An intelligent model construction method and system for trainee training data analysis

Abstract

The application relates to the technical field of intelligent training and data analysis, in particular to an intelligent model construction method and system for trainee training data analysis. The method comprises the following steps: constructing an SOP time sequence knowledge graph, acquiring standard operation procedure documents and expert-level operation demonstration data, and constructing an SOP time sequence knowledge graph; real-time feature extraction, capturing a multi-modal data stream, and generating a real-time behavior feature vector; state recognition and path deduction, inputting the real-time behavior features into an event classifier to recognize key events, and performing path deduction based on the certainty finite state automaton logic in the SOP time sequence knowledge graph to determine and output a current state pointer; deviation analysis and judgment, extracting a compliance mask based on the current state, and injecting the compliance mask into the decoding process of an attention network; feedback and adjustment, and generating corresponding deviation event signals. Deep analysis of the application makes feedback more targeted, and significantly improves the accuracy and fineness of training evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent training and data analysis, in particular to an intelligent model construction method and system for trainee training data analysis. BACKGROUND

[0002] In related technologies, the traditional method has limitations for training data analysis of complex operation skills of trainees. On the one hand, although the traditional rule engine can define rigid operation logic, it is difficult to effectively process continuous and high-dimensional multi-modal time series data such as joystick displacement and eye movement signals. On the other hand, although a pure deep learning neural network model is good at processing time series features, it lacks logical explainability in strictly following standard operating procedures and cannot impose hard process constraints.

[0003] This separation of logical constraints and feature analysis mode makes it impossible for the system to make deep attribution diagnosis of the trainee's operation. When the trainee makes a mistake, the system cannot accurately distinguish whether the deviation is due to a cognitive error of the operation process or a lack of operation execution accuracy.

[0004] In addition, due to the lack of accurate diagnosis and quantitative ability profile, the existing training system cannot dynamically adjust subsequent tasks according to the individual weaknesses of the trainee, and cannot form an efficient self-adaptive closed-loop training. Therefore, an intelligent analysis architecture that deeply couples hard logic rules and soft feature analysis is needed to solve the problems of insufficient diagnostic ability and lack of adaptive feedback. SUMMARY

[0005] To solve the above technical problems, the present application discloses an intelligent model construction method and system for trainee training data analysis. Specifically, the technical solution of the present application is:

[0006] An intelligent model construction method for trainee training data analysis, comprising:

[0007] Constructing a SOP time series knowledge graph, obtaining standard operating procedure documents and expert-level operation demonstration data, and constructing a SOP time series knowledge graph based thereon, including state migration kits, expert benchmark support and skill confidence threshold.

[0008] Real-time feature extraction, capturing multi-modal data streams, and using a multi-modal time series encoder to extract features from the data streams to generate real-time behavior feature vectors.

[0009] State recognition and path deduction, inputting real-time behavior features into an event classifier to recognize key events, and performing path deduction based on deterministic finite state automaton logic in the SOP time series knowledge graph to determine and output a current state pointer.

[0010] Bias analysis and judgment, based on the current state to extract compliance mask, inject it into the decoding process of attention network; comprehensive real-time behavior characteristics warning, compliance mask and expert benchmark warning, judge whether there is doctrine bias and skill bias.

[0011] Feedback and adjustment, generate corresponding bias event signal, update personal ability radar chart according to the signal, and dynamically adjust the scene parameters of subsequent tasks based on the radar chart.

[0012] Preferably, the SOP timing knowledge graph is constructed, comprising:

[0013] Semantic analysis is performed on the standard operating procedure document;

[0014] Identify key operation nodes and their timing logical relationship;

[0015] Construct state transition skeleton;

[0016] Define the preconditions, allowed action set and forbidden action set in each logical state;

[0017] Statistical analysis is performed on the expert-level operation demonstration data;

[0018] Calculate the high-dimensional feature mean, variance and dynamic distribution boundary of the standard action in different states;

[0019] Generate expert benchmark vector;

[0020] The expert benchmark vector is anchored on the compliance action node corresponding to the SOP timing knowledge graph.

