A fatigue driving detection method based on deep ensemble optimized fuzzy classifier

Through the deep ensemble optimized fuzzy classifier (DEO-TSK-FC), the noise, artifact and data imbalance problems in EEG signal fatigue detection are solved, the interpretability and classification accuracy of the model are improved, and efficient fatigue driving detection is achieved.

CN116616797BActive Publication Date: 2025-09-26CHENGDU UNIV OF INFORMATION TECH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310622556.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-05-29
Publication Date
2025-09-26
Estimated Expiration
2043-05-29

AI Technical Summary

Technical Problem

Existing technologies for fatigue detection in EEG signals suffer from noise, artifacts, dataset imbalance, and insufficient application capabilities of TSK fuzzy systems in different scenarios, and fail to fully explain the learned features, resulting in poor classification results.

Method used

The deep ensemble optimized fuzzy classifier (DEO-TSK-FC) is adopted. By stacking multiple zero-order fuzzy systems TSK, expanding the input data and performing fuzzy rule multiplication, combined with batch normalization and minimum learning machine algorithm, the numerical overflow and downflow problems are solved, and the model interpretability and classification accuracy are improved.

Benefits of technology

It effectively reduces the time complexity of large matrix operations, improves the detection effect of EEG data, solves the problems of data imbalance and poor model interpretability, and realizes efficient fatigue driving detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116616797B_ABST
    Figure CN116616797B_ABST
Patent Text Reader

Abstract

This invention provides a fatigue driving detection method based on a deep ensemble optimized fuzzy classifier. The method comprises: inputting preprocessed data into the deep ensemble optimized fuzzy classifier to obtain a first output vector and a first fatigue driving detection result, and simultaneously randomly selecting a first fuzzy rule; first determining whether the first fatigue driving detection result meets a preset classification accuracy, and then determining whether the second fatigue driving detection result meets the preset classification accuracy. If so, the fatigue driving detection result is output; otherwise, the method performs a multiplication operation by concatenating the output vectors and accumulating fuzzy rules. This invention solves the problems of imbalanced fatigue driving data and poor model interpretability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of fatigue driving detection, and in particular relates to a fatigue driving detection method based on a deep integrated optimized fuzzy classifier. Background Art

[0002] Using advanced machine learning methods to predict driver fatigue is crucial for reducing the incidence of traffic accidents. When a driver is fatigued, their thinking and decision-making abilities decline, leading to abnormal vehicle behavior. Therefore, a driver fatigue detection algorithm with high interpretability and classification accuracy can provide a strong guarantee for safe travel.

[0003] Early fatigue detection algorithms can be categorized into subjective and objective methods. Subjective methods are significantly influenced by human judgment and lack strong persuasiveness. Objective detection methods include physiological signals, facial features, and vehicle trajectory. The calculation of facial features and vehicle trajectory is affected by lighting technology and driver skill, respectively. However, analyzing physiological signals can effectively avoid these issues. Among physiological signals, electroencephalogram (EEG) signals can capture information about brain neurophysiological activity and accurately reflect the driver's mental state. Therefore, EEG analysis can be used to assess driver fatigue.

[0004] However, EEG signals suffer from noise, artifacts, and dataset imbalance, which hinder their classification effectiveness. This is because, under normal circumstances, most drivers unconsciously relax when they feel tired, thereby reducing fatigue. Therefore, the imbalance of EEG data needs to be considered. In most existing studies, deep learning methods have been used to detect fatigue. Although good results have been achieved, the learned features are not well explained. In recent years, TSK fuzzy systems have been widely used in EEG-based fatigue detection tasks due to their high interpretability.

[0005] Although TSK fuzzy systems have achieved remarkable results in EEG signal classification tasks, they still have certain limitations. 1) Their application capabilities in different scenarios need to be improved. 2) In practical calculations, they do not consider overflow and underflow problems during numerical calculations. 3) They do not fully account for the imbalance of EEG data. 4) They do not fully explain the features they learn. Summary of the Invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a fatigue driving detection method based on deep integrated optimized fuzzy classifier to solve the above-mentioned problems.

[0007] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0008] This solution provides a fatigue driving detection method based on deep ensemble optimized fuzzy classifier, which includes the following steps:

[0009] S1. Obtaining raw data including EEG signal data and fatigue driving dataset, and preprocessing the raw data;

[0010] S2. Inputting the preprocessed data into a deep ensemble optimized fuzzy classifier to obtain a first output vector and a first fatigue driving detection result, and randomly selecting a first fuzzy rule at the same time;

[0011] S3. Determine whether the first fatigue driving detection result reaches a preset classification accuracy. If so, output the fatigue driving detection result and terminate the process. Otherwise, merge the first output vector and the original data into first input data, send the data to the fuzzy subclassifier of the current order of the deep integrated optimization fuzzy classifier, add the cumulative multiplication of the first fuzzy rule, obtain the first output vector, the second output vector, and the second fatigue detection result, and randomly select a second fuzzy rule at the same time, and proceed to S4, wherein the first input data includes the first output vector and the original data;

