A data-enhanced winding machine spindle motion precision prediction method

Abstract

The present application relates to a kind of data enhanced winding machine spindle motion precision prediction method, first, spindle assembly process and operating parameter are collected, and rotor dynamics model is established to generate simulation data, then the sample data of different assembly process is expanded by designing conditional generative adversarial network, finally according to sample data belonging to time series data, long short-term memory network, attention mechanism network and BPNN neural network are constructed to predict the motion precision of winding machine spindle.The method considers the influence of the final assembly accuracy of spindle on the motion accuracy of winding process spindle, improves the accuracy of simulation data and the convergence speed of data expansion model;For the time series characteristics of characteristic parameters in the winding process of spindle, it is difficult to extract the features of process data before and after winding process, and the prediction model of LSTM-Attention is proposed, the feature extraction of time series parameters in the winding process of spindle is solved, the feature correlation between before and after time is mined, and the accurate prediction of the motion precision of spindle is realized.

Description

TECHNICAL FIELD

[0001] The application relates to a processing technology, in particular to a data-enhanced winding machine spindle motion precision prediction method. BACKGROUND

[0002] After years of development, China has become the world's largest chemical fiber producer and is gradually moving towards a chemical fiber power. The production process of chemical fiber fabric includes polymerization, winding, post-integration, chemical fiber fabric, etc. The filament winding stage plays a key role. In this production process, the high-speed winding machine plays the role of primary fiber stretching and winding. The high-speed winding spindle is the most core component of the winding machine. It is a thin-walled long shaft cantilever structure. The running winding stage parameters have time-varying characteristics, i.e. time-varying speed, time-varying mass and time-varying stiffness.

[0003] The single cycle working process of the winding machine spindle mainly includes rapid start, highest speed maintenance, filament winding, speed reduction stop and roll-off stage. The filament winding stage is the most significant stage affecting the dynamic characteristics of the whole system. The spindle speed can reach 14000r / min during the winding operation stage. The spindle will vibrate strongly due to high-speed operation. The spindle end vibration displacement can represent the motion precision, i.e. the spindle motion precision index can be regarded as the vibration displacement of the X and Y axis directions. If the motion precision is poor, the stability of the winding machine spindle system and the quality of the filament cake will be affected. The spindle motion precision is mainly affected by the assembly manufacturing precision and the time-varying parameters during operation. If the motion precision of the winding process spindle can be accurately predicted, it is of great significance to the winding machine performance maintenance and filament production. The winding machine spindle motion precision prediction has small sample and parameter time series data. At present, the winding machine spindle motion precision prediction mainly includes kinematic analysis, dynamic simulation to predict the precision, and manual dynamic balance adjustment based on experience. The former is difficult to accurately simulate the real working condition. Only by simplifying the model and replacing the parameters can the static simulation be carried out, which has high time cost. The latter requires difficult measurement during dynamic balance test. After several times of adjustment and debugging, the rated precision can be achieved. The motion precision is difficult to accurately predict, which makes it difficult to maintain the spindle motion precision and reduces the quality of the filament cake. SUMMARY

[0004] Aiming at the winding machine spindle motion precision prediction problem, a data-enhanced winding machine spindle motion precision prediction method is proposed, which is suitable for the winding process motion precision prediction of the winding machine single room and small batch, and is of great significance to the winding machine performance maintenance and filament production.

[0005] The technical scheme of the application is as follows: a data-enhanced winding machine spindle motion precision prediction method, specifically comprising the following steps:

[0006] 1) Establishing the spindle dynamics model: solving the mass eccentricity of assembly error in the spindle assembly process, and introducing the unbalanced force caused by it into the rotor dynamics equation; defining the initial data, time variable and time step, solving the intermediate parameters of the spindle dynamics model and the spindle vibration response at each time point, forming a one-to-one correspondence to form simulation data for sample expansion;

[0007] 2) Collecting measured data: collecting the final assembly accuracy parameters of the spindle under different assembly processes, including positioning error and orientation error, and spindle winding operation parameters, including spindle speed, winding package mass, spindle deformation and spindle vibration displacement, forming a one-to-one correspondence to form measured data samples for prediction and verification;

