Temperature decoupling method and system based on sparse representation classifier quality measure

Abstract

The application discloses a temperature decoupling method and system based on sparse representation classifier quality measurement. First, under different ambient temperatures, load different mass blocks, apply sweep excitation to one side of the piezoelectric sheet, and collect the voltage signal of the other side of the piezoelectric sheet; the voltage signal is pretreated to obtain a structural frequency domain response signal; the small sample data set composed of the structural frequency domain response signal data is divided into a training set and a test set with the quality category as a label; a neural network model based on a sparse representation classifier is established, and the training set is input into the model for model training; finally, the test set is input, and a quality classification measurement result is output. The neural network model based on the sparse representation classifier has high frequency domain position information sensitivity, can accurately measure a small quality under a large temperature difference and a small sample, and realizes temperature decoupling.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of micro mass measurement, and mainly relates to a temperature decoupling method and system based on sparse representation classifier mass measurement. BACKGROUND

[0002] A resonant piezoelectric mass sensor uses a piezoelectric material as a driver and a sensor, measures the mass of a measured object by measuring and analyzing the structural response change (frequency or phase) caused by the loading mass, and has been widely applied in the fields of biology, medicine, medical equipment, chemistry, environmental molecular detection, etc.

[0003] Traditional micro mass measurement methods are mainly realized through an impedance analyzer or a complex frequency capture circuit. The resonant frequency will shift and the phase will deviate before and after the mass is loaded. The micro mass is measured by analyzing and calculating the deviation. However, there may be a large temperature difference in the actual measurement environment, which causes the quality factor of the sensor to decrease and the resonant frequency to shift, resulting in an increase in measurement error. Therefore, the traditional measurement method is difficult to realize temperature decoupling.

[0004] A neural network is an artificial intelligence method with strong feature extraction capability, which can actively extract deep feature information of an input signal for training and learning to solve the classification or regression problem of the signal. Therefore, temperature decoupling and mass measurement can be realized by combining a neural network. However, the convolutional neural network has low sensitivity to the position information of the signal and needs a large number of data samples. It is difficult to collect a large number of data samples under a large environmental temperature difference. Therefore, there is an urgent need for a method suitable for neural network technology, which can eliminate the temperature influence and meet the small data sample collection to improve the precision and accuracy of the micro mass measurement method. SUMMARY

[0005] The present application is aimed at the problem that temperature changes interfere with micro mass measurement and it is difficult to collect a large number of data samples under a large environmental temperature difference in the prior art. A temperature decoupling method and system based on sparse representation classifier mass measurement are provided. First, different masses are loaded under different environmental temperatures, a sweep excitation is applied to one side of the piezoelectric sheet, and the voltage signal of the other side of the piezoelectric sheet is collected. The voltage signal is preprocessed to obtain a structural frequency domain response signal. A small sample data set composed of the structural frequency domain response signal is divided into a training set and a test set with the mass category as a label. A neural network model based on a sparse representation classifier is established, the training set is input into the model for model training, and finally the test set is input to output the mass classification measurement result. The neural network model based on the sparse representation classifier has high frequency domain position information sensitivity, can accurately measure the micro mass under a large temperature difference and a small sample, and realizes temperature decoupling.

[0006] In order to achieve the above object, the technical scheme adopted by the present application is: a temperature decoupling method based on sparse representation classifier quality measurement, comprising the following steps:

[0007] S1, response voltage acquisition: under different environmental temperatures, load different masses, apply a sweep excitation to one side of the side piezoelectric sheet, and collect the voltage signal of the other side piezoelectric sheet;

[0008] S2, signal preprocessing: pre-process the voltage signal collected in step S1 to obtain a structural frequency domain response signal, and the preprocessing step at least includes phase-sensitive detection, Gaussian smoothing and resampling;

[0009] S3, data set division: the structural frequency domain response signal data obtained by different temperatures and different mass blocks in steps S1 and S2 are divided into a training set and a test set; the division is labeled by mass category;

[0010] S4, establishing a neural network model based on a sparse representation classifier: the neural network model based on the sparse representation classifier includes six layers, the first layer is a sample input layer, the second and third layers are full connection layers FC1 and FC2 respectively, the fourth layer is a sparse representation classifier SRC of neurons, the fifth layer is a full connection layer FC3 of neurons, and the sixth layer is an output layer; a Dropout layer is arranged between FC1 and FC2, between FC2 and SRC, and between FC3 and the output layer; the training set in step S3 is input into the neural network model based on the sparse representation classifier for model training;

[0011] S5, result output: the test set in step S3 is input into the neural network model based on the sparse representation classifier trained in step S4, and the mass classification measurement result is output.

