Multi-modal education evaluation system and method based on super-automation and federal cognitive calculation

Abstract

The invention provides a multi-modal education evaluation system and method based on super-automation and federal cognitive calculation, and the system integrates multi-source heterogeneous data collection, a federal cognitive calculation engine, an automatic evaluation workflow engine and a privacy enhanced data processing architecture. And comprehensive acquisition, efficient processing, accurate evaluation and safety protection of education data are realized. Through the LSTM-Bayesian hybrid model of dynamic parameter adjustment, intelligent feature engineering, interpretability analysis, reinforcement learning strategy optimization and other technologies, the system can generate personalized learning path planning and an adaptive teaching intervention scheme, and the teaching effect and learning efficiency are improved. And meanwhile, the security and privacy of the education data are ensured by adopting privacy protection technologies such as differential privacy and homomorphic encryption.

Description

Technical Field

[0001] The present invention relates to the technical field of management systems, and in particular, to a multi-modal education evaluation system and method based on hyper-automation and federated cognitive computing. Background Art

[0002] With the rapid development of educational informatization, the collection, processing, and analysis of educational data have become important means to improve teaching quality and learning efficiency. However, traditional education evaluation systems have problems such as single data collection, low processing efficiency, inaccurate evaluation results, and insufficient privacy protection, and cannot meet the diversified, personalized, and secure needs of modern education evaluation. Therefore, a multi-modal education evaluation system and method based on hyper-automation and federated cognitive computing are proposed to solve the above problems. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a multi-modal education evaluation system and method based on hyper-automation and federated cognitive computing to solve at least the above problems.

[0004] The technical solution adopted in the first aspect of the present invention is as follows:

[0005] A multi-modal education evaluation system based on hyper-automation and federated cognitive computing, comprising:

[0006] A multi-source heterogeneous data collection module, integrating intelligent marking terminals, classroom behavior analysis units, and experimental equipment IoT sensors. The classroom behavior analysis unit integrates the OpenPose pose recognition algorithm for real-time capture of students' classroom behavior data;

[0007] A federated cognitive computing engine, containing an LSTM-Bayesian hybrid model with dynamically adjustable parameters. The hybrid model is jointly constructed by 128 LSTM layer units and 64 Bayesian network nodes, and the model parameters are updated in a time-varying manner through variational Bayesian inference;

[0008] An automated evaluation workflow engine, supporting multi-modal data fusion and generation of personalized intervention strategies. The workflow engine includes an intelligent feature engineering module, an interpretability analysis unit, and a reinforcement learning strategy optimization component based on the deep Q-network (DQN);

[0009] A privacy-enhanced data processing architecture, adopting a composite security mechanism of differential privacy and homomorphic encryption, where the differential privacy parameters are set to ε = 0.3 and δ = 1e-5, and the homomorphic encryption uses the Paillier algorithm to achieve ciphertext space parameter aggregation.

[0010] Further, the multi-source heterogeneous data collection module includes:

[0011] Unstructured data processing unit, applying a multimodal feature extractor based on the Transformer architecture, and the extractor supports joint embedded representation of text, image and sensor data;

[0012] Data quality verification component, implementing outlier detection based on the Isolation Forest algorithm, and the detection threshold is adaptively adjusted through a dynamic confidence interval with a confidence level ≥ 95%;

[0013] Time series data alignment module, using the Dynamic Time Warping (DTW) algorithm to achieve synchronization of multi-source signals, and the synchronization error is controlled within ±0.5 seconds.

[0014] Furthermore, the federated cognitive computing engine includes:

[0015] Dynamic parameter adjustment unit, realizing time-varying update of model parameters based on variational Bayesian inference, and the update formula is:

[0016] β k (t) = μ k + σ k .N(0,1).exp(-λt)

[0017] where μ k and σ k are the mean and variance parameters learned through backpropagation, and λ is the initial value of the decay coefficient, 0.05, which is used to control the rate of dynamic change of the parameters;

[0018] Federated aggregation controller, designing a contribution degree evaluation model based on the Shapley value, and the node weight allocation formula is:

[0019]

[0020] where φ is the model performance evaluation function, defined as the contribution degree of the federated node to the global model accuracy.

