Intelligent training scene ai adaptive guidance method and system

Abstract

The application relates to the technical field of guided content generation methods, and discloses an intelligent practical training scene AI adaptive guiding method and system, wherein the intelligent practical training scene AI adaptive guiding method comprises the following steps: mapping feature vectors output by specific encoders of various modes to a unified semantic space, training through a contrast learning loss function, and making distances of embedding vectors of the same operation intention under different modes tend to be close; fitting forgetting curve parameters of various modes based on user historical operation data, calculating a difference ratio of forgetting rates between different modes, and quantifying memory characteristic differences between the modes; converting a forgetting difference coefficient into content generation parameters such as a detail weight, an emphasis coefficient and a representation style identifier, and performing targeted modulation on guiding content; and the application makes the guiding content accurately matched through cross-modal unified representation in a semantic correlation dimension and targeted modulated through a forgetting difference coefficient in a mode adaptability dimension.

Description

Technical Field

[0001] This invention relates to the field of guided content generation methods, and more specifically, to an AI adaptive guidance method and system for intelligent training scenarios. Background Technology

[0002] In application systems that support multiple input methods, users can perform operations through various input modalities such as touch screen, mouse, and voice. When users require guidance content from the system during operation, the system needs to retrieve content related to the semantics of the current operation from the guidance knowledge base and generate guidance text adapted to the user's cognitive characteristics.

[0003] Existing methods for generating introductory content have two technical shortcomings: Firstly, during the content retrieval stage, different input modalities have significantly different raw feature spaces. Touchscreen operations feature touch point trajectories, pressure sequences, and gesture patterns; mouse operations feature cursor position sequences, click timing, and scrolling amounts; and voice operations feature audio spectra and speech recognition results. When the same operational intent is executed through different modalities, the raw feature representations differ greatly, leading to inconsistent results when semantically matching with the guidance knowledge base. For example, when a user performs the same movement operation using a touchscreen swipe versus a mouse drag, the differences in raw features may result in different matching guidance entries, causing a disjointed guidance experience.

[0004] Secondly, in the content generation stage, existing methods use a uniform content template and a fixed level of detail, failing to consider that operational knowledge acquired through different input modalities involves different memory systems. Touchscreen and mouse operations primarily rely on motor memory for encoding, while voice commands mainly rely on declarative memory. Motor memory and declarative memory have different forgetting patterns: motor memory decays more slowly but requires contextual matching for retrieval, while declarative memory decays more quickly but is easily retrieved across contexts. Adopting a uniform generation strategy leads to a mismatch between the level of detail in the guidance content and the user's actual memory state of that modality skill. Summary of the Invention

[0005] This invention provides an AI adaptive guidance method and system for intelligent training scenarios, which solves the technical problem that existing guidance content generation methods in related technologies have defects in both semantic relevance and modal adaptability.

[0006] This invention provides an AI adaptive guidance method for intelligent training scenarios, comprising the following steps: Obtain the user's historical operation record data, parse the input modality identifier and operation attributes of each record, and generate a modally annotated operation history dataset; For each operation record in the modality-annotated operation history dataset, feature extraction and semantic mapping are performed using a modality-specific encoder and a cross-modality aligned neural network to generate a modality-independent operation semantic embedding sequence; The modality-annotated operation history dataset is grouped according to the input modality. The mastery-forgetting time series of knowledge of each operation type within each modality group is statistically analyzed to generate memory decay feature data of modality grouping. The forgetting curve fitting algorithm is applied to the memory decay characteristic data of each modality group to extract the forgetting parameters unique to each modality, calculate the ratio of the forgetting rate parameters of the same operation type between different modalities, and generate a modality-specific forgetting difference coefficient matrix. Obtain the raw data of the user's current operation and its input modality identifier, apply the corresponding modality-specific encoder and cross-modality aligned neural network to generate a unified semantic embedding vector of the current operation, calculate the relevance score with the semantic embedding vector of each guidance entry in the guidance knowledge base, and generate a guidance content relevance ranking list; Based on the current input modality identifier, query the modality-specific forgetting difference coefficient matrix, calculate the detail weight and emphasis coefficient based on the forgetting difference coefficient, and generate modality-sensitive content generation configuration; Candidate guidance items are selected from the guidance content relevance ranking list. Based on the modality-sensitive content generation configuration, the candidate items are processed for content modulation to generate and output modality-adapted guidance content text.

