Cognitive feature mining method for enhancing human-intelligent cooperation mechanism

Abstract

The invention relates to the technical field of cognitive feature mining, and particularly discloses a cognitive feature mining method for enhancing a human-intelligent cooperation mechanism. The method comprises the following steps: collecting physiological and psychological characteristics of a user to construct a cognitive preference database; constructing a symbol sequence through a segmented aggregation approximation dimension reduction and symbol approximation theory; extracting an optimized population based on a multi-strategy enhanced escape algorithm, constructing an elite pool to adjust a cognitive boundary, reversely mapping a feature vector into a symbol sequence and labeling an interaction intention to form a symbol data set; and training an LARA model based on the symbol data set, and obtaining the human-intelligent cooperation mutual reliability and evaluating the human-intelligent cooperation mutual reliability in combination with the model recognition accuracy. According to the method, the comprehensiveness of cognitive features is guaranteed through the multi-modal data, the symbol screening accuracy is improved, and the human-intelligent synergy effect is enhanced.

Description

Technical Field

[0001] This invention relates to the field of cognitive feature mining technology, and more specifically to a cognitive feature mining method for enhancing human-intelligence collaborative mechanisms. Background Technology

[0002] With the development of intelligent manufacturing technology, industrial robots are being used more and more widely in production and testing processes. Among them, intelligent torque detection system robots are key equipment for accurately controlling the torque detection process.

[0003] In the current practice of human-robot collaboration in intelligent torque detection systems, cognitive feature mining is the core link connecting user needs and robot responses, but existing technologies still have significant shortcomings: Existing technologies do not take into account the specific application scenarios of intelligent torque detection system robots, such as the button layout and operation feedback logic of the torque detection interface of such robots, as well as the cognitive characteristics of users in specific tasks such as torque monitoring and threshold warning, and do not design targeted symbolic stimulus carriers (such as color and shape symbols that fit the detection scenario) and experimental schemes. As a result, the collected cognitive data is difficult to fully cover the neuropsychological and physiological dimensions of users in actual interaction, and cannot capture the cognitive orientation of users in torque detection tasks.

[0004] Therefore, this invention provides a cognitive feature mining method to enhance human-intelligence collaborative mechanisms. Summary of the Invention

[0005] The purpose of this invention is to provide a cognitive feature mining method that enhances human-intelligence collaborative mechanisms, so as to solve the aforementioned background problems.

[0006] The objective of this invention can be achieved through the following technical solutions: A method for mining cognitive features to enhance human-intelligence collaborative mechanisms includes the following steps: Collect the neuropsychological and physiological characteristics of user interactions with the robot, and establish a cognitive preference database that includes psychological and physiological characteristics; The cognitive database is segmented, aggregated, and approximate dimensionality reduced to obtain a reconstructed physiological indicator sequence. The symbolic approximation theory is then applied to symbolize the physiological indicator sequence to construct a symbolic sequence. A cognitive orientation identification model is constructed based on a multi-strategy enhanced escape algorithm. The symbol sequence is input into the cognitive orientation identification model, the optimized population is extracted and the cognitive boundary is adjusted. Functional-structural symbols that conform to cognitive orientation are selected to establish a symbol dataset. The LARA model, suitable for human-computer dialogue, is trained based on a symbolic dataset. The interaction fluency of the LARA model is mapped through the user's physiological characteristics to obtain the human-computer collaboration trust level and evaluate the degree of human-computer collaboration.

[0007] As a further aspect of the present invention, the method for performing the segmented aggregation approximate dimensionality reduction is as follows: Extract the physiological indicator sequence corresponding to each physiological indicator in the cognitive database, and divide the normalized physiological indicator sequence into n subsequences. The length of the original physiological indicator sequence is N; The mean of each subsequence is calculated as a representative value of the subsequence, which is used to reduce the length of the original physiological indicator sequence from N to n.

[0008] As a further aspect of the present invention, the symbolization process is performed as follows: The physiological index sequence is reconstructed based on the representative value of each subsequence, and the data interval of the physiological index sequence is divided into k equal probability intervals using a Gaussian distribution model. Calculate the maximum and minimum values ​​of each subsequence, and define the symbol corresponding to each subsequence based on the representative value; By combining the symbols corresponding to each subsequence, each representative value in the reconstructed physiological index sequence is mapped to an equally probable interval to construct a symbol sequence.

[0009] As a further aspect of the present invention: the cognitive orientation identification model is constructed as follows: Construct the feature space of the symbol sequence; An adaptive perturbation factor is constructed, an adaptive control factor is applied to the feature space, and a set of feature vectors is constructed. A restart mechanism is constructed based on the feature vector set to identify and update inferior individuals; Based on the updated individuals, a dynamic centroid back-learning strategy is implemented to output an optimized population.

