Behavior recognition method based on uncertain incidence relation and related device

Abstract

The invention belongs to the technical field of human behavior recognition, and discloses a behavior recognition method based on an uncertain incidence relation and a related device, and the method comprises the steps: carrying out the entropy-based adaptive fuzzy division of a plurality of information source feature data, and obtaining a feature fuzzy set corresponding to each information source; based on the plurality of information source feature data and the feature fuzzy set corresponding to each information source, evidence association rule learning is carried out, and a candidate rule set is generated; establishing an objective function based on the candidate rule set; solving the objective function to obtain an optimal information source weight value; selecting k pieces of information source feature data with maximum information source weight values from the plurality of pieces of information source feature data to obtain k pieces of information source feature data with higher information source weight values; according to the k pieces of information source feature data with the high information source weight values, a behavior recognition result based on the uncertain incidence relation is obtained; according to the method, the interpretability of an identification result is remarkably improved by quantifying the importance degree of each information source while the identification precision is ensured.

Description

Technical Field

[0001] This invention belongs to the field of human behavior recognition technology, and specifically relates to a behavior recognition method and related device based on uncertain correlation relationships. Background Technology

[0002] In diverse application scenarios such as industrial production, health monitoring, safety inspection, and smart homes, behavior recognition technology is playing an increasingly important role as a key supporting technology for promoting the intelligent development of various fields. Among them, the demand for behavior recognition is particularly urgent in the field of health monitoring, as its technological development is directly related to the real-time monitoring of individual health status and the accurate early warning of abnormal behavior. In recent years, the breakthrough development of deep learning technology has brought revolutionary changes to the field of behavior recognition. Researchers have built a deep learning architecture based on Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) network, and their hybrid models, achieving efficient extraction and pattern recognition of human behavioral features. This not only has significant advantages in conventional behavior recognition scenarios but also represents a key technological breakthrough in the medical and health field.

[0003] As application scenarios become more complex, behavior recognition technology has gradually transitioned from being based on a single sensor to a multi-sensor network to meet the demand for high-precision behavior recognition. Behavior recognition technology based on multi-sensor networks has significantly improved the accuracy and robustness of health activity recognition by integrating sensor information from different parts of the human body or different data modalities. For example, in the task of monitoring falls among the elderly, behavior recognition models based on deep learning can fuse and analyze multi-modal sensor data such as accelerometers and gyroscopes to achieve real-time detection and early warning of fall behavior, providing safety protection for the elderly.

[0004] However, existing deep learning architectures for behavior recognition still suffer from a lack of model interpretability and an exacerbated contradiction between recognition accuracy and computational complexity in practical applications, severely limiting their in-depth application in scenarios with extremely high reliability requirements, such as health monitoring. Specifically, while existing deep learning architectures can establish high-precision mappings between sensor data and behavior labels, the decision-making process is a "black box," and this lack of interpretability makes it difficult for researchers to trace the reliability basis of the model's classification predictions and to analyze the representation mechanisms of different behavioral features within the model. This problem is particularly prominent in multi-sensor deployment scenarios. Furthermore, as the number of sensors increases, the model's input dimension grows exponentially, and the contradiction between computational complexity and recognition accuracy becomes increasingly prominent, becoming a key bottleneck restricting the large-scale application of behavior recognition technology in high-reliability fields such as healthcare. Therefore, there is an urgent need to develop a behavior recognition technology solution that combines high accuracy, interpretability, and optimized computational efficiency. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a behavior recognition method and related apparatus based on uncertain correlations, in order to solve the technical problems of lack of model interpretability and the exacerbated contradiction between recognition accuracy and computational complexity in existing deep learning architecture behavior recognition technologies.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention provides a behavior recognition method based on uncertain association relationships, comprising: Acquire human behavior data collected by wearable sensors as raw data; extract features from the raw data to establish a feature dataset; divide the feature dataset into training and testing sample sets; separate the feature dataset according to the distribution location and type of wearable sensors to obtain several source feature data. Entropy-based adaptive fuzzy partitioning is performed on several source feature data to obtain the feature fuzzy set corresponding to each source; Based on the feature data of several information sources and the feature fuzzy set corresponding to each information source, evidence association rules are learned to generate a candidate rule set; Based on the candidate rule set, a database coverage strategy is used to generate the final rule set; the training sample set is classified based on the final rule set to generate a confusion matrix; and an objective function is established based on the confusion matrix. The objective function is solved using a genetic algorithm to obtain the optimal source weight values; Based on the optimal source weight value, k source feature data with the largest source weight value are selected from several source feature data, so that the cumulative contribution rate reaches the set contribution rate threshold, and k source feature data with high source weight value are obtained. The decision-level fusion strategy based on Dempster-Shafer identifies the test sample set based on k source feature data with high source weight values, and obtains the behavior recognition result based on uncertain correlation.

[0007] Furthermore, feature extraction is performed on the original data to establish a feature dataset; the feature dataset is then partitioned to obtain a training sample set and a test sample set; the feature dataset is further separated according to the distribution location and type of the wearable sensors to obtain several source feature data, as follows: The original data is preprocessed to obtain the preprocessed original data; The preprocessed raw data is segmented using a sliding window of preset width, and features are extracted and selected from the data within each sliding window to obtain a feature dataset. The feature dataset is divided according to a preset data partitioning ratio to obtain a training sample set and a test sample set. The feature dataset is separated according to the distribution location and type of wearable sensors to obtain several source feature data.

[0008] Furthermore, the process of performing entropy-based adaptive fuzzy partitioning on several source feature data to obtain the feature fuzzy set corresponding to each source is as follows: Entropy-based adaptive discretization searches the boundaries of each source feature data to obtain the discrete interval corresponding to each source feature data. For each source feature data, a trapezoidal fuzzy partitioning method is used to perform fuzzy transformation on the discrete interval corresponding to the source feature data to obtain the feature fuzzy set corresponding to each source.

[0009] Furthermore, based on several source feature data and the corresponding fuzzy feature set for each source, the process of learning evidence association rules and generating a candidate rule set is as follows: The classic association rule mining algorithm is used to calculate the item set support of several source feature data, the feature fuzzy set corresponding to each source and the preset behavior category label, and retain the item sets whose item set support is greater than the predetermined minimum support threshold to obtain frequent item sets. Frequent itemsets are divided into several rules containing premises and conclusions, and the confidence of several rules containing premises and conclusions is calculated; rules with confidence greater than a predetermined minimum confidence threshold are retained to obtain several strongly related rules. Several strongly correlated rules are sorted according to a preset priority to obtain a candidate rule set.

[0010] Furthermore, the objective function is as follows:

[0011]

[0012] in, Let be the objective function, and let represent the distance between the weighted confusion matrix and the ideal confusion matrix; For the first Weight values ​​of each source feature data; For the first Elements in the confusion matrix of source feature data The value; As an intermediate variable; For the first The first source feature data Column features; For the first The true category label of the sample of each source feature data; For elements in the ideal confusion matrix The value; Norm parameter; The number of behavior categories; This represents the quantity of source feature data.

