CN-121981147-A - Digital twin multistage memory management system and method based on emotion perception and value quantification

CN121981147ACN 121981147 ACN121981147 ACN 121981147ACN-121981147-A

Abstract

The invention discloses a digital twin multi-level memory management system and method based on emotion perception and value quantification. The system actively builds a user file through face recognition, combines real-time expression analysis to adjust dialogue emotion strategies, utilizes a fine-tuned evaluation small model to score emotion labels, emotion intensity and knowledge value, adopts a dynamic attenuation algorithm simulating human brain forgetting rules for dialogues with low scores, and is based on a formula And carrying out exponential dynamic attenuation on the memory, and realizing hierarchical management from the storage of the original data to the compression of the original data into a detailed abstract, the compression of the original data into a key core sentence and the physical deletion. The invention realizes that the digital twin has the memory capacity similar to human, and remarkably improves the interactive anthropomorphic degree of the dialogue.

Inventors

SU XUN
CUI YINGXIN
ZHANG JIALE
LIANG YIJIE

Assignees

佛山市南海区粤良友装饰材料商行

Dates

Publication Date: 20260505
Application Date: 20260123

Claims (8)

1. A digital twin person multi-level memory management method based on emotion perception and value quantification is characterized by comprising the following steps: the method comprises the steps of S1, carrying out identification and file construction, namely acquiring facial features of a user through a camera to carry out face recognition and matching the facial features of the user in a database, and if the face recognition module identifies a new user, feeding back a large language model through a prompt word by a system, triggering a function call function by the large language model to capture facial images of the user and establishing an associated file; S2, real-time emotion perception interaction, namely analyzing facial expressions of a user in real time in a dialogue process to obtain emotion labels and emotion intensities, and inputting the emotion labels and the emotion intensities as prompt words into a large language model to adjust a dynamic reply strategy of a digital twin; S3, quantitatively evaluating the dialogue value, namely, presetting a language model with smaller parameter quantity, enabling the model to score 0-100 points based on emotion labels, emotion intensities and dialogue values of the dialogue, naming the model as an evaluation small model, and scoring importance of the dialogue content by combining the emotion labels and the emotion intensities through the pre-fine-tuned evaluation small model after the dialogue is ended, wherein the score is I (I is more than or equal to 0 and less than or equal to 100); And S4, multilevel storage management, namely storing dialogue data into a short-term cache area, a long-term permanent storage area or a dynamic attenuation area respectively according to the time span t and the importance score I of the occurrence of the dialogue, and executing corresponding operation.
2. The method for managing digital twin persons in multilevel memory based on emotion perception and value quantification according to claim 1, wherein in the step of identity recognition, for a new user, the digital twin persons guide the user to introduce themselves through interactive dialogue, and meanwhile, a visual recognition tool is utilized to automatically capture image features and correlate attribute fields of the user, so that a personalized knowledge base is automatically generated.
3. The method for multi-level memory management of a digital twin based on emotion perception and value quantification as set forth in claim 1, wherein in the step of multi-level memory management, for conversations having a time span t within a preset short-term threshold t1 from which a conversation occurs, namely The system completely retains the original dialogue corpus data.
4. The method for multi-level memory management of a digital twin based on emotion perception and value quantification as set forth in claim 1, wherein in the step of multi-level memory management, for conversations having a time span t from the occurrence of a conversation exceeding a preset short-term threshold t1, the conversations are The following classification strategy was employed: when the importance score I is more than or equal to 90, permanently preserving the original dialogue data; When the I is more than or equal to 80 and less than or equal to 90, the original dialogue is compressed into a detailed abstract, and only the detailed abstract data is permanently reserved; when I < 80, a dynamic decay memory mode is entered.
5. The digital twin multi-level memory management method based on emotion perception and value quantification of claim 4, wherein the dynamic decay storage mode follows a biological heuristic forgetting algorithm, and the calculation formula of the retention score S is as follows: I is an importance score, t is a time span from the occurrence of the dialogue, and k is a preset time weight attenuation coefficient.
6. The digital twin multi-level memory management method based on emotion perception and value quantification of claim 5, wherein the dynamic attenuation storage mode is provided with a plurality of step score thresholds, the retention score S is reduced as the time span t increases, and the system sequentially judges and executes the following grading operations: When S is lower than a first preset threshold S1 for the first time, compressing the original dialogue into a detailed abstract; When the S is lower than a second preset threshold S2 for the first time, the detailed abstract is further compressed into a key core sentence; When S is below the forgetting threshold S3, the piece of data is physically deleted from the storage system.
7. The digital twin-man multi-level memory management method based on emotion perception and value quantification according to claim 1, wherein the evaluation small model is subjected to fine adjustment through a manually marked dialogue importance data set, and evaluation dimensions at least comprise emotion labels, emotion intensities and knowledge values in a dialogue.
8. A digital twin multi-level memory management system based on emotion perception and value quantification implementing the method of any of claims 1 to 7, comprising: the perception module is used for face recognition and expression analysis; the interaction module comprises a fine-tuned large language model, a digital twin-layer engine, a voice-to-text module and a text-to-voice module; the evaluation module comprises a fine-tuned evaluation small model and is used for scoring dialogue data; The storage module comprises a vector database and dynamic forgetting algorithm logic and is used for executing multi-stage storage and data cleaning logic.

