Industrial process vibration detection method and system based on lightweight convolutional neural network

Abstract

The present invention provides an industrial process oscillation detection method and system based on a lightweight convolutional neural network, which belongs to the field of oscillation detection. The method includes: determining industrial data simulation variables and generating simulation time series to construct a simulation data set; training the lightweight convolutional neural network according to the simulation data set to obtain an oscillation detection model; the lightweight convolutional neural network includes an input layer, a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a third maximum pooling layer, a fifth convolutional layer, a sixth convolutional layer, a fourth maximum pooling layer, a global average pooling layer and a fully connected layer connected in sequence; the oscillation detection model is used to determine the oscillation category of the industrial process corresponding to the industrial data to be detected. The present invention can effectively process noise sequences, multiple oscillations, intermittent oscillations, time-varying oscillations and non-stationary characteristics, and improve the accuracy and efficiency of oscillation detection.

Description

Technical Field

[0001] The present invention relates to the field of oscillation detection, and in particular to an industrial process oscillation detection method and system based on a lightweight convolutional neural network. Background Art

[0002] Industrial processes include process industries dominated by basic raw material industries such as petroleum, chemicals, steel, nonferrous metals, and building materials, and discrete industries dominated by equipment manufacturing industries such as machinery, electronics, automobiles, aerospace, rail transportation, and marine engineering. With the continuous advancement of science and technology and the rapid improvement of industrial automation levels, industrial processes are constantly developing towards scale and integration, which means that they have put forward higher requirements for safe, efficient, energy-saving, and high-quality production.

[0003] As the "brain" of the industrial production process, the control system is an indispensable part of ensuring good product quality, low material and energy consumption of the industrial process, and safe operation of the process. Therefore, it is necessary to effectively monitor the performance of the process control system. However, as time goes by, the characteristics of the production process and related equipment will change, which makes both traditional PID control and various advanced control strategies have performance degradation problems.

[0004] Oscillation is the main manifestation of controller performance deterioration and one of the most common problems in the process industry. If not handled in time, it may reduce product quality, increase the consumption of raw materials and energy, accelerate the aging of equipment, and even affect the safety of the process industry. Eliminating oscillation means reducing the volatility of process variables, which can bring stable economic benefits to the production process. The premise of eliminating oscillation is oscillation detection. A typical process industry may contain hundreds to thousands of control loops, which means that manual monitoring methods will consume a lot of manpower costs and easily lead to missed detection. On the contrary, automated detection technology is more suitable for application in industrial control process practice.

[0005] Existing methods that can be used for oscillation detection in industrial process control loops mainly include: oscillation detection algorithms based on time domain analysis, oscillation detection algorithms based on autocorrelation domain analysis, and oscillation detection algorithms based on signal decomposition.

[0006] The oscillation detection algorithm based on time domain analysis first performs statistical analysis on the time domain process data, then calculates relevant detection indicators (absolute error integral, period, regularity, etc.), and finally determines whether there is oscillation based on the calculated indicators; it is mainly based on the identifiable characteristics of process data in the time domain to design some indicators for detecting oscillation. However, the characteristics of process oscillation data are easily destroyed by noise and disturbance, which makes this type of algorithm fail.

[0007] The oscillation detection algorithm based on autocorrelation domain analysis first calculates the autocorrelation function (ACF) of the process data, then calculates the detection indicators designed based on ACF (ACF attenuation rate, ACF zero crossing point regularity, etc.), and finally determines whether there is oscillation based on the calculated indicators. Since this type of algorithm first transforms the process data from the time domain to the autocorrelation domain before analyzing it, it has better noise resistance. However, this type of algorithm is not stable enough and has difficulty in dealing with problems such as multiple oscillations.

[0008] Oscillation detection algorithms based on signal decomposition (empirical mode decomposition, intrinsic time scale decomposition, etc.) first perform time-frequency analysis on process data, then decompose the signal into several modes based on the analysis results, and finally evaluate the possibility of oscillation for each mode; this algorithm has good adaptability and can handle nonlinear and non-stationary characteristics. When the components of the signal are not complex, the oscillation components can usually be effectively separated. However, the components of process data are generally more complex and may contain intermittent oscillations and time-varying oscillation components. This makes it impossible for such algorithms to separate them well, resulting in problems of modal aliasing or splitting. In addition, there may be problems such as numerous redundant sub-signals and high computational complexity.

