Continuous casting roller fault prediction method based on time convolutional network-long and short term memory network

Abstract

The invention discloses a continuous casting roller fault prediction method based on a time convolution network-long and short term memory network, and the method specifically comprises the following steps: 1, constructing a continuous casting roller data set, and dividing the data set into a training set and a test set; step 2, establishing a fault prediction model; 3, training the sample fault prediction model constructed in the step 2 by adopting the training set divided in the step 1, and outputting a trained model; and 4, evaluating the model trained in the step 3 by using the test set divided in the step 1, and outputting a prediction result. By adopting the method provided by the invention, the fault of the continuous casting roller can be accurately predicted.

Description

Technical Field

[0001] The present invention belongs to the technical field of continuous casting roll fault diagnosis, and relates to a continuous casting roll fault prediction method based on a time convolutional network-long short-term memory network. Background Technique

[0002] The continuous casting machine is a key equipment in the production process of iron and steel enterprises, and its efficient and stable operation is crucial for the quality of steel billets and production safety. The continuous casting roll is a key part for the continuous casting machine to achieve uniform cooling and shaping of molten steel, and its performance directly determines the dimensional accuracy of steel billets and the uniformity of the internal structure. During the service process of the continuous casting roll, its roll surface will repeatedly experience periodic temperature changes and bear various mechanical stresses applied by the continuous casting billet, such as self-weight sliding force, billet pulling force, bulging force, transmission resistance, and straightening force, etc. These factors are likely to cause faults such as bending, fatigue cracks, and wear of the continuous casting roll. The faults of the continuous casting roll will reduce the production quality of steel billets, cause frequent shutdowns and repairs of the continuous casting machine, reduce production efficiency and increase maintenance costs. In extreme cases, the faults of the continuous casting roll may even trigger safety accidents, posing a serious threat to the safety of operators and equipment. Therefore, through the fault prediction of the continuous casting roll, timely replacement or repair can be carried out before the fault occurs, which can avoid sudden shutdowns caused by continuous casting roll faults, reduce the frequency and cost of planned maintenance, and increase the service life of the continuous casting roll. Summary of the Invention

[0003] The purpose of the present invention is to provide a continuous casting roll fault prediction method based on a time convolutional network-long short-term memory network, and using this method can accurately predict the faults of continuous casting rolls.

[0004] The technical solution adopted by the present invention is a continuous casting roll fault prediction method based on a time convolutional network-long short-term memory network, which specifically includes the following steps:

[0005] Step 1, construct a continuous casting roll data set and divide the data set into a training set and a test set;

[0006] Step 2, establish a fault prediction model;

[0007] Step 3, use the training set divided in Step 1 to train the sample fault prediction model constructed in Step 2, and output the trained model;

[0008] Step 4, use the test set divided in Step 1 to evaluate the model trained in Step 3, and output the prediction result.

[0009] The characteristics of the present invention also lie in:

[0010] The specific process of Step 1 is:

[0011] Step 1.1: Construct a continuous casting roll dataset based on the historical fault data of the continuous casting rolls of the continuous casting machine, and use the Z-score standardization method to standardize the original data of the samples in the continuous casting roll dataset;

[0012] Step 1.2: Label the standardized dataset, label the fault-free data as 0, and label the fault data as 1;

[0013] Step 1.3: Divide the dataset labeled in Step 1.2 into a 70% training set and a 30% test set.

[0014] In Step 1.1, the following formulas (1)-(3) are used for standardization processing:

[0015]

[0016] Among them, z represents the standardized sample data; i represents the i-th sample in the continuous casting roll dataset, N represents the total number of samples in the continuous casting roll dataset; σ is the mean difference of a certain index parameter in the continuous casting roll fault samples; μ is the mean of a certain index parameter in the continuous casting roll fault samples; x is a certain index parameter value in the continuous casting roll fault samples.

[0017] The specific process of Step 2 is as follows:

[0018] Step 2.1: Construct a Seq2Seq deep learning model based on TCN-LSTM, where the encoder module uses TCN, the decoder uses LSTM, and Self-Attention is introduced to process variable-length sequences;

[0019] Step 2.2: Input the output vector y of Step 2.1 t into the fully connected layer, and use the sigmoid activation function for conversion as the fault probability value P(y t ), that is, it represents the probability of the continuous casting roll failing.