[0021] Furthermore, the federated cognitive computing engine also includes:

[0022] Knowledge distillation module, constructing a teacher-student model architecture to achieve cross-domain knowledge transfer, where the teacher model uses a pre-trained BERT model and the student model is a lightweight LSTM network;

[0023] Dynamic curriculum learning component, automatically adjusting the model training difficulty curve according to the teaching progress, and the training difficulty is calculated by weighting the complexity of knowledge points and the historical error rate;

[0024] Forgetting mechanism controller, applying the Elastic Weight Consolidation (EWC) algorithm to retain key parameters, and the key parameter identification criterion is that the diagonal element of the Hessian matrix is greater than the threshold 0.1.

[0025] Further, the privacy-enhanced data processing architecture includes:

[0026] Gradient perturbation unit, which injects Gaussian noise satisfying (ε,δ)-differential privacy during the federated learning upload phase, and the noise standard deviation σ = 1.2;

[0027] Secure aggregation protocol, which uses the Paillier homomorphic encryption algorithm to implement ciphertext space parameter aggregation, where the key length is 2048 bits;

[0028] Data desensitization component, which performs k-anonymization processing on sensitive fields, with the privacy protection standard of k≥5 and the record difference degree within the equivalence class being greater than 30%.

[0029] Further, the automated evaluation workflow engine includes:

[0030] Intelligent feature engineering module, which automatically generates feature cross combinations, the cross depth d satisfies 2≤d≤5, and filters the Top-10 high-order features through the greedy algorithm;

[0031] Interpretability analysis unit, which applies the SHAP value algorithm to generate an evaluation decision attribution report, and the report includes the feature contribution degree ranking and confidence interval;

[0032] Policy optimization component, which realizes online reinforcement learning of intervention strategies based on the deep Q network DQN. The DQN network structure is 256-dimensional input layer, 128-64-32-dimensional hidden layer, and 12-dimensional output layer.

[0033] Further, the policy optimization component also includes:

[0034] Causal effect evaluation module, which applies the double machine learning DML method to estimate the intervention effect, and uses random forest and linear regression as the base models;

[0035] Multi-objective optimizer, which designs a Pareto front solution algorithm to balance teaching effect and resource consumption, and the objective function weights are dynamically adjusted through the expert knowledge graph;

[0036] Real-time feedback interface, which supports dynamic adjustment of policy parameters through natural language interaction, and the interaction protocol is defined based on the JSON format.

[0037] The technical solution adopted in the second aspect of the present invention is as follows:

[0038] A multi-modal education evaluation method based on hyper-automation and federated cognitive computing. This method is executed on a multi-modal education evaluation system proposed in the first aspect, and includes steps:

[0039] S1: Implement real-time multi-modal data acquisition and preprocessing through edge computing nodes. The preprocessing includes data cleaning, normalization, and feature alignment;

[0040] S2: Conduct distributed model training under the federated learning framework, satisfying privacy protection constraints. The adaptive federated average algorithm is adopted during the training process;

[0041] S3: Apply a dynamic Bayesian network to generate personalized learning path planning. The network nodes are defined as the knowledge point mastery status, and the edge weights are the transition probabilities;

[0042] S4: Output an adaptive teaching intervention plan based on the reinforcement learning policy engine. The policy reward function is the ratio of the improvement in performance to the resource consumption;

[0043] S5: Construct a digital twin of the model performance to achieve continuous optimization of algorithm parameters. The optimization objective is to minimize the prediction error, computational time consumption, and memory occupancy.