[0013] Furthermore, based on the Dempster-Shafer decision-level fusion strategy, the process of identifying the test sample set using k source feature data with high source weight values ​​to obtain the behavior recognition result based on uncertain correlation is as follows:

[0014]

[0015] in, Let the j-th information source be the category of sample y in the test sample set. The original confidence assignment; Let be the weight value of the j-th information source; For the discounted confidence allocation; Assign confidence levels for uncertainty across all categories after discounting; This is the classic Dempster-Shafer combination rule.

[0016] The present invention also provides a behavior recognition system based on uncertain correlations, comprising: The data acquisition module is used to acquire human behavior data collected by wearable sensors as raw data; extract features from the raw data to establish a feature dataset; divide the feature dataset to obtain a training sample set and a test sample set; and separate the feature dataset according to the distribution location and type of the wearable sensors to obtain several source feature data. The fuzzy partitioning module is used to perform entropy-based adaptive fuzzy partitioning on several source feature data to obtain the feature fuzzy set corresponding to each source. The association rule learning module is used to learn evidence association rules based on several source feature data and the feature fuzzy set corresponding to each source, and generate a candidate rule set. The objective function construction module is used to generate a final rule set based on the candidate rule set using a database coverage strategy; to classify the training sample set based on the final rule set to generate a confusion matrix; and to establish the objective function based on the confusion matrix. The objective function solution module is used to solve the objective function using a genetic algorithm to obtain the optimal source weight values; The source feature selection module is used to select k source feature data with the largest source weight values ​​from a number of source feature data based on the optimal source weight values, so that the cumulative contribution rate reaches a set contribution rate threshold, and obtain k source feature data with high source weight values. The fusion decision module is used to identify the test sample set based on the Dempster-Shafer decision-level fusion strategy. It identifies the test sample set based on k source feature data with high source weight values ​​and obtains the behavior recognition results based on uncertain correlation.

[0017] The present invention also provides an electronic device, comprising: A processor is used to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, performs the behavior recognition method based on uncertain association relationships.

[0018] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the behavior recognition method based on uncertain association relationships.

[0019] The present invention also provides a computer program product, the computer program product comprising a computer program, which, when executed by a processor, implements the behavior recognition method based on uncertain association relationships.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a behavior recognition method based on uncertain association relationships. It employs a hybrid architecture combining association rule mining and genetic algorithms, significantly improving the interpretability of the recognition results while ensuring recognition accuracy by quantifying the importance of each information source. Specifically, firstly, a candidate rule set is constructed for each information source. The final rule set is generated from the candidate rule set, and the confusion matrix for each information source is obtained, thus constructing the objective function. Secondly, the objective function is solved using a genetic algorithm to generate weight parameters for each information source. This not only quantifies the differences in importance between different information sources but also reveals the contribution of each information source to the final decision in a quantifiable way. Based on this, an adaptive information source selection mechanism based on cumulative contribution is adopted, optimizing the number of information sources while ensuring recognition performance, effectively reducing system energy consumption and computational complexity. Finally, the recognition results from the retained sensors are integrated within a theoretical framework, and sensor weights are considered simultaneously to implement decision fusion.

[0021] The behavior recognition system, electronic device, computer-readable storage medium, and computer program product based on uncertain association relationships provided by this invention possess all the advantages of the aforementioned behavior recognition methods based on uncertain association relationships. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart of the behavior recognition method based on uncertain association relationships provided in Example 1; Figure 2 For the two information sources in Example 1 (information source S6 and information source S) 10 A schematic diagram of the feature fuzzy set; Figure 3 This is a graph showing the change in the maximum fitness value during the iteration process of the genetic algorithm in Example 1; Figure 4 The cumulative contribution change in descending order of importance of source feature data in Example 1; Figure 5 The source weight distribution of the UCI mHealth dataset in Example 1; Figure 6 The source weight distribution of the WARD dataset in Example 1; Figure 7 The source weight distribution of the Opportunity dataset in Example 1; Figure 8 The EARL confusion matrix of the UCI mHealth dataset in Example 1; Figure 9 The ImA-EARL confusion matrix of the UCI mHealth dataset in Example 1; Figure 10 The EARL confusion matrix of the WARD dataset in Example 1; Figure 11 The ImA-EARL confusion matrix of the WARD dataset in Example 1; Figure 12 The EARL confusion matrix of the Opportunity dataset in Example 1; Figure 13 The ImA-EARL confusion matrix of the Opportunity dataset in Example 1; Figure 14This study presents a quantitative analysis of ImA-EARL and two other fusion strategies (random and reverse order) source fusion in the UCI mHealth dataset of Example 1. Figure 15 Quantitative analysis of ImA-EARL and two other fusion strategies (random and reverse order) source fusion for the WARD dataset in Example 1; Figure 16 Quantitative analysis of source fusion using ImA-EARL and two other fusion strategies (random and reverse order) on the Opportunity dataset in Example 1; Figure 17 This is a structural block diagram of the behavior recognition system based on uncertain correlation provided in Example 2; Figure 18 This is a structural block diagram of the electronic device provided in Example 3. Detailed Implementation

[0024] To make the technical problems, technical solutions, and beneficial effects solved by this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0025] The following specific embodiments further explain and illustrate the behavior recognition method based on uncertain correlation provided by the present invention: Example 1 As attached Figure 1 As shown, this embodiment 1 provides a behavior recognition method based on uncertain association relationships, including the following steps: Step 1: Acquire human behavior data collected by wearable sensors as raw data; extract features from the raw data to establish a feature dataset; partition the feature dataset to obtain a training sample set and a test sample set; separate the feature dataset according to the distribution location and type of the wearable sensors to obtain several source feature data. Specifically, read the human behavior data collected by wearable sensors as raw data; preprocess the raw data to obtain preprocessed raw data; segment the preprocessed raw data using a sliding window of preset width, and extract and select features from the data within each sliding window to obtain a feature dataset; use cross-validation to divide the feature dataset into a training sample set and a test sample set; according to the distribution location of the wearable sensors (e.g., chest, wrist, or ankle) and the type of wearable sensors (e.g., accelerometer, gyroscope, and magnetometer), separate the wearable sensor data in the feature dataset into independent source feature data to obtain several source feature data; wherein, the distribution location and type of each wearable sensor are defined as a source, and a total of [number] source feature data are obtained. Individual source feature data.

[0026] More specifically, the process is as follows: Step 11: Read the human behavior data collected by the wearable sensor to obtain the raw data.

[0027] Step 12: Using the 3σ principle and the Matlab toolkit for missing data detection, perform data cleaning on the original data to eliminate outliers and missing values, and obtain preprocessed original data.

