Psychological consultation system and method based on artificial intelligence

Abstract

The invention provides a psychological counseling system and method based on artificial intelligence, and relates to the field of intelligent psychology, comprising session node management, trust state management, observation management, antagonistic narrative management, self-organizing treatment, and asymmetric override and permission and auditing combined scale block cooperative operation. The traceability and credible management of the whole psychological consultation process are realized; the system generates an auditable session node and gives a trust state identifier at the end of each session, and a qualified observation chain is formed through user clear interaction to trigger trust judgment; the antagonistic narrative module generates an anti-narrative according to an ethical rule and collects a user structured response as a trust evidence; the self-organizing treatment module automatically executes node classification and opens artificial recheck when a critical condition is met; the override and permission module introduces an independent override to examine the trust and ethical compliance of the promotion node; the auditing module writes all events into a non-tampering account book and triggers responsibility tracing when a rule is violated; temporary intervention can be executed through multi-party rechecking in an emergency situation so as to guarantee the safety of a user.

Description

An AI-based psychological counseling system and method Technical Field

[0001] This invention relates to the field of intelligent psychology, specifically to a psychological counseling system and method based on artificial intelligence. Background Technology

[0002] Mental health services are developing rapidly due to demographic changes and increased public awareness; online and remote counseling have become the main forms; artificial intelligence and big data are empowering personalized diagnosis and treatment and resource allocation; regulatory and ethical requirements are prompting the industry to focus on transparent and traceable governance mechanisms and driving the continuous evolution and development of related technologies and compliance practices. Currently, most systems rely on human expert review and centralized records for quality control; some platforms use encrypted storage and differential privacy to protect sensitive data; some studies use statistical models or rule engines to identify abnormal behavior and trigger human intervention; there are also review processes based on legal and ethics committees, and continuous testing and evaluation are needed. The shortcomings of existing technologies are: the chain of evidence often comes from a single source or is affected by human judgment, reducing verifiability; automated decision-making lacks multi-source independent verification, resulting in insufficient robustness; the timeliness and transparency of default handling procedures and supervision and review for emergency situations need to be strengthened, and governance and review provability assurance need to be enhanced. Summary of the Invention

[0003] (I) Technical Problems to be Solved In view of the shortcomings of the prior art, the present invention provides a psychological counseling system and method based on artificial intelligence to solve the problems mentioned in the background art above, such as the evidence chain often having a single source or being affected by human judgment, which reduces verifiability; the lack of multi-source independent verification in automated decision-making, resulting in insufficient robustness; the need to strengthen the timeliness and transparency of the default handling process and supervision and review for emergency situations; and the need to strengthen governance and review of provability assurance.

[0004] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A psychological counseling system and method based on artificial intelligence, comprising: a session node management module for generating a session node according to preset rules at the end of each counseling session, initializing the session node to a candidate state, and writing the creation event of the session node into an immutable audit ledger; a trust state management module for creating a trust state identifier for the session node during generation, placing the trust state identifier in a candidate state, and triggering a trust determination procedure after receiving a qualified observation chain from the observation management module, and after determination, setting the session node as a candidate node. Nodes are labeled as trustworthy or untrustworthy, and the judgment result and the observation chain on which the judgment is based are written into an immutable audit ledger; the observation management module is used to receive and verify only interactions explicitly initiated by the user as qualified observations, forming an observation chain to trigger the trust judgment procedure, performing a qualification review on each observation, performing a multi-party review process for unconventional and urgent observations, and recording the review conclusions; the adversarial narrative management module is used to automatically generate a counter-narrative according to ethical rules when the session node is in a candidate state, present the counter-narrative to the user after obtaining the user's explicit informed consent, record the user's structured response to the counter-narrative, and The response is written into an immutable audit ledger, and a temporary unlocking process for emergency rescue exceptions is also included to trigger a temporary observation path when an immediate risk signal is detected. The self-organizing governance module receives judgment evidence from the aforementioned modules and automatically triggers a node classification procedure when preset critical conditions are reached within the system. This allows a group of session nodes to be branched into promotion paths or isolation paths according to preset criteria without initial human intervention. After classification, manual review is allowed within a limited window, and the classification result is confirmed or rolled back based on the review conclusion after the manual review is completed. The asymmetric veto and permission module is used when a session node is to be promoted to... Before long-term use, vetoers with topic matching and independence qualifications are selected according to preset rules. The vetoers are provided with observation chains, trust judgment evidence and adversarial responses as review materials, and the veto and alternative handling process is recorded. The audit and compliance governance module is used to write all events of the generation of the session node, the reception and verification of the observation, the trust judgment, the adversarial presentation and user response, the node classification and veto process into an immutable audit ledger. The module provides searchable review entry points to regulators or designated supervisors according to permissions, and triggers the accountability and rectification process when the audit finds that the rules are violated.

[0005] Preferably, at the end of each consultation session, the session node management module, through the session end detection unit, identifies the end event based on the session silence timeout or the user's active termination command, and records a high-precision timestamp and a globally unique session identifier. The session node generation unit then creates a structured data instance of the session node according to preset rules, containing a semantic vector of the session summary, interaction metadata, and a non-reusable encrypted hash identifier, and initializes its state to a candidate state. The creation event is synchronously written to an immutable audit ledger based on distributed ledger technology in a structured record format containing fields such as session identifier, creation timestamp, and initialization state. The session node management module places the node into a candidate pool and starts a configurable 72-hour timer window to wait for observational evidence reinforcement. If at least three qualified observations verified and written by the observation management module are received within this window, and these observations originate from different verification channels and cover… If at least two interaction types are present, the session node management module will submit the node to the trust status management module to trigger the trust determination procedure and add a submission event record to the audit ledger. If the above observation quantity and diversity requirements are not met within 72 hours, the node will be marked as "unstable" and migrated to the unstable pool. The reason for migration and timestamp will be recorded in the audit ledger. Under this normal process, nodes that do not meet the observation requirements cannot be marked as trustworthy or progressively designated for long-term use. However, if the observation management module reports an immediate risk signal identified by the model at any time, the session node management module will immediately interrupt the normal process, trigger a temporary unlocking process for emergency rescue exceptions, and write the trigger signal, the limited intervention permissions granted, and the timestamp into the audit ledger. Furthermore, it will provide a queryable interface based on event indexes for the self-organizing governance module and the audit compliance governance module to facilitate subsequent evidence collection and automated review.

