Temperature prediction control method for heat shrink tube winding motor

Abstract

The invention relates to a heat shrink tube winding motor temperature prediction control method, which comprises the following steps of: acquiring a historical data set, sequentially finishing data processing, data screening and dividing a training set and a test set, introducing a Circle optimization population, a nonlinear inertia weight and an asymmetric learning factor, and obtaining a more uniform and diversified initial population; according to the method, global search and local search capabilities of an algorithm are better balanced, parameters for constructing a CNN-LSTM model are optimized by using an IPSO algorithm, the generalization capability of the model is improved, the precision of the model is dynamically optimized by setting various evaluation index thresholds of the IPSO-CNN-LSTM model, optimal IPSO-CNN-LSTM prediction and residual error alarm thresholds are acquired, and related parameters of a heat shrink tube winding motor are acquired in real time. The sliding window algorithm is adopted to count the abnormal temperature frequency, and the optimal IPSO-CNN-LSTM prediction model is combined to realize the trend grading early warning of the temperature of the heat shrink tube winding motor, so that the problems of unstable parameters, large fluctuation and nonlinearity of the existing temperature monitoring of the heat shrink tube winding motor due to related influence are effectively solved, and high-precision advanced prediction and trend grading early warning are realized.

Description

Technical Field

[0001] The invention belongs to the technical field of temperature monitoring of heat shrink tube production equipment, and in particular relates to a temperature prediction and control method for a heat shrink tube winding motor. Background Art

[0002] At present, heat shrink tubing has the characteristics of insulation, sealing and protection, good chemical corrosion resistance, wear resistance, weather aging resistance, and as a marking tool for identification and classification. Heat shrink tubing has a wide range of applications, including but not limited to the wire and cable industry, electronic and electrical industry, automobile manufacturing and maintenance, chemical and petroleum industries, etc. Therefore, it is particularly urgent to improve and optimize the production process of heat shrink tubing.

[0003] The heat shrink tube winding process is one of the core processes for producing heat shrink tubes. The heat shrink tube winding process is not only related to the quality of the product, but also directly affects its use effect and life. Adjusting the winding speed can help improve the quality and efficiency of winding; the heat shrink tube winding mechanism can reduce the occurrence of folding, curling, and wrinkling problems, and further improve the quality of the finished heat shrink tube roll. As the core component of the heat shrink tube winding machine, the normal operation of the winding motor is crucial to the winding process of the heat shrink tube. The heat shrink tube winding motor is likely to cause motor temperature failure due to long-term operation and high-speed rotation. The causes of the failure may include internal motor failure, power supply problems, and the triggering of the overheating protection device. The key to solving the temperature failure of the winding motor is to diagnose the cause of the failure and perform corresponding repairs or replacements. At present, the production site of the heat shrink tube winding motor temperature failure basically relies on the experience of maintenance operators to solve it, which has a certain lag. There is a lack of methods to predict the health status of the winding motor in advance and monitor it in real time, resulting in the failure of the heat shrink tube winding motor abnormality cannot be solved in time, resulting in low overall winding quality of the heat shrink tube, and further causing a decrease in product yield. Therefore, how to achieve advance prediction and real-time monitoring of the temperature of the heat shrink tube winding motor is an urgent problem to be solved to improve the quality of the heat shrink tube winding process. Summary of the invention

[0004] The purpose of the present invention is to overcome the problems existing in the prior art and to provide a method for predicting and controlling the temperature of a heat shrink tube winding motor.

[0005] In order to achieve the above technical objectives and the above technical effects, the present invention is implemented through the following technical solutions:

[0006] A method for predicting and controlling the temperature of a heat shrink tube winding motor, the method comprising the following steps:

[0007] Step S1: historical data set collection;

[0008] Step S2: data processing;

[0009] Step S3: data screening;

[0010] Step S4: Divide the training data set and the test data set;

[0011] Step S5: construct an improved CNN-LSTM network model structure;

[0012] Step S6: Initialize particle swarm parameters and speed position information;

[0013] Step S7: population optimization;

[0014] Step S8: Optimizing inertia weight and learning factor;

[0015] Step S9: Update the particle's velocity, position, inertia weight and learning factor;

[0016] Step S10: input the parameters into the CNN-LSTM neural network for training;

[0017] Step S11: define fitness function;

[0018] Step S12: Calculate the particle fitness and update the optimal position;

[0019] Step S13: Model calculation and judgment, if the maximum number of iterations is met, continue to execute downward; otherwise, repeat step S9;

[0020] Step S14: Obtaining optimal parameters;

[0021] Step S15: The optimal parameters are input into the IPSO-CNN-LSTM prediction model;

[0022] Step S16: IPSO-CNN-LSTM model training;

[0023] Step S17: input the historical test set evaluation model and set the thresholds of various model evaluation indicators;

[0024] Step S18: Calculate various evaluation indicators of the IPSO-CNN-LSTM model and compare them with the corresponding thresholds;

[0025] Step S19: output the optimal IPSO-CNN-LSTM prediction model; at the same time, the Z-Score standardized denormalization method is used to process the residual alarm threshold δ;

[0026] Step S20: Using a sliding window algorithm to calculate the frequency of abnormal temperature of the heat shrink tube winding motor and to issue a graded warning.

[0027] Furthermore, the step S20 includes the following specific steps:

[0028] Step S20.1: real-time data collection;

[0029] Step S20.2: Initialize the current temperature anomaly frequency in the sliding window: f T =0;

[0030] Step S20.3: inputting the temperature-related parameters of the heat shrink tube winding motor collected in real time within a time period into the optimal IPSO-CNN-LSTM prediction model;

[0031] Step S20.3: inputting the temperature-related parameters of the heat shrink tube winding motor collected in real time within a time period into the optimal IPSO-CNN-LSTM prediction model;

[0032] Step S20.4: Output the corresponding predicted temperature set of the heat shrink tube winding motor: At the same time, the mean absolute error between the predicted value and the actual value of the heat shrink tube winding motor temperature is calculated:

[0033] Step S20.5: Use the Z-Score standardization and denormalization method to process the predicted temperature set of the heat shrink tube winding motor separately The mean absolute error Output corresponding to T ΔT And the mean absolute error MAE update ;

[0034] Step S20.6: If T i >maxT, where T i ∈T ΔT ; maxT is the upper limit of temperature, corresponding to the temperature limit with a larger value exceeding the threshold. Generally, the temperature provided by the heat shrink tube winding motor is combined with the various evaluation index settings of the optimal IPSO-CNN-LSTM prediction model, and jump to step S20.10; otherwise, continue to execute step S20.7;

[0035] Step S20.7: Set a sliding window length of L=10, and slide and traverse step by step along the prediction time.

