A classification system and method based on data verification and controllable feature selection

Abstract

This invention belongs to the field of data representation, classification, and electrical signal processing technology, and relates to a classification system and method based on data verification and controllable feature selection. The system includes a data acquisition module, a feature extraction module, a feature cross-validation unit, and a machine learning classification module, as well as a data preprocessing module and an I / O control module. The I / O control module achieves controllable feature selection by controlling input and output. The system and method rely on the data preprocessing module to introduce autocorrelation and active segment secondary tests, improving data quality; rely on the feature extraction module to achieve multi-dimensional feature extraction; the I / O control module introduces controllable feature selection, realizing dynamic selection and optimization of features, improving the accuracy and flexibility of diagnosis; employs machine learning models for classification, achieving automation and intelligence in signal classification; and the modular design ensures close connection between functional units and smooth data flow, which is beneficial for stable system operation and upgrade maintenance.

Description

Technical Field

[0001] This invention belongs to the field of data representation, classification and electrical signal processing technology, and relates to a classification system and method based on data verification and controllable feature selection. Background Technology

[0002] With the rapid development of information technology and the acceleration of digital transformation, data is growing at an astonishing rate. This data explosion not only brings abundant information resources but also poses unprecedented challenges to data processing, analysis, and utilization. The complexity, diversity, and massive volume of data make traditional data processing methods inadequate, urgently requiring new technologies and methods to improve the efficiency and accuracy of data processing. Classification is a fundamental and crucial task across all industries. Whether in finance, telecommunications, industrial production, or biomedicine, data classification is essential to extract valuable information or knowledge. Data quality is a critical guarantee for the accuracy of classification results. However, in practical applications, data often suffers from missing data, anomalies, and noise, which severely impact the performance of classification models. Therefore, data verification before classification is an indispensable step. Thus, a more efficient and accurate classification method is needed to address these challenges. Data verification allows for the timely identification and handling of data problems, ensuring the accuracy and reliability of the input data for the classification model. Furthermore, in classification tasks, feature selection can eliminate redundant, irrelevant, or noisy features, improving the performance and efficiency of the classification model. Simultaneously, feature selection can reduce computational costs and storage requirements, making classification models easier to implement and deploy. Summary of the Invention

[0003] The purpose of this invention is to address the problems of insufficient data verification and incomplete feature extraction in existing classification methods, and to propose a classification system and method based on data verification and controllable feature selection.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] As one aspect of the present invention, an I / O control module is proposed, which receives preprocessed data and selects outputs for feature extraction, including mutually independent pre-processing and source feature selection units, higher-order class feature selection units, and attitude selection units; the pre-processing and source feature selection units select peak envelope and / or bandpass filtered output data of channel-selected data for feature extraction; the higher-order class feature selection units select low-pass filtered output data of channel-selected data and / or output data of channel-selected data after bandpass filtering and autocorrelation test and / or output data of channel-selected data after bandpass filtering, autocorrelation test, and active segment detection and / or output data of channel-selected data after low-pass filtering and stationary detection for feature extraction; the attitude selection unit selects output data of channel-selected data after low-pass filtering and / or stationary detection and / or normalization for feature extraction.

[0006] The preceding and source feature selection unit selects data after channel selection and analog-to-digital conversion, or data after channel selection, analog-to-digital conversion, and bandpass filtering, as the data for extracting preceding features; it selects peak envelope as the data for extracting source features; the peak envelope, after channel selection, analog-to-digital conversion, and bandpass filtering, is obtained by autocorrelation test, active segment detection, and then peak extraction.

[0007] The higher-order feature selection unit uses the following data to extract dynamic and fatigue features: low-pass filtered output data of channel-selected data, output data of channel-selected data after band-pass filtering, autocorrelation test, and active segment detection, output data of channel-selected data after analog-to-digital conversion, low-pass filtering, and static detection, and output data of channel-selected data after analog-to-digital conversion, band-pass filtering, autocorrelation test, and active segment detection. It also uses the following data to extract higher-order features: output data of channel-selected data after band-pass filtering, autocorrelation test, and active segment detection, parsed envelope, and output data of channel-selected data after analog-to-digital conversion, band-pass filtering, autocorrelation test, and active segment detection.

[0008] The parsed envelope is obtained by channel selection, analog-to-digital conversion, bandpass filtering, autocorrelation test, active segment detection, and then envelope parsing extraction.

[0009] The input data for stationary detection is the output data after low-pass filtering of the data after channel selection; the input data for normalization is the output data after stationary detection after low-pass filtering of the data after channel selection.

[0010] As another aspect of the present invention, the present invention proposes a classification system based on data verification and controllable feature selection, including a data acquisition module, a data preprocessing module, a feature extraction module, a feature cross-multiplication unit, and a machine learning classification module. The output of the data acquisition module is sent to the data preprocessing module for data preprocessing. The feature extraction module receives the preprocessed data and then selects data for extracting surface electromyography (SEMG), higher-order class, and inertial features. The extracted SEMG, higher-order class, and inertial features are input to the feature cross-multiplication unit for feature cross-multiplication and then output to the machine learning classification module for classification. The invention is characterized by further including an I / O control module, which is the I / O control module described in the first aspect. The output of the data preprocessing module is the input of the I / O control module. The I / O control module receives the output of the data preprocessing module and selects data from which SEMG, higher-order class, and inertial features can be extracted and outputs it to the feature extraction module.

[0011] The system relies on a feature extraction module to achieve multi-dimensional feature extraction, including surface electromyography features, preceding features, source features, dynamic features, fatigue features, higher-order features, and posture angle features, comprehensively characterizing signal features;

[0012] The data preprocessing module is used to preprocess the acquired raw signals, including a bandpass filter unit, a low-pass filter unit, a stillness detection unit, an envelope processing unit, and a normalization unit. It is characterized in that it also includes an autocorrelation test unit and an active segment detection unit for data verification.