[0009] The main drawback of the above three oscillation detection methods is that they design certain rules based on partial features of process data to detect oscillations, which results in the algorithm's accuracy being reduced or even failing to work when these rules do not exist. In addition, each of them has problems such as poor noise resistance, unstable algorithms, low reliability, high computational complexity, and difficulty in effectively handling multiple oscillations, intermittent oscillations, time-varying oscillations, and non-stationary characteristics. Summary of the invention

[0010] The purpose of the present invention is to provide an industrial process oscillation detection method and system based on a lightweight convolutional neural network, which can improve the accuracy and efficiency of industrial process oscillation detection.

[0011] To achieve the above object, the present invention provides an industrial process oscillation detection method based on a lightweight convolutional neural network, comprising:

[0012] Determine industrial data simulation variables; the industrial data simulation variables include signal length, noise variance, disturbance amplitude, oscillation probability, number of cycles, waveform, smoothing factor, oscillation time-varying probability and frequency change factor;

[0013] Generate a simulation time series based on the industrial data simulation variable to construct a simulation data set; the simulation data set includes multiple simulation time series and an oscillation category of each simulation time series; the oscillation category is a non-oscillation category, a regular oscillation category, or an irregular oscillation category;

[0014] A lightweight convolutional neural network is trained according to the simulation data set to obtain an oscillation detection model; the lightweight convolutional neural network includes an input layer, a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a third maximum pooling layer, a fifth convolutional layer, a sixth convolutional layer, a fourth maximum pooling layer, a global average pooling layer and a fully connected layer connected in sequence;

[0015] According to the industrial data to be detected, the oscillation detection model is adopted to determine the oscillation category of the industrial process corresponding to the industrial data to be detected.

[0016] Optionally, generating a simulation time series based on the industrial data simulation variable specifically includes:

[0017] According to the signal length and the noise variance, two zero-mean Gaussian white noises are generated, namely preliminary noise data and preliminary disturbance data;

[0018] Using the preliminary disturbance data as an input of a pre-constructed first filter to smooth the preliminary disturbance data to obtain smoothed disturbance data;

[0019] Determining final disturbance data according to the smoothed disturbance data and the disturbance amplitude;

[0020] Determining the frequency of the non-frequency-converted oscillation data according to the signal length and the number of cycles;

[0021] Determining the frequency of the variable frequency oscillation data according to the frequency of the non-variable frequency oscillation data and the frequency change factor;

[0022] If the oscillation probability and the oscillation time-varying probability are both 1, determining preliminary oscillation data according to the waveform and the frequency of the variable frequency oscillation data;

[0023] constructing a second filter according to the smoothing factor, and using the preliminary oscillation data as an input of the second filter to smooth the preliminary oscillation data to obtain smoothed oscillation data;

[0024] Performing amplitude normalization processing on the smoothed oscillation data to obtain normalized oscillation data;

[0025] Determining a simulation time series according to the preliminary noise data, the final disturbance data and the normalized oscillation data;

[0026] If the oscillation probability is 1 and the oscillation time-varying probability is 0, determining preliminary oscillation data according to the waveform and the frequency of the non-frequency-varying oscillation data;

[0027] constructing a second filter according to the smoothing factor, and using the preliminary oscillation data as an input of the second filter to smooth the preliminary oscillation data to obtain smoothed oscillation data;

[0028] Performing amplitude normalization processing on the smoothed oscillation data to obtain normalized oscillation data;

[0029] Determining a simulation time series according to the preliminary noise data, the final disturbance data and the normalized oscillation data;

[0030] If the oscillation probability is 0, a time series is simulated according to the preliminary noise data and the final disturbance data.

[0031] Optionally, training a lightweight convolutional neural network according to the simulation data set to obtain an oscillation detection model specifically includes:

[0032] Using a two-dimensional coordinate mapping method to image each simulation time series in the simulation data set to obtain image data corresponding to each simulation time series;

[0033] Each image data is converted into a pixel matrix respectively, and each pixel matrix is ​​normalized to obtain a training sample set; the training sample set includes a normalized pixel matrix corresponding to each simulation time series and an oscillation category of each simulation time series;

[0034] Based on the training sample set, the stochastic gradient optimization algorithm and the classification cross entropy loss function, the lightweight convolutional neural network is optimized to obtain an oscillation detection model.