[0010] As a further aspect of the present invention, the restart mechanism is constructed as follows: Treat each feature vector in the feature vector set as an individual; Set rules for identifying substandard individuals, and mark individuals that meet the rules as substandard individuals; By inputting the feature vector of the inferior individual into the individual position update equation, the feature vector of the worst individual after the update is obtained.

[0011] As a further aspect of the present invention: the method for implementing the dynamic centroid reverse learning strategy is as follows: Based on the updated individuals, obtain the updated population; The centroids of all eigenvectors in the updated population are calculated using the centroid equation. Construct the inverse equation of the centroid, input the centroid into the inverse equation of the centroid to generate the inverse solution; The generated inverse solution is merged with the feature vector corresponding to the original candidate symbol sequence. The fitness of each merged individual is calculated. Individuals are then selected based on their fitness to construct an optimized population after the selection of individuals.

[0012] As a further aspect of the present invention, the method for adjusting the cognitive boundary is as follows: An elite pool is constructed based on an optimized population, and new solutions are generated by adjusting the boundaries of the elite pool. The new solution is adapted and updated, the elite pool boundary is redefined based on the processing results, and the elite pool is adjusted based on the elite pool boundary. A symbolic dataset is constructed by performing symbolic inverse mapping based on the adjusted elite pool.

[0013] As a further aspect of the present invention, the elite pool is constructed as follows: The fitness of the optimized population is obtained through fitness analysis. The population is then sorted in descending order of fitness, and an elite pool is constructed based on the sorting results.

[0014] As a further aspect of the present invention, the method for adaptively updating the new solution is as follows: Construct the feature vector corresponding to the new solution and calculate the fitness. Update the individuals in the elite pool based on the fitness. Based on the updated elite pool, obtain the upper and lower bounds for each dimension; By constructing a system of boundary equations, the upper and lower boundaries and the new solution are input into the system of boundary equations. The new solution and the boundary are then judged, and the boundary range is adjusted based on the judgment result.

[0015] As a further aspect of the present invention, the method for obtaining the human-intelligence collaborative mutual trust level is as follows: Acquire the physiological characteristics of users after they have engaged in dialogue through the LARA model, as well as the baseline data of their physiological characteristics before using the LARA model; Extract the alpha wave power and HRV value from the physiological characteristics of the brain, and calculate the rate of change of the alpha wave power and HRV value relative to the baseline data to obtain the power change rate and HRV change rate; The objective smoothness coefficient is obtained by summing and averaging the power change rate and HRV change rate. Obtain the accuracy of the LARA model in recognizing symbol sequences and interaction intentions, and use it as the model response matching value; Calculate the mean of the objective fluency coefficients of all users and the mean of the model response matching values, and multiply the two means to obtain the human-intelligence collaborative mutual trust degree.

[0016] The beneficial effects of this invention are: (1) By designing color symbols, shape symbols and interactive behavior symbols that fit the robot scenario of the intelligent torque detection system, and conducting interactive experiments in conjunction with emotion induction and cognitive tasks, the user's EEG, eye movement, physiological characteristics and cognitive load data are collected simultaneously. The constructed cognitive preference database can cover the neuropsychological and physiological dimensions of user cognition, providing comprehensive and practical data support for subsequent cognitive feature mining.

[0017] (2) The physiological indicator sequence in the cognitive preference database is segmented and aggregated to approximate dimensionality reduction, which can simplify the data dimension while retaining the key change trend of the physiological indicators; then, the reconstructed physiological indicator sequence is mapped to a symbol sequence through symbolic approximation theory, which can transform the abstract physiological data into a more easily analyzed symbolic form, providing a clear and processable feature carrier for the construction of subsequent cognitive orientation identification models.

[0018] (3) A cognitive orientation identification model is constructed based on the multi-strategy enhanced escape algorithm. The feature vector population is optimized by strategies such as adaptive perturbation, restart mechanism, and dynamic centroid reverse learning, which helps to increase the proportion of high-quality individuals in the population and lays the foundation for screening highly adaptable cognitive features. Subsequently, an elite pool is constructed and the cognitive boundary is adjusted. Finally, through the symbol dataset formed by symbol reverse mapping and interaction intent annotation, the selected functional-structural symbols can be associated with the actual human-computer interaction intent, providing targeted data for the subsequent training of LARA (Linguistic-Adaptive Retrieval-Augmented Language Models) model.

[0019] (4) By mapping the interaction fluency coefficient through physiological characteristics and combining it with the model recognition accuracy to calculate the human-intelligence collaboration mutual trust, we can provide an objective basis for evaluating the human-intelligence collaboration effect, help to judge whether human-intelligence collaboration meets the expectations of cognitive collaboration, and support the optimization of the human-intelligence collaboration mechanism. Attached Figure Description

[0020] The invention will now be further described with reference to the accompanying drawings.

