 point; otherwise, it is calculated as the ratio (0-1). All confidence level preset thresholds are set to 0.6. The first assessment indicator is the "quantitative value of calm tendency in emotional state" in the House-Tree-Person psychological characteristic data and the "quantitative value of psychological stress level" in the sandplay psychological characteristic data (the lower the stress level, the higher the calm tendency, which is considered a positive correlation). When the difference between the two is ≤10 points, they are considered consistent. The second assessment indicator is the "heart rate variability (HRV)" in the physiological data (the higher the HRV, the lower the emotional arousal, which is positively correlated with calm tendency). When the difference between the emotional state corresponding to the HRV indicator and the first assessment indicator is ≤15%, they are considered consistent. Abnormal interaction events in the behavioral data include "frequent pauses during training (≥3 times / 30 minutes)" and "operation response delay (≥5 seconds / time)". If there are abnormal events, the consistency comparison results need to be adjusted according to the cause of the event. When dynamically adjusting weights, if the first evaluation index of the house-tree-person data and the sand table data is consistent (difference of 5 points), and both have a confidence level of 0.8 (higher than the threshold of 0.6), then the first weight value (house-tree-person) increases from 0.3 to 0.35, and the second weight value (sand table) increases from 0.25 to 0.3. If the confidence level of the physiological data is 0.5 (lower than the threshold of 0.6), and its second evaluation index conflicts with the first evaluation index (difference of 20%), then the third weight value (physiological) decreases from 0.25 to 0.15, and the reduced weight is allocated to the data source with high confidence and consistency.

[0045] In existing technologies, multi-source data fusion often fails to calculate data confidence levels and directly uses fixed weights, leading to low-quality data affecting the evaluation results. This embodiment calculates the confidence levels of various data from multiple dimensions and dynamically adjusts the weights based on the consistency of evaluation indicators, reducing interference from low-quality data and making the weighted fusion comprehensive evaluation result more reliable and more in line with the user's true psychological state.

[0046] In another embodiment, the method further includes: periodically aggregating anonymized multi-source data vectors from multiple users and corresponding expert-confirmed evaluation results on a cloud platform to form an incremental training dataset; incrementally training the decision tree model using the incremental training dataset to generate an updated decision tree model; performing personalized fine-tuning on the updated decision tree model based on the historical multi-source data vector sequence of a specific user to generate a personalized decision tree model for the corresponding user; the personalized fine-tuning process employs temporal cross-validation to optimize the model's adaptability to the evolution trend of the user's psychological state; and deploying the personalized decision tree model or the updated decision tree model for subsequent dynamic allocation of weight values.

[0047] For example, the cloud platform can use Alibaba Cloud servers, with a periodic data aggregation period of one month. Each aggregation includes multi-source data vectors (12-dimensional, normalized to the 0-1 range) from 1000+ anonymous users and assessment results (comprehensive psychological assessment scores from 0-100) jointly confirmed by three Level 2 psychological counselors, forming an incremental training dataset (1000+ new samples added each time). When incrementally training the decision tree model (XGBoost model) using the incremental training dataset, a warm-start mode is adopted, loading the model parameters from the previous training round as initial parameters, updating the model weights with new samples, setting the training batch size to 32, and the learning rate to 0.01 to avoid resource waste caused by retraining. After training is completed, an updated decision tree model is generated. When using a specific user's historical multi-source data vector sequence, the sequence length is set to the user's previous three training sessions of multi-source data (e.g., three sets of 12-dimensional vectors spaced one month apart). The updated decision tree model is then fine-tuned: the user's historical data is divided into training and validation sets in a 7:3 ratio, 60% of the parameters at the bottom layer of the model are frozen, and only the top 40% of the parameters are trained. The fine-tuning batch size is set to 16, and the learning rate is set to 0.005. The fine-tuning process uses temporal cross-validation. The historical data is divided into three consecutive folded sets in chronological order (e.g., the first set is fold 1, the second set is fold 2, and the third set is fold 3). Two of these folded sets are used as the training set, and one is used as the validation set. The weight allocation error (mean squared error, MSE) for each validation is calculated. The model parameters with the smallest average of the three errors are taken as the final personalized model parameters, generating a personalized decision tree model for the corresponding user. When deploying the model, if the user's historical data volume is ≥3 sets (meeting the fine-tuning requirements), a personalized decision tree model is deployed for subsequent weight allocation. For example, when the user is trained again, the model can appropriately increase the weight of behavioral data (training duration, completion rate) based on the past trend of "the tendency of the user's emotional state to be calm increases with training". If the user's historical data volume is <3 sets, an updated general decision tree model is deployed. Once the data volume meets the standard, the personalized model is generated and switched. All model deployments are encapsulated in Docker containers to ensure environmental consistency. After deployment, the model is connected to the local data processing module through an API interface to achieve real-time allocation of weight values.