[0012] As an improvement of the present application, the response voltage acquisition of step S1 is carried out on a double piezoelectric sheet cantilever beam micro mass measurement platform, and the measurement platform comprises: a fixed clamp, an elastic cantilever beam, an NI data acquisition card, an amplifier and a thermostat;

[0013] The elastic cantilever beam is fixedly placed in the thermostat by the fixed clamp;

[0014] The double piezoelectric sheets are parallelly pasted at the same positions on both sides of the elastic steel cantilever beam, and there is no relative movement between them;

[0015] The sweep excitation signal is sequentially amplified by the NI data acquisition card and the amplifier and then applied to the side piezoelectric sheet on one side, and the side piezoelectric sheet causes the elastic cantilever beam to vibrate under the action of the sweep excitation signal, and the response voltage signal of the other side piezoelectric sheet is collected by the NI data acquisition card.

[0016] As an improvement of the present application, the sweep excitation signal in step S1 is a sinusoidal sweep excitation signal, and the sweep range is 50Hz-5050Hz; the temperature regulation range of the thermostat is 0℃-65℃.

[0017] As another improvement of the present application, the phase-sensitive detection in step S2 is mainly realized by building a lock-in amplifier, specifically: the sinusoidal sweep excitation signal and the response voltage signal are input into a multiplier; at the same time, the signal with a phase difference of 90° from the sinusoidal sweep excitation signal and the response voltage signal are input into another multiplier; the two signals after multiplication are respectively passed through low-pass filters to obtain direct current signals, and the amplitude and phase information of the same frequency component in the response voltage signal as the sinusoidal sweep excitation signal are calculated according to the obtained direct current signals, so as to realize the extraction of the impedance characteristic information of the elastic cantilever beam structure.

[0018] As another improvement of the present application, the Gaussian smoothing factor in step S2 is 0.1.

[0019] As still another improvement of the present application, in the neural network model of the sparse representation classifier in step S4:

[0020] The input dimension of the first layer sample input layer is 5000;

[0021] The second layer full connection layer FC1 includes 8000 neurons, and the relu function is used as the activation function;

[0022] The third layer full connection layer FC2 includes 5000 neurons, and the relu function is used as the activation function;

[0023] The fourth layer sparse representation classifier SRC includes 500 neurons, and the sigmoid function is used as the activation function;

[0024] The fifth layer full connection layer FC3 includes 1000 neurons, and the relu function is used as the activation function;

[0025] The sixth layer output layer Output adopts the softmax classifier to output the quality classification result.

[0026] As still another improvement of the present application, in the model training of step S4, the neuron dropout rate is set to 10%.

[0027] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is: a temperature decoupling system based on a sparse representation classifier quality measurement, comprising a computer program, which, when executed by a processor, realizes the steps of the method as described above.

[0028] Compared with the prior art, the present application has the beneficial effects: the present application provides a temperature decoupling method based on sparse representation classifier quality measurement, and the phase-sensitive detection processing method adopted can calculate the amplitude and phase information of the same frequency component in the voltage signal as the sinusoidal sweep excitation signal from the original voltage signal through a lock-in amplifier, so as to quickly extract the impedance characteristic information of the elastic cantilever beam structure. The Gaussian smoothing processing can effectively remove the environmental noise interference in the frequency domain response signal, and significantly improve the signal-to-noise ratio. The neural network model based on the sparse representation classifier has higher frequency domain position information sensitivity compared with the convolutional neural network, and can quickly and accurately measure the micro quality under the condition of large environmental temperature difference and small sample, and realize temperature decoupling. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The step flow chart of the method of the present application is shown in the figure.

[0030] Figure 2 The schematic diagram of the double piezoelectric sheet cantilever beam micro mass measurement platform in step S1 of the method of the present application is shown in the figure.

[0031] Figure 3 The measurement program diagram of the host computer in the method of the present application is shown in the figure.

[0032] Figure 4 The phase-sensitive detection program diagram in step S2 of the method of the present application is shown in the figure.

[0033] Figure 5 The structural frequency domain response signal obtained after step S2 of the method of the present application is shown in the figure.