[0044] Furthermore, step S2 specifically includes:

[0045] Adopt the adaptive federated average algorithm. The client selection probability formula is:

[0046]

[0047] where, Δ i is the local model update amplitude, defined as the L2 norm of the parameter gradient; α is the adjustment factor, and the initial value is set to 0.1;

[0048] Implement gradient compression transmission, set the sparsity rate threshold θ = 0.7, and only transmit the top 30% of the parameters in terms of the absolute value of the gradient;

[0049] Execute model watermark embedding. The watermark algorithm is based on the discrete cosine transform (DCT) to verify the authenticity of the node identity.

[0050] Furthermore, the optimization process in step S5 includes:

[0051] Establish a multi-objective optimization model for algorithm parameters:

[0052] min F(β) = [f1(β), f2(β), f3(β)] T

[0053] where, f1 is the prediction error, f2 is the computational time consumption in seconds, and f3 is the memory occupancy in MB;

[0054] Apply the NSGA-II algorithm to solve the Pareto optimal solution set. The population size is set to 100, and the number of iterations is 500;

[0055] Design a constraint generator based on a knowledge graph, where the constraints include knowledge point dependencies and syllabus compliance.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] The multi-modal education evaluation system and method based on hyper-automation and federated cognitive computing of the present invention have the following remarkable beneficial effects compared with traditional education evaluation systems:

[0058] 1. Comprehensive data collection: By integrating intelligent marking terminals, classroom behavior analysis units, and experimental equipment IoT sensors, the system can comprehensively collect multi-modal information such as students' learning behaviors, grades, and experimental data, providing rich data support for accurate evaluation.

[0059] 2. High processing efficiency: By adopting advanced technologies such as multi-modal feature extractors based on the Transformer architecture and dynamic time warping algorithms, the system can efficiently process and analyze multi-modal data, improving the accuracy and efficiency of evaluation.

[0060] 3. Accurate evaluation results: By generating personalized learning path planning through dynamic Bayesian networks and outputting adaptive teaching intervention plans based on a reinforcement learning-based policy engine, the system can accurately evaluate students' learning status and needs and provide personalized teaching services.

[0061] 4. Strong privacy protection: By adopting privacy protection technologies such as differential privacy and homomorphic encryption, the system can ensure the security and privacy of educational data and avoid data leakage and abuse.

[0062] 5. Strong ability to continuously optimize: By constructing a digital twin of model performance and a constraint generator based on a knowledge graph, the system can continuously optimize algorithm parameters, ensuring the accuracy and stability of evaluation results while conforming to teaching rules. Description of the Drawings

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only the preferred embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0064] Figure 1 is the overall structural schematic diagram of a multi-modal education evaluation system based on hyper-automation and federated cognitive computing proposed in an embodiment of the present invention.

[0065] Figure 2 is the overall process schematic diagram of a multi-modal education evaluation method based on hyper-automation and federated cognitive computing proposed in an embodiment of the present invention. Detailed implementation manners

[0066] The principles and features of the present invention will be described below in conjunction with the accompanying drawings. The listed embodiments are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0067] Embodiment 1

[0068] Referring to Figure 1 , a multi-modal education evaluation system based on hyper-automation and federated cognitive computing, comprising:

[0069] A multi-source heterogeneous data acquisition module, integrating intelligent marking terminals, classroom behavior analysis units and experimental equipment IoT sensors. The classroom behavior analysis unit integrates the OpenPose pose recognition algorithm for real-time capturing of students' classroom behavior data;

[0070] A federated cognitive computing engine, including an LSTM-Bayesian hybrid model with dynamically adjustable parameters. The hybrid model is jointly constructed by 128 LSTM layer units and 64 Bayesian network nodes, and realizes time-varying update of model parameters through variational Bayesian inference;

[0071] An automated evaluation workflow engine, supporting multi-modal data fusion and generation of personalized intervention strategies. The workflow engine includes an intelligent feature engineering module, an interpretability analysis unit and a reinforcement learning strategy optimization component based on the deep Q-network DQN;

[0072] A privacy-enhanced data processing architecture, adopting a composite security mechanism of differential privacy and homomorphic encryption, where the differential privacy parameter is set to ε = 0.3, δ = 1e-5, and the Paillier algorithm is used for homomorphic encryption to achieve ciphertext space parameter aggregation.