[0006] Preferably, in the trust state management module, the trust state identifier generation submodule generates an invisible and unique trust state identifier by calling an encryption algorithm according to preset rules when the session node creation event is written into the audit ledger. The specific steps include receiving the globally unique identifier of the session node, generating a fixed-length unique string identifier in the identifier generation unit using the SHA-256 algorithm combined with a timestamp and a random number seed, then establishing a bidirectional mapping relationship between this identifier and the session node identifier, forming a structured creation event entry containing the creator identifier, timestamp, and initial "candidate" label. This entry is immediately written into the immutable audit ledger to ensure the traceability and non-repudiation of its generation process. The trust determination submodule then initiates an automated determination program when the qualified observation sequence provided by the observation chain management submodule meets preset conditions. According to the determination rule table stored in the security configuration, it reads the observation chain content and verifies the independence of the observation source, the diversity of interaction types, and the logical consistency of the observation time series item by item. Based on the weights and formulas defined in the rule table, it calculates a quantified trust score. When the calculated trust score is greater than or equal to... When the trust score is above 80 and the observation chain contains at least 3 qualified observations from different verification sources with different interaction types, the trust determination submodule marks the session node as "trustworthy" and writes the complete determination result, the index of the determination rule used, and the original text summary of the observation chain data used into the immutable audit ledger. If the trust score is below 80 or the observation chain does not meet the minimum quantity and diversity requirements, the node is marked as "untrustworthy" and a migration instruction with a digital signature is immediately issued to the archiving submodule. After receiving the instruction, the archiving submodule first verifies the validity of the instruction signature, then copies the read-only image of the session node, generates an archive storage object that is write-locked at the storage layer and cannot be copied, and writes the node identifier and migration reason into its metadata. At the same time, a strict access control policy is applied to the archive object, prohibiting all regular AI training and user profiling modules from reading it, and stipulating that any subsequent recovery access request must be approved by at least 3 independent authorized parties through a contract-based signing procedure based on threshold signature technology. The entire authorization application, signing process, and approval result are recorded in detail in the audit ledger, forming a closed-loop security control.

[0007] Preferably, the observation eligibility determination submodule in the observation management module immediately initiates an automated eligibility review process upon receiving any form of interaction from the user, including explicit written confirmation, explicit verbal confirmation, explicit click confirmation on the user interface, or controlled identity verification through biometric matching. The review process specifically includes: verifying the digital identity credentials of the interaction initiator through an integrated identity service interface; verifying its digital signature using public key infrastructure technology; verifying the trust level of the credential issuing authority and the validity period of the credential; and comparing the timestamp of the interaction with the globally unique identifier of the session to ensure sequential logic. To ensure consistency, the audit process continues to calculate contextual semantic similarity, using a vector similarity algorithm based on a pre-trained language model for scoring. When the similarity score is not less than 0.85 and the difference between the interaction sequence and the current session node is within a preset 300-second tolerance window, the interaction, along with its source type, interaction type, verification evidence, signature digest, verifier identifier, and precise timestamp, is written into an immutable audit ledger in a structured record format. The writing operation uses a sequentially increasing hash chain structure and binds the timestamp to an external trusted time source to ensure the orderliness and immutability of the records, while preserving the original interaction. The cryptographic digest of the data is used for verification during auditing; the observation chain management submodule, according to preset rules, requires that a qualified observation sequence used to trigger trust determination must consist of observations from at least three different sources and covering at least two different interaction types; it is responsible for maintaining the integrity and traceability of the observation sequence by calculating the content hash value of each observation in the sequence, chaining these hash values ​​in the order of receipt, and finally writing the observation index and hash chain relationship into the audit ledger; when the observation sequence simultaneously meets the preset requirements of source independence and interaction type diversity, the submodule triggers the trust determination procedure and releases the sequence The session node is marked as a usable chain of evidence. If the sequence does not meet the requirements, the submodule marks the session node as "unstable" in the audit ledger, moves it to the unstable pool to await supplementary evidence, and records the missing element categories and the system's suggested reinforcement directions in detail in the migration log. For unconventional observations triggered by emergency situations, the system performs a post-event multi-party review process, requiring at least three independent reviewers to submit written confirmations or negative opinions on the compliance of the observations through the security review interface with electronic signatures within a limited time window. All review conclusions and reviewer identity information are solidified and written into the tamper-proof audit ledger.

[0008] Preferably, the anti-narrative generation submodule in the adversarial narrative management module executes a structured anti-narrative candidate generation process based on the semantic elements of the current session node and the historical interaction context. First, it performs natural language processing on the node content, including word segmentation, entity recognition, topic keyword extraction, sentiment vector scoring, and normalization of the user's factual statements. Then, it selects templates from a predefined anti-narrative style library that do not contain leading phrases or introduce potentially dangerous details, and generates a set of anti-narrative candidate statements through a slot-filling mechanism. The entire generation process, including timestamps, template identifiers used, semantic summaries of the original input, and generation engine identifiers, is recorded in the audit ledger. Before presenting any anti-narrative to the user, the ethics and compliance submodule performs an automated compliance check process, including matching candidate text to a blacklist of sensitive content keywords, risk scoring assessment based on rule models, and contextual inappropriateness calculation, only releasing a statement when the overall context is suitable. Publication is only permitted when the compliance score is no lower than the preset threshold of 90 points. Upon publication, the system presents the reverse narrative in the user interface in a non-intrusive, brief pop-up or highlighted text manner, while simultaneously displaying an exit button, an explanation of the exit reason, and complete instructions for withdrawal. Furthermore, prior to presentation, the system must obtain the user's active confirmation based on explicit informed consent, and the consent status, along with its timestamp and interaction version number, is written into the audit ledger in real time. The response classification submodule performs an automated, structured classification process for user responses, covering text cleaning and sentence segmentation, sentiment polarity determination based on sentiment dictionaries and models, semantic matching of key evidence items, and the application of preset mapping rules. Ultimately, responses are classified into three categories: "explicit rebuttal," "confirmation," or "vague acceptance." Based on a predefined mapping rule table, the classification results are transformed into trust evidence items that can be used for quantitative evaluation. All original response texts, classification results, and derived evidence items are fully recorded in the immutable audit ledger.