[0012] S4. Determine whether the second fatigue detection result reaches the preset classification accuracy. If so, output the fatigue driving detection result and end the process. Otherwise, incorporate the second output vector into the second input data, and incorporate the second fuzzy rule into the next-order fuzzy sub-classifier in the deep integrated optimization fuzzy classifier, increase the multiplication of the second fuzzy rule, and repeat S4 to continuously increase the splicing of the output vector and the fuzzy rule multiplication operation until the second fatigue detection result reaches the preset classification accuracy, wherein the second input data includes the first output vector, the second output vector and the original data.

[0013] The present invention has the following beneficial effects: It predicts EEG fatigue driving data by training each fuzzy subclassifier and proposing a classification process using a deep ensemble optimized fuzzy classifier, DEO-TSK-FC. The present invention has the advantage that learning can be performed through the output layer, eliminating the need for more time-consuming iterative operations. Furthermore, the present invention is primarily applicable to EEG data. The aforementioned design can reduce the large matrix operations caused by large sample data, thereby addressing the issues of high time complexity, imbalanced fatigue driving data, and poor model interpretability.

[0014] Furthermore, the S1 includes the following steps:

[0015] S101, obtaining raw data including electroencephalogram signal data and fatigue driving data set;

[0016] S102, filtering and principal component analysis dimensionality reduction processing are performed on the EEG signal data and the fatigue driving data set;

[0017] S103: Use the fuzzy mean clustering algorithm to perform fuzzification operation on the data processed in S102 to complete preprocessing.

[0018] The beneficial effect of the above further solution is that the present invention can effectively improve the detection effect by feeding the original data using fuzzy attributes into the fuzzy sub-classifier.

[0019] Furthermore, the step S2 includes the following sub-steps:

[0020] S201, randomly initialize the feature selection matrix FDM and the rule composition matrix FRM;

[0021] S202, according to the pre-processed data, the feature x is calculated using the Gaussian membership function ij Gaussian membership value of , where x ij represents the jth feature of the i-th sample;

[0022] S203, calculate each feature x according to the Gaussian membership value, the feature selection matrix FDM and the rule composition matrix FRM ij Corresponding fuzzy rules;

[0023] S204, according to each feature x ij The corresponding fuzzy rules are normalized by batch processing to obtain the processed fuzzy rules;

[0024] S205, calculating fuzzy antecedent rules based on the processed fuzzy rules, performing softmax operation on the fuzzy antecedent rules to obtain processed fuzzy antecedent rules, and randomly selecting the first fuzzy rule;

[0025] S206, constructing a fuzzy layer output matrix according to the processed fuzzy antecedent rules;

[0026] S207, calculating the posterior parameters of the k-th fuzzy subclass classifier using a minimum learning machine according to the fuzzy layer output matrix;

[0027] S208. Calculate a first fatigue driving detection result based on the posterior parameters, and obtain a first output vector based on the first fuzzy rule.

[0028] The beneficial effect of the above further solution is that the present invention can effectively obtain the classification result of the current-order fuzzy sub-classifier by randomly selecting the first fuzzy rule and calculating the first output vector.

[0029] Furthermore, the expression of the Gaussian membership value is as follows:

[0030]

[0031]

[0032] in, represents the Gaussian membership value of the rth fuzzy rule, Z r represents the membership value of the rth fuzzy rule, exp(·) represents the exponential function, R represents the total number of fuzzy rules, D represents the total number of features of the input data, and x j Represents the jth feature of the input data, m r represents the center of the Gaussian membership function, σ r Represents the Gaussian membership function deviation.

[0033] The beneficial effect of the above further solution is that the present invention can effectively avoid numerical overflow and underflow problems during data calculation by calculating the Gaussian membership value.

[0034] Furthermore, each feature x ij The corresponding fuzzy rule expression is as follows:

[0035]

[0036] Among them, f r (x ij ) represents each feature x ij The corresponding fuzzy rules, r represents the fuzzy rules, FRM k represents the fuzzy rule matrix, l represents the semantic rule, Represents each feature x ij The corresponding r-th fuzzy rule Gaussian membership value, fdm jl Indicates whether the j-th feature of the input data belongs to the l-th semantic rule and is selected. If it is 0, the semantic rule is not selected.

[0037] The beneficial effect of the above further solution is that the present invention can effectively reduce redundant operations by cutting fuzzy rules and features and retaining the main calculation components.

[0038] Furthermore, the expression of the processed fuzzy rule is as follows:

[0039]

[0040] Among them, f' r (x ij ) represents the processed fuzzy rules, B N represents the batch normalization operation Batch-normal, γ and m are both self-learning hyperparameters, m B Indicates the calculation of the mean value of different fuzzy rules, σ B Indicates the deviation of the calculated values ​​of different fuzzy rules, ∈ indicates the minimum value, fr (x ij ) represents each feature x ij The corresponding fuzzy rules.