[0034] Figure 6 The neural network model structure diagram based on the sparse representation classifier in step S4 of the method of the present application is shown in the figure.

[0035] Figure 7 The test set confusion matrix result diagram is shown in the figure. DETAILED DESCRIPTION

[0036] The present application will be further illustrated in combination with the figures and specific embodiments, and it should be understood that the following specific embodiments are only used to illustrate the present application and not to limit the scope of the present application.

[0037] Example 1

[0038] In order to overcome the problem that temperature change interferes with micro mass measurement and it is difficult to collect a large number of data samples under large environmental temperature difference, the present application discloses a temperature decoupling method based on sparse representation classifier quality measurement, as shown in the figure. Figure 1 The method comprises the following steps:

[0039] Step S1, response voltage collection: under different environmental temperatures, load different mass of mass block, apply sweep excitation to one side of the side piezoelectric sheet, collect the voltage signal of the other side piezoelectric sheet;

[0040] The voltage model collection of this step is carried out on a double piezoelectric sheet cantilever beam micro-mass measurement platform, and the specific Figure 2 As shown in the figure, on the double piezoelectric sheet cantilever beam, two piezoelectric sheets made of the same size of lead zirconate titanate (PZT) material are respectively pasted in parallel with the epoxy structural adhesive at the same position on both sides of the elastic steel cantilever beam. The double piezoelectric sheets are the sine sweep excitation side piezoelectric sheet and the response voltage signal collection side piezoelectric sheet respectively. Due to the inverse piezoelectric effect, after a sine sweep excitation signal is applied to the end of the sweep excitation side piezoelectric sheet, the sweep excitation side piezoelectric sheet will deform, and the deformation degree is proportional to the excitation voltage. The sweep excitation side piezoelectric sheet vibration causes the cantilever structure to vibrate at the same frequency, and the structure vibration amplitude is proportional to the deformation degree of the sweep excitation side piezoelectric sheet.

[0041] At the same time, due to the positive piezoelectric effect, the signal collection side piezoelectric sheet is subjected to pressure brought by the cantilever structure vibration, and the signal collection side piezoelectric sheet generates a response voltage signal, and the output voltage is proportional to the cantilever structure amplitude.

[0042] Different masses are loaded at the fixed position of the cantilever beam end, and the response voltage signal at the end of the signal collection side piezoelectric sheet under different masses is measured for subsequent processing and measurement.

[0043] One end of the double piezoelectric sheet cantilever beam is clamped and fixed by a tool steel flat tongs, and the other end is loaded with a mass block of different mass size. The structure is placed in a constant temperature oven at a fixed position. The temperature of the constant temperature oven can be adjusted in the range of 0℃-65℃ to simulate different temperature measurement environments. During the operation process of the method of this embodiment, the response voltage signal of different masses under different temperatures is obtained by setting the temperature of the constant temperature oven.

[0044] Figure 3 The upper computer Simulink measurement program diagram. The sine sweep excitation signal sweep frequency range is 50Hz-5050Hz, which is generated by the upper computer Simulink program, and is applied to the end of the sweep excitation side piezoelectric sheet through the digital amplifier, the input channel of the NI data acquisition card, and the power amplifier in turn. At the same time, the input channel of the NI data acquisition card collects the response voltage signal at the end of the signal collection side piezoelectric sheet, and the frequency domain response signal is obtained after subsequent phase-sensitive detection processing. In the experiment, all the sampling frequencies f s are 50kHz.

[0045] Step S2, signal preprocessing: the voltage signal collected in step S1 is preprocessed to obtain a structural frequency domain response signal, and the preprocessing step at least includes phase-sensitive detection, Gaussian smoothing and resampling.

[0046] As shown in Figure 4 , the phase-sensitive detection algorithm is mainly realized by a phase-locked amplifier. There are two input signals in the phase-locked amplifier. In one signal, the sinusoidal sweep excitation signal is used as the reference signal and is multiplied with the response voltage signal in a multiplier. At the same time, in the other signal, the signal with a phase difference of 90° from the sinusoidal excitation signal is used as the reference signal and is multiplied with the response voltage signal in another multiplier. Finally, the multiplied two signals are respectively passed through low-pass filters to obtain direct current signals. According to the obtained direct current components, the amplitude and phase information of the same frequency component in the response voltage signal as the excitation signal can be calculated, so as to realize the extraction of the impedance characteristic information of the cantilever structure.