[0036] Step S20.8: If MAE update If the residual error is less than the alarm threshold δ, the current temperature anomaly frequency in the sliding window is increased by 1, that is, f T =f T +1; otherwise, the current temperature anomaly frequency in the sliding window remains the original value, i.e. f T =f T +0;

[0037] Step S20.9: If the current temperature abnormality frequency f in the sliding window T ≥6, execute step S20.10; otherwise, jump to step S20.11;

[0038] Step S20.10: The heat shrink tube winding machine monitoring and warning system automatically displays a first-level warning, i.e., a red warning, indicating that the heat shrink tube winding motor is in an abnormal state, and sends a command to the winding machine equipment maintenance operator to stop the machine in time and handle the abnormality;

[0039] Step S20.11: If the current temperature abnormality frequency in the sliding window is 2≤f T ≤5, execute step S20.12; otherwise, jump to step S20.13;

[0040] Step S20.12: The heat shrink tube winding machine monitoring and warning system automatically displays a second-level warning, i.e., a yellow warning, indicating that an abnormality occurs in the heat shrink tube winding motor. A command is sent to the winder equipment maintenance operator to go to the equipment operation area to observe the real-time operation status of the equipment, evaluate the operation status, and then decide whether to take corresponding abnormality handling measures;

[0041] Step S20.13: The monitoring and early warning system of the heat shrink tube winding machine automatically displays green, indicating that the equipment temperature is normal; jump to step S20.2, and perform the next round of sliding window algorithm statistics on the frequency of abnormal temperature of the heat shrink tube winding motor for graded early warning.

[0042] Furthermore, in the step S1, during the historical data set collection process, the data interval is located during the normal operation of the heat shrink tube winder to ensure the validity of the data; the data source obtained is the winding motor speed, winding motor active power, winding motor voltage, winding motor temperature, winding machine tension, ambient temperature and ambient humidity parameters at the same time, and the parameter data at the same sampling time is 1 group, the sampling interval is set to 30s, and the heat shrink tube winder is collected continuously for 2 months to ensure a sufficient number of samples.

[0043] Furthermore, in step S2, in order to improve data quality, remove redundant information, process missing values ​​and normalize data, the following steps are specifically included:

[0044] Step S2.1: Data cleaning; use the moving window combined with the abnormal threshold of various parameters, set the moving window size to n, first, traverse all the collected parameters, find the duplicate values ​​in the data set, and delete the duplicate values; secondly, compare the abnormal threshold corresponding to the data in n consecutive time steps to determine whether the number of abnormalities is less than m. If so, use Newton polynomial interpolation to interpolate; otherwise, directly delete other characteristic parameter data in the same time step containing the abnormal data parameter; the Newton interpolation calculation formula is as follows: N n (x)=f(x0)+f[x0,x1](x-x0)+f[x0,x1,x2](x-x0)(x-x1)+...+f[x0,x1,...,x n ](x-x0)(x-x1)…(xxn-1 ) (1)

[0045]

[0046] , in formula (1) to (2), (x i ,f(x i )) represents the normal data of samples in window n, f[x0, x1, ..., x n ] represents the n-order difference quotient. For the difference quotient f[x0, x1, ..., x n ], the order of the serial number subscripts does not affect the value of the difference quotient, x i The time coordinate of the abnormal data, N n (x) represents the interpolation value calculated by Newton interpolation. In actual interpolation, only the result N of the interpolation after the last sampling needs to be calculated. n-1 (x) Save and directly calculate N next time n-1 (x) and the new term to save computation and storage space;

[0047] Step S2.2: Z-Score data normalization processing; data normalization can accelerate the speed of gradient descent optimal solution during model training and improve model accuracy. The data is uniformly mapped in the interval [0, 1]. The calculation formula is as follows:

[0048] In formula (3), y is the data value, that is, the initial value before data normalization. is the mean value of the data set, σ is the standard value, and y′ is the corresponding value of the normalized data.

[0049] Furthermore, in step S3, the parameters strongly correlated with the heat shrink tube winding motor are screened out through Pearson correlation coefficient calculation and analysis, namely, winding motor speed, winding motor active power, winding motor voltage, winding machine tension, ambient temperature and winding motor temperature, a total of 6 groups of parameters, as subsequent model training and test sample data, the screened sample data are integrated into the following matrix expression:

[0050] In formula (10), Z∈R 6×s , represents the sample data matrix after the integration of 6 sets of historical data, where n is the winding motor speed set, P is the winding motor active power set, U is the winding motor voltage set, F is the winding motor tension set, K is the ambient temperature set, T is the winding motor temperature set, and s is the sth time step of the data set.

[0051] Furthermore, in step S4, the preprocessed Z matrix sample data set is randomly divided into a training set and a test set according to conditions, wherein the training set accounts for 80% and is used to train the improved CNN-LSTM model; the test set accounts for 20% and is used to evaluate the improved CNN-LSTM model;

[0052] In step S5, the network model structure includes two one-dimensional convolutional layers, one fully connected layer and two LSTM layers. After the convolutional layer, the BN layer and the activation function Leaky-ReLU are used to realize data normalization and nonlinear operation, thereby accelerating the model convergence speed and solving the gradient vanishing problem, and extracting the deep-level features of the factors affecting the temperature of the winding motor; the extracted deep-level features are input into the LSTM to better predict the feature information.

[0053] Furthermore, in step S5, the parameters of the convolution layer are the weights of the convolution kernel and the offsets of each channel. The mathematical model of the convolution operation of the convolutional neural network CNN is as follows:

[0054] In formula (11), represents the j-th output feature map of the m-th layer, M m-1 represents the input feature set, m represents the mth layer of CNN, Represents the convolution kernel used by the m-th layer convolution operation, represents the bias of the j-th feature map, * represents the convolution operation, and f(·) represents the activation function;

[0055] The pooling layer performs a pooling operation on the features extracted by the convolution layer to achieve feature dimensionality reduction and reduce network parameters, and to a certain extent prevent the model from overfitting. The maximum pooling is used to seek the maximum value of the pooling area to obtain local features. The obtained feature map is more sensitive to texture features. For the output X of the kth filter in the i-th dimension of the convolution layer, i , the maximum pooling operation is expressed as follows:

[0056] In formula (12), p i (j) is the jth output of the pooling layer, and w is the width of the pooling kernel;

[0057] As a variant of the recurrent neural network RNN, LSTM introduces a gating unit to alleviate the gradient vanishing and gradient exploding problems in RNN training and improve prediction accuracy. At time t, the current memory unit C t It is the core part of LSTM neurons. The network's gate structure consists of a forget gate, an input gate, and an output gate, which determines whether information reaches C. t In one time step, the LSTM neuron calculates C through a series of calculations. tAnd the values ​​of other state quantities in the unit, the specific calculation formula is as follows:

[0058] f t =σ(W f [h t-1 , x t ]+b f ) (13)

[0059] i t =σ(W i [h t-1 , x t ]+b i ) (14)

[0060] o t =σ(W o [h t-1 , x t ]+b o ) (15)

[0061]

[0062] h t =o t ⊙tanh(C t ) (18), in formulas (13) to (18), f t 、i t , o t are the states of the input gate, forget gate, and output gate at time t, respectively. t-1 is the memory unit of LSTM at time t-1; and C t are the candidate memory unit of LSTM at the current moment t and the memory unit at the current moment, h t-1 is the output value of LSTM at time t-1, x t is the input value of LSTM at time t; W i , W f , W o , W C and b i 、b f 、b o 、b C are the weights and biases of the input gate, forget gate, output gate, and candidate memory, σ(·) is the sigmoid activation function, tanh(·) is the hyperbolic tangent function; ⊙ is the bitwise multiplication of each element.