[0008] 3) Generating expanded data: based on the conditional generative adversarial network, a data expansion model is established, the simulation data of step 1) and the vibration influencing factors corresponding thereto are input into the generator of the adversarial network to generate expanded data, and then the expanded data, the measured data of step 2) and the vibration influencing factors corresponding thereto are input into the discriminator of the adversarial network to distinguish true and false data; according to the adversarial training principle of the conditional generative adversarial network, the generator and the discriminator constantly update the network parameters, and finally the generator generates a sufficient number of expanded data closer to the real data;

[0009] 4) Establishing a prediction model: a spindle motion precision prediction model combining long short-term memory network, attention mechanism and BPNN neural network is established, wherein the long short-term memory network extracts time sequence feature information, the attention mechanism provides the importance of different features, and the BPNN neural network fits the complex nonlinear relationship between variables and output, the connection mode of the three is horizontal connection, and the motion precision of the winding machine spindle is predicted;

[0010] 5) Training the prediction model: the expanded data samples generated by the generator are standardized, and the training set and the test set are divided in the ratio of 8:2, the sequence data of the training set is input into the long short-term memory network for training, and when the loss function in the BPNN neural network stops optimizing or reaches the maximum training times, the training is ended, the final winding machine spindle motion precision prediction model is obtained, and the winding machine spindle motion precision prediction model is verified by the test set.

[0011] Further, the spool shaft assembly precision in step 1) includes the coaxiality of the spool shaft fixed sleeve and the elongated shaft, and the parallelism of the spool shaft fixed sleeve and the base, which are used to characterize the final precision of the spool shaft assembly process; the initial data includes the spool shaft assembly precision, the initial vibration displacement, the speed, the acceleration, the running time in the winding stage, and the angular acceleration of the spool shaft, which not only considers the error transmission of the spool shaft assembly parameters, but also considers the influence of the spool shaft running parameters on the vibration displacement; the intermediate parameters of the spool shaft dynamics model include the mass matrix, the stiffness matrix, and the damping matrix of the spool shaft system, which are used to finally calculate the spool shaft vibration response at each time point, i.e., to obtain the winding process simulation data of the winding machine spool shaft under different assembly processes.

[0012] Further, in step 2), the spool shaft speed refers to the angular velocity of the rotating shaft on the spool shaft; the winding package mass refers to the mass of the chemical fiber filament wound into a package while the spool shaft is running; and the spool shaft deformation refers to the deformation of the spool shaft due to the change in mass during the winding process.

[0013] Further, in step 3), the method for generating the extended data considers the influence of the time sequence characteristics of the spool shaft running parameters, and adds a long short-term memory network to the generator and the discriminator based on the conditional generative adversarial network, which is used to extract the time sequence characteristic parameters; the labels input to the generator and the discriminator refer to the coaxiality of the spool shaft fixed sleeve and the rotating shaft, the parallelism of the spool shaft fixed sleeve and the base, the spool shaft speed, the winding package mass, and the spool shaft deformation.

[0014] Further, in step 4), a spool shaft motion precision prediction model is established by combining a long short-term memory network, an attention mechanism, and a BPNN neural network, wherein the long short-term memory network is composed of LSTM units for transmitting cyclic information, and is subsequently connected horizontally to the attention mechanism, both of which serve the purpose of feature extraction, while the output layer is composed of a BPNN network, which is connected horizontally to the former, and finally realizes the prediction of the spool shaft motion precision;

[0015] Specific implementation method:

[0016] The sequence length of the extended data generated by the conditional generative adversarial network is set as L, the sliding window length is given as S, the extended data is cut into a series of continuous subsequences containing S data by the sliding window method, the time step of the long short-term memory network is set as T, and the batch size is set as A, then the input of the long short-term memory network is a three-dimensional tensor of (A, T, S), the long short-term memory network is trained to process the complex correlation between the data and extract the time sequence characteristic information of the running parameters;

[0017] The attention mechanism with T shared weights and biases is added after the correlation between the T time steps output by the long short-term memory network, which further explores the internal correlation between the key time sequence characteristics at different times and the motion precision, and the attention mechanism weights are summed and weighted after training.