[0007] Furthermore, the cross-modal alignment neural network includes the following components: Modal projection layer is used to project feature vectors of different dimensions output by each modality-specific encoder onto an intermediate representation space of the same dimension; The semantic alignment layer is used to map the feature vectors of the intermediate representation space to the unified semantic space through nonlinear transformation, so that the representations of the same operational intent in different modalities are close in distance in the unified semantic space. The normalized output layer is used to perform L2 normalization on the aligned semantic vectors. The modality-specific encoder and the cross-modality aligned neural network adopt an end-to-end joint training mode. During training, samples with the same operational intent executed in different modalities are constructed as positive sample pairs, and samples with different operational intents are constructed as negative samples. The loss function adopts the contrastive learning loss, which is the negative logarithm of the ratio of the exponent of the dot product of the anchor sample and the positive sample embedding vector to the sum of the exponents of the dot products of the anchor sample and all samples in the batch. Each dot product value is scaled by dividing by the temperature parameter.

[0008] Furthermore, the forgetting curve fitting algorithm adopts an exponential decay model, which means that the memory retention rate is equal to the initial memory strength parameter multiplied by the negative power of the natural constant, where the negative power is the negative product of the decay rate parameter and the time interval. The historical data was fitted using the least squares method. The optimization objective was to minimize the sum of squared residuals between the actual memory retention rate and the model prediction value for each time interval in the historical data. The decay rate parameter and initial memory strength parameter specific to each mode were then extracted. The actual memory retention rate is calculated as follows: for each time interval in the mastery-forgetting time series, count the operation records that were executed again after the last correct execution within that time interval, and calculate the ratio of the number of correctly executed records to the total number of records.

[0009] Furthermore, the modality-specific forgetting difference coefficient is calculated as follows: For the same operation type, the ratio of the decay rate parameter of the first mode to the decay rate parameter of the second mode is calculated as the forgetting difference coefficient between the first and second modes for that operation type. The forgetting difference coefficient is a positive real number, where a value of 1 indicates that the forgetting rates of the two modes are the same, a value greater than 1 indicates that the forgetting rate of the first mode is faster than that of the second mode, and a value less than 1 indicates that the forgetting rate of the first mode is slower than that of the second mode.

[0010] Furthermore, the modality-sensitive content generation configuration includes the following parameters: The detail weight is calculated by multiplying the forgetting difference coefficient of the current mode relative to the reference mode by a time modulation factor, where the time modulation factor is 1 plus the ratio of the most recent execution time interval of the user's operation on the current mode to a reference time interval threshold. The key point is the coefficient, which is used to determine the proportion of each knowledge point in the guidance content. It is calculated by dividing the decay rate parameter of the operation type corresponding to the knowledge point in the current modality by the sum of the decay rate parameters of the operation types corresponding to all knowledge points included in the current guidance entry. The expression style is identified by the action description template for the motor memory-dominated modality and the concept explanation template for the descriptive memory-dominated modality.

[0011] Furthermore, the calculation of the relevance score with the semantic embedding vector of each guidance entry in the guidance knowledge base includes: Calculate the cosine distance between the unified semantic embedding vector of the current operation and the semantic embedding vectors of each guide entry. This cosine distance is 1 minus the ratio of the dot product of the two vectors to the product of the magnitudes of the two vectors. The cosine distance is converted into a correlation score by applying the distance decay function. This correlation score is a negative power of the natural constant, which is the negative product of the distance decay coefficient and the cosine distance. The guidance items are sorted from highest to lowest relevance score to generate a sorted list of guidance content relevance.

[0012] Furthermore, the content modulation processing of the candidate entries includes: The original content of candidate entries is expanded or compressed according to the detail weight. When the detail weight is greater than 1, the extended description of the candidate guidance entry is retrieved from the guidance knowledge base to supplement it. When the detail weight is less than 1, the content is semantically compressed and the core points are retained. The length of each knowledge point is adjusted according to the emphasis coefficient, and more space is allocated to knowledge points with a faster rate of forgetting for detailed explanation; Select the corresponding expression template according to the expression style identifier. The action description template focuses on the specific action description of the operation steps, while the concept explanation template focuses on the explanation of the operation principle and purpose.

[0013] Furthermore, the modality-specific encoder adopts a corresponding network structure based on the input modality type: For touch screen mode and mouse mode, a bidirectional long short-term memory network is used as the encoder. The input layer receives the operation feature vector sequence of time step serialization, the hidden layer extracts the temporal dependency features through two LSTM units, forward and backward, and the output layer concatenates the forward hidden state and the backward hidden state of the last time step as the mode-specific feature vector. For speech modalities, a one-dimensional convolutional neural network is used as the encoder. The input layer receives the audio spectral feature sequence, the convolutional layer extracts local spectral features through multi-scale convolutional kernels, and the output layer generates modality-specific feature vectors through global average pooling.

[0014] Furthermore, it also includes the following steps: Monitor user response behavior after receiving guidance content, and statistically analyze response latency and operational correctness; The operation response data is fed back to the memory decay characteristic data of the modality group, which is used to update the forgetting curve parameters online and realize the adaptive adjustment of the modality-specific forgetting difference coefficient matrix.