[0021] Figure 1 This is a flowchart of the cognitive feature mining method for enhancing human-intelligence collaborative mechanisms according to the present invention; Figure 2 This is a module diagram of the cognitive feature mining system for enhancing human-intelligence collaborative mechanisms in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1 Please see Figure 1 As shown, this invention is a cognitive feature mining method for enhancing human-intelligence collaborative mechanisms, comprising the following steps: S1. Collect the neuropsychological and physiological characteristics of user interactions with the robot, and establish a cognitive preference database containing psychological and physiological characteristics; The method for collecting neuropsychological and physiological characteristics of user interactions with robots and establishing a cognitive preference database that includes psychological and physiological characteristics is as follows: Preferably, color symbols, shape symbols, and interactive behavior symbols are designed as the initial carriers for interactive experiments; Among them, blue color palettes with different brightness and saturation (such as Pantone 2945C and 286C) are used as color symbols; The circular and square geometric shapes are combined with blue to form shape symbols; the circular symbol corresponds to the "confirm" operation feedback on the robot torque detection interface of the intelligent torque detection system, and the square symbol corresponds to the "setting" operation entry. The button layout and operation feedback logic of the torque detection interface in the intelligent torque detection system robot are used as interactive behavior symbols. The button layout follows the principle of centering frequently used functions (such as resetting torque values) and placing secondary functions (such as querying historical data) on the side. The operation feedback logic is set to "an audio-visual prompt (low-frequency buzzer + blue flashing button) appears within 0.5 seconds after clicking". Color symbols and shape symbols are combined as symbolic stimulus variables, and combined with emotion-induced variables and cognitive task variables, the three variables are randomly presented to users to achieve an interactive experiment; It should be noted that the emotion-inducing variables include playing standardized videos; the cognitive task variables are based on applying low- or high-load interactive tasks using interactive behavioral symbols: the low-load task is to identify the color of the interface indicator light (red / green) and judge the device status; the high-load task is to monitor torque data in real time (0-500 N·m) and click the warning button when the value exceeds the threshold (±5% of the set value); before the experiment, the subjects' EEG / ERP data and HRV data in a natural state for 30 seconds were collected as baseline data. The symbol presentation time was 30 seconds (during which the subjects had to complete the corresponding cognitive task), and the interval between adjacent symbol stimulation variables was 15 seconds (during which the subjects relaxed and returned to the baseline state). The user's EEG, eye movement, and physiological indicators were collected over N monitoring periods following the interactive experiment and used as physiological characteristics. Preferably, N=50; It should be noted that the EEG indicators include the power of the user's alpha waves (8-13Hz, reflecting a relaxed state) and beta waves (14-30Hz, reflecting a focused state), as well as the P300 component in the event-related potentials (ERPs) (reflecting the cognitive decision-making process). These data were collected using the SmartBCI portable EEG system (sampling rate 250Hz, leads selected as Fz, Cz, and Pz, with the reference electrode placed behind the ear). Eye-tracking metrics included the user's fixation time ratio (fixation time in a specific symbol area / total symbol presentation time) and the rate of change in gaze heatmap (overlap area of ​​gaze heatmaps within 5 seconds / total area of ​​heatmaps), collected using an EyeLogic eye tracker (sampling rate 60Hz, fixation point recognition threshold 100ms); physiological metrics included the user's HRV value, collected using an ErgoLAB physiological acquisition module (HRV sampling rate 1000Hz). Collect users' NASA-TLX cognitive load scale and extract the overall cognitive load value as a psychological characteristic; The NASA-TLX Cognitive Load Scale scores users based on six dimensions: mental effort (the mental effort required to complete a task), physical effort (the physical effort required to complete a task), time effort (the time pressure required to complete a task), effort level (the subjective effort put into completing a task), work performance (the subject's evaluation of the quality of their task completion), and frustration level (the subject's sense of frustration due to the difficulty of the task). Each dimension has a score range of 0-100. By constructing the overall cognitive load equation: Obtain the overall cognitive load value T; n is the number of indicators in the 15 pairwise comparisons (15 pairwise comparisons are generated from 6 dimensions to determine the weight of each dimension), and M is the scale score of each dimension (fixed at 100, i.e. the full score benchmark for each dimension). The score value for each dimension i; The scale score is fixed at 100, which is the full score benchmark for each dimension and the score value for each dimension i. Outlier handling was performed on physiological and psychological characteristics, and these characteristics were standardized. It should be noted that outlier handling adopted the 3σ criterion, removing data that exceeded the mean ± 3 standard deviations; and the Z-score normalization method was used to standardize physiological and psychological characteristics. The processed data is organized according to the database structure, that is, a sequence of each physiological indicator is established according to the monitoring cycle, such as the EEG alpha wave power sequence. ,in This is data from the first period. This is data from the second period. To construct a cognitive preference database of psychological and physiological characteristics for the 50th cycle of data.

