Paddle number classification method based on principal component analysis of frequency domain characteristics of whisker sensor

Abstract

The application discloses a propeller blade number classification method based on a whisker sensor frequency domain feature principal component analysis, and the method comprises the following steps: collecting different propeller blade numbers based on a whisker sensor, and generating analog perception time sequence electric signals corresponding to periodic fluid disturbance in the process of different rotating speeds; preprocessing the analog perception time sequence electric signals to obtain optimized time sequence signals; extracting time / frequency domain features of the optimized time sequence signals to obtain multi-dimensional feature samples; performing feature dimension reduction processing on the multi-dimensional feature samples based on a frequency domain feature principal component analysis method to obtain two-dimensional feature projection samples; and realizing classification of propeller blade numbers based on the two-dimensional feature projection samples according to a K-means clustering algorithm. The application solves the problem that the existing method cannot effectively realize efficient identification and classification of propeller blade numbers.

Description

Technical Field

[0001] This invention relates to the field of underwater biomimetic sensing and intelligent signal recognition technology, and in particular to a propeller number classification method based on principal component analysis of frequency domain features of a tentacle sensor. Background Technology

[0002] Underwater propulsion devices (such as propellers) in marine environments generate periodic fluid disturbances during operation, and their spectral characteristics are closely related to the number of blades. Traditional propeller blade count identification methods mainly rely on high-precision underwater acoustic equipment or visual measurement methods, which suffer from problems such as complex structure, poor environmental adaptability, and insufficient real-time performance.

[0003] In recent years, biomimetic sensing technology has become an important development direction for underwater information detection. Among them, seal whiskers, due to their extremely high sensitivity to weak water flow disturbances, have been widely used in the field of biomimetic sensor design. Biomimetic whisker structures based on piezoelectric materials can convert fluid disturbances into electrical signals, thus providing a novel detection method for propeller feature recognition.

[0004] However, existing signal processing based on biomimetic sensors mainly focuses on flow field disturbance detection or wake reconstruction, with limited research on the automatic identification of the key parameter of propeller blade number. Furthermore, traditional signal analysis often employs single time-domain or frequency-domain methods, failing to fully utilize the multi-dimensional information such as the signal's time-frequency characteristics and envelope features, thus limiting classification accuracy.

[0005] Therefore, there is an urgent need for a method that integrates signal acquisition, filtering, multi-dimensional feature extraction, and intelligent classification to achieve efficient identification and classification of propeller blade count. Summary of the Invention

[0006] This invention provides a blade number classification method based on principal component analysis of frequency domain features of a whisker sensor to overcome the above-mentioned technical problems.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A blade number classification method based on principal component analysis of frequency domain features of a whisker sensor includes the following steps:

[0009] S1: By collecting data from propellers with different numbers of blades based on a strobe sensor, the analog sensing time-series electrical signals corresponding to the periodic fluid disturbances generated during operation at different speeds are obtained.

[0010] S2: Preprocess the analog sensing timing electrical signal to obtain an optimized timing signal;

[0011] The preprocessing includes performing analog-to-digital conversion on the analog sensing time-series electrical signal to obtain a digital signal;

[0012] The digital signal is sequentially de-trended and filtered to obtain an optimized time-domain signal;

[0013] S3: Extract time / frequency domain features from the optimized time-series signal to obtain multi-dimensional feature samples;

[0014] Furthermore, the multidimensional feature samples include the peak-to-peak ratio of the main frequency, energy of several frequency bands, frequency band energy ratio, fluctuation intensity, spectral centroid, spectral entropy, spectral kurtosis, and temporal envelope energy;

[0015] S4: Based on the frequency domain feature principal component analysis method, perform feature dimensionality reduction processing on the multidimensional feature samples to obtain two-dimensional feature projection samples;

[0016] S5: Based on the K-means clustering algorithm, the propeller blade number is classified according to the two-dimensional feature projection samples.

[0017] Furthermore, step S2 specifically includes the following steps:

[0018] S21: Perform analog-to-digital conversion on the analog sensing time-series electrical signal to obtain a digital signal;

[0019] S22: Perform detrending processing on the digital signal sequentially based on a sliding window to obtain a detrending signal;

[0020] S23: Perform initial filtering processing on the detrending signal based on a 4th-order Butterworth bandpass filter to obtain an initial filtered signal;

[0021] S24: Based on The filter performs bidirectional filtering on the initial filtered signal to eliminate phase distortion and obtain an optimized time-domain signal.

