Multi-scale residual convolution and LSTM fusion performance evaluation method for industrial processes

Abstract

The invention discloses a multi-scale residual convolution and LSTM fusion performance evaluation method for an industrial process, comprising the following steps: step 1, collecting a data set of multiple types of faults, and performing data preprocessing on the data, on this basis, numbering and labeling the fault types, setting corresponding relationships, and then dividing the fault types and labels into corresponding training sets and test sets; step 2, constructing a multi-scale residual convolution and LSTM network fault diagnosis model, comprising a residual module, a multi-scale residual module, an LSTM module and a classification module; step 3, training the model with a training set, taking the fault type label as the output of the model, taking the cross entropy of the output label and the true label as the loss function, performing back propagation with an optimizer, updating the parameters of the entire model, and optimizing the loss function; step 4, inputting the test set into the trained model, performing fault identification, and outputting the diagnosis result.

Description

Technical Field

[0001] The invention belongs to the field of automated process control, and in particular relates to a multi-scale residual convolution and long short-term memory network (MRCNN-LSTM) fusion fault diagnosis method for industrial engineering. Background Art

[0002] Modern industrial process systems are an indispensable basic equipment in current social production, and they play a vital role in the development and progress of society. In order to improve the production efficiency of industrial equipment and reduce resource waste, it is very necessary to timely discover and diagnose the type of fault. Therefore, a good industrial process fault diagnosis method is of great significance to industrial production.

[0003] The emergence of modern fault diagnosis technology has played an important role in ensuring the production safety of industrial processes and reducing resource waste. It has developed model-based methods, knowledge-based reasoning methods and data-driven methods. Due to their own limitations, model-based methods and knowledge-based reasoning methods cannot achieve satisfactory results for complex industrial process data with high dimensionality, nonlinearity, intermittency and dynamic characteristics. Data-driven methods have been well developed and applied because they have fewer limitations and only rely on past fault data.

[0004] Data-driven methods can be further divided into methods based on multivariate statistics, shallow learning methods, and deep learning methods. Methods based on multivariate statistics and shallow learning methods are commonly used in industrial processes, but they cannot achieve satisfactory results when faced with high-dimensional and nonlinear data in industrial data. Deep learning technology is developed on the basis of shallow learning. It gets rid of the tedious steps of manual feature extraction required by traditional methods and can automatically extract nonlinear high-dimensional features from raw data end-to-end. Not only that, deep learning also solves the problems of shallow learning being prone to overfitting, falling into local optimality, gradient dissipation, and weak generalization ability. It also has good diagnostic accuracy for complex industrial processes such as nonlinear, intermittent, and dynamic.

[0005] Deep learning has huge advantages over traditional methods. It uses a multi-hidden layer network structure to directly analyze input data samples and extract hidden data feature information. It is very suitable for large-scale and complex industrial process data and is currently a hot topic in the field of fault diagnosis. However, it is undeniable that each deep learning model has its own advantages and disadvantages, and its overall diagnostic accuracy needs to be further improved. At present, the application of deep learning-based fault diagnosis methods in industrial processes mainly has the following problems:

[0006] When the model depth reaches a certain number of layers, model degradation will occur;

[0007] Cannot effectively extract the temporal features hidden in the data;

[0008] It is not possible to effectively extract features of data of different sizes. Summary of the invention

[0009] In order to solve the above problems, it is necessary to provide an industrial process fault diagnosis method that can overcome the above problems.

[0010] The purpose of the present invention is to address the problems existing in the existing industrial process fault diagnosis methods and propose an industrial process fault diagnosis method that can effectively extract data features of different sizes and temporal features and avoid model degradation by integrating multi-scale residual convolution and LSTM network.

[0011] Based on the traditional convolutional neural network, the present invention constructs a multi-scale convolutional neural network and an LSTM network respectively, and embeds residual learning into the model. The entire model consists of a residual module, a multi-scale residual module, an LSTM module, and a classification module.