[0048] This embodiment updates the model through incremental training in the cloud and performs personalized fine-tuning by combining historical user data, enabling the model to continuously adapt to new data features and conform to the evolution of the psychological state of specific users. This results in more accurate subsequent weight allocation and further improves the reliability of the comprehensive evaluation results.

[0049] The embodiments of this application also provide an AI-based method and apparatus for static intelligent training and evaluation, including: a data acquisition module, which is equipped with a high-speed scanner and a microphone, for acquiring user-drawn house-tree-person images, user-built sand table scene images, and user voice descriptions of the house-tree-person images and sand table scene images; a human-computer interaction module, for providing a data acquisition entry point and displaying a dynamic shooting guidance interface; a speech-to-text module, connected to the data acquisition module, for converting the voice descriptions acquired by the microphone into text data; and a house-tree-person feature analysis module, connected to the image preprocessing module and the speech-to-text module, for using a deep learning-based object detection algorithm to extract features from the standardized house-tree-person images, obtain house-tree-person image features, and combine them with the corresponding text data to generate house-tree-person psychological feature data; The sand table feature analysis module, connected to the image preprocessing module and the speech-to-text module, uses an image segmentation algorithm based on 3D vision technology to identify sand table objects and extract spatial relationships from standardized sand table scene images, obtaining sand table scene features. Combined with corresponding text data, it generates sand table psychological feature data. The multi-source data fusion evaluation module, connected to the house-tree-person feature analysis module and the sand table feature analysis module, incorporates physiological and behavioral data collected during mindfulness training. It performs weighted fusion of the house-tree-person psychological feature data, sand table psychological feature data, physiological data, and behavioral data to generate a comprehensive psychological evaluation result. The training course matching module, connected to the multi-source data fusion evaluation module, matches corresponding mindfulness training courses from the training content database based on the comprehensive psychological evaluation result.

[0050] For example, the document scanner for the data acquisition module can be the Chengzhe Technology Elf EA3 MAX (28MP resolution, USB 3.0 interface), and the microphone can be the Logitech Blue Yeti Nano (cardioid pickup pattern, USB-C interface). Both are connected to the main unit via USB. The document scanner is fixed on an adjustable stand above the main unit (15-20cm high, facing the desktop drawing / sand table area), and the microphone is placed on the desktop next to the main unit (30-50cm away from the user's mouth). It is used to acquire A4 paper house-tree-person images, 60×40×8cm standard sand table scene images, and the user's voice description of the images (sampling rate 16000Hz). The human-computer interaction module is integrated into the main unit's 27-inch display (2560×1440 resolution), providing a data acquisition entry (start acquisition button) through a customized software interface. Clicking the button displays a dynamic shooting guide interface (black rectangular alignment box, size 21×29.7cm, with white 12-point font prompt text below). The speech-to-text module uses Baidu AI Cloud's speech recognition API and connects to the data acquisition module via the network to convert PCM format speech data collected by the microphone into UTF-8 encoded text data, achieving a recognition accuracy of ≥95%. The house-tree-person feature analysis module is a software module running on the host processor. It connects internally with the image preprocessing module (including cropping and rotation functions, resulting in a 1920×1080 image resolution) and the speech-to-text module. It uses the YOLOv8 object detection algorithm to extract features such as house outlines (error < 2 pixels) and tree branch morphology, and combines this with text data (keyword extraction uses the TF-IDF algorithm) to generate 3D house-tree-person psychological feature data (0-100 points). The sand table feature analysis module is also a software module, connected to the image preprocessing module and the speech-to-text module. It uses the Mask R-CNN image segmentation algorithm to identify sand table types (human figures, animals, etc.), calculates the 3D spatial distance of the sand table (error < 1 cm), and combines this with text semantic analysis (BERT model) to generate 3D sand table psychological feature data (0-100 points). The multi-source data fusion evaluation module connects to the first two feature analysis modules via a data bus. It receives physiological data (heart rate, blood oxygen) collected by the wristband (connected to the host via Bluetooth) and behavioral data (training duration, operation frequency) recorded by the human-computer interaction module. It uses an XGBoost decision tree model to assign weights (weights sum to 1) and generates a weighted summation to produce a 3-dimensional comprehensive psychological evaluation result (0-100 points). The training course matching module connects to the multi-source data fusion evaluation module via a data interface. It accesses the training content database (containing 200+ courses, each corresponding to an evaluation result range) stored in a local MySQL database (version 8.0) on the host. It matches corresponding courses using SQL queries (e.g., a comprehensive score of 60-70 matches a basic stress reduction course) and displays the course name and description on the human-computer interaction interface.For example, a standardized house-tree-person image refers to an image taken on A4 paper using a high-speed scanner on a dynamically guided interface (including a visual alignment frame), and after preprocessing such as cropping and rotation, with a uniform resolution of 1920×1080 pixels, meeting the standards for sharpness (PSNR≥35dB) and target integrity (unobstructed area ≥90%). It must clearly present the core elements of houses, trees, and people. Standardized sand table scene images are based on a standard 60×40×8 cm sand table, captured by a depth camera at a distance of 80-100 cm directly above, simultaneously acquiring an RGB image (1280×720 pixels) and a depth map. After segmentation processing, the images are complete in identifying the sand table objects, accurately calculate the three-dimensional spatial relationships, and have no significant shooting interference.

