Psychological interaction method and system based on sand table dynamics and healing cognitive architecture

Abstract

The present application relates to the cross field of artificial intelligence and mental health, and discloses a psychological interaction method and system based on sand table dynamics and healing cognitive architecture, which solves the technical problems that the single text mode of the existing psychological conversation system is difficult to break through speech defense, diagnosis and treatment is fragmented, and lacks long-term strategic planning. The present application constructs a multi-modal healing cognitive architecture containing six modules such as sand table dynamics analysis, constructs sand table visual evaluation dataset and healing conversation dataset through double independent pipelines, adopts double model decoupling training to obtain sand table visual evaluation model and healing cognitive conversation model, and then performs cascade joint reasoning, injects the structured psychological representation output by the sand table model into the conversation model as deep context, and guides the generation of healing replies adapted to the subconscious state of the user. The present application realizes dual-dimension insight of subconscious and consciousness, achieves diagnosis and treatment integration closed loop, establishes long-term strategic healing ability, and greatly improves the insight and intervention accuracy of automatic psychological service.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and mental health, specifically to the integrated application of multimodal large language models, computer vision, psychological sandplay therapy and human-like cognitive dialogue systems, and in particular to a psychological interaction method and system based on sandplay dynamics and a healing cognitive architecture. Background Technology

[0002] With the increasing pace of modern life and competitive pressures, mental health issues are receiving more and more attention. The "Report on the Development of National Mental Health in China (2023-2024)" shows that 6.5% of adults in my country experience three or more persistent depressive symptoms, leading to an explosive growth in public demand for mental health services. However, traditional psychological counseling services suffer from a shortage of professional counselors, uneven geographical distribution, and high service costs, creating a significant gap between the current demand and the actual needs. Therefore, utilizing artificial intelligence technology to build automated and highly available psychological healing systems has become an important direction for the industry's development.

[0003] Multimodal large language models, with their powerful multimodal information understanding and natural language dialogue capabilities, provide technological possibilities for the development of intelligent psychological healing systems. Currently, their application in the field of psychological counseling has made some progress. They mainly enhance the ability to analyze single interactions and provide emotional support for single-round responses through cue word engineering, instruction fine-tuning, and thought chain technology. For example, they enable in-depth psychological analysis of single interactions and strategy optimization for single-round emotional support dialogues.

[0004] However, existing technologies still have three core limitations when dealing with complex, long-term psychological healing tasks that involve deep subconscious issues:

[0005] 1. Single text modality is difficult to overcome verbal defenses: Existing psychological dialogue systems mostly rely on text input analysis, but in clinical practice, users have resistance or verbal defense mechanisms at the conscious level, and pure text analysis cannot reach the deep subconscious. Psychological sandplay therapy, as a non-verbal projective testing technique, can reveal the user's deep psychological structure, but the general large language model lacks professional visual-psychodynamic interpretation capabilities and cannot transform sandplay non-verbal information into dialogue context.

[0006] 2. Digital assessment and dialogue intervention are disconnected from diagnosis and treatment: Existing digital sandplay or psychological testing tools only generate static analysis reports. The assessment results exist in isolation and cannot be injected into the dialogue system in real time. The dialogue model cannot use sandplay information for targeted intervention, which violates the scientific diagnosis and treatment principle of "assessment first, intervention later" and cannot form an assessment-intervention closed loop.

[0007] 3. Lack of long-term strategic planning based on deep psychological representations: Existing large language model dialogue technology focuses on tactical responses in the current round, without initial psychological representations based on subconscious projection, making it difficult to set layered healing goals, easily getting stuck in surface dialogue, and unable to guide users to complete the full healing process from subconscious presentation to conscious integration.