[0012] The specific implementation steps of the present invention include:

[0013] Step 1: Use frequency f s Collect the operating data of industrial process system variables, establish a data set containing multiple variables and multiple types of faults, and perform data preprocessing on it. On this basis, number and label the fault types, set the corresponding relationships, and then divide the fault types and labels into corresponding training sets and test sets;

[0014] Data preprocessing refers to the Min-max normalization of the data set. This method can linearly change the collected original signal data set so that the data samples fall in the [0, 1] interval. The relevant conversion function is as follows:

[0015]

[0016] Where min{·} represents the minimum value in the original data, max{·} represents the maximum value in the sample data, and x i Represents sample data;

[0017] Fault sample labeling refers to one-hot encoding of the data set, also known as one-bit effective encoding. It uses an N-bit status register to encode N faults. Each fault has its own independent register bit, and at any time, only one of them is valid.

[0018] Step 2: Construct a multi-scale residual convolution and LSTM network fault diagnosis model;

[0019] The multi-scale residual convolution and LSTM network model is mainly composed of a residual module, a multi-scale residual module, an LSTM module and a classification module, where:

[0020] The residual module consists of two convolutional layers and a BN layer. The convolutional layer performs convolution operations on the data, using the relu activation function. Padding uses same to pad the input so that the output has the same length as the input, and uses L2 regularization technology. The convolution operation of the convolutional layer is shown in the following formula:

[0021] x [l] =f(W [l] *x [l-1] +b [l] ) (2)

[0022] Where x [l] represents the output of the lth layer, l∈[1, 2, …, n], W [l] represents the lth layer convolution kernel, “*” represents the convolution operation, b [l] represents the bias corresponding to the lth layer, and f(·) represents the activation function.

[0023] The BN layer performs batch normalization on the input data, so that the mean of the output data is 0 and the variance is 1. The relevant operations are as follows:

[0024]

[0025]

[0026] In the formula Represents the output after BN layer calculation; y i represents the input feature value; E represents the mean of the input feature value; D represents the variance of the input feature value; γ and β represent the parameters that need to be trained in the model.

[0027] L2 regularization further reduces the overfitting phenomenon of the model, which is specifically expressed as follows:

[0028]

[0029] Where N represents the number of samples, y (n) represents the actual label of the nth sample, n∈[1,2,…,N], x n represents the nth sample, θ is its parameter, l p is the norm function, p∈[1,2], represents the L2 norm, λ is the regularization coefficient, f(·) is the learned neural network, and L(·) is the loss function;

[0030] The output of the residual module is input into the maximum pooling layer, and the relevant calculations are as follows:

[0031] h=max(c(t)) (6)

[0032] Where h represents the output of the pooling layer, c(t) represents the numerical set of each neuron in the input feature, and t∈[1, 2, …, n] represents the tth neuron;

[0033] The multi-scale residual module uses residual modules with convolution kernel sizes of 1×1, 1×3, and 1×5 to be connected in parallel to extract feature information of different sizes in the input data. 1 ,y 2 and 3 Concatenate into feature vector y c =[y 1 ,y 2 ,y 3 ], the feature fusion is performed through the multi-scale feature connection module to obtain the output feature y, which is then input into the LSTM module;

[0034] The LSTM module is composed of a two-layer LSTM network, a BN layer, and a maximum pooling layer. The tanh activation function is used, and the output result is input into the classification module. The specific calculation formula of the tanh activation function is as follows:

[0035]

[0036] The classification module consists of multiple fully connected layers alternating with dropout layers. The fully connected layers flatten and weight the local features extracted by the convolutional network and input them into the dropout layers.

[0037] The specific calculation formula of the fully connected layer is as follows:

[0038] y=W*x+b (8)

[0039] Where W represents the weight, b represents the bias, x represents the input neuron, and y represents the weighted output.

[0040] The dropout layer randomly discards neurons in the network with probability p to enhance the generalization ability of the model, and then inputs the output results into the Softmax classification layer;

[0041] The Softmax classification layer performs classification operations on the input features, and its calculation formula is as follows:

[0042]

[0043] In the formula represents the value of the i-th output neuron;

[0044] Step 3: Use the divided training set to train the multi-scale residual convolution and LSTM network model;

[0045] The divided training set is input into the MRCNN-LSTM model for training, the fault type label is used as the output of the model, the cross entropy between the output label and the true label is used as the loss function, and the optimizer is used for back propagation to update the parameters of the entire model and optimize the loss function.

