A wake classification and identification method based on triboelectric sensors

Abstract

The application discloses a wake classification and identification method based on a triboelectric sensor, and the method comprises the following steps: collecting a voltage time sequence signal through a fish tail-shaped triboelectric sensor; performing time domain filtering processing on the voltage time sequence signal, presetting corresponding wake category labels and different wake intensity level labels corresponding to the same wake category, and obtaining sample data; performing data segmentation on the sample data to obtain a data sample set; constructing a wake classification and identification model; performing model training on the constructed wake classification and identification model through the data sample set to obtain an optimal wake classification and identification model; and performing wake classification and identification and wake intensity grading according to the optimal wake classification and identification model. The method can effectively capture weak wake signal and suppress background noise, and can identify the nonlinear relationship and multi-scale time dependence hidden in the waveform.

Description

A method for wake classification and identification based on triboelectric sensors Technical Field

[0001] This invention relates to the field of fluid wake measurement technology, and in particular to a wake classification and identification method based on triboelectric sensors. Background Technology

[0002] Traditional fluid wake measurements typically rely on hot-film or fiber optic sensors. These sensors require external power and are susceptible to electromagnetic interference, affecting the measurement's response amplitude and frequency response range, and failing to effectively capture weak wake signals and suppress background noise. Furthermore, traditional wake feature recognition largely depends on manual feature extraction (such as FFT main frequency and energy ratio), limiting its feature representation capabilities. It also cannot identify implicit nonlinear relationships and multi-scale time dependencies within the waveform, resulting in insufficient accuracy and low efficiency in achieving automatic fluid wake classification. Summary of the Invention

[0003] This invention provides a wake classification and identification method based on a triboelectric sensor to overcome the above-mentioned technical problems.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows:

[0005] A wake classification and identification method based on triboelectric sensors specifically includes the following steps:

[0006] S1: The voltage timing signal corresponding to the Karman vortex street formed behind the obstacle when the fluid bypasses the obstacle under different flow velocity conditions is collected by the fishtail-shaped triboelectric sensor.

[0007] The voltage time-series signal is filtered in the time domain to obtain the time-domain filtered signal, and each time-domain filtered signal is labeled with the corresponding wake current category label and wake current intensity level label to form sample data.

[0008] S2: Based on a preset sliding window, perform data segmentation on the sample data to obtain a data sample set;

[0009] S3: Construct a wake turbulence classification and recognition model, which includes an input layer, a feature extraction module, and a prediction output layer;

[0010] The input layer is used to input sample data from the data sample set into the feature extraction module;

[0011] The feature extraction module includes a cascaded CNN network module and a cascaded LSTM network module with different kernel parameters; the cascaded CNN network module is used to extract the time-domain features of the signal in the time-domain filtered signal step by step; the cascaded LSTM network module is used to extract the time relationship features of the signal in the time-domain features of the signal step by step, and obtain the time relationship feature vector.

[0012] The prediction output layer is used to perform a fully connected operation on the temporal relationship feature vector based on the fully connected layer, and output the predicted wake category label;

[0013] S4: Construct a training loss function, train the constructed wake classification and recognition model using the data sample set to obtain the optimal wake classification and recognition model; perform wake classification and recognition and wake intensity grading based on the optimal wake classification and recognition model.

[0014] Furthermore, the fishtail-shaped triboelectric sensor constructed in S1 includes a fishtail-shaped flexible housing and a triboelectric sensing unit disposed within the fishtail-shaped flexible housing.

[0015] The triboelectric sensing unit includes a first flexible electrode and a second flexible electrode arranged vertically and symmetrically.

[0016] The bottom end of the first flexible electrode is provided with a first dielectric layer, and the top end of the second flexible electrode is provided with a second dielectric layer, and the first dielectric layer and the second dielectric layer are disposed opposite to each other.

[0017] A sealing gasket is provided between the edges of the first dielectric layer and the second dielectric layer to form a cavity structure with an air gap. When the first dielectric layer and the second dielectric layer on both sides of the cavity structure are subjected to the action of the Karman vortex street, the cavity structure will swing and deform to drive the first dielectric layer and the second dielectric layer to contact / separate, thereby obtaining the corresponding wake analog electrical signal.