[0008] In summary, how to construct a system that integrates non-verbal psychodynamic analysis and cognitive architecture logic dialogue capabilities, enabling the model to understand user language, comprehend subconscious projections, and generate professional healing dialogues with long-term memory and goal planning capabilities, is a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0009] The purpose of this invention is to provide a psychological interaction method and system based on sandplay dynamics and a therapeutic cognitive framework.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] This invention provides a psychological interaction system based on sandplay dynamics and a therapeutic cognitive framework. The system constructs a multimodal therapeutic cognitive framework, which is composed of six core functional modules. The functions and configurations of each module are as follows:

[0012] Sand Table Dynamics Analysis Module: Serving as a cold start or auxiliary evaluation unit for the dialogue system, this module receives user-input electronic sand table images and information (including sand table object types, location coordinates, orientation, etc.) and performs deep visual reasoning. Specifically, it implements prototype symbol recognition, sand table dynamics layout evaluation, psychological dimension quantification and reasoning, and outputs text containing psychodynamic analysis. And psychological dimension score Initial mental representation vector This provides a subconscious basis for subsequent dialogue.

[0013] The observation and analysis module integrates and analyzes the user's multimodal input, including sand table information and text input. It first extracts key factual elements from the text input, then performs information conflict detection, compares the user's textual complaints with the sand table analysis results, and identifies consistency / contradiction points between explicit expressions and subconscious projections. Based on the conflict results, it dynamically adjusts the credibility weights of different modalities through a user state evaluation function with defensive weighting, accurately assessing the user's current psychological state.

[0014] Memory retrieval module: Stores the results of sand table dynamics analysis as the user's deep psychological characteristics into long-term memory, retrieves historical dialogue records in multi-turn dialogues, and constructs and dynamically updates user profiles by combining the consistent behavioral patterns revealed by the sand table, thereby achieving integrated retrieval of historical information and deep psychological characteristics.

[0015] Goal setting module: Adjust the priority of goals based on the psychological dimensions of the sandplay. When the risk is high, lock the immediate goal as "crisis intervention and empathic comfort" and when the risk is low, set it as "capability growth and self-exploration". At the same time, combine the deep conflicts exposed in the sandplay to set long-term healing goals, so as to achieve the hierarchical setting of healing goals.

[0016] Strategy Selection Module: Based on the set healing goals, select one or more psychological healing theories that fit the sandplay imagery as the current dialogue strategy, explain the reasons for the strategy selection and exclude other strategies, and ensure the suitability of the strategy; break down the selected strategy into specific response generation instructions to form an action plan.

[0017] Constraint Check Module: Performs risk control checks before the final response is generated. Based on preset dialogue taboo rules and high-risk points identified in the sandbox analysis, it reviews and filters the initial response content to ensure the safety and therapeutic nature of the output content.

[0018] A second aspect of this invention provides a method for generating psychological interaction data based on dual independent pipelines. This method constructs a sandplay visual evaluation dataset with high-confidence visual diagnostic capabilities and a therapeutic dialogue dataset with deep cognitive capabilities through two parallel and independent data processing flows, providing high-quality data support for subsequent model training.

[0019] Step 1: Construction of a Visual Evaluation Dataset for Sand Table Based on Expert Consensus

[0020] Data loading and consensus cleaning: Read the original sand table data annotated by multiple experts, remove samples with serious scoring disagreements through statistical methods, and use the intersection-union algorithm to perform spatial clustering of the regions of interest annotated by experts, retaining highly overlapping expert consensus regions and filtering out subjective biases of individual experts.

[0021] Multi-task hybrid instruction generation: Based on the cleaned consensus data, a multi-task hybrid instruction data covering visual localization, existence detection, local reasoning, stepwise scoring, and end-to-end full analysis is constructed to generate multimodal training samples that take into account both the accuracy of visual perception and the depth of psychological reasoning, forming a sand table visual evaluation dataset.