[0022] Furthermore, the expression for the detrending process described in S2 is:

[0023]

[0024]

[0025] In the formula: Represents digital signals; Indicates a smoothed signal; Indicates the window width; Indicates the time signature of a discrete signal; This represents the discrete signal beat after detrending; This represents the discrete signal after discretization of a digital signal. The discrete signal after the trend is the detrending signal.

[0026] Furthermore, the peak frequency ratio mentioned in S3 is the peak frequency ratio between 0-2Hz and 2-5Hz. Its expression is

[0027]

[0028] In the formula: Indicates the amplitude and frequency of the optimized timing signal; Indicates the frequency of the optimized timing signal;

[0029] The aforementioned frequency band energy includes the 0-2Hz frequency band energy. Energy in the 2-10Hz frequency band and energy in the 2-5Hz frequency band Its expression is

[0030]

[0031]

[0032]

[0033] In the formula: The time spectrum represents the optimized timing signal;

[0034] The frequency band energy ratio The expression is

[0035]

[0036] The fluctuation intensity The expression is

[0037]

[0038] The spectral centroid The expression is

[0039]

[0040] The spectrum entropy The expression is

[0041]

[0042] In the formula: Represents frequency The probability distribution;

[0043] The spectral kurtosis The expression is

[0044]

[0045] In the formula: Indicates the mean of the spectrum; Indicates the spectral variance;

[0046] The time-domain envelope energy The expression is

[0047]

[0048] In the formula: This represents the spectrum of the signal envelope.

[0049] Furthermore, S4 specifically includes the following steps:

[0050] S41: For the multidimensional feature samples Centralized processing is performed to obtain optimized samples;

[0051] And the expression for the centralization process is:

[0052]

[0053]

[0054]

[0055] In the formula: Indicates optimized sample; Represents the mean vector; It represents the mean of the same feature across all samples; Indicates the first The first multidimensional feature sample corresponding to the th _ ... n represents the number of multidimensional feature samples; p represents the number of features in each multidimensional feature sample.

[0056] S42: Calculate the covariance matrix based on the optimized samples, and the formula for calculating the covariance matrix is ​​as follows:

[0057]

[0058] In the formula: Represent the covariance matrix; Indicates transpose;

[0059] S43: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors;

[0060] And the eigenvalue decomposition expression of the covariance matrix is:

[0061]

[0062] In the formula: Represents the eigenvector; Represents the corresponding eigenvector eigenvalues;

[0063] The eigenvalues ​​obtained from eigenvalue decomposition are sorted in descending order to obtain a feature sequence list. The eigenvectors corresponding to the first two eigenvalues ​​in the feature sequence list are selected to form a two-dimensional mapping space for the data.

[0064] S44: Multidimensional feature samples Project onto a two-dimensional mapping space to obtain two-dimensional feature projection samples;

[0065] The formula for obtaining the feature projection samples is as follows:

[0066]

[0067] In the formula: Represents the feature projection sample; This represents a matrix composed of eigenvectors.

[0068] Furthermore, S5 specifically includes the following steps:

[0069] S51: Randomly divide the two-dimensional feature projection samples The initial propeller blade number clusters are determined, and the mean value corresponding to the feature projection sample in each initial propeller blade number cluster is obtained as the centroid of the initial cluster.

[0070] S52: Calculate the Euclidean distance from the feature projection samples in each initial propeller blade number cluster to the centroid and perform clustering. Divide each feature projection sample into the initial propeller blade number cluster with the smallest sum of squared Euclidean distances from the centroid and obtain the latest cluster.

[0071] S53: Obtain the mean value corresponding to the feature projection samples within each of the latest clusters, and use it as the latest centroid of the latest cluster;

[0072] S54: Obtain the current rate of change of the centroid based on the latest centroid and the centroid of the initial cluster, and determine whether the current rate of change has reached a preset index.

[0073] Furthermore, the preset indicators include the current rate of change being less than a preset threshold or reaching the maximum number of iterations;

[0074] If the preset target is met, then the latest cluster at this point is the optimal cluster.