[0051] In existing technologies, observational training devices are mostly single-function modules (such as those that only collect data or only provide courses). Each module operates independently and requires manual connection, which is cumbersome and data is not shared. This embodiment integrates multiple modules and establishes internal data connections to achieve full automation of the entire process from data collection, feature analysis, evaluation fusion to course matching. No manual intervention is required, making operation more convenient. Furthermore, data flows in a closed loop within the device, avoiding data fragmentation caused by switching between multiple devices, thus improving user experience and evaluation training efficiency.

[0052] The following is an example of a practical application.

[0053] (a) Target Groups The program selected drug addicts within the Donghai Drug Rehabilitation Center as the subjects of the program. All subjects were free from severe mental illness or physical disabilities, possessed basic image capture, speech expression, and touch operation abilities, and voluntarily participated in the program.

[0054] (II) Equipment Deployment Hardware configuration: One AI+Jingguan intelligent assessment and training all-in-one machine will be deployed in the "Psychological Health Service Center" of Donghai Drug Rehabilitation Center. The hardware parameters of each all-in-one machine are as follows: Core control module: Equipped with an ARM Cortex-A75 quad-core processor, supporting Wi-Fi 6 and Bluetooth 5.0 (ensuring stable data transmission and complying with the network security standards of drug rehabilitation centers); High-speed document scanner module: 16-megapixel high-definition lens, shooting range supports A4 size, and comes with a fill light (suitable for indoor low-light scenarios in drug rehabilitation centers). Data acquisition module: 1080P HD camera (supports facial micro-expression capture), noise-canceling microphone (pickup distance 0.5-3 meters); Interaction module: 27-inch capacitive touch screen (1920×1080 resolution), built-in 2.1 channel noise-canceling speakers (volume can be limited to 40 decibels to avoid interfering with other areas).

[0055] Software and algorithm configuration: System software: Equipped with a customized Windows operating system; Supporting environment: Each all-in-one machine is provided with an independent "training cubicle" (approximately 4 square meters in area). 2 The cubicles are equipped with comfortable seats and blackout curtains (to reduce external interference) and "Operation Instruction Posters" (including pictures and text instructions on image shooting posture, voice response precautions, etc.) to meet the self-operation needs of drug addicts.

[0056] (iii) Personnel training Six "psychological health counselors" from the Donghai Drug Rehabilitation Center received a one-day specialized training course, which included: operation of the all-in-one machine hardware (such as camera angle adjustment and sensor calibration), use of software functions (such as exporting training data and handling abnormal situations), and guidance skills for drug addicts (such as how to encourage drug addicts to express their subjective feelings truthfully). Only those who pass the assessment after the training (operational accuracy rate ≥ 95%) are allowed to participate in the implementation process.