[0035] To achieve the above object, the present invention also provides an industrial process oscillation detection system based on a lightweight convolutional neural network, comprising:

[0036] A simulation variable determination module is used to determine industrial data simulation variables; the industrial data simulation variables include signal length, noise variance, disturbance amplitude, oscillation probability, number of cycles, waveform, smoothing factor, oscillation time-varying probability and frequency change factor;

[0037] A simulation data construction module, used for generating a simulation time series based on the industrial data simulation variables to construct a simulation data set; the simulation data set includes a plurality of simulation time series and an oscillation category of each simulation time series; the oscillation category is a non-oscillation category, a regular oscillation category or an irregular oscillation category;

[0038] A neural network training module, used for training a lightweight convolutional neural network according to the simulation data set to obtain an oscillation detection model; the lightweight convolutional neural network includes an input layer, a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a third maximum pooling layer, a fifth convolutional layer, a sixth convolutional layer, a fourth maximum pooling layer, a global average pooling layer and a fully connected layer connected in sequence;

[0039] The oscillation detection module is used to determine the oscillation detection result of the industrial data to be detected by using the oscillation detection model according to the industrial data to be detected.

[0040] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the present invention constructs a simulation data set based on the definition of oscillation, and detects industrial process oscillations from a visual perspective without specific rules, so that the detection results are more intuitive, the computational complexity is reduced, and the accuracy and reliability of oscillation detection are improved. Since the convolutional neural network has a powerful feature extraction capability, the present invention can effectively process noise sequences, multiple oscillations, intermittent oscillations, time-varying oscillations and non-stationary features. In addition, the present invention uses the VGG-16 network as the basic network structure to construct a lightweight convolutional network, which greatly reduces the computational complexity and improves the efficiency and accuracy of oscillation detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0042] Figure 1 A flow chart of an industrial process oscillation detection method based on a lightweight convolutional neural network provided by the present invention;

[0043] Figure 2 Schematic diagram of the construction process of the oscillation detection model;

[0044] Figure 3 Schematic diagram of the simulation time series of different oscillation categories in the simulation data set;

[0045] Figure 4 This is a schematic diagram of the structure of the base network VGG-16;

[0046] Figure 5 A schematic diagram of the structure of a lightweight convolutional neural network constructed for the present invention;

[0047] Figure 6 This is a performance evaluation result diagram of the oscillation detection model;

[0048] Figure 7 This is a schematic diagram of the time series of non-oscillatory and oscillatory categories in the industrial open source dataset;

[0049] Figure 8 This is a test result diagram of the oscillation detection model for the first type of oscillation data;

[0050] Fig. 9 This is a graph showing the test results of the oscillation detection model on the second type of oscillation data;

[0051] Fig.10 This is a graph showing the test results of the oscillation detection model on the third type of oscillation data;

[0052] Fig.11 This is the test result diagram of the oscillation detection model on non-oscillation data;

[0053] Fig.12 A schematic diagram of an industrial process oscillation detection system based on a lightweight convolutional neural network provided by the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] The present invention re-examines the definition of oscillation: Horch proposed a widely accepted definition of oscillation, in which he specified oscillation as a periodic change that is not completely hidden in the noise, in other words, a periodic change visible to the human eye. This definition suggests that loops containing oscillatory behavior are visually obvious. However, there is currently no prior art that follows the definition of oscillation to explore automatic oscillation detection technology at the visual level. Although artificial visual detection methods provide highly intuitive detection results, they are not recommended for engineering applications considering the labor cost and missed detection issues. As one of the representative networks of deep learning, convolutional neural network (CNN) provides a solution for detecting oscillations from a visual perspective.

[0056] The purpose of the present invention is to provide an industrial process oscillation detection method and system based on a lightweight convolutional neural network, so as to solve the problem that the existing algorithms are based on the characteristics of specific rules, which easily leads to low algorithm accuracy or failure when these rules do not exist; as well as the problems that the existing algorithms have poor noise resistance, unstable algorithms, high computational complexity and difficulty in effectively processing multiple oscillations, intermittent oscillations, time-varying oscillations and non-stationary characteristics.

[0057] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] Embodiment 1

[0059] like Figure 1 and Figure 2 As shown, this embodiment provides an industrial process oscillation detection method based on a lightweight convolutional neural network, including:

[0060] S1: Determine industrial data simulation variables. Specifically, the industrial data simulation variables are determined by the simulation tool, and the industrial data simulation variables include signal length, noise variance, disturbance amplitude, oscillation probability, number of cycles, waveform, smoothing factor, oscillation time-varying probability and frequency change factor, as shown in Table 1.

[0061] Table 1 Industrial data simulation variable table

[0062]

[0063] S2: Generate a simulation time series based on the industrial data simulation variable to construct a simulation data set. The simulation data set includes multiple simulation time series and an oscillation category of each simulation time series. The oscillation category is a non-oscillation category, a regular oscillation category, or an irregular oscillation category.