[0021] Preferably, the judgment whether there is doctrine bias comprises:

[0022] Set the weight of compliance action under the current state to 1 and the weight of illegal action to 0 to generate the compliance mask;

[0023] Use the compliance mask to suppress the prediction probability of illegal action in the attention network;

[0024] Map the original maximum probability action according to the real-time behavior characteristic vector;

[0025] Check the weight of the original maximum probability action in the compliance mask;

[0026] In response to the weight being 0, it is determined that there is doctrine bias.

[0027] Preferably, the judgment whether there is skill bias comprises:

[0028] in response to the weight of the original maximum probability action in the compliance mask being non-zero, obtaining the expert reference vector mounted by the current logical state node;

[0029] calculating a similarity score between the real-time behavior feature vector and the expert reference vector;

[0030] comparing the similarity score with the skill confidence threshold;

[0031] in response to the similarity score being lower than the skill confidence threshold, determining that the skill deviation exists.

[0032] Preferably, the scene parameters of the subsequent task are adjusted, including:

[0033] quantitatively counting the frequency of each type of deviation event signal;

[0034] updating the personal ability radar chart of the trainee based on the frequency;

[0035] identifying weak links according to the personal ability radar chart;

[0036] in the next training generation, increasing the frequency of occurrence of the scene parameters corresponding to the weak links or increasing the intensity of the auxiliary prompts.

[0037] An intelligent model construction system for trainee training data analysis, comprising:

[0038] a graph construction module, configured to obtain standard operation procedure documents and expert-level operation demonstration data, and construct a SOP time sequence knowledge graph based on the standard operation procedure documents and the expert-level operation demonstration data;

[0039] a data perception module, configured to collect multi-modal data streams of trainees, and use a multi-modal time sequence encoder to extract features of the multi-modal data streams to generate a real-time behavior feature vector;

[0040] a state tracking module, configured to identify key events based on the real-time behavior feature vector, and perform path deduction on the key events in the SOP time sequence knowledge graph to determine a current logical state node and output a current state pointer;

[0041] a coupling analysis module, configured to extract a compliance mask based on the current state pointer, inject the compliance mask into a decoding process of an attention network, and determine whether a doctrine deviation or a skill deviation exists based on the real-time behavior feature vector, the compliance mask and an expert reference vector, to generate a corresponding deviation event signal;

[0042] a closed-loop feedback module, configured to update a personal ability radar chart based on the deviation event signal, and adjust scene parameters of a subsequent task based on the personal ability radar chart.

[0043] Preferably, the graph construction module is configured to:

[0044] perform semantic parsing on the SOP document to identify key operation nodes and their temporal logic relationships, and construct a state transition skeleton;

[0045] define preconditions, allowed action sets and forbidden action sets in each logical state;

[0046] perform statistical analysis on the expert-level operation demonstration data to calculate high-dimensional feature mean values, variances and dynamic distribution boundaries of standard actions in different states, and generate an expert reference vector;

[0047] anchor the expert reference vector to the compliant action nodes corresponding to the SOP temporal knowledge graph.

[0048] Preferably, the coupling analysis module is configured to:

[0049] set the weight of a compliant action in the current state to 1 and the weight of a non-compliant action in the current state to 0 to generate the compliance mask;

[0050] use the compliance mask to suppress the prediction probability of non-compliant actions in the attention network;

[0051] map the real-time behavior feature vector to the original maximum probability action;

[0052] in response to the weight of the original maximum probability action in the compliance mask being 0, determine that the doctrine deviation exists.

[0053] Preferably, the coupling analysis module is further configured to:

[0054] in response to the weight of the original maximum probability action in the compliance mask being non-zero, obtain the expert reference vector mounted by the current logical state node;

[0055] calculate the similarity score between the real-time behavior feature vector and the expert reference vector;

[0056] in response to the similarity score being lower than the skill confidence threshold, determine that the skill deviation exists.