[0024] S2. Perform segmented aggregation and approximate dimensionality reduction on the cognitive database to obtain a reconstructed physiological indicator sequence. Apply symbolic approximation theory to symbolize the physiological indicator sequence and construct a symbolic sequence. The method for obtaining the reconstructed physiological indicator sequence by segmenting, aggregating, and approximating the dimensionality of the cognitive database is as follows: Preferably, the physiological indicator sequence corresponding to each physiological indicator in the cognitive database is extracted, and the normalized physiological indicator sequence is divided into n subsequences. Preferably, n=10 (that is, 50 seconds of physiological data are divided into 10 subsequences, and each subsequence corresponds to 5 seconds of data). The mean of each subsequence is calculated as the representative value of the subsequence, which is used to reduce the length of the original physiological index sequence from N to n, that is, to achieve segmented aggregation approximate dimensionality reduction (PAA). Among them, applying symbolic approximation theory to identify cognitive orientations in physiological indicator sequences, the method for extracting function-structure symbols that conform to cognitive orientations is as follows: The physiological indicator sequence is reconstructed based on the representative value of each subsequence (the reconstructed sequence length is n=10, e.g.) ,in (The mean α-wave power of the first subsequence), and the data interval of the physiological index sequence is divided into k equal probability intervals by a Gaussian distribution model; It should be noted that the value of k is determined based on the discrimination requirements of physiological indicators, with k=5 being the preferred value; the probability of each interval is 20%, that is, the interval boundaries are determined by the quantiles of the Gaussian distribution, such as the first interval being (-∞, μ-1.28σ), the second interval being (μ-1.28σ, μ-0.52σ), the third interval being (μ-0.52σ, μ+0.52σ), the fourth interval being (μ+0.52σ, μ+1.28σ), and the fifth interval being (μ+1.28σ, +∞); Calculate the maximum and minimum values ​​of each subsequence, and define the symbol corresponding to each subsequence in combination with the representative value. For example, when the representative value is in the 3rd interval, the maximum value is <μ+0.84σ, and the minimum value is >μ-0.84σ, it is defined as the symbol "C" to ensure that the symbol can reflect the overall distribution characteristics of the subsequence. By combining the symbols corresponding to each subsequence, each representative value in the reconstructed physiological index sequence is mapped to an equal probability interval to construct a symbol sequence; For example, the mapping process is performed as follows: If the EEG alpha wave power sequence of one of the physiological index sequences is reduced to [0.2, -0.5, 0.8, -0.1, 0.3, -0.7, 1.0, -0.2, 0.6, 1.1] by PAA dimensionality reduction, and according to the mapping of 5 equally probable intervals, 0.2 is located in interval 3 (corresponding to symbol "C"), -0.5 is located in interval 2 (corresponding to symbol "B"), 0.8 is located in interval 4 (corresponding to symbol "D"), -0.1 is located in interval 3 (corresponding to symbol "C"), 0.3 is located in interval 3 (corresponding to symbol "C"), -0.7 is located in interval 1 (corresponding to symbol "A"), 1.0 is located in interval 4 (corresponding to symbol "D"), -0.2 is located in interval 3 (corresponding to symbol "C"), 0.6 is located in interval 4 (corresponding to symbol "D"), and 1.1 is located in interval 5 (corresponding to symbol "E"), then the mapped symbol sequence is [C, B, D, C, C, A, D, C, D, E].