[0018] Furthermore, the method for obtaining the voltage timing signal described in S1 is as follows:

[0019] The analog electrical signal of the wake is converted into a digital signal of the wake by a preset analog-to-digital conversion module; and the voltage timing signal is obtained by performing a Fourier transform on the digital signal of the wake using Matlab's Fast Fourier Transform.

[0020] Furthermore, the method for performing time-domain filtering on the voltage timing signal in S1 is as follows:

[0021] A second-order bandpass Butterworth filter is used to perform time-domain filtering on the voltage time-series signal to obtain the first-order filtered signal; based on The filter performs bidirectional filtering on the primary filtered signal to eliminate phase distortion and obtain a secondary filtered signal; the secondary filtered signal is then normalized to obtain a time-domain filtered signal.

[0022] Furthermore, the cascaded CNN network module in S3 includes a first CNN network layer, a second CNN network layer, and a third CNN network layer connected in sequence; the first CNN network layer is used to extract the signal temporal features from the temporal filtered signal to obtain a first signal feature; the second CNN network layer is used to extract the signal temporal features from the first signal feature to obtain a second signal feature; and the third CNN network layer is used to extract the signal temporal features from the second signal feature to obtain a third signal feature.

[0023] And the expressions for each CNN network layer in the cascaded CNN network module are as follows:

[0024]

[0025]

[0026] In the formula: Indicates data input; Represents the convolution kernel; Indicates bias; Indicates pooling operation; This represents the tensor obtained after convolution; Indicates the output data; express Activation function.

[0027] Furthermore, the cascaded LSTM network module described in S3 includes a first LSTM network layer and a second LSTM network layer connected in sequence; the first LSTM network layer is used to extract the signal time relationship features from the third signal features; the second LSTM network layer is used to further extract the signal time relationship features from the output of the first LSTM network layer to obtain a time relationship feature vector.

[0028] And the expressions for each CNN network layer of the cascaded LSTM network module are as follows:

[0029]

[0030]

[0031]

[0032]

[0033]

[0034]

[0035] In the formula: Indicates the Gate of Oblivion; express Activation function; This represents the output of a cascaded CNN network module; Indicates the forget gate bias; Indicates the forget gate weight; Indicates the input gate; Indicates the input gate weights; Indicates input gate bias; Indicates the state of candidate cells; The weights corresponding to the states of candidate cells; The bias representing the candidate cell state; This indicates a cell state update; Indicates the output gate; Indicates the output gate weights; Indicates output gate bias; express The hidden state at any given moment; express The hidden state at any given moment.

[0036] Furthermore, step S4 specifically includes the following steps:

[0037] S41: Divide the data sample set into training set, validation set and test set according to the preset ratio;

[0038] S42: Train the constructed wake classification and recognition model using the training set to obtain the trained wake classification and recognition model;

[0039] S43: Based on the constructed training loss function, the trained tailflow classification and recognition model is validated using a validation set;

[0040] And the training loss function The expression is

[0041]

[0042] In the formula: Indicates the number of samples; Indicates the number of category labels; Indicates the first The true category label of each sample data; Indicates the first Each sample data point corresponds to the unnormalized score of the true category; Indicates the first The unnormalized score of each sample data point on the kth category label;

[0043] That is, to determine whether the output of the trained wake classification and recognition model has converged;

[0044] If the output of the trained wake classification and recognition model converges, then the trained wake classification and recognition model is confirmed to be the optimal wake classification and recognition model.

[0045] Otherwise, the Adam optimizer is used to adaptively update the neural network parameters in the feature extraction module and step S42 is repeated until the neural network parameters of the trained wake classification and recognition model that has converged are confirmed to be the optimal neural network parameters and the wake classification and recognition model is reconstructed to obtain the optimal wake classification and recognition model.

[0046] S44: Perform wake classification and wake intensity grading on the test set based on the optimal wake classification and recognition model.