[0057] Compared with the prior art, the present application has the following beneficial effects:

[0058] 1、The present application realizes the fine differentiation of the operation deviation of the trainee by constructing the SOP time sequence knowledge graph and combining the compliance mask and the expert benchmark vector. The method can effectively identify the doctrine deviation and the skill deviation, and solves the problem that the traditional evaluation cannot accurately locate the error nature. This deep analysis makes the feedback more targeted, and significantly improves the accuracy and fineness of the training evaluation.

[0059] 2、The present application uses the SOP time sequence knowledge graph to carry out path deduction, and extracts the decoding process of the compliance mask injection attention network. This design provides a hard constraint of SOP rules for intelligent analysis. By real-time inhibiting the prediction probability of illegal action, it ensures that the evaluation is always anchored in the framework of the standard operating procedure. This effectively avoids the deviation of model analysis from the SOP logic, and significantly improves the compliance, objectivity and reliability of the training data analysis.

[0060] 3、The present application constructs an adaptive training closed loop from analysis to feedback. The system updates the personal ability radar chart in real time based on the deviation event signal, accurately identifies the weak link of the trainee, and dynamically adjusts the scene parameters of the subsequent task according to the weak link, such as increasing the frequency of the weak link in the next training or increasing the auxiliary prompt. This design makes the training no longer fixed, but realizes personalized guidance for thousands of people, can automatically target reinforcement, and significantly improves the training efficiency and effect.

[0061] 4、The present application realizes the unified modeling of procedures and skills by constructing the SOP time sequence knowledge graph. The graph not only contains the state transition skeleton and logic rules parsed from the SOP document, but also anchors the expert benchmark vector generated by statistical expert data; this makes the graph a single reliable data base that integrates rigid SOP procedures and flexible expert experience; all analysis is based on this, ensuring that the evaluation is both compliant and professional, and improving the robustness and authority of the model. BRIEF DESCRIPTION OF DRAWINGS

[0062] The present application will be further explained in conjunction with the accompanying drawings and embodiments:

[0063] Figure 1 is a method flowchart of the present application.

[0064] Figure 2 is a system structure diagram of the present application. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail in conjunction with specific embodiments.

[0066] Embodiment 1

[0067] Please refer to Figure 1A method for intelligent model construction of trainee training data analysis, comprising:

[0068] SOP timing knowledge graph is constructed, standard operation procedure documents and expert-level operation demonstration data are obtained, and on this basis, SOP timing knowledge graph is constructed, including state transition kit, expert benchmark support and skill confidence threshold;

[0069] Real-time feature extraction, capture multi-modal data stream, and use multi-modal timing encoder to extract features from data stream to generate real-time behavior feature vector;

[0070] State recognition and path deduction input real-time behavior features into event classifier to identify key events, and perform path deduction based on deterministic finite state automaton logic in SOP timing knowledge graph to determine and output current state pointer;

[0071] Deviation analysis and judgment, based on the current state, extract compliance mask, and inject it into the decoding process of attention network; Integrate real-time behavior feature warning, compliance mask and expert benchmark warning to determine whether there is doctrine deviation and skill deviation;

[0072] Feedback and adjustment, generate corresponding deviation event signal, update personal ability radar chart according to the signal, and dynamically adjust scene parameters of subsequent tasks based on the radar chart;

[0073] The scheme proposes an intelligent analysis architecture with logical-feature dual coupling; in the initialization stage, the system performs data import operation, retrieves electronic standard operation procedure documents as true value source of logical constraint, and synchronously loads expert-level operation demonstration data to establish quantitative quality benchmark; based on this, the system constructs the core Timing knowledge graph; The timing knowledge graph is defined as a hybrid model containing directed graph structure and high-dimensional feature embedding, the nodes represent discrete logical states, the edges represent state transition conditions, and the nodes are mounted with statistical distribution vectors representing expert operation features;