[0075] If the preset target is to be achieved, repeat steps S52 to S53.

[0076] The optimal cluster is then used as the classification result for the number of propeller blades.

[0077] Furthermore, in S52, the Euclidean distance from the feature projection samples in each initial propeller blade number cluster to the centroid is calculated, and its expression is:

[0078]

[0079] In the formula: Indicates the first Initial propeller blade number clusters; This represents the feature projection samples in the initial propeller blade number cluster; Indicates the first The mean value corresponding to the feature projection samples in each cluster.

[0080] This invention provides a beneficial effect of a blade number classification method based on principal component analysis of frequency domain features of a whisker sensor:

[0081] (1) A beard sensor is used to simulate the vibration response of marine mammal whiskers under fluid disturbance, thereby achieving highly sensitive acquisition of micro-disturbances in the underwater flow field. Compared with traditional acoustic or visual detection methods, the piezoelectric bionic beard sensor does not require the emission of sound waves or reliance on lighting conditions, and can work stably in turbid water and low visibility environments. Since the piezoelectric sensing layer has a linear voltage response to small deformations, the sensor can accurately capture the periodic vibration differences caused by different numbers of blades in the propeller wake, thereby obtaining a higher signal-to-noise ratio and dynamic sensitivity.

[0082] (2) By employing signal preprocessing and feature extraction methods based on trend term removal, bandpass filtering, and time / frequency joint analysis, digital signals are sequentially detrended, filtered, and time / frequency domain feature extracted to obtain multi-dimensional feature samples; that is, firstly, a smoothing method is used to remove low-frequency drift, i.e., trend term removal; then, a fourth-order bandpass Butterworth filter and... The filter suppresses noise interference; finally, the combination of FFT and STFT is used to extract frequency domain and time-frequency domain features to obtain multi-dimensional feature samples, which comprehensively characterizes the flow field feature differences caused by the change in the number of blades. Compared with the existing single-dimensional methods based only on amplitude or main frequency features, this invention can achieve in-depth characterization of signal patterns from multiple perspectives such as energy distribution, spectral structure and time stability, which significantly improves feature discrimination.

[0083] (3) The PCA algorithm can effectively remove redundant correlations between features by reducing the dimensionality of multidimensional feature samples, compressing the ten-dimensional features into two principal components, and realizing low-dimensional visualization and efficient classification. Subsequently, the K-means algorithm automatically classifies the sample categories based on the principle of minimizing Euclidean distance, and can achieve cluster recognition of different blade numbers without the need for supervised labels. Attached Figure Description

[0084] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0085] Figure 1 This is a flowchart of the blade number classification method based on principal component analysis of frequency domain features of the whisker sensor according to the present invention;

[0086] Figure 2 This is a flowchart illustrating the technical process of the method described in this embodiment;

[0087] Figure 3 This is a frequency spectrum analysis diagram (0.1-30Hz) of a four-bladed propeller at a rotational speed of 500 rpm in this embodiment.

[0088] Figure 4 This is a heatmap showing the combined spectral and time-frequency characteristics in this embodiment;

[0089] Figure 5 This is a diagram showing the K-means clustering distribution results based on principal components in this embodiment. Detailed Implementation

[0090] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0091] This embodiment provides a blade number classification method based on principal component analysis of frequency domain features of a whisker sensor, such as... Figures 1 to 2 As shown, the steps include:

[0092] S1: By collecting data from propellers with different numbers of blades based on a strobe sensor, the analog sensing time-series electrical signals corresponding to the periodic fluid disturbances generated during operation at different speeds are obtained.

