A multi-dimensional behavior clustering-based professional skill practical training path dynamic generation method and system
By using a composite distance metric and a reinforcement learning path decision model, the problem of feature distortion and path lag in multi-dimensional heterogeneous data in existing technologies is solved, and the accurate dynamic generation and real-time optimization of vocational skills training paths are realized.
Patent Information
- Application Number
- CN202611325737.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-10
AI Technical Summary
Existing vocational skills training platforms lack effective measurement methods when faced with multi-dimensional and heterogeneous micro-behavioral data, resulting in the failure of user group feature extraction. Furthermore, dynamic path recommendation models lack underlying execution logic topological constraints, making it impossible to accurately identify operational error patterns and implicit cognitive bottlenecks, leading to delayed training path generation and illegal node output.
A composite distance metric algorithm is adopted to integrate dynamic time-warped distance, relative entropy divergence and Euclidean distance to accurately measure and topologically cluster multi-dimensional behavioral data. Combined with professional skills knowledge graph and deep Q network, a reinforcement learning path decision model is constructed to realize real-time dynamic path generation.
It achieves accurate measurement and topological clustering of heterogeneous training behavior data, improves the accuracy and convergence efficiency of training paths, and enables step-by-step path optimization and real-time intervention while strictly following the teaching logic.
Smart Images

Figure CN123071121A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of data processing and smart education technology, specifically to a method and system for dynamically generating vocational skills training paths based on multi-dimensional behavior clustering. Background Technology
[0002] In the current fields of artificial intelligence and smart education technology, vocational skills training platforms primarily provide training environments for users through virtual simulations or online simulators. To improve training efficiency, these platforms typically need to assign personalized training paths to different users. Existing training path generation systems mostly rely on static rule mapping or single-dimensional collaborative filtering recommendation mechanisms. This means that users are categorized into fixed ability levels based on their periodic test scores or subjectively filled-in static tags, and then pre-set static course packages are distributed. However, with the increasing complexity of training data, this conventional path generation technology exposes significant underlying technical deficiencies, specifically manifested in the following two core technical issues: First, existing systems lack effective measurement methods for multi-dimensional, heterogeneous micro-behavioral data in the feature extraction and clustering analysis stages. Vocational skills training not only includes static performance results but also massive amounts of time-series discrete operational data (such as trigger sequences of specific interaction events and the frequency of undo / redo) and time-series continuous cognitive data (such as gaze trajectories and operation pause times). Traditional clustering algorithms (such as K-Means) typically only use Euclidean distance to measure static, uniform-dimensional data. When faced with heterogeneous data matrices containing discrete time series and continuous probability distributions, conventional Euclidean distance cannot measure the misalignment similarity of time series steps and the differences in probability density distribution. Forced data alignment and splicing lead to severe distortion of information in the high-dimensional feature space. This deficiency in measurement algorithms prevents existing systems from accurately identifying user groups in the same "micro-operational error pattern" or "implicit cognitive bottleneck," resulting in the failure of group feature extraction.
[0003] Secondly, existing dynamic path recommendation models lack hard constraints on the underlying execution logic topology when making action decisions, and cannot perform step-by-step real-time optimization based on multimodal micro-features. When using conventional reinforcement learning (such as standard DQN) for path recommendation, the model blindly explores and tries in a vast action space, which can easily lead to the output of illegal training nodes that violate the strict pre-requirement logic of professional skills (e.g., issuing troubleshooting tasks directly without mastering the basic equipment startup steps). At the same time, due to the lack of accurate prior group feature inputs and combined reward function constraints for micro-indicators such as operation fluency and thinking time, the convergence speed of the decision engine is extremely slow and the output has obvious lag, making it impossible to intervene and reconstruct the path in real time based on the user's real-time operation delays or sudden changes in error frequency during a single training task. Summary of the Invention
[0004] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a method for dynamically generating vocational skills training paths based on multi-dimensional behavior clustering, including the following steps: S100, collecting multi-dimensional behavior data of target users during the training process, wherein the multi-dimensional behavior data includes at least state dimension features, time-series discrete operation dimension features and time-series continuous cognitive dimension features, and performing timestamp alignment processing on the multi-dimensional behavior data to generate a time series feature vector; S200, based on the time series feature vector, a composite distance metric algorithm is used to calculate the behavioral similarity distance between the target user and historical user samples; wherein, the composite distance metric algorithm integrates the dynamic time warping distance used to measure the operational dimension features, the relative entropy divergence used to measure the cognitive dimension features, and the Euclidean distance used to measure the state dimension features. S300, perform clustering processing on users based on the calculated behavioral similarity distance, determine the target cluster to which the target user currently belongs, and extract the centroid features of the target cluster; S400, a reinforcement learning path decision model is constructed based on a pre-built vocational skills knowledge graph and a deep Q-network, wherein the nodes of the vocational skills knowledge graph represent training tasks and the edges represent the execution dependencies between tasks; S500, the centroid features of the target cluster and the current knowledge node state of the target user are input into the reinforcement learning path decision model as the environment state space. The reinforcement learning path decision model outputs the next optimal training task node to dynamically generate a training path for the target user.
[0005] Furthermore, the specific features in the multi-dimensional behavioral data include: The state dimension features include the static basic profile data of the target user and the skill node mastery index calculated based on historical training records. The time-series discrete operation dimension features include the discrete interaction event sequence triggered within the training operation interface, operation error type classification identifiers, and step undo / redo frequency. The temporally continuous cognitive dimension features include the gaze-staying coordinate stream collected by the sensors of the training terminal, the attention heat value generated based on the cursor trajectory, and the pause and thinking time in key operation stages.
[0006] Furthermore, the process of performing timestamp alignment on the multi-dimensional behavioral data to generate a time-series feature vector specifically includes the following sub-steps: S101, use a time sliding window of preset length to slice the single training cycle of the target user to obtain several consecutive time windows; S102, within each time window, the time-series discrete operation dimension features are one-hot encoded according to the timestamp of occurrence to obtain operation feature sub-vectors; S103, a polynomial interpolation algorithm is used to smooth and resample the temporally continuous cognitive dimension features within the time window to unify the data frequency and obtain cognitive feature sub-vectors. S104, the state dimension features, the operation feature sub-vectors and the cognitive feature sub-vectors within the same time window are concatenated, and feature dimensionality reduction is performed through a fully connected network layer to generate the time series feature vector corresponding to the time window.