[0074] In the training execution stage, the data perception interface continuously collects the multi-modal data stream of the trainee, which covers the joystick displacement, the discrete sequence of keys, the eye gaze point coordinates and the heterogeneous data of physiological electrical signals; in order to realize the machine readability of the data, the system calls the pre-trained multi-modal timing encoder to perform noise reduction, space-time alignment and feature mapping processing on the original signal, and outputs continuous high-dimensional real-time behavior feature vector;

[0075] Specifically, the encoder can adopt an attention mechanism-based late fusion architecture; for example, for continuous time series signals such as joystick displacement, a one-dimensional convolutional neural network is used to extract local features; for discrete sequences of keys, an embedding layer is used to map them to a high-dimensional vector space; for eye gaze point coordinates and physiological electrical signals, they are also processed by their respective feature extractors, and the modal features are aligned and spliced in the time dimension, and finally input into the backbone network of a Transformer encoder, and through the self-attention mechanism, deep fusion and time series modeling are performed;

[0076] The dynamic tracking logic is triggered; the system analyzes the mutation features in the real-time behavior feature vector to identify key events, and performs a path search algorithm in the

[0077] The system performs hierarchical decision logic; at the cognitive compliance level, the system checks whether the action intention mapped by the real-time behavior feature vector falls within the non-zero interval of the compliance mask, to determine whether there is a doctrine deviation that violates the procedure logic; at the skill quality level, if the action is compliant, the system further calculates the distance measure between the real-time behavior feature vector and the expert benchmark vector preloaded on the graph node, to determine whether there is a skill deviation of insufficient operation precision; once any type of deviation is identified, the system generates a deviation event signal carrying an attribution label; these signals are accumulated for dynamic updating of the individual ability radar chart of the trainee, which is used as a quantitative ability portrait and directly input as a feedback parameter to the training generation engine; in the next round, the system automatically reconstructs the scene parameters according to the weak dimensions indicated by the radar chart, thereby realizing adaptive closed-loop reinforcement training;

[0078] The technical solution constructs a graph structure with coexisting logical skeleton and feature flesh, and seamlessly embeds hard logic constraints into the attention calculation of soft neural networks using the compliance mask mechanism; this design breaks the binary opposition between traditional rule engines that cannot handle continuous data and pure neural networks that lack logical explainability, and realizes deep attribution diagnosis of complex operation skills.

[0079] Embodiment 2

[0080] Constructing a time-series knowledge graph, including:

[0081] ​Semantic parsing of standard operating procedure documents;

[0082] Identifying key operation nodes and their timing logic relationships;

[0083] Building state transition skeleton;

[0084] Defining preconditions, allowed action set and forbidden action set for each logic state;

[0085] Statistical analysis of expert-level operation demonstration data;

[0086] Calculating high-dimensional feature mean, variance and dynamic distribution boundary of standard actions in different states;

[0087] Generating expert reference vectors;

[0088] Anchoring expert reference vectors to compliant action nodes in the timing knowledge graph;

[0089] Building The process of building the timing knowledge graph is designed as a parallel processing flow of logical layer construction and feature layer construction. For unstructured standard operating procedure documents, the system uses natural language processing techniques to perform deep semantic parsing. This process involves entity extraction and relationship extraction algorithms, aiming to accurately isolate key operation nodes and the strictly defined timing logic relationships between these nodes from the text. Based on the parsed topology, the system instantiates and generates a state transition skeleton. For each logic state node in the skeleton, the system explicitly defines its state attributes according to rules, including preconditions that must be met to enter the state, the allowed action set consisting of legal operation instructions in that state, and the forbidden action set that explicitly triggers a violation determination;

[0090] The system performs high-dimensional statistical analysis on a large amount of expert-level operation demonstration data. Considering the slight physical fluctuations within the compliance range of expert operations, this scheme does not use single trajectory matching, but uses probability distribution modeling;

[0091] The system calculates the multi-dimensional statistics of standard actions in a specific state, including high-dimensional feature mean, variance reflecting operation stability, and dynamic distribution boundary defining the range of compliance fluctuations. Based on these statistical parameters, the system generates expert reference vectors, which are essentially a parameter set containing feature center points and covariance matrices. Through graph mapping algorithms, the system anchors these generated expert reference vectors to the corresponding compliant action nodes in the timing knowledge graph, thereby giving the logic skeleton a quantitative discrimination standard.