[0047] This invention provides a wake classification and identification method based on triboelectric sensors, with the following advantages:

[0048] (1) The present invention constructs a fish tail-shaped triboelectric sensor based on triboelectric sensor. Based on the principle of triboelectric charging and electrostatic induction, it realizes passive sensing. It can work stably for a long time without external power supply, which significantly improves the system reliability and energy self-consistency. By simulating the force characteristics of a real fish tail in the wake flow field, it is extremely sensitive to the periodic disturbance caused by vortex shedding. Compared with rigid probes or fixed piezoelectric sheets, the fish tail-shaped triboelectric sensor has a higher response amplitude and a wider frequency response range, which can effectively capture weak wake signals and suppress background noise.

[0049] (2) By performing time-domain filtering on the voltage timing signal, the signal-to-noise ratio is significantly improved, and the dominant frequency components and energy characteristics of the wake signal are preserved. Compared with traditional low-pass or high-pass filtering schemes, this invention can maintain the consistency of signal waveform characteristics under different working conditions and ensure data comparability.

[0050] (3) The wake classification and recognition model constructed in this invention uses CNN-LSTM neural network to automatically extract time-series features, which can identify the nonlinear relationships and multi-scale time dependencies hidden in the waveform, greatly improving the accuracy and efficiency of wake classification and recognition. Moreover, the model still has good generalization ability and stability under different flow velocities and sensor distances. Attached Figure Description

[0051] 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.

[0052] Figure 1 is a flowchart of the wake classification and identification method based on triboelectric sensors of the present invention;

[0053] Figure 2 is a schematic diagram of the fishtail-shaped triboelectric sensor in this embodiment;

[0054] Figure 3 is a time-domain waveform diagram of the original signal of the wake measured by the sensor in this embodiment;

[0055] Figure 4 shows the original frequency domain waveform of the wake measured by the sensor in this embodiment;

[0056] Figure 5 shows the time-domain waveform after signal filtering and truncation in this embodiment;

[0057] Figure 6 is a comparison of the time-domain waveforms of the original signal and the filtered signal in this embodiment;

[0058] Figure 7 is a flowchart of the neural network model classification and recognition process in this embodiment;

[0059] Figure 8 shows the classification and recognition accuracy of the neural network model in this embodiment.

[0060] Figure 9 is a confusion matrix diagram of the classification results of the neural network model in this embodiment.

[0061] In the figure: 1. Fishtail-shaped flexible shell; 2. Triboelectric sensing unit; 3. Flexible electrode; 4. First dielectric layer; 5. Second dielectric layer; 6. Sealing gasket. Detailed Implementation

[0062] 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.

[0063] This embodiment provides a wake classification and identification method based on triboelectric sensors, including the following steps:

[0064] S1: The voltage timing signal corresponding to the Karman vortex street formed behind the obstacle when the fluid bypasses the obstacle under different flow velocity conditions is collected by the fishtail-shaped triboelectric sensor.

[0065] The voltage time-series signal is filtered in the time domain to obtain the time-domain filtered signal, and each time-domain filtered signal is labeled with the corresponding wake current category label and wake current intensity level label to form sample data.

[0066] The fishtail-shaped triboelectric sensor described in this embodiment includes a fishtail-shaped flexible housing and a triboelectric sensing unit disposed within the fishtail-shaped flexible housing. The triboelectric sensing unit includes a first flexible electrode and a second flexible electrode arranged vertically symmetrically. The bottom end of the first flexible electrode is provided with a first dielectric layer, and the top end of the second flexible electrode is provided with a second dielectric layer, with the first dielectric layer and the second dielectric layer disposed opposite to each other. A sealing gasket is provided between the edges of the first dielectric layer and the second dielectric layer to form a cavity structure with an air gap. When the first dielectric layer and the second dielectric layer on both sides of the cavity structure are subjected to the action of the Karman vortex street, the cavity structure swings and deforms to drive the first dielectric layer and the second dielectric layer to contact / separate, thereby acquiring the corresponding wake analog electrical signal.