[0007] Furthermore, the step of calculating the behavioral similarity distance between the target user and historical user samples using a composite distance metric algorithm based on the time-series feature vector specifically includes: Extract the time-series feature vectors of the target user and the historical user samples respectively; For the extracted time series feature vector, the dynamic time warping algorithm is used to calculate the dynamic time warping distance of the operation dimension feature sequence between the target user and the historical user sample; The relative entropy divergence of the probability distribution of the cognitive dimension feature sequence between the target user and the historical user samples is calculated using the KL divergence algorithm. The Euclidean distance between the target user and the historical user sample is calculated using the Euclidean distance formula. Obtain the adaptive weight coefficients corresponding to the operation dimension feature, the cognition dimension feature and the state dimension feature respectively, and sum each of the adaptive weight coefficients with the corresponding dynamic time warp distance, the relative entropy divergence and the Euclidean distance to output the final behavior similarity distance.
[0008] Furthermore, the specific weight allocation logic for obtaining the adaptive weight coefficients corresponding to the operational dimension features, the cognitive dimension features, and the state dimension features is as follows: The numerical distribution characteristics of historical user samples in the operation dimension, the cognition dimension, and the state dimension are statistically analyzed, and the information entropy corresponding to each dimension is calculated respectively. The dispersion and information content of each feature dimension are evaluated based on the information entropy. Based on the inverse mapping rule that the smaller the information entropy, the greater the information content and the higher the assigned weight, the initial weight of each dimension is determined respectively. The initial weights of each dimension are normalized so that the sum of the adaptive weight coefficients of the normalized operational dimension features, cognitive dimension features, and state dimension features is one.
[0009] Further, the step of clustering users based on the calculated behavioral similarity distance to determine the target cluster to which the target user currently belongs specifically includes: Construct a global distance matrix containing the target user and each historical user sample, where the matrix elements are the corresponding behavioral similarity distances; Based on the global distance matrix and the preset cutoff distance, the local density value of each user is calculated. The local density value represents the number of other users adjacent to the user within the cutoff distance range. Calculate the minimum relative distance between each user and other neighboring users with higher local density values; A cluster center decision graph is constructed based on the product of the local density value and the minimum relative distance, and several cluster centers are selected from the cluster center decision graph according to the numerical descent gradient to form multiple candidate clusters. The target user is recursively assigned to a candidate cluster containing the nearest neighboring user with a higher local density value, and this candidate cluster is used as the target cluster to which the target user currently belongs.
[0010] Furthermore, the extraction of the centroid features of the target cluster specifically includes: Obtain the time-series feature vectors of all sample users within the target cluster; Calculate the distance decay factor from the cluster center to each sample user within the cluster, and use it as the fusion weight; Based on the fusion weight, all the time series feature vectors within the target cluster are weighted and averaged to generate the centroid feature that can characterize the common micro-operational bottlenecks and cognitive patterns of the cluster. During the training process, background monitoring is initiated. If the variance change rate of the number of newly gathered user samples or the behavioral similarity distance within the target cluster exceeds the preset dynamic update threshold, the cluster center is recalculated and the centroid feature is updated.
[0011] Furthermore, the construction of the reinforcement learning path decision model based on the pre-built vocational skills knowledge graph and the deep Q-network specifically includes the definition of the attributes of the vocational skills knowledge graph: Each node in the professional skills knowledge graph is configured with a task attribute vector, which includes the task difficulty coefficient, standard operation time, and skill category label. Each edge in the vocational skills knowledge graph is configured with a pre-dependency weight matrix, which is used to characterize the correlation between hard preconditions and soft recommendations when moving from the current training task node to the next training task node. An action mask matrix is constructed based on the aforementioned pre-dependency weight matrix to filter out illegal paths that violate the execution logic order when the deep Q network makes output decisions; The reinforcement learning path decision model constructed based on the pre-built vocational skills knowledge graph and deep Q-network also includes setting the Markov decision environment of the deep Q-network, with the specific rules as follows: Define a state space, which includes the centroid features of the target cluster, the task attribute vector corresponding to the knowledge node where the target user is currently located, and the pass rate sequence of the target user's historical completed nodes; An action space is defined, which is a set of legal next-level candidate task nodes that can be reached in the professional skills knowledge graph after being filtered by the action mask matrix based on the knowledge node where the target user is currently located. A reward function is defined, which is a combined evaluation function. Its input variables include the percentage decrease in the target user's operation error rate after executing the next level task node, the closeness of the standard operation time, and the decay value of the step cancellation frequency. The neural network parameters in the deep Q-network are iteratively optimized with the training objective of maximizing the cumulative expected discounted value of the reward function.
[0012] Furthermore, the step of outputting the next optimal training task node through the reinforcement learning path decision model to dynamically generate a training path for the target user specifically includes: The reinforcement learning path decision model calculates the action value function Q value corresponding to each legal next-level candidate task node in the action space based on the current input environmental state space through its internal neural network layer. An exploration and utilization strategy based on confidence upper limit is adopted to select the node with the highest target Q value from the candidate task node set as the next optimal training task node; The topological relationship between the next optimal training task node and the target user's current knowledge node is analyzed and mapped to the corresponding dynamic intervention strategy. The dynamic intervention strategies include: a prior knowledge advancement strategy executed when the topological relationship is a successor node, a specialized reinforcement and retraining strategy executed when the topological relationship is a sibling node, and a basic downgrade review strategy executed when the topological relationship is a predecessor node. According to the determined dynamic intervention strategy, the corresponding training task scenario instructions are rendered in real time and seamlessly issued to the training terminal of the target user. After dynamically generating the training path for the target user, the process also includes a real-time closed-loop feedback optimization step for the reinforcement learning path decision model: Real-time monitoring of the latest multi-dimensional behavioral data of the target user during the execution of the next optimal training task node; Based on the latest multi-dimensional behavioral data, calculate the target user's real state transition characteristics and the actual reward value obtained after performing the task; The interactive experience tuple containing the environmental state space, the decision action corresponding to the target Q value, the real reward value, and the real state transition features is stored in the priority experience replay pool. Experience batches are extracted from the priority experience replay pool, and the neural network parameters of the reinforcement learning path decision model are updated using backpropagation of temporal difference error. Simultaneously, the steps of claims 1 to 3 are triggered to re-evaluate whether the target user has undergone cluster transition.