[0093] Specifically, the whisker sensor is a biomimetic seal whisker sensor. Inspired by the highly sensitive fluid sensing mechanism of seal whiskers in nature, this sensor combines a biomimetic structure with piezoelectric sensing technology to achieve high-precision sensing and classification of wake signals from propellers with different blade numbers. Seal whiskers (Vibrissae) are a typical passive fluid sensing organ; their unique wave-like geometry allows them to stably capture flow field disturbance signals in turbulent waters. They can sense not only changes in fluid velocity but also vortex structures in the wake. Even in completely dark underwater environments, seals can perceive the trajectory and shape of objects in front of them using only their whiskers, thanks to their highly sensitive response to weak fluid disturbances. This embodiment uses this as a biomimetic prototype to design a biomimetic whisker-type flow field sensor based on piezoelectric materials. The sensor uses a flexible substrate layer to support the biomimetic whisker structure, which undergoes periodic bending deformation under fluid action. This deformation is converted into an electrical signal output in the piezoelectric layer, realizing the electrical detection of underwater wake disturbance signals. By acquiring, preprocessing, and extracting features from signals, characteristic wake patterns generated by propellers with different numbers of blades can be identified. It includes a flexible shell housing a sensing unit, with a hollow cavity inside the shell housing the sensing unit; a sensing unit for generating analog sensing timing electrical signals (the sensing unit includes, in sequence, an external support material, a biomimetic seal whisker sensor flexible shell made primarily of PDMS, and a piezoelectric crystal, the piezoelectric crystal being embedded at the root of the biomimetic seal whisker sensor flexible shell, which is embedded within the external support material), and sending the analog sensing timing electrical signals to a preset preprocessing unit for preprocessing. The whisker sensor is implemented based on a preset STM32 control unit, with a sampling frequency set to 400Hz. The acquired raw signals are converted from analog to digital and stored in a 16-bit unsigned integer data format. After signal acquisition, the data is preprocessed using a Matlab platform.

[0094] S2: Preprocess the analog sensing timing electrical signal to obtain an optimized timing signal;

[0095] The preprocessing includes performing analog-to-digital conversion on the analog sensing time-series electrical signal to obtain a digital signal; and performing detrending and filtering processing on the digital signal in sequence to obtain an optimized time-domain signal.

[0096] The specific steps include:

[0097] S21: Perform analog-to-digital conversion on the analog sensing time-series electrical signal to obtain a digital signal;

[0098] S22: The LOWESS smoothing method is used to extract and remove low-frequency trend terms to eliminate the influence of background drift. Specifically, the digital signal is sequentially detrended using a sliding window to obtain a detrended signal. The expression for the detrending process is as follows:

[0099]

[0100]

[0101] In the formula: Represents digital signals; Indicates a smoothed signal; Indicates the window width; Indicates the time signature of a discrete signal; This represents the discrete signal beat after detrending; This represents the discrete signal after discretization of a digital signal. The discrete signal after representing the trend is the detrended signal; in this embodiment, the original signal, i.e., the digital signal, may have slow drift and low-frequency baseline fluctuations, which will mask the periodic signal generated by the propeller wake. The overall trend is removed by smoothing and subtraction.

[0102] S23: Perform initial filtering processing on the detrending signal based on a 4th-order Butterworth bandpass filter to obtain an initial filtered signal;

[0103] The expression for the fourth-order Butterworth bandpass filter is:

[0104]

[0105] In the formula: Indicates the center frequency; Indicates bandwidth; Indicates the high cutoff frequency (0.1Hz); Indicates a low cutoff frequency (30Hz); Represents the filter coefficients; Indicates the imaginary part; Indicates the filter order;

[0106] In this embodiment, the original signal is subject to drift below 0.1Hz caused by environmental disturbances and temperature changes, and noise above 30Hz caused by electronic noise and high-frequency vibration interference. Therefore, only the effective wake signal component within 0.1-30Hz of the detrending signal is retained. Figure 3 As shown;

[0107] S24: Based on The filter performs bidirectional filtering on the initial filtered signal to eliminate phase distortion and reduce transient effects, thereby obtaining an optimized time-domain signal. In this embodiment, to eliminate boundary effects, only the stable segment of the middle 40%-60% of the signal is retained for analysis, and the data is normalized to ensure the comparability of data from different channels.

