CN111811617B - A liquid level prediction method based on short-time Fourier transform and convolutional neural network - Google Patents

Abstract

The invention discloses a non-contact liquid level prediction method based on a convolutional neural network. The method comprises the steps of firstly, obtaining a resonance wave data set required by an experiment through an acoustic resonance liquid level measuring instrument. And secondly, performing short-time Fourier transform on the time domain data set, and converting the one-dimensional time domain waveform signal into a two-dimensional frequency spectrum pattern form to be used as the input of the convolutional neural network. Then, a regression model is constructed based on the convolutional neural network, the mean square error is selected as a loss function, and an Adam function is selected as an optimization algorithm. And finally, inputting the training set into the constructed convolutional neural network model for modeling, and then inputting the test set into the trained convolutional neural network model to obtain the predicted liquid level height. The method realizes the mapping process from the original data to the task target by constructing a convolution neural network regression model; by iterative optimization of model parameters, the generalization capability of the convolutional neural network model and the accuracy of a prediction result are improved.

Description

A liquid level prediction method based on short-time Fourier transform and convolutional neural network

technical field

The invention belongs to the field of deep learning, and designs a liquid level prediction method based on short-time Fourier transform and convolutional neural network.

Background technique

Liquid level is a kind of material level, which refers to the level of liquid level in a closed container or an open container. The measurement of the liquid level is mainly by detecting the unique properties of the substances on both sides of the liquid level, or the changes of some of the same physical parameters (such as resistance, capacitance, pressure difference, sound speed, light energy, etc.) during the propagation process, so as to determine the liquid level height. method. According to whether the measuring device is in contact with the measured liquid, it can be divided into contact measurement method and non-contact measurement method. Commonly used non-contact measurement methods include radar, laser, ultrasonic and other measurement methods.

(1) Liquid level measurement method based on ultrasonic reflection: It has the characteristics of good directionality and easy operation, and has become one of the most commonly used methods. Its measurement principle is to transmit ultrasonic waves to the liquid surface and receive echoes, and multiply the speed of sound by the time of the round-trip range to calculate the distance from the liquid surface to the acoustic wave receiving device. However, in practical industrial applications, foreign objects such as foam, residues and deposits often appear on the surface of the liquid to be tested. When ultrasonic waves encounter these obstacles, parasitic reflections are prone to occur, changing the propagation path, which seriously affects the measurement effect and greatly reduces the measurement accuracy of ultrasonic waves.

(2) Method based on the wavelength of low-frequency sound waves: the wavelength of low-frequency sound waves is longer, and diffraction will occur when encountering obstacles, that is, sound waves can continue to propagate around obstacles, avoiding parasitic reflections. Based on the resonance principle of low-frequency sound waves, the liquid level height was converted from the initial resonance frequency (Resonance Frequency RF), and a better measurement effect was obtained. However, the maximum range of this method depends on the initial RF, and the minimum value of this frequency is limited by factors such as speaker principle, type, sound source volume and quality, and is generally only 20Hz. If measured at the standard sound speed c=331.45m/s, the maximum range is only 8.28m. Moreover, this also puts forward higher requirements on the sensitivity of the microphone, and the lowest audio frequency that can be sensed by a general microphone is about 20 Hz. These factors greatly limit the application of this method in long-distance measurement.

(3) In order to solve the problem of limited measurement distance in the low-frequency acoustic resonance measurement method, the liquid level measurement method based on the fixed-frequency acoustic resonance changes the transmitted low-frequency acoustic signal into a fixed-frequency sweep signal. The improved method does not need to extract the basic resonance frequency points, but instead extracts a series of resonance frequency points, and converts the distance between the measuring device and the liquid surface according to the difference between the adjacent resonance frequency points. The liquid level prediction method based on acoustic resonance deletes a large number of irrelevant original data in the process of feature extraction such as extraction, which reduces the amount of computation. However, this artificially set feature extraction only analyzes the data information from one angle of the signal. This method may obtain relatively optimal features, but discards a lot of original information. In a noisy environment, there will be many resonance points. The problem of selection or omission will greatly reduce the accuracy of liquid level measurement.

In recent years, deep learning has been widely used in the fields of image processing and natural language processing. Its "end-to-end" learning method is the most important aspect that distinguishes other algorithms. It makes the whole learning process free from artificial sub-problems. Instead, it is completely handed over to the deep learning model to directly learn the mapping from the original input to the desired output, and it is more likely to obtain the global optimal solution.

