Large-model-based score validity determination method for various complex scenes

Abstract

The invention relates to the technical field of sports performance judgment, in particular to a method for judging performance validity of various complex scenes based on a large model, which comprises the following steps: S1, integrating multi-source data, collecting data from multiple fields, preprocessing the collected data, and integrating the processed data into a structured data set; s2, recording and processing environmental conditions, generating a score data set containing the environmental conditions, analyzing environmental condition data, extracting key factors influencing score evaluation, and inputting the key factors as features into the large model; s3, constructing a dynamic adaptability evaluation model; and S4, high-dimensional feature processing: extracting key features in the data by adopting a principal component analysis method, reducing the dimensionality of the features, and improving the interpretability and understandability of the model. By effectively integrating data in multiple fields, especially processing multivariable data in a complex scene, adaptability and accuracy to different data sources are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of sports performance determination, and particularly to a method for determining the validity of various complex scenario scores based on a large model. Background Art

[0002] With the development of educational evaluation technology, score evaluation methods based on large models have gradually attracted attention. These methods aim to determine the validity of scores in various complex scenarios through big data and machine learning technologies, providing more scientific and objective evaluations. However, existing evaluation methods still have some deficiencies in dealing with complex scenarios, multi-variable data, and high-dimensional features, which affect the accuracy and reliability of the evaluation.

[0003] Limitations in data collection and processing: Existing technical solutions mainly rely on data collection and analysis in a single field and are difficult to handle complex scenarios across different fields. For example, the system of Patent CN111837123B mainly creates a conditional probability model based on uncertain data of sports scores but lacks the ability to integrate data from other relevant fields. In addition, problems with data quality and limited ability to record and process environmental conditions may lead to a decrease in the accuracy of score evaluation.

[0004] Lack of model adaptability and flexibility: Some evaluation models are mainly based on fixed rules and standard actions and lack the dynamic adaptation ability to individual differences in complex scenarios. For example, the system of Patent CN118887060B mainly evaluates the standardization and normality of athletes' movements by analyzing their postures but has insufficient adaptability to individual differences. In addition, existing technologies have limitations in dealing with multi-variable data and are difficult to effectively integrate multiple factors for comprehensive evaluation.

[0005] Computational complexity and real-time issues: When dealing with high-dimensional features, the computational complexity of existing systems is relatively high, which may affect the real-time performance and accuracy of the evaluation. For example, the system of Patent CN118887060B has a relatively high computational complexity when dealing with high-dimensional features. In addition, existing technologies have deficiencies in real-time performance and may not be able to meet actual requirements. Summary of the Invention

[0006] Based on the above objectives, the present invention provides a method for determining the validity of various complex scenario scores based on a large model, including the following: S1: Multi-source data integration, collecting data from multiple fields, preprocessing the collected data, and integrating the processed data into a structured data set; S2: Environmental condition recording and processing. Record and evaluate various conditions in the environment through sensors, associate the recorded environmental condition data with the performance data to generate a performance dataset containing environmental conditions, analyze the environmental condition data, extract key factors affecting performance evaluation, and use them as features to input into the large model; S3: Build a dynamic adaptive evaluation model; S4: High-dimensional feature processing. Adopt the principal component analysis method to extract key features in the data, reduce the dimension of the features, and generate new combined features through feature engineering to improve the interpretability and understandability of the model.

[0007] Preferably, the following steps are further included: S5: Real-time and computing optimization. In S3, adopt a distributed computing framework to improve computing efficiency and real-time performance, and use GPU to accelerate computing to reduce computing time; S6: Feedback mechanism and model optimization. Establish a feedback mechanism for evaluation results, continuously optimize the evaluation model by collecting user feedback information, update the model regularly, introduce new data and features, and display the decision-making process and basis of the model through visualization tools.

[0008] Preferably, in S1, the following steps are specifically included: S1.1: Input data from multiple fields, including student profile tables in relational databases, sports event records stored in unstructured document storage, API interfaces of third-party education platforms, and JSON-format real-time stream data from wearable device sensors; The processing process is as follows: Use the Python crawler framework to scrape the CSV files of previous sports performance from public event websites; Call the RESTful API to regularly obtain the student GPA historical records in the academic performance management system; Deploy the Kafka data pipeline to receive Protobuf-format stream data from IoT sensors distributed in sports venues in real time; Use the Apache NiFi visual data flow tool to configure the regular incremental synchronization of heterogeneous data sources; Output a multi-modal original heterogeneous dataset containing text, numerical values, time series signals, and images; S1.2: Process the original heterogeneous dataset output by S1.1 to output a normalized intermediate dataset; S1.3: Perform spatio-temporal alignment and fusion of multi-source data on the normalized intermediate dataset output by S1.2 to output a structured data cube for complex scenario analysis.

[0009] Preferably, in S1.2, data processing is performed on the original heterogeneous dataset output in S1.1, and the processing process is as follows: Use Pandas to implement outlier filtering for tabular data; Adopt sliding window median filtering for sensor stream data to eliminate high-frequency noise; Normalize the grayscale of image data through OpenCV; Interpolate the discontinuous records in academic achievements using time series interpolation; Use the Transformer model based on the attention mechanism to perform multi-dimensional semantic filling for the unanswered questions in the psychological assessment questionnaire; Implement RobustScaler standardization for numerical fields; For text-based evaluation reports, perform L2 normalization after TF-IDF vectorization; Convert time series sensor data into a time-frequency domain feature matrix through wavelet transform; Output the normalized intermediate dataset.

[0010] Preferably, in S1.3, multi-source data spatio-temporal alignment and fusion are performed on the normalized intermediate dataset output in S1.2, and the processing process is as follows: Spatio-temporal alignment: Add a unified timestamp to all data records, the UTC time synchronized by the NTP server, with a precision of milliseconds; Establish a cross-data source entity resolution index based on the student ID, and use the Levenshtein distance to match name ambiguous records; Perform spatial matching on the venue sensor data and competition records, and map the GPS coordinates to the specific competition venue number; Feature fusion: Adopt a graph neural network to construct a multi-relationship knowledge graph of students - courses - environment; Use a multi-modal fusion Transformer to align data streams with different sampling frequencies: Aggregate high-frequency sensor data into minute-level statistics through time slicing; Perform time series interpolation on low-frequency psychological assessment data to generate continuous features; Create a unified data view: Core entity table, including student ID, biometric fingerprint, school status; Behavior fact table, including exam event key, environmental parameters, original score, standardized score; Associated dimension table, including course metadata, venue configuration, evaluation standard version; Output a structured data cube for complex scenario analysis, including: Time domain dimension: an event time axis accurate to milliseconds; Spatial dimension: 3D coordinate mapping of the competition venue topology; Entity dimension: fully connected relationship network of students - teachers - equipment; Feature dimension: 587 derived feature fields after orthogonalization processing.