[0028] Step 13: Using a sliding window of a preset width, segment the preprocessed raw data and extract features from the data in each sliding window to obtain feature-extracted data; preferably, the width of the sliding window of the preset width is 2.56s and the overlap rate of two adjacent sliding windows is 50%.

[0029] Step 14: Select features from the extracted data using the maximum relevance and minimum redundancy method, retaining the subset of features that are most relevant to the behavior category and have the least redundancy among them, to obtain the feature dataset. The maximum relevance and minimum redundancy method can effectively reduce the data dimensionality of the extracted data, thereby improving the training efficiency and recognition accuracy of the behavior recognition model based on evidence association rules, while also enhancing the interpretability of the behavior recognition model based on evidence association rules.

[0030] Step 15: Divide the wearable sensor data in the feature dataset into an 8:2 ratio and evaluate its performance using a 5-fold cross-validation method to obtain a training sample set and a test sample set. The training sample set is used to train the behavior recognition model based on evidence association rules, and the test sample set is used to test the behavior recognition model based on evidence association rules.

[0031] Step 16: Based on the distribution location of the wearable sensors (e.g., chest, wrist, or ankle) and the type of wearable sensors (e.g., accelerometer, gyroscope, and magnetometer), separate the wearable sensor data in the feature dataset into independent source feature data, obtaining several source feature data sets; wherein, the distribution location of each wearable sensor and the type of each wearable sensor are defined as a source, resulting in a total of [number missing] source feature data sets. Individual source feature data.

[0032] It should be noted that the feature dataset contains human behavior data collected by multiple wearable sensors. This data is typically distinguished by columns; for example, columns 1 to 3 contain data from the accelerometer. The data is divided according to the distribution location and type of the wearable sensors. Each wearable sensor, or several wearable sensors, is treated as an independent source of feature data. Each data point in the feature dataset is extracted in column form, thus separating the feature dataset into... Individual source feature data; among which... This indicates the number of source feature data, where each source feature data is a continuous numerical value.

[0033] Step 2: Perform entropy-based adaptive fuzzy partitioning on several source feature data to obtain the feature fuzzy set corresponding to each source. Specifically, based on entropy-based adaptive discretization, the boundary of each source feature data is searched to obtain the discrete interval corresponding to each source feature data; the trapezoidal fuzzy partitioning method is used to perform fuzzy transformation on the discrete interval corresponding to each source feature data to obtain the feature fuzzy set corresponding to each source.

[0034] It is worth noting that human behavior data collected by wearable sensors is typically represented as time-series data. By using entropy-based adaptive discretization, the boundary points of each source feature data are adaptively searched, transforming continuous source feature data into discrete intervals, which are then used to represent human behavior features. Subsequently, a gradient fuzzy partitioning method is employed to transform the discrete intervals into fuzzy intervals, obtaining the feature fuzzy set corresponding to each source. Each feature fuzzy set contains the fuzzy interval corresponding to each source feature data and its membership function. This entropy-based adaptive fuzzy partitioning method can effectively address the uncertainty of human behavior data collected by wearable sensors, providing a foundation for subsequent evidence association rule learning.

[0035] Specifically, the steps are as follows: Step 21: The entropy-based adaptive discretization method uses the minimum description length criterion (MDLP) to search for the boundary points of each source feature data to obtain the discrete interval corresponding to each source feature data. Using the entropy-based adaptive discretization method, the optimal bisection boundary is recursively selected based on the entropy minimization principle to search for the feature interval with the minimum entropy value between human behavior features and categories as much as possible. Among them, the discrete interval corresponding to each source feature data constitutes a hard partition of each source feature data. Hard partition means that each feature value in each source feature data belongs to only one discrete interval, that is, the membership degree is 0 or 1.

[0036] Step 22: Using the gradient fuzzy partitioning method, the discrete intervals corresponding to the feature data of each information source are fuzzy transformed to obtain the feature fuzzy set corresponding to each information source. Specifically, the feature fuzzy set corresponding to each information source (such as "VeryLow", "Low", "Medium", "High") is defined by the trapezoidal membership function, which describes the degree to which continuous features belong to the corresponding fuzzy interval, and the value range is [0,1]. Among them, a feature value can belong to multiple adjacent fuzzy intervals at the same time, but the membership degree is different, which can effectively handle the uncertainty of human behavior data collected by wearable sensors.

[0037] As attached Figure 2 As shown, attached Figure 2 The document provides two information sources (S6 and S). 10 A schematic diagram of the feature fuzzy set; from the attached... Figure 2As can be seen, by using a series of overlapping triangular fuzzy sets, continuous feature values ​​(horizontal coordinates) are mapped to discrete linguistic variables (such as "Very Low", "Low", "Lower medium", "Higher medium", "High", "Very High"). The feature fuzzy set corresponding to each source represents a specific linguistic value and covers a certain range of feature values. Through fuzzy transformation, the uncertainty and fuzziness of sensor data can be processed, and numerical data can be transformed into symbolic representations that can be used for rule-based reasoning. For example, for source S6, the feature value is -3.62, which belongs to both the "Very Low" and "Low" fuzzy sets, but with different membership degrees, thus providing a foundation for subsequent evidence association rule learning.

[0038] Step 3: Based on several source feature data and the corresponding feature fuzzy set for each source, evidence association rules are learned to generate a candidate rule set. Specifically, a classic association rule mining algorithm is used to calculate the itemset support of several source feature data, the feature fuzzy set corresponding to each source, and the preset behavior category label. Itemsets with support greater than a predetermined minimum support threshold are retained to obtain frequent itemsets. Frequent itemsets are then divided into several rules containing premises and conclusions, and the confidence of these rules is calculated. Rules with confidence greater than a predetermined minimum confidence threshold are retained to obtain several strongly associated rules. These strongly associated rules are then sorted according to a preset priority to obtain the candidate rule set.

[0039] It should be noted that, based on the source feature data obtained in step 1 and the feature fuzzy set corresponding to each source obtained in step 2, the association rule learning method is used to mine the association relationship between the behavior category labels in the preset training set and the source feature data and the fuzzy features in the feature fuzzy set. The behavior category labels in the preset training set are obtained through pre-labeling and represent the true category of human behavior (e.g., standing still or walking). The fuzzy features refer to the features after fuzzy transformation in step 2. Each fuzzy feature represents its fuzzy interval (e.g., "Very Low", "Low", "Lower medium", "Higher medium", "High", "Very High") and its corresponding membership degree. Using a classic association rule mining algorithm, itemsets with support greater than the minimum support are retained as frequent itemsets, and a reliable candidate rule set is generated based on these frequent itemsets to establish a human behavior recognition model. The rule-based classification method identifies samples of unknown categories by extracting rules from the data. Each rule in the generated candidate rule set is a refined knowledge module, thus possessing good interpretability.