[0013] The autocorrelation test unit removes data with no significant changes by calculating the autocorrelation of surface electromyography data after channel selection and bandpass filtering. The active segment detection unit divides the data after channel selection and bandpass filtering autocorrelation test, and retains active segments based on sample entropy and slope change frequency. The division uses a sliding window method to segment the data, and the value range of the data segments is 40 to 160; the sliding step size ranges from 10 to 60.

[0014] The data acquisition module is used to acquire surface electromyography (EMG) and attitude angle analog signals, including a multi-channel surface muscle electrode, a channel selection unit, a nine-axis inertial measurement unit, and an analog-to-digital conversion unit. After channel and axis selection, the analog signals are converted into digital signals. The channel selection unit is connected to the multi-channel surface muscle electrode and the nine-axis inertial measurement unit, respectively, and is used to select the corresponding channel for surface EMG, the nine-axis inertial measurement unit, and the number of axes, and output the surface EMG and attitude angle analog signals. The analog-to-digital conversion unit is connected to a bandpass filter unit and a low-pass filter unit, and is used to convert the surface EMG and attitude angle analog signals into digital signals. The envelope processing unit performs parsing envelope and peak envelope extraction.

[0015] The feature extraction module includes a surface electromyography (SEMG) feature extraction unit, a fatigue feature extraction unit, a higher-order class feature extraction unit, and an inertial feature extraction unit, used to extract multi-dimensional features. The SEMG feature extraction unit is used for pre-sequence feature extraction and source feature extraction. The inertial feature extraction unit is used for posture feature extraction. The pre-sequence and source feature extraction are connected to the pre-sequence and source feature selection units in the I / O control module. The dynamic feature extraction unit and the higher-order feature extraction unit are simultaneously connected to the fatigue feature extraction unit. The inertial feature extraction unit is connected to the posture feature selection unit in the I / O control module. The feature cross-referencing unit is used for feature filtering and combination. The outputs of all feature extraction modules are connected to the feature cross-referencing unit. The machine learning classification module uses the cross-referencing features to perform intelligent signal classification.

[0016] As a third aspect of the present invention, a classification method based on data verification and controllable feature selection is proposed, which applies the classification system based on data verification and controllable feature selection provided in the second aspect, and includes the following steps:

[0017] S1. Collect the surface electromyography (EMG) signal and posture angle signal of the subject, and obtain the surface EMG and posture data through analog-to-digital conversion; in specific implementation, S1 is implemented through the data acquisition module;

[0018] S2. The data obtained from S1 is filtered, autocorrelation tested, active segment detected, stationary segment detected, and envelope processed to obtain preprocessed data; in practice, S2 is implemented through the data preprocessing module.

[0019] S3. Select the feature type to be extracted according to preset conditions; in practice, S3 is implemented through the I / O control module.

[0020] S4. Based on the preset conditions in S3, extract multi-dimensional features from the preprocessed data; in practice, S4 is implemented through the feature extraction module.

[0021] S5. The extracted multi-dimensional features are filtered and combined for optimization to obtain the data to be classified; in practice, S5 is implemented through feature cross-units.

[0022] S6. Classify the data to be classified.

[0023] In practice, S6 is implemented through a machine learning classification module;

[0024] The data filtering described in S2 specifically involves: performing bandpass filtering and low-pass filtering on the electromyography (EMG) data and posture data after channel selection and analog-to-digital conversion, respectively; the autocorrelation test specifically involves: performing an autocorrelation test on the EMG data after channel selection and analog-to-digital conversion, and using the output data of the autocorrelation test for active segment detection; the active segment detection specifically involves: performing active segment detection on the data after performing autocorrelation test on the EMG data after channel selection and analog-to-digital conversion, and using the output data of the active segment detection for envelope extraction and higher-order class feature extraction; the stillness detection specifically involves: performing stillness detection on the posture data after low-pass filtering of the posture data after channel selection and analog-to-digital conversion, and using the output data of the stillness detection for higher-order class feature extraction or normalization before using it for posture feature selection; the envelope processing specifically involves: extracting the parsed envelope and peak envelope from the data obtained after performing autocorrelation test on the EMG data after channel selection and analog-to-digital conversion and then active segment detection, and then using them for higher-order class feature selection and preceding feature selection, respectively.

[0025] S3 describes selecting the feature types to be extracted based on preset conditions, specifically: using the low-pass filtered output data of the channel-selected data, the output data of the channel-selected data after band-pass filtering, autocorrelation test, and active segment detection, the output data of the channel-selected data after analog-to-digital conversion, low-pass filtering, and static detection, and the output data of the channel-selected data after analog-to-digital conversion, band-pass filtering, autocorrelation test, and active segment detection as data for extracting dynamic and fatigue features; and using the output data of the channel-selected data after band-pass filtering, autocorrelation test, and active segment detection, the parsed envelope, and the output data of the channel-selected data after analog-to-digital conversion, band-pass filtering, autocorrelation test, and active segment detection as data for extracting higher-order features.

[0026] The S1 acquires the subject's surface electromyography (EMG) signals and posture angle signals, and obtains the EMG and posture data through analog-to-digital conversion. In specific implementation, this is achieved through a data acquisition module, specifically by acquiring the subject's EMG signals and posture angle signals through a multi-channel surface muscle electrode or a nine-axis inertial sensing unit (IMU).

[0027] The surface electromyography (EMG) signal is the forearm muscle EMG signal acquired by the subject through a multi-channel surface muscle electrode; the posture angle signal is the posture angle signal acquired by the subject through a nine-axis IMU.

[0028] The multi-channel surface muscle electrode includes, but is not limited to, selecting 4-channel and 8-channel electrodes;

[0029] The attitude angle signal includes, but is not limited to, the attitude angle signal of the selected forearm node and thigh node; the attitude angle signal is a three-axis, six-axis or nine-axis output signal acquired by a nine-axis IMU;

[0030] The method and the system it relies on employ multi-channel surface muscle electrodes and a nine-axis inertial sensing unit (IMU), which enriches the types and dimensions of data acquisition and provides a more comprehensive data foundation for subsequent feature extraction.

[0031] The preprocessing described in S2, when implemented, includes the following sub-steps:

[0032] S21. Bandpass filter is applied to the surface electromyography signal collected in S1 to obtain bandpass filtered surface electromyography data.

