Simulation training data analysis processing method based on big data

Abstract

The invention discloses a simulation training data analysis processing method based on big data, and belongs to the technical field of simulation training. Comprising the following steps: constructing a multi-modal data synchronous acquisition framework, and synchronously acquiring operation logs, physiological signals and audio and video data of at least one trained object in a simulation training process so as to construct a time-space consistent training data set; the training data set is preprocessed, time alignment of data of different sampling rates is achieved through a dynamic time warping algorithm, and an incidence relation network of operation behaviors, physiological reactions and environmental parameters is constructed based on the knowledge graph technology; extracting spatio-temporal characteristics based on the incidence relation network, constructing an operation sequence prediction model based on a Transform architecture, and taking a historical operation sequence as input and a future operation state prediction result as output by the sequence prediction model; a historical normal behavior model is constructed through a variational auto-encoder, and the historical normal behavior model takes a historical operation sequence as input.

Description

Technical Field

[0001] This invention relates to the field of simulation training technology, specifically to a method for analyzing and processing simulation training data based on big data. Background Technology

[0002] Simulation training, as a core means of improving skill levels, has been widely applied in fields such as medicine, aviation, and industry. With the development of sensor technology and artificial intelligence, the data generated by simulation training is experiencing explosive growth, encompassing multimodal information such as operational trajectories, physiological signals, and environmental parameters. However, existing technologies face the following core challenges when processing large-scale simulation training data:

[0003] Challenges of multimodal data fusion: Operation logs, physiological signals, and audio / video data often have different sampling rates and semantic spaces. Traditional methods struggle to achieve high-precision time alignment and semantic association, making it difficult to conduct in-depth cross-modal analysis of behavior, physiology, and environment.

[0004] Insufficient modeling of complex operation modes: The operation sequences in modern simulation training are becoming increasingly complex. Traditional time series models are unable to capture long-distance dependencies and have limited accuracy in predicting future operation states, which cannot meet the needs of high real-time analysis.

[0005] Personalized training support is lacking: Existing training systems generally adopt a standardized model, ignoring individual differences in trainees' operating habits, cognitive abilities, physiological stress, etc., and lack a data-driven mechanism for generating personalized training programs.

[0006] Weak real-time anomaly detection capability: Traditional anomaly detection relies on rule engines or manual inspection, which is slow to respond to subtle deviations in the operation sequence and makes it difficult to detect and correct violations or inefficient behaviors in a timely manner during the training process.

[0007] In summary, existing simulation training data analysis methods have significant shortcomings in terms of multimodal data processing accuracy, complex behavior modeling capabilities, personalized training support, and real-time feedback efficiency. There is an urgent need for an intelligent solution based on big data to achieve in-depth analysis of training data and dynamic optimization of the training process. Summary of the Invention

[0008] The purpose of this invention is to provide a method for analyzing and processing simulation training data based on big data, so as to solve the problems mentioned in the background art.

[0009] In view of the above problems, the technical solution proposed by the present invention is as follows:

[0010] A method for analyzing and processing simulation training data based on big data includes the following steps:

[0011] S1: Construct a multimodal data synchronous acquisition framework to synchronously acquire operation logs, physiological signals, and audio-visual data of at least one trainee during the simulated training process, thereby constructing a spatiotemporally consistent training dataset. By synchronously acquiring multi-dimensional data, the one-sidedness of a single data type is avoided, providing a complete data foundation for subsequent cross-modal analysis and solving the "multimodal data fusion problem" in the background technology.

[0012] S2: The training dataset is preprocessed, and the time alignment of data with different sampling rates is achieved through dynamic time warping algorithm. Based on knowledge graph technology, a network of relationships between operational behavior, physiological response and environmental parameters is constructed to achieve spatiotemporal alignment and semantic association of heterogeneous data, improve the accuracy of cross-modal analysis, and solve the problems of time synchronization and semantic fragmentation in the "multimodal data fusion problem".

[0013] S3: Based on the aforementioned relationship network, extract spatiotemporal features and construct an operation sequence prediction model based on the Transformer architecture. The sequence prediction model takes historical operation sequences as input and future operation state prediction results as output, captures long-distance dependencies of operation sequences, improves the prediction accuracy of complex operation modes, and solves the problem of "insufficient modeling of complex operation modes".

[0014] S4: A historical normal behavior model is constructed through a variational autoencoder. The historical normal behavior model takes the historical operation sequence as input and the historical normal operation mode as output. The prediction result of the future operation state is compared with the historical normal operation mode. When the deviation exceeds a preset threshold, an alarm mechanism is triggered to realize the real-time detection and alarm of operation anomalies. The response latency is as low as milliseconds, which solves the problem of "weak real-time anomaly detection capability".