[0092] Example 3

[0093] determining whether there is a doctrine deviation, comprising:

[0094] setting the compliance action weight in the current state to 1 and the violation action weight to 0 to generate a compliance mask;

[0095] suppressing the prediction probability of the violation action in the attention network by using the compliance mask;

[0096] mapping the original maximum probability action according to the real-time behavior feature vector;

[0097] checking the weight of the original maximum probability action in the compliance mask;

[0098] in response to the weight being , determining that there is a doctrine deviation;

[0099] The determination logic of the doctrine deviation depends on the mathematical blocking mechanism of the hard constraint mask. After determining the current logical state node, the system retrieves the allowed action set of the node, constructs a vector with the same length as the model output vocabulary, that is, the compliance mask. In the construction process of the vector, the system sets the weight of the index position corresponding to the allowed action set to 1, and forcibly sets the weight of the index position corresponding to the prohibited action set and irrelevant action to 0. In the reasoning stage of the attention network, the system applies this compliance mask to the layer before the layer, which aims to physically suppress the prediction probability of the violation action in the attention network and force the attention focus of the model to only distribute within the compliance operation space;

[0100] To diagnose the real intention of the student, the system processes the real-time behavior feature vector in parallel, and maps the original maximum probability action through the classifier not disturbed by the mask. The action represents the operation instruction that the student actually tries to perform at the current time. The determination logic then performs conflict checking, that is, retrieving the weight value corresponding to the original maximum probability action in the compliance mask. If the checking result shows that the weight is , it mathematically proves that the student's intention falls into the prohibited space or logical blind area of the current state, indicating that the student does not meet the preconditions or tries to operate or execute the wrong process steps. Accordingly, in response to the checking result that the weight is , the system triggers an interrupt signal and determines that there is a doctrine deviation. This mechanism ensures zero-tolerance identification of violations of principles and avoids ambiguous explanations that may be produced by the probability model.

[0101] Embodiment 4

[0102] determining whether there is a skill deviation, comprising:

[0103] in response to the weight of the original maximum probability action in the compliance mask being , obtain an expert reference vector mounted on the current logical state node;

[0104] calculate a similarity score between the real-time behavior feature vector and the expert reference vector;

[0105] compare the similarity score with a skill confidence threshold;

[0106] in response to the similarity score being lower than the skill confidence threshold, determine that there is a skill deviation;

[0107] the determination of the skill deviation is initiated after the process compliance confirmation; once the system confirms that the weight of the original maximum probability action in the compliance mask is not , it indicates that the trainee passes the cognitive logic check, and the evaluation focus is immediately switched to the operation performance dimension; the system immediately accesses the time sequence knowledge graph, and calls the expert reference vector mounted on the current logical state node; the vector carries the time sequence form feature of the specific action in the ideal state;

[0108] the system performs vector space analysis to calculate a similarity score between the real-time behavior feature vector and the expert reference vector; the score is obtained by calculating the cosine similarity between the two vectors or the normalized reciprocal based on the Mahalanobis distance, and the numerical range is usually normalized to to to quantitatively represent the degree of coincidence between the trainee's operation and the expert standard;

[0109] The system then compares the real-time calculated similarity score with a preset skill confidence threshold; the skill confidence threshold is not arbitrarily set, but is set according to the statistical distribution characteristics of the expert data, and is usually set to the value of the mean of the expert features minus two or three times the standard deviation, to ensure that only the operation within the reasonable statistical interval is recognized as qualified; when the comparison result shows that the similarity score is lower than the skill confidence threshold, the system identifies that the trainee's operation quality deviates significantly although the action intention is correct, and determines that there is a skill deviation.