[0033] The bandpass filter has a filtering range of 20-500Hz;

[0034] S22. Perform a 10Hz low-pass filter on the acquired attitude angle signal to obtain the low-pass filtered attitude data.

[0035] S23. Calculate the autocorrelation of the surface electromyography data after bandpass filtering and filter it to remove data with no significant change, and obtain the data after autocorrelation test.

[0036] The autocorrelation is calculated using the lag autocorrelation function and the discrete cross-correlation function.

[0037] S24. After dividing the data following the autocorrelation test, retain the active segment based on the sample entropy and the number of slope changes to obtain the data after the active segment detection.

[0038] The data is segmented using a sliding window method, with the window value ranging from 0.4 to 0.8 seconds and the step size ranging from 0.1 to 0.3 seconds.

[0039] S25. Perform parsing envelope extraction and peak envelope extraction on the data after the active segment detection to obtain the parsing envelope and peak envelope;

[0040] The analytical envelope extraction includes, but is not limited to, envelope extraction based on Hilbert transform and short-time energy method; the peak envelope extraction uses a processing window with a value range of 2-10.

[0041] S26. Perform static segment removal on the attitude data after low-pass filtering in S22 to obtain the data after removing the static segment;

[0042] The exclusion of static segments is based on the sliding window method for data partitioning, including but not limited to the exclusion of static data segments based on the threshold method, standard deviation method, and energy integration method; the window value range is 0.4-0.8s, and the step size ranges from 0.1-0.3s.

[0043] S27. Normalize the data after excluding static segments to obtain normalized data.

[0044] In practice, S3 consists of three parts:

[0045] (31) The preceding and source feature selection unit controls whether the signal after bandpass filtering and peak envelope processing is used for surface electromyography feature extraction.

[0046] (32) The higher-order feature selection unit controls whether the low-pass filtering, autocorrelation test, static detection, active segment detection and parsing envelope processing signal is used for dynamic features, higher-order features and fatigue extraction.

[0047] (33) Whether the normalized signal of the attitude feature selection unit is used for inertial feature extraction;

[0048] S3. Select the feature type to be extracted according to preset conditions; in practice, S3 is implemented through the I / O control module.

[0049] S4. Based on the preset conditions in S3, extract multi-dimensional features from the preprocessed data; in specific implementation, this is achieved through a feature extraction module; the multi-dimensional features are extracted from surface electromyography features, higher-order class features, and inertial features, specifically as follows:

[0050] S41. Extract the preceding features and source features of the input signal; the preceding features are the signal features when the muscle begins to contract but the action has not yet started; the source features include, but are not limited to, root mean square, average amplitude, and average power frequency; S41 is specifically implemented through a surface electromyography feature extraction unit.

[0051] S42. Extract the dynamic features, fatigue features, and higher-order features of the input signal;

[0052] The dynamic features include, but are not limited to, activity level, momentum, and relative rate of change; the higher-order features include, but are not limited to, skewness, kurtosis, entropy, and fractal dimension; the fatigue features include, but are not limited to, dynamic features, higher-order features, median frequency, mean absolute value, and mean amplitude; in specific implementation, S42 is achieved through a higher-order feature extraction unit.

[0053] S43. Extract the attitude features of the input signal; the attitude features include, but are not limited to, acceleration, angular velocity, relative position, and the features in S42; in specific implementation, S43 is achieved through an inertial feature extraction unit.

[0054] S5. Filter and optimize the extracted multi-dimensional features to obtain the data to be classified; in practice, this is achieved through feature cross-units, including the following sub-steps:

[0055] S51. The feature cross-combination unit performs cross-combination on the filtered features to generate new composite features;

[0056] S52. Use the feature selection method again to perform a second screening on the feature set containing composite features to obtain the final optimized feature set;

[0057] The cross-combination includes, but is not limited to, feature cross-combination based on factorization machine, Kronecker product, and attention decomposition machine;

[0058] The feature selection methods include, but are not limited to, NCFS, maximum correlation minimum redundancy, Relief algorithm, and random forest algorithm.

[0059] S6. Classify the data to be classified, specifically including the following sub-steps:

[0060] S61. Input the optimized features into the pre-trained machine learning model;

[0061] S62. The machine learning model outputs classification results, including classification categories and confidence levels;

[0062] S63. If the confidence level is lower than the preset threshold, the system will mark the sample as "pending confirmation".

[0063] The machine learning models include, but are not limited to, SVM, decision trees, multilayer perceptron models, logistic regression, and neural networks.

[0064] Beneficial effects

[0065] The classification system based on data verification and controllable feature selection described in this invention has the following advantages compared with existing classification systems:

[0066] 1. The system employs a multi-channel surface muscle electrode and a nine-axis inertial measurement unit, which enriches the types and dimensions of data acquisition and provides a more comprehensive data foundation for subsequent feature extraction;

[0067] 2. The system relies on the feature extraction module to achieve multi-dimensional feature extraction, including surface electromyography features, preceding features, source features, dynamic features, fatigue features, higher-order features and posture angle features, comprehensively characterizing signal features;

[0068] 3. The system introduces a controllable feature selection mechanism based on the I / O control module. Through the I / O control module and the feature intersection and selection module, dynamic selection and optimized combination of features are realized, which improves the accuracy and flexibility of subsequent diagnosis.

[0069] 4. The machine learning classification module of the system uses machine learning methods to classify signals, realizing the automation and intelligence of the signal classification process;

[0070] 5. The modular design of the system ensures close connection between various functional units and smooth data flow, which is conducive to the stable operation of the system and subsequent upgrades and maintenance. Attached Figure Description

[0071] Figure 1 This is a classification system based on data verification and controllable feature selection in the embodiments of the present invention;

[0072] Figure 2 This is a schematic diagram of the I / O control interface of a classification system based on data verification and controllable feature selection in an embodiment of the present invention. Detailed Implementation

[0073] The following description, in conjunction with the accompanying drawings and embodiments, provides a further explanation and detailed description of the classification system based on data verification and controllable feature selection according to the present invention.