[0013] On the other hand, a dynamic generation system for vocational skills training paths based on multi-dimensional behavior clustering includes: a multi-dimensional data perception and alignment module, used to collect multi-dimensional behavior data of target users during the training process, wherein the multi-dimensional behavior data includes at least state dimension features, time-series discrete operation dimension features and time-series continuous cognitive dimension features, and timestamp alignment processing is performed on the multi-dimensional behavior data to generate a time series feature vector. The composite distance metric and fusion module is used to calculate the behavioral similarity distance between the target user and historical user samples based on the time series feature vector using a composite distance metric algorithm; wherein, the composite distance metric algorithm fuses the dynamic time warping distance used to measure the operational dimension features, the relative entropy divergence used to measure the cognitive dimension features, and the Euclidean distance used to measure the state dimension features. The target group clustering and feature extraction module is used to cluster users based on the calculated behavioral similarity distance, determine the target cluster to which the target user currently belongs, and extract the centroid features of the target cluster. A reinforcement learning path decision building module is used to build a reinforcement learning path decision model based on a pre-built vocational skills knowledge graph and a deep Q-network, wherein the nodes of the vocational skills knowledge graph represent training tasks and the edges represent the execution dependencies between tasks. The dynamic path mapping and closed-loop output module is used to input the centroid features of the target cluster and the current knowledge node state of the target user as the environmental state space into the reinforcement learning path decision model, and output the next optimal training task node through the reinforcement learning path decision model to dynamically generate a training path for the target user.
[0014] Beneficial effects: This invention introduces a composite distance metric algorithm that integrates dynamic time warp distance, relative entropy divergence and Euclidean distance, and achieves accurate measurement and topological clustering of heterogeneous data such as time-series discrete operational features, time-series continuous cognitive features and static state features. It effectively solves the problems of feature distortion and clustering failure caused by traditional Euclidean distance when processing multidimensional heterogeneous training behavioral data. This invention uses the centroid features of the target cluster obtained based on local density optimization as the prior input of the environmental state, and combines the pre-dependency weights of the professional skills knowledge graph to construct an action mask matrix to constrain the discrete action space of the deep Q network. This enables the reinforcement learning decision model to perform step-by-step closed-loop parameter iteration and path distribution based on the user's real-time micro-operation feedback, while strictly following the underlying teaching execution logic. This improves the accuracy, convergence efficiency and adaptive optimization capability of the dynamic decision-making of training paths. Attached Figure Description
[0015] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a flowchart of the multi-dimensional behavioral data alignment and feature generation process of the present invention; Figure 3 This is a schematic diagram illustrating the principle of composite distance measurement and dynamic weight allocation in this invention. Detailed Implementation
[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.
[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but includes other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0018] The present invention will now be described in further detail with reference to the accompanying drawings: Example: like Figures 1 to 3 As shown, a method for dynamically generating vocational skills training paths based on multi-dimensional behavior clustering includes the following steps: S100 collects multi-dimensional behavioral data of the target user during the training process. The multi-dimensional behavioral data includes at least state dimension features, time-series discrete operation dimension features, and time-series continuous cognitive dimension features. The multi-dimensional behavioral data is then processed with timestamp alignment to generate a time series feature vector. S200, based on time series feature vectors, uses a composite distance metric algorithm to calculate the behavioral similarity distance between target users and historical user samples; the composite distance metric algorithm integrates dynamic time warping distance for measuring operational dimension features, relative entropy divergence for measuring cognitive dimension features, and Euclidean distance for measuring state dimension features. S300: Based on the calculated behavioral similarity distance, cluster users to determine the target cluster to which the target user currently belongs, and extract the centroid features of the target cluster. S400 is a reinforcement learning path decision model built on a pre-constructed vocational skills knowledge graph and a deep Q-network. In this model, the nodes of the vocational skills knowledge graph represent training tasks, and the edges represent the execution dependencies between tasks. S500 inputs the centroid features of the target cluster and the current knowledge node state of the target user as the environmental state space into the reinforcement learning path decision model. The reinforcement learning path decision model outputs the next optimal training task node to dynamically generate a training path for the target user.
[0019] Furthermore, the specific implementation process of step S100 is as follows: Step S100: Collect multi-dimensional behavioral data of the target user during the training process. The multi-dimensional behavioral data includes at least state dimension features, time-series discrete operation dimension features, and time-series continuous cognitive dimension features. The multi-dimensional behavioral data is then processed with timestamp alignment to generate a time series feature vector.
[0020] In practice, the training system terminal is configured with a multimodal data acquisition probe engine at the bottom layer. The probe engine starts the status reading thread, event listening thread and sensor capture thread in parallel.
[0021] First, the state reading thread acquires the target user's state dimension features. These features include static basic profile data and skill node mastery indices calculated based on historical training records. The execution logic is as follows: The system database retrieves the target user's unique identifier, and the corresponding initial profile matrix composed of demographic parameters and occupational background parameters is obtained. Simultaneously, the system reads the target user's score records for all prerequisite knowledge nodes in historical training assessments, calculates a skill node mastery vector consisting of multiple floating-point values using a weighted summation algorithm, and concatenates the initial profile matrix with the skill node mastery vector to form the state dimension features.
[0022] Secondly, the event listening thread records the discrete operational characteristics of the target user within the training interface in real time. These operational characteristics include discrete interaction event sequences, operation error type classification identifiers, and step undo / redo frequency. The implementation steps include: embedding an interceptor in the underlying business logic of the training terminal's front-end code; when the system detects a mouse click, keyboard input, or 3D model dragging command, the interceptor generates a discrete event log with an absolute system timestamp. The discrete event log records the event identifier, trigger control coordinates, and input parameter values. The input parameter value is read and compared with a pre-stored table of valid operation thresholds in memory. If the input parameter value falls into an invalid range, the error mapping dictionary is retrieved, and the current discrete interaction event is marked with the corresponding operation error type classification identifier. The system backend sets a statistical time window, accumulating the absolute number of undo and redo command identifiers received within the current statistical time window. The real-time step undo / redo frequency is calculated by dividing the absolute number by the duration of the statistical time window.
[0023] Furthermore, the sensor capture thread synchronously collects temporally continuous cognitive dimension features. These features include gaze-stationary coordinate streams, attention heatmaps generated based on cursor trajectories, and pause / thinking durations during key operations. Specifically, the implementation involves: accessing the underlying driver interface of an external eye-tracking device or front-facing vision sensor connected to the training terminal; continuously outputting the two-dimensional spatial coordinates of the user's pupil gaze point on the screen at a preset first sampling rate (e.g., 120Hz); arranging these coordinates temporally to form a continuous gaze-stationary coordinate stream. A cursor tracking buffer is allocated in system memory; the absolute screen coordinate stream of the mouse or stylus is extracted at a preset second sampling rate; a two-dimensional Gaussian kernel function is used to smooth the spatial distribution of the absolute coordinate stream, generating an attention heatmap matrix representing the regional interaction density. Time-determination logic is set within the discrete interaction event sequence; the timestamps of two adjacent discrete interaction events are extracted, and the time difference is calculated. If the time difference exceeds a preset pause threshold, the time difference is stored as an independent scalar in the pause / thinking duration vector.