[0047] The frequency domain response amplitude signal sample obtained by the phase-sensitive detection algorithm is sequentially subjected to Gaussian smoothing with a smoothing factor of 0.1 and resampling processing to obtain a structural frequency domain response signal sample with a data length of 5000 as shown in Figure 5 .

[0048] In the general micro-mass measurement method based on the sparse representation classifier, the vibration frequency f n of the nth order modal of the cantilever structure is:

[0049]

[0050] wherein is the effective mass of the cantilever structure under the n-th vibration modal, and k is the elastic coefficient. When a certain mass is loaded at the end of the cantilever, the resonance frequency of the cantilever under each vibration modal will be shifted, and the frequency domain response amplitude will also change. In theory, the size of the end-loaded mass can be measured by analyzing the shift of the multi-order resonance frequency and the frequency domain response amplitude. However, in the real measurement process, the change of the environmental temperature will also cause the shift of the structural resonance frequency and the change of the frequency domain response amplitude, which will interfere with the micro-mass measurement. Therefore, the method of the present application proposes a network model based on the sparse representation classifier to realize temperature decoupling and accurate measurement of the micro-mass.

[0051] Step S3, data set division: the structural frequency domain response signal data obtained by the steps S1 and S2 of the mass block at different temperatures and different masses are divided into a training set and a test set; the division is labeled with the mass category.

[0052] The pre-processed structural frequency domain response signal samples are labeled with mass categories to form a small sample data set, which contains 140 sub-samples and is divided into 10 mass categories, including 0.1g, 0.2g, 0.3g, 0.4g, 0.5g, 0.6g, 0.7g, 0.8g, 0.9g, and 1g. Each mass type has 14 sub-samples at different temperatures, including 0℃, 5℃, 15℃, 20℃, 25℃, 30℃, 35℃, 40℃, 45℃, 50℃, 55℃, 60℃, and 65℃. The data set is divided into a training set and a test set in a 7:3 ratio; the training set is input into a neural network model based on a sparse representation classifier for training and learning.

[0053] Step S4, a neural network model based on a sparse representation classifier is established. As shown in Figure 6 The neural network model structure diagram based on the sparse representation classifier is shown in FIG. 4. The first layer is a sample input layer Input with an input dimension of 5000; the second layer is a fully connected layer FC1 with 8000 neurons, and the relu function is used as the activation function; the third layer is a fully connected layer FC2 with 5000 neurons, and the relu function is used as the activation function; the fourth layer is a sparse representation classifier SRC with 500 neurons, and the sigmoid function is used as the activation function; the fifth layer is a fully connected layer FC3 with 1000 neurons, and the relu function is used as the activation function; and the sixth layer is an output layer Output, which adopts a softmax classifier to output 10 mass classification results. A Dropout layer is arranged between FC1 and FC2, between FC2 and SRC, and between FC3 and Output, and the neuron dropout rate is set to 10% during the training process to prevent overfitting during the training.

[0054] Step S5, result output: the test set in step S3 is input into the neural network model based on the sparse representation classifier trained in step S4 for testing, and the mass classification measurement result is output.

[0055] In a specific case, the oven is used to simulate the real measurement environment temperature as described in step S1, the mass measurement device is placed in the oven, the temperature in the oven is adjusted, 10 different mass blocks are loaded at 14 different environmental temperatures, a sinusoidal sweep excitation signal is applied to the cantilever beam side piezoelectric sheet under each condition, and the response voltage signal of the other side piezoelectric sheet is measured; the obtained voltage signal is processed into a structural frequency domain response signal through the data preprocessing method described in step S2; the processed data samples are prepared into a small sample data set as described in step S3, and the training set and the test set are divided, then the training set is input into the neural network model based on the sparse representation classifier described in step S4 for training, finally the test set is input into the trained network model for testing, and the mass measurement result confusion matrix is output as shown in FIG. 5. Figure 7As shown, the average measurement accuracy is as high as 94.99%, the network can quickly realize accurate measurement of different quality under the interference of extreme environmental temperature, and has strong temperature decoupling ability.

[0056] In summary, the method adopts a double piezoelectric cantilever beam micro mass measurement platform and a small sample micro mass measurement method based on a sparse representation classifier for training and testing, realizes fast and accurate measurement of micro mass under large environmental temperature difference and small sample data set, and temperature decoupling, the proposed network model has high position information sensitivity to frequency domain signals, realizes fast extraction of structure frequency domain response information, has strong practicability, and can adapt to measurement conditions under bad temperature conditions.