[0020] The specific process of Step 2.1 is as follows:

[0021] Step 2.1.1: Create a TCN encoder model, and input the sample data sequence {p 1 , p 2 , ···, p 16 , p 17} of the continuous casting roll fault training set in Step 1.3 into the TCN encoder model, and output the sequence H t = {h 1 , h 2 , … h t-1};

[0022] Step 2.1.2: For the output sequence H of Step 2.1.1 tDynamically adjust the weights using self-attention to obtain the context vector D t ={d 1 ,d 2 ,…d t-1};

[0023] Step 2.1.3: Create an LSTM decoder model and use the gated weighted fusion mechanism to combine the context vector D obtained in Step 2.1.2 t with the output H of the previous time step t as the input sequence X of the decoder LSTM t , and the formula is as follows:

[0024] X t =g t *D t +(1 - g t )*H t (4)

[0025] where g t is the dynamic contribution weight of the context vector;

[0026] Use the memory cell and gated mechanism of the decoder LSTM to output the vector y of the final time step t , and the process is as follows:

[0027] f t =σ(W f ·[h t-1 ,x t +b f ) (5)

[0028] o t =σ(W o *[h t-1 ,x t +b o ) (6)

[0029]

[0030] i t =σ(W i ·[h t-1 ,x t +b i ) (8)

[0031]

[0032] y t =o t *tanh(C t ) (11)

[0033] where ft is the weight of the forget gate, with a value range of [0, 1]; W f , W o , W i represent the weight matrices of the forget gate, output gate, and input gate respectively, and b f , b o , b i are the corresponding bias terms respectively; C t are the candidate memory unit and updated memory unit at the current time respectively; x t is the input information at the current time; h t-1 is the short-term hidden information at the previous time; σ is the sigmoid function; b f is the bias vector of the forget gate; i t is the output of the forget gate.

[0034] In step 2.2, the failure probability value P(y t ) is represented by the sigmoid activation function δ(y) as follows:

[0035]

[0036] The specific process of step 3 is as follows:

[0037] Step 3.1, hyperparameter setting: including learning rate, training batch size, number of iterations, number of convolutional layers, filters, convolutional kernel size, dilation rate of each convolutional layer, residual connection, number of LSTM units;

[0038] Step 3.2, model training: By using the training set to train the failure model established in step 2, continuously optimize the parameters in the model until the loss value of the loss function in the failure model tends to be stable.

[0039] In step 4, the model accuracy, recall rate, accuracy rate, and F1 score are calculated as evaluation indicators, and the specific calculation process is as follows:

[0040] Model accuracy: The proportion of the number of correctly predicted samples in the total number of samples:

[0041]

[0042] Model accuracy rate: The ratio of the number of correctly predicted continuous casting roll failure samples TP to the positive class samples:

[0043]

[0044] Model recall rate: The ratio of the number of samples predicted as failures to the number of samples that are actually failures:

[0045]

[0046] Calculate the F1 score: Considering the balance between Precision and Recall, take the harmonic mean of the two:

[0047]

[0048] Among them, TP is the number of correctly predicted continuous casting rolls as the fault class, FP is the number of wrongly predicted continuous casting roll fault classes as non-fault classes, FN is the number of wrongly predicted continuous casting roll non-fault classes as fault classes, TN is the number of correctly predicted continuous casting roll non-fault classes, and the sum of the four is the total number m of test samples.

[0049] The beneficial effect of the present invention is to predict the fault state of the continuous casting roll in order to achieve timely maintenance of the continuous casting roll, reduce the maintenance cost, and increase its service life. Using the TCN as the encoder and the LSTM as the decoder, introducing the self-attention mechanism to dynamically adjust the attention weights, and using the grid search algorithm to optimize the parameters, a continuous casting roll fault prediction model based on TCN-LSTM is established, which further improves the accuracy of continuous casting roll fault prediction. Brief Description of the Drawings

[0050] Figure 1 is the flow chart of the continuous casting roll fault prediction method based on the time convolutional network-long short-term memory network of the present invention;

[0051] Figure 2 is the schematic structural diagram of the continuous casting roll fault prediction model constructed in the continuous casting roll fault prediction method based on the time convolutional network-long short-term memory network of the present invention;

[0052] Figure 3 is the schematic structural diagram of the long short-term memory network in the continuous casting roll fault prediction method based on the time convolutional network-long short-term memory network of the present invention. Detailed Embodiments

[0053] The present invention will be described in detail below with reference to the drawings and specific embodiments.