[0074] Example 1

[0075] This embodiment illustrates the specific implementation of a classification system based on data verification and controllable feature selection, as described in this invention, for the auxiliary diagnosis and detection of Attention Deficit Hyperactivity Disorder (ADHD). ADHD is a common neurodevelopmental disorder characterized by difficulty concentrating, hyperactivity, and increased impulsive behavior. Initial diagnosis of ADHD often begins in childhood, but its effects are far-reaching, frequently extending into adulthood and hindering daily life and social interactions. Currently, the diagnosis of ADHD heavily relies on detailed clinical evaluation, a process that is not only time-consuming but also faces the challenge of a global shortage of professional diagnostic resources, often leading to delays in diagnosis. Given these challenges, recent innovations in machine learning and artificial intelligence have opened new avenues for rapid and accurate diagnosis of ADHD. However, existing auxiliary diagnostic systems largely rely on high-end medical equipment, such as MRI and EEG. These devices are not only expensive but also require patients to remain still for extended periods, increasing the complexity of diagnosis and the difficulty of patient cooperation. Although they demonstrate potential in improving diagnostic accuracy, their high cost, complex operation, and limited availability cannot be ignored. In contrast, combining surface electromyography (SEMG) and accelerometer nodes to acquire signals is simpler, providing comprehensive monitoring of motion status and offering advantages such as ease of use and low cost. Introducing a controllable feature selection mechanism allows for flexible adjustment of the feature set according to specific needs, resulting in more flexible and diverse feature selection. Combined with machine learning, this enables efficient and accurate diagnosis of ADHD, overcoming the shortcomings of existing methods.

[0076] This embodiment illustrates the application of the system described in this invention to the diagnosis of adult ADHD. In specific implementation of the system, as follows... Figure 1The classification system shown, based on data verification and controllable feature selection, includes a data acquisition module, a data preprocessing module, an I / O control module, a feature extraction module, a feature cross-referencing unit, and a machine learning classification module. The system introduces a controllable feature selection mechanism through the I / O control module. Through the I / O control module and the feature cross-referencing and selection module, dynamic selection and optimized combination of features are achieved, improving the accuracy and flexibility of subsequent diagnosis. The system uses the machine learning classification module for signal classification, realizing the automation and intelligence of the signal classification process. The modular design of the system ensures close integration between functional units and smooth data flow, which is beneficial for stable system operation and subsequent upgrades and maintenance.

[0077] The I / O control module receives preprocessed data and selects outputs for feature extraction. It includes independent pre-processing and source feature selection units, higher-order feature selection units, and attitude selection units. The pre-processing and source feature selection units select peak envelope and / or bandpass-filtered output data of channel-selected data for feature extraction. The higher-order feature selection units select low-pass-filtered output data of channel-selected data and / or output data of channel-selected data after bandpass filtering and autocorrelation testing and / or output data of channel-selected data after bandpass filtering, autocorrelation testing, and active segment detection and / or output data of channel-selected data after low-pass filtering and stationary detection for feature extraction. The attitude selection unit selects output data of channel-selected data after low-pass filtering and / or stationary detection and / or normalization for feature extraction.

[0078] The preceding and source feature selection unit selects data after channel selection and analog-to-digital conversion, or data after channel selection, analog-to-digital conversion, and bandpass filtering, as the data for extracting preceding features; it selects peak envelope as the data for extracting source features; the peak envelope, after channel selection, analog-to-digital conversion, and bandpass filtering, is obtained by autocorrelation test, active segment detection, and then peak extraction.

[0079] The higher-order feature selection unit uses the following data to extract dynamic and fatigue features: low-pass filtered output data of channel-selected data, output data of channel-selected data after band-pass filtering, autocorrelation test, and active segment detection, output data of channel-selected data after analog-to-digital conversion, low-pass filtering, and static detection, and output data of channel-selected data after analog-to-digital conversion, band-pass filtering, autocorrelation test, and active segment detection. It also uses the following data to extract higher-order features: output data of channel-selected data after band-pass filtering, autocorrelation test, and active segment detection, parsed envelope, and output data of channel-selected data after analog-to-digital conversion, band-pass filtering, autocorrelation test, and active segment detection.

[0080] The parsed envelope is obtained by channel selection, analog-to-digital conversion, bandpass filtering, autocorrelation test, active segment detection, and then envelope parsing extraction.

[0081] The input data for stationary detection is the output data after low-pass filtering of the data after channel selection; the input data for normalization is the output data after stationary detection after low-pass filtering of the data after channel selection.

[0082] On the other hand, a classification system based on data verification and controllable feature selection is proposed, including a data acquisition module, a data preprocessing module, a feature extraction module, a feature cross-multiplication unit, and a machine learning classification module. The output of the data acquisition module is sent to the data preprocessing module for data preprocessing. The feature extraction module receives the preprocessed data and then selects data for extracting surface electromyography (SEMG), higher-order class, and inertial features. The extracted SEMG, higher-order class, and inertial features are input to the feature cross-multiplication unit for feature cross-multiplication and then output to the machine learning classification module for classification. The system is characterized by further including an I / O control module, which is the I / O control module described in the first aspect. The output of the data preprocessing module is the input of the I / O control module. The I / O control module receives the output of the data preprocessing module and selects data from which SEMG, higher-order class, and inertial features can be extracted and outputs it to the feature extraction module.

[0083] The system relies on a feature extraction module to achieve multi-dimensional feature extraction, including surface electromyography features, preceding features, source features, dynamic features, fatigue features, higher-order features, and posture angle features, comprehensively characterizing signal features;

[0084] The data preprocessing module is used to preprocess the acquired raw signals, including a bandpass filter unit, a low-pass filter unit, a stillness detection unit, an envelope processing unit, and a normalization unit. It is characterized in that it also includes an autocorrelation test unit and an active segment detection unit for data verification.