[0046] Step 4: Input the segmented test set into the trained model to perform fault identification and output the diagnosis results. The output of the Softmax regression classifier can reflect the fault type predicted by the model.

[0047] Compared with the existing industrial process fault diagnosis method, the present invention has the following beneficial effects:

[0048] The present invention proposes an industrial process fault diagnosis model that integrates multi-scale convolution and LSTM network, which overcomes the shortcoming that a single deep learning model easily causes feature information loss. The multi-scale residual module can adaptively perform multi-branch parallel extraction of rich and complementary nonlinear and high-dimensional spatial features in the data along the time axis of the signal, while retaining the temporal characteristics of the data; the LSTM module can use its good ability to process temporal data to further extract the temporal characteristics of the input data;

[0049] The introduction of residual learning enables the model to better learn the deep and shallow features of the data, accelerate the convergence of the network, and avoid the performance degradation of the model;

[0050] Since the proposed model is based on CNN, it inherits the advantage of CNN's powerful feature extraction ability. It can extract high-dimensional features of data from unprocessed raw signals in an end-to-end manner and identify fault types. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of the implementation of the fault diagnosis model proposed by the present invention;

[0052] Figure 2 It is a structural schematic diagram of the model proposed by the present invention;

[0053] Figure 3a and Figure 3b They are respectively the loss rate and accuracy curves of the model proposed in the present invention during the training process;

[0054] Figure 4a and Figure 4b They are two-dimensional scatter plots of the original data of the test set and the data output by the model proposed in the present invention;

[0055] Figure 5 Shown are test set accuracy curves of the model of the present invention and four comparative models. DETAILED DESCRIPTION

[0056] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0057] Industrial coking furnace is a typical industrial process system. It is a vertical tube heating furnace. It plays an irreplaceable role in the deep processing of chemical raw materials. Its fuel is generally high-pressure gas. When working, the gas enters the coking furnace from the north and south sides respectively. The raw material residual oil is sent to the coking furnace convection chamber from the north and south sides respectively to be preheated to about 330℃, and then sent to the bottom of the fractionation tower together, contacting with the oil and gas from the top of the coking furnace tower and transferring heat and mass; in this process, the lighter components in the mixed oil evaporate and rise to the distillation section for separation, while the fractions above the wax oil in the raw material flow into the bottom of the tower together with the condensed components in the oil and gas from the top of the coking tower. The fractionation tower bottom oil at about 360℃ is sent to the heating furnace radiation chamber in two ways through the heating furnace radiation feed pump and quickly heated to 495℃, and then enters the coking tower for cracking reaction.

[0058] In this example verification, several common faults of different types were pre-designed in the coking furnace. The specific fault conditions and labels are shown in Table 1.

[0059] Table 1:

[0060]

[0061]

[0062] Figure 1 It is an implementation flow chart of the industrial process fault diagnosis method integrating multi-scale residual convolution and LSTM network proposed in the present invention. Figure 2 It is a specific structural diagram, which specifically includes the following steps:

[0063] Step 1: Collect the operating data of 8 main process variables in the industrial coking furnace, establish a data set containing 8 variables and 5 types of faults, collect 1850 sample data under each fault condition, and collect 2000 normal state samples, from which 1850 samples are randomly selected to form experimental data together with the fault samples. Perform data preprocessing on the collected data set, and on this basis, number and label the fault types, set the corresponding relationship, and then randomly select 80% of the samples as the training set, and the remaining 20% ​​of the samples as the test set.

[0064] Data preprocessing refers to the Min-max normalization of the data set. This method can linearly change the collected original signal data set so that the data samples fall in the [0, 1] interval. The relevant conversion function is as follows:

[0065]

[0066] Where min{·} represents the maximum value in the original data, max{·} represents the minimum value in the sample data, and x i Represents sample data;

[0067] Fault sample labeling refers to one-hot encoding of the data set, also known as one-bit effective encoding. It uses an N-bit status register to encode N faults. Each fault has its own independent register bit, and at any time, only one of them is valid.