[0063] Furthermore, the step S6 specifically includes the following steps:

[0064] Step S6.1: Initialize IPSO hyperparameters and determine the initial values ​​of IPSO parameters;

[0065] Step S6.2: Initialize particle positions and velocities, and randomly generate populations: x i =(n1,m1,m2,m3,α,p,batch size ), where n1 is the number of CNN convolution kernels, m1 and m2 are the number of neurons in the two LSTM hidden layers, m3 is the number of neurons in the fully connected layer, α is the initialization learning rate, p is the neuron inactivation rate, and batch size is the batch size (BatchSize), and the value ranges of the above parameters are as follows:

[0066] n1∈[2, 64], m1∈[1, 30], m2∈[1, 30], m3∈[1, 20], α∈[0.001, 0.005], p∈[0.01, 0.90], batch size ∈[1,50].

[0067] Furthermore, in step S7, the Circle mapping method is introduced to initialize the particle swarm to obtain a more uniform and diverse initial population to improve the convergence speed and accuracy of the algorithm. The expression of the chaotic sequence generated by the Circle mapping is as follows:

[0068]

[0069] In, X i represents the number of the i-th chaotic sequence, X i+1 represents the number of the i+1th chaotic sequence, e=0.5, f=0.2, mod represents the modulus operator, and the population initialization operation using Circle mapping includes the initialization of velocity and position. The specific initial operations are as follows:

[0070] v i，j =v lower_b +(v up_b -v lower_b )*X i，j (twenty two)

[0071] x i，j -x lower_b +(x up_b -x lower_b )*x′ i，j (25), in formulas (22) to (23), v lower_b and v up_b are the upper and lower bounds of the particle velocity, x lower_b and x up_b are the upper and lower bounds of the particle position, respectively, i，j and X′ i，j They are the chaotic sequence values ​​generated by the Circle mapping on the corresponding particle dimensions.

[0072] Furthermore, the step S8 specifically includes the following steps:

[0073] Step S8.1: Introduce the dynamic nonlinear change inertia weight, and select the dynamic nonlinear change inertia weight ω, as follows:

[0074]

[0075] Where ω(k) is the inertia weight of the kth iteration, ω max and ω min The maximum and minimum inertia weights are 0.9 and 0.4, respectively. max is the maximum number of iterations, k is the number of iterations;

[0076] Step S8.2: Introduce asymmetric learning factors, including individual learning factor c1 and social learning factor c2. c1 and c2 are adaptively adjusted according to the different search periods. The specific expressions are as follows:

[0077]

[0078] In formulas (25) and (26), c1(k) and c2(k) are the individual learning factor and social learning factor of the kth iteration, respectively. 1_e 、c 1_s The initial value and final value of the individual learning factor are 2.5 and 0.5, respectively. 2_e 、c 2_s The initial value of social learning factor is 1 and the final value is 2.25, t max is the maximum number of iterations, and k is the number of iterations.

[0079] Furthermore, in step S11, the optimized initial particles are used as parameters of the CNN-LSTM model, and the mean square error MSE is used as the fitness of the IPSO algorithm, as follows:

[0080] In formula (27): is the predicted value, y i is the actual value, and N is the number of samples.

[0081] Furthermore, in step S17, the mean absolute error (MAE) threshold (MAE) is introduced. th_max , mean absolute percentage error MAPE threshold MAPE th_max , RMSE threshold RMSE th_max and the coefficient of determination R 2 Threshold R 2 th_minEvaluation indicators comprehensively evaluate the prediction effect of IPSO-CNN-LSTM model.

[0082] Furthermore, the step S18 specifically includes the following steps:

[0083] Step S18.1: Calculate the evaluation index, calculate and evaluate various indicators of the IPSO-CNN-LSTM model, as follows:

[0084]

[0085] In formulas (28) to (31), is the predicted value, y i is the actual value, is the mean, N is the number of samples;

[0086] Step S18.2: Threshold range judgment, compare the various evaluation indicators of the IPSO-CNN-LSTM model in turn to see if they are within the specified threshold range, as follows:

[0087] If the requirements of the above inequality group (32) are met, it means that the prediction accuracy of the IPSO-CNN-LSTM model is high. At this time, the MAE at the end of the model training is recorded. find As the absolute upper limit of the predicted residual alarm threshold, continue to execute downward; otherwise, repeat step S14.

[0088] Furthermore, in step S19, the specific calculation formula is as follows:

[0089] δ = MAE final ×σ test +μ test (33), where MAE final is the final mean absolute error MAE of the optimal IPSO-CNN-LSTM prediction model, δ test is the standard deviation of the test set, μ test is the mean of the test set.

[0090] The beneficial effects of the present invention are:

[0091] The method model of the present invention combines the sliding window algorithm to count the frequency of temperature anomalies and realizes the intelligent graded warning of the real-time collected temperature of the heat shrink tube winding motor. In view of the shortcomings of the traditional PSO algorithm, by introducing the Circle optimization population, a more uniform and diverse initial population is obtained, which is conducive to improving the convergence speed and accuracy of the algorithm; by introducing nonlinear inertia weights, the global search and local search capabilities of the algorithm can be better balanced. By introducing asymmetric learning factors, particles have stronger global convergence capabilities during search. For the highly nonlinear heat shrink tube winding motor temperature-related parameter data, the IPSO algorithm is used to optimize the parameters of the CNN-LSTM model, so that the model avoids the local optimal situation and improves the generalization ability of the model. Aiming at the fact that the temperature warning of the heat shrink tube winding motor has a certain lag, the real-time collection of relevant parameters, such as winding motor temperature, winding motor speed, winding motor active power, winding motor voltage, winding machine tension, ambient temperature, etc., is adopted. The sliding window algorithm is used to count the frequency of temperature anomalies and the optimal IPSO-CNN-LSTM prediction model is combined to realize the trend warning of the temperature of the heat shrink tube winding motor. This effectively solves the problem that the existing winding motor temperature monitoring cannot achieve high-precision advance prediction and trend classification warning due to the unstable, large fluctuation and nonlinear characteristics of the winding setting parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 It is a schematic diagram of the CNN-LSTM network structure of the present invention;

[0093] Figure 2 It is a schematic diagram of the LSTM neuron structure of the present invention;

[0094] Figure 3 It is a schematic diagram of the temperature prediction and monitoring process of the heat shrink tube winding motor based on IPSO-CNN-LSTM of the present invention;

[0095] Figure 4 It is a schematic diagram of the process of graded warning of temperature anomaly frequency statistics using the sliding window algorithm of the present invention. DETAILED DESCRIPTION

[0096] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0097] A method for predicting and controlling the temperature of a heat shrink tube winding motor, the method comprising the following steps:

[0098] Step S1: historical data set acquisition. All data of the embodiment of the present invention are taken from a SCADA system supporting a heat shrinkable tube winder. The data interval is from May 1, 2024 to June 30, 2024 during normal operation. According to the operation characteristics of the heat shrinkable tube winder and the coupling characteristics between the main parameters, the data sources obtained are seven types of parameters at the same time, including the speed of the winder motor, the active power of the winder motor, the voltage of the winder motor, the temperature of the winder motor, the tension of the winder, the ambient temperature, and the ambient humidity. The parameter data at the same sampling time is 1 group, and the sampling interval between each group of data is 30s, totaling 164160 groups of data.