[0009] Preferably, when the archiving submodule in the trust state management module receives a migration instruction for an untrusted node with a digital signature from the trust determination submodule, it immediately executes a standardized archiving migration process. First, it verifies the digital signature of the migration instruction to ensure its legality and integrity. Then, it creates a read-only data mirror of the node and generates a non-replicable archive storage object with a write lock flag enabled at the object storage layer. Simultaneously, it writes the globally unique identifier of the session node and the specific migration reason code into the object's metadata. The archiving submodule applies immutable storage strategies to such archive objects, such as using copy-on-write snapshot technology or blockchain storage anchoring, and implements object-level hash locking at the storage layer to ensure its content cannot be tampered with. To strengthen access control, the system divides the key used to decrypt the object into multiple ciphertext fragments using a threshold signature scheme and distributes them to a preset number of independent authorized parties for safekeeping. The system's access control rules explicitly prohibit all regular AI model training and user profile analysis modules from reading objects in the archive pool, allowing only the audit and compliance review interfaces to access them in read-only mode, and each access request must pass a two-factor authentication. The system employs an authentication and pre-approval process. Any request to restore access to archived content must be accompanied by a formal application document detailing the purpose, scope, timeframe, and a commitment to assume responsibility. The restoration approval process requires at least three independent authorized parties to complete the approval through a legally binding electronic contract signing procedure. The signing record must include the public key signatures of all parties, a precise timestamp, and a clear statement of responsibility. These signing records are aggregated into a formal restoration approval certificate using threshold signature technology. The archiving submodule then writes the complete restoration application, evidence of the signing process, and the final approval result into an immutable audit ledger. Upon approval, the system generates a time-limited and scope-limited temporary access token within a defined timeframe and records the token information and authorized access scope in the audit ledger. After the token expires, the key management module automatically performs key sealing and access permission revocation registration. If the restoration application is rejected or the contract signing fails, the archiving submodule maintains the read-only and non-replicable attributes of the object and records the reasons for rejection and related responsibility allocation information in detail in the audit ledger to form a complete control loop.

[0010] Preferably, the self-organizing governance module, through the collaborative operation of its critical decision-making submodule, node classification and execution submodule, and manual review submodule, achieves automated routing and closed-loop supervision of session nodes. The critical decision-making submodule continuously runs within the system's pre-defined decision network graph, collecting and summarizing standardized evidence items from the session node management module, trust state management module, observation management module, and adversarial narrative management module in real time. Standardized preprocessing is performed on each evidence item, including unifying timestamps of different formats to international standard time, assigning dynamic weights based on the credibility level of the evidence source, and calculating the semantic and logical consistency of the evidence cluster. The system's preset critical triggering conditions are quantified as follows: within any rolling 24-hour time window, the cumulative number of associated independent evidence sources is not less than 10, and the overall consistency score of the evidence cluster is not less than 0.9, or the semantic connectivity index between decision network nodes exceeds 0.75; when either condition is met, the critical judgment submodule sends a classification trigger signal to the node classification execution submodule; after receiving the signal, the node classification execution submodule automatically performs the diversion operation according to the pre-loaded branch criterion rule set, where the promotion path requires the overall trust score of associated session nodes to be not less than 85 points and there are no "unstable" or risk labels, while the interval The off-path is activated when any key evidence fails the eligibility review or the session content reaches a preset risk content threshold. Each automatic routing operation generates a node classification decision summary, including a list of evidence indexes that triggered the signal, the specific matching criterion entries, the scores calculated by the system, and a precise timestamp. The full summary is immediately written into an immutable audit ledger. Subsequently, the manual review submodule submits the above decision summary to a designated independent supervisor via a secure communication channel within a preset 60-hour review window. The supervisor must review the summary and related evidence in a dedicated review interface and submit it in legally binding electronic form. The system uses a sub-signature to make a written decision of "acceptance" or "rollback". If the decision is "acceptance", the manual review submodule updates the final status in the audit ledger and attaches the review comments and supervisor identification. If the decision is "rollback", the submodule executes the rollback process, restores the relevant nodes to their original state before classification, and generates detailed operation records and responsibility allocation instructions in the audit ledger. If no manual review conclusion is received within a 60-hour window, the manual review submodule will automatically trigger temporary protection measures, immediately freeze all access permissions of the relevant nodes, and send a priority notification to the audit and compliance governance module to initiate subsequent special reviews and accountability procedures.

[0011] Preferably, the asymmetric veto and approval module, through the collaboration of its vetoer selection submodule, material review and provision submodule, and judgment and processing submodule, constructs an accountable human oversight mechanism. The vetoer selection submodule executes an automated selection process according to preset rules. First, it screens candidates from a pool of vetoers qualified with expertise in the subject area. Key screening criteria include that the candidate's professional subject tags must perfectly match the subject marked on the conversation node, their digital independence credentials must be verified, and the interval between their most recent contact with similar subjects or related users must be greater than 30 days to ensure independence. During the selection process… During the process, the system performs rigorous conflict of interest avoidance verification by cross-checking the records of candidates with current session participants, including users and associated consultants, in the pre-declared conflict of interest list to ensure that there are no direct economic transactions, employment relationships, or familial ties between the parties. Finally, the system uses an auditable random number generator to select three vetoes from the pool of candidates that have passed the screening, and writes the complete selection logic, random number seed, and result list into an immutable audit ledger. After the selection is completed, the review materials provision submodule automatically pushes a structured review material to the security review interface of each selected vetoe. The package includes the complete observation chain text and verifiable digital digest, an automatically generated trust determination evidence table, and a structured record of user adversarial responses. All materials are accompanied by proof of origin and tamper-proof timestamps. The judgment processing submodule requires the rejector to make a written judgment using electronic signatures based on the provided materials in the review interface. Options include "agree," "disagree," or "propose an alternative solution." When a rejector selects "disagree," the system automatically triggers a locked review process, mandating that the rejector simultaneously submit a detailed and actionable alternative solution. The solution must be clearly defined. The system outlines the steps, responsible parties, and execution timeline, with the vetoer signing a commitment letter in digital signature to assume joint liability. The entire veto process, the full text of the submitted alternative, and the subsequent multi-party review conclusions are all recorded as key events in an immutable audit ledger. Furthermore, the system continuously monitors the behavior of vetoers. If a vetoer raises "objections" more than five times within a preset statistical period, or if their submitted alternative is deemed unqualified in subsequent reviews, the system will automatically trigger a monitoring and retraining process for the vetoer, recording the triggering reason and the processing result in the audit ledger.

