Search

CN-121987222-A - Electroencephalogram signal characteristic layering optimization method for special equipment operator fatigue detection

CN121987222ACN 121987222 ACN121987222 ACN 121987222ACN-121987222-A

Abstract

A layering optimization method for electroencephalogram characteristics oriented to fatigue detection of a special equipment operator is characterized by comprising the steps of collecting original electroencephalograms of the special equipment operator under different fatigue grades, carrying out multi-step preprocessing on the original electroencephalograms to obtain a high-purity electroencephalogram sample set, dividing the high-purity electroencephalogram sample set according to frequency bands, extracting multi-dimensional and multi-type characteristics, carrying out standardized normalization processing on the characteristics to construct a multi-dimensional characteristic set, carrying out accurate optimization on the multi-dimensional characteristic set through a four-stage layering collaborative optimization mechanism to obtain an optimal core characteristic subset, wherein the four-stage layering collaborative optimization mechanism comprises intra-class pre-screening, cross-type colinear elimination, physiological weighting strengthening screening and attention-directed recursive characteristic elimination optimization, inputting the optimal core characteristic subset into a detection model for detection, and outputting a fatigue state detection result.

Inventors

  • CHEN JICHI
  • YUAN GUANGXIANG
  • WEI CHUNFENG
  • FAN FUCHANG
  • HE ENQIU

Assignees

  • 沈阳工业大学

Dates

Publication Date
20260508
Application Date
20260114

Claims (9)

  1. 1. An electroencephalogram signal characteristic layering optimization method for special equipment operator fatigue detection is characterized by comprising the following steps of: Collecting original electroencephalogram signals of operators of special equipment under different fatigue levels; carrying out multi-step pretreatment on the original electroencephalogram signals to obtain a high-purity electroencephalogram signal sample set; Dividing a high-purity electroencephalogram signal sample set according to frequency bands, extracting multi-dimensional and multi-type features, and carrying out standardized normalization processing on the features to obtain a multi-dimensional feature set; accurately optimizing the multi-dimensional feature set through a four-stage hierarchical collaborative optimization mechanism to obtain an optimal core feature subset, wherein the four-stage hierarchical collaborative optimization mechanism comprises intra-class pre-screening, cross-type collinearity elimination, physiological weighting strengthening screening and attention-directed recursive feature elimination optimization; and inputting the optimal core feature subset into a detection model for detection, and outputting a fatigue state detection result.
  2. 2. The electroencephalogram signal characteristic layering optimization method for special equipment operator fatigue detection according to claim 1, wherein, The acquisition process of the original brain electrical signals comprises the following steps: According to international 10-20 system expansion standard, arranging silver-silver chloride dry electrodes in 24 key brain areas such as frontal lobe, top lobe, temporal lobe, occipital lobe, central area and the like of the scalp of a special equipment operator; And synchronously collecting original electroencephalogram signals of an operator of special equipment in four fatigue levels of awake state, mild fatigue, moderate fatigue and severe fatigue, wherein the sampling frequency is 500Hz.
  3. 3. The electroencephalogram signal characteristic layering optimization method for special equipment operator fatigue detection according to claim 1, wherein, The process for carrying out multi-step preprocessing on the original electroencephalogram signal comprises the following steps: filtering low-frequency drift and high-frequency noise by adopting a band-pass filter of 0.3-48 Hz; removing physiological artifacts by a method of combining independent component analysis and wavelet threshold denoising; and (3) carrying out sample segmentation on the continuous signals after artifact removal through a sliding time window with the window length of 2.5s and the overlapping proportion of 55%, so as to obtain a high-purity electroencephalogram signal sample set.
  4. 4. The electroencephalogram signal characteristic layering optimization method for special equipment operator fatigue detection according to claim 1, wherein, Dividing a high-purity electroencephalogram signal sample set according to frequency bands, and extracting multi-dimensional and multi-type features comprises the following steps: the high-purity EEG signal is divided into four frequency bands according to physiological functions, namely a full frequency band (0.3-48 Hz), Wave band (4-8 Hz), Wave band (8-14 Hz), Band (14-30 Hz); And respectively extracting four types of characteristics, namely a time domain characteristic, a frequency domain characteristic, a nonlinear characteristic and a functional connection characteristic, from the signals of each frequency band.
  5. 5. The electroencephalogram signal characteristic layering optimization method for special equipment operator fatigue detection according to claim 1, wherein, The process for carrying out standardized normalization processing on the characteristics comprises the following steps: the multi-type characteristics are normalized through an improved Z-score standardization method to obtain a multi-dimensional characteristic set, and the formula is as follows: Wherein z i is a normalized characteristic value, x i is an original characteristic value, mu is a mean value of the characteristic, and sigma is a standard deviation of the characteristic.
  6. 6. The electroencephalogram signal characteristic layering optimization method for special equipment operator fatigue detection according to claim 1, wherein, The process for precisely optimizing the multi-dimensional feature set through the four-stage hierarchical collaborative optimization mechanism comprises the following steps: In the intra-class pre-screening stage, variance expansion factors and mutual information are calculated for each class of features of each frequency band, and co-linear features and weak correlation features with variance expansion factors larger than a preset threshold are removed by combining a maximum correlation minimum redundancy criterion to obtain an intra-class screening feature subset; In the cross-type collinearity elimination stage, four kinds of intra-class screening feature subsets of the same frequency band are spliced into a comprehensive feature matrix, invalid features with variance smaller than 10 -7 are eliminated, cross-type collinearity features are eliminated by iterative calculation of variance expansion factor values until all feature variance expansion factors are smaller than or equal to a preset threshold value, and a non-collinear feature subset is obtained; The physiological weighting strengthening screening stage constructs a dynamic weighting coefficient system based on brain physiological knowledge, distributes differential weighting coefficients for the non-collinear feature subsets of different frequency bands, brain regions and feature types, calculates the weighted mutual information value of each feature, sorts the weighted mutual information values in descending order, and reserves the features of the pre-set proportion to obtain a physiological strengthening feature subset; And (3) only calculating the differential type weight of the physiological enhancement feature subset in the attention-directed recursive feature elimination optimization stage, calculating the weighted importance score of the feature based on the improved support vector machine model, and iteratively eliminating the feature with the lowest score by the attention-directed recursive feature elimination algorithm until the feature dimension reaches a preset target dimension to obtain the optimal core feature subset.
  7. 7. The electroencephalogram signal characteristic layering optimization method for special equipment operator fatigue detection according to claim 6, wherein, The calculation process of the in-class pre-screening stage comprises the following steps: calculating a variance expansion factor: Wherein VIF k is the variance expansion factor of the kth feature, Constructing a decision coefficient of a linear regression model for the kth feature by taking all other features in the class as independent variables; mutual information calculation formula: Wherein MI (X, Y) is the mutual information value of the feature X and the fatigue label Y, P (X, Y) is the joint probability distribution of the feature discrete value and the fatigue label, P (X) is the feature edge probability, and P (Y) is the label edge probability.
  8. 8. The electroencephalogram signal characteristic layering optimization method for special equipment operator fatigue detection according to claim 6, wherein, The calculation process of the physiological weighted strengthening screening stage and the attention-directed recursive feature elimination optimization stage comprises the following steps: Physiological weighted mutual information value calculation formula: Wherein WMI k is the weighted mutual information value of the kth feature, MI k is the basic mutual information value of the kth feature, W b is the weighting coefficient of the frequency band to which the feature belongs, and W r is the weighting coefficient of the brain region to which the feature belongs; weighted importance score: Where WIS k IS the weighted importance score for the kth feature, IS k IS the original importance score for the kth feature, and W t IS the weight of the type to which the feature belongs.
  9. 9. The electroencephalogram signal characteristic layering optimization method for special equipment operator fatigue detection according to claim 1, wherein, The detection model adopts a machine learning model.