[0041] The beneficial effect of the above further solution is that the present invention makes each fuzzy rule consistent by calculating the processed fuzzy rules, thereby increasing the computational adaptability of the model in different data sets.

[0042] Furthermore, the expression of the fuzzy antecedent rule is as follows:

[0043]

[0044] Among them, W ir represents the fuzzy antecedent rule, f' r (x ij ) represents the processed fuzzy rules, D represents the total number of features of the input data, and j represents the number of features of the input data;

[0045] The expression of the fuzzy antecedent rule after the processing is as follows:

[0046]

[0047] Among them, W' ir represents the processed fuzzy antecedent rule, r represents the rth fuzzy rule, R represents the total number of fuzzy rules, and exp(·) represents the exponential function.

[0048] The beneficial effect of the above further solution is that the present invention facilitates the solution operation of fatigue driving detection results by calculating fuzzy rule values ​​and constructing a parameter matrix.

[0049] Furthermore, the expression of the fatigue driving detection result is as follows:

[0050]

[0051]

[0052]

[0053] Among them, y represents the fatigue driving detection result, Q k Represents the matrix constructed by the fuzzy rule output matrix Wir, Represents the matrix parameter result solved by the minimum learning machine algorithm LLM, w N5 It represents the output value of the 5th fuzzy rule of the Nth sample, N represents the total number of samples, C represents a constant, and I represents the unit matrix. Represents Q k The transpose of , L represents the label vector.

[0054] The beneficial effect of the above further scheme is: the present invention constructs a matrix equation of the fatigue driving detection result and uses iterative solution to calculate the fatigue driving detection result. At the same time, if the result does not meet the expectation, the output vector of the fatigue driving detection result is retained, and the previous input data is combined as the input of the next-order fuzzy sub-classifier to obtain an accurate fatigue driving detection result. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Flow chart of the method of the present invention.

[0056] Figure 2 It is a flowchart of the present invention. DETAILED DESCRIPTION

[0057] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0058] Example

[0059] Before describing the present invention, the following parameters are described:

[0060]

[0061] To address these issues in the aforementioned background technology, the present invention proposes a fatigue driving detection method based on a deep ensemble optimized fuzzy classifier. First, the present invention stacks multiple zero-order fuzzy systems (TSKs) to achieve the same effect as a high-order fuzzy system (TSK). Second, the present invention expands the input data, using both the first and second output vectors as inputs. The output of each layer is retained for input to the next layer, while a random fuzzy rule is retained for calculation in the next layer to enrich the fuzzy rule base. Finally, batch normalization (BN) and softmax operations are introduced to address value overflow and downflow issues. Finally, the LLM algorithm is proposed to calculate the posterior parameters to accelerate the training process.

[0062] like Figure 1 As shown, the present invention provides a fatigue driving detection method based on deep ensemble optimized fuzzy classifier, and its implementation method is as follows:

[0063] S1. Obtain the raw data including the EEG signal data and the fatigue driving dataset, and preprocess the raw data. The implementation method is as follows:

[0064] S101, obtaining raw data including electroencephalogram signal data and fatigue driving data set;

[0065] S102, filtering and principal component analysis dimensionality reduction processing are performed on the EEG signal data and the fatigue driving data set;

[0066] S103, using the fuzzy mean clustering algorithm to perform fuzzification operation on the data processed in S102 to complete preprocessing;

[0067] S2. Input the preprocessed data into a deep ensemble optimized fuzzy classifier to obtain a first output vector and a first fatigue driving detection result, and randomly select a first fuzzy rule at the same time. The implementation method is as follows:

[0068] S201, randomly initialize the feature selection matrix FDM and the rule composition matrix FRM;

[0069] S202, according to the pre-processed data, the feature x is calculated using the Gaussian membership function ij Gaussian membership value of :

[0070]

[0071]

[0072] in, represents the Gaussian membership value of the rth fuzzy rule, Z r represents the membership value of the rth fuzzy rule, exp(·) represents the exponential function, R represents the total number of fuzzy rules, D represents the total number of features of the input data, and x j Represents the jth feature of the input data, m r represents the center of the Gaussian membership function, σ r represents the Gaussian membership function deviation, x ij represents the jth feature of the i-th sample;

[0073] S203, calculate each feature x according to the Gaussian membership value, the feature selection matrix FDM and the rule composition matrix FRM ij The corresponding fuzzy rules are:

[0074]