[0012] Preferably, the audit and compliance governance module, through the collaboration of the event logging submodule, the review entry submodule, and the rectification trigger submodule, constructs a fully traceable audit and supervision system. The event logging submodule monitors and solidifies all key operational nodes in the system, including the generation of session nodes, the receipt and verification of observations, the trust determination process, the presentation of adversarial narratives and user responses, and the results of each step in the node classification and veto process. Each event is converted into a structured data entry, with standard fields including a unique event number, a high-precision timestamp, the identity identifier of the event trigger source, a globally unique identifier of the associated session node, and a hash-calculated summary of the event content. The system includes the hash value of the previous event used to ensure chain continuity; before writing, the system uses the SHA-256 algorithm to calculate a digital digest of the complete event record, and the responsible event logger digitally signs it using its private key. The signed record is then submitted to an immutable audit ledger maintained based on a distributed consensus mechanism to ensure data integrity and non-repudiation; the audit entry submodule provides a secure audit access interface for external regulators or designated independent supervisors, and implements strict access control. Accessors must pass two-factor authentication and match a preset minimum privilege role policy before logging in; the search function supports searching by time interval, event type, and session node. The system combines multiple dimensions, including identifiers and evidence indexes, for querying and filtering. All retrieval operations are recorded as audit events, including the retrieval user, search criteria, and timestamp, ensuring complete transparency of the review process. The rectification trigger submodule incorporates a continuously running audit rule checking engine. It explicitly defines rule violations as several detectable technical anomalies, such as failed digital signature verification, broken event sequence hash chains, failed identity verification of the observed source, automatic trust determination results exceeding the allowed range of the preset judgment table, or failure to complete the necessary manual review process within a specified timeframe. Once the engine detects any violation... Upon activation, the rectification trigger submodule immediately and automatically executes the response procedure, including freezing all access permissions of the relevant session nodes and sending a written notification with event details to the designated responsible person via an encrypted communication channel within 24 hours. The system further requires that within 72 hours, at least three independent reviewers be convened to review the event through a secure interface and form a written review conclusion in the form of electronic signatures. The responsible party must submit a detailed rectification plan within 30 days, specifying the rectification measures, responsible persons, and completion timetable. After the rectification measures are implemented and verified, the entire process record will be written into the audit ledger, thus forming a complete management closed loop from violation discovery to rectification completion.

[0013] Preferably, the temporary unlocking process for emergency rescue exceptions is designed to address immediate high-risk situations identified during consultations. It is continuously activated by a risk identification unit, which monitors multiple signals in real time during the session. This includes analyzing user input text for emotion and crisis keywords based on natural language processing, detecting anomalies in user interaction patterns, and receiving data from integrated external alarm system interfaces. A quantified immediate risk score is calculated based on a pre-set risk assessment model. When the calculated immediate risk score is greater than or equal to 0.95, or when any single risk signal, such as a clear expression of suicidal intent or a violent threat, reaches the pre-set "severity level 3," the system will automatically trigger a temporary observation pathway, interrupting regular automated observation. The process involves a temporary observation access authorization system that uses manual intervention for emergency verification. Observations made by authorized hotline operators accessing the system via an encrypted voice channel, or by controlled human observers undergoing strict identity verification and synchronous monitoring, are considered valid emergency observations. Once the hotline is connected, the operator must complete a standard controlled identity verification process and record the user's explicit verbal confirmation via audio recording and electronic signature. If a controlled human observation mode is used, two independently qualified observers with no conflict of interest are required to record the observation content separately in synchronous or sequential mode. The original observation text, the observers' digital identity credentials, and the timestamp are then written into an immutable audit ledger. Upon triggering the temporary access, the system immediately notifies the authorized personnel... The crisis intervention team grants a strictly limited set of access permissions, typically including only temporary read access to the current high-risk session node and a 24-hour time-limited intervention operation permission, based on Coordinated Universal Time (UTC) and starting a countdown. Any request to extend the default 24-hour duration must be reviewed and approved in writing by at least three independent reviewers in the security review interface, and each approver must sign an electronic joint liability commitment. After the emergency intervention is implemented, the system automatically initiates a mandatory multi-party review process within 72 hours, requiring at least three independent reviewers to provide written reviews based on the complete materials provided, including risk-triggered evidence, summaries of human communication recordings, and controlled observation records, in the review interface. The system determines whether to "maintain" the temporary access permissions granted due to intervention or to "roll them back." The review conclusion must list the reasons for accepting or rejecting each piece of evidence and confirm the reviewer's responsibility with their electronic signature. The entire review conclusion, the identification of all participating reviewers, and the final responsibility allocation entries are irreversibly written into the audit ledger, forming an immutable post-event traceability record. If the review conclusion decides to roll back, the system immediately restores the node to its original access control state before the trigger and records the rollback operation. If the review process is not completed within 72 hours, the system will automatically freeze all temporary access permissions and mark the case as an unresolved event, transferring it to the audit entry point to initiate a deeper responsibility tracing and rectification process.