[0040] Specifically, the steps are as follows: Step 31: Using the classic association rule mining algorithm, calculate the item set support of several source feature data, the feature fuzzy set corresponding to each source, and the preset behavior category label. Keep the item sets whose support is greater than the predetermined minimum support threshold to obtain frequent item sets. Here, an item set refers to the combination relationship between source feature data, fuzzy features, and preset behavior category label, such as: {X1=Very Low, X2=High, Category=Walking}. The support represents the frequency of occurrence of the corresponding item set.

[0041] Step 32: Divide each frequent itemset into an antecedent and a consequent to obtain several rules containing an antecedent and a consequent. The antecedent contains feature-fuzzy interval pairs, such as "X1 is Very Low and X2 is High". The consequent is the behavior category, such as "the behavior is walking". This forms candidate rules in the form of "antecedent → consequent".

[0042] Step 33: Calculate the confidence scores of several rules that include premises and conclusions; retain the rules with confidence scores greater than a predetermined minimum confidence score threshold to obtain several strongly correlated rules; sort the several strongly correlated rules according to a preset priority to obtain a candidate rule set.

[0043] It should be noted that the number of rules generated during the process of generating rules containing premises and conclusions is relatively large. The main reason for this is that a certain number of rules are redundant and useless. Therefore, before selecting strongly related rules, existing rules need to be reduced, and redundant and useless rules should be filtered out. Specifically, by comparing the premise and conclusion parts of the rules containing premises and conclusions, rules with identical conclusions but more specific premises are deleted. The rules containing premises and conclusions are arranged from low to high according to the number of premises, and each rule is compared pairwise from top to bottom. Assume there exists a rule R1 (… ) and rule R2 ( As can be seen, the conclusions of R1 and R2 are of the same category, while the premises... Compared to More specifically, R2 is deleted while R1 is retained, resulting in a reduced rule set. All association rules are then sorted according to priority. Specifically, firstly, they are sorted in descending order of confidence, where higher confidence indicates a more reliable rule. Secondly, if the confidence levels are the same, they are sorted in descending order of support, where higher support indicates a rule covers more samples. Finally, if the support levels are still the same, they are sorted in ascending order of the number of prerequisite features, where fewer features indicate a simpler rule. After sorting the rules, a candidate rule set is obtained.

[0044] The confidence score of a rule containing premises and a conclusion represents the conditional probability that the conclusion holds given that the premises are met; rules with confidence scores greater than the minimum confidence threshold Minconf are retained; in fuzzy partitioning scenarios, support and confidence scores are calculated by weighting the corresponding features using their fuzzy membership functions; the calculated confidence score is used as the rule weight β. j The reliability of the rules is described as follows:

[0045]

[0046] in, For the sample Regarding the premise of the rule The matching degree is calculated using the fuzzy membership function; Let j be the set of samples of category j; Let A be the total number of samples in the training set; T is the set of all samples that satisfy premise A, i.e., the set of samples with a matching degree greater than zero.

[0047] It should also be noted that the model is constructed by learning association rules based on human behavior data; the association relationship between features and behavior categories is learned from the training data through association rule mining algorithms, generating a candidate rule set for each information source; the candidate rule set contains several association rules in the form of "if the features meet certain conditions, then the behavior belongs to a certain category"; in the process of building the model, two key parameters need to be determined first, such as the preset minimum support threshold (defined as Minsup) and the preset minimum confidence threshold (defined as Minconf); preferably, , .

[0048] Secondly, to achieve the interpretability of the recognition model, the internal definitions of the computer are displayed in a human-understandable rule form; that is, for a classification problem with M classes and P features, it is assumed that the j-th association rule in the rule base has k prerequisite features; its basic structure is represented as: Rule R j If X1 is AF1, j1 And...and X p It is AF p , jp The conclusion is ω j The rule weight is β j ;in: , representing the data feature set, which consists of various feature variables (such as chest acceleration X-axis, ankle gyroscope Y-axis, etc.) in the source feature data obtained after feature selection in step 1. , represents the premise fuzzy set, that is, the set of fuzzy intervals corresponding to each feature; Corresponding features A specific fuzzy division, such as the "Very Low" range of feature X1; ω j The conclusion of the rule, i.e., the predicted behavior category, such as standing still, walking, etc.; β j is the rule weight, used to characterize the reliability of the rule, and is measured by confidence; k is the rule R j The number of features contained in the premise.

[0049] Step 4: Generate the final rule set using a database coverage strategy based on the candidate rule set; generate a confusion matrix based on the final rule set; and establish an objective function based on the confusion matrix. Specifically, for the candidate rule set obtained in step 3, a database coverage strategy is used to generate the final rule set. The final rule set is then used to classify the training samples in the training sample set, identifying the categories of the training samples and generating confusion matrices for each source, thus providing a performance evaluation basis for constructing the objective function. Since in fuzzy partitioning, the feature value of a training sample in the training sample set may simultaneously belong to multiple fuzzy intervals, it is necessary to activate rules by calculating the matching degree between the feature value of the training sample and the feature fuzzy set. Specifically, when the matching degree is greater than zero, rules in the final rule set that match the features of the training sample are activated. When the categories of the activated rules are inconsistent, the uncertain label information in all activated rules is fused using the Dempster-Shafer combination rule to determine the final predicted category of the training sample. Through the above rule reasoning process, the final rule sets of each source are used to identify the training data, and a confusion matrix corresponding to the identification results is established. Finally, an optimization objective function is established based on the obtained confusion matrix. This objective function is used to evaluate the merits of different source weight combinations and serves as the target for optimization by the genetic algorithm in step 5.

[0050] Specifically, the process is as follows: Step 41: Calculate the matching degree between training samples and rules. Specifically, for each training sample in the training sample set, calculate its matching degree with each rule in the candidate rule set. Unlike hard partitioning, in fuzzy partitioning, the feature points of a sample and the corresponding partition interval are not necessarily a perfect match (i.e., the membership degree is 1). Therefore, it is necessary to calculate the matching degree between the rules in the candidate rule set and the training samples (defined as the membership function between the matched features and the corresponding interval) using the membership function of the matched features and the corresponding interval. Specifically, calculate the p-th feature value of the i-th sample. On the rules The p-th fuzzy interval in the premise membership degree Then, calculate the sample according to formula (4-1). Regarding the rules Overall matching degree.

[0051] The matching degree calculation process is as follows:

[0052] Where k is the rule The number of features included in the premise; this matching degree is used as the weight of the activation rule in subsequent step 43, participating in the Dempster's rule combination fusion.

[0053] Step 42: Calculate the class confidence assignment of the rule. Specifically, to determine the weight of the rule, it is necessary to calculate the degree of confidence of the sample in the uncertain class of the rule, defined as the mass function. The calculation of overheating in terms of the degree of trust of the sample in the uncertain category in the rule is as follows:

[0054] in, The rules that are activated; The rule weights are represented by the rule confidence scores. These rule weights are used to discount the confidence assignment of information sources in subsequent decision fusion. For rules Category Confidence assignment; Assign confidence levels for uncertainty across all categories; This is the complete set of categories.