[0024] Because the data collected through the three independent threads exhibit heterogeneity in time frequency and data format, this embodiment performs timestamp alignment processing on the multi-dimensional behavioral data to generate a unified time series feature vector. The specific alignment and vectorization processing includes the following steps: Step S101: Slice the single training cycle of the target user into several consecutive time windows using a time sliding window of a preset length. On the global time axis of the training cycle, set the absolute length of the time sliding window to T and the sliding step size to S. Cut along the global time axis according to the sliding step size S, dividing the entire continuous training cycle into a set of overlapping or contiguous time intervals, with each time interval corresponding to an independent time window unit.
[0025] Step S102: Within each time window unit, the time-series discrete operation dimension features are one-hot encoded according to the timestamp of occurrence to obtain an operation feature sub-vector. The system's predefined global operation event dictionary is read. Assuming the global operation event dictionary contains N independent legal operation events, an N-dimensional all-zero row vector is initialized in memory. All discrete interaction event logs within the current time window unit are traversed, event identifiers are read, their index positions in the global operation event dictionary are determined, and the values at the corresponding index positions in the all-zero row vector are overwritten with 1, completing the spatial mapping of discrete events and outputting an N-dimensional operation feature sub-vector.
[0026] Step S103: A polynomial interpolation algorithm is used to smooth and resample the temporally continuous cognitive dimension features within the time window unit, resulting in a cognitive feature sub-vector with uniform frequency. The sampling time points and corresponding coordinate values of all gaze-stationary coordinate flows within the time window unit are extracted, and a piecewise cubic Hermitian polynomial is constructed. A globally uniform target alignment sampling frequency is set, and a new sampling timestamp sequence with equal time intervals is regenerated within the time window unit. The new sampling timestamp sequence is input into the constructed piecewise cubic Hermitian polynomial, and the interpolated resampled coordinate values and resampled heatmap values are calculated and output. The resampled coordinate values and resampled heatmap values under the same new sampling timestamp are combined to form a cognitive feature sub-vector that is strictly equidistant and continuous on the time scale.
[0027] Step S104 involves concatenating the state dimension features, operational feature sub-vectors, and cognitive feature sub-vectors within the same time window unit, and then performing feature dimensionality reduction through a fully connected network layer to generate the time series feature vector corresponding to the time window unit. A vector concatenation operation is performed, stringing the one-dimensional state dimension feature array, the N-dimensional operational feature sub-vectors, and the resampled cognitive feature sub-vectors in the vector space along the column dimension to construct an initial joint feature matrix. A fully connected network layer with multilayer perceptron weight matrices and bias vectors is pre-configured in the system. The initial joint feature matrix is input into the fully connected network layer, performing matrix multiplication and tensor addition operations. A fixed-dimensional multi-dimensional floating-point array is output through a non-linear activation function (e.g., ReLU activation function), and this multi-dimensional floating-point array is recognized as the standard time series feature vector and stored in the feature database as the basic data input source for subsequent clustering and reinforcement learning decision models.
[0028] Furthermore, the specific implementation process of step S200 is as follows: Step S200: Based on the time series feature vector, a composite distance metric algorithm is used to calculate the behavioral similarity distance between the target user and historical user samples; wherein, the composite distance metric algorithm integrates the dynamic time warping distance used to measure operational dimension features, the relative entropy divergence used to measure cognitive dimension features, and the Euclidean distance used to measure state dimension features.
[0029] In practice, the system incorporates a composite distance calculation engine to extract time-series feature vectors from both the target user and historical user samples. After extraction, multi-dimensional distance calculations are performed independently and in parallel. The implementation steps are as follows: First, the dynamic time warping distance of the operational dimension feature sequence is calculated. The target operational feature sub-vector sequence is extracted from the time-series feature vector of the target user, and the historical operational feature sub-vector sequence corresponding to the time period of the historical user samples is extracted simultaneously. A two-dimensional cumulative distance matrix is initialized in memory, where the row dimension corresponds to the length of the target operational feature sub-vector sequence, and the column dimension corresponds to the length of the historical operational feature sub-vector sequence. Euclidean distance is used to calculate the local spatial distance between feature points corresponding to each discrete time point in the two sequences, and the calculated local spatial distance is filled into the corresponding matrix grid. Then, a dynamic programming algorithm is called to search for a warped path to the endpoint grid, starting from the starting grid of the two-dimensional cumulative distance matrix, according to a preset step size constraint rule (moving one grid to the right, down, or down to the right). Along the search path, the local spatial distances in all matrix grids along the way are numerically accumulated. All possible warped paths are compared, and the shortest warped path with the minimum accumulated sum is recorded. The minimum accumulated sum corresponding to the shortest path is directly output as the dynamic time warping distance.
[0030] Secondly, the relative entropy divergence of the probability distribution of the cognitive dimension feature sequence is calculated. The target cognitive feature sub-vector sequence of the target user and the historical cognitive feature sub-vector sequence of historical user samples are obtained. The continuous cognitive feature vectors are discretized and binned. The frequency of feature values within each bin is counted, and the frequency is divided by the absolute number of total sampling points to construct the target probability distribution sequence and the historical reference probability distribution sequence as a benchmark. Following the KL divergence calculation rule, all bins are traversed, and the quotient obtained by dividing the probability value of the target probability distribution sequence within the same bin by the probability value of the historical reference probability distribution sequence is calculated. The natural logarithm of the quotient is then calculated. Further, the obtained natural logarithm is multiplied by the corresponding probability value of the target probability distribution sequence, and the sum of the iterative calculation results for all bins is obtained. The final sum is confirmed as the relative entropy divergence.
[0031] Next, the Euclidean distance of the state dimension feature vectors is calculated. The target state dimension features of the target user and the historical state dimension features of historical user samples are read. The values of the target state dimension features and historical state dimension features on the same feature axis are subtracted to obtain the difference. All differences are squared and summed. Finally, the square root of the sum is taken, and the scalar Euclidean distance is output.