[0014] (III) Beneficial Effects This invention provides an artificial intelligence-based psychological counseling system and method. It has the following beneficial effects: 1. By generating a session node after each psychological counseling session and writing it into an immutable audit ledger, this invention ensures the traceability of all counseling process behaviors; and by collaboratively establishing a qualified observation chain through trust status management and observation management modules, it effectively avoids interference from false or manipulated data in trust determination.

[0015] 2. This invention operates in synergy between adversarial narrative management and self-organizing governance modules; it intervenes in ethical compliance and anti-narrative generation at the candidate state stage; it ensures that user responses are structured and recorded and incorporated into trust assessment; it forms a multi-layered review loop in conjunction with asymmetric veto and audit governance mechanisms; it achieves full verification and self-correction from generation to archiving; and it helps to improve the ethical compliance and social trust of AI-based psychological counseling. Attached Figure Description

[0016] To make the content of the present invention easier to understand, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 is a flowchart of session node generation and trust determination of the present invention; Figure 2 is a flowchart of emergency rescue exception temporary unlocking of the present invention. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: This embodiment of the invention provides an artificial intelligence-based psychological counseling system and method, including: at the end of a psychological counseling session, a session end detection unit identifies the end event and records the end time and a unique session identifier; a session node generation unit reads the end record, creates a session node instance according to preset rules, and assigns a non-reusable unique identifier and a candidate status; the session node generation unit immediately writes the creation event as a structured entry into an immutable audit ledger and starts a 72-hour waiting window for observation reinforcement; the observation qualification determination submodule performs identity credential verification and verifies the identity through public key signature and credential issuing authority when it receives any interaction such as explicit written confirmation, explicit verbal confirmation, explicit click confirmation on the user interface, or controlled identity verification. To verify the authenticity of the source, the observation qualification determination submodule, when the context similarity score is not less than 0.85 and the time difference is not greater than 300 seconds, writes the interaction as a structured record into an immutable audit ledger and transmits the observation item to the observation chain management submodule. The observation chain management submodule requires that a qualified observation consists of an observation sequence with at least 3 different sources and at least 2 different interaction types. The observation chain management submodule calculates the sequence hash of each observation in the order of receipt and writes the sequence hash and observation index into the audit ledger to maintain integrity. When the observation sequence meets the requirements of source independence and interaction diversity, the observation chain management submodule triggers the trust determination procedure and hands the observation chain over to the trust determination submodule for further determination. The trust status identifier generation submodule generates the following when the session node is created: The unique trust status identifier is invisible and written into the audit ledger as a creation entry. The trust determination submodule receives the observation chain and verifies the independence of the observation source, the diversity of observation interaction types, and the consistency of the time series item by item according to the preset determination table. It also calculates the trust score according to the determination rules. When the trust score is greater than or equal to 80 and the observation chain contains at least 3 qualified observations from different sources with different interaction types, the trust determination submodule marks the session node as trustworthy, records the determination basis in the audit ledger, and writes it into the complete text of the observation chain. When the trust score is less than 80 or there are fewer than 3 observation chains, the trust determination submodule marks the session node as untrustworthy and issues a migration instruction to the archiving submodule. The archiving submodule verifies the migration instruction, signs and copies the read-only image of the session node to generate an uncopyable archive. The object is processed by writing a unique session node identifier and migration reason into its metadata. The archiving submodule divides the archived object key into three parts and distributes them to independent authorized parties using a threshold signature scheme. The archive recovery process requires at least three independent authorized parties to complete a contractual signing and generate a recovery approval certificate before temporary access is allowed. The reverse narrative generation submodule, based on the semantic elements of the session node, performs word segmentation, entity recognition, topic extraction, sentiment scoring, and factual statement normalization. It then selects templates from the approved reverse narrative style library that do not contain leading wording or introduce dangerous details to fill slots and generate a reverse narrative candidate set. The ethics and compliance submodule performs sensitive keyword blacklist matching, risk scoring assessment, and contextual inappropriateness calculation for each candidate, requiring a compliance score of at least 90 before publication.Before release, the interface displays an exit path and enforces explicit informed consent from users, recording the consent timestamp and version number. The response classification submodule performs text cleaning, sentiment polarity assessment, and evidence matching on user responses, categorizing them into three types: explicit rebuttal / confirmation and ambiguous acceptance. The classification results are mapped to trust evidence and written into the audit ledger. The vetoer selection submodule selects three vetoers from the vetoer pool based on topic tag complete matching, independence certificate verification, and a recent contact record interval greater than 30 days. The review interface provides observation chain trust judgment evidence and adversarial responses as review materials. In the review interface, vetoers choose to agree / disagree or propose alternative actions in writing. When disagreeing, they submit an executable alternative and a commitment to assume joint liability. Disagreement triggers a locked review process, requiring the reviewer and vetoer to submit review conclusions. The node classification execution submodule receives evidence from the preset decision network; when the number of sources is not less than 10 and the consistency score is not less than 0.9 or the connectivity index exceeds 0.75, a classification trigger signal is issued. The system automatically generates decision summaries based on promotion and isolation path rules. The manual review submodule submits the decision summary to a designated independent supervisor within a 60-hour window. The supervisor makes a written acceptance or rollback decision and records the review opinion and signatory identifier in the audit ledger. If the manual review is not completed within 60 hours, the manual review submodule automatically freezes access permissions for relevant nodes and sends a notification to the audit and compliance governance module to initiate subsequent review procedures. An emergency rescue exception is triggered when the risk identification unit calculates an immediate risk score greater than or equal to 0.95 or any single risk signal reaches severity level 3. This triggers a temporary observation path, allowing manual hotline or controlled manual observation as qualified observation, and grants limited access for a default 24-hour duration. Any extension must be approved in writing by at least three independent reviewers and accompanied by a signed joint liability commitment. All trigger evidence observation text trust judgment records, rejection judgment classification decisions, and review conclusions are recorded in the audit ledger as unalterable event records to achieve a complete and traceable chain of responsibility and a closed-loop rectification process.