[0075] Among them, f r (x ij ) represents each feature x ij The corresponding fuzzy rules, r represents the fuzzy rules, FRM k represents the fuzzy rule matrix, l represents the semantic rule, Represents each feature x ijThe corresponding r-th fuzzy rule Gaussian membership value, fdm jl Indicates whether the jth feature of the input data belongs to the lth semantic rule and is selected. If it is 0, the semantic rule is not selected;

[0076] S204, according to each feature x ij The corresponding fuzzy rules are processed by batch normalization to obtain the fuzzy rules:

[0077]

[0078] Among them, f' r (x ij ) represents the processed fuzzy rules, B N represents the batch normalization operation Batch-normal, γ and m are both self-learning hyperparameters, m B Indicates the calculation of the mean value of different fuzzy rules, σ B Indicates the deviation of the calculated values ​​of different fuzzy rules, ∈ indicates the minimum value, f r (x ij ) represents each feature x ij Corresponding fuzzy rules;

[0079] S205, calculating fuzzy antecedent rules based on the processed fuzzy rules, performing softmax operation on the fuzzy antecedent rules to obtain processed fuzzy antecedent rules, and randomly selecting the first fuzzy rule;

[0080] The expression of the fuzzy antecedent rule is as follows:

[0081]

[0082] Among them, W ir represents the fuzzy antecedent rule, f' r (x ij ) represents the processed fuzzy rules, D represents the total number of features of the input data, and j represents the number of features of the input data;

[0083] The expression of the fuzzy antecedent rule after the processing is as follows:

[0084]

[0085] Among them, W' ir represents the processed fuzzy antecedent rule, r represents the rth fuzzy rule, R represents the total number of fuzzy rules, and exp(·) represents the exponential function;

[0086] S206, constructing a fuzzy layer output matrix according to the processed fuzzy antecedent rules;

[0087] S207, calculating the posterior parameters of the k-th fuzzy subclass classifier using a minimum learning machine according to the fuzzy layer output matrix;

[0088] S208. Calculate a first fatigue driving detection result based on the posterior parameters, and obtain a first output vector based on a first fuzzy rule;

[0089] S3. Determine whether the first fatigue driving detection result reaches a preset classification accuracy. If so, output the fatigue driving detection result and terminate the process. Otherwise, merge the first output vector and the original data into first input data, send the data to the fuzzy subclassifier of the current order of the deep integrated optimization fuzzy classifier, add the cumulative multiplication of the first fuzzy rule, obtain the first output vector, the second output vector, and the second fatigue detection result, and randomly select a second fuzzy rule at the same time, and proceed to S4, wherein the first input data includes the first output vector and the original data;

[0090] S4. Determine whether the second fatigue detection result reaches the preset classification accuracy. If so, output the fatigue driving detection result and end the process. Otherwise, incorporate the second output vector into the second input data, and incorporate the second fuzzy rule into the next-order fuzzy sub-classifier in the deep integrated optimization fuzzy classifier, increase the multiplication of the second fuzzy rule, and repeat S4 to continuously increase the splicing of the output vector and the fuzzy rule multiplication operation until the second fatigue detection result reaches the preset classification accuracy, wherein the second input data includes the first output vector, the second output vector and the original data.

[0091] The expression of the output fatigue driving detection result is as follows:

[0092]

[0093]

[0094]

[0095] Among them, y represents the fatigue driving detection result, Q k Represents the matrix constructed by the fuzzy rule output matrix Wil, Represents the matrix parameter result solved by the minimum learning machine algorithm LLM, w N5 It represents the output value of the 5th fuzzy rule of the Nth sample, N represents the total number of samples, C represents a constant, and I represents the unit matrix. Represents Q k The transpose of , L represents the label vector.

[0096] In this embodiment, the classification task of EEG signal data generally includes two steps: signal preprocessing and classification. In the signal preprocessing stage, the original signal will be converted into classification attributes, but this conversion process will lead to information loss and data ambiguity. However, when using a fuzzy classifier in the classification stage, these losses can be taken into account. Therefore, the present invention can first perform fuzzy processing on the original data and then input it into the DEO-TSK-FC fuzzy classifier for classification. Figure 2 As shown, Figure 2 In the figure, the EEG signal EEG preprocessing is as follows: first, the original data (360, 310000) is filtered and the principal component analysis dimension reduction operation is performed to obtain the cleaned data (360, 25), and then the cleaned data is fuzzified. The fuzzified data (360, 25) is transmitted to the DEO-TSK-FC fuzzy classifier, and the final output is Yn+1, where Yn+1 represents the classification result, ∑ represents the sum operation, ∏ represents the cumulative multiplication operation, X represents the input data, L represents the label, and Y i Represents the features generated by the kth layer, Op-TSK i represents the kth zero-order TSK classifier operation, R i It represents the fuzzy rules selected by the k-th layer. Each fuzzy sub-classifier is superimposed, that is, the k-th one is the k-th fuzzy sub-layer classifier.