[0015] This invention provides an AI adaptive guidance system for intelligent training scenarios, used to execute the aforementioned AI adaptive guidance method for intelligent training scenarios, comprising: The data acquisition module is used to acquire the user's historical operation records and the raw data of the current operation; The modality coding module includes modality-specific encoders corresponding to each input modality, which are used to convert the raw operational data of each modality into modality-specific feature vectors; The cross-modal alignment module, including the cross-modal alignment neural network, is used to map modality-specific feature vectors to a unified semantic space to generate modality-independent semantic embedding vectors; The forgetting modeling module is used to statistically analyze the mastery-forgetting time series by modality and to generate a modality-specific forgetting difference coefficient matrix by applying a forgetting curve fitting algorithm. The semantic matching module is used to calculate the relevance score between the unified semantic embedding vector of the current operation and each guide entry in the guide knowledge base and generate a sorted list; The content generation module is used to modulate the candidate guidance items according to the modality-sensitive content generation configuration and generate modality-adapted guidance content text; The content presentation interface is used to output modally adapted introductory text.

[0016] The beneficial effects of this invention are as follows: by mapping the operations of different input modalities to a unified semantic space through a cross-modal alignment neural network, the same operation intention can be executed by touch screen, mouse or voice, and a semantically consistent embedding vector representation can be obtained. This overcomes the factor that the difference in the original feature space of different modalities leads to the inconsistency of semantic matching results with the guiding knowledge base, and solves the technical problem of inconsistent semantic matching.

[0017] The differences in memory decay characteristics of each modality are quantified by a modality-specific forgetting difference coefficient matrix. The forgetting difference coefficient is converted into modulation parameters for content generation, so that the generated guidance content matches the memory characteristics of the modality in terms of detail, key distribution and expression. This overcomes the factor that the unified generation strategy ignores the differences between motor memory and declarative memory, and solves the technical problem of insufficient modality adaptability of guidance content.

[0018] The synergistic effect of the two mechanisms mentioned above enables the guidance content to achieve accurate matching in the semantic relevance dimension through cross-modal unified representation, and to achieve targeted modulation in the modality adaptability dimension through forgetting difference coefficient, thereby solving the problem that existing guidance content generation methods have defects in both the semantic relevance and modality adaptability dimensions. Attached Figure Description

[0019] Figure 1 This is a flowchart of the AI ​​adaptive guidance method for intelligent training scenarios of the present invention; Figure 2 This is a line graph comparing the multimodal memory decay curves of the present invention, showing the distribution of different input modalities performing the same operation intent in a unified semantic space after cross-modal alignment neural network mapping; Figure 3This is a heatmap of the modality-specific forgetting difference coefficients of the present invention, showing the modality-specific forgetting difference coefficient matrix. ; Figure 4 This is a timing diagram of the user operation accuracy and guidance effect feedback of the present invention, showing the trend of the operation accuracy of user U001 after receiving modal adaptation guidance within a 30-day experimental period. Detailed Implementation

[0020] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0021] This implementation provides an AI adaptive guidance method for intelligent training scenarios, such as... Figure 1 As shown, the method includes the following steps: Step 100: Obtain the user's historical operation record data, parse the input modality identifier and operation attributes of each record, and generate a modally labeled operation history dataset.

[0022] The system acquires user operation records within a defined historical time window from the data acquisition module. Each operation record is parsed to extract the following fields: input modality identifier (touchscreen, mouse, voice, etc.), operation type label, operation execution timestamp, and operation correctness label. The parsed records are then organized chronologically to generate a modally labeled historical operation dataset. in Indicates the first The input modality identifier of each record. Indicates the operation type label. Represents a timestamp. Labels indicating correct operation This indicates the total number of records.

[0023] Furthermore, operational correctness labels It is a binary variable, taking the value 0 or 1, where This indicates that the operation was executed correctly. This indicates that the operation failed.

[0024] Taking a CNC machine tool operation training system as an example, the system supports three input modes: touch screen, mouse, and voice. Users learn operation skills such as "spindle start," "feed adjustment," and "tool changeover." The system obtains the operation record data of user ID U001 over the past 30 days, parses it, and generates a modally annotated operation history dataset. Some records are shown in the table below: Table 1. Examples of historical datasets for modality annotation operations. Step 200: For each operation record in the modality-annotated operation history dataset, perform feature extraction and semantic mapping through a modality-specific encoder and a cross-modality aligned neural network to generate a modality-independent operation semantic embedding sequence.

[0025] For each operation record in the modal-annotated operation history dataset, the raw operation data is first preprocessed. For the raw data of the touch screen modality, the touch point coordinate sequence is normalized according to the screen resolution, and the pressure value sequence is normalized according to the device's maximum pressure threshold. For the raw data of the mouse modality, the cursor position sequence is normalized according to the display area size. For the raw data of the speech modality, the audio spectral features are Z-score normalized.

[0026] After preprocessing, based on the input modality identifier Select the corresponding mode-specific encoder The modality-specific encoder transforms the preprocessed operational data into modality-specific feature vectors. .