[0019] Example 2: This embodiment of the invention provides an artificial intelligence-based psychological counseling system and method, including: the entire process of collaborative operation of various modules of the system and embedding specific technical implementation details in a complete psychological counseling session; when a user initiates a counseling request for emotion management through a mobile application client, the system first completes two-way identity authentication with the client certificate through the OAuth2.0 protocol, and assigns a globally unique session identifier generated based on the UUIDv4 algorithm to the session; during the counseling process, all user interaction text, speech-to-text data, and interaction metadata are transmitted in real time to the backend processing cluster through a TLS1.3 encrypted channel, and are analyzed in real time, with emotion vector calculation and topic tagging performed by a lightweight stream processing engine. The emotion analysis adopts a fine-tuned version based on the RoBERTa model, outputting a real-time emotion vector containing dimensions such as pleasure, excitement, and anxiety; when the system determines that the session has ended through the session silence detection algorithm or when the user clicks the "end" button, the session end detection unit records a Coordinated Universal Time timestamp accurate to nanoseconds and triggers the core process of the session node management module; the node generation unit then starts, first calling a pre-trained text summarization model to analyze all data in this session. The interactive content generates a 256-dimensional semantic vector summary, extracting key entities and topic tags. This information, along with session identifiers, end timestamps, and initial state "candidate states," is then combined to form a structured data object conforming to the Apache Avro pattern. This object is then submitted to an Apache Kafka-based event bus, where the consumer service writes it as a "creation event" into the system's core immutable audit ledger. This is implemented using a permissioned blockchain network, specifically a private chain based on the Hyperledger Fabric architecture. Each organization acts as a peer node. Events are encapsulated as transactions via chaincode, verified through an endorsement policy, sorted, and written into blocks, thus ensuring immutability. After successful writing, the node is placed in a Redis cache candidate pool with a 72-hour lifespan and begins waiting for observation evidence. Almost simultaneously, the observation management module begins operation. Its observation eligibility determination submodule listens for user confirmation interactions from clients. For example, when a user clicks "confirm" on the satisfaction confirmation pop-up pushed by the system after a consultation, this click event, along with its accompanying security token, timestamp, and device fingerprint information, is encapsulated into a JSON request and sent to the backend.The backend verification service first verifies the JWT signature of the token to confirm that it comes from a trusted client and has not expired. Then, it calls the face recognition microservice to perform liveness verification, ensuring that the operation is performed by the user. Finally, it uses a graph neural network model to calculate the relevance of the contextual semantics of this click action to the current session topic. When the comprehensive verification passes and the relevance score is greater than 0.85, the interaction is marked as a "qualified observation," and its complete data is serialized into Protocol Buffers format and written to the blockchain ledger via Fabric chaincode calls, forming an observation record. The observation chain management submodule continuously scans the ledger, querying a state database to collect attributes... All observation records within the same session node are aggregated and analyzed according to preset rules. A trust determination requires at least three observations, originating from at least two distinct and independent verification channels: "in-app clicks," "biometric verification," and "SMS verification code replies." When the rules are met, the observation chain management submodule generates an observation chain integrity proof—a root hash calculated using a Merkle tree from the hash values ​​of multiple observation records—and publishes this proof along with the trigger signal to the event bus. The trust determination submodule of the trust state management module subscribes to this, and upon triggering, loads all observation data and the original session content from the cache and ledger. The determination submodule then... A hybrid decision engine based on rules and statistics is run. The engine first checks the independence of the observation chain, verifying that the IP address, device ID, and user account of each observation source do not overlap. Then it checks diversity, confirming that the interaction types cover at least two preset types. Finally, it performs the core trust score calculation. The calculation model is a gradient boosting decision tree model, whose input features include the number of observations, source independence, interaction type entropy, user historical trust baseline, and sentiment stability score of the conversation content. The model outputs a score between 0 and 100. If the score is greater than or equal to 80 and the observation chain fully meets the hard requirements for independence and diversity, the system determines it to be "trustworthy." The results, score details, and the Merkle root hash of the observation chain used are packaged together into a single transaction, signed by the decision service's private key, and submitted to the blockchain ledger for permanent storage. If the transaction is deemed "untrustworthy," the archiving submodule is triggered, migrating all data from the cache to an object storage service configured with a WORM policy, i.e., write-once, read-many. For example, a storage bucket compatible with the S3 interface is used with object locking enabled. The access key for the stored object is split into three parts using the Shamir secret sharing algorithm and distributed to three independent authorization management services. Only by obtaining more than two key fragments can the access key be reconstructed, thus achieving a technical "uncopyable" nature.During the window period when a node is in the candidate state, the adversarial narrative management module is also launched in parallel. Its anti-narrative generation submodule calls a large language model that has been rigorously trained with ethical alignment, such as a GPT-NeoX variant fine-tuned on a specific psychological counseling corpus. The cue words are carefully designed to strictly exclude leading and dangerous wording. After receiving the conversation summary and emotion vector, the model generates anti-narrative candidate sentences such as "You just mentioned feeling a lot of stress. Could this also mean that you are in a period that needs adjustment?" After generation, the ethics compliance submodule uses a filtering engine containing a list of sensitive words and regular expression matching, combined with a lightweight risk classifier, to score the candidate sentences. Only sentences with a score higher than 9 are allowed to pass. Only sentences with a value of 0 are allowed to pass; subsequently, the system presents a counter-narrative to the user via an in-app notification in the form of a non-modal pop-up, forcing the user to make an explicit choice of "informed and continue" or "skip," and the selection action along with the timestamp is immediately recorded; any text reply from the user is processed by the response classification submodule, which uses a BERT-based text classification model to quickly classify the reply into three categories: "explicit rebuttal," "confirmation," or "vague acceptance," and the classification results are written into the ledger as important qualitative evidence; after the system has run for a period of time and accumulated enough nodes and evidence, the self-organizing governance module begins to play its role; its critical decision submodule continuously retrieves data from the state database of the blockchain ledger through a G... The raphQL interface queries and retrieves relevant events, maintaining an in-memory knowledge graph. The Neo4j graph database stores the relationships between nodes and evidence. The system calculates graph metrics every five minutes. A critical condition is reached when the number of evidence nodes associated with a user or topic exceeds 10 within 24 hours, and the subgraph consistency score calculated using the graph embedding algorithm is greater than 0.9. At this point, the node classification execution submodule is triggered. Based on a preset YAML format criterion file, it performs batch automated classification of relevant nodes. The classification logic itself is a decision tree script, marking nodes as "promoted" or "isolated," generating a classification report, which is then stored in the ledger and triggers human intervention. The manual review process involves the following steps: The manual review submodule sends the report to the designated supervisor's to-do list via an internal workflow system. The supervisor must review the report within 60 hours through a secure management backend built with React. All actions taken, including clicking the "Accept" or "Rollback" button, are captured by the front-end SDK and automatically signed with their digital certificate, submitted as a transaction to the blockchain. For nodes marked as "promoted" and passing the review, the asymmetric veto and permission module intervenes. The vetoor selection submodule selects three experts from an Elasticsearch index containing expert information, using a verifiable random function algorithm, who are topic-matched and have no recent conflicts of interest.The system sends links to the review materials via encrypted email. Rejectors access the materials through a dedicated portal and make their decisions in a smart contract-driven review interface. If any rejector clicks "oppose," a form pops up requiring a structured alternative solution; once submitted, it's locked, triggering a review process requiring endorsement from at least two other experts. All rejections and reviews are wallet-signed via a browser plugin similar to MetaMask, ensuring non-repudiation and recording the transactions on-chain. Finally, the event logging submodule of the audit and compliance governance module, actually composed of the blockchain network itself and an off-chain monitoring log system, ensures all events are traceable. One omission; its review entry submodule provides regulatory agencies with a blockchain explorer customized based on Hyperledger Explorer. Regulators can log in via multi-factor authentication and retrieve all historical records by block, transaction ID, or custom index. The rectification trigger submodule is a continuously running monitoring daemon that periodically scans on-chain data and compares it with a predefined compliance rule base. Once anomalies such as invalid signatures or unreviewed timeouts are detected, it automatically calls the workflow engine's API to initiate a complete rectification process that includes task allocation, time limit reminders, and result verification. Each step of the process is then anchored back to the blockchain, thus forming a completely closed-loop, transparent, and verifiable system from technical implementation to compliance supervision.