[0015] S5: Extract the feature vector of the training object from the training dataset. The feature vector includes at least operational accuracy features, temporal stability features, and physiological stress features. Map the feature vector to the feature vector space through a clustering algorithm to form a feature profile of the training object, quantify the individual differences of the training object, provide data support for personalized training, and solve the problem of "gap in personalized training support".

[0016] S6: Construct a training parameter optimization model based on reinforcement learning. Using the feature profile as input, dynamically adjust the training scenario parameters to generate a personalized training scheme for the trainee. Optimize training parameters dynamically based on individual characteristics to improve training efficiency and solve the problem of "lack of personalized training support".

[0017] As a preferred technical solution of the present invention, the multimodal data synchronous acquisition framework includes edge computing nodes and a central server. The edge computing nodes are deployed at the training site and are used to perform feature extraction and dimensionality reduction preprocessing on audio and video data. The central server adopts a distributed storage architecture, which supports parallel reading and writing and real-time retrieval of the training dataset. Through the edge-cloud collaborative architecture, the data transmission pressure is reduced and the processing efficiency is improved, solving the bandwidth bottleneck and real-time problem during large-scale data acquisition.

[0018] As a preferred technical solution of the present invention, the dynamic time warping algorithm is used to align heterogeneous sampling rate data. It generates the optimal alignment path by calculating the distance matrix of the time series, ensuring that the timestamp error of the multimodal data does not exceed a preset threshold, achieving multimodal data alignment with millisecond-level accuracy, providing a time reference for cross-modal correlation analysis, and solving the problem of "large time synchronization error of multimodal data".

[0019] As a preferred technical solution of the present invention, the knowledge graph technology identifies operational behavior entities, physiological response entities, and environmental parameter entities from the training dataset through entity extraction and relational reasoning, and constructs a causal relationship network between entities based on co-occurrence relationships, thereby constructing a semantic association network of operation-physiology-environment, supporting causal analysis of complex behaviors, and solving the problem of "semantic fragmentation of multimodal data".

[0020] As a preferred technical solution of the present invention, the operation sequence prediction model based on the Transformer architecture includes a multi-head self-attention mechanism to capture long-distance dependencies in the operation sequence, and improves the prediction accuracy of rare operation patterns and the modeling ability of complex operation sequences through a contrastive learning algorithm.

[0021] As a preferred technical solution of the present invention, the normal behavior model constructed by the variational autoencoder constrains the distribution of latent variables through the KL divergence loss function. When the reconstruction error between the predicted result of the future operation state and the historical normal operation mode exceeds a preset threshold, a hierarchical alarm mechanism is triggered to realize the quantitative assessment and hierarchical response of operation anomalies, reduce the false alarm rate and improve the pertinence of alarms.

[0022] As a preferred technical solution of the present invention, the training parameter optimization model of the reinforcement learning adopts the policy gradient algorithm. The feature vector space includes a state space and an action space. The state space includes the feature profile, the current training progress and the historical training effect. The action space includes adjustable parameters for adjusting the training scenario complexity and selecting the feedback type. The training parameters are dynamically optimized based on individual features to improve training efficiency and solve the inefficiency problem of traditional training modes.

[0023] As a preferred embodiment of the present invention, the response time of the alarm mechanism does not exceed a preset threshold. The preset threshold is dynamically adjusted according to the risk level of the training task to ensure the real-time response to abnormal operations of high-risk operations. The higher the risk level, the faster the alarm response.

[0024] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention achieves the construction of a spatiotemporally consistent training dataset through synchronous acquisition and preprocessing of multimodal data; it improves the prediction accuracy of complex operations and the real-time anomaly detection capability by utilizing the Transformer architecture and variational autoencoder; it generates personalized training schemes that quantify individual differences based on clustering algorithms and reinforcement learning; and it optimizes data processing efficiency by combining edge computing and distributed storage. Ultimately, it forms a complete closed loop from data acquisition to dynamic optimization, significantly improving the analytical accuracy, real-time performance, and personalization level of simulation training. It solves core problems such as multimodal data fusion, complex behavior modeling, real-time feedback, and personalized training, and promotes the intelligent upgrade of simulation training. Attached Figure Description

[0025] Figure 1 This is a flowchart of a big data-based simulation training data analysis and processing method disclosed in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Please see Figure 1 This invention provides a technical solution: a method for analyzing and processing simulation training data based on big data, comprising the following steps:

[0028] S1: Construct a multimodal data synchronous acquisition framework to synchronously collect operation logs, physiological signals, and audio-visual data of at least one trainee during the simulated training process, thereby constructing a spatiotemporally consistent training dataset. High-precision clock synchronization technology is used to ensure the consistency of timestamps of operation logs, physiological signals, and audio-visual data. Invalid data is initially filtered through edge computing nodes to reduce transmission pressure, and finally a spatiotemporally aligned training dataset is formed on the central server.