[0011] Preferably, in S2, it specifically includes the following steps: S2.1: Take the structured data cube output by S1.3 as input, and perform synchronous acquisition of multi - modal environmental parameters. The processing process is as follows: Deploy multi - type sensor arrays: The temperature and humidity sensor collects the micro - climate data of the competition venue at a frequency of 10Hz; The nine - axis inertial measurement unit is embedded in the sports equipment to capture three - dimensional acceleration / angular velocity in real - time; The panoramic camera records the global visual information of the competition venue; The distributed sound pressure meter constructs the sound field model; Establish a time synchronization mechanism: Adopt the IEEE 1588 precise time protocol to achieve μs - level clock synchronization across devices; Inject the atomic clock timestamp generated by the NTP server into each frame of data; Spatial calibration processing: Use the Leica total station to establish a three - dimensional coordinate system for the competition venue; Realize the real - time pose solution of the mobile sensor through the AprilTag visual marker; Finally, output the multi - modal environmental parameter stream; S2.2: Take the multi - modal environmental parameter stream output by S2.1 and the behavior fact table in the S1.3 data cube as input, perform dynamic environment - event spatio - temporal association, and output the spatio - temporally bound environment - event association data set; S2.3: Take the environment - event association data set output by S2.2 as input, and perform environment - sensitive feature engineering. The processing process is as follows: Physical field modeling: Build a three - dimensional simulation model of the temperature / humidity field based on computational fluid dynamics; Use the wave superposition method to reconstruct the sound field propagation path; Quantify the dynamic interference factors in the visual environment through the optical flow method; Multi - scale feature extraction: Calculate the cumulative effect of the environmental parameters deviating from the reference value; Statistically analyze the joint distribution of multi - dimensional environmental parameters; Extract frequency - domain features from inertial data; Use YOLOv8 to detect sudden interference events in the video; Causal reasoning analysis: Apply Bayesian network to construct an environment - performance causal relationship diagram; Execute do - calculus intervention analysis to quantify the marginal effect of key environmental factors on performance; Determine the contribution weights of each environmental feature through SHAP value decomposition; Output the environment - sensitive evaluation feature set, including: Physical field features, including temperature gradient vector and time - varying curve of sound pressure level; Dynamic response features, including spectral entropy of equipment vibration and offset of visual attention heat map; Causal reasoning features, including do - operator effect values of environmental interference factors and p - value matrix of conditional independence test.

[0012] Preferably, in S2.2, the multi - modal environmental parameter stream output in S2.1 and the behavior fact table in the data cube of S1.3 are used as inputs for dynamic environment - event spatio - temporal association, and the processing process is as follows: Event - driven environmental slice extraction: For each performance event, obtain the following data centered on the event occurrence point: Time window: Dynamically adjust the forward - tracing time and backward - extending time according to the event type; Spatial range: Construct a three - dimensional bounding box; Then use Flink SQL to query the parameter sequence in the environment stream that matches the spatio - temporal range in real - time; Heterogeneous data alignment: High - frequency sensor data is aligned to the event time axis through cubic B - spline interpolation; Extract key frames from the video stream; Extract the equivalent continuous sound level and 1 / 3 octave spectrum within the event window from the acoustic data; Establish an association index: Use GeoHash coding to convert the spatial coordinates into a 7 - bit string; Output the spatio - temporally bound environment - event association dataset, including: Environmental parameter snapshots, including temperature gradient matrix, motion trajectory tensor, visual scene graph, and acoustic fingerprint; Event metadata, including participating subjects, evaluation standard version, and equipment calibration status; Spatio - temporal correlation metrics, including spatial overlap rate and time coverage of environmental data and events.

[0013] Preferably, in S3, it specifically includes the following steps: S3.1: Model architecture design, including multi - modal feature embedding, cross - modal interaction, and establishing a dynamic adaptive mechanism; S3.2: Loss function design, including supervised loss, unsupervised loss, and hybrid loss functions; S3.3: Model training, including distributed training framework, optimizer selection and learning rate adjustment, and training monitoring and policy optimization; S3.4: Model verification and optimization, including model performance evaluation, model optimization, and model deployment and application; S3.5: Dynamic adjustment and update of the dynamic adaptability evaluation model, including environmental feature monitoring, model parameter adjustment, model structure optimization, model performance monitoring, and model update strategy.

[0014] Preferably, in S3.1, it specifically includes: Multi-modal feature embedding: Physical field feature embedding: Input the temperature gradient vector and the time-varying curve of sound pressure level, and extract spatial-frequency domain features through positional encoding embedding and channel attention mechanism; The temperature gradient vector extracts local spatial features through 3D convolution; The time-varying curve of sound pressure level extracts frequency domain features through one-dimensional convolution and combines the attention mechanism to capture important frequency bands; Dynamic response feature embedding: Input the vibration spectrum entropy of the instrument and the offset of the visual attention heat map, and extract temporal features using a convolutional neural network with residual connections; Perform frequency domain expansion on the spectrum entropy and extract the frequency band energy distribution features; Perform spatial encoding on the offset of the visual attention heat map to capture dynamic changes; Causal inference feature embedding: Input the effect value of the environmental interference factor do-operator and the p-value matrix of conditional independence test, and model the causal relationship through a graph attention network; Use graph embedding technology to map the causal relationship to a low-dimensional space and extract key causal features; Cross-modal interaction: Introduce a multi-head attention mechanism to achieve information interaction between different feature towers; Design learnable feature fusion weights to ensure that the weights of different features can be adaptively optimized; Use a cross-attention mechanism to capture the correlation between physical field features and dynamic response features; Dynamic adaptive mechanism: Parameter adjustment driven by environment-sensitive features: Dynamically adjust model parameters based on environmental features so that the model can adapt to the evaluation requirements under different environmental conditions; Achieve dynamic adjustment of the model through an adaptive layer, including: Learning rate adaptation: Dynamically adjust the model learning rate according to environmental changes; Model structure adaptation: Dynamically adjust the number of model layers and width according to environmental characteristics; Feature enhancement: Random noise injection: Add Gaussian noise to the input features to simulate environmental noise interference; Feature reordering: Randomly reorder the input features to enhance the model's robustness to the feature order; Feature combination: Generate new feature combinations through a learnable combination method to enrich feature representation.