[0032] After obtaining the three distance values mentioned above, to avoid local features being masked due to fixed weight allocation, an adaptive weight allocation logic based on information entropy is introduced. Specifically, adaptive weight coefficients corresponding to operational dimension features, cognitive dimension features, and state dimension features are obtained. The weight allocation logic includes the following specific calculation steps: The system reads the numerical sets of all historical user samples from the historical database across the operational, cognitive, and state dimensions. It then statistically analyzes the distribution of feature values within each dimension and calculates the information entropy for each dimension. The specific calculation logic is as follows: It calculates the probability of different feature values occurring discretely within each dimension, multiplies the probability by the negative of its natural logarithm, and sums the products of all value categories to calculate the initial information entropy for each dimension. After obtaining the initial information entropy for each dimension, based on the inverse mapping rule that smaller information entropy corresponds to greater information content and higher assigned weights, it takes the reciprocal of the initial information entropy and directly determines the initial weight for each dimension. Subsequently, it extracts the initial weights for the operational, cognitive, and state dimensions, and sums these three initial weights to obtain the overall weight sum. Each dimension's initial weight is independently divided by the overall weight sum, and linear normalization is performed. The three floating-point numbers output by the normalization process are confirmed as the adaptive weight coefficients corresponding to the operational, cognitive, and state dimension features, respectively, and the linear normalization process ensures that the sum of the three adaptive weight coefficients is absolutely equal to one.
[0033] Finally, the weighted summation operator is invoked to multiply the adaptive weight coefficients of the corresponding operational dimension features (calculated under normalization) with the dynamic time warp distance, the adaptive weight coefficients of the corresponding cognitive dimension features with the relative entropy divergence, and the adaptive weight coefficients of the corresponding state dimension features with the Euclidean distance. The products obtained from these three independent multiplication operations are then scalar-summed, and the final sum is confirmed as the behavioral similarity distance between the target user and a specific historical user sample. The system iterates through all historical user samples, calculates and outputs a complete spatial distance set, and stores this set in the global distance matrix as the basis for subsequent clustering operations.
[0034] Furthermore, the specific implementation process of step S300 is as follows: Step S300: Cluster users based on the calculated behavioral similarity distance, determine the target cluster to which the target user currently belongs, and extract the centroid features of the target cluster.
[0035] In practice, the system's internal clustering analysis engine constructs a group topology structure based on discrete individual features. It reads the behavioral similarity distance values between all users cached in global memory to construct a two-dimensional global distance matrix. In this matrix, row and column indices correspond to different training user identifiers, and the intersecting grids within the matrix store the corresponding behavioral similarity distance scalars. A fixed constant is set as the truncation distance parameter. For each training user entity in the global distance matrix, taking the currently traversed training user entity as the center, the absolute number of other training user entities whose distance values in the global distance matrix are strictly less than the truncation distance parameter is counted. This absolute number is directly assigned to the currently traversed training user entity as a local density value.
[0036] After calculating the local density value, for each training user entity, a set of candidate users with a local density value significantly greater than the current training user entity is searched in the global space. Within this candidate user set, the spatial distance value of the nearest associated entity to the current training user entity is extracted, and this extracted spatial distance value is defined as the minimum relative distance. When a training user entity is determined to have the highest global local density value, the maximum distance extremum in the global distance matrix is extracted and directly assigned to the training user entity with the highest global local density value, serving as the corresponding minimum relative distance.
[0037] The underlying multiplier logic is invoked to perform a floating-point multiplication operation on the local density value of each training user entity and its corresponding minimum relative distance, outputting the product parameter for the corresponding entity. A two-dimensional cluster center decision graph space is constructed in memory with the local density value as the horizontal axis and the minimum relative distance as the vertical axis. Within the cluster center decision graph space, the entities are sorted in descending order according to the magnitude of the product parameter, and the product difference between adjacent sorted entities is extracted and defined as the numerical descent gradient. The numerical descent gradient is compared with the system's preset curvature inflection point threshold, and several training user entities that are ahead of the curvature inflection point threshold are extracted. These extracted training user entities are directly marked as the cluster centers of their respective independent candidate clusters.
[0038] After confirming the cluster centers of all candidate clusters, a recursive allocation algorithm is executed. All user entities other than the cluster centers are retrieved and assigned to their respective clusters in descending order of local density values. The candidate cluster to which the target individual belongs is determined, and the target individual is directly merged into the candidate cluster with the same state. Following this recursive allocation logic, the affiliation of the target user is determined, and the candidate cluster to which the target user ultimately falls is confirmed as the target cluster.
[0039] After obtaining the target cluster, the centroid feature extraction logic is triggered. A list of identifiers for all sample users within the target cluster is defined in memory, and the time-series feature vectors associated with each sample user are retrieved based on the list index. An exponential decay function is called, substituting the spatial distance value of each sample user within the cluster to its respective cluster center into the exponential decay function to calculate and output the corresponding distance decay factor, which is then used as the fusion weight. Matrix multiplication is performed on all time-series feature vectors within the target cluster, multiplying each feature vector by its corresponding fusion weight. Subsequently, matrix addition is performed on all weighted feature vectors along their vector dimensions, and the resulting matrix is divided by the sum of the fusion weights to complete the weighted average aggregation process. The resulting multidimensional feature tensor is then labeled as the centroid feature of the target cluster.
[0040] In addition, a guardian monitoring thread runs resident in the background of the training system. This thread scans the underlying database at fixed intervals, calculating the absolute value of newly aggregated user samples within the target cluster, and simultaneously reads the updated behavioral similarity distance data and calculates the variance change rate. When the absolute value of newly aggregated user samples exceeds a preset warning threshold, or the variance change rate exceeds the dynamic update threshold, the guardian monitoring thread sends an interrupt update signal to the system master process. Upon receiving the interrupt update signal, the system master process clears the existing cluster center markers in memory, resets the global distance matrix, and re-initiates the preceding local density value calculation and cluster center selection operations. This generates the updated centroid features of the target cluster and overwrites them in the feature database.
[0041] Furthermore, the specific implementation process of step S400 is as follows: Step S400: Construct a reinforcement learning path decision model based on a pre-built vocational skills knowledge graph and a deep Q-network. In this model, nodes in the vocational skills knowledge graph represent training tasks, and edges represent the execution dependencies between tasks.
[0042] In practice, the system initializes the network topology of the vocational skills knowledge graph at the underlying graph database. First, multiple independent node entities are instantiated in the graph database, each mapping to a specific training task. A fixed-dimensional task attribute vector is configured for each node entity. Quantitative index parameters stored in the lesson plan database are read, the task difficulty coefficient is converted into a first floating-point number, the standard operation time is converted into a second floating-point number, and the skill category label is mapped into a discrete one-hot encoded sequence using one-hot encoding logic. The tensor concatenation operator is called to concatenate the first floating-point number, the second floating-point number, and the one-hot encoded sequence bit-by-bit along the row vector dimension to generate the task attribute vector, which is then fixed into the memory address space of the corresponding node entity.