[0085] The autocorrelation test unit removes data with no significant changes by calculating the autocorrelation of surface electromyography data after channel selection and bandpass filtering. The active segment detection unit divides the data after channel selection and bandpass filtering autocorrelation test, and retains active segments based on sample entropy and slope change frequency. The division uses a sliding window method to segment the data, with the segment value ranging from 40 to 160 and the sliding step size ranging from 10 to 60. Specifically, in this embodiment, the data segment value is 40, 100, or 160, and the sliding step size is 10, 30, or 60.

[0086] The data acquisition module is used to acquire surface electromyography (EMG) and attitude angle analog signals, including a multi-channel surface muscle electrode, a channel selection unit, a nine-axis inertial measurement unit, and an analog-to-digital conversion unit. After channel and axis selection, the analog signals are converted into digital signals. The channel selection unit is connected to the multi-channel surface muscle electrode and the nine-axis inertial measurement unit, respectively, and is used to select the corresponding channel for surface EMG, the nine-axis inertial measurement unit, and the number of axes, and output the surface EMG and attitude angle analog signals. The analog-to-digital conversion unit is connected to a bandpass filter unit and a low-pass filter unit, and is used to convert the surface EMG and attitude angle analog signals into digital signals. The envelope processing unit performs parsing envelope and peak envelope extraction.

[0087] The feature extraction module includes a surface electromyography (SEMG) feature extraction unit, a fatigue feature extraction unit, a higher-order class feature extraction unit, and an inertial feature extraction unit, used to extract multi-dimensional features. The SEMG feature extraction unit is used for pre-sequence feature extraction and source feature extraction. The inertial feature extraction unit is used for posture feature extraction. The pre-sequence and source feature extraction are connected to the pre-sequence and source feature selection units in the I / O control module. The dynamic feature extraction unit and the higher-order feature extraction unit are simultaneously connected to the fatigue feature extraction unit. The inertial feature extraction unit is connected to the posture feature selection unit in the I / O control module. The feature cross-referencing unit is used for feature filtering and combination. The outputs of all feature extraction modules are connected to the feature cross-referencing unit. The machine learning classification module uses the cross-referencing features to perform intelligent signal classification.

[0088] As a third aspect, a classification method based on data verification and controllable feature selection is proposed, which applies the classification system based on data verification and controllable feature selection provided in the second aspect, and includes the following steps:

[0089] S1. Collect the surface electromyography (EMG) signal and posture angle signal of the subject, and obtain the surface EMG and posture data through analog-to-digital conversion; in specific implementation, S1 is implemented through the data acquisition module;

[0090] S2. The data obtained from S1 is filtered, autocorrelation tested, active segment detected, stationary segment detected, and envelope processed to obtain preprocessed data; in practice, S2 is implemented through the data preprocessing module.

[0091] S3. Select the feature type to be extracted according to preset conditions; in practice, S3 is implemented through the I / O control module.

[0092] S4. Based on the preset conditions in S3, extract multi-dimensional features from the preprocessed data; in practice, S4 is implemented through the feature extraction module.

[0093] S5. The extracted multi-dimensional features are filtered and combined for optimization to obtain the data to be classified; in practice, S5 is implemented through feature cross-units.

[0094] S6. Classify the data to be classified.

[0095] In practice, S6 is implemented through a machine learning classification module;

[0096] The data filtering described in S2 specifically involves: performing bandpass filtering and low-pass filtering on the electromyography (EMG) data and posture data after channel selection and analog-to-digital conversion, respectively; the autocorrelation test specifically involves: performing an autocorrelation test on the EMG data after channel selection and analog-to-digital conversion, and using the output data of the autocorrelation test for active segment detection; the active segment detection specifically involves: performing active segment detection on the data after performing autocorrelation test on the EMG data after channel selection and analog-to-digital conversion, and using the output data of the active segment detection for envelope extraction and higher-order class feature extraction; the stillness detection specifically involves: performing stillness detection on the posture data after low-pass filtering of the posture data after channel selection and analog-to-digital conversion, and using the output data of the stillness detection for higher-order class feature extraction or normalization before using it for posture feature selection; the envelope processing specifically involves: extracting the parsed envelope and peak envelope from the data obtained after performing autocorrelation test on the EMG data after channel selection and analog-to-digital conversion and then active segment detection, and then using them for higher-order class feature selection and preceding feature selection, respectively.

[0097] S3 describes selecting the feature types to be extracted based on preset conditions, specifically: using the low-pass filtered output data of the channel-selected data, the output data of the channel-selected data after band-pass filtering, autocorrelation test, and active segment detection, the output data of the channel-selected data after analog-to-digital conversion, low-pass filtering, and static detection, and the output data of the channel-selected data after analog-to-digital conversion, band-pass filtering, autocorrelation test, and active segment detection as data for extracting dynamic and fatigue features; and using the output data of the channel-selected data after band-pass filtering, autocorrelation test, and active segment detection, the parsed envelope, and the output data of the channel-selected data after analog-to-digital conversion, band-pass filtering, autocorrelation test, and active segment detection as data for extracting higher-order features.

[0098] The S1 acquires the subject's surface electromyography (EMG) signals and posture angle signals, and obtains the EMG and posture data through analog-to-digital conversion. In specific implementation, this is achieved through a data acquisition module, specifically by acquiring the subject's EMG signals and posture angle signals through a multi-channel surface muscle electrode or a nine-axis inertial measurement unit (IMU).

[0099] The surface electromyography (EMG) signal is the forearm muscle EMG signal acquired by the subject through a multi-channel surface muscle electrode; the posture angle signal is the posture angle signal acquired by the subject through a nine-axis IMU.

[0100] The multi-channel surface muscle electrode includes, but is not limited to, selecting 4-channel and 8-channel electrodes; specifically in this embodiment, the surface muscle electrode has 4, 6 or 8 channels.

[0101] The attitude angle signal includes, but is not limited to, the attitude angle signal of the selected forearm node and thigh node; the attitude angle signal is a three-axis, six-axis or nine-axis output signal acquired by a nine-axis IMU; specifically in this embodiment, the IMU is a nine-axis output.

[0102] The method and the system it relies on employ multi-channel surface muscle electrodes and a nine-axis inertial sensing unit (IMU), which enriches the types and dimensions of data acquisition and provides a more comprehensive data foundation for subsequent feature extraction.