[0027] Input modality-specific feature vectors into the cross-modality alignment neural network Mapping to a modality-independent unified semantic space generates unified semantic embedding vectors: in, Indicates the first The original operation data of the record, This represents the mapped unified semantic embedding vector.

[0028] Perform the above processing on all operation records to generate a modality-independent operation semantic embedding sequence. .

[0029] The aforementioned modality-specific encoder The appropriate network structure is adopted based on the input modality type. For touchscreen and mouse modalities, a bidirectional long short-term memory network is used as the encoder. The input layer receives a sequence of operation feature vectors serialized over time steps. The hidden layer extracts temporal dependency features through two LSTM units (forward and backward). The output layer concatenates the forward and backward hidden states of the last time step as the modality-specific feature vector. For speech modalities, a one-dimensional convolutional neural network is used as the encoder. The input layer receives the audio spectral feature sequence, the convolutional layer extracts local spectral features through multi-scale convolutional kernels, and the output layer generates modality-specific feature vectors through global average pooling. .

[0030] The aforementioned cross-modal alignment neural network It includes the following components: Modal projection layer: projects the feature vectors of different dimensions output by each modality-specific encoder onto an intermediate representation space of the same dimension; Semantic alignment layer: Makes representations of the same operational intent in different modalities close in distance within a unified semantic space; Normalized output layer: The aligned semantic vectors are L2 normalized to facilitate subsequent cosine distance calculation.

[0031] The data flow relationship between the above-mentioned components is as follows: the modality-specific feature vector is first input into the modality projection layer for dimensional unification, the projection result is input into the semantic alignment layer for cross-modality semantic alignment, and the alignment result is input into the normalization output layer to generate the final unified semantic embedding vector.

[0032] Furthermore, the modality projection layer is implemented using a fully connected layer, mapping the feature vectors output by each modality-specific encoder to a unified intermediate representation space through a linear transformation. The semantic alignment layer is implemented using a multi-layer fully connected network, mapping the feature vectors of the intermediate representation space to the unified semantic space through a non-linear transformation. The network parameters are learned through a gradient descent optimization process of the contrastive learning loss function, so that the embedding vector distances of different modal samples with the same operational intent in the unified semantic space are brought closer.

[0033] The aforementioned modality-specific encoder and cross-modality aligned neural network employ an end-to-end joint training mode. During training, samples with the same operational intent performed in different modalities are constructed as positive sample pairs, while samples with different operational intents are constructed as negative samples. The loss function uses a contrastive learning loss: in, Represents an exponential function. The unified semantic embedding vector representing the anchor sample. This represents the positive sample embedding vector that has the same operational intent as the anchor sample but a different input modality. Indicates the index of the sample within the batch. Indicates the first in the batch The embedding vector of each sample, This represents the temperature parameter. The optimization strategy employs the Adam optimization algorithm.

[0034] Furthermore, temperature parameters The range of values ​​is This is used to control the smoothness of the similarity distribution in the contrastive learning loss function. When When the value is smaller, the loss function penalizes hard-to-distinguish samples more severely; when... When a larger value is taken, the penalty of the loss function on each sample tends to be uniform.

[0035] Step 300: Group the modality-annotated operation history dataset according to the input modality, and statistically analyze the mastery-forgetting time series of knowledge of each operation type within each modality group to generate memory decay feature data of the modality group.

[0036] Modality-annotated operation history dataset Grouping by input modality identifier to generate a subset of operation records for each modality. ,in This represents the set of input modal types.

[0037] For each modal subset The operation records within are aggregated by operation type label. For each operation type... Statistical operation types Construct a mastery-forgetting time series by analyzing the distribution of correct and incorrect executions on the timeline: in, This represents the index of a data point in a mastery-forgetting time series. Representing modes Next operation type Mastery-Forgetting Time Series.

[0038] Memory decay characteristics were calculated for mastery-forgetting time series analysis of various operation types, including: the distribution of time intervals between adjacent correct executions, the distribution of time points at which forgetting occurs (from correct to incorrect), and the number of reviews required for memory retrieval (from incorrect to correct). These characteristics were then organized into modal grouped memory decay characteristic data. .

[0039] Step 400: Apply the forgetting curve fitting algorithm to the memory decay characteristic data of each modality group to extract the forgetting parameters unique to each modality and generate a modality-specific forgetting difference coefficient matrix.

[0040] Memory decay characteristics of modal grouping Each modality-operation type combination A forgetting curve fitting algorithm was applied to fit the function of memory retention rate over time for modality-operation type combinations. The forgetting curve adopted an exponential decay model. in, Representing modes Next operation type Memories after a time interval Retention rate after This represents the initial memory strength parameter. This represents the decay rate parameter.

[0041] Furthermore, time interval The range of values ​​is The unit is seconds, representing the length of time elapsed since the last successful execution of this operation type until the current moment. This indicates the immediate state immediately after the operation is completed.