[0055] Step 43: Classify samples through evidence combination. Specifically, using Dempster's rule from evidence reasoning theory, the mass function generated by fusing all activated rules is combined. By combining the uncertain category information from all activated rules, the final label of the sample is determined. Through the fusion process, the final confidence assignment of sample x to each category is obtained. The predicted class of the sample is the class with the highest confidence assignment value; the fusion process is as follows:

[0056] in, R represents the classic Dempster-Shafer combination rules; R is the set of rules that are activated.

[0057] Step 44: Generate a confusion matrix. Specifically, select the rule set with the best recognition performance for the training samples as the human behavior recognition model under the current information source; here, the best rule set refers to the final rule set generated based on the candidate rule set generated in step 3 through a database coverage strategy, which is the candidate rule set that can optimize the subsequent objective function, and its performance is reflected in the classification result of the training samples being closest to the ideal state.

[0058] The training samples are classified using a recognition model, and the corresponding confusion matrix is ​​calculated. ;in, The number of behavior categories, the elements in the matrix This represents the number of samples with true category i that are classified as category j. Specifically, for each sample in the training sample set, the recognition model under the current information source is used for classification, and the classification result is compared with the true label of the sample to count the recognition status of each category. The confusion matrix is ​​calculated for all information sources according to the above method. In an ideal case, all activities can be correctly identified by using the generated rule base, thus obtaining the following ideal confusion matrix.

[0059]

[0060] in, This is the ideal confusion matrix.

[0061] Step 45: Establish the objective function for optimization. Specifically, assume variables... Let represent the weight value assigned to the s-th sensor (source) in the wearable human activity monitoring system under study; to measure the recognition strength of the weighted multi-sensor human behavior monitoring system, an S-tuple multivariate mapping function is designed using the definition of the matrix p-norm; based on the obtained confusion matrices of each source, the following objective function is established:

[0062]

[0063] in, Let be the objective function, and let represent the distance between the weighted confusion matrix and the ideal confusion matrix; For the first Weight values ​​of each source feature data; For the first Elements in the confusion matrix of source feature data The value; As an intermediate variable; For the first The first source feature data One sample; For the first The true category label of the sample of each source feature data; For elements in the ideal confusion matrix The value; Norm parameter; The number of behavior categories; This represents the number of source feature data.

[0064] Therefore, the source weight learning task is transformed into how to solve a constrained optimization problem with the following objective function, namely, Min( ; where each weight value It meets the following conditions:

[0065] That is, all source weights are positive and normalized, and the sum of the weights is 1.

[0066] Step 5: Solve the objective function to obtain the optimal source weight values. Specifically, based on the genetic algorithm, solve the objective function to obtain the optimal source weight values. To find the minimum value of the objective function established in Step 4, a set of optimal source weight values ​​is obtained. The optimal source weight values ​​quantify the contribution of each source to behavior recognition. In this embodiment 1, while ensuring performance, the input of sources is reduced to reduce model complexity and improve interpretability under multiple sources. To screen the sources, a genetic algorithm is used to determine the optimal solution by simulating the natural evolution process. The optimization is continuously iterated through population initialization, chromosome evaluation, selection, crossover, and mutation to generate weight values ​​for each source.

[0067] Specifically, the process is as follows: Step 51: Based on the objective function constructed in Step 4, design a genetic algorithm. Specifically, binary encoding is used for gene pool initialization. The purpose of gene pool initialization is to randomly generate an initial population (set to 100) for the genetic algorithm. Each individual (chromosome) is a possible solution. Then, the fitness function is designed. In the genetic algorithm, the fitness value is used to evaluate the individual's survival and reproductive ability in a specific environment and to decide whether to retain the individual. The fitness value is calculated by the fitness function. For a dataset with S information sources, the fitness function is as follows:

[0068] in, This is the fitness function.

[0069] In addition, a classic "death" strategy is introduced to handle convex constraints, in which any candidate solution that violates the constraints is reset to negative infinity (or zero) and thus excluded, ensuring that only feasible solutions can participate in selection, crossover and mutation, thereby maintaining the integrity of the optimization process.

[0070] Step 52: The population selection process employs a roulette wheel selection mechanism based on fitness values. The selection probability is proportional to the individual's fitness, ensuring that high-quality chromosomes have a higher probability of inheritance. After the selection operation, a new generation of the population is generated through single-point crossover (crossover rate set to 0.6) and mutation (mutation rate set to 1 / L, where L is the chromosome length) operators. During the single-point crossover process, a breakpoint on the parent chromosome is randomly selected, and the gene sequences of the corresponding segments are exchanged, thereby achieving the recombination of genetic information. After multiple generations of iteration, the maximum fitness value in the population gradually converges, ultimately yielding a set of optimal source weights {W1, ..., W...}. s} is the optimal source weight value.

[0071] As attached Figure 3 As shown, attached Figure 3 The figure shows the change in the maximum fitness value during the iteration process of the genetic algorithm; where the horizontal axis represents the number of iterations of the genetic algorithm, and the vertical axis represents the maximum fitness value in the population; from the attached figure... Figure 3 As can be seen, with the increase of the number of iterations (from 0 to about 120), the maximum fitness value shows a step-like increase and gradually tends to stabilize, indicating that the genetic algorithm can effectively search the solution space and gradually converge to the optimal or suboptimal solution. The stability of the fitness value indicates that the algorithm has found a relatively ideal combination of information source weights, which can maximize the objective function (such as recognition accuracy), thus verifying the effectiveness and convergence of using the genetic algorithm to learn the importance weights of information sources.

[0072] Step 6: Based on the optimal source weight values, select k source feature data with the largest source weight values ​​from several source feature data, so that the cumulative contribution rate reaches the set contribution rate threshold, thus obtaining k source feature data with high source weight values. It should be noted that in order to effectively select important sources, a weight threshold needs to be set first; this process is based on the optimal source weight values ​​obtained in Step 5; by calculating the cumulative contribution of the optimal source weight values, a minimum source subset can be determined; this source subset optimizes the number of sources while ensuring recognition performance; the purpose of using a preset threshold is to select sources with higher importance, thereby ensuring that the selected sources can provide valuable information in subsequent analysis; in this embodiment 1, a method based on cumulative contribution is adopted, which, compared with the traditional hard threshold setting method, can adapt to changes in the number of sources in different datasets.

[0073] Specifically, the process is as follows: Step 61: To further evaluate the importance of the sensor, the genetic algorithm from Step 5 is first used to optimize the objective function and obtain the optimal weight values. Sort all the derived optimal weight values ​​in descending order; the sorting result of the optimal weight values ​​is as follows:

[0074] in, Let (i) be the source weight value at the i-th position after sorting in descending order of weight, where i is the index of the renumbered source. Here, (i) represents the sorted position, not the original source number s.

[0075] Step 62: Based on the ranking results of the most weighted values, the cumulative contribution of the source feature data with the largest weight values ​​among the top k sources can be measured by their proportion relative to the overall weight distribution; the calculation process is as follows:

[0076] in, It represents the cumulative contribution of the source feature data with the largest weight value among the top k sources.