[0064] In this embodiment, both disturbance and noise are Gaussian white noise, and the noise variance N var and the disturbance amplitude D amp Determine the size of the disturbance and noise. The amplitude of all oscillating time series is 1 to ensure that the noise variance N var and the disturbance amplitude D amp effectiveness.

[0065] Specifically, we first initialize an empty time series and set t = [0, 1, ..., S L -1]. Step S2 comprises:

[0066] S21: According to the signal length S L and the noise variance N var , generate two zero-mean Gaussian white noises, which are the preliminary noise data D noiseAnd preliminary disturbance data D disturbance That is D noise and D disturbance The length is S L , the variance is N var .

[0067] S22: Using the preliminary disturbance data as the input of a pre-constructed first filter to smooth the preliminary disturbance data to obtain smoothed disturbance data.

[0068] The transfer function of the first filter is:

[0069]

[0070] Wherein, H1(s) is the transfer function of the first filter.

[0071] Specifically, the smoothing process of the preliminary disturbance data is as follows: First, the IIR digital filter design method is used to construct the transfer function: Since H1(s) here is continuous, but the preliminary disturbance data is discrete, the IIR digital filter design method is used for discretization.

[0072] Secondly, the preliminary disturbance data is Z-transformed and then multiplied by the transfer function of the first filter. The transfer function at this time is the discretized result.

[0073] Finally, perform an inverse Z transform on the multiplied result to obtain the result in the time domain, that is, the smoothed disturbance data.

[0074] S23: According to the smoothed disturbance data and the disturbance amplitude D amp , determine the final perturbation data

[0075]

[0076] S24: According to the signal length S L And the number of cycles N per , determine the frequency f0 of the non-frequency oscillation data: f0 = 1 / (S L / N per ).

[0077] S25: Determine the frequency of the variable frequency oscillation data according to the frequency of the non-variable frequency oscillation data and the frequency change factor.

[0078] Specifically, first, according to the frequency f0 of the non-frequency-converted oscillation data and the frequency change factor F cf , using the formula Determine the highest variable frequency f h and the minimum variable frequency f l, and then use the formula Determine the frequency f(t) of the variable frequency oscillation data at time t.

[0079] S26: If the oscillation probability O prob and the oscillation time-varying probability VF prob are all 1, then according to the waveform O wf and the frequency of the variable frequency oscillation data, determine the preliminary oscillation data O d The second filter is constructed according to the smoothing factor SF, and the preliminary oscillation data O d As the input of the second filter, the preliminary oscillation data O d Perform smoothing to obtain smooth oscillation data To smooth the oscillating data Perform amplitude normalization to obtain normalized oscillation data According to the preliminary noise data D noise , the final disturbance data And the normalized oscillation data Determine the simulation time series O ts .

[0080] Specifically, in step S26, the following formula is used to determine the preliminary oscillation data O at time t: d (t):

[0081]

[0082] Among them, sin() is the sine function, square() is the square wave function, and sawtooth() is the sawtooth wave function.

[0083] S27: If the oscillation probability O prob is 1, and the oscillation time-varying probability VF prob is 0, then according to the waveform O wf and the frequency f0 of the non-frequency-converted oscillation data, determine the preliminary oscillation data O d The second filter is constructed according to the smoothing factor SF, and the preliminary oscillation data O d As the input of the second filter, the preliminary oscillation data O d Perform smoothing to obtain smooth oscillation data To smooth the oscillating data Perform amplitude normalization to obtain normalized oscillation data According to the preliminary noise data D noise , the final disturbance data And the normalized oscillation data Determine the simulation time series O ts.

[0084] Specifically, in step S27, the following formula is used to determine the preliminary oscillation data O at time t: d (t):

[0085]

[0086] The transfer function of the second filter constructed in step S26 and step S27 is:

[0087]

[0088] Wherein, H2(s) is the transfer function of the second filter.

[0089] Specifically, the smoothing process of the preliminary oscillation data is as follows: First, the IIR digital filter design method is used to construct the transfer function: The second filter. Secondly, the preliminary oscillation data is Z-transformed and then multiplied with the transfer function of the second filter. Finally, the multiplication result is inversely Z-transformed to obtain the result in the time domain, that is, the smoothed oscillation data.

[0090] Using formula For smooth oscillating data Perform amplitude normalization.

[0091] In step S26 and step S27, the formula is used Determine the simulation time series.