[0018] The BPNN neural network has at least a three-layer network structure, which is divided into an input layer, a hidden layer and an output layer, is responsible for mapping the parameter time sequence characteristic information extracted by the former two from a high-dimensional space to a low-dimensional space, estimating the error of the directly preceding layer of the output layer by using the error of the output layer, and then estimating the error of the more preceding layer by using the error, and thus the error estimation of all other layers is obtained, forming a process of transmitting the error shown at the output end to the input end of the network in the direction opposite to the transmission of the input signal.

[0019] The data enhancement winding machine spindle motion precision prediction method has the advantages that, compared with the traditional deep learning prediction method, the method introduces rotor dynamics to solve the spindle vibration response on the basis of the traditional deep learning prediction method, the method considers the influence of the final assembly precision of the spindle on the motion precision of the spindle in the winding process, improves the accuracy of the simulation data and the convergence speed of the data expansion model, and solves the problem that it is difficult to extract the features before and after the winding process data due to the time sequence characteristics of the characteristic parameters of the spindle winding process, mines the feature correlation between the before and after time, and based on the data enhancement method, the motion precision of the spindle can be accurately predicted, and the method has remarkable engineering practical value for the research on the maintainability of the motion precision of the winding machine spindle. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The figure is a whole idea diagram of the data enhancement winding machine spindle motion precision prediction method;

[0021] Figure 2 The figure is a winding machine spindle motion precision index diagram;

[0022] Figure 3 The figure is a winding machine spindle winding process vibration response solving flow diagram;

[0023] Figure 4a The figure is a winding machine spindle structure diagram Figure 1 ;

[0024] Figure 4b The figure is a winding machine spindle structure diagram Figure 2 ;

[0025] Figure 5 The figure is a quality eccentricity received unbalanced force diagram;

[0026] Figure 6 The figure is a data enhancement principle diagram;

[0027] Figure 7A winding machine spindle motion precision prediction model schematic diagram for the present application;

[0028] Figure 8 An experimental result graph of the data enhancement method under different assembly processes of the winding machine spindle for the present application;

[0029] Figure 9 A comparison effect diagram of the winding machine spindle motion precision prediction model for the present application. DETAILED DESCRIPTION

[0030] The present application will be described in detail below in conjunction with the drawings and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present application, and gives a detailed implementation manner and specific operation process, but the protection scope of the present application is not limited to the following embodiments.

[0031] As Figure 1 indicated in the data enhancement winding machine spindle motion precision prediction method overall idea graph, specifically including the following steps:

[0032] Establishing a spindle dynamics model: solving the mass eccentricity of the assembly error in the spindle assembly process, and introducing the unbalanced force caused by it into the rotor dynamics equation; defining initial data, time variable and time step, solving the intermediate parameters of the spindle dynamics model and the spindle vibration response at each time point, forming a one-to-one correspondence relationship to form simulation data for sample expansion;

[0033] Collecting measured data: collecting the final assembly precision parameters of the spindle under different assembly processes, such as positioning error and orientation error, and spindle winding operation parameters, such as spindle speed, winding package mass, spindle deformation, and Figure 2 indicated in the spindle X-axis and Y-axis direction vibration displacement, forming a one-to-one correspondence relationship to form measured data samples for prediction verification;

[0034] Generating expanded data: based on the conditional generative adversarial network, establishing a data expansion model, inputting the simulation data solved by the spindle dynamics model and the vibration influencing factors corresponding thereto as labels into the generator in the adversarial network to generate expanded data, and then inputting the expanded data, measured data and vibration influencing factors corresponding thereto as labels into the discriminator in the adversarial network to distinguish true and false data; according to the adversarial training principle of the conditional generative adversarial network, the generator and the discriminator constantly update the network parameters in opposition, and finally the generator generates a sufficient number of expanded data closer to the real data;

[0035] Establish a prediction model: Establish a spindle motion accuracy prediction model that combines a long short-term memory network, an attention mechanism, and a BPNN neural network. The long short-term memory network extracts time series feature information, the attention mechanism provides the importance of different features, and the BPNN neural network fits the complex nonlinear relationship between variables and output. The three are connected laterally to predict the spindle motion accuracy of the winding machine.