[0103] The preprocessing described in S2, when implemented, includes the following sub-steps:

[0104] S21. Bandpass filter is applied to the surface electromyography signal collected in S1 to obtain bandpass filtered surface electromyography data.

[0105] The bandpass filter has a filtering range of 20-500Hz;

[0106] S22. Perform a 10Hz low-pass filter on the acquired attitude angle signal to obtain the low-pass filtered attitude data.

[0107] S23. Calculate the autocorrelation of the surface electromyography data after bandpass filtering and filter it to remove data with no significant change, and obtain the data after autocorrelation test.

[0108] The autocorrelation is calculated using the lag autocorrelation function and the discrete cross-correlation function.

[0109] S24. After dividing the data following the autocorrelation test, retain the active segment based on the sample entropy and the number of slope changes to obtain the data after the active segment detection.

[0110] The data is segmented using a sliding window method, with the window value ranging from 0.4 to 0.8 seconds and the step size ranging from 0.1 to 0.3 seconds. In practice, the window value is 0.8 seconds and the step size is 0.1 or 0.2 seconds.

[0111] S25. Perform parsing envelope extraction and peak envelope extraction on the data after the active segment detection to obtain the parsing envelope and peak envelope;

[0112] The analytical envelope extraction includes, but is not limited to, envelope extraction based on Hilbert transform and short-time energy method; the peak envelope extraction uses a processing window with a value range of 2-10; in specific implementations, the processing window size used is 2, 6 or 10.

[0113] S26. Perform static segment removal on the attitude data after low-pass filtering in S22 to obtain the data after removing the static segment;

[0114] The exclusion of static segments is based on the sliding window method for data segmentation, including but not limited to the exclusion of static data segments based on the threshold method, standard deviation method, and energy integration method; the window value ranges from 0.4 to 0.8 s, and the step size ranges from 0.1 to 0.3 s; in specific implementation, the window value is 0.4 or 0.6 s, and the step size is 0.3 s.

[0115] S27. Normalize the data after excluding static segments to obtain normalized data.

[0116] In practice, S3 consists of three parts:

[0117] (31) The preceding and source feature selection unit controls whether the signal after bandpass filtering and peak envelope processing is used for surface electromyography feature extraction.

[0118] (32) The higher-order feature selection unit controls whether the low-pass filtering, autocorrelation test, static detection, active segment detection and parsing envelope processing signal is used for dynamic features, higher-order features and fatigue extraction.

[0119] (33) Whether the normalized signal of the attitude feature selection unit is used for inertial feature extraction;

[0120] S3. Select the feature type to be extracted according to preset conditions; in practice, S3 is implemented through the I / O control module.

[0121] S4. Based on the preset conditions in S3, extract multi-dimensional features from the preprocessed data; in specific implementation, this is achieved through a feature extraction module; the multi-dimensional features are extracted from surface electromyography features, higher-order class features, and inertial features, specifically as follows:

[0122] S41. Extract the preceding features and source features of the input signal; the preceding features are the signal features when the muscle begins to contract but the action has not yet started; the source features include, but are not limited to, root mean square, average amplitude, and average power frequency; S41 is specifically implemented through a surface electromyography feature extraction unit.

[0123] S42. Extract the dynamic features, fatigue features, and higher-order features of the input signal;

[0124] The dynamic features include, but are not limited to, activity level, momentum, and relative rate of change; the higher-order features include, but are not limited to, skewness, kurtosis, entropy, and fractal dimension; the fatigue features include, but are not limited to, dynamic features, higher-order features, median frequency, mean absolute value, and mean amplitude; in specific implementation, S42 is achieved through a higher-order feature extraction unit.

[0125] S43. Extract the attitude features of the input signal; the attitude features include, but are not limited to, acceleration, angular velocity, relative position, and the features in S42; in specific implementation, S43 is achieved through an inertial feature extraction unit.

[0126] S5. Filter and optimize the extracted multi-dimensional features to obtain the data to be classified; in practice, this is achieved through feature cross-units, including the following sub-steps:

[0127] S51. The feature cross-combination unit performs cross-combination on the filtered features to generate new composite features;

[0128] S52. Use feature selection again to perform a second screening on the feature set containing composite features to obtain the final optimized feature set;

[0129] The cross-combination includes, but is not limited to, feature cross-combination based on factorization machine, Kronecker product, and attention decomposition machine; in specific implementation, feature cross-combination is performed using Kronecker product.

[0130] The feature selection includes, but is not limited to, the NCFS method, the maximum correlation minimum redundancy method, the Relief algorithm, and the random forest algorithm. In specific implementation: the NCFS method is used for feature selection.

[0131] S6. Classify the data to be classified, specifically including the following sub-steps:

[0132] S61. Input the optimized features into the pre-trained machine learning model;

[0133] S62. The machine learning model outputs classification results, including classification categories and confidence levels;

[0134] S63. If the confidence level is lower than the preset threshold, the system will mark the sample as "pending confirmation".

[0135] The machine learning model includes, but is not limited to, SVM, multilayer perceptron model, decision tree, logistic regression and neural network; in specific implementation, SVM is used for classification.

[0136] In practice, the multi-channel surface muscle electrodes in the data acquisition module use the Delsys Trigno wireless surface electromyography (EMG) sensor, with a sampling frequency set to 600Hz; the nine-axis inertial measurement unit uses the Xsens MTw Awinda inertial measurement unit, with a sampling frequency set to 100Hz. The channel selection unit uses the NI USB-6343 multifunction data acquisition card for signal selection and acquisition.

[0137] In the data preprocessing module, the bandpass filter unit uses a 4th-order Butterworth filter with a frequency of 20-500Hz; the low-pass filter unit uses a 4th-order Butterworth low-pass filter with a cutoff frequency of 20Hz.

[0138] The autocorrelation test unit calculates the discrete cross-correlation of each channel of the signal; the calculation result is a symmetric sequence of correlations containing positive and negative delays. The first half of the symmetric sequence is extracted to obtain the extracted sequence and the mean of the extracted sequence is calculated; it is determined whether the mean is less than 0.66 and greater than 0.54. If so, the channel is valid; otherwise, the channel is deleted, that is, the data of this channel will not be used for subsequent active segment detection.