[0068] Step 2: Construct a multi-scale residual convolution and LSTM network fault diagnosis model;

[0069] The multi-scale residual convolution and LSTM network model is mainly composed of a residual module, a multi-scale residual module, an LSTM module and a classification module, where:

[0070] The residual module consists of two convolutional layers and a BN layer. The convolutional layer performs convolution operations on the data, using the relu activation function. Padding uses same to pad the input so that the output has the same length as the input, and uses L2 regularization technology. The number of convolution kernels is 64, and the convolution kernel size is 1x5. The convolution operation of the convolutional layer is shown in the following formula:

[0071] x [l] =f(W [l] *x [l-1] +b [l] ) (2)

[0072] Where x [l] represents the output of the lth layer, l∈[1, 2, …, n], W [l] represents the lth layer convolution kernel, “*” represents the convolution operation, b [l] represents the bias corresponding to the lth layer, and f(·) represents the activation function.

[0073] The BN layer performs batch normalization on the input data, making the mean of the output data 0 and the variance 1, thereby improving the generalization ability of the model, accelerating the convergence and training speed of the model, and to a certain extent avoiding the gradient explosion or gradient disappearance problems during the training process of the model, so that the stability and anti-overfitting ability of the network model during training are enhanced. The relevant calculations are as follows:

[0074]

[0075]

[0076] In the formula Represents the output after BN layer calculation; y i represents the input feature value; E represents the mean of the input feature value; D represents the variance of the input feature value; γ and β represent the parameters that need to be trained in the model.

[0077] L2 regularization technology can further reduce the overfitting phenomenon of the model, and its optimization problem can be expressed as follows:

[0078]

[0079] Where N represents the number of samples; y (n) represents the actual label of the nth sample, n∈[1,2,…,N]; x n represents the nth sample, θ is its parameter; l p is the norm function, p∈[1,2], represents the L2 norm, λ is the regularization coefficient, f(·) is the learned neural network, and L(·) is the loss function.

[0080] The output of the residual module is input into the maximum pooling layer, which can prevent the network from overfitting, improve the generalization ability of the model, reduce the data dimension, and speed up the calculation. The relevant calculations are as follows:

[0081] h=max(c(t)) (6)

[0082] Where h represents the output of the pooling layer, c(t) represents the numerical set of each neuron in the input feature, and t∈[1, 2, …, n] represents the tth neuron;

[0083] The multi-scale residual module uses residual modules with convolution kernel sizes of 1×1, 1×3, and 1×5, and the number of convolution kernels is 64, which are connected in parallel to extract feature information of different sizes in the input data. 1 ,y 2 and 3 Concatenate into feature vector y c =[y 1 ,y 2 ,y 3 ], the feature fusion is performed through the multi-scale feature connection module to obtain the output feature y, which is then input into the LSTM module;

[0084] The LSTM module is composed of a two-layer LSTM network, a BN layer, and a maximum pooling layer. It uses the tanh activation function, and the number of units is 64 and 128 respectively. It can further extract the temporal characteristics of the data and reduce the information loss of the data characteristics. The output results are input into the classification module. The specific calculation formula of the tanh activation function is as follows:

[0085]

[0086] The classification module consists of two layers of dropout layers and fully connected layers connected alternately. The sizes of the two dropout layers are 0.3 and 0.2 respectively, and the number of nodes of the two fully connected layers are 256 and 6 respectively. The fully connected layer flattens and weights the local features extracted by the convolutional network and inputs them into the dropout layer;

[0087] The calculation formula of the fully connected layer is as follows:

[0088] y=W*x+b (8)

[0089] Where W represents weight, b represents bias, x represents input neuron, and y represents weighted output;

[0090] The dropout layer randomly discards neurons in the network with probability p to enhance the generalization ability of the model, and then inputs the output results into the Softmax classification layer;

[0091] The Softmax classification layer performs classification operations on the input features, and its calculation formula is as follows:

[0092]

[0093] In the formula represents the value of the i-th output neuron;

[0094] Step 3: Use the divided training set to train the multi-scale residual convolution and LSTM network model;

[0095] The divided training set is input into the MRCNN-LSTM model for training, the fault type label is used as the output of the model, the cross entropy between the output label and the true label is used as the loss function, and the Adam optimizer is used for back propagation to update the parameters of the entire model and optimize the loss function. This optimization algorithm can quickly reach the global optimal point of the model along the direction with the fastest gradient descent, and can achieve the best training effect with fewer network iterations.