[0099] Step S2: Data processing, in order to improve data quality, remove redundant information, process missing values ​​and normalize data, specifically includes the following steps:

[0100] Step S2.1: Data cleaning; The abnormal data in the original data are mainly divided into two categories: duplication and missing, which will cause deviations in the model. The moving window combined with the abnormal threshold of various parameters is used, and the moving window size is set to n. First, all the collected parameters are traversed to find the duplicate values ​​in the data set and delete the duplicate values; secondly, the abnormal threshold corresponding to the data in n consecutive time steps is compared to determine whether the number of abnormalities is less than m. If so, the Newton polynomial interpolation method is used for interpolation; otherwise, other characteristic parameter data in the same time step containing the abnormal data parameters are directly deleted; the Newton interpolation calculation formula is as follows: N n (x)=f(x0)+f[x0,x1](x-x0)+f[x0,x1,x2](x-x0)(x-x1)+...+f[x0,x1,...,x n ](x-x0)(x-x1)…(xx n-1 ) (1)

[0101]

[0102] , in formula (1) to (2), (x i ,f(x i )) represents the normal data of samples in window n, f[x0, x1, ..., x n ] represents the n-order difference quotient. For the difference quotient f[x0, x1, ..., x n ], the order of the subscripts does not affect the value of the difference quotient, x i The time coordinate of the abnormal data, N n (x) represents the interpolation value calculated by Newton interpolation. In actual interpolation, only the result N of the interpolation after the last sampling needs to be calculated. n-1 (x) Save and directly calculate N next time n-1(x) and the sum of the new term can be used to save computing and storage space. In this embodiment, n=10, m=2;

[0103] Step S2.2: Z-Score data normalization processing; data normalization can accelerate the speed of gradient descent optimal solution during model training and improve model accuracy. The data is uniformly mapped in the interval [0, 1]. The calculation formula is as follows:

[0104] In formula (3), y is the data value, that is, the initial value before data normalization. is the mean value of the data set, σ is the standard value, and y′ is the corresponding value of the normalized data;

[0105] Step S3: Data screening: Pearson correlation coefficient is calculated for the normalized winding motor speed, winding motor active power, winding motor voltage, winding machine tension, ambient temperature, and ambient humidity. Parameters with strong correlation with winding motor temperature are selected for subsequent model training. The correlation coefficient is calculated as follows:

[0106]

[0107] In formulas (4) to (9), s represents the number of data sets, n j 、n i They represent the jth speed and the ith speed in the winding motor speed data set n respectively; P j , P i They represent the jth active power and the ith active power in the active power data set P of the winding motor respectively; U j , U i They represent the j-th voltage and the i-th voltage of the winding motor voltage data set U respectively; F j 、F i They represent the jth tension and the ith tension of the winder tension data set F respectively; AH j , AH i They represent the j-th ambient temperature and the i-th ambient temperature of the ambient temperature dataset AH respectively; K j , K i They represent the j-th and i-th ambient humidity of the ambient humidity dataset K respectively; ρ nT , PT , UT , FT , AHT , KTThey respectively represent the Pearson correlation coefficient between the motor speed set n and the motor temperature set T, the Pearson correlation coefficient between the motor active power set P and the motor temperature set T, the Pearson correlation coefficient between the motor voltage set U and the motor temperature set T, the Pearson correlation coefficient between the tension set F and the motor temperature set T, the Pearson correlation coefficient between the ambient humidity K and the motor bearing set T, and the Pearson correlation coefficient between the ambient temperature AT and the motor bearing temperature set T; Table 1 specifically gives the correlation coefficient values ​​and significance of the Pearson correlation analysis of each parameter with the winding motor temperature:

[0108] Table 1 Pearson correlation coefficient analysis of each parameter and motor bearing temperature

[0109] Parameter Category Correlation coefficient (ρ) Significance level Winding motor speed 0.912 0.000 Active power of winding motor 0.855 0.000 Winding motor voltage 0.824 0.000 Winder tension 0.645 0.000 Ambient humidity 0.158 -0.256 Ambient temperature 0.769 0.000

[0110] By comparing the size of the correlation coefficient in Table 1, generally the Pearson correlation coefficient in [0.6, 0.8] indicates a strong correlation; [0.8, 1.0] indicates an extremely strong correlation; and less than 0.6 indicates a low correlation strength. Therefore, a total of 6 groups of parameters, including winding motor speed, winding motor active power, winding motor voltage, winding machine tension, ambient temperature and winding motor temperature, are selected as sample data for subsequent model training and testing; the selected sample data are integrated into the following matrix expression:

[0111] In formula (10), Z∈R 6×s , represents the sample data matrix after the integration of 6 sets of historical data, where n is the winding motor speed set, P is the winding motor active power set, U is the winding motor voltage set, F is the winding motor tension set, K is the ambient temperature set, T is the winding motor temperature set, and s is the sth time step of the data set;

[0112] Step S4: Divide the training data set and the test data set. The preprocessed Z matrix sample data set is randomly divided into a training set and a test set according to conditions, wherein the training set accounts for 80% and is used to train the improved CNN-LSTM model; the test set accounts for 20% and is used to evaluate the improved CNN-LSTM model;

[0113] Step S5: Construct an improved CNN-LSTM network model structure, such as Figure 1As shown, in order to solve the problem of insufficient feature extraction capability of the traditional winding motor temperature prediction model, this embodiment introduces a winding motor temperature prediction model based on the combination of CNN and LSTM. The network model structure includes 2 one-dimensional convolution layers, 1 fully connected layer and 2 LSTM layers. After the convolution layer, the BN layer and the activation function Leaky-ReLU are used to realize data normalization and nonlinear operation, thereby accelerating the model convergence speed and solving the gradient vanishing problem, and extracting the deep-level features of the factors affecting the winding motor temperature; the extracted deep features are input into the LSTM to better predict the feature information. The convolutional neural network (CNN) generally includes an input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer. The convolution layer is the core module of CNN, and its parameters are the weights of the convolution kernel and the offset of each channel. The mathematical model of the convolution operation of the convolutional neural network CNN is as follows:

[0114] In formula (11), represents the j-th output feature map of the m-th layer, M m-1 represents the input feature set, m represents the mth layer of CNN, Represents the convolution kernel used by the m-th layer convolution operation, represents the bias of the j-th feature map, * represents the convolution operation, and f(·) represents the activation function. In this embodiment, the Leaky-ReLU activation function is used;

[0115] The pooling layer performs a pooling operation on the features extracted by the convolution layer to achieve feature dimensionality reduction and reduce network parameters, and to a certain extent prevent the model from overfitting. The maximum pooling is used to seek the maximum value of the pooling area to obtain local features. The obtained feature map is more sensitive to texture features. For the output X of the kth filter in the i-th dimension of the convolution layer, i ,The maximum pooling operation is expressed as follows:

[0116] In formula (12), p i (j) is the jth output of the pooling layer, and w is the width of the pooling kernel;

[0117] As a variant of the recurrent neural network (RNN), LSTM introduces gated units to alleviate the gradient vanishing and gradient exploding problems in RNN training and improve prediction accuracy. Figure 2 As shown, this is the internal structure of an LSTM neuron at time t, where the current memory unit C t It is the core part of LSTM neurons. The network's gate structure consists of a forget gate, an input gate, and an output gate, which determines whether information reaches C. tIn one time step, the LSTM neuron calculates C through a series of calculations. t And the values ​​of other state quantities in the unit, the specific calculation formula is as follows:

[0118] f t =σ(W f [h t-1 , x t ]+b f ) (13)

[0119] i t =σ(W i [h t-1 , x t ]+b i ) (14)

[0120] o t =σ(W o [h t-1 , x t ]+b o ) (15)

[0121]

[0122] h t =o t ⊙tanh(C t ) (18), in formulas (13) to (18), f t 、i t , o t are the states of the input gate, forget gate, and output gate at time t, respectively. t-1 is the memory unit of LSTM at time t-1; and C t are the candidate memory unit of LSTM at the current moment t and the memory unit at the current moment, h t-1 is the output value of LSTM at time t-1, x t is the input value of LSTM at time t; W i , W f , W o , W C and b i 、b f 、b o 、b C are the weights and biases of the input gate, forget gate, output gate, and candidate memory, σ(·) is the sigmoid activation function, tanh(·) is the hyperbolic tangent function; ⊙ is the bitwise multiplication of each element;

[0123] Step S6: Initialize the particle swarm parameters and velocity position information, and build the IPSO-CNN-LSTM model. The process of the model is as follows: Figure 3 As shown, when CNN-LSTM is used to predict the temperature of the winding motor, its accuracy is strongly correlated with the number of convolution kernels, the number of neurons in the fully connected layer, and the number of hidden neurons in the LSTM. Therefore, the present invention uses an improved particle swarm optimization (PSO) algorithm (the improved PSO algorithm is referred to as "IPSO") to optimize these parameters and improve the generalization ability of the model;

[0124] PSO is a swarm intelligence optimization algorithm that obtains the global optimal value through the collaboration and information sharing between particles to solve multi-objective, nonlinear and multi-variable problems. Each particle is composed of three indicators: speed, position and fitness value. The size of the fitness value indicates the quality of the particle. This value is calculated based on the fitness function. Suppose the space dimension is N, the number of particles is m, and the particle positions are x1, x2, ..., x m , the corresponding speeds are v1, v2, …, v m ; At time t, the position of the jth particle is X j =(x j1 , x j2 , …, x jN ), the speed corresponds to v j =(v j1 , v j2 , …, v jN ), the i-th particle searches for the local optimal position and is recorded as p i =(p i1 , p i2 ,…,p iN ), all particles search to the global optimal position and record it as p g =(p g1 , p g2 ,…,p gN ), the formula for updating the position and velocity of the jth component of the ith particle is as follows:

[0125] v ij (k+1)=ωv ij (k))+c1*r1[p ij (k)-x ij (k)]+c2*r2[p gj (k)-x ij (k)] (19)

[0126] x ij (k+1)=x ij (k)+v ij (k+1) (20)

[0127] In formulas (19) to (20), k is the number of iterations, ω is the inertia weight, which is used to balance and adjust the local search and global search capabilities of the PSO algorithm, r1 and c2 are learning factors, which reflect the ability of particle self-learning and social learning, and are generally set to 2. r1 and r2 are random numbers distributed in [0, 1];

[0128] Step S7: population optimization;

[0129] Step S8: Optimize the inertia weight and learning factor. The inertia weight and learning factor in the PSO algorithm are usually set to a constant or linearly decreasing value, which may cause the PSO algorithm to fall into the local optimal stage prematurely, causing a trade-off between the global and local search capabilities. Therefore, the particle swarm optimization algorithm is improved from the perspective of the inertia weight and learning factor.

[0130] Step S9: Update the particle's velocity, position, inertia weight and learning factor;

[0131] Step S10: input the parameters into the CNN-LSTM neural network for training;

[0132] Step S11: define fitness function;

[0133] Step S12: Calculate the particle fitness and update the optimal position;

[0134] Step S13: Model calculation and judgment, if the maximum number of iterations is met, continue to execute downward; otherwise, repeat step S9;

[0135] Step S14: Obtain optimal parameters, including the number of convolution kernels, the number of neurons, the learning rate, the number of iterations, the batch size, and other optimal parameters optimized by IPSO;

[0136] Step S15: input the optimal parameters found by the IPSO algorithm into the IPSO-CNN-LSTM prediction model;

[0137] Step S16: IPSO-CNN-LSTM model training;

[0138] Step S17: input the historical test set evaluation model and set the thresholds of various model evaluation indicators;

[0139] Step S18: Calculate various evaluation indicators of the IPSO-CNN-LSTM model and compare them with the corresponding thresholds;

[0140] Step S19: output the optimal IPSO-CNN-LSTM prediction model; at the same time, the Z-Score standardized denormalization method is used to process the residual alarm threshold δ;

[0141] Step S20: Using a sliding window algorithm to calculate the frequency of abnormal temperature of the heat shrink tube winding motor and to issue a graded warning.

[0142] In step S20, the process is as follows Figure 4 As shown, the specific steps include:

[0143] Step S20.1: real-time data collection;

[0144] Step S20.2: Initialize the current temperature anomaly frequency in the sliding window: f T =0;

[0145] Step S20.3: inputting the temperature-related parameters of the heat shrink tube winding motor collected in real time within a time period into the optimal IPSO-CNN-LSTM prediction model;

[0146] Step S20.3: inputting the temperature-related parameters of the heat shrink tube winding motor collected in real time within a time period into the optimal IPSO-CNN-LSTM prediction model;

[0147] Step S20.4: Output the corresponding predicted temperature set of the heat shrink tube winding motor: At the same time, the mean absolute error between the predicted value and the actual value of the heat shrink tube winding motor temperature is calculated:

[0148] Step S20.5: Use the Z-Score standardization and denormalization method to process the predicted temperature set of the heat shrink tube winding motor separately The mean absolute error Output corresponding to T ΔT And the mean absolute error MAE update ;

[0149] Step S20.6: If T i >maxT, where T i ∈T ΔT ; maxT is the upper limit of temperature, corresponding to the temperature limit with a larger value exceeding the threshold. Generally, the temperature provided by the heat shrink tube winding motor is combined with the various evaluation index settings of the optimal IPSO-CNN-LSTM prediction model, and jump to step S20.10; otherwise, continue to execute step S20.7;

[0150] Step S20.7: Set a sliding window length of L=10, and slide and traverse step by step along the prediction time.

[0151] Step S20.8: If MAE update If the residual error is less than the alarm threshold δ, the current temperature anomaly frequency in the sliding window is increased by 1, that is, f T =f T +1; otherwise, the current temperature anomaly frequency in the sliding window remains the original value, i.e. f T =f T+0;

[0152] Step S20.9: If the current temperature abnormality frequency f in the sliding window T ≥6, execute step S20.10; otherwise, jump to step S20.11;

[0153] Step S20.10: The heat shrink tube winding machine monitoring and warning system automatically displays a first-level warning, i.e., a red warning, indicating that the heat shrink tube winding motor is in an abnormal state, and sends a command to the winding machine equipment maintenance operator to stop the machine in time and handle the abnormality;

[0154] Step S20.11: If the current temperature abnormality frequency in the sliding window is 2≤f T ≤5, execute step S20.12; otherwise, jump to step S20.13;

[0155] Step S20.12: The heat shrink tube winding machine monitoring and warning system automatically displays a second-level warning, i.e., a yellow warning, indicating that an abnormality occurs in the heat shrink tube winding motor. A command is sent to the winder equipment maintenance operator to go to the equipment operation area to observe the real-time operation status of the equipment, evaluate the operation status, and then decide whether to take corresponding abnormality handling measures;

[0156] Step S20.13: The monitoring and early warning system of the heat shrink tube winding machine automatically displays green, indicating that the equipment temperature is normal; jump to step S20.2, and perform the next round of sliding window algorithm statistics on the frequency of abnormal temperature of the heat shrink tube winding motor for graded early warning.