[0036] Training the prediction model: The augmented data samples generated by the generator are standardized and divided into training and test sets in an 8:2 ratio. The training set sequence data is input into the Long Short-Term Memory network for training. The training ends when the loss function in the BPNN neural network stops optimizing or reaches the maximum number of training iterations, thus obtaining the motion accuracy prediction model.

[0037] Experimental comparison: A portion of the measured data was divided into a validation set and input into the pre-trained prediction model. The prediction accuracy of the prediction model was verified by comparing the difference between the prediction results and the measured results.

[0038] like Figure 3 The diagram shows the flow chart for solving the vibration response of the winding machine spindle during the winding process. It calculates and solves for the intermediate parameters of the spindle dynamics model and the vibration response of the spindle at each time point. Figure 4a , 4b The schematic diagram of the winding machine spindle structure shown below illustrates that the spindle assembly consists of components such as bearings, bearing bases, a fixed-length sleeve, a short shaft, a slender shaft, and a rotating long sleeve. The fixed-length sleeve in the sleeve assembly directly participates in the winding of the yarn cake during the winding process. Therefore, the assembly accuracy of the fixed-length sleeve can be considered as the final assembly accuracy of the spindle, especially the coaxiality between the fixed-length sleeve and the slender shaft, and the parallelism between the fixed-length sleeve and the base. These can be measured using an error measuring instrument after the spindle assembly is completed. The following section, with reference to the attached diagram, further elaborates on the data-enhanced method for predicting the motion accuracy of the winding machine spindle:

[0039] Step 1: To illustrate the influence of spindle assembly accuracy on the motion accuracy of the winding process, such as... Figure 5 As shown, the mass eccentricity e caused by the assembly accuracy of the spindle shaft is solved based on the inertial coordinate system X1O1Y1 and the rotating coordinate system X2O1Y2. The centroids of the fixed-length sleeve and the bearing base are denoted as O2 and O3, respectively. Due to the mass eccentricity, an unbalanced force will be generated during the movement of the spindle shaft. The matrix F of the unbalanced force caused by the mass imbalance is shown. u Its expression is as follows:

[0040]

[0041] In the formula, m is the mass of the eccentric point, ω is the rotational speed of the spindle shaft, and e x2 e y2are the coordinates of the eccentric point in the system coordinate, t is the running time.

[0042] According to the running characteristics of the winding machine spindle, the dynamic equation of the spindle rotor system is established, and its expression is shown as follows:

[0043]

[0044] In the formula, M represents the mass matrix of the spindle system, C represents the system damping matrix, K represents the stiffness matrix, F g represents the generalized force matrix of the spindle system, and S(t) represents the displacement vector matrix.

[0045] If the winding process of the winding machine spindle is regarded as a function of time or the rotating speed of the spindle, the variable mass parameter can be introduced into the dynamic equation of the spindle system, and the dynamic expression of the variable package mass unit can be represented as:

[0046] The solving formula of each parameter is shown as follows:

[0047]

[0048]

[0049]

[0050]

[0051] In the formula, l, r, r1, r(t), and p respectively represent the width of the fixed-length sleeve on the spindle, the radius of the cheese, the radius of the spindle when the cheese is empty, the real-time winding radius of the variable package, the density of the spindle, [N] is the displacement function of the centroid of the fixed-length sleeve on the spindle relative to time, [ω ji ] is the absolute angular velocity of the rotating coordinate system relative to the inertial coordinate system, and [ε ji ] is the projection component of the angular acceleration on the coordinate axis of the rotating coordinate system.