[0139] The activity segment detection unit first uses the sliding window method to segment the data, with the window size set to 0.6s and the step size to 0.2s. The slope change count (SSC) is calculated for each channel of the segmented data. If the SSC is greater than 2 and the sample entropy value is greater than 0.7, it indicates that there is muscle activity; otherwise, the channel is deleted.

[0140] In the envelope processing unit, the Hilbert transform method is used to parse the envelope, and the sliding window method is used for the peak envelope, with the window length set to 100ms.

[0141] The stationary detection unit first uses the sliding window method to segment the data, consistent with the active segment detection unit, with the window size set to 0.6s and the step size to 0.2s. The average acceleration value of each channel of the segmented data is calculated. If the average value of each channel is less than 10.8g / ㎡ and greater than 8.8g / ㎡, this segment of data is marked as a stationary segment.

[0142] The normalization unit calculation subtracts the mean from each channel of the data and divides by the standard deviation to transform the original data into a distribution with a mean of 0 and a standard deviation of 1.

[0143] The I / O control module is implemented using LabVIEW software programming. Dynamic control of feature selection is achieved through front panel controls, such as... Figure 2As shown. In specific implementation, clicking the "Preceding Features" button on the panel will cause the software to divide the data into equal-length segments of 0.6 seconds. For each segment, the software will calculate its root mean square, mean absolute value, and variance, as well as the data from the 0.2-second window before it, and add the results to the feature list. Clicking the "Higher-Order Features" button on the panel will cause the software to calculate the cross-entropy and fractal dimension of the data and add them to the feature list. Clicking the "Fatigue Features" button on the panel will cause the software to calculate the kurtosis and median frequency of the data and add them to the feature list. Finally, click "Feature Extraction Complete".

[0144] In the feature crossover and selection module, the feature crossover unit uses the attention decomposition machine method to perform feature crossover, and the feature selection unit uses the random forest algorithm in the scikit-learn library to obtain the feature importance ranking, and selects the top N features as the output features to the classifier; the value of N ranges from 10 to 200; in specific implementation, N is 30.

[0145] The machine learning ADHD diagnostic unit uses the TensorFlow framework to build a multilayer perceptron model. The hidden layer has a 3-layer structure with 64, 32 and 16 neurons in each layer, respectively. The activation function is ReLU and the output layer uses the Softmax function.

[0146] During system operation, surface electromyography (EMG) and posture angle signals are first collected from the subject while performing a standardized computer task, with a data acquisition time of 15 minutes. After preprocessing, the raw data is used by the I / O control module to select the types of features to be extracted. The feature extraction module extracts the selected multi-dimensional features, which are then optimized by the feature cross-validation and selection module and input into the machine learning ADHD diagnostic unit. The diagnostic unit outputs a diagnosis result of ADHD or no ADHD, along with a confidence level; the confidence threshold is set to 0.85.

[0147] In this embodiment, the system achieved a diagnostic accuracy of 88.0%, a sensitivity of 89.5%, and a specificity of 91.2% on a sample of 100 adult volunteers (50 ADHD patients and 50 healthy controls), and was able to provide objective auxiliary diagnostic results in a short time.

[0148] Table 1 shows the impact of the number of features selected in the feature cross unit on classification performance.

[0149] Table 1. Impact of the number of features selected in the feature cross unit on classification performance.

[0150]

[0151] As shown in Table 1, the effect is best when the number of features selected for the feature cross unit is 30.

[0152] Table 2 shows a comparison of the effects of feature cross units.

[0153] Table 2 Comparison of the effects of characteristic cross units

[0154]

[0155] As shown in Table 2, the feature crossover unit significantly improves the auxiliary diagnosis effect of ADHD after crossover of features and further feature selection, with an accuracy increase of 3.5%, sensitivity increase of 3.2%, and specificity increase of 1.7%.

[0156] The above description is merely a preferred embodiment of the present invention, and the present invention should not be limited to the content disclosed in this embodiment and the accompanying drawings. Any equivalent or modified embodiments made without departing from the spirit of the present invention fall within the scope of protection of the present invention.

Claims

1. An I / O control module, applied to a classification system based on data verification and controllable feature selection, characterized in that, The system incorporates a controllable feature selection mechanism based on the I / O control module. The output of the data preprocessing module serves as the input to the I / O control module. The I / O control module receives the output of the data preprocessing module and selects data from which surface electromyography, higher-order class features, and inertial features can be extracted, outputting this data to the feature extraction module. It receives preprocessed data and selects outputs for feature extraction, including independent pre-processing and source feature selection units, higher-order class feature selection units, and attitude selection units. The pre-processing and source feature selection unit selects peak envelope and / or bandpass-filtered output data from channel-selected data for feature extraction. The higher-order class feature selection unit selects low-pass-filtered output data from channel-selected data and / or output data from channel-selected data after bandpass filtering and autocorrelation testing, and / or output data from channel-selected data after bandpass filtering, autocorrelation testing, and active segment detection, and / or output data from channel-selected data after low-pass filtering and stationary detection for feature extraction. The attitude selection unit selects output data from channel-selected data after low-pass filtering and / or stationary detection and / or normalization for feature extraction.

2. The I / O control module according to claim 1, characterized in that, The preceding and source feature selection unit selects data after channel selection and analog-to-digital conversion, or data after channel selection, analog-to-digital conversion, and bandpass filtering, as the data for extracting preceding features; and selects peak envelope as the data for extracting source features. The peak envelope is obtained by channel selection, analog-to-digital conversion, bandpass filtering, autocorrelation test, active segment detection, and peak extraction.