[0077] Step 63: Select the first time the cumulative contribution rate reaches the preset threshold. (Default value is 0.8) The source feature data with the largest source weight value, that is, satisfying... Theoretically, increasing the preset threshold It can improve recognition accuracy, but it increases computational complexity; while lowering the preset threshold... While it saves computation time, it sacrifices accuracy; in practical applications, user preferences must be considered when determining the cumulative threshold, as shown in the attached figure. Figure 4 As shown.

[0078] Step 7: Based on the Dempster-Shafer decision-level fusion strategy, the behavior recognition result based on uncertain association is obtained according to the feature data of k source sources with high source weight values. To improve the accuracy and robustness of the HAR system, within the framework of evidence theory, the recognition results of the retained sources are fused using a Dempster-Shafer decision-level fusion strategy by simultaneously considering sensor weights. Specifically, for the test sample set, the candidate rule set of each retained source (as the rule base of that source) is used to independently identify and obtain the initial confidence assignment of each source. Then, the confidence assignment of each source is discounted according to the source weights obtained in Step 6, so that the source with higher weight has a greater impact on the final decision. Finally, the discounted confidence assignments of all retained sources are fused by Dempster-Shafer combination rules to obtain the final confidence assignment, and the category with the highest confidence value is selected as the behavior recognition result based on uncertain association.

[0079] Specifically, the process is as follows: Step 71: Establish an evidence association rule base to preserve information sources. Specifically, firstly, through the confidence rule reasoning process, independently utilize the source feature data from the k sources with higher source weight values ​​determined in Step 6, i.e. For each of the reserved information sources with higher weight values, the feature data and fuzzy intervals of each source are independently identified using the candidate rule set generated for each reserved source in step 3 (used here as the rule base for that source). This rule base contains several rules in the form of "if the features meet certain conditions, then the behavior belongs to a certain category", and each rule is accompanied by a confidence level (rule weight β). j This is used for classification reasoning on new samples; the generated evidence association rule base is used to identify the newly acquired, unidentified input sample y of human behavior sensor data. For input sample y, the matching degree between its features and the premises of each rule is calculated, rules with a matching degree greater than zero are activated, and the confidence assignments of all activated rules are fused using the evidence combination method in step 4 to obtain the category confidence assignment of the information source for sample y. We obtain the class confidence assignment of the s-th preserved information source to sample y. ;in .

[0080] Step 72: Introduce source weights for confidence discounting. Specifically, assume... Indicates from the previous A class confidence assignment derived from each sensor; wherein, the focus element { } and Ω; to construct an accurate and compact wearable behavior recognition system, each derived category confidence assignment needs to be based on its contribution (i.e., the weight value W obtained in step 6). s By applying a discount, the weight values ​​of the information sources are incorporated into the confidence assignment, so that the higher the weight of the information source, the greater its influence on the final decision, thus realizing the quantification of the importance of the information sources. Through the discount mechanism, information sources with higher weights contribute more to the final decision, while information sources with lower weights have a smaller impact, thus realizing weighted fusion based on the importance of information sources.

[0081] The formula for calculating the discounted confidence allocation is as follows:

[0082] in, Let j be the category of the j-th source pair of sample y. The original confidence assignment; The weight value of the j-th information source (obtained by the genetic algorithm in step 5); The discounted confidence assignment reflects the impact of source weights on decision-making; Assign confidence levels for uncertainty across all categories after discounting.

[0083] Step 73: Decision fusion using the Dempster-Shafer combination rule; specifically, within the evidence framework, through integration of pre-dempster and post-dempster rules... The discounted quality function corresponding to each sensor is used to complete the following decision-level fusion process:

[0084] in, The Dempster-Shafer combination rule applies; therefore, the identification result corresponds to the class focal element with the highest quality value. Especially when the highest focal element is a composite element, the quality m(·|y) should be converted into a Pignistic probability associated with each single class, and then the final identification result is obtained using the highest class probability value.

[0085] Performance verification experiment: (1) Experimental data; The UCI mHealth public dataset was used, which mainly contains the physical movements and vital signs of ten subjects during twelve physical activities, including standing, sitting and relaxing, lying down, walking, climbing stairs, lumbar flexion, arm extension, knee bending, cycling, jogging, running, and jumping forward and backward (represented in this dataset as...). In this dataset, three triaxial accelerometers (connected to the chest, right wrist, and left ankle, respectively), two triaxial gyroscopes and magnetometers (connected to the right wrist and left ankle, respectively), and one dual-lead electrocardiogram sensor (connected to the chest) recorded sensory signals at a sampling rate of 50 Hz. The dataset contains eight signal sources, as shown in Table 1. Before training the recognition model, the data needs to be preprocessed, including feature extraction and feature selection. The collected raw data is time-series data. For this type of data, preprocessing is required. A fixed-width (2.56 seconds) sliding window (50% coverage) technique is used to extract time-domain and frequency-domain features. After feature extraction, the maximum correlation minimum redundancy method is used to select appropriate features. Finally, the recognition model is trained using Example 1 of this paper. After experimental evaluation, the recognition model is imported into the corresponding module.

[0086] Table 1. Distribution of sources in the mHaelth dataset

[0087] (2) Experimental evaluation indicators; The classification evaluation indicators used in Example 1 include classification accuracy, classification precision, classification recall, and the harmonic mean that considers both precision and recall. These can be expressed by the following formula:

[0088]

[0089]

[0090]

[0091] Wherein, TP, FP, TN and FN represent positive samples being identified as positive samples, negative samples being incorrectly identified as positive samples, negative samples being identified as negative samples and negative samples being incorrectly identified as positive samples, respectively.

[0092] (3) Experimental Results; Based on the experimental results, a visualization analysis of the source weight distribution of the multi-source dataset was conducted; see attached. Figure 5 As shown, attached Figure 5 The source weight distribution of the UCI mHealth dataset is given in the appendix; Figure 5 As can be seen, the weights of the eight sources in the UCI mHealth dataset are intuitively represented by color gradients (color depth is positively correlated with weight value); three sources in this dataset were filtered out due to low contribution, ultimately retaining five sources; furthermore, to systematically evaluate the performance advantages of the proposed importance-aware recognition model based on evidence association rules (ImA-EARL), a comparative experiment was designed to compare its performance with the baseline model (EARL) that does not incorporate a sensor importance-aware mechanism; see attached. Figure 6-7 As shown, attached Figure 6 The source weight distribution of the WARD dataset is given in the appendix. Figure 7 The source weight distribution of the Opportunity dataset is given in the figure.