[0029] S2: The training dataset is preprocessed, and the time alignment of data at different sampling rates is achieved through a dynamic time warping algorithm. A network of relationships between operational behaviors, physiological responses, and environmental parameters is constructed based on knowledge graph technology. By calculating the distance matrix of time series at different sampling rates, the optimal alignment path is generated to accurately align physiological signals with operational sequences. Named entity recognition technology is used to extract operational behavior entities, physiological response entities, and environmental parameter entities from the data, and causal relationship edges are constructed based on co-occurrence frequency.

[0030] S3: Based on the correlation network, spatiotemporal features are extracted, and an operation sequence prediction model based on the Transformer architecture is constructed. The sequence prediction model takes the historical operation sequence as input and the prediction result of the future operation state as output. It extracts the spatial trajectory and time sequence features of the operation from the correlation network. The semantic associations of different time steps are captured in parallel through a multi-head self-attention mechanism. The sequence order is processed by position encoding to achieve the prediction of the operation state in the next 1 to 5 seconds.

[0031] S4: A historical normal behavior model is constructed through a variational autoencoder. The historical normal behavior model takes the historical operation sequence as input and the historical normal operation mode as output. The prediction result of the future operation state is compared with the historical normal operation mode. When the deviation exceeds a preset threshold, an alarm mechanism is triggered. The historical normal operation sequence is encoded as a latent variable distribution and the operation mode is reconstructed through a decoder. The reconstruction error between the current prediction result and the historical normal mode is calculated. When the error exceeds a threshold, a graded alarm is triggered.

[0032] S5: Extract feature vectors of the training subjects from the training dataset. The feature vectors should include at least operational accuracy features, temporal stability features, and physiological stress features. Map the feature vectors to the feature vector space using a clustering algorithm to form a feature profile of the training subjects. Calculate indicators such as operational accuracy, temporal stability, and physiological stress from the training data. Use DBSCAN or K-means algorithms to map the feature vectors to a three-dimensional feature space to form clusters such as "novice," "skilled," and "expert," generating a feature profile of the training subjects.

[0033] S6: Construct a training parameter optimization model based on reinforcement learning. Taking feature profiles as input, dynamically adjust the training scenario parameters to generate personalized training schemes for the trainees. Employ a proximal strategy optimization algorithm. The state space includes feature profiles, training progress, and historical results, while the action space includes parameters such as scenario complexity and feedback type. By maximizing the reward function, generate customized schemes for different clusters.

[0034] As an embodiment of the present invention, the multimodal data synchronous acquisition framework further includes edge computing nodes and a central server. The edge computing nodes are deployed at the training site and are used for feature extraction and dimensionality reduction preprocessing of audio and video data. The central server adopts a distributed storage architecture, supports parallel reading and writing and real-time retrieval of the training dataset, collects audio and video data in real time at the training site, extracts key features through a lightweight model, and extracts frames from the original video to compress the data volume by more than 80%. The central server uses a Hadoop distributed file system to store data and achieves parallel reading and writing through a Spark cluster, supporting concurrent processing of thousands of data streams and ensuring that the data write latency is less than 1 second.

[0035] As an embodiment of the present invention, the Dynamic Time Warping (DTW) algorithm is further used to align heterogeneous sampling rate data. It generates the optimal alignment path by calculating the distance matrix of the time series, ensuring that the timestamp error of the multimodal data does not exceed a preset threshold. Taking operation logs and electrocardiogram signals as examples, the DTW algorithm first constructs a two-dimensional distance matrix, with the matrix elements being the data differences at corresponding time points. Then, it uses dynamic programming to search for the minimum cumulative distance path from the starting point to the ending point of the matrix, generating a time alignment mapping table, and finally controlling the error within the preset threshold.

[0036] As an embodiment of the present invention, the knowledge graph technology further identifies operational behavior entities, physiological reaction entities, and environmental parameter entities from the training dataset through entity extraction and relation reasoning, and constructs a causal relationship network between entities based on co-occurrence relationships. The BERT-NER model is used to identify entities from text logs. Causal edges are constructed through co-occurrence analysis to form a knowledge graph, providing semantic support for subsequent prediction models.

[0037] As an embodiment of the present invention, the operation sequence prediction model based on the Transformer architecture further includes a multi-head self-attention mechanism to capture long-distance dependencies in the operation sequence, and improves the prediction accuracy of rare operation patterns through a contrastive learning algorithm. The operation sequence is mapped to multiple subspaces, and attention weights in different subspaces are calculated in parallel to capture dependencies across subspaces. By bringing similar operation sequences closer in the feature space and pushing dissimilar sequences further apart, the model's ability to distinguish rare patterns is enhanced.