[0043] Next, directed connection edges are constructed between connected node entities. The system's preset course outline topology is read, and a prerequisite dependency weight matrix is assigned to node entities with sequential execution logic. In the prerequisite dependency weight matrix, hard prerequisite condition parameters of Boolean type and soft recommendation relevance parameters with values between zero and one are set. An action mask matrix is constructed using the prerequisite dependency weight matrix. The specific execution logic is as follows: traverse all node entities in the graph database, extract the connection edges with a negative hard prerequisite condition parameter, and fill the corresponding intersection coordinates with zero; extract the connection edges with a positive hard prerequisite condition parameter and fill the corresponding intersection coordinates with one, thereby generating an action mask matrix with path passage interception and filtering functions.
[0044] Subsequently, the system initializes the underlying tensor flow graph of the deep Q-network and strictly defines the Markov decision environment. First, the state space is set. The system retrieves the centroid feature tensor of the target cluster generated in step S300, reads the task attribute vector stored in the memory address of the target user's current knowledge node, and simultaneously queries the underlying performance record table to extract the target user's historical completed node sequence and its corresponding pass rate percentage, arranging it into a continuous pass rate sequence of historical completed nodes. In the tensor concatenation layer, a one-dimensional tensor concatenation and fusion operation is performed on the centroid feature tensor of the target cluster, the task attribute vector, and the pass rate sequence of historical completed nodes, uniformly outputting a complete state space feature input matrix.
[0045] Next, an action space is defined. Based on the global index coordinates of the target user's current knowledge node in the professional skills knowledge graph, the entire set of next-level related task nodes is retrieved. A matrix dot product is performed between the discrete action identifier sequence of all next-level related task nodes and the corresponding row vector in the action mask matrix. Illegal node entities with a dot product result of zero are filtered out and discarded, while legal node entities with a dot product result of one are extracted and retained. A set of legal next-level candidate task nodes is constructed from all retained legal node entities, and this set is mapped to the discrete action space output register range of the deep Q-network.
[0046] Next, a reward function composed of combined evaluation functions is defined. Real-time operation feedback data of the target user after executing any legal next-level candidate task node is read. A subtraction operator is invoked to subtract the current operation error rate from the target user's previous historical average operation error rate, calculating the operation error rate reduction ratio parameter. The current actual completion time and the standard operation time are extracted, and the absolute value of the difference is calculated. This absolute value is then substituted into a Gaussian decay function to calculate the standard operation time proximity parameter. The step cancellation / redo frequency value collected in step S100 is retrieved, and the frequency difference between the previous node and the current node is calculated to generate a decay value parameter for the step cancellation frequency. A weighted accumulator is invoked to multiply the operation error rate reduction ratio parameter, the standard operation time proximity parameter, and the decay value parameter for the step cancellation frequency by preset independent constant weights at the system's underlying level. The single-precision scalar result of the summation is confirmed as the true reward value in the reinforcement learning process.
[0047] Finally, based on the constructed state-space feature input matrix, discrete action space, and true reward value, the neural network parameters in the deep Q-network are iteratively optimized with the training objective of maximizing the cumulative expected discounted value of the reward function. The state-space feature input matrix is forward-propagated through multiple fully connected hidden layers within the deep Q-network. After activation function mapping, the predicted action value Q-value corresponding to each discrete action identifier is calculated. The target Q-value is logically derived based on the Bellman equation, and the mean squared error loss function between the predicted action value Q-value and the target Q-value is calculated. The gradient descent optimization operator is invoked to backpropagate the mean squared error loss function value to all fully connected hidden layers. The weight tensors and bias parameters within each neuron node are updated using the chain rule of differentiation until the mean squared error loss function value converges to a pre-set lower bound threshold, completing the construction of the reinforcement learning path decision model and the permanent fixation of the model parameters.
[0048] Furthermore, the specific implementation process of step S500 is as follows: Step S500: Input the centroid features of the target cluster and the current knowledge node state of the target user as the environmental state space into the reinforcement learning path decision model, and output the next optimal training task node through the reinforcement learning path decision model to dynamically generate the training path of the target user.
[0049] In practice, the system's main control program extracts the environmental state space feature input tensor constructed in the preceding steps and feeds it into the forward propagation network of the reinforcement learning path decision model. The fully connected hidden layer nodes and nonlinear activation functions within the reinforcement learning path decision model perform matrix multiplication and spatial mapping calculations on the environmental state space feature input tensor, generating an array of action value functions (Q-values) corresponding to all legal next-level candidate task nodes in the action space at the output layer.
[0050] The system memory loads an exploration and exploitation strategy algorithm module based on a confidence upper limit. It reads the predicted action values from the Q-value array of the action value function and simultaneously queries the underlying runtime logs to extract the historical absolute frequency of exploration for each legitimate next-level candidate task node. The predicted action values and historical absolute frequencies of exploration are then substituted into the confidence upper limit mathematical model to calculate the comprehensive target Q-value set. A descending sorting operator is called to perform a permutation operation on the comprehensive target Q-value set, extracting the node entity mapped to the first value in the permutation. This first node entity is then identified as the next optimal training task node.
[0051] After obtaining the next optimal training task node, the underlying adjacency matrix of the professional skills knowledge graph is read. The directed edge attribute parameters between the target user's current knowledge node and the next optimal training task node are extracted from the adjacency matrix, and their absolute topological relationship is parsed. The system memory has a pre-built intervention strategy mapping dictionary. The main control program matches the corresponding code logic based on the parsed absolute topological relationship: when the topological relationship is a directed edge directly pointing from the current knowledge node to the next optimal training task node, it is determined to be a successor node relationship, and the main control program calls the preceding new knowledge advancement strategy code module; when the topological relationship is that both nodes share a common parent preceding node and have no direct preceding constraints, it is determined to be a sibling node relationship, and the main control program calls the specialized reinforcement retraining strategy code module; when the topological relationship is a directed edge pointing from the next optimal training task node to the current knowledge node, it is determined to be a preceding node relationship, and the main control program calls the basic downgrade review strategy code module.
[0052] Based on the invoked code modules, the system extracts the corresponding 3D model asset files and interaction logic scripts from the asset database, packages them into a training task scene instruction data package, and calls the underlying network transmission protocol to send the training task scene instruction data package to the target user's training terminal. The training terminal's rendering pipeline receives the training task scene instruction data package, clears the current video memory, and re-executes the 3D space rendering operation to achieve seamless presentation of the new task scene.