[0108] S3: Perform time / frequency domain feature extraction on the optimized time-series signal to obtain multi-dimensional feature samples; and the multi-dimensional feature samples include the peak-to-peak ratio of the main frequency, energy of several frequency bands, frequency band energy ratio, wave intensity, spectral centroid, spectral entropy, spectral kurtosis, and time-domain envelope energy; in this embodiment, FFT and STFT techniques are used to extract time / frequency domain features to comprehensively characterize the differences in flow field features caused by changes in the number of blades;

[0109] Specifically, the peak frequency ratio is the ratio of the peak frequencies of 0-2Hz to 2-5Hz. Its expression is

[0110]

[0111] In the formula: Indicates the amplitude and frequency of the optimized timing signal; Indicates the frequency of the optimized timing signal;

[0112] The aforementioned frequency band energy includes the 0-2Hz frequency band energy. Energy in the 2-10Hz frequency band and energy in the 2-5Hz frequency band Its expression is

[0113]

[0114]

[0115]

[0116] In the formula: The time spectrum represents the optimized timing signal;

[0117] The frequency band energy ratio The expression is

[0118]

[0119] The fluctuation intensity The expression is

[0120]

[0121] The spectral centroid The expression is

[0122]

[0123] The spectrum entropy The expression is

[0124]

[0125] In the formula: Represents frequency The probability distribution;

[0126] The spectral kurtosis The expression is

[0127]

[0128] In the formula: Indicates the mean of the spectrum; Indicates the spectral variance;

[0129] The time-domain envelope energy The expression is

[0130]

[0131] In the formula: The spectrum of the signal envelope;

[0132] In this embodiment, (1) the peak ratio of the 0-2Hz to the 2-5Hz main frequency reflects the characteristics of the main vibration range of the signal; (2) the 0-2Hz energy reflects the proportion of the slowly changing low-frequency energy in the signal throughout the time domain, describing the slow change or drift trend of the system; (3) the 2-10Hz energy reflects which frequency band the signal energy is mainly distributed in. If it is concentrated in this range, it indicates that the vibration frequency is stable; (4) the 2-5Hz energy reflects the energy intensity of the main working frequency band in the signal, which is a measure of the effective vibration energy of the system; (5) the 0-2Hz to 2-5Hz energy ratio measures the proportion of high-frequency and low-frequency energy, reflecting whether the system vibration is affected by low-frequency energy. (6) 2-5Hz fluctuation, which measures the stability of the energy in this frequency band; (7) Spectral centroid, which represents the "center frequency" of the spectrum energy distribution. A higher centroid indicates more high-frequency components, while a lower centroid indicates more low-frequency components; (8) Spectral entropy, which measures the complexity of the spectrum energy distribution. Low entropy indicates that the energy is concentrated on a few frequencies, indicating that the signal is simple; high entropy indicates that the energy distribution is wide, indicating that the signal is complex and contains more noise; (9) Spectral kurtosis, which measures the "sharpness" of the spectrum shape. High kurtosis indicates that there is a strong single peak in the spectrum; low kurtosis indicates that the spectrum is flat; (10) Envelope energy, which reflects the trend of signal amplitude change over time;

[0133] S4: Based on the frequency domain feature principal component analysis method, perform feature dimensionality reduction processing on the multidimensional feature samples to obtain two-dimensional feature projection samples, specifically including the following steps:

[0134] S41: For the multidimensional feature samples Centralized processing is performed to obtain optimized samples;

[0135] And the expression for the centralization process is:

[0136]

[0137]

[0138]

[0139] In the formula: Indicates optimized sample; Represents the mean vector; It represents the mean of the same feature across all samples; Indicates the first The first multidimensional feature sample corresponding to the th _ ... n represents the number of multidimensional feature samples; p represents the number of features in each multidimensional feature sample.

[0140] S42: Calculate the covariance matrix based on the optimized samples, and the formula for calculating the covariance matrix is ​​as follows:

[0141]

[0142] In the formula: Represent the covariance matrix; Indicates transpose;

[0143] S43: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors;

[0144] And the eigenvalue decomposition expression of the covariance matrix is:

[0145]

[0146] In the formula: Represents the eigenvector; Represents the corresponding eigenvector eigenvalues;

[0147] The eigenvalues ​​obtained from eigenvalue decomposition are sorted in descending order to obtain a feature sequence list. The eigenvectors corresponding to the first two eigenvalues ​​in the feature sequence list are selected to form a two-dimensional mapping space for the data.