[0067] Specifically, the fishtail-shaped flexible shell, i.e., the flexible housing 1 of the triboelectric sensor, can be made of flexible, dense, and waterproof materials such as silicone or polydimethylsiloxane (PDMS). A shielding layer can be added between the flexible housing 1 and the triboelectric sensing unit 2 to reduce the interference of water charges on the signal of the triboelectric sensing unit. The triboelectric sensing unit 2 is mainly composed of a flexible electrode 3 and a dielectric layer. The flexible electrode 3 can be made of flexible electrodes such as conductive fabric doped with silver wire or screen-printed carbon electrodes; the materials of the first dielectric layer 4 and the second dielectric layer 5 are selected based on the triboelectric sequence, choosing two materials that are far apart in the sequence. For example, the first dielectric layer 4 is made of fluorinated ethylene propylene (FEP), and the second dielectric layer 5 is made of polyamide (Nylon). A double-sided adhesive gasket 6, less than 1 mm wide, is used within the triboelectric sensing unit 2 to separate the edges of the first dielectric layer 4 and the second dielectric layer 5, creating a certain air gap. The thickness of the sealing gasket 6 should be approximately 0.3 mm. Alternatively, a 1 mm wide polytetrafluoroethylene double-sided tape can be used, tightly adhering to the edge of the sensing unit. This contact-separation mode triboelectric sensor utilizes frictional charging and electrostatic induction between two materials with different electronegativity to generate current. Taking fluorinated ethylene propylene copolymer and polyamide as examples, they serve as dielectric layers, respectively. Contact charging and triboelectric charging are generated through the contact-separation of the two materials, inducing corresponding charges on the electrodes on both sides. When the sensor is placed in the wake current, under the excitation of the wake current, the flexible sensor bends and swings, causing the first dielectric layer 4 and the second dielectric layer 5 to come into contact and separate. The friction between them carries opposite charges, inducing opposite charges on their respective electrodes. When an external force causes the two layers to separate, a tiny gap is created. Because charges cannot jump across the gap, positive and negative charges are separated on either side of the gap. To balance this potential difference, free electrons in the external circuit connecting the back electrodes of the two friction layers begin to flow directionally under electrostatic induction, thus forming a current. When an external force brings the two materials into contact again, the gap disappears, and the potential difference caused by separation weakens or disappears. To regain charge balance, electrons in the external circuit flow in the opposite direction, generating a reverse current. This "contact-separation" cycle repeats continuously as long as a continuous reciprocating mechanical motion is provided, thus constantly outputting alternating current, i.e., a wake analog electrical signal.

[0068] The method for obtaining the voltage timing signal in this embodiment is as follows: the analog electrical signal of the wake is converted into a digital signal of the wake by a preset analog-to-digital converter module in STM32; and the digital signal of the wake is Fourier transformed by the Fast Fourier Transform in Matlab to obtain the voltage timing signal. Specifically, by comparing the signal-time frequency diagrams of the corresponding voltage timing signals under different operating conditions (different flow velocity environments and different types of obstacles), it can be seen that as the size of the corresponding obstacle of the same type increases, the shedding frequency of the Karman vortex street decreases and the wake period becomes longer. The main period of the sensor output signal increases accordingly, that is, the peak spacing increases, and the vortex shedding frequency in the wake around the obstacle of the same type increases. The relation approximately satisfies:

[0069]

[0070] in Represents the Strouhal number; Indicates the size corresponding to the category of obstacle (e.g., if the category of obstacle is a cylinder, the corresponding size is the cylinder diameter). This indicates the incoming velocity of the free fluid.

[0071] At the same flow rate, when the cylinder diameter As the frequency increases, the vortex shedding frequency The main period of the sensor signal decreases, T=1 / Therefore, as the diameter increases, the peak spacing in the corresponding time-domain waveform also increases. A larger diameter cylinder generates stronger wake kinetic energy in the flow field, resulting in greater force and displacement amplitude on the tail, thus increasing the peak voltage or charge of the triboelectric signal. The wake's action differs depending on the diameter; smaller cylinders generate more stable vortex streets, resulting in a relatively regular, near-sinusoidal signal waveform; while larger cylinders exhibit stronger wake energy and more complex flow structures, leading to asymmetrical peaks, distortion, or multi-peak characteristics in the signal waveform. For example, the tail-shaped triboelectric sensor measures wake signals from cylinders of different diameters with distinct characteristics, making them suitable for neural network models to extract these features for classification and recognition.