[0097] In this embodiment, the DEO-TSK-FC fuzzy classifier includes a stack structure, a learning algorithm, and a prediction process. Before the obtained EEG signal data is sent to the DEO-TSK-FC fuzzy classifier for classification, the artifacts are removed and blurred. The main method is filtering and PCA principal component analysis dimensionality reduction, and the fuzzy c-means clustering algorithm (FCM) is used for fuzzification operation. After preprocessing, the data set {X, L} is obtained, where X∈Rd represents the input data, Rd represents d-dimensional data, that is, x∈n*d, n represents the total number of samples, x represents a certain input data, X={x1,x2,…,xn}, xn represents the amount of input data, and L represents the corresponding actual label. The DEO-TSK-FC fuzzy classifier consists of several fuzzy sub-classifiers. For the convenience of description, each fuzzy sub-classifier is represented as Op-TSK k , each fuzzy sub-classifier (Op-TSK1, Op-TSK2, ..., Op-TSK k ) is a zero-order TSK fuzzy classifier.

[0098] In this example, the first layer of input data X is the original data. The input to each subsequent fuzzy subclassifier is the output of the previous fuzzy subclassifier appended to the input data. Furthermore, each layer randomly selects a fuzzy rule, multiplies it with the fuzzy rule of the previous layer, and calculates the output. The output of the kth layer is determined by the features generated by the kth layer.

[0099] In this embodiment, during the training process of the DEO-TSK-FC fuzzy classifier, the training starts with the zero-order fuzzy sub-classifier Op-T SK0 with the original data as input. The initial assumption is that Op-TSK0 consists of only 5 fuzzy rules, and its output is the output vector Y1, which is different from the subsequent fuzzy sub-classifiers. Next, the DEO-TSK-FC fuzzy classifier provides input for the next TSK fuzzy sub-classifier. When the TSK fuzzy sub-classifier fails to achieve the expected results in the zero-order fuzzy sub-classifier, the DEO-TSK-FC fuzzy classifier uses the input and output of the current layer to form the input of the next layer of fuzzy sub-classifier, and randomly selects the fuzzy rules of the current layer to multiply the fuzzy rules of the next layer. At this time, the input value is (X, Y1)∈Rd+1, X represents the input data, and Rd represents the d-dimensional data. The performance of the first fuzzy sub-classifier can be calculated by using each element in the output vector Y1. If the TSK fuzzy sub-classifier in the kth layer fails to achieve the desired performance, DEO-TSK-FC will continue training on the data in the k+1th layer and repeat the process until the expected result is obtained in the Kth layer, where K represents the total number of layers, i.e., the total number of fuzzy sub-classifiers.

[0100] In this embodiment, based on the characteristics of the stack structure, the method proposed by the present invention differs from the current deep structure of the fuzzy classifier. The present invention uses the calculated features of the fuzzy sub-classifier as the extended features of the next layer and retains a fuzzy rule to jointly calculate the output of the DEO-TSK-FC fuzzy classifier.

[0101] In this embodiment, from Figure 2 It can be seen that the interpretability of the DEO-TSK-FC fuzzy classifier is determined by the interpretability of each fuzzy sub-classifier. From the interpretability of each fuzzy sub-classifier, it can be seen that: (1) DEO-TSK-FC, which is composed of 0-order TSK fuzzy classifiers, has only one output and is therefore interpretable. (2) The calculation process and results of each layer are traceable, which helps to understand the classification process. In order to understand its interpretability, 5 Gaussian membership functions are used to represent each fuzzy rule. The centers of these membership functions (MF1, MF2, MF3, MF4, MF5) are 0.25, 0.5, 0.75 and 1 respectively. Therefore, the fuzzy rules can be semantically interpreted, that is, (very low, low, medium, high, very high). Based on these results, the interpretability of the 0-order TSK fuzzy classifier can be guaranteed.

[0102] In this embodiment, in the fuzzy rule generation of TSK fuzzy sub-classifier, not all rules of each fuzzy sub-classifier need to be calculated. Therefore, a feature selection matrix [FDM] is added. D×R To each op-TSKk (k=1,2,…,K), K represents the total number of fuzzy sub-classifiers. In addition, the present invention also adds a fuzzy rule matrix [FRM] D×5×R In this matrix, the value of each element is randomly assigned to 0 or 1, which determines which Gaussian membership function is used for each fuzzy rule. For example, FRM[4,2,3]=1 means that the fourth input feature in the third fuzzy rule uses a "low" membership function. Therefore, the output of the first-level zero-order fuzzy subclassifier of the DEO-TSK-FC fuzzy classifier can be expressed as:

[0103]

[0104] Among them, μX r (x j ) represents x j In the membership of the rth fuzzy rule, the Gaussian membership calculation function is used, exp(·) represents the exponential calculation, x j Represents the jth feature of the input data, m r represents the center FRM of the Gaussian membership function determined by the fuzzy rule matrix, σ r Denotes the Gaussian membership function deviation in the j-th feature.