[0157] The step S6 specifically includes the following steps:

[0158] Step S6.1: Initialize IPSO hyperparameters and determine the initial values ​​of IPSO parameters, as shown in Table 2:

[0159] Table 2 IPSO parameter settings

[0160] IPSO Parameters Parameter Value Population size 30 Particle Dimension 7 Maximum number of iterations 50 <![CDATA[Initial value c of individual learning factor 1_s > 2.5 <![CDATA[Final value c of individual learning factor 1_e > 0.5 <![CDATA[Initial value c of social learning factor 2_s > 1 <![CDATA[Final value c of the social individual learning factor 2_e > 2.25 <![CDATA[Inertia weight ω max > 0.9 <![CDATA[Inertia weight ω min > 0.4

[0161] Step S6.2: Initialize particle positions and velocities, and randomly generate populations: x i =(n1,m1,m2,m3,α,p,batch size ), where n1 is the number of CNN convolution kernels, m1 and m2 are the number of neurons in the two LSTM hidden layers, m3 is the number of neurons in the fully connected layer, α is the initialization learning rate, p is the neuron inactivation rate (Dropout Rate), batch size is the batch size (Batch Size), and the value ranges of the above parameters are as follows:

[0162] n1∈[2, 64], m1∈[1, 30], m2∈[1, 30], m3∈[1, 20], α∈[0.001, 0.005], p∈[0.01, 0.90], batch size ∈[1,50].

[0163] In step S7, a random initialization operation is generally used to generate an initial population. However, the initial population obtained in this way has a fundamental limitation on the convergence performance of the algorithm due to the uneven distribution of individuals therein. Based on the characteristics of randomness, ergodicity and regularity of chaotic mapping, the initial population can be optimized. Therefore, in this embodiment, the Circle mapping method is selected to initialize the particle swarm to obtain a more uniform and diversified initial population to improve the convergence speed and accuracy of the algorithm. Circle mapping is a chaotic mapping that can be used to generate chaotic numbers between 0 and 1. The expression of the chaotic sequence generated by Circle mapping is as follows:

[0164]

[0165] In, X i represents the number of the i-th chaotic sequence, X i+1 represents the number of the i+1 chaotic sequence, e=0.5, f=0.2, mod represents the modulus operator, and the population initialization operation using Circle mapping includes the initialization of velocity and position. The specific initial operations are as follows:

[0166] v i，j =v lower_b +(v up_b -v lower_b )*X i，j (twenty two)

[0167] x i,j =x lower_b +(x up_b -x lower_b )*X′ i，j (23), in formulas (22) to (23), v lower_b and v up_b are the upper and lower bounds of the particle velocity, x lower_b and x up_b are the upper and lower bounds of the particle position, respectively, i，j and X′ i，j They are the chaotic sequence values ​​generated by the Circle mapping on the corresponding particle dimensions.

[0168] The step S8 specifically includes the following steps:

[0169] Step S8.1: Introduce a dynamic nonlinear inertia weight. The size of the inertia weight represents the ability of the particle to maintain the motion state of the previous moment. When ω is large, the global search ability of the PSO algorithm is guaranteed, but the convergence accuracy is low and the convergence speed is slow; when ω is small, the local search ability of the PSO algorithm is guaranteed, the convergence speed is fast, but it is easy to fall into the local optimum; when the inertia weight ω is a fixed value, the global search ability and convergence speed of the PSO are limited. In order to better balance the global search and local search capabilities of the PSO algorithm in the exploration stage and improve the convergence accuracy, this embodiment selects the dynamic nonlinear inertia weight ω, which is as follows:

[0170]

[0171] Where ω(k) is the inertia weight of the kth iteration, ω max and ω min The maximum and minimum inertia weights are 0.9 and 0.4, respectively. max is the maximum number of iterations, k is the number of iterations;

[0172] Step S8.2: Introduce asymmetric learning factors, including individual learning factor c1 and social learning factor c2. The larger the value of individual learning factor c1, the more favorable it is for overall search. The larger the value of social learning factor c2, the more favorable it is for local search. c1 and c2 are adaptively adjusted according to different search periods. The specific expressions are as follows:

[0173]

[0174] In formulas (25) and (26), c1(k) and c2(k) are the individual learning factor and social learning factor of the kth iteration, respectively. 1_e 、c 1_s The initial value and final value of the individual learning factor are 2.5 and 0.5, respectively. 2_e 、c 2_s The initial value of social learning factor is 1 and the final value is 2.25, t max is the maximum number of iterations, and k is the number of iterations.

[0175] In step S11, the optimized initial particles are used as the parameters of the CNN-LSTM model, and the mean square error MSE is used as the fitness of the IPSO algorithm, as follows:

[0176] In formula (27): is the predicted value, y i is the actual value, and N is the number of samples.

[0177] In step S12, p is calculated and determined. i and pg , according to equations (19) and (20), the individual optimum and global optimum of the population are updated.

[0178] In step S17, the mean absolute error (MAE) threshold (MAE) is introduced. th_max , mean absolute percentage error MAPE threshold MAPE th_max , RMSE threshold MASE th_max and the coefficient of determination R 2 Threshold R 2 th_min Evaluation indicators comprehensively evaluate the prediction effect of the IPSO-CNN-LSTM model. The specific thresholds of various evaluation indicators are shown in Table 3:

[0179] Table 3 Thresholds of various evaluation indicators of IPSO-CNN-LSTM model

[0180] <![CDATA[MAE th_max ]]> <![CDATA[MAPE th_max / %]]> <![CDATA[RMSEt h_max ]]> <![CDATA[R 2 th_min ]]> 0.5 4.5 0.5 0.9 .

[0181] The step S18 specifically includes the following steps:

[0182] Step S18.1: Calculate the evaluation index, calculate and evaluate various indicators of the IPSO-CNN-LSTM model, as follows:

[0183]

[0184] In formulas (28) to (31), is the predicted value, y i is the actual value, is the mean, N is the number of samples;

[0185] Step S18.2: Threshold range judgment, compare the various evaluation indicators of the IPSO-CNN-LSTM model in turn to see if they are within the specified threshold range, as follows:

[0186] If the requirements of the above inequality group (32) are met, it means that the prediction accuracy of the IPSO-CNN-LSTM model is high. At this time, the MAE at the end of the model training is recorded. find As the absolute upper limit of the predicted residual alarm threshold, continue to execute downward; otherwise, repeat step S14.

[0187] In step S19, the specific calculation formula is as follows:

[0188] δ = MAE final ×σ test +μ test (33), where MAE finalis the final mean absolute error MAE of the optimal IPSO-CNN-LSTM prediction model, δ test is the standard deviation of the test set, μ test is the mean of the test set.