[0015] Preferably, in S3.2, the supervised loss includes: Use the mean squared error to measure the difference between the model output and the true evaluation result; For classification tasks, use cross-entropy loss; For regression tasks, use a weighted combination of the root mean squared error and the mean absolute error.

[0016] Advantages of the present invention: 1. This method adopts a multi-domain data fusion strategy, overcoming the limitation of mainly relying on single-domain data analysis in the prior art. By effectively integrating data from multiple domains, especially processing multi-variable data in complex scenarios, the adaptability and accuracy to different data sources are improved. Compared with the system of the existing patent CN111837123B, it can evaluate the performance in complex scenarios more comprehensively. In addition, through technologies such as unsupervised learning and variational autoencoder (VAE), it can better capture environmental characteristics and high-dimensional data, providing strong support for cross-domain evaluation.

[0017] 2. This method overcomes the defect of evaluating the model based on fixed rules in the prior art and can dynamically adjust the evaluation criteria according to the changes in the actual scenario and individual differences. Through the optimization of hybrid supervised and unsupervised losses, the model can better adapt to the changes of different individuals and scenarios. For example, by adjusting the weight coefficients of the loss function and using strategies such as variational autoencoder (VAE), the model can effectively capture individual differences and dynamically adjust the evaluation strategy, thus providing a more accurate performance evaluation.

[0018] 3. This method significantly reduces the computational complexity through an optimized distributed training framework (such as Horovod) and efficient computational resource management, thereby improving the training speed and real-time evaluation ability of the model.

[0019] In the processing of high-dimensional features, an optimization method based on efficient computing (such as AdamW optimizer and dynamic learning rate adjustment strategy) is adopted to reduce the impact of computational complexity on the real-time performance and accuracy of evaluation, ensuring the efficiency and real-time response ability of the model in practical applications.

[0020] 4. This method adopts a series of multi-dimensional evaluation metrics (such as MSE, RMSE, F1-score, etc.) to evaluate performance from multiple dimensions, ensuring the comprehensiveness and accuracy of the evaluation results. Especially when dealing with high-dimensional data and complex scenarios, these diverse metrics can effectively capture the changes in different types of data, avoiding the limitations that may be brought by a single evaluation criterion.

[0021] 6. By adopting a dynamic learning rate adjustment and early stopping mechanism, this method can effectively prevent overfitting during the training process of the model and improve the generalization ability of the model. At the same time, through strategies such as cross-validation, the consistency and stability of the model on different datasets are ensured. When dealing with complex performance evaluation tasks, the optimization strategy of the model enables it to provide highly accurate prediction results in a variety of tasks and scenarios, thus enhancing the reliability and effectiveness of the evaluation.

[0022] 7. The evaluation framework proposed by this technical solution has strong adaptability and can handle the performance evaluation requirements in a variety of complex scenarios. By combining the advantages of supervised and unsupervised learning, the model can adjust the evaluation strategy when facing various dynamic environmental conditions and maintain stable evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only those of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0024] Figure 1 is a flowchart of the method of the present invention; Figure 2 is a flowchart of steps S5 and S6 of the method of the present invention; Figure 3 is a flowchart of step S3 of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] The following will describe the present invention in detail with reference to the drawings and specific embodiments. At the same time, it should be noted here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments. For some well-known technologies, those skilled in the art can also adopt other alternative methods for implementation; moreover, the drawings are only for more specifically describing the embodiments and are not intended to specifically limit the present invention.

[0026] Please refer to Figures 1 - 3, embodiments of the present invention provide a method for determining the effectiveness of scores in various complex scenarios based on large models. Through the integration of multi-source data, the recording and processing of environmental conditions, the construction of a dynamic adaptability evaluation model, and the processing of high-dimensional features, accurate assessment of score effectiveness is achieved.

[0027] Specifically, first, multi-dimensional raw data is collected from multiple fields, such as student scores, behavior data, environmental conditions, device sensor data, etc., to ensure that the obtained information has rich background information. The preprocessing process includes data cleaning, missing value filling, outlier detection, etc., aiming to provide a high-quality data basis for subsequent analysis. After the integration of data, a structured data set is formed, preparing for further analysis.

[0028] Furthermore, through tools such as sensors and environmental monitoring devices, various condition data in the actual environment, such as temperature, humidity, noise, etc., are collected, and correlation analysis is carried out in combination with score data. By generating a score data set containing environmental conditions, the potential impact of environmental factors on score assessment can be analyzed. During the analysis of environmental data, statistical methods are applied to extract key influencing factors, such as the negative or positive effects of environmental temperature, humidity, etc. on scores, and these factors will be used as features and input into the subsequent large model.

[0029] Furthermore, using the aforementioned data set and analysis results, the present invention constructs a dynamic adaptability evaluation model. This model can adjust the evaluation method in real time according to different environmental conditions, data dimensions and their changes, and has strong adaptability. The beneficial effect of this step is that it can dynamically process data in different situations and optimize model judgment according to real-time feedback, which helps to improve the accuracy and reliability of score effectiveness determination.

[0030] Furthermore, for the high-dimensional features in the data set, the principal component analysis (PCA) method is used for dimensionality reduction, and the most representative key features are extracted therefrom. These key features can not only reduce redundant information, but also improve the computational efficiency of the model. On this basis, new combined features are generated through feature engineering to further enrich the model input, making the prediction ability of the model stronger. In addition, the generated new features improve the interpretability of the model, making the analysis results more transparent and facilitating decision-makers to understand and use. Through the integration of multi-source data and the correlation of environmental conditions, comprehensive consideration of external factors is ensured, enhancing the representativeness and scientific nature of the assessment results; the dynamic adaptability model and high-dimensional feature processing improve the flexibility and interpretability of the model, making the determination of score effectiveness not only more accurate, but also easy to understand and apply.

[0031] In a possible implementation, to improve the efficiency and real-time performance of model calculations, the dynamic adaptability evaluation model constructed in step S3 needs to process a large amount of data and complex calculation tasks. Therefore, adopting a distributed computing framework for data processing and model training can disperse the calculation tasks to multiple computing nodes for parallel processing, significantly improving the calculation speed. In addition, combined with GPU accelerated computing, the parallel processing ability of the GPU greatly reduces the calculation time. Especially when dealing with large-scale data sets and complex models, the acceleration advantage of the GPU is particularly obvious. Through these calculation optimization methods, the model evaluation time can be effectively reduced, and the real-time performance can be improved, so as to adapt to the rapidly changing environment and requirements, and ensure that the effectiveness determination of the results in complex scenarios can be completed in a short time.