[0092] S28: If the oscillation probability is 0, then according to the preliminary noise data D noise And the final disturbance data Determine the simulation time series O ts :

[0093] In this embodiment, a total of 10,000 simulation time series are generated to construct a simulation data set. Figure 3 The simulation dataset is used for training and performance evaluation of the oscillation detection model.

[0094] S3: Training a lightweight convolutional neural network according to the simulation data set to obtain an oscillation detection model.

[0095] The present invention first builds the basic network VGG-16 based on Keras and TensorFlow deep learning framework, and its structure is as follows Figure 4As shown in FIG, the structure of the base network is lightened by the model pruning operation to form a light convolutional neural network. The model pruning operation is as follows: first, the last convolutional layer in module 1, module 2, module 3, and module 4 is removed respectively; second, module 5 is removed; and finally, fully connected layer 1 and fully connected layer 2 are removed.

[0096] The lightweight convolutional neural network includes an input layer, a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a third maximum pooling layer, a fifth convolutional layer, a sixth convolutional layer, a fourth maximum pooling layer, a global average pooling layer and a fully connected layer, which are connected in sequence. Figure 5 As shown in the figure. The number of output channels of the first convolutional layer is 64, the number of data channels of the second convolutional layer is 128, the number of output channels of the third and fourth convolutional layers is 256, and the number of output channels of the fifth and sixth convolutional layers is 512. The convolution kernel size of all convolutional layers is 3×3, the convolution stride is 1, and ReLU activation is used. The pooling size of all maximum pooling layers is 2×2, and the pooling stride is 2. The number of hidden neuron units in the fully connected layer is 3.

[0097] Specifically, step S3 includes:

[0098] S31: Use a two-dimensional coordinate drawing method to image each simulation time series in the simulation data set to obtain image data corresponding to each simulation time series. That is, use the time information of the simulation time series as the X-axis coordinate and the amplitude information as the Y-axis coordinate to draw data points in a two-dimensional space coordinate system, and then connect these data points in sequence by straight lines. The imaging process must meet the following imaging rules: ① When the data points or cycles of the simulation time series are very long, distinguishable features can still be observed; ② The original structure of the data is not destroyed; ③ The resolution of the imaging should match the performance of the experimental equipment.

[0099] After the simulation time series is successfully drawn based on the above imaging rules, the imaged simulation time series is saved.

[0100] S32: Convert each image data into a pixel matrix respectively, and normalize each pixel matrix to obtain a training sample set. The training sample set includes a normalized pixel matrix corresponding to each simulation time series and an oscillation category of each simulation time series.

[0101] Specifically, the present invention uses the function in Keras to realize the mapping from image to pixel matrix.

[0102] Since the pixel values ​​in the pixel matrix are between 0 and 255, it is not suitable for convolutional neural network training. Because CNN usually uses smaller weights for fitting. When the values ​​of the training data are larger integer values, it will slow down the training process and also cause overfitting problems. Therefore, the present invention uses the following formula to normalize the pixel matrix:

[0103]

[0104] Among them, X (i) is the i-th pixel matrix, is the i-th normalized pixel matrix.

[0105] S33: Based on the training sample set, the stochastic gradient optimization algorithm and the classification cross entropy loss function, the lightweight convolutional neural network is optimized to obtain an oscillation detection model.

[0106] Specifically, the above training sample set is randomly divided into a training set and a verification set in a 4:1 manner to train and verify the lightweight convolutional neural network.

[0107] After setting the hyperparameters such as the number of convolution output channels, the number of hidden neuron units, the activation function, the loss function, the optimizer, the learning rate, and the number of training iterations, the lightweight convolutional neural network is trained using the training set. In this embodiment, the learning rate is set to 0.001, the training batch size is 4, the softmax classification cross entropy is selected as the loss function, and the stochastic gradient optimization (Stochastic Gradient Descent, SGD) is used as the optimizer.

[0108] The training process is divided into two processes: forward and reverse. In the forward process, the output result after the data undergoes a series of linear and nonlinear transformations can be expressed as:

[0109]

[0110] Among them, Z (i) is the result of linear and nonlinear transformation of the ith normalized pixel matrix, F(·) represents a series of linear and nonlinear operations, K is the number of oscillation categories, is the result of the i-th normalized pixel matrix corresponding to the first oscillation category.

[0111] The last step of the forward process is the softmax classification operation, which converts the output of the neuron into a probability distribution of the category. The prediction result after softmax conversion can be expressed as:

[0112]

[0113] in, is the probability that the i-th normalized pixel matrix is ​​of the j-th oscillation category.