[0096] Step 4: Input the segmented test set into the trained model to perform fault identification and output the diagnosis results. The output of the Softmax regression classifier can reflect the fault type predicted by the model.

[0097] In order to better illustrate the technical effect of the present invention, this experiment uses CNN, ResCNN, CNN-LSTM and MRCNN as comparative experiments. The above are all network models that have achieved good results. For the comparability of the experimental results, the parameter settings of the four comparative models are basically the same as those of the present invention.

[0098] In this experiment, Keras library and python3.5 were used to implement the fusion multi-scale residual convolution and LSTM network model proposed in this invention and four comparison models. The number of iterations during model training was set to 100, the batch size was set to 64, the cross entropy loss function was selected as the loss function, the initial learning rate was set to 0.001, and the learning rate decreased by one tenth every 30 iterations.

[0099] The training set accuracy and loss rate and the test set accuracy and loss rate of this experiment are as follows Figure 3a and Figure 3b As shown, it can be seen that there is no overfitting or underfitting phenomenon in the training process, indicating that the structure and parameter setting of the model proposed in the present invention are reasonable. Figure 4a and Figure 4b The two-dimensional scatter plots are respectively the original data of the test set and the data after the output of the model proposed in the present invention. It can be seen from the two-dimensional scatter plots that the model proposed in the present invention can classify the fault types well, and only a very small number of samples are misclassified.

[0100] Table 2 shows the specific precision (P) and recall (R) of the five models for each fault type. It can be seen from the two evaluation indicators of micro-average precision and recall that the micro-average precision and recall of the model proposed in the present invention are the highest among the five models, reaching 97.80%, showing that the model proposed in the present invention has outstanding diagnostic accuracy. Specifically for each fault type, it can be found that, except for fault 2, the micro-average precision and recall of MRCNN-LSTM for the other five types are improved to varying degrees.

[0101] Table 2:

[0102]

[0103]

[0104] The test set accuracy curves of the present invention and four comparison models are shown in Figure 5 As shown, it can be clearly seen that the accuracy curve of the model of the present invention has been significantly improved compared with the comparison model, and its curve is almost always above the comparison model curve. Especially after 30 iterations, the accuracy of the model begins to comprehensively lead the four comparison models and remains so until the end of training.

[0105] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Multi-scale residual convolution and LSTM fusion performance evaluation method for industrial processes, characterized by The steps include: Step 1: Use frequency f s Collect the operating data of industrial process system variables, establish a data set containing multiple variables and multiple types of faults, and perform data preprocessing on it. On this basis, number and label the fault types, set the corresponding relationships, and then divide the fault types and labels into corresponding training sets and test sets; Step 2: Construct a multi-scale residual convolution and LSTM network fault diagnosis model, including a residual module, a multi-scale residual module, an LSTM module, and a classification module. Specifically, the following steps are included: Step 2.1, the residual module includes a convolutional layer and a BN layer. The convolutional layer performs convolution operations on the data, uses the relu activation function, and padding uses same to pad the input so that the output has the same length as the input, and uses regularization; The BN layer performs batch normalization on the input data, so that the mean of the output data is 0 and the variance is 1; The output of the residual module is input into the maximum pooling layer; Step 2.2, the multi-scale residual module uses 64 residual modules with different convolution kernel sizes to connect in parallel, concatenates the extracted feature data into feature vectors, and performs feature fusion through the multi-scale feature connection module to obtain the output feature y, which is then input into the LSTM module; Step 2.3, the LSTM module includes two layers of LSTM networks with 64 and 128 units respectively, a BN layer and a maximum pooling layer, an activation function is used, and the output result is input into the classification module; Step 2.4, the classification module includes a fully connected layer and a dropout layer. The fully connected layer flattens and weights the local features extracted by the convolutional network and inputs them into the dropout layer; The dropout layer randomly discards neurons in the network with probability p to enhance the generalization ability of the model, and then inputs the output results into the classification layer; Step 3: Use the training set to train the multi-scale residual convolution and LSTM network fault diagnosis model, use the fault type label as the output of the model, use the cross entropy between the output label and the true label as the loss function, use the optimizer for back propagation, update the parameters of the entire model, and optimize the loss function; Step 4: Input the test set into the trained multi-scale residual convolution and LSTM network fault diagnosis model to perform fault identification and output the diagnosis results.