[0073] The multi-source heterogeneous data acquisition module includes:

[0074] An unstructured data processing unit, applying a multi-modal feature extractor based on the Transformer architecture. The extractor supports joint embedding representation of text, image and sensor data;

[0075] A data quality verification component, implementing outlier detection based on the isolation forest algorithm, and the detection threshold is adaptively adjusted through a dynamic confidence interval with a confidence level ≥ 95%;

[0076] A time series data alignment module, using the dynamic time warping DTW algorithm to achieve multi-source signal synchronization, and the synchronization error is controlled within ±0.5 seconds.

[0077] The federated cognitive computing engine includes:

[0078] The dynamic parameter adjustment unit realizes the time-varying update of model parameters based on variational Bayesian inference, and the update formula is:

[0079] β k (t) = μ k + σ k .N(0, 1).exp(-λt)

[0080] Among them, μ k 、σ k are the mean and variance parameters learned through backpropagation, λ is the initial value of the decay coefficient 0.05, which is used to control the rate of dynamic change of parameters;

[0081] The federated aggregation controller designs a contribution evaluation model based on the Shapley value, and the node weight allocation formula is:

[0082]

[0083] Among them, φ is the model performance evaluation function, which is defined as the contribution of the federated node to the global model accuracy.

[0084] The federated cognitive computing engine further includes:

[0085] The knowledge distillation module constructs a teacher-student model architecture to achieve cross-domain knowledge transfer. The teacher model uses the pre-trained BERT model, and the student model is a lightweight LSTM network;

[0086] The dynamic curriculum learning component automatically adjusts the model training difficulty curve according to the teaching progress, and the training difficulty is calculated by weighting the knowledge point complexity and the historical error rate;

[0087] The forgetting mechanism controller applies the Elastic Weight Consolidation (EWC) algorithm to retain key parameters, and the key parameter identification criterion is that the diagonal element of the Hessian matrix is greater than the threshold 0.1.

[0088] The privacy-enhanced data processing architecture includes:

[0089] The gradient perturbation unit injects Gaussian noise that satisfies (ε, δ)-differential privacy during the federated learning upload phase, and the noise standard deviation σ = 1.2;

[0090] The secure aggregation protocol uses the Paillier homomorphic encryption algorithm to achieve ciphertext space parameter aggregation, where the key length is 2048 bits;

[0091] The data desensitization component performs k-anonymization processing on sensitive fields, and the privacy protection standard is k ≥ 5, and the record difference degree within the equivalence class is greater than 30%.

[0092] The automated evaluation workflow engine includes:

[0093] An intelligent feature engineering module that automatically generates feature cross - combinations, with the cross - depth d satisfying 2 ≤ d ≤ 5, and selects the top - 10 high - order features through a greedy algorithm;

[0094] An interpretability analysis unit that applies the SHAP value algorithm to generate an evaluation decision attribution report, and the report includes the feature contribution degree ranking and confidence interval;

[0095] A policy optimization component that implements online reinforcement learning of intervention policies based on the deep Q - network (DQN). The DQN network structure has an input layer of 256 dimensions, hidden layers of 128 - 64 - 32 dimensions, and an output layer of 12 dimensions.

[0096] The policy optimization component further includes:

[0097] A causal effect evaluation module that applies the double - machine learning (DML) method to estimate the intervention effect, and uses random forest and linear regression as the base models;

[0098] A multi - objective optimizer that designs a Pareto - front - solving algorithm to balance teaching effects and resource consumption, and the weights of the objective functions are dynamically adjusted through an expert knowledge graph;

[0099] A real - time feedback interface that supports dynamic adjustment of policy parameters through natural language interaction, and the interaction protocol is defined based on the JSON format.