[0038] As an embodiment of the present invention, the normal behavior model constructed by the variational autoencoder further constrains the distribution of latent variables through the KL divergence loss function. When the reconstruction error between the prediction result of the future operation state and the historical normal operation mode exceeds a preset threshold, a hierarchical alarm mechanism is triggered. KL divergence ensures that the distribution of latent variables is close to a normal distribution, and the reconstruction error measures the difference between the prediction result and the historical mode. Three thresholds are set to correspond to different response strategies.

[0039] As an embodiment of the present invention, the reinforcement learning training parameter optimization model further adopts the policy gradient algorithm. The feature vector space includes a state space and an action space. The state space includes feature profiles, current training progress and historical training results. The action space includes adjustable parameters for adjusting training scenario complexity and selecting feedback type. The feature profiles, training progress and historical results of the trainee are encoded into state vectors. Actions are output through the policy network, and environmental feedback rewards drive model iteration to form a personalized training strategy.

[0040] As an embodiment of the present invention, the response time of the alarm mechanism does not exceed a preset threshold. The preset threshold is dynamically adjusted according to the risk level of the training task, and a mapping relationship between the risk level and the response threshold is established. Real-time anomaly detection results are cached in Redis, and the priority scheduling of alarm messages is achieved using a message queue to ensure that alarm responses for high-risk tasks are processed first.

Claims

1. A method for analyzing and processing simulation training data based on big data, characterized in that, Includes the following steps: S1: Construct a multimodal data synchronous acquisition framework to synchronously acquire the operation logs, physiological signals, and audio and video data of at least one trainee during the simulated training process, thereby constructing a spatiotemporally consistent training dataset; S2: The training dataset is preprocessed, and the time alignment of data with different sampling rates is achieved through dynamic time warping algorithm. A network of relationships between operational behavior, physiological response and environmental parameters is constructed based on knowledge graph technology. S3: Extract spatiotemporal features based on the aforementioned relationship network, and construct an operation sequence prediction model based on the Transformer architecture. The sequence prediction model takes historical operation sequences as input and future operation state prediction results as output. S4: Construct a historical normal behavior model through a variational autoencoder. The historical normal behavior model takes the historical operation sequence as input and the historical normal operation mode as output. The prediction result of the future operation state is compared with the historical normal operation mode. When the deviation exceeds a preset threshold, an alarm mechanism is triggered. S5: Extract the feature vector of the training object from the training dataset. The feature vector includes at least operational accuracy features, temporal stability features, and physiological stress features. Map the feature vector to the feature vector space using a clustering algorithm to form a feature profile of the training object. S6: Construct a training parameter optimization model based on reinforcement learning, take the feature profile as input, dynamically adjust the training scenario parameters, and generate a personalized training scheme for the training object.

2. The method for analyzing and processing simulation training data based on big data according to claim 1, characterized in that, The multimodal data synchronous acquisition framework includes edge computing nodes and a central server. The edge computing nodes are deployed at the training site and are used to perform feature extraction and dimensionality reduction preprocessing on audio and video data. The central server adopts a distributed storage architecture to support parallel reading and writing and real-time retrieval of the training dataset.

3. The method for analyzing and processing simulation training data based on big data according to claim 1, characterized in that, The dynamic time warping algorithm is used to align heterogeneous sampling rate data. It generates the optimal alignment path by calculating the distance matrix of the time series, ensuring that the timestamp error of the multimodal data does not exceed a preset threshold.

4. The method for analyzing and processing simulation training data based on big data according to claim 1, characterized in that, The knowledge graph technology identifies operational behavior entities, physiological response entities, and environmental parameter entities from the training dataset through entity extraction and relational reasoning, and constructs a causal relationship network between entities based on co-occurrence relationships.

5. The method for analyzing and processing simulation training data based on big data according to claim 1, characterized in that, The Transformer-based operation sequence prediction model includes a multi-head self-attention mechanism to capture long-distance dependencies in operation sequences and improves the prediction accuracy for rare operation patterns through a contrastive learning algorithm.

6. The method for analyzing and processing simulation training data based on big data according to claim 1, characterized in that, The normal behavior model constructed by the variational autoencoder constrains the distribution of latent variables through the KL divergence loss function. When the reconstruction error between the predicted future operating state and the historical normal operating mode exceeds a preset threshold, a hierarchical alarm mechanism is triggered.

7. The method for analyzing and processing simulation training data based on big data according to claim 1, characterized in that, The reinforcement learning training parameter optimization model adopts the policy gradient algorithm. The feature vector space includes a state space and an action space. The state space includes the feature profile, the current training progress and the historical training effect. The action space includes adjustable parameters for adjusting the training scenario complexity and selecting the feedback type.

8. The method for analyzing and processing simulation training data based on big data according to claim 3, characterized in that, The response time of the alarm mechanism does not exceed a preset threshold, which is dynamically adjusted according to the risk level of the training task.

Images