[0032] Furthermore, to enhance the accuracy and adaptability of the evaluation model, a feedback mechanism is established. Specifically, by collecting the feedback information of users on the model evaluation results, the performance and effectiveness of the model can be understood in a timely manner, providing data support for model optimization. This feedback information can include users' satisfaction with the evaluation results, the deviation between the model prediction results and the actual situation, etc. These data are used to update the model regularly. By retraining and optimizing the algorithm, the performance and prediction accuracy of the model are gradually improved. At the same time, to help users better understand and trust the decision-making process of the model, a visualization tool is introduced to display the decision-making process, basis, and key features affecting the evaluation results of the model. In this way, users can not only see the final evaluation results but also understand how the model makes these judgments, thereby enhancing the transparency and interpretability of the model.

[0033] Step S5 is closely connected to S6. The former ensures that the model can quickly respond to various dynamically changing inputs through optimizing the computing power and real-time processing. Especially in large-scale data processing and complex situations, timely evaluation can be achieved. The latter continuously improves the accuracy and robustness of the evaluation model through establishing a feedback mechanism and model optimization, enabling the evaluation model to continuously adapt to new data and scenario changes during long-term use. Through this organic combination, the model can not only provide accurate evaluation results in the short term but also maintain high effectiveness and adaptability in the long term through continuous feedback and optimization.

[0034] In a possible implementation, data is first obtained from multiple heterogeneous data sources, including student profile tables, sports event records, academic performance histories, real-time stream data from wearable device sensors, etc. The data sources involve relational databases (such as student profile tables), unstructured data (such as sports event records), third-party APIs (such as APIs of educational platforms), and real-time data streams from IoT sensors. To obtain data from these different sources, first, the historical score CSV files on public event websites are crawled through the Python scraping framework Scrapy, and then the GPA historical records of students are obtained regularly through RESTful APIs (using the Requests library). For real-time stream data, a Kafka data pipeline is deployed to receive the data streams in Protobuf format sent by IoT sensors (such as heart rate monitors and motion capture cameras) distributed in sports venues. To perform data synchronization, the Apache NiFi tool is used to configure the incremental synchronization of heterogeneous data sources (such as MySQL databases, MongoDB databases, and S3 buckets). Finally, all the collected data is stored in the / raw_data partition of the distributed file system HDFS in multimodal forms such as text, numerical values, time series signals, and images.

[0035] After data collection, it enters the data preprocessing stage. Through normalization processing, the original heterogeneous data set can be converted into a standard format that can be used for subsequent analysis. For example, for numerical data, normalization or standardization processing can be performed; for text data, word segmentation, stop word removal, etc. can be carried out; for image data, size unification and feature extraction can be done. The goal of data processing is to eliminate the differences between different data sources, making subsequent spatio-temporal alignment and data fusion smoother.

[0036] Perform spatio-temporal alignment and fusion of multi-source data on the normalized data. The information provided by different data sources has different timeliness and spatial characteristics. Therefore, it is necessary to align the data in the time dimension and the space dimension. For example, the update of academic performance may be on a semester basis, sports event scores are on a per-game basis, while sensor data may be in real-time streaming. After spatio-temporal alignment, the information from multiple data sources can be effectively fused together to form a unified data cube, which has a structured feature and is suitable for further analyzing the performance validity in complex scenarios.

[0037] In a possible implementation, S1.2 is the data processing stage, which includes a series of preprocessing steps for different data types, aiming to improve data quality and standardize data formats for subsequent analysis and fusion.

[0038] Specifically, for tabular data, the Pandas library is used to detect and filter outliers in the data. The specific method is to calculate the Z-score for each field. If the Z-score of a certain value is greater than 3, then this value is considered an outlier and marked for review. In this way, extreme values in the data can be automatically screened out, ensuring the quality and accuracy of the data. For sensor stream data, moving window median filtering is used to eliminate high-frequency noise. Specifically, a moving window with a window size of 50 sampling points is used to calculate the median, thereby removing instantaneous fluctuations and noise in the signal. For image data, the OpenCV library is used to uniformly scale video frames of different resolutions to 224x224 pixels and perform grayscale processing. This makes the image data have a unified resolution and grayscale level, facilitating subsequent feature extraction and model training.

[0039] Gaps in academic performance records are filled using cubic spline interpolation. This method uses a mathematical interpolation algorithm to smooth out the gaps in the data, thereby generating continuous performance records and avoiding data incompleteness caused by gaps. Cubic spline interpolation provides smooth and realistic performance predictions, helping to improve the coherence and integrity of academic data, enabling the analysis model to better handle academic performance data. Further, for unanswered questions in psychological assessment questionnaires, a Transformer model based on the attention mechanism (BERT-MAX) is used for multi-dimensional semantic filling. This model infers possible answers to unanswered questions by understanding the context and the user's historical data, thereby filling in the missing content. The BERT-MAX model can intelligently fill in the missing answers according to the context of the questionnaire, avoiding invalid data caused by missing answers and improving the integrity and accuracy of psychological assessment data.

[0040] For differences between different scoring systems (such as the 100-point system and the 5-point system), RobustScaler is used for standardization. This method eliminates the differences between scoring systems, making numerical fields comparable across different data sources and avoiding biases in data analysis caused by inconsistent scoring criteria. Further, for text-based assessment reports, first TF-IDF is used for vectorization to convert the text into a numerical vector representation, and then L2 normalization is performed. The text data processed in this way is convenient for subsequent feature analysis and machine learning model processing.

[0041] For time-series sensor data, wavelet transform is used to convert the time-series signal into a time-frequency domain feature matrix. Wavelet transform can reveal multiple frequency components in the signal, extract local features of the signal, and is suitable for complex pattern recognition of time-series data. After all data processing steps are completed, a normalized intermediate data set is finally output and stored in Parquet format. This format facilitates efficient storage and retrieval and is stored in the / processed directory of the HBase cluster, along with a schema metadata description file for describing the field structure. Using the Parquet format can effectively compress data and improve query efficiency, while the storage method of the HBase cluster ensures efficient and persistent access to data. The accompanying schema metadata file helps ensure the structuring and understandability of the data, facilitating subsequent analysis and model training.

[0042] In a possible implementation, first, all data records are synchronized to a unified UTC time through an NTP server to ensure that the time dimension of the data is accurate to milliseconds. This step solves the problem of inconsistent data collection times from different sources. Second, an entity resolution index is established based on the student ID, and the Levenshtein distance is used to handle the ambiguity of the student name to eliminate problems such as duplicate names. Finally, the venue sensor data is spatially matched with the competition records, and the GPS coordinates are mapped to specific competition venue numbers to ensure the consistency and accuracy of the data in space.