[0114] The backward process is the process of gradient descent and weight parameter update. It is connected with the forward process through the loss function The specific operations are as follows:

[0115]

[0116] Among them, θ is all relevant parameters of the lightweight convolutional neural network (such as weight vector and bias term), y (i) yes The label of , that is, the oscillation category of the i-th normalized pixel matrix (using one-hot encoding), N is the number of normalized pixel matrices.

[0117] The process of gradient optimization using SGD can be expressed as:

[0118]

[0119] θ m+1 =θ m +v m+1 ;

[0120] Among them, v m+1 is the velocity vector of the objective function during the m+1th optimization, v m is the velocity vector of the objective function during the mth optimization, γ is the momentum term, γ is set to 0.9, η m is the learning rate, θ m is the parameter of the lightweight convolutional neural network during the mth optimization, θ m+1 are the parameters of the lightweight convolutional neural network during the m+1th optimization.

[0121] After sufficient iterations in the forward and backward propagation stages (such as 100 iterations of training), the network learning is stopped, and the model is automatically saved during the learning period.

[0122] Then the validation set is input into the oscillation detection model to evaluate the performance of the oscillation detection model. The evaluation index uses the confusion matrix, as shown in Table 2.

[0123] Table 2 Confusion matrix

[0124] Predicted as positive sample Predicted as negative sample True positive sample TP FN True negative samples FP TN

[0125] Among them, TP, TN, FP and FN represent the number of correctly predicted positive samples, the number of correctly predicted negative samples, the number of incorrectly predicted positive samples, and the number of incorrectly predicted negative samples, respectively.

[0126] The performance evaluation results of the oscillation detection model are shown in Figure 6 As shown, Figure 6 The 0 in the horizontal and vertical coordinates represents the non-oscillation data in the simulation data set, 1 represents the regular oscillation data in the simulation data set, and 2 represents the irregular oscillation data in the simulation data set.

[0127] Furthermore, the industrial open source dataset ISDB is obtained from the literature "Dambros JWV, Trierweiler JO, Farenzena M. Industrial datasets and a tool for SISO control loops data visualization and analysis [J]. Computers & Chemical Engineering, 2021, 146: 107198." Figure 7 ISDB is used to test the oscillation detection effect of the oscillation detection model.

[0128] Multiple industrial data are randomly extracted from ISDB and input into the oscillation detection model after imaging, pixel matrix conversion and normalization. After the forward process processing, the output result M of the forward process can be obtained. o =[p1,p2,...,p j ,...,p K ]. Among them, M o is a vector containing the predicted probabilities of each oscillation category, p j is the probability that the extracted industrial data is the jth oscillation category.

[0129] Then, by obtaining the index value IV with the maximum probability, the oscillation category to which the extracted industrial data belongs can be determined. The operation can be expressed as:

[0130]

[0131] The test results of the oscillation detection model constructed by the present invention on oscillation data and non-oscillation data are as follows: Figures 8 to 11 shown.

[0132] S4: According to the industrial data to be detected, the oscillation detection model is used to determine the oscillation category of the industrial process corresponding to the industrial data to be detected.

[0133] In addition, the oscillation detection model can be updated with additional training data to handle new loop characteristics. The present invention does not require prior knowledge of the loop or process to be detected. The oscillation detection model will automatically infer whether there is oscillation based on the input data, and has high adaptability, thereby solving the problem of instability and low reliability of some existing technologies.

[0134] Compared with the existing oscillation detection algorithms, the present invention starts from the definition of oscillation and establishes a lightweight convolutional neural network. It detects industrial process oscillations from a visual perspective without specific rules, making the detection results more intuitive, reducing the computational complexity, and improving the accuracy and reliability of oscillation detection. Since the convolutional neural network has a powerful feature extraction capability, the present invention can effectively process noise sequences, multiple oscillations, intermittent oscillations, time-varying oscillations, and non-stationary features, solving the problem that the prior art is difficult to handle the above features. In addition, the present invention uses the VGG-16 network as the base network structure and uses model pruning technology to construct a lightweight convolutional network, which greatly reduces the computational complexity, improves the efficiency of oscillation detection, and solves the problem of high computational complexity of the prior art.

[0135] Embodiment 2

[0136] In order to execute the method corresponding to the above-mentioned embodiment 1 to achieve the corresponding functions and technical effects, an industrial process oscillation detection system based on a lightweight convolutional neural network is provided below.