[0020] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A psychological counseling system based on artificial intelligence, characterized in that, include: The session node management module is used to generate a session node according to preset rules at the end of each consultation session, initialize the session node to a candidate state, and write the creation event of the session node into an immutable audit ledger. The trust state management module is used to create a trust state identifier for the session node when it is generated, place the trust state identifier in the candidate state, and trigger a trust determination procedure after receiving a qualified observation chain from the observation management module. After the determination, the session node is marked as trustworthy or untrustworthy, and the determination result and the observation chain on which the determination is based are written into an immutable audit ledger. The observation management module is used to receive and verify only interactions explicitly initiated by the user as qualified observations, forming an observation chain to trigger the trust determination procedure, performing a qualification review on each observation, performing a post-event multi-party review process on non-routine and emergency observations, and recording the review conclusions. The adversarial narrative management module is used to automatically generate a counter-narrative according to ethical rules when the session node is in the candidate state, present the counter-narrative to the user after obtaining the user's explicit informed consent, record the user's structured response to the counter-narrative, and write the response into an immutable audit ledger. It also includes a temporary unlocking process for emergency rescue exceptions, which is used to trigger a temporary observation path when an immediate risk signal is identified. The self-organizing governance module receives judgment evidence from the aforementioned modules and automatically triggers a node classification procedure when preset critical conditions are met within the system. It can branch a group of session nodes into promotion paths or isolation paths according to preset criteria without initial human intervention. After classification, it opens for manual review within a limited window and confirms or rolls back the classification results based on the review conclusion after the manual review is completed. The asymmetric veto and permission module is used to select vetoers with topic matching and independence qualifications according to preset rules before the session node is to be promoted to long-term use. It provides the vetoers with observation chains, trust judgment evidence, and adversarial responses as review materials and records the veto and alternative handling process. The audit and compliance governance module is used to write all events of the generation of the session node, the reception and verification of the observation, the trust determination, the adversarial presentation and user response, and the node classification and veto process into an immutable audit ledger. It provides searchable review entry points to regulators or designated supervisors according to their permissions, and triggers the accountability and rectification process when the audit finds that the rules have been violated.

2. The artificial intelligence-based psychological counseling system according to claim 1, characterized in that: Among the observations The management module includes an observation eligibility determination submodule and an observation chain management submodule; The observation qualification determination submodule is used to clearly define the interaction types that constitute qualified observations according to preset rules as explicit written confirmation from users, explicit verbal confirmation from users, explicit click confirmation on the user interface, and controlled identity verification. When any of the above interactions are received, the source, timing and context of the interaction are independently verified, and after the verification is passed, the interaction is written as a qualified observation in the form of an auditable record and the trust determination procedure is triggered. The observation chain management submodule is used to require qualified observations to consist of observation sequences from different sources and with different interaction types according to preset rules, maintain the integrity of the observation sequences according to preset rules, allow the trust determination when the observation sequences meet the independence requirements of the preset rules, and mark the session nodes as unstable in the audit ledger when the observation sequences do not meet the requirements of the preset rules, and move the session nodes into the unstable pool for supplementary evidence.

3. The artificial intelligence-based psychological counseling system according to claim 1, characterized in that: The adversarial narrative management module includes a counter-narrative generation submodule, an ethical compliance submodule, and a response classification submodule. The anti-narrative generation submodule is used to generate a set of anti-narrative candidates that do not contain leading wording or introduce dangerous details based on the semantic elements and interaction context of the session node, and to record the generation trajectory; the ethics compliance submodule is used to perform a compliance check on the anti-narrative before it is put into user interaction, and to publish the anti-narrative to the user in a non-intrusive and concise manner after compliance is passed, and to display an exit path when it is published. The response classification submodule is used to structure the user's response to the counter-narrative into three categories: explicit rebuttal, confirmation, or ambiguous acceptance. The classification results are mapped to trust evidence according to preset rules, and the response and the trust evidence are recorded in the audit ledger.