[0100] Exemplarily, the deployment on which the present invention depends includes: a private cloud cluster, an edge computing gateway, multi - modal acquisition devices, intelligent marking terminals. The intelligent marking terminals support OMR / OCR hybrid recognition, classroom behavior analysis cameras, and the classroom behavior analysis cameras integrate the OpenPose pose recognition algorithm;

[0101] The data governance architecture includes constructing an educational data ontology model, which contains 12 core classes (students, knowledge points, ability dimensions, etc.) and 56 object properties. Apache NiFi is used to build a data pipeline to implement ETL processing of multi - source data, establish a data quality monitoring dashboard, and abnormal data automatically triggers a re - acquisition process.

[0102] The implementation of federated cognitive computing can specifically include:

[0103] Dynamic parameter model training

[0104] Initialize the parameters of the LSTM - Bayesian hybrid model: the number of units in the LSTM layer: 128; the number of nodes in the Bayesian network: 64; the initial value of the attenuation coefficient λ: 0.05; perform variational Bayesian inference to obtain the training effect. Among them, the main code for performing variational Bayesian inference is:

[0105] class DynamicParameter(nn.Module):

[0106] def __init__(self):

[0107] super().__init__()

[0108] self.mu = nn.Parameter(torch.randn(64)) # Learnable mean parameter

[0109] self.sigma = nn.Parameter(torch.ones(64)) # Learnable variance parameter

[0110] def forward(self, t):

[0111] return self.mu + self.sigma * torch.randn(64) * torch.exp(-0.05 * t)

[0112] For the federated learning process, it can specifically include:

[0113] Steps for implementing the secure aggregation protocol: Each node encrypts the local model gradient using the Paillier algorithm; the aggregation server performs ciphertext space addition operations; after the global model is updated, it is distributed to each node; Privacy protection parameter settings: Differential privacy budget ε = 0.3, δ = 1e-5; Standard deviation of Gaussian noise σ = 1.2;

[0114] Execution of the automated evaluation workflow

[0115] The following are examples of multi-modal data fusion:

[0116] Feature engineering processing:

[0117] Data source Original feature Derived feature Exam score Math score Knowledge point mastery index (K = Σ(w_i * s_i)) Classroom video Attention duration Learning participation degree (E = 0.6A + 0.4B) Experimental data Operation duration Practical ability score (P = αT + βE)

[0118] Feature cross depth d = 3, generating high-order features such as "attention fluctuation × experimental error correlation"

[0119] Generation of personalized intervention strategies

[0120] The DQN policy network structure, where the main code of the DQN policy network structure is:

[0121]

[0122] By applying the technology of the present invention, the implementation effects of students in a certain university are specifically as follows:

[0123] 15 potential students with learning difficulties were identified (only 6 were found by traditional methods), and 42 personalized learning plans were generated. The average execution rate of the plans reached 89%, the average class score increased by 14.5 points, and the standard deviation decreased by 38%.

[0124] Example Two

[0125] Refer to Figure 2 , a multi-modal education evaluation method based on hyper-automation and federated cognitive computing, which is executed on a multi-modal education evaluation system proposed in Example One, including the steps:

[0126] S1: Implement real-time multi-modal data collection and preprocessing through edge computing nodes. The preprocessing includes data cleaning, normalization, and feature alignment;

[0127] S2: Conduct distributed model training under the federated learning framework to meet privacy protection constraints. The adaptive federated average algorithm is used in the training process;

[0128] S3: Apply a dynamic Bayesian network to generate personalized learning path planning. The network nodes are defined as the knowledge point mastery status, and the edge weights are transition probabilities;

[0129] S4: Output an adaptive teaching intervention plan based on the policy engine of reinforcement learning. The policy reward function is the ratio of the score improvement to the resource consumption;

[0130] S5: Build a digital twin of the model performance to achieve continuous optimization of algorithm parameters. The optimization goal is to minimize the prediction error, calculation time, and memory occupancy.