[0025] Example 2 Please see Figure 1 As shown, this invention is a cognitive feature mining method for enhancing human-intelligence collaborative mechanisms, comprising the following steps: S3. Construct a cognitive orientation identification model based on a multi-strategy enhanced escape algorithm. Input the symbol sequence into the cognitive orientation identification model, extract the optimized population and adjust the cognitive boundary, select functional-structural symbols that conform to the cognitive orientation, and establish a symbol dataset. The method for constructing a cognitive orientation identification model based on a multi-strategy enhanced escape algorithm is as follows: S301. Construct the feature space of the symbol sequence; Preferably, the symbol sequence is converted into a feature vector, where each symbol in the symbol sequence represents a dimension, and the value of each dimension is the interval index corresponding to the symbol; For example, the feature vector corresponding to the symbol sequence [C,B,D,C,C,A,D,C,D,E] is [3,2,4,3,3,1,4,3,4,5]; S302. Construct an adaptive perturbation factor, apply an adaptive control factor to the feature space, and construct a set of feature vectors; Preferably, through the adaptive perturbation equation: ( Obtain the adaptive perturbation factor (ADF); Where RAND is a random number between 0 and 1. The maximum number of iterations (determined by the dimension of the feature space); Preferred, =200, meaning each feature vector needs to undergo 200 iterations of optimization; This represents the current iteration number. The initial value is 1. It is incremented by 1 after each candidate solution generation and evaluation, until... = Stop iteration when the time is right; Obtain the feature vectors corresponding to all physiological indicator sequences, and apply ADF perturbation to each dimension of each feature vector; For example, the ADF perturbation can be applied as follows: if the original dimension is 3 and ADF=0.6, the perturbation value is 3±round(ADF×1), i.e., 2, 3 or 4, to ensure that the perturbation amplitude is moderate and to reduce the deviation of the symbol sequence from the original cognitive features. Based on the eigenvectors with applied ADF perturbation, new candidate solutions are regenerated, and a candidate symbol sequence containing all candidate solutions is constructed. The candidate symbol sequences are transformed into feature vectors to construct a feature vector set containing all candidate symbol sequences. S303. Construct a restart mechanism based on the feature vector set to identify and update inferior individuals; Each feature vector in the feature vector set is treated as an individual, representing a combination of symbols with cognitive characteristics; It should be noted that each user corresponds to one set of candidate symbol sequences, that is, one feature vector individual. The set of all individuals constitutes the initial population, and the population size is the total number of individuals. The population size is denoted as Num. Based on the cognitive orientation identification objective (screening symbol sequences that are strongly correlated with the user's cognitive load), establish rules for judging inferior individuals; Individuals that meet the individual determination rules are marked as inferior individuals; The method for setting the criteria for identifying inferior individuals is as follows: calculate the Pearson correlation coefficient r between the symbol sequence corresponding to each feature vector individual and the overall cognitive load value. If r < 0.3 (i.e., the symbol sequence has no effective representation of cognitive orientation and belongs to redundant or interfering cognitive features), then the feature vector individual is identified as an inferior individual. Preferably, the worst individual factor is set to 0.05, and the number of worst individuals is 0.1 × Num; Equation updated via individual location: Obtain the feature vector of the worst individual i after the update; in, , The best individual in the current population (i.e., the feature vector with the highest correlation between the symbol sequence and the overall cognitive load value, such as the vector [3,2,4,3,3,1,4,3,4,5] with r=0.75); For poor-quality feature vectors to be optimized (e.g., vectors [1,1,2,1,1,3,2,1,2,3] that are judged to be poor-quality); The distance between the cognitive features of the worst and best individuals is used to determine the standard deviation of the Gaussian distribution, ensuring that the updated individuals are distributed around the best individual. Follow For the mean, A Gaussian distribution with standard deviation; It is understandable that by updating the worst individual's position equation, the worst individual can move closer to the best individual. Introducing randomness avoids complete consistency with the best individual, increases the probability of the algorithm finding the global optimal solution, and reduces the slow convergence of the algorithm due to some poor individuals. S304. Implement a dynamic centroid reverse learning strategy based on the updated individuals to output an optimized population. Based on the updated individuals, obtain the updated population; Preferably, using the centroid equation: Calculate the centroid C of all eigenvectors in the updated population; Where Num represents the population size. Let be the i-th eigenvector in the population (each dimension takes the mean; for example, if the first dimension of 55 eigenvectors takes values ​​of 3, 2, 4, ..., 3, and the mean is 3.1, then the first dimension of the centroid C is 3.1). It should be noted that the centroid C represents the average cognitive characteristics of the population, which can reflect the cognitive orientation of most subjects. Construct the inverse equation of the centroid: Inverse solution generated based on centroid ; Where U is the upper boundary of each dimension of the feature vector (if the maximum interval index corresponding to the sign is 5, then U=5), and L is the lower boundary of each dimension of the feature vector (if the minimum interval index corresponding to the sign is 1, then L=1). It is understandable that the physical meaning of generating inverse solutions is that inverse solutions can expand the search space, for example, a certain feature vector. The first dimension is 2, and the first dimension of the centroid C is 3.1. Then the first dimension of the reverse solution is 3.1 + (5-1) - 2 = 5.1, which is rounded down to 5, corresponding to the symbol "E". This symbol may represent the special cognitive orientation of a few subjects, reducing the algorithm from missing potential high-quality symbols. The generated inverse solution is merged with the feature vector corresponding to the original candidate symbol sequence, and the fitness of each merged individual is calculated (fitness is determined based on the correlation between the symbol sequence corresponding to the individual and the overall cognitive load value; the higher the correlation, the higher the fitness). The top 50% of individuals by fitness were selected and retained (the number of individuals retained remained at 55-70, consistent with the original population size), thus obtaining an optimized population. Among them, the method for adjusting cognitive boundaries and selecting functional-structural symbols that conform to cognitive orientation is as follows: S311. Construct an elite pool based on optimized population, and generate new solutions by adjusting the boundaries of the elite pool; Preferably, the fitness of the optimized population is obtained by fitness analysis based on the output, and the individuals are sorted in descending order according to fitness. The top 20% of individuals in the sorting are then used to construct an elite pool. The fitness analysis is performed as follows: Obtain the cognitive load sequence, which is composed of the total cognitive load values ​​within the monitoring period, and calculate the Pearson correlation coefficient r between each feature vector in the optimized population and the cognitive load sequence; Where r takes values ​​in the range [-1, 1], r > 0 indicates a positive correlation, and the larger the r is, the better the symbol sequence reflects cognitive load). Obtain the matching degree S of the user's cognitive orientation using the fitness equation: Obtain fitness F; It should be noted that the matching degree of user cognitive orientation was scored by three experts in the field of human-computer interaction. The scoring criteria were the relevance of the symbol to the torque detection task (0-4 points), the recognizability of the symbol (0-3 points), and the emotional regulation effect of the symbol (0-3 points), with a maximum score of 10 points. Through the guiding equation: Calculate new solutions in the elite pool ; in, and For weighting coefficients, the preferred ones are... and They are 0.8 and 0.2 respectively; The weight is higher, which is used to make new solutions mainly converge on individuals in the elite pool (E_pool,d) to ensure the consistency of cognitive orientation; With lower weights, random individuals are introduced ( This information helps reduce homogenization among individuals within the elite pool. , These are adjustment coefficients, all of which are set to 1. E_pool,3 is a candidate solution in the d-th dimension of the elite pool (e.g., if the 3rd dimension of an individual in the elite pool is 4, then E_pool,3=4). Let d be the data of a single individual randomly selected from the entire candidate sequence. This represents the d-th dimension of the data for individuals in the current elite pool. S312. Perform adaptive update processing on the new solution, redetermine the elite pool boundary based on the processing result, and adjust the elite pool based on the elite pool boundary. Construct the feature vector corresponding to the new solution and calculate its fitness; if the fitness of the new solution is higher than that of the individual with the lowest fitness in the elite pool, then replace the individual with the lowest fitness in the elite pool with the new solution to complete the elite pool update. If the fitness of a new solution is less than 80% of the average fitness of individuals in the elite pool, then discard the new solution and... Adjust to a new value within the range of 0.7-0.9, as originally... =0.8, adjust to 0.85. Adjust to a new value within the range of 0.1-0.3 (as in the original). After adjusting the value from 0.2 to 0.15, the new solution is recalculated by substituting it into the guiding equation. Based on the updated elite pool, obtain the upper boundary U and lower boundary L for each dimension, where U is the maximum value of the data in that dimension of the elite pool, and L is the minimum value of the data in that dimension of the elite pool. For example, if the second dimension of the 5 individuals in the elite pool has values ​​of 2, 2, 3, 2, 3, then U=3 and L=2; By constructing a system of boundary equations, the upper and lower boundaries and the new solution are input into the system of boundary equations. The new solution and the boundary are judged, and the boundary range is adjusted based on the judgment result. Preferably, the method for constructing the boundary equation system is as follows: Through formula one: ; Through formula two: Determine if the new solution exceeds the current boundary. If FlagU=1 (e.g., y_new,i,d=4,U=3,4 / 3>1) or =1 (e.g., y_new,i,d=1, L=2, 1 / 2<1), then readjust the boundary range and update U to (e.g., U is updated to 4), update L to (e.g., L is updated to 1); S314. Perform symbol reverse mapping based on the adjusted elite pool to construct a symbol dataset; Extract all feature vectors from the updated elite pool, and according to the set "symbol-numerical mapping rules" (A=1, B=2, C=3, D=4, E=5), convert each feature vector in reverse into the corresponding symbol sequence: Example: The feature vector [3,2,4,3,3,1,4,3,4,5] in the elite pool is inversely mapped to the symbol sequence [C,B,D,C,C,A,D,C,D,E]. This symbol sequence corresponds to a complete "color-shape-interaction" combination, such as C corresponding to "aviation blue circle", B corresponding to "aviation blue square", and D corresponding to "aviation blue flashing circle". Based on the actual task scenarios of the intelligent torque detection system robot (such as torque zeroing, parameter setting, and anomaly warning), each symbol sequence is labeled with a unique "interaction intent tag"—the tag must be strongly correlated with the symbol's functional-structural attributes, and the specific rules are as follows: If the symbol sequence contains "aviation blue circle" (corresponding to the value 3) and appears ≥ 3 times: it is marked as "operation confirmation intention" (such as confirmation feedback of torque zeroing operation); If the symbol sequence contains "aviation blue square" (corresponding to the value 2) and is accompanied by "low-frequency repetitive structure" (such as two consecutive 2s): it is marked as "parameter setting intention" (such as the adjustment entry of torque threshold). The symbol dataset is constructed by combining all symbol sequences and interaction intent labels from the elite pool.