[0093] As attached Figure 8-13 As shown, attached Figure 8 The image shows the EARL confusion matrix for the UCI mHealth dataset. Figure 9 The ImA-EARL confusion matrix for the UCI mHealth dataset is given below. Figure 10 The EARL confusion matrix for the WARD dataset is given below. Figure 11 The ImA-EARL confusion matrix for the WARD dataset is given below. Figure 12 The text provides the EARL confusion matrix for the Opportunity dataset, along with related information. Figure 13 The appendix presents the ImA-EARL confusion matrix for the Opportunity dataset; where (a) represents the EARL method and (b) represents the ImA-EARL method; from the appendix... Figure 8-13 As can be seen, compared with the EARL method, even with a small number of sources, the ImA-EARL method performs better in activity categories ω5 and ω. 12 It also has a better recognition rate.

[0094] (4) Classification performance comparison experiment; In order to verify the advantages of the behavior recognition method proposed in Example 1 compared with existing methods, the method was compared with the baseline method and other advanced methods under the same experimental conditions; the comparison results are shown in Table 2.

[0095] Table 2 Comparison Results of Classification Performance

[0096] As can be seen from Table 2 above, the behavior recognition method described in Embodiment 1 shows significant improvement in all four indicators compared to other rule-based and neural network-based recognition methods, indicating that the method described in Embodiment 1 has a good effect on human behavior recognition. Meanwhile, Table 3 below compares the test time, number of rules (#rule count), and number of sources (#source count) of the recognition model without source selection (EARL) and ImA-EARL. On the mHealth dataset, ImA-EARL's test time is shorter than EARL's, decreasing from 8.31 seconds to 4.94 seconds, the number of generated rules decreases from 1340 to 847, and the number of data sources used decreases from 8 to 5.

[0097] Table 3 Comparison of test time, number of rules, and number of sources between EARL and ImA-EARL

[0098] To illustrate the effectiveness of the method described in Example 1 in quantifying sensor importance, the accuracy of the identification of three datasets was analyzed as a function of the number of fused sensors. Furthermore, the accuracy changes were recorded when two other specific sensor fusion strategies were employed based on sensor importance (or weight values): a reverse-order sensor fusion strategy and a random-sequence sensor fusion strategy (average values ​​from Monte Carlo simulations). See attached... Figure 4-5 As shown, the accuracy trends across the three datasets are largely consistent. For the method described in Example 1 (ImA-EARL), retaining sensors with higher weight values ​​enables the best recognition accuracy across all datasets, with the number of sensors being 5, 6, and 6, respectively. From a complexity perspective, our scheme achieves a trade-off between accuracy and complexity by reducing the number of sensors fused in the recognition process. Furthermore, for each number of sensors, the method described in Example 1 demonstrates better recognition accuracy than the other two fusion strategies, and the improvement is more significant when the number of sensors is small. This intuitively shows that the learned sensor importance distribution is feasible and effective for practical recognition tasks.

[0099] To verify the rationality of the method described in Example 1 for assessing sensor importance, the method performance was tested by sequentially increasing the number of sensor sources from the most important to the least important. The test results are attached. Figure 14-16 As shown; from the appendix Figure 14-16 As can be seen, performance improves with the increase in the number of input sensor signals; in some cases, fewer input sources can achieve better performance than using all sources, indicating that selecting input sources in HAR helps improve recognition performance; when selecting input sources from the least important sensor sources, recognition performance is significantly lower than in other cases; furthermore, by considering the attached... Figure 9 The analysis of the results shows that, on the one hand, when the number of fusion sources is small, fusion starting from important sources is significantly better than fusion starting from unimportant sources, and both show a stable upward trend overall. This comparison confirms the rationality of the method and source selection described in Example 1. On the other hand, when a certain number of signal sources are fused, the change in accuracy tends to stabilize or even decreases slightly. It can be said that using appropriate input signal sources can improve recognition performance, while irrelevant input sources will hinder recognition performance. In summary, it can be said that using an appropriate subset of input sources can improve recognition performance, while irrelevant input sources will hinder recognition performance, which also proves the rationality and effectiveness of the proposed source selection method.

[0100] Example 2 As attached Figure 17 As shown, a behavior recognition system based on uncertain association relationships includes: a data acquisition module, a fuzzy segmentation module, an association rule learning module, an objective function construction module, an objective function solving module, a source feature selection module, and a fusion decision module.

[0101] The system comprises the following modules: a data acquisition module, which acquires human behavior data collected by wearable sensors as raw data; a feature extraction module, which separates the raw data according to the distribution of the wearable sensors to obtain several source feature data; a fuzzy partitioning module, which performs entropy-based adaptive fuzzy partitioning on the source feature data to obtain a feature fuzzy set for each source; an association rule learning module, which learns evidence association rules based on the source feature data and the feature fuzzy set for each source to generate a candidate rule set; an objective function construction module, which establishes an objective function based on the candidate rule set; an objective function solving module, which solves the objective function to obtain the optimal source weight value; a source feature selection module, which selects k source feature data with the largest source weight values ​​from the source feature data based on the optimal source weight values, such that the cumulative contribution rate reaches a set contribution rate threshold, to obtain k source feature data with high source weight values; and a fusion decision module, which uses a Dempster-Shafer decision-level fusion strategy to obtain the behavior recognition result based on uncertain association relationships from the k source feature data with high source weight values.

[0102] Example 3 As attached Figure 18 As shown, this embodiment 3 provides an electronic device, including: a memory for storing a computer program; a processor for executing the computer program to implement the steps of a static trusted verification function test method for a trusted controller; or, the processor for executing the computer program to implement the functions of each module in the above-mentioned static trusted verification function test system for a trusted controller.

[0103] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a preset function, the instruction segments describing the execution process of the computer program in the electronic device.

[0104] The electronic device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above are examples of electronic devices and do not constitute a limitation on the electronic device. It may include more components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0105] The processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor, or any conventional processor, etc. The processor is the control center of the electronic device, connecting various parts of the entire electronic device through various interfaces and lines.

[0106] The memory can be used to store the computer program and / or module. The processor implements various functions of the electronic device by running or executing the computer program and / or module stored in the memory and by calling the data stored in the memory.

[0107] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function (such as sound playback, image playback, etc.). The data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards, secure digital cards, flash memory cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0108] Example 4 This embodiment 4 also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the static trusted verification function test method for a trusted controller.

[0109] If the modules / units integrated in the static trusted verification function test system for the trusted controller are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0110] Based on this understanding, the present invention can implement all or part of the processes in the above-described static trusted verification functional testing method for a trusted controller, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above-described static trusted verification functional testing method for a trusted controller. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or a preset intermediate form, etc.

[0111] The computer-readable storage medium may include any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory, random access memory, electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0112] Example 5 This embodiment 5 provides a computer product, which includes a computer program stored in a computer-readable storage medium. The processor of the electronic device reads the computer program from the computer-readable storage medium and executes the computer program, so that the electronic device can perform the static trusted verification function test method for trusted controllers described in embodiment 1, which will not be repeated here.

[0113] It should be noted that those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above methods.