[0137] like Fig.12 As shown, the industrial process oscillation detection system based on lightweight convolutional neural network provided in this embodiment includes: a simulation variable determination module 21, a simulation data construction module 22, a neural network training module 23 and an oscillation detection module 24.

[0138] The simulation variable determination module 21 is used to determine industrial data simulation variables, including signal length, noise variance, disturbance amplitude, oscillation probability, number of cycles, waveform, smoothing factor, oscillation time-varying probability and frequency change factor.

[0139] The simulation data construction module 22 is used to generate a simulation time series based on the industrial data simulation variables to construct a simulation data set. The simulation data set includes multiple simulation time series and an oscillation category of each simulation time series. The oscillation category is a non-oscillation category, a regular oscillation category, or an irregular oscillation category.

[0140] The neural network training module 23 is used to train the lightweight convolutional neural network according to the simulation data set to obtain an oscillation detection model. The lightweight convolutional neural network includes an input layer, a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a third maximum pooling layer, a fifth convolutional layer, a sixth convolutional layer, a fourth maximum pooling layer, a global average pooling layer and a fully connected layer connected in sequence.

[0141] The oscillation detection module 14 is used to determine the oscillation detection result of the industrial data to be detected by using the oscillation detection model according to the industrial data to be detected.

[0142] Compared with the prior art, the industrial process oscillation detection system based on a lightweight convolutional neural network provided in this embodiment has the same beneficial effects as the industrial process oscillation detection method based on a lightweight convolutional neural network provided in Example 1, which will not be repeated here.

[0143] Embodiment 3

[0144] This embodiment provides an electronic device, including a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the industrial process oscillation detection method based on a lightweight convolutional neural network of embodiment one.

[0145] Optionally, the above-mentioned electronic device may be a server.

[0146] In addition, an embodiment of the present invention further provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the industrial process oscillation detection method based on a lightweight convolutional neural network of embodiment one.

[0147] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0148] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for industrial process oscillation detection based on lightweight convolutional neural network, characterized in that: The industrial process oscillation detection method based on lightweight convolutional neural network includes: Determine industrial data simulation variables; the industrial data simulation variables include signal length, noise variance, disturbance amplitude, oscillation probability, number of cycles, waveform, smoothing factor, oscillation time-varying probability and frequency change factor; Generate a simulation time series based on the industrial data simulation variable to construct a simulation data set; the simulation data set includes multiple simulation time series and an oscillation category of each simulation time series; the oscillation category is a non-oscillation category, a regular oscillation category, or an irregular oscillation category; Generating a simulation time series based on the industrial data simulation variable specifically includes: According to the signal length and the noise variance, two zero-mean Gaussian white noises are generated, namely preliminary noise data and preliminary disturbance data; Using the preliminary disturbance data as an input of a pre-constructed first filter to smooth the preliminary disturbance data to obtain smoothed disturbance data; Determining final disturbance data according to the smoothed disturbance data and the disturbance amplitude; Determining the frequency of the non-frequency-converted oscillation data according to the signal length and the number of cycles; Determining the frequency of the variable frequency oscillation data according to the frequency of the non-variable frequency oscillation data and the frequency change factor; If the oscillation probability and the oscillation time-varying probability are both 1, determining preliminary oscillation data according to the waveform and the frequency of the variable frequency oscillation data; constructing a second filter according to the smoothing factor, and using the preliminary oscillation data as an input of the second filter to smooth the preliminary oscillation data to obtain smoothed oscillation data; Performing amplitude normalization processing on the smoothed oscillation data to obtain normalized oscillation data; Determining a simulation time series according to the preliminary noise data, the final disturbance data and the normalized oscillation data; If the oscillation probability is 1 and the oscillation time-varying probability is 0, determining preliminary oscillation data according to the waveform and the frequency of the non-frequency-varying oscillation data; constructing a second filter according to the smoothing factor, and using the preliminary oscillation data as an input of the second filter to smooth the preliminary oscillation data to obtain smoothed oscillation data; Performing amplitude normalization processing on the smoothed oscillation data to obtain normalized oscillation data; Determining a simulation time series according to the preliminary noise data, the final disturbance data and the normalized oscillation data; If the oscillation probability is 0, determining a simulation time series according to the preliminary noise data and the final disturbance data; A lightweight convolutional neural network is trained according to the simulation data set to obtain an oscillation detection model; the lightweight convolutional neural network includes an input layer, a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a third maximum pooling layer, a fifth convolutional layer, a sixth convolutional layer, a fourth maximum pooling layer, a global average pooling layer and a fully connected layer connected in sequence; According to the industrial data to be detected, the oscillation detection model is adopted to determine the oscillation category of the industrial process corresponding to the industrial data to be detected.