[0043] Furthermore, a multi-relational knowledge graph is constructed using a graph attention network (GAT) to comprehensively model the relationships among students, courses, and the environment. Combining with a multi-modal fusion Transformer (MuIT), high-frequency sensor data (such as 100Hz data) is aggregated into minute-level statistics through time slicing, while low-frequency data (such as monthly psychological assessment data) is interpolated temporally to generate continuous features. This can solve the problem of data fusion with different sampling frequencies and ensure the consistency of the data in the time dimension.

[0044] A core entity table (student ID, biometric fingerprint, school status, etc.), a behavior fact table (including exam events, environmental parameters, grades, etc.), and an associated dimension table (course metadata, venue configuration, etc.) are constructed. These tables provide comprehensive structured data support for subsequent analysis. Through the association of these tables, strong data support can be provided for grade determination in complex scenarios.

[0045] Through the spatio-temporal fusion engine, the system can achieve millimeter-level alignment of multi-source heterogeneous data and eliminate the evaluation bias caused by differences in data sampling frequencies in traditional methods. The multi-modal neural network processes unstructured data, such as combining video action analysis with heart rate changes, to capture potential influencing factors that cannot be reflected in traditional transcripts.

[0046] Considering the existence of student information conflicts in the actual scenario (such as cases of the same name or cross-school exchanges), this method ensures the accuracy of data association through a dynamic entity resolution mechanism. In addition, based on the version control and lineage tracking functions of Apache Atlas, the auditability of the data evolution process is ensured, thereby enhancing the transparency and credibility of the entire system.

[0047] Finally, the output structured data cube supports multi-dimensional query analysis, such as analyzing complex scenarios through time domain, space, entity, and feature dimensions, supporting queries for complex scenarios like "the change in the three-point shooting percentage of left-handed players in a basketball game when the humidity > 80%", providing a fine-grained analysis basis.

[0048] In a possible implementation manner, in step S2.1, first, multi-modal environmental data of the competition venue is collected through a multi-type sensor array. These sensors include temperature and humidity sensors, nine-axis inertial measurement units, panoramic cameras, and distributed sound pressure meters, which respectively record the microclimate of the competition venue, the dynamics of sports equipment, the full-field visual information, and the sound field characteristics. By deploying these sensors and achieving high-precision time and space synchronization, especially using the IEEE 1588 protocol for μs-level clock synchronization, it is ensured that the data collected by different devices is time-consistent, laying a foundation for subsequent data fusion and analysis.

[0049] In step S2.2, by combining the environmental parameter stream output in step S2.1 with the behavior fact table data, dynamic environment-event spatio-temporal association is performed. The core of this step is to ensure the precise matching of environmental data and the behavior data of athletes in time and space. This spatio-temporal binding helps to understand the real-time impact of environmental factors on athletes' performance, especially how it affects the competition results under rapidly changing environmental conditions (such as changes in wind speed, temperature, and humidity).

[0050] Step S2.3 is based on the environment-event association data set output in step S2.2 to extract environment-sensitive features. At this time, through physical field modeling techniques (such as temperature and humidity field, sound field modeling, etc.), combined with various methods such as wave superposition method and optical flow method, the interference effects of dynamic factors in the environment on athletes' performance are obtained. At the same time, through multi-scale feature extraction, the changes in the environment are quantified, so as to obtain the key factors that have a significant impact on the results.

[0051] Among them, causal inference analysis is a key step in S2.3. By constructing a Bayesian network, the causal relationship between various environmental factors and the results is analyzed, and do-calculus is applied for intervention analysis to quantify the marginal effects of different environmental factors (such as light, temperature, etc.) on the results. This analysis method enables the model not only to identify significant environmental factors but also to reveal the complex interaction between these factors.

[0052] In a possible implementation, an accurate spatio-temporal correlation model is established through event-driven environmental slice extraction and heterogeneous data alignment to ensure an exact match between environmental data and behavioral events.

[0053] Specifically, when an event occurs (e.g., student A completes a physics experiment), the time window and spatial range are dynamically adjusted centered around the moment of the event. The time window varies dynamically according to the type of event and may involve extending the moments before and after. The spatial range is defined by constructing a three-dimensional bounding box, where the size of the event site is multiplied by a safety factor (such as 1.2) to ensure that all possible environmental impact areas are covered. Flink SQL is used to query the data in the environmental stream in real time, and relevant sequences of environmental parameters are extracted according to the spatio-temporal range. The advantage of this step is that it can accurately capture the environmental changes before and after the event, especially suitable for complex experimental scenarios or competition environments, where minor environmental changes may have an important impact on the results.

[0054] Furthermore, for multi-modal data (such as high-frequency sensor data, video streams, and acoustic data), different techniques are used for alignment: Specifically, high-frequency sensor data (e.g., 100Hz data of inertial sensors) is aligned to the event time axis through cubic B-spline interpolation to ensure that the sensor data at each moment matches the event time and avoid data misalignment caused by sampling frequency differences.

[0055] The video stream extracts key frames through the OpenCV DNN module when key experimental operations are detected, which can accurately capture the visual information related to the event.

[0056] The acoustic data provides detailed information about the environmental sound by extracting the equivalent continuous sound level (Leq) and 1 / 3 octave spectrum within the event window, further supplementing the audio data when the event occurs.

[0057] To improve the efficiency of spatial data processing, the spatial coordinates are converted into GeoHash codes, which can transform the geographical location into a 7-bit string (with an accuracy of approximately 1.2 meters). In this way, the geographical location information related to the event can be quickly searched, while ensuring efficient processing of large-scale environmental data streams.

[0058] Finally, through the above steps, a data set including environmental parameter snapshots, event metadata, and spatio-temporal correlation metrics is generated. The data set includes: Environmental parameter snapshots: such as temperature gradient matrices, motion trajectory tensors, visual scene graphs, and acoustic fingerprints, which help provide an all-round perspective of the environment.

[0059] Event metadata: including participating entities, evaluation standard version, and device calibration status, to ensure clear background information for each event.

[0060] Spatio-temporal correlation metrics: such as the spatial overlap rate and time coverage between environmental data and events, to ensure an exact match between environmental and event data.

[0061] By combining the environmental parameter stream and the behavior fact table data, step S2.2 can provide more accurate event analysis. For example, dynamically adjusting the time window and spatial range enables the environmental data to precisely match the moment and location of the event occurrence, enhancing the reliability of spatio-temporal correlation. Meanwhile, the heterogeneous data alignment technology can eliminate time errors between different data sources, ensuring the synchronization of various types of data. GeoHash encoding makes the processing of spatial data more efficient, avoiding the complexity brought by traditional coordinate matching.