[0131] Step S2 specifically includes:

[0132] Using the adaptive federated average algorithm, the client selection probability formula is:

[0133]

[0134] where, Δ i is the local model update amplitude, defined as the L2 norm of the parameter gradient; α is the adjustment factor, and the initial value is set to 0.1;

[0135] Implement gradient compression transmission, set the sparsity rate threshold θ = 0.7, and only transmit the top 30% of the parameters of the absolute value of the gradient;

[0136] Execute model watermark embedding. The watermark algorithm is based on the discrete cosine transform DCT to verify the authenticity of the node identity.

[0137] The optimization process in Step S5 includes:

[0138] Establish a multi-objective optimization model for algorithm parameters:

[0139] Minimize F(β) = [f1(β), f2(β), f3(β)] T

[0140] Where f1 is the prediction error, f2 is the computing time in seconds, and f3 is the memory occupancy in MB;

[0141] Apply the NSGA-II algorithm to solve the Pareto optimal solution set, set the population size to 100, and the number of iterations to 500; design a knowledge graph-based constraint condition generator, and the constraint conditions include knowledge point dependency and syllabus compliance.

[0142] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multimodal education evaluation system based on hyper-automation and federated cognitive computing, characterized in that, Including: A multi-source heterogeneous data acquisition module, integrating intelligent marking terminals, classroom behavior analysis units, and experimental equipment IoT sensors. The classroom behavior analysis unit integrates the OpenPose pose recognition algorithm for real-time capture of students' classroom behavior data; A federated cognitive computing engine, containing an LSTM-Bayesian hybrid model with dynamically adjustable parameters. The hybrid model is jointly constructed by 128 LSTM layer units and 64 Bayesian network nodes, and the model parameters are updated in a time-varying manner through variational Bayesian inference; An automated evaluation workflow engine, supporting multi-modal data fusion and generation of personalized intervention strategies. The workflow engine includes an intelligent feature engineering module, an interpretability analysis unit, and a reinforcement learning strategy optimization component based on the deep Q-network DQN; A privacy-enhanced data processing architecture, adopting a composite security mechanism of differential privacy and homomorphic encryption. The differential privacy parameter is set to ε = 0.3, δ = 1e-5, and the Paillier algorithm is used for homomorphic encryption to achieve ciphertext space parameter aggregation.

2. The system according to claim 1, wherein The multi-source heterogeneous data acquisition module includes: An unstructured data processing unit, applying a multi-modal feature extractor based on the Transformer architecture. The extractor supports joint embedded representation of text, images, and sensor data; A data quality verification component, implementing outlier detection based on the isolation forest algorithm. The detection threshold is adaptively adjusted through a dynamic confidence interval with a confidence level ≥ 95%; A time series data alignment module, using the dynamic time warping DTW algorithm to achieve synchronization of multi-source signals, with the synchronization error controlled within ±0.5 seconds.

3. The system according to claim 1, wherein The federated cognitive computing engine contains: A dynamic parameter adjustment unit, realizing time-varying update of model parameters based on variational Bayesian inference. The update formula is: β k (t) = μ k + σ k .N(0,1).exp(-λt) Among them, μ k , σ k are the mean and variance parameters learned through backpropagation, and λ is the initial value of the decay coefficient 0.05, which is used to control the rate of dynamic change of the parameters; A federated aggregation controller, designing a contribution degree evaluation model based on the Shapley value. The node weight allocation formula is: where φ is the model performance evaluation function, defined as the contribution degree of the federated node to the global model accuracy.

4. The system according to claim 3, characterized in that The federated cognitive computing engine also includes: A knowledge distillation module, constructing a teacher-student model architecture to achieve cross-domain knowledge transfer. The teacher model uses a pre-trained BERT model, and the student model is a lightweight LSTM network; A dynamic curriculum learning component, automatically adjusting the model training difficulty curve according to the teaching progress. The training difficulty is calculated by weighting the knowledge point complexity and the historical error rate; A forgetting mechanism controller, applying the elastic weight consolidation EWC algorithm to retain key parameters. The key parameter identification criterion is that the diagonal element of the Hessian matrix is greater than the threshold 0.