Description

Electroencephalogram signal characteristic layering optimization method for special equipment operator fatigue detection Technical Field The invention belongs to the technical field of artificial intelligence, and particularly relates to an electroencephalogram signal characteristic layering optimization method for special equipment operator fatigue detection. Background In the operation and inspection operation of special equipment such as boilers, pressure vessels, elevators, hoisting machinery and the like, operators need to keep high concentration under high risk, high load and long-time on duty conditions, and the fatigue state is a core risk source for causing safety accidents, reducing production efficiency and damaging product quality. The traditional fatigue monitoring means mainly depend on subjective evaluation, behavior characteristic observation or environment parameter indirect inference, have inherent defects of strong subjectivity, easiness in external environment interference, early warning lag, high misjudgment rate and the like, and are difficult to realize objective, accurate and real-time monitoring of the fatigue state. With the deep fusion of biomedical engineering and artificial intelligence technology, fatigue detection technology based on brain electrical signals becomes a research hotspot by virtue of unique advantages. The brain electrical signal is used as a physiological electrical signal directly reflecting the brain nerve activity, and the characteristic parameters of different frequency bands and the fatigue state have clear physiological association, in the fatigue process,The power of the wave band (4-8 Hz) has a remarkable rising trend along with the fatigue deepening,The power of the wave band (8-14 Hz) is gradually increased,The power of the wave band (14-30 Hz) is continuously reduced, and meanwhile, the functional connection modes of key brain regions such as frontal lobe, temporal lobe and the like are regularly changed. The physiological characteristics directly related to the fatigue state enable the electroencephalogram signal to become an ideal information carrier for fatigue detection, and the method has the natural advantages of objectivity, real-time and early warning in advance. However, the existing fatigue detection technology based on the electroencephalogram signals still faces the following key technical bottlenecks: The multi-dimensional characteristic colinear redundancy is serious, and the generalization performance of the model is weak. In order to comprehensively capture fatigue related information in electroencephalogram signals, the prior art generally extracts multiple types and high-dimensional characteristics such as time domain, frequency domain, nonlinearity, functional connection and the like, but a mode of directly inputting a model after simple splicing is adopted, so that the problem of collinearity of cross-type and cross-frequency band characteristics is not effectively solved, the redundancy of characteristic space is high, the computational complexity of the model is greatly increased, the generalization capability is obviously reduced in cross-test and cross-scene test, and the requirements of embedded terminal equipment on instantaneity and stability cannot be met. Feature screening lacks physiological prior guidance and core features are easy to lose or redundant. The existing feature screening method is mostly based on a pure data driving strategy, physiological significance of brain electrical signals is not fully fused, and differential screening rules are not designed aiming at fatigue discrimination characteristics of different frequency bands, brain regions and feature types. This results in redundant features that are prominent in some data planes but weak in physiological significance being over-retained, while core features that have a key physiological indication effect on fatigue detection may be erroneously deleted, ultimately resulting in an insufficient fatigue characterization capability of the feature set. The feature optimization flow lacks a layering cooperative mechanism, and the precision and the instantaneity are difficult to balance. The existing method mostly adopts a simple process of 'one-time screening' or 'single-stage dimension reduction', and does not construct a layered collaborative optimization framework of 'redundancy removal-strong core-fine optimization', so that the feature set has too high dimension, increases calculation cost, influences real-time performance, loses key information due to too low dimension, reduces detection precision, and is difficult to realize effective balance between detection precision and real-time performance. Therefore, how to effectively solve the co-linearity problem among the multi-type electroencephalogram features and establish a feature screening framework integrating physiological priori knowledge and data driving so as to accurately extract a core feature set w