[0042] By fitting historical data using the least squares method, the decay rate parameters specific to each mode are extracted. and initial memory strength parameters .

[0043] Furthermore, the objective function for least squares fitting is: in This represents the total number of time interval samples in the historical data. For the first The time interval value of each time interval sample. For the first The actual memory retention rate corresponding to each time interval sample.

[0044] Furthermore, the decay rate parameter The range of values ​​is This represents the rate at which memory retention decays per unit of time; initial memory strength parameter. The range of values ​​is , represents the memory retention rate immediately after the operation is completed, where a value of 1 indicates a state of complete mastery.

[0045] Calculate the same operation type The modality-specific forgetting difference coefficient is generated by calculating the ratio of parameter differences between different modalities. in, Indicates the operation type In modality With mode The forgetting difference coefficient between them.

[0046] The forgetting difference coefficients of all modality pairs and operation types are organized into a modality-specific forgetting difference coefficient matrix. .

[0047] Furthermore, the forgetting variance coefficient The range of values ​​is ,in This indicates that the forgetting rates of the two modalities are the same. Representing modes The rate of forgetting is faster than that of modality. , Representing modes The rate of forgetting is slower than that of modality. .

[0048] The aforementioned forgetting curve fitting algorithm employs the Ebbinghaus forgetting curve fitting method. The input to the Ebbinghaus forgetting curve fitting method is the mastery-forgetting time series. Each data point contains a timestamp and a corresponding operation correctness label. The output of the Ebbinghaus forgetting curve fitting method is the parameters of an exponential decay model, including the decay rate parameter. and initial memory strength parameters The fitting process employs the least squares method, with the optimization objective being to minimize the sum of squared residuals between the actual memory retention rate and the model prediction value for each time interval in the historical data.

[0049] Furthermore, the actual memory retention rate is calculated as follows: for each time interval in the mastery-forgetting time series... Statistics on the time interval since the last successful execution The operation records that were subsequently executed were then used to calculate the ratio of the number of correctly executed records to the total number of records, which was used as the time interval. The corresponding actual memory retention rate.

[0050] After grouping the historical data of user U001 by modality, the actual memory retention rate distribution under each modality was statistically analyzed for the "spindle start" operation type. The forgetting curve parameters were then fitted using the least squares method. The results are shown in the table below: Table 2. Parameters for fitting the forgetting curve of the "spindle start" operation under each modality. Using the touchscreen modality as the baseline modality, calculate the forgetting difference coefficient between the speech modality and the touchscreen modality: This indicates that the forgetting rate of "spindle activation" operation knowledge learned through voice commands is approximately 2.32 times that of knowledge learned through touch screen operations. The average forgetting difference coefficients for each operation type were used to generate the voice-touch screen elements of a modality-specific forgetting difference coefficient matrix: Step 500: Obtain the raw data of the user's current operation and its input modality identifier, apply the corresponding modality-specific encoder and cross-modality aligned neural network to generate a unified semantic embedding vector for the current operation, calculate the relevance score with the semantic embedding vector of each guidance entry in the guidance knowledge base, and generate a guidance content relevance ranking list.

[0051] Obtain raw data of the user's current operation from the data acquisition module. and its input modality identifier .

[0052] Based on input modality identifier Select the corresponding mode-specific encoder The raw data of the current operation is input into a modality-specific encoder, and then processed through a cross-modal alignment neural network. Mapping to a unified semantic space generates a unified semantic embedding vector for the current operation: Obtain the pre-computed semantic embedding vector set for each guidance entry from the guidance knowledge base. ,in Indicates the index of the guide entry. Indicates the first The semantic embedding vector of each guiding entry, Indicates the total number of guide entries.

[0053] Calculate the cosine distance between the unified semantic embedding vector of the current operation and the semantic embedding vectors of each guiding entry: Furthermore, since both the unified semantic embedding vector and the guiding entry semantic embedding vector have undergone L2 normalization, the range of cosine similarity values ​​is [missing value]. Therefore, cosine distance The range of values ​​is ,in This indicates that the semantics are completely consistent. It indicates that the meanings are completely opposite.

[0054] Applying the distance decay function to the cosine distance converts it to a correlation score: in, This is the distance attenuation coefficient.

[0055] Furthermore, the distance attenuation coefficient The range of values ​​is This is used to control the sensitivity of the correlation score to changes in cosine distance. When When the value is large, the relevance score is more sensitive to changes in distance, with only the closer guiding items receiving higher scores; when When smaller values ​​are taken, the correlation scores are distributed more evenly. Because... and Correlation score The range of values ​​is ,in The case where the corresponding semantics are completely identical.

[0056] The guidance items are sorted from highest to lowest relevance score to generate a ranking list of guidance content relevance. .