2. The industrial process oscillation detection method based on lightweight convolutional neural network according to claim 1 is characterized in that: The transfer function of the first filter is: Where H1(s) is the transfer function of the first filter; The final perturbation data is determined using the following formula: in, To smooth the disturbance data, is the final perturbation data, D amp is the disturbance amplitude.

3. The industrial process oscillation detection method based on lightweight convolutional neural network according to claim 1 is characterized in that: Using the formula f0 = 1 / (S L / N per ) determining the frequency of the non-frequency-converted oscillation data; According to the frequency and frequency change factor of the non-frequency oscillation data, the formula is used Determine the highest variable frequency and the lowest variable frequency using the formula Determine the frequency of variable frequency oscillation data; Where f0 is the frequency of the non-frequency oscillation data, f(t) is the frequency of the frequency oscillation data at time t, S L is the signal length, N per is the number of cycles, f h is the highest variable frequency, f l is the minimum variable frequency, F cf is the frequency change factor.

4. The industrial process oscillation detection method based on lightweight convolutional neural network according to claim 1 is characterized in that: When the oscillation probability and the oscillation time-varying probability are both 1, the following formula is used to determine the preliminary oscillation data: When the oscillation probability is 1 and the oscillation time-varying probability is 0, the following formula is used to determine the preliminary oscillation data: Among them, O d (t) is the preliminary oscillation data at time t, f(t) is the frequency of the variable frequency oscillation data at time t, f0 is the frequency of the non-variable frequency oscillation data, O wf is a waveform, sin() is a sine function, square() is a square wave function, and sawtooth() is a sawtooth wave function.

5. The industrial process oscillation detection method based on lightweight convolutional neural network according to claim 1 is characterized in that: The transfer function of the second filter is: Where H2(s) is the transfer function of the second filter, and SF is the smoothing factor; The smooth oscillation data is normalized using the following formula: in, is the normalized oscillation data, To smooth out oscillating data.

6. The industrial process oscillation detection method based on lightweight convolutional neural network according to claim 1 is characterized in that: When the oscillation probability is 1, the simulation time series is determined using the following formula: When the oscillation probability is 0, the simulation time series is determined using the following formula: Among them, O ts is the simulation time series, D noise is the preliminary noise data, is the final perturbation data, is the normalized oscillation data.

7. The industrial process oscillation detection method based on lightweight convolutional neural network according to claim 1 is characterized in that: The lightweight convolutional neural network is trained according to the simulation data set to obtain an oscillation detection model, which specifically includes: Using a two-dimensional coordinate mapping method to image each simulation time series in the simulation data set to obtain image data corresponding to each simulation time series; Each image data is converted into a pixel matrix respectively, and each pixel matrix is ​​normalized to obtain a training sample set; the training sample set includes a normalized pixel matrix corresponding to each simulation time series and an oscillation category of each simulation time series; Based on the training sample set, the stochastic gradient optimization algorithm and the classification cross entropy loss function, the lightweight convolutional neural network is optimized to obtain an oscillation detection model.

8. An industrial process oscillation detection system based on a lightweight convolutional neural network, applied to the industrial process oscillation detection method based on a lightweight convolutional neural network according to any one of claims 1 to 7, characterized in that: The industrial process oscillation detection system based on lightweight convolutional neural network includes: A simulation variable determination module is used to determine industrial data simulation variables; the industrial data simulation variables include signal length, noise variance, disturbance amplitude, oscillation probability, number of cycles, waveform, smoothing factor, oscillation time-varying probability and frequency change factor; A simulation data construction module, used for generating a simulation time series based on the industrial data simulation variables to construct a simulation data set; the simulation data set includes a plurality of simulation time series and an oscillation category of each simulation time series; the oscillation category is a non-oscillation category, a regular oscillation category or an irregular oscillation category; A neural network training module, used for training a lightweight convolutional neural network according to the simulation data set to obtain an oscillation detection model; the lightweight convolutional neural network includes an input layer, a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a third maximum pooling layer, a fifth convolutional layer, a sixth convolutional layer, a fourth maximum pooling layer, a global average pooling layer and a fully connected layer connected in sequence; The oscillation detection module is used to determine the oscillation detection result of the industrial data to be detected by using the oscillation detection model according to the industrial data to be detected.

Images