[0054] Example 1

[0055] The fault prediction method of continuous casting rolls based on the Temporal Convolutional Network-Long Short-Term Memory Network constructs a fault prediction model of continuous casting rolls based on TCN-LSTM, which can overcome the problems of local information loss and insufficient grasp of global information caused by the use of the Recurrent Neural Network (RNN) in the Sequence-to-Sequence (Seq2Seq). By adopting the advantages of the Temporal Convolutional Network (TCN) and the Long Short-Term Memory (LSTM), combining the prediction framework of Seq2Seq and the self-attention mechanism, the ability of the model to process the operation data of continuous casting rolls is improved, ensuring that long-term and short-term data patterns can be comprehensively captured and analyzed in fault prediction, improving the accuracy of continuous casting roll fault prediction, and providing technical support for the continuity and stability of billet production.

[0056] Example 2

[0057] The fault prediction method of continuous casting rolls based on the Temporal Convolutional Network-Long Short-Term Memory Network-Sequence-to-Sequence, the overall process of fault prediction is as Figure 1 , this method mainly includes 4 processes: constructing a continuous casting roll data set, establishing a fault prediction model, sample training, sample testing and evaluation. Specifically, it includes the following steps:

[0058] Step 1, constructing a continuous casting roll data set;

[0059] Step 2, establishing a fault prediction model;

[0060] Step 3, sample training: using the training samples in Step 1 to train the continuous casting roll fault prediction model established in Step 2;

[0061] Step 4, sample testing and evaluation: testing the model trained in Step 3.

[0062] Example 3

[0063] The specific process of Step 1 is as follows:

[0064] Step 1.1, taking the historical fault data of the continuous casting rolls of a 2150-type continuous casting machine of a certain company as an example, the index parameters of the continuous casting roll data set are constructed as shown in Table 1; a partial continuous casting roll fault data set is shown in Table 2. The Z-score standardization method is used to standardize the original data of the continuous casting roll data set samples. The standardization formula is as follows:

[0065]

[0066] Among them, z represents the standardized sample data; i represents the i-th sample in the continuous casting roll dataset, N represents the total number of samples in the continuous casting roll dataset; σ is the mean difference of a certain (Table 1) index parameter in the continuous casting roll fault samples; μ is the mean of a certain (Table 1) index parameter in the continuous casting roll fault samples; x is the value of a certain (Table 1) index parameter in the continuous casting roll fault samples.

[0067] Step 1.2, label the dataset after standardization; mark the fault-free data as 0 and the fault data as 1 to construct the classification labels for model training. The processed continuous casting roll fault samples are shown in Table 3;

[0068] Step 1.3, divide the fault samples. Randomly divide the labeled dataset in Step 1.2 into a 70% training set and a 30% test set.

[0069] Example 4

[0070] Step 2, as Figure 2 shown, establish a fault prediction model;

[0071] Step 2.1, construct a Seq2Seq deep learning model based on TCN-LSTM. Among them, the encoder module uses TCN, the decoder uses LSTM, and Self-Attention is introduced to process variable-length sequences. Seq2Seq is a processing framework, mainly consisting of an encoder (Encoder) and a decoder (Decoder). The input sequence is processed by the encoder, and the decoder generates the output sequence. Its encoder generally uses recurrent neural networks such as RNN, LSTM, or GRU, and the decoder also uses a similar structure, and the performance is improved by combining the attention mechanism.

[0072] Step 2.1.1, create a TCN encoder model. Take the sample data sequence {p 1 , p 2 , ···, p 16 , p 17} of the standardized continuous casting roll fault training set in Step 1.3 as the input, use the convolutional operation of TCN to expand the receptive field, capture long-term dependence relationships, extract more accurate fault features, and output the sequence H t = {h 1 , h 2 , … h t-1};

[0073] Step 2.1.2, dynamically adjust the weights of the output sequence H t in Step 2.1.1 using self-attention to obtain the context vector D t = {d1 , d 2 , … d t-1}, so as to adapt it to input sequences of different lengths, better solve the information bottleneck, and avoid information loss.