[0057] At 10:20:15 on October 25, 2024, user U001 issued the "Start spindle" command via voice. After the system collected the voice audio data, it extracted the spectral feature vector through a voice modal encoder, and then obtained a unified semantic embedding vector through cross-modal alignment neural network mapping. (Dimensionality is 128). Let the distance attenuation coefficient be... The relevance scores of each item in the knowledge base were calculated and sorted, and the results are shown in the table below: Table 3. List of Guided Content Relevance Ranking Step 600: Query the modality-specific forgetting difference coefficient matrix based on the current input modality identifier, obtain the content generation modulation parameter group corresponding to the modality, and generate the modality-sensitive content generation configuration.

[0058] Based on the current input modality identifier Query modality-specific forgetting difference coefficient matrix Obtain the forgetting difference coefficient of this mode relative to the reference mode. .

[0059] Furthermore, the baseline mode is a pre-specified reference input mode used as a benchmark for comparing forgetting rates. Indicates the current mode Relative to the reference mode The forgetting difference coefficient, i.e. The mean value was obtained by taking the forgetting difference coefficient for each operation type.

[0060] Based on the forgetting difference coefficient calculation, a modulation parameter set is generated, including: Detail weight Modes with faster forgetting rates ( A higher value corresponds to a higher detail weight, generating more detailed guidance content to compensate for faster memory decay; Emphasis on coefficients The length allocation ratio of each knowledge point is determined based on the characteristics of modal memory. Expression style identifier Motor memory-dominated modalities (touchscreen, mouse) use action description templates, while declarative memory-dominated modalities (voice) use concept explanation templates.

[0061] Organize the above parameters into a modally sensitive content generation configuration: Furthermore, the detail weight The range of values ​​is ,in This indicates that the original level of detail should be maintained. This indicates an increase in level of detail. This indicates a reduction in detail. The formula for calculating the detail weight is: in This is the forgetting difference coefficient between the current mode and the baseline mode.

[0062] Furthermore, the emphasis is placed on the coefficient. The range of values ​​is This is used to determine the proportion of each knowledge point in the guiding content; the sum of the proportions of all knowledge points within the same guiding item is 1. For the first knowledge point in a guiding item... The key knowledge point is that the formula for calculating the coefficient is: in For the current mode, the first Each knowledge point corresponds to the decay rate parameter of the operation type. Index representing the summation of knowledge points This represents the total number of knowledge points contained in the current guide entry.

[0063] In this embodiment, to adapt to the memory recall characteristics of different modalities, the calculation of the above-mentioned detail weight also considers the most recent execution time interval of the user's operation in that modality. For motor memory-dominated modalities, if the most recent execution time interval is long, contextual description content is added to assist memory recall.

[0064] Furthermore, the formula for calculating the detail weight of the time interval is as follows: in The most recent execution time interval for the user's current modal operation. This is a preset reference time interval threshold, used to control the impact of time intervals within a reasonable range.

[0065] Step 700: Select the top-ranked candidate guidance items from the guidance content relevance ranking list, perform content modulation processing on the candidate items based on the modality-sensitive content generation configuration, and generate modality-adapted guidance content text.

[0066] List sorted by relevance of guiding content Selecting the top-ranked correlation scores Candidate boot entries for each bit, generating a candidate boot entry set. ,in This indicates the candidate guide item that ranks first in relevance score. This indicates the candidate guide item that ranks second in relevance score. Indicates the ranking of relevance scores Candidate guide entries for the position.

[0067] Furthermore, the number of candidate entries It is a positive integer, and its value range is 1. ,in This refers to the total number of guiding entries in the guiding knowledge base.

[0068] Modality-sensitive content generation configuration Content modulation processing is performed on each candidate guide entry: Step 701: Weight by Detail Expand or compress the original content of the candidate entries. When When necessary, supplement the candidate guidance entries by retrieving extended descriptions from the guidance knowledge base; when At that time, semantic compression is performed on the content to retain the core points.

[0069] Step 702: Emphasize by key factor Adjust the length allocation for each knowledge point. Allocate more space to explain knowledge points that are forgotten more quickly.

[0070] Step 703: Identify by expression style Select the corresponding expression template. Action description templates focus on the specific action descriptions of the operation steps; concept explanation templates focus on the explanation of the operation principles and purposes.

[0071] The modulated content is then integrated to generate modally adapted guiding text. .

[0072] The current user U001 uses voice modal input. Querying the modality-specific forgetting difference coefficient matrix yields: The user last performed a spindle-related operation via voice modality on October 15, 2024. The time elapsed since then is... Day, set reference time interval threshold Day, calculate the detail weight: Select candidate item G-003, ranked first from the relevance sorting list, and generate the content as shown in the table below: Table 4 Modality-Sensitive Content Generation Configuration Based on the above configuration, content modulation is performed on boot entry G-003: due to Supplement the content by retrieving extended instructions for G-003 from the guidance knowledge base; allocate the length of each knowledge point according to its emphasis coefficient, with the "operation steps" knowledge point being given more emphasis. The highest level of detail is achieved by using a conceptual explanation template, focusing on explaining the working principle and operational purpose of the spindle startup. The generated modal adaptation guidance text... It includes a complete conceptual explanation of spindle startup, detailed operating procedures, and safety precautions.