SUMMARY OF THE INVENTION

The purpose of the present invention is to propose a liquid level prediction method based on short-time Fourier transform and convolutional neural network in view of the shortcomings of the prior art.

The present invention includes the following steps:

Step (1) Get the dataset

The data used in the present invention are all collected by an acoustic resonance liquid level measuring instrument (abbreviated as a liquid level instrument). The loudspeaker of the equipment measuring system emits 1000 to 2500 Hz sine waves, and the microphone collects and transmits the resonance waves in real time and stores them. Select the complete and undisturbed acoustic wave data from the measured liquid level database, and use the corresponding measurement length as the data label.

Step (2) Feature extraction

The collected 5000 groups of original signals are subjected to short-time Fourier transform, so that the one-dimensional time-domain waveform signal is converted into a time-frequency spectrum of two-dimensional data; the mathematical formula of the short-time Fourier transform (STFT) is as follows :

Among them, y(n) is the audio signal, g(n) is the window function, f is the frequency, t is the time, e and π are the natural logarithmic base and pi respectively, both of which are constants.

Step (3) Build a Convolutional Neural Network

The present invention builds a three-layer convolutional neural network regression model; selects Adam as the optimization algorithm, and selects the mean square error MSE as the loss function.

Each layer of the network in the established convolutional neural network model includes a convolutional layer, a pooling layer, and a nonlinear activation layer. The convolution kernel size of the first layer of convolutional neural network convolution layer Conv1 is 5 × 5, the number of channels is 6, the stride is 1, the maximum pooling layer kernel size is 2, and the stride is 1; the second layer convolution The convolution kernel size of the convolutional layer Conv2 of the convolutional neural network is 5 × 5, the stride is 1, the number of channels is 16, the kernel size of the maximum pooling layer is 2, and the stride is 1; the third layer convolutional neural network volume The size of the convolution kernel of the convolutional layer Conv3 is 5×5, the stride is 1, the number of channels is 120, the kernel size of the maximum pooling layer is 2×2, the stride is 1, and the number of hidden nodes of the fully connected layer Fc1 is 120; The number of hidden nodes in the fully connected layer Fc2 is 84; the number of hidden nodes in the fully connected layer Fc3 is 1.

Step (4) Training and Evaluation of Convolutional Neural Networks

The 5000 sets of data collected were divided into training set, validation set and test set according to 8:1:1. First, input the training set data into the model established in step 3, and then input the test set into the trained convolutional neural network to output the predicted liquid level height.

Adam is an extension of the optimization algorithm stochastic gradient descent, which iteratively updates the network weights based on training data. In recent years, this algorithm has been widely used in the field of deep learning, especially in tasks such as computer vision and natural language processing. Therefore, Adam algorithm is selected as the optimization algorithm of the present invention.

In machine learning, the loss function is used to estimate the inconsistency between the predicted value of the model and the actual value. The smaller the loss function, the better the robustness of the model. It is the loss function that guides the model parameters. study.

The present invention selects the mean square error as the loss function, which is the mean value of the sum of squares of the difference between the predicted value and the label, and its formula is as follows:

是预测值，n表示样本总量。where y _i is the sample label of the ith sample,

is the predicted value, and n is the total sample size.

According to the above evaluation results, a convolutional neural network that takes into account prediction error and network complexity (parameter amount and calculation amount) is preferably used as the final model.

Beneficial effects of the present invention: The present invention adopts the acoustic resonance liquid level prediction method of a fixed frequency band, which can expand the liquid level measurement range, and converts each audio signal into a two-dimensional frequency domain through short-time Fourier transform as a convolutional neural network. Input; by building a convolutional neural network regression model, the mapping process from the original data to the task target is realized; by iterative optimization of the model parameters, the generalization ability of the convolutional neural network model and the accuracy of the prediction results are improved.

Description of drawings

Fig. 1. the flow chart of the present invention;

Figure 2. Liquid level prediction system based on acoustic resonance principle;

Figure 3. The time-domain waveform and frequency graph collected by the microphone;

Figure 4. Training loss curve.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings.

According to the algorithm flow chart shown in Figure 1, and combined with the actual measurement environment and the example of liquid level measurement, each step of the method is introduced in detail.

Step 1: Get the data

The prediction model data set is constructed on the acoustic resonance liquid level measurement device. As shown in Figure 2, the speaker plays a sine wave with a fixed frequency band, and the microphone collects the resonance wave data and stores the data in the built-in memory card of the processor. In the experiment, audio data with tags of 0.6m, 0.8m, 1.0m... For the original one-dimensional time domain waveform signal, filtering processing such as Gaussian average filtering and median filtering can be performed, and then min-max normalization processing is performed.