[0074] Step 2.1.3, create an LSTM decoder model, the structure of which is as Figure 3 shown. Use the gated weighted fusion mechanism to combine the context vector D t obtained in Step 2.1.2 with the output H t of the previous time step as the input sequence X t of the decoder LSTM, which can be expressed as:

[0075] X t = g t * D t + (1 - g t ) * H t (4)

[0076] where gt is the dynamic contribution weight of the context vector.

[0077] Use the memory unit and gating mechanism of the decoder LSTM to output the vector y t at the final time step, and its calculation is as follows:

[0078] f t = σ(W f · [h t-1 , x t + b f ) (5)

[0079] o t = σ(W o * [h t-1 , x t + b o ) (6)

[0080]

[0081] i t = σ(W i · [h t-1 , x t + b i ) (8)

[0082]

[0083] y t = o t * tanh(C t ) (11)

[0084] where f tis the weight of the forget gate, with a value range of [0, 1]; W f 、W o 、W i represent the weight matrices of the forget gate, output gate, and input gate respectively, and b f 、b o 、b i are the corresponding bias terms respectively; C t are the candidate memory unit and updated memory unit at the current moment respectively; x t is the input information at the current moment; h t-1 is the short-term hidden information at the previous moment; σ is the sigmoid function; b f is the bias vector of the forget gate; i t is the output of the forget gate.

[0085] Step 2.2, input the output vector y t of Step 2.1.3 into the fully connected layer, and use the sigmoid activation function for conversion as the fault probability value P(y t ), which represents the probability of the continuous casting roll failing. The fault probability value can be expressed by the sigmoid activation function δ(y) as:

[0086]

[0087] Example 5

[0088] The specific process of Step 3 is as follows:

[0089] Step 3.1, hyperparameter setting; in order to find the optimal hyperparameters in the training set, mainly including the learning rate, training batch size, number of iterations, number of convolutional layers, filters, convolutional kernel size, dilation rate of each convolutional layer, residual connection, number of LSTM units, the parameter settings are shown in Table 4. Using the grid search algorithm, by presetting several groups of hyperparameter combinations, each group of hyperparameters uses cross-validation for fault model training and evaluation, and the optimal parameter combination is selected according to the evaluation effect of the fault model.

[0090] Step 3.2, model training; by using the training set to train the fault model established in Step 2, continuously optimize the parameters in the model until the loss value of the loss function (mean squared error MSE is used as the loss function) in the fault model tends to be stable.

[0091] Example 6

[0092] The specific process of Step 4 is as follows:

[0093] By calculating the model accuracy, recall, precision, and F1-score as evaluation metrics, the results are shown in Table 5; the accuracy of the improved model on the test set reaches 96.94%, and the F1-score is as high as 96.10%.

[0094] Model accuracy (Accuracy): The proportion of the number of correctly predicted samples in the total number of samples:

[0095]

[0096] Model precision: The ratio of the number of correctly predicted continuous casting roll fault samples TP to the positive class samples:

[0097]

[0098] Model recall: The ratio of the number of samples predicted as faults to the number of samples actually being faults:

[0099]

[0100] Calculate the F1-score: Considering the balance between Precision and Recall, the harmonic mean of the two is taken. The closer the value is to 1, the better the performance of Precision and Recall:

[0101]

[0102] Among them, TP is the number of correctly predicted continuous casting rolls as fault classes, FP is the number of mis-predicted continuous casting roll fault classes as non-fault classes, FN is the number of mis-predicted continuous casting roll non-fault classes as fault classes, TN is the number of correctly predicted continuous casting roll non-fault classes, and the sum of the four (TP + FP + FN + TN) is the total number of test samples m.