[0073] Step 800: Output the modally adapted guide content text and its generation parameters to the content presentation interface, and output the guide content that is cross-modal semantically matched and modally adapted.

[0074] Modality-adapted guidance text Output to the content presentation interface. Simultaneously, record the parameter configuration generated during this guidance session, including: the unified semantic embedding vector of the current operation, the matched guidance entry identifier, and the applied content generation modulation parameter group, for subsequent guidance effect analysis and parameter optimization.

[0075] In this embodiment of the application, in addition to the above steps, the following steps are also included: Step 810: Monitor the user's operation response behavior after receiving the guidance content, and statistically analyze the response latency and operation correctness.

[0076] Step 820: Feedback the operation response data to the memory decay characteristic data of the modal grouping, so as to update the forgetting curve parameters online and realize the adaptive adjustment of the modality-specific forgetting difference coefficient matrix.

[0077] It is understood that data preprocessing methods known to those skilled in the art include data cleaning, data transformation, and data reduction. Data transformation includes type conversion and normalization and standardization. Although the dimensions and types of data were omitted in the description of the preceding embodiments, data preprocessing is a technical knowledge known to those skilled in the art and a prerequisite step in data processing. Therefore, the previously described well-known data preprocessing steps were not described independently.

[0078] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An AI adaptive guidance method for intelligent training scenarios, characterized in that, Includes the following steps: Obtain the user's historical operation record data, parse the input modality identifier and operation attributes of each record, and generate a modally annotated operation history dataset; For each operation record in the modality-annotated operation history dataset, feature extraction and semantic mapping are performed using a modality-specific encoder and a cross-modality aligned neural network to generate a modality-independent operation semantic embedding sequence; The modality-annotated operation history dataset is grouped according to the input modality. The mastery-forgetting time series of knowledge of each operation type within each modality group is statistically analyzed to generate memory decay feature data of modality grouping. The forgetting curve fitting algorithm is applied to the memory decay characteristic data of each modality group to extract the forgetting parameters unique to each modality, calculate the ratio of the forgetting rate parameters of the same operation type between different modalities, and generate a modality-specific forgetting difference coefficient matrix. Obtain the raw data of the user's current operation and its input modality identifier, apply the corresponding modality-specific encoder and cross-modality aligned neural network to generate a unified semantic embedding vector of the current operation, calculate the relevance score with the semantic embedding vector of each guidance entry in the guidance knowledge base, and generate a guidance content relevance ranking list; Based on the current input modality identifier, query the modality-specific forgetting difference coefficient matrix, calculate the detail weight and emphasis coefficient based on the forgetting difference coefficient, and generate modality-sensitive content generation configuration; Candidate guidance items are selected from the guidance content relevance ranking list. Based on the modality-sensitive content generation configuration, the candidate items are processed for content modulation to generate and output modality-adapted guidance content text.

2. The AI ​​adaptive guidance method for intelligent training scenarios according to claim 1, characterized in that, The cross-modal alignment neural network comprises the following components: Modal projection layer is used to project feature vectors of different dimensions output by each modality-specific encoder onto an intermediate representation space of the same dimension; The semantic alignment layer is used to map feature vectors from the intermediate representation space to a unified semantic space through nonlinear transformations. The normalized output layer is used to perform L2 normalization on the aligned semantic vectors. The modality-specific encoder and the cross-modality aligned neural network adopt an end-to-end joint training mode. During training, samples with the same operational intent executed in different modalities are constructed as positive sample pairs, and samples with different operational intents are constructed as negative samples. The loss function adopts the contrastive learning loss, which is the negative logarithm of the ratio of the exponent of the dot product of the anchor sample and the positive sample embedding vector to the sum of the exponents of the dot products of the anchor sample and all samples in the batch. Each dot product value is scaled by dividing by the temperature parameter.

3. The AI ​​adaptive guidance method for intelligent training scenarios according to claim 1, characterized in that, The forgetting curve fitting algorithm adopts an exponential decay model, which means that the memory retention rate is equal to the initial memory strength parameter multiplied by the negative power of the natural constant. The negative power is the negative value of the product of the decay rate parameter and the time interval. The historical data was fitted using the least squares method. The optimization objective was to minimize the sum of squared residuals between the actual memory retention rate and the model prediction value for each time interval in the historical data. The decay rate parameter and initial memory strength parameter specific to each mode were then extracted. The actual memory retention rate is calculated as follows: for each time interval in the mastery-forgetting time series, count the operation records that were executed again after the last correct execution within that time interval, and calculate the ratio of the number of correctly executed records to the total number of records.