Step 2: Feature Extraction

According to prior knowledge and expert experience in related fields, the original signal is processed by short-time Fourier transform (STFT), and the one-dimensional time-domain waveform signal processed in step 1 is converted into a time-frequency domain spectrum with a two-dimensional structure Graphical form. The mean value, variance, and integral features of the spectrogram are extracted, which can reflect the characteristics of the frequency domain signal. The mathematical formula for the Short Time Fourier Transform (STFT) is as follows:

Among them, y(n) is the audio signal, and g(n) is the window function. f is the frequency, t is the time, e and π are the natural logarithmic base and pi respectively, both of which are constants. Figure 3 shows the time-domain waveform graph with the label 10 collected by the microphone and the time-frequency graph after short-time Fourier transform.

Step 3: Build the Convolutional Neural Network Model

The established three-layer convolutional neural network model. The designed convolutional neural network structure and its hyperparameters are shown in Table 1.

Table 1: CNN network architecture

Among them, "f" represents the convolution kernel, "s" is the stride, and "p" is the padding parameter.

Step 4: Training and Optimization of Convolutional Neural Networks

(1) Using the data set obtained in step 2, the convolutional neural network is trained and optimized under the pytorch deep learning framework, and a convolutional neural network model that can predict the liquid level height is obtained. In this embodiment, the height of the liquid level is a continuous value, so a regression model based on a convolutional neural network is constructed. The mean square error (MSE) was used to measure the difference between the predicted value and the true value.

(2) The neural network trains and tests the generated data set, and the effect evaluation of deep learning mainly depends on the training loss curve and the correct rate curve. Figure 4 is the training loss function curve with a learning rate of 0.001 and 20 iterations.

(3) Evaluation and Improvement of Deep Convolutional Neural Networks

Based on the classification accuracy, the performance of the above deep convolutional neural network model is evaluated.

Adjust the model hyperparameters (number of convolution layers, number of convolution kernels, type of nonlinear activation function, learning rate, etc.), retrain the model, and evaluate its performance.

According to the evaluation results, a structure that takes into account the accuracy and network complexity (parameter amount and calculation amount) is selected as the final model to improve the model representation ability and generalization ability. Table 2 randomly selects five groups of data from the training set and outputs the prediction results, and Table 3 randomly selects five groups of data from the test set and outputs the prediction results.

Table 2: Training set prediction results

Table 3: Test set prediction results

Claims (1)

1. a liquid level prediction method based on short-time Fourier transform and convolutional neural network, is characterized in that following steps: Step (1), get the dataset The data is collected by the acoustic resonance liquid level measuring instrument; the loudspeaker of the liquid level measuring instrument emits 1000 to 2500 Hz sine waves, and the microphone collects and returns the resonance waves in real time and stores them; select thousands of complete sets from the measured resonance wave data , undisturbed resonant waves, and use their corresponding measurement lengths as data labels to construct a data set; Step (2), feature extraction Perform short-time Fourier transform on all groups of resonant wave signals in the dataset, and convert one-dimensional time-domain waveform signals into two-dimensional spectrogram form; Step (3), build a convolutional neural network Build a three-layer convolutional neural network model; select Adam as the optimization algorithm and MSE as the objective function; Each layer of the network in the established convolutional neural network model includes a convolutional layer, a pooling layer, and a nonlinear activation layer; The convolution kernel size of the first layer of convolutional neural network convolution layer Conv1 is 5 × 5, the number of channels is 6, the stride size is 1, the maximum pooling layer kernel size is 2, and the stride size is 1; The convolution kernel size of the second layer of convolutional neural network convolution layer Conv2 is 5 × 5, the stride is 1, the number of channels is 16, the maximum pooling layer kernel size is 2, and the stride is 1; The third layer of convolutional neural network convolution layer Conv3 convolution kernel size is 5 × 5, the stride is 1, the number of channels is 120, the maximum pooling layer kernel size is 2 × 2, the stride is 1, The number of hidden nodes in the fully connected layer Fc1 is 120; the number of hidden nodes in the fully connected layer Fc2 is 84; the number of hidden nodes in the fully connected layer Fc3 is 1. Step (4) Training and Evaluation of Convolutional Neural Networks Divide all the data in the dataset into training set, validation set, and test set according to 8:1:1; input the training set data into the model established in step 3, and the validation set evaluates the model trained on the basis of the test set, and then the The test set is input into the trained convolutional neural network, and the predicted liquid level height is output.

Images