4. The artificial intelligence-based psychological counseling system according to claim 1, characterized in that: Trust The status management module includes a trust status identifier generation submodule, a trust determination submodule, and an archiving submodule; The trust status identifier generation submodule is used to create a unique and invisible trust status identifier according to preset rules when the session node is generated, place the session node in the candidate state, and write the creation event into the audit ledger. The trust determination submodule is used to perform trust determination in the audit record according to the preset determination table after receiving a qualified observation sequence from the observation chain management submodule and passing the observation qualification review according to preset rules. The trust determination result is written to the session node with a preset label of trustworthy or untrustworthy, and the determination basis and the observation chain are recorded in the audit ledger. The archiving submodule is used to migrate the session node to a read-only, non-replicable archiving pool according to preset rules when the trust determination result is untrustworthy. It restricts access to the archive by any regular training or profiling module based on preset rules, and requires the recovery of the archive to be performed through a multi-party contract-based authorization procedure based on preset rules. The authorization process and authorization results are recorded in the audit ledger.

5. The artificial intelligence-based psychological counseling system according to claim 1, characterized in that: The self-organizing governance module includes a criticality determination submodule, a node classification and execution submodule, and a manual review submodule. The critical decision submodule is used to continuously collect and summarize evidence from the session node management module, trust state management module, observation management module, and adversarial narrative management module in a decision network with preset rules. When a preset critical connectivity or consistency condition is reached within the system, a node classification trigger signal is issued. The node classification execution submodule is used to automatically divert the involved session nodes into promotion paths or isolation paths according to preset rules after receiving the trigger signal, generate a node classification decision summary, and write the decision summary into the audit ledger. The manual review submodule is used to submit the node classification decision summary to a designated independent supervisor for review within a window defined by preset rules. After the supervisor makes an acceptance or rollback decision, the final state is updated in the audit ledger according to the supervisor's decision, and the review opinion is recorded.

6. The artificial intelligence-based psychological counseling system according to claim 1, characterized in that: The asymmetric veto and approval module includes a vetoer selection submodule, a review material provision submodule, and a judgment processing submodule; the vetoer selection submodule is used to select vetoers with the qualifications of topic matching and independence according to preset rules, and to avoid conflicts of interest; The review materials providing submodule is used to provide the vetoer with an observation chain, trust determination evidence, and adversarial responses as review materials; The judgment processing submodule is used to require the vetoer to make a written judgment of agreement, objection, or proposed alternative disposal in a preset review interface. When the vetoer raises an objection, the locking review process is automatically triggered, requiring the vetoer to submit an executable alternative disposal plan along with the reasons for the objection, and to assume joint liability for the alternative plan. The veto behavior, the alternative plan, and the review result are recorded in the audit ledger. For vetoers who raise more than a threshold number of objections within a preset time or whose proposed alternative plan fails the review, a supervision and retraining procedure is triggered according to preset rules.

7. The artificial intelligence-based psychological counseling system according to claim 1, characterized in that: The audit and compliance governance module includes an event logging submodule, an audit entry submodule, and a rectification triggering submodule. The event logging submodule records all events related to the generation of the session node, the reception and verification of observations, the trust determination, the adversarial presentation and user response, and the node classification and veto process in an immutable audit ledger. The audit entry submodule provides a searchable audit entry to regulators or designated supervisors according to access control. The rectification triggering submodule triggers the accountability and rectification process when the audit finds a rule violation and writes the rectification record into the audit ledger.

8. The artificial intelligence-based psychological counseling system according to claim 1, characterized in that: Among them emergency The temporary unlocking process for rescue exceptions triggers a temporary observation pathway when an immediate risk signal defined by preset rules is detected; The temporary observation channel allows manual hotline or controlled manual observation as a qualified observation to trigger the trust determination, immediately granting limited access permissions to implement intervention, and subsequently, within a time limit defined by preset rules, multiple independent reviewers confirm in writing in the review interface whether to maintain the node access permissions caused by the intervention or roll back the access, and write the review conclusion and responsibility record into an immutable audit ledger.

9. The method of an artificial intelligence-based psychological counseling system according to claim 1, characterized in that, Includes the following steps: The session node management steps are as follows: at the end of each consultation session, a session node is generated according to a preset rule, the session node is initialized to a candidate state, and the creation event of the session node is written into an immutable audit ledger. The trust state management step creates a trust state identifier when the session node is generated, places the trust state identifier in the candidate state, and triggers a trust determination procedure after receiving a qualified observation chain from the observation management step. After the determination, the session node is marked as trustworthy or untrustworthy, and the determination result and the observation chain on which the determination is based are written into an immutable audit ledger. The observation management steps involve receiving and verifying only interactions explicitly initiated by the user as qualified observations, forming an observation chain to trigger the trust determination procedure, performing a qualification review on each observation, performing a post-event multi-party review process for unconventional and emergency observations, and recording the review conclusions. The adversarial narrative management steps involve automatically generating a counter-narrative according to ethical rules when the session node is in a candidate state, presenting the counter-narrative to the user after obtaining the user's explicit informed consent, recording the user's structured response to the counter-narrative, and writing the response into an immutable audit ledger. It also includes a temporary unlocking process for emergency rescue exceptions, which is used to trigger a temporary observation path when an immediate risk signal is identified. The self-organizing governance step involves receiving judgment evidence from the aforementioned steps and automatically triggering a node classification procedure when a preset critical condition is reached within the system. This procedure can branch a group of session nodes into promotion or isolation paths according to preset criteria without initial human intervention. After classification, manual review is opened within a limited window, and the classification results are confirmed or rolled back based on the review conclusion after the manual review is completed. The asymmetric veto and permission step involves selecting vetoers with topic matching and independence qualifications according to preset rules before the session node is to be promoted to long-term use. The vetoers are provided with observation chains, trust judgment evidence, and adversarial responses as review materials, and the veto and alternative handling process is recorded. The audit and compliance governance steps involve writing all events of the session node generation, observation reception and verification, trust determination, adversarial presentation and user response, node classification and veto process into an immutable audit ledger, providing searchable review entry points to regulators or designated supervisors according to permissions, and triggering accountability and rectification processes when audits discover rule violations.

Images