[0057] II. Specific Implementation Steps The training frequency was set at 60 minutes twice a week for a total of 6 months. The process was divided into three phases: assessment, training, and follow-up. The steps for each phase are as follows: (I) First stage: Psychological assessment (Week 1, 30 minutes each time) Image Acquisition: After entering the training cubicle, the drug addicts follow the prompts on the all-in-one machine's touchscreen to place their self-drawn "House-Tree-Person" drawing (paper and brushes are provided by the drug rehabilitation center one day in advance) in the high-speed scanner's shooting area, and click the "Image Acquisition" button. The high-speed scanner will automatically complete the shooting (if the image is blurry, the system will prompt "retake" until the image clarity meets the standard); for sand table scenes, the counselor will assist in setting up a simple sand table (including basic sand figures such as people, houses, and trees). After the drug addicts arrange the sand table themselves, the image will be taken through the camera (the system will automatically stitch the images together).

[0058] Subjective feedback collection: After image acquisition is completed, the all-in-one machine automatically asks questions through the voice interaction module. The questions are designed around "image-related psychological feelings". For the house-tree-person drawing, the questions include "Does your family live in the house in your drawing?" "What does the growth state of the trees represent to you?" For the sand table, the questions include "What is the item you are most concerned about in the sand table?" "What past experience does this scene remind you of?" After the drug addict answers the voice, the system transcribes the voice into text in real time and stores it locally.

[0059] Physiological data collection: During and after the subjective feedback process, drug addicts wear a wristband device (wristband type, no need to stick) and remain seated for 5 minutes. The integrated machine automatically collects physiological data such as heart rate variability (HRV) and respiratory rate, and uploads them to the AI ​​algorithm module simultaneously.

[0060] AI assessment report generation: After data collection is completed, the system outputs the user psychological assessment module.

[0061] (II) Second stage: Personalized training (weeks 2-24, 60 minutes each time) Training program matching: Before each training session, the all-in-one machine automatically matches training content based on the previous evaluation report, for example: For drug addicts with "weak belief in rehabilitation," "mindfulness belief strengthening training" is provided. For drug addicts with "significant anxiety tendencies", "breathing relaxation training" is recommended; For drug addicts with "low family support", we recommend "family mindfulness meditation".

[0062] Training data recording: After each training session, the all-in-one machine automatically records the "training completion rate" (e.g., "breathing training completed 90%)", "real-time physiological changes" (e.g., "heart rate dropped from 95 beats / minute to 75 beats / minute after training"), and "user subjective evaluation" ("training satisfaction rating" is completed via touch screen, 1-5 points). All data is synchronized to the drug rehabilitation center's management server.

[0063] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and embodiments shown and described herein.

Claims

1. An AI-based method for training and evaluating observable intelligence, characterized in that, include: S1: Collect the house-tree-person image drawn by the user and the sand table scene image built by the user through a high-speed document scanner, and collect the user's voice description of the house-tree-person image and the sand table scene image through a microphone, and convert the voice description into text data; S2: Use a deep learning-based object detection algorithm to extract features from the house-tree-person image, obtain the house-tree-person image features, and combine them with the corresponding text data to generate house-tree-person psychological feature data; S3: Use an image segmentation algorithm based on 3D vision technology to identify sand table objects and extract spatial relationships from the sand table scene image to obtain sand table scene features, and combine them with corresponding text data to generate sand table psychological feature data; S4: Weighted fusion of House-Tree-Person psychological characteristic data, sandplay psychological characteristic data, physiological data and behavioral data collected during mindfulness training to generate a comprehensive psychological assessment result; S5: Match the corresponding mindfulness training course from the training content database based on the comprehensive psychological assessment results.

2. The AI-based observable intelligence training and evaluation method as described in claim 1, characterized in that, In S1, the human-computer interaction interface provides a dedicated data acquisition entry point; In response to the user's trigger operation, the document scanner is invoked, and a dynamic shooting guidance interface is displayed on the human-computer interaction interface. The shooting guidance interface includes a visual alignment box and text prompts to guide the user to place the drawn house-tree-person image or the sand table scene image built by the user in the optimal shooting area. Upon receiving an image capture command, the system controls the document scanner to capture images and obtain the original images. Provide an image preprocessing toolset, which includes at least image cropping, rotation, brightness adjustment, and contrast adjustment functions, and receive one or more interactive preprocessing operations performed by the user on the original image through the image preprocessing toolset.