[0103] Table 1 Index parameters of the continuous casting roll dataset

[0104] Serial number Number Parameter name Parameter meaning description Unit 1 <![CDATA[p 1 > Continuous casting billet temperature Surface temperature of continuous casting billet ℃ 2 <![CDATA[p 2 > Segment position Sector segment position where the current continuous casting roll is located - 3 <![CDATA[p 3 > Inner and outer roller mark 0 is the inner frame roll, 1 is the outer frame roll - 4 <![CDATA[p 4 > Roll gap Distance between the upper and lower continuous casting rolls mm 5 <![CDATA[p 5 > Speed Rotational speed of continuous casting roll r / min 6 <![CDATA[p 6 > Fault-free online time Fault-free working time of the continuous casting roll in the current upper segment after the previous maintenance h 7 <![CDATA[p 7 > Cooling water inlet temperature Temperature when the cooling water enters the continuous casting roll core ℃ 8 <![CDATA[p 8 > Cooling water outlet temperature Temperature when the cooling water flows out of the continuous casting roll core ℃ 9 <![CDATA[p 9 > Cooling water flow rate Flow rate of the cooling water in the continuous casting roll core L / s 10 <![CDATA[p 10 > Cooling water pressure Water pressure of the cooling water in the continuous casting roll core MPa 11 <![CDATA[p 11 > Upper row pressure of the roll group Hydraulic cylinder pressure at the upper row of the roll group to which the continuous casting roll belongs MPa 12 <![CDATA[p 12 > Lower row pressure of the roll group Hydraulic cylinder pressure at the lower row of the roll group to which the continuous casting roll belongs MPa 13 <![CDATA[p 13 > Accumulative upper segment time Accumulative working time of the continuous casting roll in the upper segment h 14 <![CDATA[p 14 > Diameter Outer diameter of the continuous casting roll mm 15 <![CDATA[p 15 > Length Length of the continuous casting roll mm 16 <![CDATA[p 16 > Maintenance times Historical maintenance times of the continuous casting roll - 17 <![CDATA[p 17 > Billet drawing speed Speed at which the continuous casting billet is drawn out from the sector segment m / min

[0105] Table 2 Original dataset for continuous casting roll fault prediction

[0106] Continuous casting roll code <![CDATA[p 1 > … <![CDATA[p 8 > <![CDATA[p 9 > <![CDATA[p 10 > … <![CDATA[p 16 > <![CDATA[p 17 > 11010011 1050.06 … 78.50 1.15 5.10 … 5 1.2 11010011 1050.24 … 78.58 1.18 5.07 … 5 1.2 11010011 1050.55 … 77.96 1.31 5.08 … 5 1.2 … … … … … … … … … 11000011 1076.25 … 84.85 1.81 4.91 … 11 1.4 11000011 1076.43 … 86.44 1.90 4.99 … 11 1.4 11000011 1076.54 … 88.53 1.95 4.97 … 11 1.4

[0107] Table 3 Data after Z-score normalization processing and marking

[0108]

[0109]

[0110] Table 4 Z-score model parameter settings

[0111] Parameter name Value Learning rate 0.01 Training batch size 64 Number of iterations 400 Convolutional layer 4 Filter 64 Convolution kernel size 2 Dilation rate of each convolutional layer 1、2、4、8 Residual connection True LSTM cell 64 Activation function Sigmoid

[0112] Table 5 Evaluation Index Results

[0113]

Claims

1. A continuous casting roller fault prediction method based on a temporal convolutional network-long short-term memory network, characterized in that: The specific steps include: Step 1, construct a continuous casting roller dataset and divide the dataset into a training set and a test set; Step 2: Establish a fault prediction model; Step 3, using the training set divided in step 1 to train the sample fault prediction model constructed in step 2, and outputting the trained model; Step 4: Use the test set divided in step 1 to evaluate the model trained in step 3 and output the prediction results.

2. The continuous casting roller fault prediction method based on time convolution network-long short-term memory network according to claim 1 is characterized in that: The specific process of step 1 is as follows: Step 1.1, construct a continuous casting roller dataset based on the historical failure data of the continuous casting roller of the continuous casting machine, and use the Z-score standardization method to standardize the original data of the continuous casting roller dataset samples; Step 1.2, label the standardized data set, label the fault-free data as 0, and the faulty data as 1; Step 1.3, divide the labeled dataset in step 1.2 into 70% training set and 30% test set.