[0022] Step 2: Construction of a multi-turn dialogue dataset based on a healing cognitive architecture

[0023] Initialization and State Recovery: Load the original psychological counseling dialogue dataset, read the progress file and temporary data file, determine the interruption point, and realize breakpoint resumption;

[0024] Iterative traversal and context extraction: Traverse the dataset, locate the counselor's response, extract historical dialogues, and form a complete dialogue context sequence;

[0025] The thought chain generation request construction combines the dialogue context, the counselor's original response, and the instructions of the healing cognitive architecture system, requiring an external high-performance, high-parameter language model to perform reverse reasoning based on five modules;

[0026] Response validation and structure standardization: The generated thought chain is formatted to ensure that the module labels are complete and unqualified results are processed;

[0027] Data augmentation and persistent storage: The validated thought chain is spliced ​​with the original response to form augmented data, which is incrementally saved to a temporary file;

[0028] Completion and Cleanup: After processing is complete, a healing dialogue dataset is generated, and the progress file is deleted.

[0029] A third aspect of this invention provides a psychological interaction method based on sandplay dynamics and a therapeutic cognitive framework, implemented using decoupled data training and cascaded joint reasoning, specifically comprising two stages: independent training of dual models and cascaded joint reasoning.

[0030] Phase 1: Independent Supervised Fine-Tuning Training of Two Models

[0031] Sand table visual evaluation model training: A pre-trained multimodal large model is selected as the base, and supervised fine-tuning is carried out using the above sand table visual evaluation dataset to enable the model to establish a mapping relationship from sand table visual input to standardized psychological diagnosis reports, giving the model the ability to "understand" subconscious projections and output a structured representation containing psychological dimension scores and dynamic analysis text.

[0032] Training of the healing cognitive dialogue model: A pre-trained large language model is selected as the foundation, and supervised fine-tuning is carried out using the above-mentioned healing dialogue dataset. The multimodal healing cognitive architecture is internalized into the implicit thinking process of the model, enabling the model to generate professional healing responses based on the dialogue context and deep psychological state.

[0033] Phase Two: Cascaded Joint Reasoning

[0034] Pre-visual diagnosis: The system receives sand table images and information uploaded by the user, calls the sand table visual evaluation model to perform reasoning, and generates structured mental representation vectors. ;

[0035] Cross-modal state injection: The above-mentioned structured mental representation vectors are used as deep mental state contexts and injected into the input of the healing cognitive dialogue model to provide the dialogue model with prior knowledge of the user's subconscious state;

[0036] Healing Dialogue Generation: The healing cognitive dialogue model combines the user's current text input with the deep psychological state context, activates the internal multimodal healing cognitive architecture to reason, and sequentially completes observation and analysis, memory retrieval, goal setting, strategy selection and constraint checking, and finally generates and outputs a professional healing response that is adapted to the user's subconscious state.

[0037] The beneficial effects of this invention are as follows:

[0038] 1. This invention introduces an independent sandplay dynamics analysis pipeline, which utilizes the non-verbal projection medium of the sandplay to directly obtain the user's subconscious psychological representation before the dialogue; through cross-modal consistency verification, it identifies the user's verbal defense state, bypasses conscious resistance, and directly intervenes in deep psychological facts, greatly improving the depth and accuracy of psychological insight.

[0039] 2. Restore the professional consultation process and realize the integrated diagnosis and treatment closed loop: Construct a cascading reasoning mechanism based on "state injection" to replicate the clinical process of "assessment first, intervention later" in human psychological counseling; transform the static diagnosis report of sandplay into dynamic dialogue strategy parameters, inject them into the dialogue model in real time, break down the barriers between assessment and healing, and ensure that every response is based on scientific psychological assessment, rather than being blindly generated.

[0040] 3. Establish long-term strategic collaboration and avoid short-sighted interactions: Achieve deep collaboration of results in each stage through a multimodal healing cognitive framework. The psychological dimension score of the sandplay directly determines the level of the dialogue target, and the core imagery of the sandplay constrains the choice of strategies. Internalize the deep psychological structure revealed by the sandplay into a long-term user profile, so that the model can anchor deep healing goals in multiple rounds of interaction, achieve the leap from "tactical response" to "strategic healing", and avoid superficial and repetitive dialogue.

