CN-121981409-A - Labor education collaborative evaluation method based on large-model intelligent body recognition game

CN121981409ACN 121981409 ACN121981409 ACN 121981409ACN-121981409-A

Abstract

The invention provides a labor education collaborative evaluation method based on a large-model intelligent body recognition game, and belongs to the technical field of education evaluation. Firstly, sensor data is converted into labor semantic terms through multi-mode signal decoupling and feature construction and semantic alignment. Secondly, building a four-dimensional literacy state vector, and solving Nash equilibrium strategies and dynamic contribution rates of schools, families and social parties. Then, constructing a multi-mode large model intelligent body, integrating the information to generate a cognitive state vector, deducing and outputting a prediction result through a thinking chain, and calculating uncertainty. And finally, establishing a self-adaptive intervention system, generating intervention suggestions for the rational anchor points by using Nash equilibrium, checking the alignment degree, and carrying out decision distribution according to uncertainty and the alignment degree to form a collaborative feedback closed loop. The method realizes full-link modeling from physical behaviors to educational evaluation, and remarkably improves the accuracy and the interpretability of the evaluation.

Inventors

PAN HONG
ZHANG HUAYING
GUO JINGWEI
Ren Mengshi
WANG YU

Assignees

山东劳动职业技术学院（山东劳动技师学院）

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (10)