[0105]

[0106]

[0107] Among them, f r (x) represents the fuzzy rule result, D represents the total number of features of the input data, fdm il Represents the feature composition matrix, y represents the predicted output, f r (x) represents the rth fuzzy rule result, R represents the total number of fuzzy rules, β r represents the subsequent parameters of the rth fuzzy rule, m r The present invention represents the center FRM of the Gaussian membership function determined by the fuzzy rule matrix. il Random initialization was performed.

[0108] In this embodiment, we will analyze the interpretability of each fuzzy sub-classifier in DEO-TSK-FC. Figure 2As can be seen from the figure, the first TSK fuzzy subclassifier only has the original input, and the output value of its sample can truly reflect its distinguishing information. Therefore, the first TSK fuzzy subclassifier is interpretable. The kth (k>1) TSK fuzzy subclassifier adds k-1 features and multiplication operations. For the input sample (X, Y1, ..., Yk-1) of the kth layer, the recursive expression of Op-TSKk of the current layer fuzzy subclassifier is:

[0109]

[0110]

[0111] in, represents the rth fuzzy rule value of the Kth layer, R k represents the fuzzy rule randomly retained at the kth level, y k represents the classification result of the kth layer, R represents the total number of fuzzy rules, β k represents the fuzzy consequent parameter calculated at the kth layer, R k is a fuzzy rule randomly selected in the kth layer. If k is 0, R k is 1. Based on the above analysis, it can be seen that the entire DEO-TSK-FC has strong interpretability.

[0112] In this embodiment, it is assumed that the TSK fuzzy subclassifier has R rules, and the corresponding inference rule is: If the first component of the input data X is X (r,1) , the second component is X (r,2) , until the D'th component is X (r,D') , then the output variable y r is equal to the function f r (x). This rule applies to all r values, ranging from 1 to R. In other words, when the input variables meet these conditions, the corresponding function will be used to calculate the value of the output variable.

[0113] In this embodiment, there are many related membership function sets in the TSK fuzzy subclassifier algorithm. At the same time, the fuzzy sets constructed by different membership function sets and the results of the initialization center are different. In order to facilitate calculation, the present invention uses the Gaussian membership function set, and the membership value μX r (x j ) can be as follows:

[0114]

[0115] Among them, m r and σ r Represents the center and standard deviation of the Gaussian membership function. Five Gaussian membership function centers are used in the present invention.

[0116] In this embodiment, μX is calculated r (x j ), since fuzzy AND operation is used, the emission level is calculated for each fuzzy rule, that is:

[0117]

[0118] Among them, f r (x i ) represents the emission level of the rth fuzzy rule, and D represents the total number of features of the input data. Based on the above, the weight calculation and normalization of each fuzzy rule are obtained:

[0119]

[0120] in, Indicates the calculation result of fuzzy rules, r indicates the rth fuzzy rule, and R indicates the number of fuzzy rules. Since the fuzzy rules may have numerical overflow and downflow problems during the calculation process, thus affecting the classification accuracy of the model, Rewritten as a softmax function:

[0121]

[0122]

[0123]

[0124] in, represents the Gaussian membership value of the rth fuzzy rule, Z r represents the membership value of the rth fuzzy rule, exp(·) represents the exponential function, R represents the total number of fuzzy rules, D represents the total number of features of the input data, and x j Represents the jth feature of the input data, m r represents the center of the Gaussian membership function, σ r Represents the Gaussian membership function deviation, y represents the output result, represents the result of fuzzy antecedent calculation, β r Represents the fuzzy consequent parameter, which is calculated by the LLM minimum learning machine algorithm.

[0125] Each feature x ij The corresponding fuzzy rule expression is as follows:

[0126]

[0127] Among them, f r (x ij ) represents each feature x ijThe corresponding fuzzy rules, r represents the fuzzy rules, FRM k represents the fuzzy rule matrix, l represents the semantic rule, represents the Gaussian membership value of the rth fuzzy rule, fdm jl Indicates whether the j-th feature of the input data belongs to the l-th semantic rule and is selected. If it is 0, the semantic rule is not selected.

[0128] The output obtained here is that the input data X is calculated to obtain the output result y, that is, the fatigue detection result. In order to optimize the TSK fuzzy subclassifier, a softmax activation layer is added to avoid numerical overflow and underflow, thereby improving the generalization performance of the model and reducing its overfitting problem. Here, β r Trained by the minimum learning machine LLM.