[0041] 4. Decoupled training and joint reasoning enhance model professionalism and stability: The dual-model decoupled training method allows the sand table visual evaluation model and the healing cognitive dialogue model to focus on visual diagnosis and cognitive dialogue respectively, ensuring their respective professionalism and optimal performance; cascaded joint reasoning effectively solves the problems of insufficient professionalism of a single model and the separation of visual illusion and diagnosis and treatment in multimodal fusion, improving the overall stability and intervention accuracy of the system.

[0042] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the multimodal healing cognitive architecture of Embodiment 1 of the present invention;

[0044] Figure 2This is a schematic diagram of the dual independent pipeline data generation process in Embodiment 2 of the present invention;

[0045] Figure 3 This is a schematic diagram of the decoupled training and cascaded joint inference process of Embodiment 3 of the present invention.

[0046] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below. Detailed Implementation

[0047] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1: Implementation of a Multimodal Healing Cognitive Architecture

[0049] The core of the psychological interaction system in this embodiment is a multimodal therapeutic cognitive framework, which consists of six modules: sandplay dynamics analysis, observation analysis, memory retrieval, goal setting, strategy selection, and constraint checking. The specific implementation details of each module are as follows:

[0050] Sand table dynamics analysis module: Receives electronic sand table images input by the user. and sand table information Perform deep visual reasoning:

[0051] Prototype symbol recognition: using a visual encoder to identify sandplay sets And it reflects the symbolic meaning of the prototypes of each sand figurine;

[0052] Dynamic layout assessment: Analyze the local themes and overall layout of the sand table, calculate the visual entropy and spatial distribution characteristics of the images, and assess the blockage / flow state of psychological energy;

[0053] Psychological Dimension Quantification and Reasoning: Generating Psychodynamic Analysis Text Calculate psychological scores such as depression dimension. (Values ​​range from 0 to 2), output the initial mental representation vector:

[0054] .

[0055] Observation and Analysis Module: Extracting User Text Input Calculate the psychological defense / conflict index based on key elements such as events, people, and emotions. Then through Assess the user's overall psychological state; among which For defense adjustment function, The larger, The lower the weight, the higher the sand table score. The higher the weight, the more adaptive the verbal defense state can be.

[0056] Memory retrieval module: By storing data in long-term memory, retrieving historical dialogue records, combining behavioral patterns revealed by the sand table, and associating them with current user behavior, a dynamically updated user profile is constructed, identifying consistent behavioral patterns at both the conscious and subconscious levels.

[0057] Target setting module: based on Setting targets for risk levels, if For high-risk cases, such as a depression score ≥1.5, the immediate goal is crisis intervention and empathetic reassurance, while challenging questions should be suspended. For low-risk cases, the immediate goal is capacity building and self-exploration. At the same time, based on the deep conflicts such as energy blockage and self-division revealed by the sandplay, long-term healing goals should be set.

[0058] Strategy selection module: via Quantitative assessment strategies are employed, appropriate psychological healing theories are selected, and specific response instructions are broken down into these instructions. It can be adjusted according to the actual application scenario, and the best option can be selected. .

[0059] Constraint Check Module: Based on preset dialogue taboo rules, such as avoiding provocative language, not making absolute judgments, and high-risk points in sandplay analysis, such as suicidal tendencies and severe anxiety, the module reviews the initial response content, filters risky content, and ensures the safety and therapeutic nature of the output.

[0060] Example 2: Generation of psychological interaction data in dual independent pipelines

[0061] This embodiment constructs a sandbox visual evaluation dataset and a therapeutic dialogue dataset through two parallel pipelines. The specific implementation steps are as follows:

[0062] Pipeline 1: Construction of the Visual Evaluation Dataset for Sand Table

[0063] Data loading and consensus cleaning: Read 1000 sand table image data with multi-expert annotations. Each data includes scores and visual region annotations from 3 to 5 psychological experts. Calculate the standard deviation of the scores for each psychological dimension and remove samples with a standard deviation > 0.5. Use the IoU algorithm to cluster the regions of interest annotated by experts and retain the regions with IoU ≥ 0.7 as the expert consensus regions.