[0148] S44: Multidimensional feature samples Project onto a two-dimensional mapping space to obtain two-dimensional feature projection samples;

[0149] The formula for obtaining the feature projection samples is as follows:

[0150]

[0151] In the formula: Represents the feature projection sample; This represents a matrix composed of eigenvectors;

[0152] This embodiment employs Principal Component Analysis (PCA) algorithm. It achieves feature dimensionality reduction by centering the original signal data, calculating the covariance matrix, performing eigenvalue decomposition, and projective transformation. PCA, while minimizing information loss, uses linear transformation to "rotate" the data to a new coordinate system. Dimensionality reduction effectively removes redundant correlations between features, compressing ten-dimensional features into two principal components, facilitating low-dimensional visualization and efficient classification. PCA calculates the feature covariance matrix and extracts the two principal components with the largest variance contribution rates, mapping multi-dimensional features to a two-dimensional plane, achieving feature compression and visualization. The output is displayed visually as a heatmap. Figure 4 As shown;

[0153] S5: Classification of propeller blade numbers based on two-dimensional feature projection samples using the K-means clustering algorithm, specifically including the following steps:

[0154] S51: Randomly divide the two-dimensional feature projection samples Clusters of initial propeller blade counts , ,…, And obtain the mean value corresponding to the feature projection samples in each initial propeller blade number cluster as the centroid of the initial cluster;

[0155] S52: Calculate the Euclidean distance from the feature projection samples in each initial propeller blade number cluster to the centroid and perform clustering. Divide each feature projection sample into the initial propeller blade number cluster with the smallest sum of squared Euclidean distances from the centroid and obtain the latest cluster.

[0156] The Euclidean distance from the feature projection samples in each initial propeller blade number cluster to the centroid is calculated as follows:

[0157]

[0158] In the formula: Indicates the first Initial propeller blade number clusters; This represents the feature projection samples in the initial propeller blade number cluster; Indicates the first The mean value corresponding to the feature projection samples in each cluster;

[0159] S53: Obtain the mean value corresponding to the feature projection samples within each of the latest clusters, and use it as the latest centroid of the latest cluster;

[0160] S54: Obtain the current rate of change of the centroid based on the latest centroid and the centroid of the initial cluster, and determine whether the current rate of change has reached a preset index.

[0161] Furthermore, the preset indicators include the current rate of change being less than a preset threshold or reaching the maximum number of iterations;

[0162] If the preset target is met, then the latest cluster at this point is the optimal cluster.

[0163] If the preset target is to be achieved, repeat steps S52 to S53.

[0164] The optimal cluster is then used as the classification result for the number of propeller blades.

[0165] In this embodiment, the dimensionality-reduced principal component data, i.e., the two-dimensional feature projection samples, are input into the K-means clustering algorithm for automatic classification. The number of clusters is set to 3, corresponding to three-leaf, five-leaf, and seven-leaf paddle classes. The algorithm uses Euclidean distance as the similarity measure and optimizes clustering by minimizing the within-group squared error. The final output is visualized as a scatter plot. Figure 5 As shown, various samples form distinct clusters on the principal component plane, with clear cluster boundaries and accurate classification results. This embodiment verifies the results through experiments using propeller models with different numbers of blades arranged in a still water environment. Operating various propellers at different speeds, the sensor output signals, after being processed by the method described in this embodiment, can accurately distinguish the wake characteristics of different blade numbers. Experimental results show that the classification accuracy exceeds 98%, the signal analysis is stable, and the computational efficiency is high, verifying the feasibility and reliability of the system.

[0166] Compared with the prior art, the method described in this embodiment has the following advantages:

[0167] (1) A piezoelectric sensing unit with a biomimetic seal whisker structure is used to achieve high sensitivity to micro-disturbances in underwater flow fields by simulating the vibration response of marine mammal whiskers under fluid disturbance. Compared with traditional acoustic or visual detection methods, the piezoelectric biomimetic whisker sensor does not require the emission of sound waves or reliance on lighting conditions, and can work stably in turbid water and low visibility environments. Since the piezoelectric sensing layer has a linear voltage response to small deformations, the sensor can accurately capture the periodic vibration differences caused by different numbers of blades in the propeller wake, thereby obtaining a higher signal-to-noise ratio and dynamic sensitivity.