3. The I / O control module according to claim 1, characterized in that, The higher-order feature selection unit uses the following data to extract dynamic and fatigue features: low-pass filtered output data of channel-selected data, output data of channel-selected data after band-pass filtering, autocorrelation test, and active segment detection, output data of channel-selected data after analog-to-digital conversion, low-pass filtering, and static detection, and output data of channel-selected data after analog-to-digital conversion, band-pass filtering, autocorrelation test, and active segment detection. It also uses the following data to extract higher-order features: output data of channel-selected data after band-pass filtering, autocorrelation test, and active segment detection, parsed envelope, and output data of channel-selected data after analog-to-digital conversion, band-pass filtering, autocorrelation test, and active segment detection.

4. The I / O control module according to claim 3, characterized in that, The parsed envelope is obtained by channel selection, analog-to-digital conversion, bandpass filtering, autocorrelation test, active segment detection, and then envelope parsing extraction.

5. The I / O control module according to claim 1, characterized in that, The input data for stationary detection is the output data after low-pass filtering of the data after channel selection; the input data for normalization is the output data after stationary detection after low-pass filtering of the data after channel selection.

6. A classification system based on data verification and controllable feature selection, comprising a data acquisition module, a data preprocessing module, a feature extraction module, a feature cross-multiplication unit, and a machine learning classification module, wherein the output of the data acquisition module is sent to the data preprocessing module for data preprocessing; the feature extraction module receives the preprocessed data and then selects features for extracting surface electromyography (SEMG), higher-order classes, and inertial features; the extracted SEMG, higher-order classes, and inertial features are input to the feature cross-multiplication unit for feature cross-multiplication, and then output to the machine learning classification module for classification, characterized in that... It also includes an I / O control module, wherein the I / O control module is the I / O control module according to any one of claims 1 to 5.

7. A classification system based on data verification and controllable feature selection according to claim 6, wherein the data preprocessing module is used to preprocess the acquired raw signal, including a bandpass filtering unit, a low-pass filtering unit, a stillness detection unit, an envelope processing unit, and a normalization unit, characterized in that, It also includes an autocorrelation testing unit and an active segment detection unit for data verification; The autocorrelation test unit removes data with no significant changes by calculating the autocorrelation of surface electromyography data after channel selection and bandpass filtering; the active segment detection unit divides the data after channel selection and bandpass filtering, and retains the active segment based on sample entropy and slope change count. The data is segmented using a sliding window method, with the data segment values ​​ranging from 40 to 160; the sliding step size ranges from 10 to 60; and the envelope processing unit performs parsing and peak envelope extraction. The data acquisition module is used to acquire surface electromyography (EMG) and attitude angle analog signals, including a multi-channel surface muscle electrode, a channel selection unit, a nine-axis inertial measurement unit, and an analog-to-digital conversion unit. After channel and axis selection, the analog signals are converted into digital signals. The channel selection unit is connected to the multi-channel surface muscle electrode and the nine-axis inertial measurement unit, respectively. It is used to select the corresponding channel of surface EMG, the nine-axis inertial measurement unit, and the number of axes, and output the surface EMG and attitude angle analog signals. The analog-to-digital conversion unit is connected to a bandpass filter unit and a low-pass filter unit, and is used to convert the surface EMG and attitude angle analog signals into digital signals. The feature extraction module includes a surface electromyography (SEMG) feature extraction unit, a fatigue feature extraction unit, a higher-order class feature extraction unit, and an inertial feature extraction unit, used to extract multi-dimensional features. The SEMG feature extraction unit is used for pre-sequence feature extraction and source feature extraction. The inertial feature extraction unit is used for posture feature extraction. The pre-sequence and source feature extraction are connected to the pre-sequence and source feature selection units in the I / O control module. The higher-order feature extraction unit is connected to the fatigue feature extraction unit. The inertial feature extraction unit is connected to the posture feature selection unit in the I / O control module. The feature cross-referencing unit is used for feature selection and combination. The output of the feature extraction module is connected to the feature cross-referencing unit. The machine learning classification module uses the cross-referencing features to perform intelligent signal classification.

8. A classification method based on data verification and controllable feature selection, employing the classification system based on data verification and controllable feature selection as described in claim 6 or 7, characterized in that, Includes the following steps: S1. Collect the surface electromyography (EMG) signal and posture angle signal of the subject, and obtain the surface EMG and posture data through analog-to-digital conversion; S2. Filter, autocorrelation test, active segment detection, stationary detection and envelope processing are performed on the data obtained in S1 to obtain preprocessed data; S3. Select the feature type to be extracted based on preset conditions; S4. Based on the preset conditions in S3, extract multi-dimensional features from the preprocessed data. S5. Filter and combine the extracted multi-dimensional features to obtain the data to be classified; S6. Classify the data to be classified.

9. A classification method based on data verification and controllable feature selection according to claim 8, characterized in that, S2 filters the data, specifically by performing band-pass filtering and low-pass filtering on the electromyography (EMG) data and posture data after channel selection and analog-to-digital conversion, respectively. The autocorrelation test is performed on the EMG data after channel selection and analog-to-digital conversion, and the output data is used for active segment detection. The active segment detection is performed on the EMG data after autocorrelation testing, and the output data is used for envelope extraction and higher-order feature extraction. The stationary detection is performed on the posture data after low-pass filtering, and the output data is used for higher-order feature extraction or normalization before being used for posture feature selection. The envelope processing involves extracting the parsed envelope and peak envelope from the EMG data after channel selection and analog-to-digital conversion, and then using them for higher-order feature selection and preceding feature selection, respectively.

10. A classification method based on data verification and controllable feature selection according to claim 8, characterized in that, S3 describes selecting the feature types to be extracted based on preset conditions, specifically: using the low-pass filtered output data of the channel-selected data, the output data of the channel-selected data after band-pass filtering, autocorrelation test, and active segment detection, the output data of the channel-selected data after analog-to-digital conversion, low-pass filtering, and static detection, and the output data of the channel-selected data after analog-to-digital conversion, band-pass filtering, autocorrelation test, and active segment detection as data for extracting dynamic and fatigue features; and using the output data of the channel-selected data after band-pass filtering, autocorrelation test, and active segment detection, the parsed envelope, and the output data of the channel-selected data after analog-to-digital conversion, band-pass filtering, autocorrelation test, and active segment detection as data for extracting higher-order features.

Images