[0064] Multi-task hybrid instruction generation: Based on 680 cleaned consensus data sets, multi-task hybrid instructions are generated, including "identifying the core sand table objects", "determining whether there are images representing loneliness in the sand table", "analyzing the psychological meaning of the sand table object layout", "gradually calculating depression dimension scores", and "conducting a full psychodynamic analysis of the sand table", forming a sand table visual evaluation dataset containing 6,800 training samples.

[0065] Pipeline 2: Construction of the Healing Dialogue Dataset

[0066] Initialization and State Recovery: Load the original psychological counseling dialogue dataset of 5000 rounds, read the progress file, and determine that the last processing was interrupted at round 2300;

[0067] Iterative traversal and context extraction: Starting from round 2300, traverse the dialogue until the counselor responds, then extract all previous historical dialogues to form a complete context sequence;

[0068] The thought process of generating a request is constructed by reverse deducing the thought process of the response based on the context sequence, the counselor's original response and the system instructions, using the modules of observation and analysis, memory retrieval, goal setting, strategy selection and constraint checking.

[0069] Response validation and structure standardization: The external large model is called to generate the thought chain. The regular expression is used to validate whether it contains complete labels of observation, memory, goal, strategy and constraint. The 210 results that failed the validation are manually extracted and filled with default values.

[0070] Data augmentation and persistent storage: The validated thought chain is spliced ​​with the original response to form augmented data. The progress file is updated every 100 rounds and a temporary data file is written every 500 rounds.

[0071] Completion and Cleanup: After processing 5000 rounds of data, save the contents of the temporary data file as a healing dialogue dataset containing 5000 enhanced data points, and delete the progress file.

[0072] Example 3: A Psychological Interaction Method for Decoupling Training and Cascaded Joint Reasoning

[0073] This embodiment uses decoupled training to obtain two independent models, and then achieves psychological interaction through cascaded joint reasoning. The specific implementation steps are as follows:

[0074] Phase 1: Independent Supervised Fine-Tuning Training of Two Models

[0075] Sand table visual evaluation model training: A pre-trained basic multimodal large model was selected as the base, and supervised fine-tuning was performed using the sand table visual evaluation dataset constructed in Example 2. The training batch size was 32, the learning rate was 1e-5, and the training epoch was 10. After training, the model can output a structured diagnostic report containing depression dimension scores and dynamic analysis text from sand table image input. The visual feature recognition accuracy is ≥85%, and the consistency of psychological reasoning is ≥80%.

[0076] Training of the healing cognitive dialogue model: ChatGLM was selected as the pre-trained large language model base. Supervised fine-tuning was performed using the healing dialogue dataset constructed in Example 2. The training batch size was 16, the learning rate was 2e-5, and the training epoch was 8. After training, the model internalized the multimodal healing cognitive architecture into an implicit thought process and could generate professional healing responses based on the dialogue context.

[0077] Phase 2: Cascaded Joint Reasoning

[0078] Pre-operative visual diagnosis: Users upload images of the electronic sand table and information about the sandplay objects. The system then uses a sand table visual assessment model to generate a psychodynamic analysis text: "The sandplay objects are scattered, indicating significant energy blockage, representing a psychological state of interpersonal isolation and self-isolation in the user," along with a depression dimension score. , forming a representation vector ;

[0079] Cross-modal state injection: As a deep psychological state context, it is injected into the input of the healing cognitive dialogue model;

[0080] Healing dialogue generation: The user's text input in this round is "I think I'm usually quite cheerful, and I have quite a few friends." The healing cognitive dialogue model combines this text with the injected... Observation and analysis detected a conflict between the user's explicit expression and subconscious projection. The system constructs a user profile through memory retrieval. The goal setting module sets the immediate goal as "gentle exploration, guiding self-awareness". The strategy selection module selects humanistic healing strategies. After constraint checking and filtering, the system generates a response: "I can feel that you want to maintain a cheerful state. Behind this is your concern for your own state. However, sometimes our outward state and our true inner feelings may be different. Do you sometimes feel a sense of loneliness when you are alone?"