[0062] In a possible implementation, in S3.1, the design of the model architecture is carried out first. This link includes three key parts: multi-modal feature embedding, cross-modal interaction, and dynamic adaptive mechanism. Multi-modal feature embedding is to uniformly encode information from different data sources (such as videos, sensor data, voice, etc.) into a common feature space, which can ensure that various types of data can effectively interact in the same model. Cross-modal interaction enables data of different modalities to interact with each other through designing appropriate network layers (such as the Transformer architecture), fully exploring the correlations between them. The dynamic adaptive mechanism refers to automatically adjusting the model structure and parameters according to different environmental changes and data distributions, ensuring that the model can maintain good performance in different scenarios.

[0063] In step S3.2, three types of loss functions are designed: supervised loss, unsupervised loss, and hybrid loss. Supervised loss is used to process the training data with known labels, ensuring that the model can obtain correct prediction results on the labeled data. Unsupervised loss is applied to the data lacking clear labels, and through techniques such as clustering and generative adversarial networks (GANs), the latent structures in the data are mined. Hybrid loss combines the supervised loss and the unsupervised loss to further improve the generalization ability of the model. Through the design of these three types of losses, the model can be optimized at multiple levels, thereby enhancing its overall performance.

[0064] Step S3.3 focuses on the training process of the model, including the use of a distributed training framework, optimizer selection and learning rate adjustment, as well as training monitoring and policy optimization. The distributed training framework enables parallel processing of large-scale data and complex computational tasks, greatly improving the training efficiency. Optimizer selection determines the speed and stability of model convergence, while learning rate adjustment helps the model avoid overfitting or underfitting at different stages through dynamic changes. Training monitoring and policy optimization is to monitor the training process of the model in real time and dynamically adjust the training policy according to loss values, gradient information, etc., to ensure the efficiency and stability of the training process.

[0065] In Step S3.4, the performance of the model needs to be strictly evaluated and optimized. Model performance evaluation quantifies the performance of the model through multiple metrics (such as accuracy, recall, F1 value, etc.) to judge its effectiveness in various scenarios. Model optimization fine-tunes the model through methods such as backpropagation and hyperparameter tuning to improve its adaptability to complex scenarios. Model deployment and application is to apply the optimized model to the actual environment for real-time prediction and decision-making.

[0066] Finally, Step S3.5 focuses on the continuous adaptability of the model in actual applications. In this step, the changes in data distribution are tracked in real time through environmental feature monitoring, and the parameters and structure of the model are adjusted according to the monitoring results. This dynamic adjustment mechanism can ensure that the model always maintains good performance in a changing environment. Model performance monitoring helps the team determine whether the model needs to be further updated by providing real-time feedback on the difference between the predicted results and the actual results of the model. Model update strategies include methods such as periodic retraining and incremental learning to ensure that the model continuously adapts to new scenarios or data.

[0067] In a possible implementation, first, in the multi-modal feature embedding part, the system receives multiple input features, including the temperature gradient vector, the time-varying curve of sound pressure level, the spectral entropy of instrument vibration, and the offset of the visual attention heat map. These features undergo different embedding processes. The temperature gradient vector first extracts local spatial features through 3D convolution to capture the changes in the physical environment; the time-varying curve of sound pressure level extracts frequency-domain features through one-dimensional convolution and selects key frequency bands with the help of the attention mechanism; the spectral entropy of instrument vibration extracts the energy distribution of frequency bands through frequency-domain expansion technology and performs spatial encoding in combination with the offset of the visual attention heat map to capture the dynamic response of the instrument and environmental changes. These embedded features help provide accurate input information for the subsequent determination process.

[0068] Furthermore, in the causal inference feature embedding part, the causal relationship is modeled through a Graph Attention Network (GAT), which can explore the mutual influence between different features. This ability of causal inference is particularly important in complex environments, as it can effectively eliminate redundant and irrelevant features, strengthen the role of key causal features, and thus improve the decision-making accuracy of the model.

[0069] Furthermore, in the cross-modal interaction stage, the multi-head attention mechanism is adopted to ensure that features of different modalities can interact with each other and share information. Through the dynamically adjusted feature fusion weights, the model can optimize the contribution of each feature according to the actual situation of the data. On this basis, the cross-attention mechanism is used to capture the complex correlations between the physical field features and the dynamic response features, enhancing the model's understanding of the interaction between the environment and the device.

[0070] According to different environmental conditions (such as temperature gradient, sound pressure level), the model can automatically adjust its parameters. The introduction of the adaptive layer enables the learning rate, model structure, and feature representation to be optimized according to the changes in real-time data. In addition, feature enhancement methods, such as random noise injection, feature reordering, and feature combination, further improve the robustness and adaptability of the model.

[0071] In a possible implementation, in step S3.2, first, the difference between the model output and the true evaluation result is measured through the supervised loss. Specifically, for different tasks, different loss functions are selected: Mean Squared Error (MSE): Used to measure the difference between the model output and the true value in regression tasks.

[0072] Cross-Entropy Loss: Used for classification tasks to measure the difference between the class probabilities predicted by the model and the actual class labels.

[0073] For regression tasks, the weighted combination of Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) is also combined to more comprehensively evaluate the prediction performance of regression tasks.

[0074] Meanwhile, the unsupervised loss uses a Variational Autoencoder (VAE) and KL divergence to capture environmental features. The VAE reconstructs the original features, enabling the model to understand and represent these features, and the KL divergence is used to measure the difference between the reconstructed features and the original features, thereby improving the effect of unsupervised learning. For high-frequency sensor data, the model performs time-domain reconstruction to ensure that the high-frequency information in the data can be captured.

[0075] The final hybrid loss function combines the supervised and unsupervised losses and is obtained through weighted summation: ; where the weight coefficient and are weight coefficients optimized through a dynamic adjustment strategy. This weighted sum loss function can effectively combine the advantages of supervised learning and unsupervised learning, improving the training effect of the model and its adaptability to the environment.

[0076] In S3.3, the adoption of a distributed training framework (such as Horovod) enables multi-GPU parallel training, improving the model training efficiency. The data parallel strategy can divide the data into multiple batches and train them on different GPUs respectively, thus accelerating the training process.

[0077] The model parameters are synchronously updated through the parameter server, ensuring the parameter consistency of each node in distributed training and avoiding data deviation during the training process.

[0078] In terms of optimizer selection and learning rate adjustment, the AdamW optimizer is adopted, with the initial learning rate set to 10^-4. The learning rate adopts a dynamic adjustment strategy - Cosine Annealing Warm Restarts, which can make the learning rate gradually decay during the training process and occasionally reheat during the training process to help the model jump out of local optima, thus obtaining better training effects.