[0057] It should be noted that the above content only illustrates the technical idea of the present application, and cannot limit the protection scope of the present application. For ordinary skilled persons in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which fall within the protection scope of the claims of the present application.

Claims

1. A temperature decoupling method based on sparse representation classifier quality measures, characterized in that, The method comprises the following steps: S1, response voltage collection: under different environmental temperatures, different masses are loaded, a sweep excitation is applied to the side piezoelectric sheet on one side, and the voltage signal of the piezoelectric sheet on the other side is collected; S2, signal preprocessing: the voltage signal collected in step S1 is preprocessed to obtain a structural frequency domain response signal, and the preprocessing step at least comprises phase-sensitive detection, Gaussian smoothing and resampling; S3, data set division: the structural frequency domain response signal data obtained by different temperatures and different mass blocks in steps S1 and S2 are divided into a training set and a test set; the division is labeled by mass category; S4, establishment of a neural network model based on a sparse representation classifier: the neural network model based on the sparse representation classifier comprises six layers, the first layer is a sample input layer, the second and third layers are full connection layers FC1 and FC2 respectively, the fourth layer is a sparse representation classifier SRC of neurons, the fifth layer is a full connection layer FC3 of neurons, and the sixth layer is an output layer; a Dropout layer is arranged between FC1 and FC2, between FC2 and SRC, and between FC3 and the output layer; the training set in step S3 is input into the neural network model based on the sparse representation classifier for model training; S5, result output: the test set in step S3 is input into the neural network model based on the sparse representation classifier trained in step S4, and a mass classification measurement result is output.

2. The method of claim 1, wherein: The response voltage collection in step S1 is performed on a double piezoelectric sheet cantilever beam micro-mass measurement platform, and the measurement platform comprises a fixed clamp, an elastic cantilever beam, an NI data acquisition card, an amplifier and a thermostat; The elastic cantilever beam is placed in the thermostat by the fixed clamp; The double piezoelectric sheets are parallelly pasted at the same positions on both sides of the elastic steel cantilever beam, and there is no relative movement between the two; The sweep excitation signal is sequentially amplified by the NI data acquisition card and the amplifier and then applied to the side piezoelectric sheet on one side; the side piezoelectric sheet causes the elastic cantilever beam to vibrate under the action of the sweep excitation signal, and the response voltage signal of the piezoelectric sheet on the other side is collected by the NI data acquisition card.

3. The method of claim 2, wherein the temperature decoupling is based on a quality measure of the sparse representation classifier. The sweep excitation signal in step S1 is a sinusoidal sweep excitation signal, and the sweep range is 50Hz-5050Hz; the temperature regulation range of the thermostat is 0℃-65℃.

4. The method of claim 3, wherein: The phase-sensitive detection in step S2 is realized by building a lock-in amplifier, specifically: the sinusoidal sweep excitation signal and the response voltage signal are multiplied in a multiplier; at the same time, the signal with a phase difference of 90° from the sinusoidal sweep excitation signal and the response voltage signal are multiplied in another multiplier; the two multiplied signals are respectively passed through low-pass filters to obtain direct current signals, the amplitude and phase information of the same frequency component in the response voltage signal as the sinusoidal sweep excitation signal are calculated according to the obtained direct current signals, and the extraction of the structural impedance characteristic information of the elastic cantilever beam is realized.

5. The method of claim 3, wherein: The Gaussian smoothing factor in step S2 is 0.

1.

6. The method of claim 5, wherein: In the neural network model based on the sparse representation classifier in step S4: the input dimension of the first layer sample input layer is 5000; The second layer of fully connected layer FC1 includes 8000 neurons, and a relu function as an activation function; The third layer of fully connected layer FC2 includes 5000 neurons, and a relu function as an activation function; The fourth layer of sparse representation classifier SRC includes 500 neurons, and a sigmoid function as an activation function; The fifth layer of fully connected layer FC3 includes 1000 neurons, and a relu function as an activation function; The sixth layer of output layer Output adopts a softmax classifier, and outputs a quality classification result.

7. The method of claim 6, wherein the temperature decoupling is based on a quality measure of the sparse representation classifier. In the model training of the step S4, a neuron dropout rate is set to 10%.

8. A temperature decoupling system based on sparse representation classifier quality measure comprising a computer program, characterized in that: The computer program is executed by a processor to implement the steps of the method in any one of claims 1-7.

Images