[0114] The behavior recognition method based on uncertain correlation described in this invention acquires data collected by wearable sensors and separates it to obtain source feature data. Then, it performs adaptive fuzzy partitioning to generate a feature fuzzy set, learns evidence association rules to generate a candidate rule set, and uses a genetic algorithm to solve the objective function to obtain the optimal source weight value. This allows for the selection of source feature data with higher source weight values. Finally, based on a specific fusion strategy, the behavior recognition result is obtained. This method effectively solves the problem of lack of model interpretability in existing deep learning architecture behavior recognition technologies, enabling the tracing of the reliability basis of model classification prediction and the analysis of the representation mechanism of different behavioral features within the model. Simultaneously, by reasonably selecting source feature data, it alleviates the contradiction between recognition accuracy and computational complexity, optimizing computational efficiency while ensuring high recognition accuracy. This provides strong support for the large-scale application of behavior recognition technology in scenarios with extremely high reliability requirements, such as health monitoring.

[0115] The above embodiments are merely one of the implementation methods for achieving the technical solution of the present invention. The scope of protection claimed by the present invention is not limited to this embodiment, but also includes any variations, substitutions and other implementation methods that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention.

Claims

1. A behavior recognition method based on uncertain correlations, characterized in that, include: Acquire human behavior data collected by wearable sensors as raw data; Feature extraction is performed on the original data to establish... Feature dataset; The feature dataset is divided into training and testing sample sets; the feature dataset is further separated according to the distribution location and type of wearable sensors to obtain several source feature data. Entropy-based adaptive fuzzy partitioning is performed on several source feature data to obtain the feature fuzzy set corresponding to each source; Based on the feature data of several information sources and the feature fuzzy set corresponding to each information source, evidence association rules are learned to generate a candidate rule set; Based on the candidate rule set, a database coverage strategy is used to generate the final rule set; A confusion matrix is ​​generated by classifying the training sample set based on the final rule set; Based on the confusion matrix, establish the objective function; The objective function is solved using a genetic algorithm to obtain the optimal source weight values; Based on the optimal source weight value, k source feature data with the largest source weight value are selected from several source feature data, so that the cumulative contribution rate reaches the set contribution rate threshold, and k source feature data with high source weight value are obtained. The decision-level fusion strategy based on Dempster-Shafer identifies the test sample set based on k source feature data with high source weight values, and obtains the behavior recognition result based on uncertain correlation.

2. The behavior recognition method based on uncertain correlation according to claim 1, characterized in that, Extract features from the raw data to create a feature dataset; The process of partitioning the feature dataset to obtain training and testing sample sets, and then further separating the feature dataset according to the distribution location and type of wearable sensors to obtain several source feature data is as follows: The original data is preprocessed to obtain the preprocessed original data; The preprocessed raw data is segmented using a sliding window of preset width, and features are extracted and selected from the data within each sliding window to obtain a feature dataset. The feature dataset is divided according to a preset data partitioning ratio to obtain a training sample set and a test sample set. The feature dataset is separated according to the distribution location and type of wearable sensors to obtain several source feature data.

3. The behavior recognition method based on uncertain correlation according to claim 1, characterized in that, The process of performing entropy-based adaptive fuzzy partitioning on several source feature data to obtain the feature fuzzy set corresponding to each source is as follows: Entropy-based adaptive discretization searches the boundaries of each source feature data to obtain the discrete interval corresponding to each source feature data. For each source feature data, a trapezoidal fuzzy partitioning method is used to perform fuzzy transformation on the discrete interval corresponding to the source feature data to obtain the feature fuzzy set corresponding to each source.

4. The behavior recognition method based on uncertain correlation according to claim 1, characterized in that, The process of learning evidence association rules and generating a candidate rule set based on several source feature data and the corresponding fuzzy feature set for each source is as follows: The classic association rule mining algorithm is used to calculate the item set support of several source feature data, the feature fuzzy set corresponding to each source and the preset behavior category label, and retain the item sets whose item set support is greater than the predetermined minimum support threshold to obtain frequent item sets. Frequent itemsets are divided into several rules containing premises and conclusions, and the confidence of several rules containing premises and conclusions is calculated; rules with confidence greater than a predetermined minimum confidence threshold are retained to obtain several strongly related rules. Several strongly correlated rules are sorted according to a preset priority to obtain a candidate rule set.

5. The behavior recognition method based on uncertain correlation according to claim 1, characterized in that, The objective function is as follows: in, Let be the objective function, and let represent the distance between the weighted confusion matrix and the ideal confusion matrix; For the first Weight values ​​of each source feature data; For the first Elements in the confusion matrix of source feature data The value; As an intermediate variable; For the first The first source feature data Column features; For the first The true category label of the sample of each source feature data; For elements in the ideal confusion matrix The value; Norm parameter; The number of behavior categories; This represents the number of source feature data.

6. The behavior recognition method based on uncertain correlation according to claim 1, characterized in that, The decision-level fusion strategy based on Dempster-Shafer identifies the test sample set based on k source feature data with high source weight values, and obtains the result of behavior recognition based on uncertain correlations, as follows: in, Let the j-th information source be the category of sample y in the test sample set. The original confidence assignment; Let be the weight value of the j-th information source; For the discounted confidence allocation; Assign confidence levels for uncertainty across all categories after discounting; This is the classic Dempster-Shafer combination rule.

7. A behavior recognition system based on uncertain correlations, characterized in that, include: The data acquisition module is used to acquire human behavior data collected by wearable sensors as raw data; Feature extraction is performed on the original data to establish... Feature dataset; The feature dataset is divided into training and testing sample sets; the feature dataset is further separated according to the distribution location and type of wearable sensors to obtain several source feature data. The fuzzy partitioning module is used to perform entropy-based adaptive fuzzy partitioning on several source feature data to obtain the feature fuzzy set corresponding to each source. The association rule learning module is used to learn evidence association rules based on several source feature data and the feature fuzzy set corresponding to each source, and generate a candidate rule set. The objective function construction module is used to generate the final rule set based on the candidate rule set using a database coverage strategy; A confusion matrix is ​​generated by classifying the training sample set based on the final rule set; Based on the confusion matrix, establish the objective function; The objective function solution module is used to solve the objective function using a genetic algorithm to obtain the optimal source weight values; The source feature selection module is used to select k source feature data with the largest source weight values ​​from a number of source feature data based on the optimal source weight values, so that the cumulative contribution rate reaches a set contribution rate threshold, and obtain k source feature data with high source weight values. The fusion decision module is used to identify the test sample set based on the Dempster-Shafer decision-level fusion strategy. It identifies the test sample set based on k source feature data with high source weight values ​​and obtains the behavior recognition results based on uncertain correlation.

8. An electronic device, characterized in that, include: A processor is used to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, performs the behavior recognition method based on uncertain association relationships as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the behavior recognition method based on uncertain association relationships as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the behavior recognition method based on uncertain association relationships as described in any one of claims 1-6.

Images