[0053] While the target user is executing the next optimal training task node, the system background synchronously wakes up the closed-loop feedback optimization daemon thread. It calls a multi-dimensional data perception probe to continuously collect the latest state dimension features, operation dimension features, and cognitive dimension features. After the target user completes the current node task, it calls dimensionality reduction and concatenation operators to generate a new time-series feature vector, and performs a matrix concatenation operation with this vector and the updated knowledge node state to construct the true state transition feature tensor. Simultaneously, it calls the preset reward function calculation logic to calculate and output the actual single-precision floating-point true reward value.
[0054] Initialize the interaction experience tuple structure in memory space, concatenating the previously used environment state space feature input tensor, the decision action identifier corresponding to the target Q value, the single-precision floating-point true reward value, and the true state transition feature tensor in a fixed memory alignment order. Push the concatenated interaction experience tuple into the system's priority experience replay pool queue. The priority experience replay pool reassigns priority weight values to all tuples in the queue based on the absolute value of the temporal difference error between the true reward value and the expected Q value.
[0055] After reaching the preset training iteration cycle, a neural network parameter update instruction is triggered. The random number generator, based on the probability distribution characteristics of the priority weight values, batch-extracts a fixed number of interactive experience tuples from the priority experience replay pool queue to form an experience batch tensor. The global mean square temporal difference error value corresponding to the experience batch tensor is calculated, and the backpropagation algorithm is invoked. The adaptive moment estimation gradient descent optimizer is used to update the internal weight matrix and bias parameters of the reinforcement learning path decision model's neural network.
[0056] After completing the parameter update calculation, the closed-loop feedback optimization daemon thread sends a synchronization trigger signal to the clustering analysis engine. Upon receiving the synchronization trigger signal, the clustering analysis engine extracts new behavioral data indicators from the real state transition feature tensor, re-executes the distance calculation and topology center calculation steps, and verifies the target user's coordinate offset in the multi-dimensional feature space. If the coordinate offset causes a change in the node with the highest local density belonging to the target user, a cluster transition event is determined to have occurred. The centroid features of the new cluster are extracted to replace the original centroid feature input source, and the process proceeds to the next Markov decision cycle.
[0057] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for dynamically generating vocational skills training paths based on multi-dimensional behavioral clustering, characterized in that, Includes the following steps: S100: Collect multi-dimensional behavioral data of the target user during the training process. The multi-dimensional behavioral data includes at least state dimension features, time-series discrete operation dimension features, and time-series continuous cognitive dimension features. The multi-dimensional behavioral data is then processed with timestamp alignment to generate a time series feature vector. S200, based on the time series feature vector, a composite distance metric algorithm is used to calculate the behavioral similarity distance between the target user and historical user samples; wherein, the composite distance metric algorithm integrates the dynamic time warping distance used to measure the operational dimension features, the relative entropy divergence used to measure the cognitive dimension features, and the Euclidean distance used to measure the state dimension features. S300, perform clustering processing on users based on the calculated behavioral similarity distance, determine the target cluster to which the target user currently belongs, and extract the centroid features of the target cluster; S400, a reinforcement learning path decision model is constructed based on a pre-built vocational skills knowledge graph and a deep Q-network, wherein the nodes of the vocational skills knowledge graph represent training tasks and the edges represent the execution dependencies between tasks; S500, the centroid features of the target cluster and the current knowledge node state of the target user are input into the reinforcement learning path decision model as the environment state space. The reinforcement learning path decision model outputs the next optimal training task node to dynamically generate a training path for the target user.
2. The method for dynamically generating vocational skills training paths based on multi-dimensional behavior clustering according to claim 1, characterized in that, The specific features in the multi-dimensional behavioral data include: The state dimension features include the static basic profile data of the target user and the skill node mastery index calculated based on historical training records. The time-series discrete operation dimension features include the discrete interaction event sequence triggered within the training operation interface, operation error type classification identifiers, and step undo / redo frequency. The temporally continuous cognitive dimension features include the gaze-staying coordinate stream collected by the sensors of the training terminal, the attention heat value generated based on the cursor trajectory, and the pause and thinking time in key operation stages.
3. The method for dynamically generating vocational skills training paths based on multi-dimensional behavior clustering according to claim 2, characterized in that, The process of aligning the multi-dimensional behavioral data with timestamps to generate a time-series feature vector specifically includes the following sub-steps: S101, use a time sliding window of preset length to slice the single training cycle of the target user to obtain several consecutive time windows; S102, within each time window, the time-series discrete operation dimension features are one-hot encoded according to the timestamp of occurrence to obtain operation feature sub-vectors; S103, a polynomial interpolation algorithm is used to smooth and resample the temporally continuous cognitive dimension features within the time window to unify the data frequency and obtain cognitive feature sub-vectors. S104, the state dimension features, the operation feature sub-vectors and the cognitive feature sub-vectors within the same time window are concatenated, and feature dimensionality reduction is performed through a fully connected network layer to generate the time series feature vector corresponding to the time window.
4. The method for dynamically generating vocational skills training paths based on multi-dimensional behavior clustering according to claim 3, characterized in that, The step of calculating the behavioral similarity distance between the target user and historical user samples using a composite distance metric algorithm based on the time series feature vector specifically includes: Extract the time-series feature vectors of the target user and the historical user samples respectively; For the extracted time series feature vector, the dynamic time warping algorithm is used to calculate the dynamic time warping distance of the operation dimension feature sequence between the target user and the historical user sample; The relative entropy divergence of the probability distribution of the cognitive dimension feature sequence between the target user and the historical user samples is calculated using the KL divergence algorithm. The Euclidean distance between the target user and the historical user sample is calculated using the Euclidean distance formula. Obtain the adaptive weight coefficients corresponding to the operation dimension feature, the cognition dimension feature and the state dimension feature respectively, and sum each of the adaptive weight coefficients with the corresponding dynamic time warp distance, the relative entropy divergence and the Euclidean distance to output the final behavior similarity distance.
5. The method for dynamically generating vocational skills training paths based on multi-dimensional behavior clustering according to claim 4, characterized in that, The specific weight allocation logic for obtaining the adaptive weight coefficients corresponding to the operational dimension feature, the cognitive dimension feature, and the state dimension feature is as follows: The numerical distribution characteristics of historical user samples in the operation dimension, the cognition dimension, and the state dimension are statistically analyzed, and the information entropy corresponding to each dimension is calculated respectively. The dispersion and information content of each feature dimension are evaluated based on the information entropy. Based on the inverse mapping rule that the smaller the information entropy, the greater the information content and the higher the assigned weight, the initial weight of each dimension is determined respectively. The initial weights of each dimension are normalized so that the sum of the adaptive weight coefficients of the normalized operational dimension features, cognitive dimension features, and state dimension features is one.