1.

5. The system according to claim 1, wherein The privacy-enhanced data processing architecture includes: A gradient perturbation unit, injecting Gaussian noise that satisfies (ε,δ)-differential privacy during the federated learning upload stage. The noise standard deviation σ = 1.2; A secure aggregation protocol, using the Paillier homomorphic encryption algorithm to achieve ciphertext space parameter aggregation, where the key length is 2048 bits; A data desensitization component, performing k-anonymization processing on sensitive fields. The privacy protection standard is k ≥ 5, and the difference degree within the equivalence class is greater than 30%.

6. The system according to claim 1, wherein The automated evaluation workflow engine contains: Intelligent Feature Engineering Module, which automatically generates feature cross combinations, with the cross depth d satisfying 2 ≤ d ≤ 5, and filters the Top-10 high-order features through a greedy algorithm; Interpretability Analysis Unit, which applies the SHAP value algorithm to generate an evaluation decision attribution report, and the report includes the feature contribution degree ranking and confidence interval; Policy Optimization Component, which realizes the online reinforcement learning of intervention policies based on the Deep Q-Network (DQN). The DQN network structure is 256-dimensional for the input layer, 128-64-32-dimensional for the hidden layer, and 12-dimensional for the output layer.

7. The system according to claim 6, wherein The said Policy Optimization Component further includes: Causal Effect Evaluation Module, which applies the Double Machine Learning (DML) method to estimate the intervention effect, and uses random forest and linear regression as the base models; Multi-objective Optimizer, which designs a Pareto front solution algorithm to balance teaching effect and resource consumption, and the weights of the objective functions are dynamically adjusted through an expert knowledge graph; Real-time Feedback Interface, which supports dynamic adjustment of policy parameters through natural language interaction, and the interaction protocol is defined based on the JSON format.

8. An education evaluation method based on the system according to any one of claims 1-7, characterized in that, Including the steps: S1: Implement real-time collection and preprocessing of multi-modal data through edge computing nodes. The preprocessing includes data cleaning, normalization, and feature alignment; S2: Conduct distributed model training under the federated learning framework, satisfying privacy protection constraints, and the adaptive federated average algorithm is adopted during the training process; S3: Apply a dynamic Bayesian network to generate personalized learning path planning. The network nodes are defined as the knowledge point mastery status, and the edge weights are transition probabilities; S4: Output an adaptive teaching intervention plan based on the reinforcement learning policy engine. The policy reward function is the ratio of the improvement in grades to resource consumption; S5: Construct a digital twin of model performance to achieve continuous optimization of algorithm parameters. The optimization goal is to minimize the prediction error, computing time consumption, and memory occupancy.

9. The method according to claim 8, wherein Step S2 specifically includes: Adopting the adaptive federated average algorithm, the client selection probability formula is: where, Δ i is the local model update amplitude, defined as the L2 norm of the parameter gradient; α is the adjustment factor, and its initial value is set to 0.1; Implement gradient compression transmission, set the sparsity rate threshold θ = 0.7, and only transmit the parameters of the top 30% of the absolute values of the gradients; Execute model watermark embedding. The watermark algorithm is based on the Discrete Cosine Transform (DCT) to verify the authenticity of the node identity.

10. The method according to claim 8, characterized in that, The optimization process in Step S5 includes: Establish a multi-objective optimization model for algorithm parameters: min F(β) = [f1(β), f2(β), f3(β)] T Among them, f1 is the prediction error, f2 is the computing time consumption in seconds, f3 is the memory occupancy in MB; Apply the NSGA-II algorithm to solve the Pareto optimal solution set, set the population size to 100, and the number of iterations to 500; Design a constraint condition generator based on the knowledge graph. The constraint conditions include knowledge point dependency relationships and compliance with the teaching syllabus.

Images