[0072] In this embodiment, to extract the voltage time-series signal collected by the fishtail-shaped triboelectric sensor with obvious characteristics, the method further includes time-domain filtering of the voltage time-series signal. This is achieved by performing time-domain filtering on the voltage time-series signal based on a second-order bandpass Butterworth filter to obtain a first-order filtered signal. The expression for the second-order bandpass Butterworth filter is as follows:

[0073]

[0074] In the formula: Indicates the center frequency. Indicates bandwidth. This indicates the high cutoff frequency (2Hz). This indicates the low cutoff frequency (0.1Hz).

[0075] Based on Matlab The filter performs bidirectional filtering on the primary filtered signal to eliminate phase distortion and reduce transient effects to obtain the secondary filtered signal; the secondary filtered signal is then normalized to obtain the time-domain filtered signal.

[0076] S2: Based on a preset sliding window, perform data segmentation on the sample data to obtain a data sample set;

[0077] In this embodiment, after obtaining a stable voltage timing signal by filtering the original voltage timing signal data, a sliding window is used to divide the data into sample data of the same size to obtain a data sample set.

[0078] S3: Construct a wake bleed classification and recognition model, which includes an input layer, a feature extraction module, and a prediction output layer. The input layer is used to input sample data from the data sample set into the feature extraction module. The feature extraction module includes a cascaded CNN network module and a cascaded LSTM network module with different kernel parameters. The cascaded CNN network module is used to extract the temporal features of the signal from the temporal filtered signal step by step. The cascaded LSTM network module is used to extract the temporal relationship features of the signal from the temporal features step by step, and obtain the temporal relationship feature vector. The prediction output layer is used to perform a fully connected operation on the temporal relationship feature vector based on the fully connected layer, and output the predicted wake bleed category label.

[0079] Specifically, the cascaded CNN network module includes a first CNN network layer, a second CNN network layer, and a third CNN network layer connected in sequence, with convolutional kernel sizes of 25, 15, and 7 for each layer, and the activation function of each layer using... The activation function and pooling layer size are set to 2; the first CNN network layer is used to extract the signal temporal features from the temporal filtered signal to obtain the first signal feature; the second CNN network layer is used to extract the signal temporal features from the first signal feature to obtain the second signal feature; the third CNN network layer is used to extract the signal temporal features from the second signal feature to obtain the third signal feature;

[0080] And the expressions for each CNN network layer in the cascaded CNN network module are as follows:

[0081]

[0082]

[0083] In the formula: Indicates data input; Represents the convolution kernel; Indicates bias; Indicates pooling operation; This represents the tensor obtained after convolution; Indicates the output data; express Activation function.

[0084] Following the cascaded CNN network module, a cascaded LSTM network module, namely a two-layer Long Short-Term Memory Recurrent Neural Network (LSTM), is connected. Randomly deactivated neurons are configured in the LSTM layers to ensure good generalization ability. Through the LSTM network layers, the model extracts signal temporal relationship features. The cascaded LSTM network module includes a first LSTM network layer and a second LSTM network layer connected sequentially. The first LSTM network layer is used to extract the signal temporal relationship features from the third signal features; the second LSTM network layer is used to further extract the signal temporal relationship features from the output of the first LSTM network layer to obtain a temporal relationship feature vector.

[0085] And the expressions for each CNN network layer of the cascaded LSTM network module are as follows:

[0086]

[0087]

[0088]

[0089]

[0090]

[0091]

[0092] In the formula: Indicates the Gate of Oblivion; express Activation function; This represents the output of a cascaded CNN network module; Indicates the forget gate bias; Indicates the forget gate weight; Indicates the input gate; Indicates the input gate weights; Indicates input gate bias; Indicates the state of candidate cells; The weights corresponding to the states of candidate cells; The bias representing the candidate cell state; This indicates a cell state update; Indicates the output gate; Indicates the output gate weights; Indicates output gate bias; express The hidden state at any given moment; express The hidden state at any given moment.