[0110] Embodiment 5

[0111] adjust the scene parameters of the subsequent tasks, including:

[0112] quantitatively count the frequency of each type of deviation event signal;

[0113] update the personal ability radar chart of the trainee based on the frequency;

[0114] identify the weak link according to the personal ability radar chart;

[0115] when generating the next training, increase the appearance frequency of the scene parameters corresponding to the weak link or increase the intensity of the auxiliary prompt;

[0116] To build an adaptive closed-loop training system, the system implements a data-based feedback adjustment strategy; the system listens and captures all deviation event signals generated during the entire training process in real time, classifies and aggregates these signals, and quantitatively counts the frequency of various types of deviation event signals; based on these statistical data, the system executes an update algorithm to dynamically correct the individual ability radar chart of the trainee; this radar chart is represented as a multi-dimensional state vector in data structure, with each dimension corresponding to a specific skill module in the individual ability radar chart, and the numerical value is positively correlated with the probability of unbiased operation under this module;

[0117] In the initialization phase of the next round of training tasks, the scenario generation engine analyzes the updated individual ability radar chart, identifies weak links through a threshold filtering algorithm, i.e., those ability dimensions that are significantly lower than the pre-set baseline; for these identified weaknesses, the system dynamically adjusts the parameters when generating the configuration file, increasing the frequency of scenario parameters corresponding to weak links, or increasing the intensity of auxiliary prompts in the visualization interface; this parameter reconstruction mechanism based on the ability profile ensures that every minute of training time is focused on the areas most in need of improvement for the trainee, maximizing training efficiency.

[0118] Embodiment 6

[0119] Referring to Figure 2 , the atlas construction module is used to obtain standard operation procedure documents and expert-level operation demonstration data, and based on the standard operation procedure documents and expert-level operation demonstration data, to construct a time-series knowledge graph;

[0120] The data perception module is used to collect the multi-modal data stream of the trainee, and use a multi-modal time-series encoder to extract features from the multi-modal data stream to generate real-time behavior feature vectors;

[0121] The state tracking module is used to identify key events based on real-time behavior feature vectors, and to perform path deduction of key events in the time-series knowledge graph, determine the current logical state node and output the current state pointer;

[0122] The coupling analysis module is used to extract a compliance mask based on the current state pointer, inject the compliance mask into the decoding process of the attention network, and based on the real-time behavior feature vectors, the compliance mask, and the expert benchmark vector, determine whether there is a doctrine deviation or a skill deviation, and generate corresponding deviation event signals;

[0123] The closed-loop feedback module is used to update the individual ability radar chart based on the deviation event signals, and adjust the scenario parameters of subsequent tasks based on the individual ability radar chart;

[0124] The atlas construction module is used to:

[0125] Semantic parsing of standard operating procedure documents, identification of key operation nodes and their timing logic relationships, and construction of state transition skeletons;

[0126] Definition of preconditions, allowed action sets, and prohibited action sets in each logical state;

[0127] Statistical analysis of expert-level operation demonstration data, calculation of high-dimensional feature mean, variance, and dynamic distribution boundaries of standard actions in different states, and generation of expert reference vectors;

[0128] Anchoring expert reference vectors to compliant action nodes corresponding to the timing knowledge graph;

[0129] The coupling analysis module is used to:

[0130] Set the weight of the compliant action in the current state to , and set the weight of the non-compliant action in the current state to , generating a compliance mask;

[0131] Using the compliance mask to suppress the prediction probability of non-compliant actions in the attention network;

[0132] Mapping the original maximum probability action according to the real-time behavior feature vector;

[0133] In response to the weight of the original maximum probability action in the compliance mask being , it is determined that there is a doctrine deviation;

[0134] The coupling analysis module is also used to:

[0135] In response to the weight of the original maximum probability action in the compliance mask being non- , the expert reference vector mounted by the current logical state node is obtained;

[0136] Calculate the similarity score between the real-time behavior feature vector and the expert reference vector;