3. The AI-based observable intelligence training and evaluation method as described in claim 1, characterized in that, In S2, an object detection algorithm is used to identify and extract the outlines of houses, trees, people, the number of doors and windows of houses, the shape of tree branches and trunks, and the facial expression features of people from house-tree-person images. At the same time, the color distribution histogram of the house-tree-person image is extracted as color features, and the relative distance between the center point of the person outline and the center point of the house outline is calculated as composition features. The extracted features are input into a pre-set psychological feature mapping table for matching. The psychological feature mapping table stores the correspondence between features and psychological feature entries, and outputs the initial psychological feature vector. Keyword extraction and sentiment analysis are performed on text data to generate user subjective feedback vectors; The initial psychological feature vector and the user's subjective feedback vector are weighted and fused to generate the final house-tree-person psychological feature data. The house-tree-person psychological feature data includes quantitative values ​​of family relationship intimacy, self-awareness and self-confidence, and emotional state calmness tendency.

4. The AI-based observable intelligence training and evaluation method as described in claim 3, characterized in that, An initial psychological feature vector and a user subjective feedback vector are input into an attention mechanism network, which is trained to output dynamic weights corresponding to the consistency between objective image features and user subjective descriptions. When the sentiment polarity value in the user's subjective feedback vector is highly consistent with the predicted sentiment tendency in the initial psychological feature vector, a higher first weight is assigned to the user's subjective feedback vector, and a lower second weight is assigned to the initial psychological feature vector; when there is a significant inconsistency between the two, the first weight is reduced and the second weight is increased. Based on a pre-built psychological rule base, feature conflicts are detected between the initial psychological feature vector and the user's subjective feedback vector; If a conflict is detected, the conflict resolution procedure is initiated: similar conflict cases in historical data and their final evaluation results confirmed by experts are retrieved, the bias weight of the current conflict is calculated through similarity matching, and the dynamic weight is adjusted accordingly. Using modified or unmodified dynamic weights, the initial psychological feature vector and the user's subjective feedback vector are weighted and summed, and the summation result is mapped onto a predefined psychological scale to generate House-Tree-Person psychological feature data that includes quantitative values ​​of family relationship intimacy, self-awareness and self-confidence, and emotional state calmness tendency.

5. The AI-based observable intelligence training and evaluation method as described in claim 1, characterized in that, In S3, a depth image of the sand table scene is acquired by a depth camera based on the structured light principle. Combined with the RGB image, an image segmentation algorithm is used to identify and classify the sand figures, which include human figures, animal figures, and architectural figures. The three-dimensional spatial distance between the sand figures and the sequence of lines connecting their center points are calculated based on the depth image. Each identified sandplay object type and its three-dimensional spatial relationship are matched with a pre-set sandplay object symbolic meaning database, and an initial symbolic weight is assigned to each sandplay object and its spatial relationship. Semantic analysis is performed on the text data to extract users' subjective interpretations of specific sandplay objects and their placement. Based on these subjective interpretations, the initial symbolic weights of the corresponding sandplay objects are dynamically adjusted to generate a weighted symbolic feature vector. The weighted symbolic feature vector is input into a pre-trained long short-term memory network model. The long short-term memory network model is trained using a dataset of real sandplay cases and corresponding professional psychological assessment results. It outputs sandplay psychological feature data, which includes quantitative values ​​of emotional need intensity, psychological stress level, and subconscious conflict index.

6. The AI-based observable intelligence training and evaluation method as described in claim 5, characterized in that, Based on the temporal data of sand table placement recorded by the depth camera when the user builds the sand table, the weighted symbolic feature vector is constructed into a spatiotemporal input sequence according to the time order. Each time step in the sequence contains the type of sand table placement, symbolic weight, and its three-dimensional coordinates in the sand table. The sequence sorted by placement time is input into the first long short-term memory network model, which specifically learns the dynamic evolution pattern of user building behavior and outputs a temporal feature vector. Cluster analysis is performed on the final sand table layout to identify the core and peripheral areas. A spatial structure sequence is generated in order from the core to the edge. The spatial structure sequence is input into the second long short-term memory network model. The second long short-term memory network model learns the static geometric and symbolic structure of the overall sand table layout and outputs a spatial feature vector. The temporal feature vector and spatial feature vector are concatenated and then input into a fully connected neural network, which outputs quantitative values ​​of emotional need intensity, psychological stress level, and subconscious conflict index.