[0168] (2) A signal preprocessing and feature extraction method based on trend term removal, bandpass filtering, and time-frequency joint analysis is proposed: First, a smoothing method is used to remove low-frequency drift; second, a fourth-order bandpass Butterworth filter is designed to suppress noise interference; finally, FFT and STFT are combined to extract frequency domain and time-frequency domain features. The feature parameters include ten indicators such as frequency interval energy ratio, spectral centroid, spectral entropy, spectral kurtosis, and envelope energy distribution, which comprehensively characterize the flow field feature differences caused by the change in the number of blades. Compared with the existing single-dimensional methods based only on amplitude or dominant frequency features, this invention can achieve in-depth characterization of signal patterns from multiple perspectives such as energy distribution, spectral structure, and time stability, significantly improving feature discrimination.

[0169] (3) The PCA algorithm and K-means clustering algorithm are used for blade number classification. The PCA algorithm can effectively remove redundant correlations between features by reducing dimensionality, compressing the ten-dimensional features into two principal components, and realizing low-dimensional visualization and efficient classification. Subsequently, the K-means algorithm automatically classifies the samples based on the principle of minimizing Euclidean distance, and can realize the clustering identification of different blade numbers without supervised labels. Experimental results show that the system of the present invention is significantly better than the traditional amplitude-frequency feature threshold segmentation method for three-bladed, five-bladed, and seven-bladed propellers.

[0170] (4) The method described in this embodiment can form a heat map after normalizing the extracted multidimensional features, and can also visualize the results through PCA two-dimensional projection and K-means clustering, intuitively reflecting the differences in signal features of different blade numbers. The clustering results have clear boundaries and concentrated distribution, which verifies the effectiveness of feature design and dimensionality reduction algorithm, and provides a quantitative basis for subsequent model optimization and sensor array layout.

[0171] In summary, the method described in this embodiment constructs a complete technology chain from fluid disturbance perception to propeller number identification through a biomimetic seal whisker sensor, intelligent signal analysis, and clustering classification algorithms, achieving high-precision classification and intelligent identification of underwater propeller parameters. The method described in this embodiment uses frequency domain analysis for propeller number identification, avoiding the problems caused by changes in propeller speed and size, reducing noise interference, and significantly improving the robustness and accuracy of classification. It can be applied to the identification and tracking of underwater targets, and is particularly suitable for distinguishing and dynamically tracking fixed-type underwater vehicles performing specific tasks.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A blade number classification method based on principal component analysis of frequency domain features of a whisker sensor, characterized in that, The specific steps include: S1: By collecting data from propellers with different numbers of blades based on a strobe sensor, the analog sensing time-series electrical signals corresponding to the periodic fluid disturbances generated during operation at different speeds are obtained. S2: Preprocess the analog sensing timing electrical signal to obtain an optimized timing signal; The preprocessing includes performing analog-to-digital conversion on the analog sensing time-series electrical signal to obtain a digital signal; The digital signal is sequentially de-trended and filtered to obtain an optimized time-domain signal; S3: Extract time / frequency domain features from the optimized time-series signal to obtain multi-dimensional feature samples; Furthermore, the multidimensional feature samples include the peak-to-peak ratio of the main frequency, energy of several frequency bands, frequency band energy ratio, fluctuation intensity, spectral centroid, spectral entropy, spectral kurtosis, and temporal envelope energy; S4: Based on the frequency domain feature principal component analysis method, perform feature dimensionality reduction processing on the multidimensional feature samples to obtain two-dimensional feature projection samples; S5: Based on the K-means clustering algorithm, the propeller blade number is classified according to the two-dimensional feature projection samples.

2. The blade number classification method based on principal component analysis of frequency domain features of a whisker sensor as described in claim 1, characterized in that, S2 specifically includes the following steps: S21: Perform analog-to-digital conversion on the analog sensing time-series electrical signal to obtain a digital signal; S22: Perform detrending processing on the digital signal sequentially based on a sliding window to obtain a detrending signal; S23: Perform initial filtering processing on the detrending signal based on a 4th-order Butterworth bandpass filter to obtain an initial filtered signal; S24: Based on The filter performs bidirectional filtering on the initial filtered signal to eliminate phase distortion and obtain an optimized time-domain signal.