[0093] S4: Construct a training loss function, train the constructed wake classification and recognition model using the data sample set to obtain the optimal wake classification and recognition model; perform wake classification and recognition and wake intensity grading based on the optimal wake classification and recognition model, specifically including the following steps:

[0094] S41: Divide the data sample set into training set, validation set and test set according to the preset ratio;

[0095] S42: Train the constructed wake classification and recognition model using the training set to obtain the trained wake classification and recognition model;

[0096] S43: Based on the constructed training loss function, namely the cross-entropy loss function, the trained tailflow classification and recognition model is validated using a validation set. This loss function calculates the negative log probability of the true class based on the unnormalized score output by the network, thereby measuring the inconsistency between the predicted result and the true label. The expression is

[0097]

[0098] In the formula: Indicates the number of samples; Indicates the number of category labels; Indicates the first The true category label of each sample data; Indicates the first Each sample data point corresponds to the unnormalized score of the true category; Indicates the first The unnormalized score of each sample data point on the kth category label;

[0099] That is, to determine whether the output of the trained wake classification and recognition model has converged;

[0100] If the output of the trained wake classification and recognition model converges, then the trained wake classification and recognition model is confirmed to be the optimal wake classification and recognition model.

[0101] Otherwise, the Adam optimizer is used to adaptively update the neural network parameters in the feature extraction module and step S42 is repeated until the neural network parameters of the trained wake classification and recognition model that has converged are confirmed to be the optimal neural network parameters and the wake classification and recognition model is reconstructed to obtain the optimal wake classification and recognition model.

[0102] The method for adaptive gradient update of neural network parameters in the feature extraction module is as follows:

[0103] S431: Obtain the gradient of the neural network parameters in the feature extraction module at time step t. : ;in Indicates the parameters of the neural network The gradient operator; Represents the loss function; This represents the model parameters at time step t-1;

[0104] S432: Based on the gradient Obtaining first-order moment estimates and second-order moment estimates:

[0105]

[0106]

[0107] In the formula: This represents the first moment estimate for time step t; This represents the decay rate estimated by the first moment; This represents the first moment estimate at time step t-1; This represents the second-order moment estimate for time step t; This represents the decay rate estimated by the second moment; This represents the second moment estimate at time step t-1;

[0108] S433: Obtain the deviation between the first and second moments based on the first and second moment estimates; and adaptively update the network parameters of the neural network in the feature extraction module based on the deviation between the first and second moments.

[0109] And the expression for the deviation between the first and second moments is:

[0110]

[0111]

[0112] In the formula: This represents the first-order moment estimate after bias correction; This represents the second-order moment estimate after bias correction; This represents the t-th power of the decay rate estimated by the first moment; The t-th power represents the decay rate estimated by the second moment. The learning rate is dynamically adjusted to monitor the loss of the validation set. When the loss of the validation set fails to decrease for ten consecutive times, the learning rate is halved to escape local optima and oscillations.

[0113] S44: Perform wake classification and wake intensity grading on the test set based on the optimal wake classification and recognition model.

[0114] In this embodiment, an experimental sample with a cylinder (obstacle) diameter (obstacle size) of 1 cm, a fishtail-shaped triboelectric sensor 3 cm away from the cylinder, and a water flow velocity of 100 mm / s is used as an example. The time-domain waveform of the original voltage timing signal of the experimental sample is shown in Figure 2. In the figure, between 0.5 s and 12 s, the wake signal waveform is clear and stable, with a high signal-to-noise ratio and obvious periodicity; between 12 s and 30 s, the wake signal waveform noise (environmental noise, electromagnetic noise) is relatively obvious and does not have obvious periodicity; after 30 s, the noise is strong, and the signal shows an upward drift trend. The frequency domain waveform of the original signal of the experimental sample is shown in Figure 3. In the figure, the main frequency range of the signal is distributed between 0.1-2 Hz. The time-domain waveform of the filtered and truncated experimental sample signal is shown in Figure 4. In this embodiment, a second-order bandpass Butterworth filter (0.1-2 Hz) is used. Based on the original signal time-domain waveform, the period with obvious periodicity and high signal-to-noise ratio can be extracted. The experimental sample extracts the signal between 0.5 s and 12 s. By extracting signals during periods of high signal-to-noise ratio, the training accuracy of the neural network model can be improved. Figure 5 shows a comparison of the time-domain waveforms of the original and filtered signals. The flowchart of the wake bleed classification and recognition model used in this embodiment is shown in Figure 6. After filtering and extracting all experimental data, a sliding window is used to extract the data to generate suitable samples. The wake bleed classification and recognition model uses a 3-layer convolutional neural network (CNN) and a 2-layer long short-term memory recurrent neural network (LSTM), with the classification result finally output by a fully connected layer. The model is trained using a training set, the neural network hyperparameters are adjusted using a test set, and the model accuracy is obtained using a validation set. Figure 7 (left) shows the training set loss and test set loss of the neural network model, and Figure 7 (right) shows the training set accuracy and test set accuracy. In this embodiment, the model is trained using a training set, and the hyperparameters are adjusted using a validation set. The batch size is set to 30, the learning rate to 0.001, and the random deactivation to 0.5. The model accuracy is tested using a test set, achieving a classification and recognition accuracy of over 96.8% for the neural network model.