[0026] S4. Train an LARA model suitable for human-computer dialogue based on a symbolic dataset, map the interaction fluency of the LARA model through the user's physiological characteristics, obtain the human-computer collaboration trust level, and evaluate the human-computer collaboration level. The method for training an LARA model suitable for human-computer dialogue based on a symbolic dataset is as follows: Preferably, based on the symbol dataset, the psychological and physiological characteristics of the corresponding symbol sequences in the associated cognitive preference database and the interactive intent labels are preprocessed using one-hot encoding, and then divided into training, validation and test sets in a 7:2:1 ratio; The LARA (Linguistic-Adaptive Retrieval-Augmented Language Models) model adopts an Encoder-Decoder architecture. The Encoder module extracts local cognitive features of symbol sequences through CNN algorithm, fuses psychophysiological temporal features with LSTM and concatenates them into 128-dimensional fused features. The Decoder module first completes the interactive intent recognition with a fully connected layer, and then generates natural language dialogue response through LSTM + attention mechanism. The initial hyperparameter values ​​for the LARA model are as follows: CNN layers (2 layers, 3×3 kernel size, 64 neurons), LSTM layers (128 hidden neurons, dropout=0.2); the adjustment method is as follows: if the validation set mean squared error (MSE) > 5%, the number of CNN layers is increased to 3, the number of hidden neurons in the LSTM layer is adjusted to 160, and the model is retrained until the MSE ≤ 5%. The LARA model learns parameters using the training set and tunes hyperparameters using the validation set. The method for evaluating the mutual trust in human-intelligence collaboration by mapping the interaction fluency coefficient of the LARA model using physiological features is as follows: Acquire the physiological characteristics of users after they have engaged in dialogue through the LARA model, as well as the baseline data of their physiological characteristics before using the LARA model; Extract the alpha wave power and HRV value from the physiological characteristics of the brain, and calculate the rate of change of the alpha wave power and HRV value relative to the baseline data to obtain the power change rate and HRV change rate; The objective smoothness coefficient is obtained by summing and averaging the power change rate and HRV change rate. Obtain the accuracy of the LARA model in recognizing symbol sequences and interaction intentions, and use it as the model response matching value; Calculate the mean of the objective fluency coefficient of all users and the mean of the model response matching value, and multiply the two means to obtain the human-intelligence collaborative mutual trust degree. The human-intelligence synergy trust level is compared with a preset synergy trust threshold. If the human-intelligence synergy trust level is higher than or equal to the preset synergy trust threshold, the LARA model is determined to meet the expectation of cognitive synergy, that is, the degree of human-intelligence synergy meets the expectation. It should be noted that the preset collaborative trust threshold is 0.7. This threshold is determined based on pre-experimental data from 55-70 healthy subjects. When the threshold is ≥0.7, the user's operational dependence on the intelligent torque detection system robot reaches more than 85%, which meets the basic requirements of human-intelligent collaboration in industrial scenarios. If the mutual trust level of human-intelligence collaboration is lower than the preset collaboration trust threshold, then the LARA model is determined to not meet the expectations of cognitive collaboration.