[0129] In this embodiment, batch normalization, as an optimization technique widely used in deep learning, is introduced in TSK sub-classification optimization, which can help the model converge faster and improve the generalization performance of the model. r (x i ), the output batch normalization BN result is:

[0130]

[0131] Among them, f' r (x ij ) represents the processed fuzzy rules, B N represents the batch normalization operation Batch-normal, γ and m are both self-learning hyperparameters, m B Indicates the calculation of the mean value of different fuzzy rules, σ B Indicates the deviation of the calculated values ​​of different fuzzy rules, ∈ indicates the minimum value, f r (x ij ) represents each feature x ij Corresponding fuzzy rules. Due to data dispersion across different experimental subjects, batch normalization (BN) can be used to optimize the TSK fuzzy subclassifier. Specifically, the present invention calculates the mean and standard deviation of the emission levels of different rules, then uses BN to normalize the rules to obtain a new rule level. Finally, the new rule level is used to calculate the classification result.

[0132] In this embodiment, the DEO-TSK-FC fuzzy classifier consists of two parts. The first part is to solve the k-th Op-TSK-FC through LLM, and the second part is to continuously call Op-TSK-FC starting from the k-th layer. k The training algorithm trains the entire DEO-TSK-FC fuzzy classifier. Assume that R1, R2, ..., R krepresents the fuzzy rule for random selection of sub-layers of layer k, β1,β2,...,β k Represents the output weight of the kth layer. For the training set where X n =[x1,x2,...,x n ] T ∈R N×D , Y n =[y1,y2,...,y n ],y n ∈{1,2,...,C'},y n Indicates the classification label corresponding to the C' classification problem, the intermediate emission layer H n =[R1,R2,...,R k ], posterior parameters X n Represents input data, x n Represents the characteristics of the input data, T represents transposition, R N×D Represents the real number space, the data dimension is N*D, D represents the total number of features of the input data, N represents the total number of samples, Y n represents the output feature, y n Indicates the classification result, C' indicates the number of classifications, which is 2 here, and H n Represents the fuzzy rule matrix of the intermediate calculation, R k represents the fuzzy rule, β k Indicates the consequent parameters calculated by LLM.

[0133] In this embodiment, in order to quickly solve the posterior parameters The number of sub-layers is determined in advance, the parameters of the current layer are calculated after obtaining the output of the previous layer, and finally the LLM algorithm is used to quickly solve the ridge regression problem with a given constant C.

[0134]

[0135] Among them, K represents the number of fuzzy sub-classifiers, y represents the output result of the last sub-classifier, that is, the fatigue driving detection result, R k represents the fuzzy rule selected at the kth level, represents the calculation result of the k-th layer fuzzy rule, β k Represents the subsequent parameters of the kth layer. If described in matrix form, it is:

[0136]

[0137] Among them, Y represents the output vector, Q k represents the fuzzy rule calculation of the kth sub-classifier, represents the consequent parameter of the fuzzy rule, which can be obtained by LLM solution. f r (x i ) represents the result of fuzzy rule calculation, R k Represents the fuzzy rule selected at the kth layer, r represents the fuzzy rule, R represents the total number of fuzzy rules, k represents the kth layer or the kth fuzzy sub-classifier, and the solution formula of the minimum learning machine LLM algorithm is obtained. The analytical expression is:

[0138]

[0139] Among them, y represents the fatigue driving detection result, Q k Represents the matrix constructed by the fuzzy rule output matrix Wir, Represents the matrix parameter result solved by the minimum learning machine algorithm LLM, w N5 It represents the output value of the 5th fuzzy rule of the Nth sample, N represents the total number of samples, C represents a constant, and I represents the unit matrix. Represents Q k The transpose of , L represents the label vector.

[0140] The advantage of using the minimum learning machine (LLM) algorithm is that learning can be performed through the output layer without the need for more time-consuming iterative operations. At the same time, since the present invention is mainly used for EEG data, which is a high-dimensional data, this form can reduce the large matrix operations generated by large sample data.

[0141] In summary, the main contributions of this invention are as follows: 1) To address data imbalance, oversampling is performed in the deep integration layer to expand features. 2) A copy of the fuzzy rules is retained at each layer to improve model reasoning capabilities. 3) To ensure interpretability, the resulting fuzzy rule matrix and feature selection matrix are constructed into a fuzzy rule base, providing explanatory power for the model's judgment of fatigue driving. Compared with existing methods, this invention has better interpretability and judgment capabilities.

Claims

1. A fatigue driving detection method based on deep ensemble optimized fuzzy classifier, characterized in that: The following steps are involved: S1. Obtaining raw data including EEG signal data and fatigue driving dataset, and preprocessing the raw data; S2. Inputting the preprocessed data into a deep ensemble optimized fuzzy classifier to obtain a first output vector and a first fatigue driving detection result, and randomly selecting a first fuzzy rule at the same time; S3. Determine whether the first fatigue driving detection result reaches a preset classification accuracy. If so, output the fatigue driving detection result and terminate the process. Otherwise, merge the first output vector and the original data into first input data, send the data to the fuzzy subclassifier of the current order of the deep integrated optimization fuzzy classifier, add the cumulative multiplication of the first fuzzy rule, obtain a second output vector and a second fatigue detection result, and randomly select a second fuzzy rule at the same time, and proceed to S4, wherein the first input data includes the first output vector and the original data; S4. Determine whether the second fatigue detection result reaches the preset classification accuracy. If so, output the fatigue driving detection result and end the process. Otherwise, incorporate the second output vector into the second input data, and incorporate the second fuzzy rule into the next-order fuzzy sub-classifier in the deep integrated optimization fuzzy classifier, increase the multiplication of the second fuzzy rule, and repeat S4 to continuously increase the splicing of the output vector and the fuzzy rule multiplication operation until the second fatigue detection result reaches the preset classification accuracy, wherein the second input data includes the first output vector, the second output vector and the original data.