[0115] The beneficial effects of the method described in this embodiment are as follows: (1) Traditional fluid wake measurement usually relies on hot film or fiber optic sensors, which require external power supply and are susceptible to electromagnetic interference. The method described in this embodiment uses a triboelectric biomimetic fish tail sensor, which achieves passive sensing based on the principle of triboelectric generation and electrostatic induction. It can work stably for a long time without external power supply, significantly improving system reliability and energy self-sufficiency. Taking the fish tail-shaped triboelectric sensor as an example, it can simulate the force characteristics of a real fish tail in the wake flow field and is extremely sensitive to the periodic disturbances caused by vortex shedding. Compared with rigid probes or fixed piezoelectric sheets, the biomimetic fish tail has a higher response amplitude and a wider frequency response range, which can effectively capture weak wake signals and suppress background noise. (2) Through bandpass filtering, denoising and normalization, the method described in this embodiment significantly improves the signal-to-noise ratio and retains the dominant frequency components and energy characteristics of the wake. Compared with traditional low-pass or high-pass filtering schemes, the method described in this embodiment can maintain the consistency of signal waveform characteristics under different working conditions and ensure data comparability. (3) Traditional wake feature recognition often relies on manual feature extraction (such as FFT main frequency, energy ratio, etc.), which has limited feature expression capabilities. This invention uses a CNN-LSTM neural network to automatically extract temporal features, which can identify the nonlinear relationships and multi-scale time dependencies hidden in the waveform, and achieve high-precision automatic classification of cylinder diameter. Experimental verification shows that the model still has good generalization ability and stability under different flow velocities and sensor distances.

[0116] 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 wake classification and identification method based on triboelectric sensors, characterized in that, Specifically, the following steps are included: S1: Using a fishtail-shaped triboelectric sensor, the voltage timing signal corresponding to the Karman vortex street formed behind the obstacles when the fluid bypasses different types of pre-set obstacles under different flow velocity environments is collected; the voltage timing signal is processed by time-domain filtering to obtain a time-domain filtered signal, and each time-domain filtered signal is labeled with a corresponding wake category label and wake intensity level label to form sample data; the constructed fishtail-shaped triboelectric sensor includes a fishtail-shaped flexible shell and a triboelectric sensing unit disposed in the fishtail-shaped flexible shell; the triboelectric sensing unit includes a first flexible electrode and a second flexible electrode arranged vertically symmetrically; the bottom end of the first flexible electrode is provided with a first dielectric layer, and the top end of the second flexible electrode is provided with a second dielectric layer, and the first dielectric layer and the second dielectric layer are arranged opposite to each other; a sealing gasket is provided between the edges of the first dielectric layer and the second dielectric layer to form a cavity structure with an air gap. When the first dielectric layer and the second dielectric layer on both sides of the cavity structure are subjected to the action of the Karman vortex street, the cavity structure swings and deforms to drive the first dielectric layer and the second dielectric layer to contact / separate, and the corresponding wake analog electrical signal is obtained; S2: Based on a preset sliding window, perform data segmentation on the sample data to obtain a data sample set; S3: Construct a wake turbulence classification and recognition model, which includes an input layer, a feature extraction module, and a prediction output layer; the input layer is used to input the sample data from the data sample set into the feature extraction module; The feature extraction module includes a cascaded CNN network module and a cascaded LSTM network module with different kernel parameters. The cascaded CNN network module is used to extract the temporal features of the signal from the temporal filtered signal step by step. The cascaded LSTM network module is used to extract the temporal relationship features of the signal from the temporal features step by step, and obtain the temporal relationship feature vector. The prediction output layer is used to perform a fully connected operation on the temporal relationship feature vector based on the fully connected layer, and output the predicted wake category label. S4: Construct a training loss function, and train the constructed wake classification and recognition model with a data sample set to obtain the optimal wake classification and recognition model. Perform wake classification and recognition and wake intensity classification based on the optimal wake classification and recognition model.