3. The continuous casting roller fault prediction method based on time convolution network-long short-term memory network according to claim 2 is characterized in that: In step 1.1, the following formulas (1) to (3) are used for standardization: Among them, z represents the standardized sample data; i represents the i-th sample in the continuous casting roller data set, N represents the total number of samples in the continuous casting roller data set; σ is the mean difference of a certain indicator parameter in the continuous casting roller failure samples; μ is the mean value of a certain indicator parameter in the continuous casting roller failure samples; x is the value of a certain indicator parameter in the continuous casting roller failure samples.

4. The continuous casting roller fault prediction method based on time convolution network-long short-term memory network according to claim 3 is characterized in that: The specific process of step 2 is as follows: Step 2.1: Build a Seq2Seq deep learning model based on TCN-LSTM, where the encoder module uses TCN, the decoder uses LSTM, and Self-Attention is introduced to process sequences of variable length; Step 2.2: convert the output vector y of step 2.1 into t Input the fully connected layer and use the sigmoid activation function to convert it into the fault probability value P(y t ), which means the probability of failure of the continuous casting roll.

5. The continuous casting roller fault prediction method based on time convolution network-long short-term memory network according to claim 4 is characterized in that: The specific process of step 2.1 is as follows: Step 2.1.1, create a TCN encoder model, and transform the continuous casting roller fault training set sample data sequence {p1, p2, ···, p 16 ,p 17 } Input TCN encoder model, output sequence Ht = {h1,h2,…h t-1 }; Step 2.1.2, for the output sequence H of step 2.1.1 t Use self-attention to dynamically adjust the weights and get the context vector D t ={d1,d2,…d t-1 }; Step 2.1.3, create an LSTM decoder model, and use the gated weighted fusion mechanism to transform the context vector D obtained in step 2.1.2 t The output H of the previous time step t Combined, as the input sequence X of the decoder LSTM t , the formula is as follows: X t =g t *D t +(1-g t )*H t (4) Among them, g t Dynamic contribution weights for context vectors; Using the memory unit and gating mechanism of the decoder LSTM, the vector y of the final time step is output t , the process is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) (5) the t =σ(W o *[h t-1 ,x t ]+b o ) (6) i t =σ(W i ·[h t-1 ,x t ]+b i ) (8) y t =o t *tanh(C t ) (11) Among them, f t is the forget gate weight, ranging from [0,1]; W f , W o , W i Respectively represent the weight matrices of the forget gate, output gate, and input gate, b f , b o , b i are the corresponding bias terms respectively; C t They are the candidate memory unit and the updated memory unit at the current moment respectively; x t is the input information at the current moment; h t-1 is the short-term hidden information of the previous moment; σ is the sigmoid function; b f is the bias vector of the forget gate; i t is the output of the forget gate.

6. The continuous casting roller fault prediction method based on time convolution network-long short-term memory network according to claim 5 is characterized in that: In step 2.2, the failure probability value P(y t ) is expressed as follows through the sigmoid activation function δ(y):

7. The continuous casting roller fault prediction method based on time convolution network-long short-term memory network according to claim 6 is characterized in that: The specific process of step 3 is as follows: Step 3.1, hyperparameter settings: including learning rate, training batch size, number of iterations, number of convolution layers, filters, convolution kernel size, convolution expansion rate per layer, residual connection, and number of LSTM units; Step 3.2, model training: Use the training set to train the fault model established in step 2, and continuously optimize the parameters in the model until the loss value of the loss function in the fault model tends to be stable.

8. The continuous casting roller fault prediction method based on time convolution network-long short-term memory network according to claim 7 is characterized in that: In step 4, the model accuracy, recall rate, precision rate and F1 score are calculated as evaluation indicators. The specific calculation process is as follows: Model accuracy: the proportion of correctly predicted samples in the total number of samples: Model accuracy: The ratio of the number of correctly predicted continuous casting roller failure samples to the number of positive samples: Model recall: the ratio of the number of samples predicted to be faulty to the number of samples that are actually faulty: Calculate the F1 score: Consider the balance between Precision and Recall and reconcile the two: Among them, TP is the number of continuous casting rollers correctly predicted as faulty, FP is the number of continuous casting rollers incorrectly predicted as fault-free, FN is the number of continuous casting rollers incorrectly predicted as faulty, TN is the number of continuous casting rollers correctly predicted as fault-free, and the sum of the four is the total number of test samples m.

Images