[0079] The model will monitor performance metrics in real time during the training process, including accuracy, recall, and F1-score. These metrics help evaluate the overall performance of the model in classification tasks and ensure that the training strategy can be dynamically adjusted to improve the model's performance.

[0080] The early stopping mechanism is adopted to prevent overfitting and ensure the generalization ability during the training process.

[0081] In S3.4, the performance of the model is cross-validated to ensure the stability and adaptability of the model on different datasets.

[0082] A series of performance evaluation metrics are designed, such as MSE, RMSE, F1-score, etc., to comprehensively measure the prediction ability of the model.

[0083] According to the evaluation results, the model structure and training strategy will be optimized. Through model compression and quantization, the inference efficiency is improved, enabling the model to run more efficiently in practical applications.

[0084] The optimized model will be deployed to the actual environment, and the environmental features will be processed in real time through an efficient inference engine and the evaluation results will be output. At the same time, through a continuous model monitoring and updating mechanism, the model can adapt to environmental changes and maintain its long-term effectiveness and adaptability.

[0085] The present invention encompasses any alternatives, modifications, equivalent methods, and solutions made within the spirit and scope of the present invention. To enable the public to have a thorough understanding of the present invention, specific details are elaborated in the following preferred embodiments of the present invention. However, those skilled in the art can fully understand the present invention even without the description of these details. Additionally, well-known methods, processes, procedures, components, and circuits, etc. are not described in detail to avoid unnecessary confusion to the essence of the present invention.

[0086] The above description is only a preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for determining the effectiveness of scores in various complex scenarios based on large models, characterized in that, Including the following: S1: Multi-source data integration, collecting data from multiple fields, preprocessing the collected data, and integrating the processed data into a structured dataset; S2: Environmental condition recording and processing, recording various conditions in the evaluation environment through sensors, associating the recorded environmental condition data with the performance data to generate a performance dataset containing environmental conditions, analyzing the environmental condition data, extracting key factors affecting performance evaluation, and using them as features to input into the large model; S3: Constructing a dynamic adaptive evaluation model; S4: High-dimensional feature processing, using the principal component analysis method to extract key features in the data, reducing the dimension of the features, and generating new combined features through feature engineering to improve the interpretability and understandability of the model.

2. The method for determining the effectiveness of scores in various complex scenarios based on a large model according to claim 1, wherein, It also includes the following steps: S5: Real-time and computational optimization, in S3, adopting a distributed computing framework to improve computational efficiency and real-time performance, and using GPU to accelerate the calculation to reduce the calculation time; S6: Feedback mechanism and model optimization, establishing a feedback mechanism for evaluation results, continuously optimizing the evaluation model by collecting user feedback information, regularly updating the model, introducing new data and features, and using visualization tools to display the decision-making process and basis of the model.

3. A method for determining the effectiveness of scores in various complex scenarios based on a large model according to claim 1, characterized in that, In S1, it specifically includes the following steps: S1.1: Inputting data from multiple fields, including student profile tables in relational databases, sports event records stored in unstructured document storage, API interfaces of third-party education platforms, and real-time stream data in JSON format from wearable device sensors; The processing process is as follows: Using the Python crawler framework to crawl CSV files of previous sports results from public event websites; Calling the RESTful API to regularly obtain the student GPA historical records in the academic performance management system; Deploying a Kafka data pipeline to receive Protobuf format stream data from IoT sensors distributed in sports venues in real time; Using the Apache NiFi visual data flow tool to configure regular incremental synchronization of heterogeneous data sources; Outputting a multi-modal original heterogeneous dataset containing text, numerical values, time series signals, and images; S1.2: Performing data processing on the original heterogeneous dataset output by S1.1 and outputting a normalized intermediate dataset; S1.3: Performing multi-source data spatio-temporal alignment and fusion on the normalized intermediate dataset output by S1.2 and outputting a structured data cube for complex scenario analysis.

4. A method for determining the effectiveness of scores in various complex scenarios based on a large model according to claim 3, characterized in that, In S1.2, the data processing of the original heterogeneous dataset output by S1.1 is as follows: Using Pandas to implement outlier filtering for tabular data; Adopting sliding window median filtering for sensor stream data to eliminate high-frequency noise; Performing grayscale normalization on image data through OpenCV; Supplementary interpolation of discontinuous records in academic performance using time series interpolation; Using a Transformer model based on the attention mechanism to perform multi-dimensional semantic filling for unanswered questions in psychological evaluation questionnaires; Implementing RobustScaler standardization for numerical fields; Performing L2 normalization after TF-IDF vectorization for text-based evaluation reports; The time-series sensor data is transformed into a time-frequency domain feature matrix through wavelet transform; Output the normalized intermediate data set.

5. The method for determining the effectiveness of various complex scenario scores based on a large model according to claim 4, wherein, In S1.3, the normalized intermediate data set output in S1.2 is subjected to multi-source data spatio-temporal alignment and fusion, and the processing process is as follows: Spatio-temporal alignment: Add a unified timestamp to all data records, the UTC time synchronized by the NTP server, with a precision to milliseconds; Establish an entity resolution index across data sources based on the student ID, and use the Levenshtein distance to match name ambiguous records; Perform spatial matching on the venue sensor data and the competition records, and map the GPS coordinates to the specific competition venue number; Feature fusion: Adopt a graph neural network to construct a multi-relational knowledge graph of students-courses-environments; Use a multi-modal fusion Transformer to align data streams with different sampling frequencies: The high-frequency sensor data is aggregated into minute-level statistics through time slicing; The low-frequency psychological assessment data is interpolated in time series to generate continuous features; Create a unified data view: Core entity table, including student ID, biometric fingerprint, school status; Behavior fact table, including exam event key, environmental parameters, original score, standardized score; Associated dimension table, including course metadata, venue configuration, evaluation standard version; Output a structured data cube for complex scenario analysis, including: Time domain dimension: an event time axis accurate to milliseconds; Spatial dimension: a three-dimensional coordinate mapping of the competition venue topology structure; Entity dimension: a fully connected relational network of students-teachers-devices; Feature dimension: 587 derived feature fields processed by orthogonalization.