[0027] Example 3 Please see Figure 2 As shown, this invention is a cognitive feature mining system for enhancing human-intelligence collaborative mechanisms, comprising the following modules: Feature acquisition module: used to collect the neuropsychological and physiological characteristics of user interaction with the robot, and to establish a cognitive preference database containing psychological and physiological characteristics; Symbolic processing module: used to perform segmented aggregation and approximate dimensionality reduction on the cognitive database to obtain a reconstructed physiological indicator sequence, and to apply symbolic approximation theory to symbolize the physiological indicator sequence to construct a symbolic sequence; Boundary Adjustment Module: Based on the multi-strategy enhanced escape algorithm, a cognitive orientation identification model is constructed. The symbol sequence is input into the cognitive orientation identification model, the optimized population is extracted and the cognitive boundary is adjusted. Functional-structural symbols that conform to the cognitive orientation are selected to establish a symbol dataset. Collaborative evaluation module: Based on a symbolic dataset, an LARA model suitable for human-computer dialogue is trained. The interaction fluency of the LARA model is mapped through the user's physiological characteristics to obtain the human-computer collaborative trust level and evaluate the degree of human-computer collaboration.

[0028] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A cognitive feature mining method to enhance human-intelligence collaborative mechanisms, characterized by: Includes the following steps: Collect the neuropsychological and physiological characteristics of user interactions with the robot, and establish a cognitive preference database that includes psychological and physiological characteristics; The cognitive database is segmented, aggregated, and approximate dimensionality reduced to obtain a reconstructed physiological index sequence. The symbolic approximation theory is then applied to symbolize the physiological index sequence to construct a symbolic sequence. A cognitive orientation identification model is constructed based on a multi-strategy enhanced escape algorithm. The symbol sequence is input into the cognitive orientation identification model, the optimized population is extracted and the cognitive boundary is adjusted. Functional-structural symbols that conform to cognitive orientation are selected to establish a symbol dataset. The LARA model, suitable for human-computer dialogue, is trained based on a symbolic dataset. The interaction fluency of the LARA model is mapped through the user's physiological characteristics to obtain the human-computer collaboration trust level and evaluate the degree of human-computer collaboration.