[0081] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0082] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A psychological interaction system based on sandplay dynamics and a therapeutic cognitive framework, characterized in that, The system constructs a multimodal healing cognitive architecture, which consists of six functional modules: sandplay dynamics analysis module, observation and analysis module, memory retrieval module, goal setting module, strategy selection module, and constraint checking module. The sand table dynamics analysis module, as a cold start or auxiliary evaluation unit of the dialogue system, performs deep visual reasoning on the electronic sand table images and sand table information input by the user, identifies and interprets the meaning of the sand table prototypes, analyzes the sand table layout and psychological state, outputs psychological dimension scores and psychodynamic analysis text, and forms initial psychological representations. The observation and analysis module performs fusion analysis on the user's multimodal input, extracts key input elements, detects information conflicts between the user's textual complaints and the results of the sand table dynamics analysis, and evaluates the user's current psychological state through a user state evaluation function based on the conflict results. The memory retrieval module stores the results of the sand table dynamics analysis as the user's deep psychological characteristics into long-term memory, retrieves historical dialogue records and combines them with the behavioral patterns revealed by the sand table to construct a dynamically updated user profile. The goal setting module adjusts the priority of goals based on the sandplay psychological dimension score, and sets immediate healing goals and long-term healing goals in layers based on the deep conflicts exposed in the sandplay. The strategy selection module selects a psychological healing theory that matches the sandplay imagery as a dialogue strategy based on the set healing goals, explains the reasons for the strategy selection, and breaks it down into specific response generation instructions. Before a response is generated, the constraint inspection module reviews and filters the initial response content based on preset dialogue taboo rules and high-risk points analyzed in the sandbox, thereby achieving risk control.

2. The psychological interaction system according to claim 1, characterized in that, The deep visual reasoning process of the sand table dynamics analysis module is formalized as follows: Prototype symbol recognition: Using a visual encoder to identify representative sets of sand table objects and map their prototype symbolic meanings; Dynamic layout assessment: Analyze the local themes and overall layout of the sand table, calculate the visual entropy and spatial distribution characteristics of the images, and assess the state of blockage or flow of psychological energy. Psychological Dimension Quantification and Reasoning: Generating Psychodynamic Analysis Text Based on the Above Features And calculate psychological dimension scores. The output is represented as a vector: This vector is passed as an initial mental representation to subsequent modules.

3. The psychological interaction system according to claim 1, characterized in that, The observation and analysis module performs cross-modal consistency verification and psychological state assessment in the following ways: Define psychological defense / conflict index Quantify the deviation between users' conscious and subconscious minds; Based on δ, the overall psychological state is evaluated using a user state evaluation function: ， in, For the user's latest input text, For sentiment analysis functions, For psychological energy assessment function, Scoring the psychological dimensions of the sandplay therapy. Here, α and γ are the defense adjustment functions, and they are the weighting coefficients. This is the normalization function.

4. The psychological interaction system according to claim 1, characterized in that, The strategy selection module quantifies and selects dialogue strategies using the following formula: ; Among them, Goal refers to the set immediate and long-term healing goals. The score represents the alignment between strategy and objectives. The estimated acceptance score for the strategy by users. The score represents the degree of match between the strategy and the sand table dynamics analysis text. These are the dimension weight coefficients.