7. The AI-based observable intelligence training and evaluation method as described in claim 1, characterized in that, Construct a multi-source data vector, which includes house-tree-person psychological characteristic data, sandplay psychological characteristic data, physiological data collected by sensors during mindfulness training, and behavioral data recorded through the human-computer interaction interface; Multi-source data vectors are input into a pre-trained decision tree model, which is trained based on users’ historical evaluation data and expert annotation results. The decision tree model is used to dynamically allocate the first weight value of the house-tree-person psychological characteristic data, the second weight value of the sand table psychological characteristic data, the third weight value of the physiological data, and the fourth weight value of the behavioral data. Based on the first, second, third, and fourth weight values ​​output by the decision tree model, the multi-source data vectors are weighted and summed to generate the comprehensive psychological assessment result.

8. The AI-based observable intelligence training and evaluation method as described in claim 7, characterized in that, Image data confidence is calculated based on the clarity and integrity of house-tree-person images and sand table scene images; text data confidence is calculated based on the emotional word density and syntactic complexity of speech description text data; physiological data confidence is calculated based on the signal stability of physiological data within the acquisition time window; and behavioral data confidence is calculated based on the event record integrity of behavioral data. The consistency comparison of the first assessment index of emotional state in the house-tree-person psychological characteristic data and the sand table psychological characteristic data was performed. The consistency comparison of the second assessment index of emotional arousal in the physiological data was performed with the first assessment index. The consistency comparison results were verified based on the abnormal interaction events recorded in the behavioral data. In response to the consistency comparison results, the first weight value is dynamically adjusted to the fourth weight value: when the evaluation indicators of two or more data sources are consistent, the weight of the corresponding weight value of the data source is increased; when the evaluation indicators of a certain data source conflict significantly with those of other data sources and its data quality confidence is lower than a preset threshold, the weight of the corresponding weight value of that data source is decreased.

9. The AI-based observable intelligence training and evaluation method as described in claim 7, characterized in that, Also includes: The cloud platform periodically aggregates anonymized multi-source data vectors from multiple users and corresponding expert-confirmed evaluation results to form an incremental training dataset. The decision tree model is then incrementally trained using the incremental training dataset to generate an updated decision tree model. Based on the historical multi-source data vector sequence of a specific user, the updated decision tree model is fine-tuned to generate a personalized decision tree model for the corresponding user. The personalization fine-tuning process uses time-series cross-validation to optimize the model's adaptability to the evolution trend of the user's psychological state. Deploy personalized or updated decision tree models for subsequent dynamic allocation of weights.

10. An AI-based observable intelligent training and evaluation device, characterized in that, include: The data acquisition module is equipped with a document scanner and a microphone to acquire images of the house-tree-person (HTP) drawn by the user, images of the sand table scene built by the user, and the user's voice description of the HTP images and sand table scene images. The human-computer interaction module provides a data acquisition entry point and displays a dynamic shooting guidance interface; The speech-to-text module, connected to the data acquisition module, is used to convert the speech descriptions captured by the microphone into text data; The House-Tree-Person Feature Analysis Module, connected to the Image Preprocessing Module and the Speech-to-Text Module, is used to extract features from standardized House-Tree-Person images using a deep learning-based object detection algorithm, obtain House-Tree-Person image features, and combine them with corresponding text data to generate House-Tree-Person psychological feature data. The sand table feature analysis module, connected to the image preprocessing module and the speech-to-text module, is used to perform sand table object recognition and spatial relationship extraction on standardized sand table scene images using image segmentation algorithms based on 3D vision technology, to obtain sand table scene features, and combine them with corresponding text data to generate sand table psychological feature data. The multi-source data fusion and evaluation module is connected to the house-tree-person feature analysis module and the sand table feature analysis module, and incorporates physiological and behavioral data collected during the mindfulness training process. It is used to weight and fuse the house-tree-person psychological feature data, sand table psychological feature data, physiological data, and behavioral data to generate a comprehensive psychological evaluation result. The training course matching module is connected to the multi-source data fusion evaluation module and is used to match the corresponding mindfulness training course from the training content database based on the comprehensive psychological evaluation results.

Images