6. The method for determining the effectiveness of scores in various complex scenarios based on a large model according to claim 5, wherein, In S2, it specifically includes the following steps: S2.1: Take the structured data cube output in S1.3 as input, and perform multi-modal environmental parameter synchronous acquisition. The processing process is as follows: Deploy multi-type sensor arrays: The temperature and humidity sensors collect the microclimate data of the competition venue at a frequency of 10Hz; The nine-axis inertial measurement unit is embedded in the sports equipment to capture three-dimensional acceleration / angular velocity in real time; The panoramic camera records the global visual information of the competition venue; The distributed sound pressure meter constructs a sound field model; Establish a time synchronization mechanism: Adopt the IEEE 1588 precise time protocol to achieve sub-microsecond clock synchronization across devices; Inject the atomic clock timestamp generated by the NTP server into each frame of data; Spatial calibration processing: Use a Leica total station to establish a three-dimensional coordinate system for the competition venue; Realize real-time pose solution of the mobile sensor through the AprilTag visual marker; Finally, output the multi-modal environmental parameter stream; S2.2: Take the multi-modal environmental parameter stream output in S2.1 and the behavior fact table in the S1.3 data cube as input, perform dynamic environment-event spatio-temporal association, and output a spatio-temporally bound environment-event association data set; S2.3: Take the environment-event association data set output in S2.2 as input, and perform environment-sensitive feature engineering. The processing process is as follows: Physical field modeling: Build a three-dimensional simulation model of the temperature / humidity field based on computational fluid dynamics; Use the wave superposition method to reconstruct the sound field propagation path; Quantify the dynamic interference factors in the visual environment through the optical flow method; Multi-scale feature extraction: Calculate the cumulative effect of the environmental parameters deviating from the reference value; Statistical joint distribution of multi-dimensional environmental parameters; Extract frequency-domain features from inertial data; Detect sudden interference events in videos using YOLOv8; Causal inference analysis: Apply Bayesian networks to construct an environment-performance causal relationship graph; Perform do-calculus intervention analysis to quantify the marginal effect of key environmental factors on performance; Determine the contribution weights of each environmental feature through SHAP value decomposition; Output an environment-sensitive evaluation feature set, including: Physical field features, including temperature gradient vectors and time-varying curves of sound pressure levels; Dynamic response features, including entropy of instrument vibration spectra and offsets of visual attention heatmaps; Causal inference features, including do-operator effect values of environmental interference factors and p-value matrices of conditional independence tests.

7. A method for determining the effectiveness of scores in various complex scenarios based on a large model according to claim 6, characterized in that, In S2.2, use the multi-modal environmental parameter stream output in S2.1 and the behavior fact table in the data cube of S1.3 as inputs for dynamic environment-event spatio-temporal association. The processing steps are as follows: Event-driven environmental slice extraction: For each performance event, obtain the following data centered on the event occurrence point: Time window: Dynamically adjust the forward trace time and backward extension time according to the event type; Spatial range: Construct a three-dimensional bounding box; Then use Flink SQL to query the parameter sequence in the environmental stream that matches the spatio-temporal range in real time; Heterogeneous data alignment: High-frequency sensor data is aligned to the event time axis through cubic B-spline interpolation; Extract key frames from the video stream; Extract the equivalent continuous sound level and 1 / 3 octave spectrum within the event window from the acoustic data; Establish an association index: Use GeoHash coding to convert spatial coordinates into 7-bit strings; Output a spatio-temporally bound environment-event association data set, including: Environmental parameter snapshots, including temperature gradient matrices, motion trajectory tensors, visual scene graphs, and acoustic fingerprints; Event metadata, including participating entities, evaluation standard versions, and device calibration status; Spatio-temporal association degree indicators, including the spatial overlap rate and time coverage of environmental data and events.

8. A method for determining the effectiveness of scores in various complex scenarios based on a large model according to claim 7, characterized in that, In S3, it specifically includes the following steps: S3.1: Model architecture design, including multi-modal feature embedding, cross-modal interaction, and establishment of a dynamic adaptive mechanism; S3.2: Loss function design, including supervised loss, unsupervised loss, and mixed loss functions; S3.3: Model training, including a distributed training framework, optimizer selection and learning rate adjustment, and training monitoring and strategy optimization; S3.4: Model verification and optimization, including model performance evaluation, model optimization, and model deployment and application; S3.5: Dynamic adjustment and update of the dynamic adaptability evaluation model, including environmental feature monitoring, model parameter adjustment, model structure optimization, model performance monitoring, and model update strategy.

9. A method for determining the effectiveness of scores in various complex scenarios based on a large model according to claim 6, characterized in that, In S3.1, it specifically includes: Multi-modal feature embedding: Physical field feature embedding, input temperature gradient vectors and time-varying curves of sound pressure levels, and extract spatio-frequency domain features through position encoding embedding and channel attention mechanisms; Extract local spatial features from the temperature gradient vector through 3D convolution; Extract frequency domain features from the time-varying curve of sound pressure level through one-dimensional convolution and capture important frequency bands by combining attention mechanisms; Dynamic response feature embedding, inputting the vibration spectrum entropy of the device and the offset of the visual attention heatmap, and using a convolutional neural network with residual connections to extract temporal features; Performing frequency-domain expansion on the spectrum entropy to extract the frequency band energy distribution features; Performing spatial encoding on the offset of the visual attention heatmap to capture dynamic changes; Causal inference feature embedding, inputting the do-operator effect value of the environmental interference factor and the p-value matrix of conditional independence test, and modeling the causal relationship through a graph attention network; Using graph embedding technology to map the causal relationship to a low-dimensional space and extract key causal features; Cross-modal interaction: Introducing a multi-head attention mechanism to achieve information interaction between different feature towers; Designing learnable feature fusion weights to ensure that the weights of different features can be adaptively optimized; Using a cross-attention mechanism to capture the correlation between physical field features and dynamic response features; Dynamic adaptive mechanism: Parameter adjustment driven by environment-sensitive features, dynamically adjusting model parameters based on environmental features so that the model can adapt to the evaluation requirements under different environmental conditions; Achieving dynamic adjustment of the model through an adaptive layer, including: Learning rate adaptation: Dynamically adjusting the model learning rate according to environmental changes; Model structure adaptation: Dynamically adjusting the number of layers and width of the model according to environmental features; Feature enhancement: Random noise injection: Adding Gaussian noise to the input features to simulate environmental noise interference; Feature reordering: Randomly reordering the input features to enhance the model's robustness to the feature order; Feature combination: Generating new feature combinations through a learnable combination method to enrich feature expression.

10. A method for determining the effectiveness of scores in various complex scenarios based on a large model according to claim 9, characterized in that, In S3.2, the supervised loss includes: Using the mean square error to measure the difference between the model output and the true evaluation result; For classification tasks, using cross-entropy loss; For regression tasks, using a weighted combination of the root mean square error and the mean absolute error.

Images