[0137] In response to the similarity score being lower than the skill confidence threshold, it is determined that there is a skill deviation;

[0138] The intelligent model construction system is architected as a software and hardware deep collaboration analysis platform, aiming to realize intelligent alignment and diagnosis of standard operating procedures and student behavior data;

[0139] The atlas construction module is configured to perform fusion processing of double-source data; it receives standard operating procedure document input, performs semantic analysis using a natural language processing engine, accurately identifies key operation nodes and their embedded timing logic relationships, and instantiates a state transition skeleton in memory; to implement strict logical constraints, the module is programmed to define the preconditions, allowed action set and prohibited action set in each logical state; simultaneously, the module performs high-dimensional statistical modeling on expert-level operation demonstration data, calculates the characteristic mean, variance and dynamic distribution boundary of the standard action in each state, generates an expert reference vector based thereon, and anchors it on the compliance action node corresponding to the timing knowledge graph;

[0140] The data perception module is configured with a high-frequency sensor array to collect multi-modal data streams of the trainee; the module is built-in with a multi-modal timing encoder that uses deep neural network technology to denoise, synchronize and feature extract heterogeneous raw data, and outputs standardized real-time behavior feature vectors in real time, providing high signal-to-noise ratio input for subsequent analysis;

[0141] The state tracking module runs a real-time path matching algorithm; it analyzes real-time behavior feature vectors to identify key events, and performs path deduction in the timing knowledge graph based on these events, thereby accurately locating the trainee's progress in the logical space, determining the current logical state node and outputting the current state pointer;

[0142] The coupling analysis module performs dual constraint interpretation; it extracts a compliance mask based on the state pointer and injects the mask into the decoding process of the attention network; by setting the compliance action weight to and the violation action weight to in matrix operations, the compliance mask mathematically suppresses the prediction probability of violation actions in the attention network; in the judgment logic, the module maps the original maximum probability action from the real-time behavior feature vector, and if the weight of the action in the compliance mask is , it is determined that there is a doctrine deviation according to the logic conflict principle; if the action is compliant, the module further acquires the corresponding expert reference vector, calculates its similarity score with the real-time vector, and when the score is lower than the skill confidence threshold, it is determined that there is a skill deviation;

[0143] The closed-loop feedback module is responsible for closing the training loop; it collects all deviation event signals, quantitatively updates the trainee's personal ability radar chart, and automatically adjusts the scene parameters of subsequent tasks based on the skill short board identified by the radar chart, achieving targeted teaching intervention.

[0144] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced, without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. An intelligent model building method for trainee training data analysis, characterized in that, Comprise: SOP timing knowledge graph is constructed, standard operation procedure documents and expert level operation demonstration data are obtained, and on this basis, SOP timing knowledge graph is constructed, including state transition kit, expert benchmark support and skill confidence threshold; Real-time feature extraction, capture multi-modal data stream, and use multi-modal timing encoder to extract features from data stream to generate real-time behavior feature vector; State recognition and path deduction, input real-time behavior features into event classifier to identify key events, and perform path deduction based on deterministic finite state automaton logic in SOP timing knowledge graph to determine and output current state pointer; Deviation analysis and judgment, extract compliance mask based on current state, and inject it into the decoding process of attention network; Integrate real-time behavior feature vector, compliance mask and expert benchmark vector to determine whether there is doctrine deviation and skill deviation; Feedback and adjustment, generate corresponding deviation event signal, update personal ability radar chart according to the signal, and dynamically adjust scene parameters of subsequent tasks based on the radar chart; The SOP timing knowledge graph is constructed, comprising: Semantic analysis of the standard operation procedure documents; Identify key operation nodes and their timing logic relationship; Construct state transition skeleton; Define the preconditions, allowed action set and forbidden action set in each logical state; Statistical analysis of the expert level operation demonstration data; Calculate the high-dimensional feature mean, variance and dynamic distribution boundary of the standard action in different states; Generate expert benchmark vector; Anchoring the expert benchmark vector on the compliance action node corresponding to the SOP timing knowledge graph; The judgment whether there is doctrine deviation, comprising: Set the compliance action weight in the current state to 1 and the illegal action weight to 0 to generate the compliance mask; Use the compliance mask to suppress the prediction probability of illegal actions in the attention network; Map the original maximum probability action according to the real-time behavior feature vector; Check the weight of the original maximum probability action in the compliance mask; In response to the weight being 0, it is determined that there is doctrine deviation; The judgment whether there is skill deviation, comprising: In response to the weight of the original maximum probability action in the compliance mask being non-zero, obtain the expert benchmark vector mounted on the current logical state node; Calculate the similarity score between the real-time behavior feature vector and the expert benchmark vector; Compare the similarity score with the skill confidence threshold; In response to the similarity score being lower than the skill confidence threshold, it is determined that there is skill deviation. 2.The intelligent model building method for trainee training data analysis of claim 1, wherein, Adjusting the scene parameters of subsequent tasks, comprising: Quantitative statistics of the frequency of each type of deviation event signal; Update the personal ability radar chart of the trainee based on the frequency; Identify weak links according to the personal ability radar chart; When generating the next training, increase the appearance frequency of the scene parameters corresponding to the weak links or increase the intensity of the auxiliary prompts.