1. A labor education collaborative evaluation method based on large-model intelligent recognition game is characterized by comprising the following steps: the acquired multi-modal sensor signals are adaptively decoupled through a skill eigenmode, dynamic phase space feature construction is aligned with the manifold projection of the plum clusters and cross-modal semantic mutual information, and the acquired multi-modal sensor signals are converted into discrete labor semantic word elements; establishing a four-dimensional literacy state vector driven by physical feedback by taking the labor semantic word elements as state observation variables, solving Nash equilibrium strategies of three-party education subjects of schools, families and society, and calculating dynamic contribution rates of all parties based on inverse fact deduction; Constructing a multi-mode large-model intelligent body, constructing a cognitive state vector based on the labor semantic word elements, the Nash equilibrium strategy, the dynamic contribution rate and the four-dimensional literacy state vector, generating a prediction result by deduction of a thinking chain, and calculating prediction uncertainty; And constructing a multi-mode large model intelligent body self-adaptive intervention system based on value alignment, taking the Nash equilibrium strategy as a rational anchor point, performing intervention suggestion generation and alignment degree check on the prediction result, performing self-adaptive decision distribution based on prediction uncertainty and strategy alignment degree, and forming a collaborative feedback closed loop.
2. The method according to claim 1, characterized in that the skill eigenmode adaptive decoupling specifically comprises: Constructing a variation optimization objective function; converting the variation optimization objective function into an augmented Lagrangian function, and carrying out iterative solution in a frequency domain by using an alternate direction multiplier method; After iterative convergence, the final spectrum is then obtained Performing inverse Fourier transform to obtain pure time domain eigenmodes ; The dynamic phase space feature structure and the lie group manifold projection specifically comprise: Calculating physical eigenvectors including operational stability and energy efficiency in dynamic phase space; deep features are extracted by utilizing a Leu convolution network, feature points on a curved manifold are projected to a tangent space through Leu logarithmic mapping operators, and geometric alignment features are obtained ; The cross-modal semantic mutual information alignment specifically comprises: According to geometrically-aligned features Constructing a contrast alignment loss function based on a skill topology mask, and training a projection matrix of a physical feature space to a public hidden space by minimizing the contrast alignment loss function, wherein the contrast alignment loss function is as follows: , Wherein, the Representing contrast learning alignment loss; Representing mathematical expectations; A projection matrix representing the physical feature space into a public hidden space; A projection matrix representing a semantic feature space into a public hidden space; Semantic vectors representing standard process description texts corresponding to the current actions; representing a cosine similarity function; Representing the temperature coefficient; represent the first Semantic vectors of the negative samples; Is an indication function; A set of topology masks for skills; Acquiring aligned features based on projection matrix from physical feature space to public hidden space ; And using an adaptive vector quantizer to align the features Mapping into discrete labor semantic tokens 。
3. The method according to claim 2, characterized in that the four-dimensional state vector comprises in particular: Defining a physical feedback mapping function, and converting discrete labor semantic terms into instantaneous correction vectors for literacy states, wherein the physical feedback mapping function is as follows: , Wherein, the Representative time of day Is used for correcting the vector by physical feedback; is a pre-trained high-dimensional mapping matrix; is the single-hot coding representation of the labor semantic word elements; is the current system time; The specific moment when the physical action occurs; Is a memory attenuation coefficient; and integrating the physical feedback mapping function and the collaborative tensor term into an Eyew random differential equation to obtain a four-dimensional literacy state vector.
4. A method according to claim 3, characterized in that solving the nash equalization strategy specifically comprises: Constructing a full-period accumulated value functional of each education subject; introducing a value function, solving an optimal feedback control strategy by utilizing a Hamiltonian-Jacobian-Bellman-Isaacs equation to obtain a Nash equilibrium strategy constrained by physical feedback and synergistic effect 。
5. The method of claim 4, wherein calculating the dynamic contribution rate of each party based on the counterfactual deduction specifically comprises: Nash-based equalization strategy Defining a feature utility function of the sub-coalition: , Wherein, the As a characteristic utility function, integration interval Is the current evaluation window; representing any child federation that does not contain a principal; Representing norms of the vectors; representing the linear contributions of members within the federation; representing the synergistic effect produced by the members within the federation; To represent an integral intermediate variable, expressed in an evaluation time window; Represents the first A single-sided gain matrix of the individual bodies, Represents the natural drift term of the human body, Is a state transition matrix; is a literacy state vector of the device, Represents the intra-alliance item The individual bodies are at the moment Is set to be the most suitable strategy control vector, Representing a physical feedback term; Weighting and summing all possible sub-alliance combinations by using a Shapley formula, and calculating expected values of marginal contributions of each party to obtain Time main body Dynamic contribution ratio of (2) 。
6. The method of claim 5, wherein constructing a cognitive state vector based on the labor semantic terms, nash equalization strategy, dynamic contribution rate, four-dimensional literacy state vector specifically comprises: Mapping Nash equilibrium strategy and dynamic contribution rate to semantic hidden space of large model through a learning projection network, aligning with the labor semantic word elements on the same manifold, and constructing initial cognitive state vector : , Wherein, the Is a layer normalization operation; Is a physical mode projection matrix; the physical perception result of the labor semantic word is called; is a mathematical modal projection matrix; Is a multi-layer perceptron; Is a four-dimensional state vector that is a vector of states, Is the optimal strategy under the nash equilibrium, Is the dynamic contribution rate; coding for a position; First, the Layer cognitive state vector The update formula is as follows: , Wherein, the Represents a standard self-attention mechanism function; Respectively the first Layer query, key, value matrix; As a learnable saliency weighting matrix, The function normalizes the weighting values to The interval is used as a gating signal, Representing element-by-element multiplication; Is a residual connection.
7. The method of claim 6, wherein the generating the prediction result by the multi-modal large model agent through mental chain deduction specifically comprises: inputting the cognitive state vector into a large intelligent agent, forcing the large intelligent agent to execute autoregressive thinking chain deduction, and generating a word element sequence comprising an intermediate reasoning step; the multi-mode large-model intelligent body introduces a logic consistency check loss rewarding optimization objective function, and causal logic in a constraint thinking chain and mathematical trend of a literacy state evolution equation are kept consistent The following are provided: , Wherein, the Is the total training loss function of the multi-mode large-model intelligent body; Is a standard cross entropy generation penalty; Is the final predictive label; is a weight coefficient of the logical constraint; extracting emotion tendency scores of the thinking chain text by using an emotion analysis model; Is a generated sequence of thought chain text; Is the true literacy change direction, Representing the mathematically true amount of change.
8. The method of claim 7, wherein the computational prediction uncertainty calculation formula is as follows: , Wherein, the Is a cognitive uncertainty score; The number of Ensemble samples; represent the first Model parameter state at sub-sampling; Is the first A prediction distribution obtained by sub-sampling; Is that An average distribution of sub-samples; Is KL divergence.
9. The method of claim 8, wherein generating and aligning a verification specification by taking the nash equalization strategy as a rational anchor for intervention advice comprises: Injecting the Nash equilibrium strategy into a generation prompt of a large model, constructing a number-document mapping generation equation, and generating an intervention proposal manuscript; And calculating cosine similarity of the text embedding of the intervention proposal manuscript and the Nash equilibrium policy vector in a semantic space as an alignment degree score.
10. The method of claim 1, wherein the adaptively distributing decisions based on the prediction uncertainty and the policy alignment specifically comprises automatically issuing a dry pre-instruction when the cognitive uncertainty score is below a safety upper limit and the policy alignment is above a safety lower limit; The forming the cooperative feedback loop specifically includes: collecting a new round of labor semantic word elements after the intervention instruction is executed, calculating the real increment of the literacy state, updating the cognitive parameters of the multi-mode large model intelligent body according to the real increment by utilizing the reinforcement learning idea, and reinforcing or punishing the corresponding intervention strategy generation tendency.