2. The multi-scale residual convolution and LSTM fusion performance evaluation method for industrial processes as described in claim 1 is characterized in that The convolution operation in step 2.1 is as follows: x [l] =f(W [l] *x [l-1] +b [l] ) (2) Where x [l] represents the output of the lth layer, l∈[1, 2, …, n], W [l] represents the convolution kernel of the first layer, "*" represents the convolution operation, b [l] represents the bias corresponding to the lth layer, and f(·) represents the activation function.

3. The multi-scale residual convolution and LSTM fusion performance evaluation method for industrial processes as described in claim 1 is characterized in that The BN layer in step 2.1 is specifically operated as follows: In the formula Represents the output after BN layer calculation; y i represents the input feature value; E represents the mean of the input feature value; D represents the variance of the input feature value; γ and β represent the parameters that need to be trained in the model.

4. The multi-scale residual convolution and LSTM fusion performance evaluation method for industrial processes as described in claim 1 is characterized in that The regularization in step 2.1 is L2 regularization, which is specifically expressed as follows: Where N represents the number of samples, y (n) represents the actual label of the nth sample, n∈[1,2,…,N], x n represents the nth sample, θ is its parameter, l p is the norm function, p∈[1,2], represents the L2 norm, λ is the regularization coefficient, f(·) is the learned neural network, and L(·) is the loss function.

5. The multi-scale residual convolution and LSTM fusion performance evaluation method for industrial processes as described in claim 1 is characterized in that The maximum pooling layer in step 2.1 is specifically calculated as follows: h=max(c(t)) (6) Where h represents the output of the pooling layer, c(t) represents the numerical set of each neuron in the input feature, and t∈[1, 2, …, n] represents the tth neuron.

6. The multi-scale residual convolution and LSTM fusion performance evaluation method for industrial processes as described in claim 1, characterized in that The activation function in step 2.3 is the tanh activation function, and the specific calculation formula is as follows:

7. The multi-scale residual convolution and LSTM fusion performance evaluation method for industrial processes as described in claim 1 is characterized in that The specific calculation formula of the fully connected layer in step 2.4 is as follows: y=W*x+b (8) Where W represents the weight, b represents the bias, x represents the input neuron, and y represents the weighted output.

8. The multi-scale residual convolution and LSTM fusion performance evaluation method for industrial processes as described in claim 1, characterized in that The classification layer in step 2.4 is a Softmax classification layer, which performs classification operations on the input features. The calculation formula is as follows: In the formula represents the value of the i-th output neuron.

9. The multi-scale residual convolution and LSTM fusion performance evaluation method for industrial processes as described in claim 1, characterized in that The data preprocessing in step 1 refers to performing deviation standardization on the data set, and performing linear changes on the collected original signal data set so that the data samples fall within the interval [0, 1]. The relevant conversion function is as follows: Where min{·} represents the minimum value in the original data, max{·} represents the maximum value in the sample data, and x i Represents sample data.

10. The multi-scale residual convolution and LSTM fusion performance evaluation method for industrial processes as claimed in claim 1, characterized in that The labeling in step 1 refers to one-hot encoding of the data set, that is, using an N-bit status register to encode N faults, each fault has its own independent register bit, and at any time, only one of them is valid.

Images