4. The AI ​​adaptive guidance method for intelligent training scenarios according to claim 1, characterized in that, The modality-specific forgetting difference coefficient is calculated as follows: For the same operation type, the ratio of the decay rate parameter of the first mode to the decay rate parameter of the second mode is calculated as the forgetting difference coefficient between the first and second modes for that operation type. The forgetting difference coefficient is a positive real number, where a value of 1 indicates that the forgetting rates of the two modes are the same, a value greater than 1 indicates that the forgetting rate of the first mode is faster than that of the second mode, and a value less than 1 indicates that the forgetting rate of the first mode is slower than that of the second mode.

5. The AI ​​adaptive guidance method for intelligent training scenarios according to claim 1, characterized in that, The modality-sensitive content generation configuration includes the following parameters: The detail weight is calculated by multiplying the forgetting difference coefficient of the current mode relative to the reference mode by a time modulation factor, where the time modulation factor is 1 plus the ratio of the most recent execution time interval of the user's operation on the current mode to a reference time interval threshold. The key point is the coefficient, which is used to determine the proportion of each knowledge point in the guidance content. It is calculated by dividing the decay rate parameter of the operation type corresponding to the knowledge point in the current modality by the sum of the decay rate parameters of the operation types corresponding to all knowledge points included in the current guidance entry. The expression style is identified by the action description template for the motor memory-dominated modality and the concept explanation template for the descriptive memory-dominated modality.

6. The AI ​​adaptive guidance method for intelligent training scenarios according to claim 1, characterized in that, The calculation of the relevance score with the semantic embedding vector of each guidance entry in the guidance knowledge base includes: Calculate the cosine distance between the unified semantic embedding vector of the current operation and the semantic embedding vectors of each guide entry. This cosine distance is 1 minus the ratio of the dot product of the two vectors to the product of the magnitudes of the two vectors. The cosine distance is converted into a correlation score by applying the distance decay function. This correlation score is a negative power of the natural constant, which is the negative product of the distance decay coefficient and the cosine distance. The guidance items are sorted from highest to lowest relevance score to generate a sorted list of guidance content relevance.

7. The AI ​​adaptive guidance method for intelligent training scenarios according to claim 1, characterized in that, The content modulation processing of the candidate entries includes: The original content of candidate entries is expanded or compressed according to the detail weight. When the detail weight is greater than 1, the extended description of the candidate guidance entry is retrieved from the guidance knowledge base to supplement it. When the detail weight is less than 1, the content is semantically compressed and the core points are retained. Adjust the length allocation of each knowledge point according to the emphasis coefficient; Select the corresponding expression template according to the expression style identifier.

8. The AI ​​adaptive guidance method for intelligent training scenarios according to claim 1, characterized in that, The modality-specific encoder adopts a corresponding network structure based on the input modality type: For touch screen mode and mouse mode, a bidirectional long short-term memory network is used as the encoder. The input layer receives the operation feature vector sequence of time step serialization, the hidden layer extracts the temporal dependency features through two LSTM units, forward and backward, and the output layer concatenates the forward hidden state and the backward hidden state of the last time step as the mode-specific feature vector. For speech modalities, a one-dimensional convolutional neural network is used as the encoder. The input layer receives the audio spectral feature sequence, the convolutional layer extracts local spectral features through multi-scale convolutional kernels, and the output layer generates modality-specific feature vectors through global average pooling.

9. The AI ​​adaptive guidance method for intelligent training scenarios according to claim 1, characterized in that, It also includes the following steps: Monitor user response behavior after receiving guidance content, and statistically analyze response latency and operational correctness; The operation response data is fed back to the memory decay characteristic data of the modality group, which is used to update the forgetting curve parameters online and realize the adaptive adjustment of the modality-specific forgetting difference coefficient matrix.

10. An AI adaptive guidance system for intelligent training scenarios, used to execute the AI ​​adaptive guidance method for intelligent training scenarios according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire the user's historical operation records and the raw data of the current operation; The modality coding module includes modality-specific encoders corresponding to each input modality, which are used to convert the raw operational data of each modality into modality-specific feature vectors; The cross-modal alignment module, including the cross-modal alignment neural network, is used to map modality-specific feature vectors to a unified semantic space to generate modality-independent semantic embedding vectors; The forgetting modeling module is used to statistically analyze the mastery-forgetting time series by modality and to generate a modality-specific forgetting difference coefficient matrix by applying a forgetting curve fitting algorithm. The semantic matching module is used to calculate the relevance score between the unified semantic embedding vector of the current operation and each guide entry in the guide knowledge base and generate a sorted list; The content generation module is used to modulate the candidate guidance items according to the modality-sensitive content generation configuration and generate modality-adapted guidance content text; The content presentation interface is used to output modally adapted introductory text.

Images