3. The blade number classification method based on principal component analysis of frequency domain features of a whisker sensor according to claim 2, characterized in that, The expression for detrending processing described in S2 is: In the formula: Represents digital signals; Indicates a smoothed signal; Indicates the window width; Indicates the time signature of a discrete signal; This represents the discrete signal beat after detrending; This represents the discrete signal after discretization of a digital signal. The discrete signal after the trend is the detrending signal.

4. The blade number classification method based on principal component analysis of frequency domain features of a whisker sensor according to claim 3, characterized in that, The peak frequency ratio mentioned in S3 is the peak frequency ratio of 0-2Hz to 2-5Hz. Its expression is In the formula: Indicates the amplitude and frequency of the optimized timing signal; Indicates the frequency of the optimized timing signal; The aforementioned frequency band energy includes the 0-2Hz frequency band energy. Energy in the 2-10Hz frequency band and energy in the 2-5Hz frequency band Its expression is In the formula: The time spectrum represents the optimized timing signal; The frequency band energy ratio The expression is The fluctuation intensity The expression is The spectral centroid The expression is The spectrum entropy The expression is In the formula: Represents frequency The probability distribution; The spectral kurtosis The expression is In the formula: Indicates the mean of the spectrum; Indicates the spectral variance; The time-domain envelope energy The expression is In the formula: This represents the spectrum of the signal envelope.

5. The blade number classification method based on principal component analysis of frequency domain features of a whisker sensor according to claim 4, characterized in that, S4 specifically includes the following steps: S41: For the multidimensional feature samples Centralized processing is performed to obtain optimized samples; And the expression for the centralization process is: In the formula: Indicates optimized sample; Represents the mean vector; It represents the mean of the same feature across all samples; Indicates the first The first multidimensional feature sample corresponding to the th _ ... n represents the number of multidimensional feature samples; p represents the number of features in each multidimensional feature sample. S42: Calculate the covariance matrix based on the optimized samples, and the formula for calculating the covariance matrix is ​​as follows: In the formula: Represent the covariance matrix; Indicates transpose; S43: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors; And the eigenvalue decomposition expression of the covariance matrix is: In the formula: Represents the eigenvector; Represents the corresponding eigenvector eigenvalues; The eigenvalues ​​obtained from eigenvalue decomposition are sorted in descending order to obtain a feature sequence list. The eigenvectors corresponding to the first two eigenvalues ​​in the feature sequence list are selected to form a two-dimensional mapping space for the data. S44: Multidimensional feature samples Project onto a two-dimensional mapping space to obtain two-dimensional feature projection samples; The formula for obtaining the feature projection samples is as follows: In the formula: Represents the feature projection sample; This represents a matrix composed of eigenvectors.

6. The blade number classification method based on principal component analysis of frequency domain features of a whisker sensor according to claim 5, characterized in that, S5 specifically includes the following steps: S51: Randomly divide the two-dimensional feature projection samples The initial propeller blade number clusters are determined, and the mean value corresponding to the feature projection sample in each initial propeller blade number cluster is obtained as the centroid of the initial cluster. S52: Calculate the Euclidean distance from the feature projection samples in each initial propeller blade number cluster to the centroid and perform clustering. Divide each feature projection sample into the initial propeller blade number cluster with the smallest sum of squared Euclidean distances from the centroid and obtain the latest cluster. S53: Obtain the mean value corresponding to the feature projection samples within each of the latest clusters, and use it as the latest centroid of the latest cluster; S54: Obtain the current rate of change of the centroid based on the latest centroid and the centroid of the initial cluster, and determine whether the current rate of change has reached a preset index. Furthermore, the preset indicators include the current rate of change being less than a preset threshold or reaching the maximum number of iterations; If the preset target is met, then the latest cluster at this point is the optimal cluster. If the preset target is to be achieved, repeat steps S52 to S53. The optimal cluster is then used as the classification result for the number of propeller blades.

7. The blade number classification method based on principal component analysis of frequency domain features of a whisker sensor according to claim 6, characterized in that, In S52, the Euclidean distance from the feature projection samples in each initial propeller blade number cluster to the centroid is calculated, and its expression is as follows: In the formula: Indicates the first Initial propeller blade number clusters; This represents the feature projection samples in the initial propeller blade number cluster; Indicates the first The mean value corresponding to the feature projection samples in each cluster.

Images