5. A method for generating psychological interaction data based on dual independent pipelines, characterized in that, The sand table visual evaluation dataset and the healing dialogue dataset were constructed through two parallel data processing workflows, including the following steps: S1. Construction of a visual evaluation dataset for sand table based on expert consensus Data loading and consensus cleaning: Read the original sand table data annotated by multiple experts, calculate the standard deviation of scores for each psychological dimension at the scoring level, and remove samples with serious disagreements; at the visual level, use the intersection-union algorithm to perform spatial clustering of expert-annotated regions of interest, and retain highly overlapping expert consensus regions. Multi-task hybrid instruction generation: Based on the cleaned consensus data, construct multi-task hybrid instruction data covering visual localization, existence detection, local reasoning, stepwise scoring, and end-to-end full analysis, generate multimodal training samples that take into account both visual perception accuracy and psychological reasoning depth, and form a sand table visual evaluation dataset. S2. Construction of a multi-turn dialogue dataset based on a healing cognitive architecture Initialization and State Recovery: Load the original psychological counseling dialogue dataset, read the processing progress file and temporary data file, and determine the location where data processing was interrupted; Iterative traversal and context extraction: Traverse the dataset sequentially, locate the counselor's response, and extract all historical dialogues up to that response to form a complete dialogue context sequence; Cognitive architecture thought chain generation request construction: combining the dialogue context sequence, the counselor's original response text and system instructions containing the definition of the healing cognitive architecture, the system instructions require an external high-performance, high-parameter language model to perform reverse reasoning based on observation and analysis, memory retrieval, goal setting, strategy selection and constraint checking modules; Response validation and structural standardization: The format of the thought chain generated by the external high-performance, large-parameter language model is validated to ensure that it contains complete module tags, and the results of validation failure are extracted or filled with default values. Data augmentation and incremental persistent storage: The validated thought chain is spliced ​​with the original response to form dialogue cognitive augmentation data, which is updated in real time and written to temporary data files periodically; Completion and Cleanup: After processing, save the contents of the temporary data file as a healing dialogue dataset and delete the progress file.

6. A psychological interaction method based on sandplay dynamics and a therapeutic cognitive framework, characterized in that, This is achieved by decoupling data training and joint inference, and includes the following steps: S1, Independent Supervised Fine-Tuning Training of Two Models Sand table visual evaluation model training: A pre-trained multimodal large model is selected as the base, and the sand table visual evaluation dataset constructed by the method described in claim 5 is used to perform supervised fine-tuning on it, so that the model establishes a mapping relationship from sand table visual input to standardized psychological diagnosis report, and outputs a structured representation containing psychological dimension scores and dynamic analysis text. Training of the healing cognitive dialogue model: A pre-trained large language model is selected as the foundation, and the healing dialogue dataset constructed by the method described in claim 5 is used to perform supervised fine-tuning on it, so as to internalize the multimodal healing cognitive architecture into the implicit thinking process of the model and establish a mapping relationship from dialogue context to professional healing response. S2, Cascaded Joint Reasoning Pre-visual diagnosis: Receives electronic sand table images and information uploaded by users, calls upon the trained sand table visual evaluation model for reasoning, and generates structured mental representation vectors. ; Cross-modal state injection: The structured mental representation vector is injected as a deep mental state context into the input of the trained healing cognitive dialogue model; Healing Dialogue Generation: The healing cognitive dialogue model combines the deep psychological state context of the user's current text input and injection, activates the internal multimodal healing cognitive architecture to reason, and generates and outputs healing responses that are adapted to the user's subconscious state.

7. The psychological interaction method according to claim 6, characterized in that, The logic of the cascaded joint reasoning is isomorphic to the clinical workflow of human psychological counselors. First, the user's psychological assessment is completed through the sandplay visual assessment model, and then the assessment results are injected as prior knowledge into the healing cognitive dialogue model to guide the dialogue intervention process.

8. The psychological interaction system according to claim 1 or the psychological interaction method according to claim 6, characterized in that, The multi-dimensional mental health score of the sandplay dynamics analysis module has a value range of 0 to 2 for each dimension, where a higher score indicates a higher level of psychological risk. When any dimension score is high risk, the goal setting module will immediately lock the healing goal to crisis intervention and empathic reassurance, and suspend challenging questions.

Images