Description

Labor education collaborative evaluation method based on large-model intelligent body recognition game Technical Field The invention relates to a labor education collaborative evaluation method based on a large-model intelligent body recognition game, and belongs to the technical field of education evaluation. Background With the continuous development of education ideas, labor education is an important link for cultivating students' overall literacy, and the scientificity and effectiveness of an evaluation system are increasingly focused. Labor education has been transformed from a traditional single training mode to a new "234" personal care mode of "in-school + out-of-school bilinear, family + school + social three-party collaboration". In the complex educational ecosystem, the labor practice scene of students is greatly widened, and massive associated data with different forms are generated. The data show remarkable multi-mode heterogeneous characteristics on physical properties, namely physical sensor signals (such as current load waveforms of a digital processing center and voltage time sequence data of a welding robot) and monitoring video streams with high frequency, continuous and often mixed environmental noise are generated in an internal calibration training link, and low-frequency discrete text logs and subjective evaluation data are generated in an external calibration enterprise training and home labor link. The physical-semantic isomerism of the data causes a huge cognitive gap, namely, although the action waveform data at the bottom layer is objective, the action waveform data are dumb data, and the action waveform data lack of semantic labels, so that the action waveform data are difficult to map directly to higher-order literacy indexes such as craftsman spirit, ecological consciousness and the like, and an evaluation result often flows in a form, and the skill internalization degree cannot be deeply explored. The existing labor education evaluation technical means mainly depend on static statistical analysis or independent single-mode machine learning models, and two major core technical pain points exist. Firstly, a deep alignment mechanism of physical behaviors and mental literacy is lacking, the existing algorithm cannot adaptively extract intrinsic characteristics capable of representing the skill level of students from non-steady training signals, and normal process adjustment in the operation of the students is often misjudged as illegal operation, or fine action characteristics conforming to standard process (SOP) cannot be identified. Secondly, the problem of black box of a collaborative people raising mechanism is that under the mode of 234, the teaching strategy of a school, the emotion support of a family and the post guidance of an enterprise are not simply and linearly overlapped, but are a complex dynamic game system. The strategies of the education subjects are selected to be interwoven and cause each other, the traditional linear regression evaluation model can not simulate the nonlinear multi-agent interaction process, and the attribution difficult problem that ' whether the literacy promotion of students benefits from the investment of which party ' can not be accurately answered ', so that the education resource allocation is unbalanced. In addition, while Large Language Models (LLMs) exhibit strong capabilities in terms of natural language processing and logical reasoning, providing new opportunities for educational evaluation, the direct application thereof to the field of craftsman labor education remains a challenge. The general base large model lacks domain sensing capability for specific industrial processes (such as turning parameters and welding current waveforms), and direct input of raw sensing data often results in "phantom" evaluation that the model output deviates from the process standard. Meanwhile, when the big model processes literacy evaluation of long-period continuous evolution, an effective state memory and evolution reasoning mechanism is lacked. Therefore, there is an urgent need to develop an intelligent system capable of deeply decoupling the semantics of physical behaviors, dynamically deducting the multi-subject collaborative gaming process, and generating an adaptive intervention strategy so as to realize precise quantification and scientific decision on the whole process of labor education. Disclosure of Invention The invention aims to provide a labor education collaborative evaluation method based on large-model intelligent recognition game, which solves the technical problems of lack of a cross-modal feature mapping mechanism of physical consistency constraint, lack of a dynamic collaborative evolution model based on differential game, lack of a literacy evolution prediction mechanism of long time sequence memory and causal deduction capability of a traditional model, existence of machine illusion and violation of education ethics specifi