3. An intelligent model building system for trainee training data analysis based on the intelligent model building method for trainee training data analysis according to any one of claims 1-2, characterized in that, Comprise: A graph construction module is configured to obtain standard operation procedure documents and expert level operation demonstration data, and construct a SOP timing knowledge graph based on the standard operation procedure documents and the expert level operation demonstration data; The data perception module is configured to collect multi-modal data streams of the trainee and extract features from the multi-modal data streams by using a multi-modal time sequence encoder to generate a real-time behavior feature vector; The state tracking module is configured to identify a key event based on the real-time behavior feature vector, perform path deduction on the key event in the SOP time sequence knowledge graph, determine a current logical state node, and output a current state pointer; The coupling analysis module is configured to extract a compliance mask based on the current state pointer, inject the compliance mask into a decoding process of an attention network, and determine whether there is a doctrine deviation or a skill deviation based on the real-time behavior feature vector, the compliance mask, and an expert benchmark vector, and generate a corresponding deviation event signal; The closed-loop feedback module is configured to update a personal ability radar chart based on the deviation event signal, and adjust scene parameters of a subsequent task based on the personal ability radar chart.

4. The intelligent model building system for trainee training data analysis as claimed in claim 3 wherein, The graph construction module is configured to: perform semantic analysis on the standard operation procedure document, identify key operation nodes and their time sequence logical relationships, and construct a state transition skeleton; define preconditions, an allowed action set, and a prohibited action set in each logical state; perform statistical analysis on the expert-level operation demonstration data, calculate high-dimensional feature mean values, variances, and dynamic distribution boundaries of standard actions in different states, and generate an expert benchmark vector; anchor the expert benchmark vector on a compliance action node corresponding to the SOP time sequence knowledge graph.

5. The intelligent model building system for trainee training data analysis as claimed in claim 4 wherein, The coupling analysis module is configured to: set a compliance action weight in a current state to 1 and a non-compliance action weight in the current state to 0 to generate the compliance mask; suppress a prediction probability of a non-compliance action in the attention network by using the compliance mask; map an original maximum probability action according to the real-time behavior feature vector; in response to a weight of the original maximum probability action in the compliance mask being 0, determine that there is the doctrine deviation.

6. The intelligent model building system for trainee training data analysis as claimed in claim 5 wherein, The coupling analysis module is further configured to: in response to the weight of the original maximum probability action in the compliance mask being non-0, obtain the expert benchmark vector mounted on the current logical state node; calculate a similarity score between the real-time behavior feature vector and the expert benchmark vector; in response to the similarity score being lower than a skill confidence threshold, determine that there is the skill deviation.

Images