Description

Digital twin multistage memory management system and method based on emotion perception and value quantification Technical Field The invention relates to the technical field of artificial intelligence and digital twin, in particular to a digital twin multi-level memory management system and method based on multi-modal emotion perception, automatic quantitative scoring of dialogue value and a simulated biological forgetting mechanism. Background With the development of large language models and multi-modal interaction technologies, digital twins have been widely used in the fields of virtual assistants, emotion accompanying, digital diversion, and the like. Existing digital twin interactive systems typically focus on speech cloning and visual image restoration, but the following limitations still exist in terms of memory management and depth of interaction: The identity perception is passive, namely the existing digital person mostly relies on the user to manually input information or to log in through a single account to identify the identity. Under the multi-user switching or off-line real-time interaction scene, the capability of actively identifying users and automatically constructing dynamic files is lacking; Interaction lacks emotion feedback-although the prior art can realize smooth dialogue, the facial expression change of the user cannot be perceived in real time. This results in the digital person's replies often being mood neutral, making it difficult to adjust the speech based on the user's immediate mood; The storage mechanism is mechanized, traditional digital human memory typically adopts a "full memory" or "sliding window forget" mode: full-scale storage can lead to high data redundancy, reduced retrieval efficiency with time and huge storage cost; The forgetting of the sliding window can mechanically discard the old dialogue, so that the digital person can forget the extremely important event happening a plurality of months ago; Lack of value ratings-existing systems cannot distinguish the importance of conversational content. For the user, a profound personal conversation is clearly different from asking for trivial things for memory value. The current system lacks a mechanism to dynamically adjust memory depth based on importance and time decay coefficients that resembles the human brain. Therefore, how to construct a digital twin system which can actively perceive identity and emotion and can perform scientific and efficient memory management like human brain according to information value and time span is a technical problem to be solved in the current industry. Disclosure of Invention Object of the Invention The invention aims to solve the problems of passive identification, interaction emotion deletion and mechanical inefficiency of a memory storage mechanism in the existing digital twin technology, and provides a digital twin multistage memory management system and method based on emotion perception and value quantification. The system can actively identify and establish the user file, adjust the dialogue strategy according to the real-time emotion of the user, simulate the human brain memory mechanism and dynamically manage mass dialogue data. Technical proposal In order to achieve the above purpose, the invention adopts the following technical scheme: A digital twin multi-level memory management method based on emotion perception and value quantification mainly comprises the following steps. The first step is that an automatic construction system for identification and archives based on multi-mode perception Firstly, capturing facial features of an interactive object in real time through a face recognition module; If the identification result matches the existing user in the database, automatically calling the personal file of the user, wherein the personal file comprises basic information, historical dialogue abstract and preference setting as the context background of the current dialogue; If the user is identified as a strange user, the system triggers a preset 'first-time meeting' interaction script, actively activates a camera snapshot instruction through the function calling capability of the large language model, inputs the face characteristics of the user, simultaneously constructs information such as the name of the user acquired in the dialogue, and automatically builds a special data file of the user. Second step, emotion perception based on facial expression recognition module In the interactive process, the system utilizes the facial expression recognition module to analyze the micro expression of the user in real time, and generates a corresponding emotion label and emotion intensity. The tag is used as a dynamic supplement of system prompt words to input a large language model. The digital twin adjusts the replying expression according to the digital words, and emotion self-adaptive interaction is realized. Third step, dialog value quantification based on the evaluat