[0189] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting and controlling the temperature of a heat shrink tube winding motor, characterized in that: The method comprises the following steps: Step S1: historical data set collection; Step S2: data processing; Step S3: data screening; Step S4: Divide the training data set and the test data set; Step S5: construct an improved CNN-LSTM network model structure; Step S6: Initialize particle swarm parameters and speed position information; Step S7: population optimization; Step S8: Optimizing inertia weight and learning factor; Step S9: Update the particle's velocity, position, inertia weight and learning factor; Step S10: input the parameters into the CNN-LSTM neural network for training; Step S11: define fitness function; Step S12: Calculate the particle fitness and update the optimal position; Step S13: Model calculation and judgment, if the maximum number of iterations is met, continue to execute downward; otherwise, repeat step S9; Step S14: Obtaining optimal parameters; Step S15: The optimal parameters are input into the IPSO-CNN-LSTM prediction model; Step S16: IPSO-CNN-LSTM model training; Step S17: input the historical test set evaluation model and set the thresholds of various model evaluation indicators; Step S18: Calculate various evaluation indicators of the IPSO-CNN-LSTM model and compare them with the corresponding thresholds; Step S19: output the optimal IPSO-CNN-LSTM prediction model; at the same time, the Z-Score standardized denormalization method is used to process the residual alarm threshold δ; Step S20: Using a sliding window algorithm to calculate the frequency of abnormal temperature of the heat shrink tube winding motor and to issue a graded warning.

2. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 1, characterized in that: The step S20 includes the following specific steps: Step S20.1: real-time data collection; Step S20.2: Initialize the current temperature anomaly frequency in the sliding window: f T =0; Step S20.3: inputting the temperature-related parameters of the heat shrink tube winding motor collected in real time within a time period into the optimal IPSO-CNN-LSTM prediction model; Step S20.3: inputting the temperature-related parameters of the heat shrink tube winding motor collected in real time within a time period into the optimal IPSO-CNN-LSTM prediction model; Step S20.4: Output the corresponding predicted temperature set of the heat shrink tube winding motor: At the same time, the mean absolute error between the predicted value and the actual value of the heat shrink tube winding motor temperature is calculated: Step S20.5: Use the Z-Score standardization and denormalization method to process the predicted temperature set of the heat shrink tube winding motor separately The mean absolute error Output corresponding to T ΔT And the mean absolute error MAE update ; Step S20.6: If T i >maxT, where T i ∈T ΔT ; maxT is the upper limit of temperature, corresponding to the temperature limit with a larger value exceeding the threshold. Generally, the temperature provided by the heat shrink tube winding motor is combined with the various evaluation index settings of the optimal IPSO-CNN-LSTM prediction model, and jump to step S20.10; otherwise, continue to execute step S20.7; Step S20.7: Set a sliding window length of L=10, and slide and traverse step by step along the prediction time. Step S20.8: If MAE update If the residual error is less than the alarm threshold δ, the current temperature anomaly frequency in the sliding window is increased by 1, that is, f T =f T +1; otherwise, the current temperature anomaly frequency in the sliding window remains the original value, i.e. f T =f T +0; Step S20.9: If the current temperature abnormality frequency f in the sliding window T ≥6, execute step S20.10; otherwise, jump to step S20.11; Step S20.10: The heat shrink tube winding machine monitoring and warning system automatically displays a first-level warning, i.e., a red warning, indicating that the heat shrink tube winding motor is in an abnormal state, and sends a command to the winding machine equipment maintenance operator to stop the machine in time and handle the abnormality; Step S20.11: If the current temperature abnormality frequency in the sliding window is 2≤f T ≤5, execute step S20.12; otherwise, jump to step S20.13; Step S20.12: The heat shrink tube winding machine monitoring and warning system automatically displays a second-level warning, i.e., a yellow warning, indicating that an abnormality occurs in the heat shrink tube winding motor. A command is sent to the winder equipment maintenance operator to go to the equipment operation area to observe the real-time operation status of the equipment, evaluate the operation status, and then decide whether to take corresponding abnormality handling measures; Step S20.13: The monitoring and early warning system of the tube shrinking and winding machine automatically displays green, indicating that the equipment temperature is normal; Jump to step S20.2 and perform the next round of sliding window algorithm to count the frequency of abnormal temperature of the heat shrink tube winding motor and provide graded warning.

3. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 2, characterized in that: In the step S1, during the historical data set collection process, the data interval is located during the normal operation of the heat shrink tube winder to ensure the validity of the data; the data source obtained is the winding motor speed, winding motor active power, winding motor voltage, winding motor temperature, winding machine tension, ambient temperature and ambient humidity parameters at the same time, and the parameter data at the same sampling time is 1 group, the sampling interval is set to 30s, and the heat shrink tube winder is collected continuously for 2 months to ensure a sufficient number of samples.

4. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 3, characterized in that: In step S2, in order to improve data quality, remove redundant information, process missing values ​​and normalize data, the following steps are specifically included: Step S2.1: Data cleaning; use a moving window combined with various parameter abnormality thresholds, set the moving window size to n, first, traverse all collected parameters, find duplicate values ​​in the data set, and delete the duplicate values; second, determine whether the number of abnormalities is less than m in the data of n consecutive time steps and compare them with the corresponding abnormality thresholds. If so, use Newton polynomial interpolation to interpolate; otherwise, directly delete other feature parameter data in the same time step containing abnormal data parameters; Newton interpolation calculation formula is as follows: N n (x)=f(x0)+f[x0,x1](x-x0)+f[x0,x1,x2](x-x0)(x-x1)+..+f[x0,x1,...,x n ](x-x0)(x-x1)…(xx n-1 ) (1) In formulas (1) to (2), (x i ,f(x i )) represents the normal data of samples in window n, f[x0, x1..., x n ] represents the n-order difference quotient. For the difference quotient f[x0, x1, ..., x n ], the order of the serial number subscripts does not affect the value of the difference quotient, x i The time coordinate of the abnormal data, N n (x) represents the interpolation value calculated by Newton interpolation. In actual interpolation, only the result N of the interpolation after the last sampling needs to be calculated. n-1 (x) Save and directly calculate N next time n-1 (x) and the new term to save computation and storage space; Step S2.2: Z-Score data normalization processing; data normalization can accelerate the speed of gradient descent optimal solution during model training and improve model accuracy. The data is uniformly mapped in the interval [0, 1]. The calculation formula is as follows: In formula (3), y is the data value, that is, the initial value before data normalization. is the mean value of the data set, σ is the standard value, and y′ is the corresponding value of the normalized data.

5. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 4, characterized in that: In step S3, the parameters strongly correlated with the heat shrink tube winding motor are screened out through Pearson correlation coefficient calculation and analysis, namely, winding motor speed, winding motor active power, winding motor voltage, winding machine tension, ambient temperature and winding motor temperature, a total of 6 groups of parameters, which are used as sample data for subsequent model training and testing. The screened sample data are integrated into the following matrix expression: In formula (10), Z∈R 6×s , represents the sample data matrix after the integration of 6 sets of historical data, where n is the winding motor speed set, P is the winding motor active power set, U is the winding motor voltage set, F is the winding motor tension set, K is the ambient temperature set, T is the winding motor temperature set, and s is the sth time step of the data set.

6. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 5, characterized in that: In the step S4, the preprocessed Z matrix sample data set is randomly divided into a training set and a test set according to conditions, wherein the training set accounts for 80% and is used to train the improved CNN-LSTM model; the test set accounts for 20% and is used to evaluate the improved CNN-LSTM model; In step S5, the network model structure includes two one-dimensional convolutional layers, one fully connected layer and two LSTM layers. After the convolutional layer, the BN layer and the activation function Leaky-ReLU are used to realize data normalization and nonlinear operation, thereby accelerating the model convergence speed and solving the gradient vanishing problem, and extracting the deep-level features of the factors affecting the temperature of the winding motor; The extracted deep features are input into LSTM to better predict the feature information.

7. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 1, characterized in that: In step S5, the parameters of the convolution layer are the weights of the convolution kernel and the offsets of each channel. The mathematical model of the convolution operation of the convolutional neural network CNN is as follows: In formula (11), represents the j-th output feature map of the m-th layer, M m-1 represents the input feature set, m represents the mth layer of CNN, Represents the convolution kernel used by the m-th layer convolution operation, represents the bias of the j-th feature map, * represents the convolution operation, and f(·) represents the activation function; The pooling layer performs a pooling operation on the features extracted by the convolution layer to achieve feature dimensionality reduction and reduce network parameters, and to a certain extent prevent the model from overfitting. The maximum pooling is used to seek the maximum value of the pooling area to obtain local features. The obtained feature map is more sensitive to texture features. For the output X of the kth filter in the i-th dimension of the convolution layer, i , the maximum pooling operation is expressed as follows: In formula (12), p i (j) is the jth output of the pooling layer, and w is the width of the pooling kernel; As a variant of the recurrent neural network RNN, LSTM introduces a gating unit to alleviate the gradient vanishing and gradient exploding problems in RNN training and improve prediction accuracy. At time t, the current memory unit C t It is the core part of LSTM neurons. The network's gate structure consists of a forget gate, an input gate, and an output gate, which determines whether information reaches C. t In one time step, the LSTM neuron calculates C through a series of calculations. t And the values ​​of other state quantities in the unit, the specific calculation formula is as follows: f t =σ(W f [h t-1 ,x t ]+b f ) (13) i t =σ(W i [h t-1 ,x t ]+b i ) (14) the t =σ(W o [h t-1 ,x t ]+b o ) (15) h t =o t ⊙tanh(C t ) (18), in formulas (13) to (18), f t 、i t , o t are the states of the input gate, forget gate, and output gate at time t, respectively. t-1 is the memory unit of LSTM at time t-1; and C t are the candidate memory unit of LSTM at the current moment t and the memory unit at the current moment, h t-1 is the output value of LSTM at time t-1, x t is the input value of LSTM at time t; W i , W f , W o , W C and b i 、b f 、b o 、b C are the weights and biases of the input gate, forget gate, output gate, and candidate memory, σ(·) is the sigmoid activation function, tanh(·) is the hyperbolic tangent function; ⊙ is the bitwise multiplication of each element.

8. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 7, characterized in that: The step S6 specifically includes the following steps: Step S6.1: Initialize IPSO hyperparameters and determine the initial values ​​of IPSO parameters; Step S6.2: Initialize particle positions and velocities, and randomly generate populations: x i =(n1,m1,m2,m3,α,p,batch size ), where n1 is the number of CNN convolution kernels, m1 and m2 are the number of neurons in the two LSTM hidden layers, m3 is the number of neurons in the fully connected layer, α is the initialization learning rate, p is the neuron inactivation rate, and batch size is the batch size (Batch Size), and the value ranges of the above parameters are as follows: n1∈[2,64],m1∈[1,30],m2∈[1,30],m3∈[1,20],α∈[0.001,0.005],p∈[0.01,0.90],batch size ∈[1,50]。 9. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 8, characterized in that: In step S7, the Circle mapping method is introduced to initialize the particle swarm to obtain a more uniform and diverse initial population to improve the convergence speed and accuracy of the algorithm. The expression of the chaotic sequence generated by the Circle mapping is as follows: In formula (21), X i represents the number of the i-th chaotic sequence, X i+1 represents the number of the i+1th chaotic sequence, e=0.5, f=0.2, mod represents the modulo operator, and the population initialization operation using Circle mapping includes the initialization of velocity and the initialization of position. The specific initial operations are as follows: v i，j =v lower_b +(v up_b -v lower_b )*X i，i (22) x i，j =x lower_b +(x up_b -x lower_b ))*X' i，j (23), in formulas (22) to (23), v lower_b and v up_b are the upper and lower bounds of the particle velocity, x lower_b and x up_b are the upper and lower bounds of the particle position, respectively, i，j and X' i，j They are the chaotic sequence values ​​generated by the Circle mapping on the corresponding particle dimensions.

10. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 9, characterized in that: The step S8 specifically includes the following steps: Step S8.1: Introduce the dynamic nonlinear change inertia weight, and select the dynamic nonlinear change inertia weight ω, as follows: In formula (24), ω(k) is the inertia weight of the kth iteration, ω max and ω min The maximum and minimum inertia weights are 0.9 and 0.4, respectively. max is the maximum number of iterations, k is the number of iterations; Step S8.2: Introduce asymmetric learning factors, including individual learning factor c1 and social learning factor c2. c1 and c2 are adaptively adjusted according to the different search periods. The specific expressions are as follows: In formulas (25) and (26), c1(k) and c2(k) are the individual learning factor and social learning factor of the kth iteration, respectively. 1_e 、c 1_s The initial value and final value of the individual learning factor are 2.5 and 0.5, respectively. 2_e 、c 2_s The initial value of social learning factor is 1 and the final value is 2.25, t max is the maximum number of iterations, and k is the number of iterations.

11. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 10, characterized in that: In step S11, the optimized initial particles are used as the parameters of the CNN-LSTM model, and the mean square error MSE is used as the fitness of the IPSO algorithm, as follows: In formula (27): is the predicted value, y i is the actual value, and N is the number of samples.

12. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 11, characterized in that: In step S17, the mean absolute error (MAE) threshold (MAE) is introduced. th_max , mean absolute percentage error MAPE threshold MAPE th_max , RMSE threshold RMSE th_max and the coefficient of determination R 2 Threshold R 2 th_min Evaluation indicators comprehensively evaluate the prediction effect of IPSO-CNN-LSTM model.

13. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 12, characterized in that: The step S18 specifically includes the following steps: Step S18.1: Calculate the evaluation index, calculate and evaluate various indicators of the IPSO-CNN-LSTM model, as follows: In formulas (28) to (31), is the predicted value, y i is the actual value, is the mean, N is the number of samples; Step S18.2: Threshold range judgment, compare the various evaluation indicators of the IPSO-CNN-LSTM model in turn to see if they are within the specified threshold range, as follows: If the requirements of the above inequality group (32) are met, it means that the prediction accuracy of the IPSO-CNN-LSTM model is high. At this time, the MAE at the end of the model training is recorded. final As the absolute upper limit of the predicted residual alarm threshold, continue to execute downward; otherwise, repeat step S14.

14. The method for predicting and controlling the temperature of a heat shrink tube winding motor according to claim 13, characterized in that: In step S19, the specific calculation formula is as follows: δ = MAE final ×σ test +μ test (33), where MAE final is the final mean absolute error MAE of the optimal IPSO-CNN-LSTM prediction model, σ test is the standard deviation of the test set, μ test is the mean of the test set.

Images