2. The cognitive feature mining method for enhancing human-intelligence collaborative mechanisms according to claim 1, characterized in that: The method for performing the approximate dimensionality reduction of the segmented aggregation is as follows: Extract the physiological indicator sequence corresponding to each physiological indicator in the cognitive database, and divide the normalized physiological indicator sequence into n subsequences. The length of the original physiological indicator sequence is N; The mean of each subsequence is calculated as a representative value for the subsequence, which is used to reduce the length of the original physiological indicator sequence from N to n.

3. The cognitive feature mining method for enhancing human-intelligence collaborative mechanisms according to claim 1, characterized in that: The symbolization process is performed as follows: The physiological index sequence is reconstructed based on the representative value of each subsequence, and the data interval of the physiological index sequence is divided into k equal probability intervals using a Gaussian distribution model. Calculate the maximum and minimum values ​​of each subsequence, and define the symbol corresponding to each subsequence based on the representative value; By combining the symbols corresponding to each subsequence, each representative value in the reconstructed physiological index sequence is mapped to an equally probable interval to construct a symbol sequence.

4. The cognitive feature mining method for enhancing human-intelligence collaborative mechanisms according to claim 1, characterized in that: The cognitive orientation identification model is constructed as follows: Construct the feature space of the symbol sequence; An adaptive perturbation factor is constructed, an adaptive control factor is applied to the feature space, and a set of feature vectors is constructed. A restart mechanism is constructed based on the feature vector set to identify and update inferior individuals; Based on the updated individuals, a dynamic centroid back-learning strategy is implemented to output an optimized population.

5. The cognitive feature mining method for enhancing human-intelligence collaborative mechanisms according to claim 4, characterized in that: The restart mechanism is constructed as follows: Treat each feature vector in the feature vector set as an individual; Set rules for identifying substandard individuals, and mark individuals that meet the rules as substandard individuals; By inputting the feature vector of the inferior individual into the individual position update equation, the feature vector of the worst individual after the update is obtained.

6. The cognitive feature mining method for enhancing human-intelligence collaborative mechanisms according to claim 4, characterized in that: The method for implementing the dynamic centroid back-learning strategy is as follows: Based on the updated individuals, obtain the updated population; The centroids of all eigenvectors in the updated population are calculated using the centroid equation. Construct the inverse equation of the centroid, input the centroid into the inverse equation of the centroid to generate the inverse solution; The generated inverse solution is merged with the feature vector corresponding to the original candidate symbol sequence. The fitness of each merged individual is calculated. Individuals are then selected based on their fitness to construct an optimized population after the selection of individuals.

7. The cognitive feature mining method for enhancing human-intelligence collaborative mechanisms according to claim 1, characterized in that: The method for adjusting the cognitive boundary is as follows: An elite pool is constructed based on an optimized population, and new solutions are generated by adjusting the boundaries of the elite pool. The new solution is adapted and updated, the elite pool boundary is redefined based on the processing results, and the elite pool is adjusted based on the elite pool boundary. A symbolic dataset is constructed by performing symbolic inverse mapping based on the adjusted elite pool.

8. The cognitive feature mining method for enhancing human-intelligence collaborative mechanisms according to claim 7, characterized in that: The elite pool is constructed as follows: The fitness of the optimized population is obtained through fitness analysis. The population is then sorted in descending order of fitness, and an elite pool is constructed based on the sorting results.

9. The cognitive feature mining method for enhancing human-intelligence collaborative mechanisms according to claim 7, characterized in that: The method for adaptively updating the new solution is as follows: Construct the feature vector corresponding to the new solution and calculate the fitness. Update the individuals in the elite pool based on the fitness. Based on the updated elite pool, obtain the upper and lower bounds for each dimension; By constructing a system of boundary equations, the upper and lower boundaries and the new solution are input into the system of boundary equations. The new solution and the boundary are then judged, and the boundary range is adjusted based on the judgment result.

10. The cognitive feature mining method for enhancing human-intelligence collaborative mechanisms according to claim 1, characterized in that: The method for obtaining the aforementioned human-intelligence collaborative mutual trust level is as follows: Acquire the physiological characteristics of users after they have engaged in dialogue through the LARA model, as well as the baseline data of their physiological characteristics before using the LARA model; Extract the alpha wave power and HRV value from the physiological characteristics of the brain, and calculate the rate of change of the alpha wave power and HRV value relative to the baseline data to obtain the power change rate and HRV change rate; The objective smoothness coefficient is obtained by summing and averaging the power change rate and HRV change rate. Obtain the accuracy of the LARA model in recognizing symbol sequences and interaction intentions, and use it as the model response matching value; Calculate the mean of the objective fluency coefficients of all users and the mean of the model response matching values, and multiply the two means to obtain the human-intelligence collaborative mutual trust degree.

Images