2. The fatigue driving detection method based on deep ensemble optimized fuzzy classifier according to claim 1 is characterized in that: Said S1 comprises the following steps: S101, obtaining raw data including electroencephalogram signal data and fatigue driving data set; S102, filtering and principal component analysis dimensionality reduction processing are performed on the EEG signal data and the fatigue driving data set; S103: Use the fuzzy mean clustering algorithm to perform fuzzification operation on the data processed in S102 to complete preprocessing.

3. The fatigue driving detection method based on deep ensemble optimized fuzzy classifier according to claim 2 is characterized in that: The S2 comprises the following sub-steps: S201, randomly initialize the feature selection matrix FDM and the rule composition matrix FRM; S202, according to the pre-processed data, the feature is calculated using the Gaussian membership function The Gaussian membership value of Indicates the i The first sample j Features S203, calculate each feature according to the Gaussian membership value, feature selection matrix FDM and rule composition matrix FRM Corresponding fuzzy rules; S204, according to each feature The corresponding fuzzy rules are normalized by batch processing to obtain the processed fuzzy rules; S205, calculating fuzzy antecedent rules based on the processed fuzzy rules, performing softmax operation on the fuzzy antecedent rules to obtain processed fuzzy antecedent rules, and randomly selecting the first fuzzy rule; S206, constructing a fuzzy layer output matrix according to the processed fuzzy antecedent rules; S207, according to the fuzzy layer output matrix, use the minimum learning machine to calculate the k The posterior parameters of the fuzzy subclass classifier; S208. Calculate a first fatigue driving detection result based on the posterior parameters, and obtain a first output vector based on the first fuzzy rule.

4. The fatigue driving detection method based on deep ensemble optimized fuzzy classifier according to claim 3 is characterized in that: The expression of the Gaussian membership value is as follows: in, Indicates the r The Gaussian membership value of the fuzzy rules, Indicates the r The membership value of a fuzzy rule, represents the exponential function, represents the total number of fuzzy rules, represents the total number of features of the input data, Indicates the input data Features, represents the center of the Gaussian membership function, Represents the Gaussian membership function deviation.

5. The fatigue driving detection method based on deep ensemble optimized fuzzy classifier according to claim 3 is characterized in that: Each of the characteristics The corresponding fuzzy rule expression is as follows: in, Represents each feature The corresponding fuzzy rules are represents fuzzy rules, represents the fuzzy rule matrix, Represents semantic rules, Represents each feature The corresponding r The Gaussian membership value of the fuzzy rules, Indicates the input data j The feature belongs to l Whether a semantic rule is selected. If it is 0, the semantic rule is not selected.

6. The fatigue driving detection method based on deep ensemble optimized fuzzy classifier according to claim 3 is characterized in that: The expression of the processed fuzzy rule is as follows: in, represents the processed fuzzy rules, Indicates the batch normalization operation Batch-normal, and Both represent self-learning hyperparameters, Indicates that different fuzzy rules are input to calculate the mean value, Indicates the deviation of the calculated values ​​of different fuzzy rules, represents the minimum value, Represents each feature The corresponding fuzzy rules.

7. The fatigue driving detection method based on deep ensemble optimized fuzzy classifier according to claim 3 is characterized in that: The expression of the fuzzy antecedent rule is as follows: in, represents the fuzzy antecedent rule, represents the processed fuzzy rules, represents the total number of features of the input data, Indicates the number of features of the input data; The expression of the fuzzy antecedent rule after the processing is as follows: in, represents the processed fuzzy antecedent rule, r Indicates the r fuzzy rules, R represents the total number of fuzzy rules, Represents the exponential function.

8. The fatigue driving detection method based on deep ensemble optimized fuzzy classifier according to claim 1 is characterized in that: The expression of the fatigue driving detection result is as follows: in, Indicates the fatigue driving detection result. Represents the processed fuzzy antecedent rule The constructed matrix, Represents the matrix parameter result solved by the minimum learning machine algorithm LLM, Indicates the N The output value of the fifth fuzzy rule of samples, represents the total number of samples, represents a constant, represents the identity matrix, express The transpose of Represents a label vector.