6. The method for dynamically generating vocational skills training paths based on multi-dimensional behavior clustering according to claim 5, characterized in that, The step of clustering users based on the calculated behavioral similarity distance to determine the target cluster to which the target user currently belongs specifically includes: Construct a global distance matrix containing the target user and each historical user sample, where the matrix elements are the corresponding behavioral similarity distances; Based on the global distance matrix and the preset cutoff distance, the local density value of each user is calculated. The local density value represents the number of other users adjacent to the user within the cutoff distance range. Calculate the minimum relative distance between each user and other neighboring users with higher local density values; A cluster center decision graph is constructed based on the product of the local density value and the minimum relative distance, and several cluster centers are selected from the cluster center decision graph according to the numerical descent gradient to form multiple candidate clusters. The target user is recursively assigned to a candidate cluster containing the nearest neighboring user with a higher local density value, and this candidate cluster is used as the target cluster to which the target user currently belongs.
7. The method for dynamically generating vocational skills training paths based on multi-dimensional behavioral clustering according to claim 6, characterized in that, The extraction of the centroid features of the target cluster specifically includes: Obtain the time-series feature vectors of all sample users within the target cluster; Calculate the distance decay factor from the cluster center to each sample user within the cluster, and use it as the fusion weight; Based on the fusion weight, all the time series feature vectors within the target cluster are weighted and averaged to generate the centroid feature that can characterize the common micro-operational bottlenecks and cognitive patterns of the cluster. During the training process, background monitoring is initiated. If the variance change rate of the number of newly gathered user samples or the behavioral similarity distance within the target cluster exceeds the preset dynamic update threshold, the cluster center is recalculated and the centroid feature is updated.
8. The method for dynamically generating vocational skills training paths based on multi-dimensional behavioral clustering according to claim 7, characterized in that, The reinforcement learning path decision model built based on a pre-constructed vocational skills knowledge graph and a deep Q-network specifically includes the definition of the attributes of the vocational skills knowledge graph: Each node in the professional skills knowledge graph is configured with a task attribute vector, which includes the task difficulty coefficient, standard operation time, and skill category label. Each edge in the vocational skills knowledge graph is configured with a pre-dependency weight matrix, which is used to characterize the correlation between hard preconditions and soft recommendations when moving from the current training task node to the next training task node. An action mask matrix is constructed based on the aforementioned pre-dependency weight matrix to filter out illegal paths that violate the execution logic order when the deep Q network makes output decisions; The reinforcement learning path decision model constructed based on the pre-built vocational skills knowledge graph and deep Q-network also includes setting the Markov decision environment of the deep Q-network, with the specific rules as follows: Define a state space, which includes the centroid features of the target cluster, the task attribute vector corresponding to the knowledge node where the target user is currently located, and the pass rate sequence of the target user's historical completed nodes; An action space is defined, which is a set of legal next-level candidate task nodes that can be reached in the professional skills knowledge graph after being filtered by the action mask matrix based on the knowledge node where the target user is currently located. A reward function is defined, which is a combined evaluation function. Its input variables include the percentage decrease in the target user's operation error rate after executing the next level task node, the closeness of the standard operation time, and the decay value of the step cancellation frequency. The neural network parameters in the deep Q-network are iteratively optimized with the training objective of maximizing the cumulative expected discounted value of the reward function.
9. The method for dynamically generating vocational skills training paths based on multi-dimensional behavior clustering according to claim 8, characterized in that, The step of outputting the next optimal training task node through the reinforcement learning path decision model to dynamically generate a training path for the target user specifically includes: The reinforcement learning path decision model calculates the action value function Q value corresponding to each legal next-level candidate task node in the action space based on the current input environmental state space through its internal neural network layer. An exploration and utilization strategy based on confidence upper limit is adopted to select the node with the highest target Q value from the candidate task node set as the next optimal training task node; The topological relationship between the next optimal training task node and the target user's current knowledge node is analyzed and mapped to the corresponding dynamic intervention strategy. The dynamic intervention strategies include: a prior knowledge advancement strategy executed when the topological relationship is a successor node, a specialized reinforcement and retraining strategy executed when the topological relationship is a sibling node, and a basic downgrade review strategy executed when the topological relationship is a predecessor node. According to the determined dynamic intervention strategy, the corresponding training task scenario instructions are rendered in real time and seamlessly issued to the training terminal of the target user. After dynamically generating the training path for the target user, the process also includes a real-time closed-loop feedback optimization step for the reinforcement learning path decision model: Real-time monitoring of the latest multi-dimensional behavioral data of the target user during the execution of the next optimal training task node; Based on the latest multi-dimensional behavioral data, calculate the target user's real state transition characteristics and the actual reward value obtained after performing the task; The interactive experience tuple containing the environmental state space, the decision action corresponding to the target Q value, the real reward value, and the real state transition features is stored in the priority experience replay pool. Experience batches are extracted from the priority experience replay pool, and the neural network parameters of the reinforcement learning path decision model are updated using backpropagation of temporal difference error. Simultaneously, the steps of claims 1 to 3 are triggered to re-evaluate whether the target user has undergone cluster transition.
10. A dynamic generation system for vocational skills training paths based on multi-dimensional behavior clustering, applied to the dynamic generation method for vocational skills training paths based on multi-dimensional behavior clustering as described in any one of claims 1-9, characterized in that, include: The multi-dimensional data perception and alignment module is used to collect multi-dimensional behavioral data of target users during the training process. The multi-dimensional behavioral data includes at least state dimension features, time-series discrete operation dimension features, and time-series continuous cognitive dimension features. The module performs timestamp alignment processing on the multi-dimensional behavioral data to generate a time series feature vector. The composite distance metric and fusion module is used to calculate the behavioral similarity distance between the target user and historical user samples based on the time series feature vector using a composite distance metric algorithm; wherein, the composite distance metric algorithm fuses the dynamic time warping distance used to measure the operational dimension features, the relative entropy divergence used to measure the cognitive dimension features, and the Euclidean distance used to measure the state dimension features. The target group clustering and feature extraction module is used to cluster users based on the calculated behavioral similarity distance, determine the target cluster to which the target user currently belongs, and extract the centroid features of the target cluster. A reinforcement learning path decision building module is used to build a reinforcement learning path decision model based on a pre-built vocational skills knowledge graph and a deep Q-network, wherein the nodes of the vocational skills knowledge graph represent training tasks and the edges represent the execution dependencies between tasks. The dynamic path mapping and closed-loop output module is used to input the centroid features of the target cluster and the current knowledge node state of the target user as the environmental state space into the reinforcement learning path decision model, and output the next optimal training task node through the reinforcement learning path decision model to dynamically generate a training path for the target user.