2. The wake classification and identification method based on a triboelectric sensor according to claim 1, characterized in that, The method for obtaining the voltage timing signal described in S1 is as follows: the analog electrical signal of the wake is converted into a digital signal of the wake through a preset analog-to-digital conversion module; and the digital signal of the wake is Fourier transformed by the fast Fourier transform of Matlab to obtain the voltage timing signal.

3. The wake classification and identification method based on a triboelectric sensor according to claim 2, characterized in that, The method for time-domain filtering of voltage time-series signals in S1 is as follows: the voltage time-series signals are filtered in the time domain based on a second-order bandpass Butterworth filter to obtain a first-order filtered signal. based on The filter performs bidirectional filtering on the primary filtered signal to eliminate phase distortion and obtain a secondary filtered signal; the secondary filtered signal is then normalized to obtain a time-domain filtered signal.

4. The wake classification and identification method based on a triboelectric sensor according to claim 3, characterized in that, The cascaded CNN network module described in S3 includes a first CNN network layer, a second CNN network layer, and a third CNN network layer connected sequentially. The first CNN network layer is used to extract temporal features from the time-domain filtered signal to obtain a first signal feature. The second CNN network layer is used to extract temporal features from the first signal feature to obtain a second signal feature. The third CNN network layer is used to extract temporal features from the second signal feature to obtain a third signal feature. Furthermore, the expression for each CNN network layer in the cascaded CNN network module is: In the formula: Indicates data input; Represents the convolution kernel; Indicates bias; Indicates pooling operation; This represents the tensor obtained after convolution; Indicates the output data; express Activation function.

5. The wake classification and identification method based on a triboelectric sensor according to claim 4, characterized in that, The cascaded LSTM network module described in S3 includes a first LSTM network layer and a second LSTM network layer connected sequentially. The first LSTM network layer is used to extract the signal time relationship feature from the third signal feature. The second LSTM network layer is used to further extract the signal time relationship feature from the output of the first LSTM network layer to obtain a time relationship feature vector. The expression for each LSTM network layer of the cascaded LSTM network module is as follows: In the formula: Indicates the Gate of Oblivion; express Activation function; This represents the output of a cascaded CNN network module; Indicates the forget gate bias; Indicates the forget gate weight; Indicates the input gate; Indicates the input gate weights; Indicates input gate bias; Indicates the state of candidate cells; The weights corresponding to the states of candidate cells; The bias representing the candidate cell state; This indicates a cell state update; Indicates the output gate; Indicates the output gate weights; Indicates output gate bias; express The hidden state at any given moment; express The hidden state at any given moment.

6. The wake classification and identification method based on a triboelectric sensor according to claim 5, characterized in that, S4 specifically includes the following steps: S41: Dividing the data sample set into a training set, a validation set, and a test set according to a preset ratio; S42: Training the constructed wake bleed classification and recognition model using the training set to obtain the trained wake bleed classification and recognition model; S43: Validating the trained wake bleed classification and recognition model using the validation set based on the constructed training loss function; and the training loss function... The expression is In the formula: Indicates the number of samples; Indicates the number of category labels; Indicates the first The true category label of each sample data; Indicates the first Each sample data point corresponds to the unnormalized score of the true category; Indicates the first The unnormalized score of each sample data point on the k-th category label is used to determine whether the output of the trained wake classification and recognition model has converged. If the output of the trained wake classification and recognition model has converged, then the trained wake classification and recognition model is confirmed to be the optimal wake classification and recognition model. Otherwise, the Adam optimizer is used to adaptively update the neural network parameters in the feature extraction module and step S42 is repeated until the neural network parameters of the trained wake classification and recognition model with converged output are confirmed to be the optimal neural network parameters, and the wake classification and recognition model is reconstructed to obtain the optimal wake classification and recognition model. S44: Based on the optimal wake classification and recognition model, the test set is classified and the wake intensity is graded.

Images