[0052] By using the Taylor series expansion, the recursive relationship of the displacement S, velocity acceleration of the system at time t and the motion state quantity at time t+Δt is established. Thus, the motion state vector of the system at each time can be obtained by the recursive formula given the initial state of the system (at time t=0). The Newmark direct integration method is adopted to program the spindle system by using Matlab, and the vibration response of the winding process of the spindle system is calculated.

[0053] Step 2: Record the coaxiality of the winding machine spindle fixed sleeve and the rotating shaft, the parallelism of the spindle fixed sleeve and the base under different assembly processes; collect the spindle speed, the wound bobbin mass and the spindle deformation during the winding operation of the winding machine spindle.

[0054] Step 3: As shown in Figure 6 , according to the spindle vibration response simulation data and the measured data obtained in steps 1 and 2, the coaxiality of the spindle fixed sleeve and the rotating shaft, the parallelism of the spindle fixed sleeve and the base, the spindle speed during the winding operation of the spindle, the wound bobbin mass and the spindle deformation are taken as vibration influencing factor labels K and the spindle vibration displacement data X as input to the generator to generate augmented data G, and the augmented data G and the measured data Y and the vibration influencing factor labels K are taken as input to the discriminator, and the discriminator outputs the probability values of "false" and "true". Through the continuous training game of the model, the error of each training is calculated, and the generator and discriminator parameter updating is generated through the objective function, so as to realize that the generated augmented data Z is very similar to the measured data. The objective function expression of the conditional generative adversarial network is as follows:

[0055]

[0056]

[0057]

[0058] V G (D(Y), G(F)) is the objective function; E is the target expectation; D(Y) is the probability that the discriminator judges whether the real data is real, and the output value is between 0 and 1. The stronger the ability of the discriminator, the larger D(Y) is, that is, the larger the objective function value is; D(G(F)) is the probability that the discriminator judges whether the generated data is real. The generator G(F) continuously approaches the measured sample data distribution, and D(G(F)) becomes larger, and then the objective function becomes smaller. Such a maximum and minimum value game alternates to improve the ability of the generator and the discriminator. Finally, the augmented data generated by the generator is infinitely close to the measured data, and the model balance is achieved.

[0059] The goodness of the augmented data Z relative to the measured data is usually judged by correlation, mean absolute error and mean square error. The correlation R 2 , the mean absolute error MAE and the mean square error RMSE are expressed as follows. The augmented data is and the measured data Y i :

[0060]

[0061]

[0062]

[0063] Step 4: As Figure 7 As shown, a prediction model combining a long short-term memory (LSTM) network, an attention mechanism, and a back propagation neural network (BPNN) is established. The LSTM network extracts time-series features, the attention mechanism provides the importance of different features, and the BPNN fits the complex nonlinear relationship between variables and outputs. The three are connected laterally. Regarding dimensions, due to the significant differences between the assembly and operation data of the winding machine spindle, the input features need to be preprocessed to eliminate the influence of dimensions. The input features are normalized by minima to eliminate the influence of dimensions. The normalization formula is shown below:

[0064]

[0065] Let the sequence length of the expanded data Z obtained in step 3 be L, and the sliding window length be S. Using a sliding window, a series of continuous subsequences containing S data points are extracted from the expanded data Z. Let the time step of the Long Short-Term Memory (LSTM) network be T, and the batch size be A. Then, the input to the LSTM network is a three-dimensional tensor of (A, T, S). The LSTM network is trained to handle complex relationships between data and extract time-series feature information of the operating parameters. The hyperparameters of the LSTM network consist of the learning rate, batch size, time step, and number of hidden neurons, with values ​​of 0.001, 30, 12, and 128, respectively. The data at time t during the LSTM winding process is represented by x. t Therefore, the update mechanism for LSTM to extract feature information can be represented by the following formula:

[0066] f t =σ(W f ·[x t h t-1 ]+b f )

[0067] i t =σ(W i ·[x t h t-1 ]+b i )

[0068] O t =σ(W o ·[x t h t-1 ]+bo )

[0069]

[0070]

[0071] h t =O t *tanh(C t )

[0072] In the formula, i t , o t , f t respectively represent the calculation values of the input, output and forget gate at time t during the winding process of the spindle, C t , C t-1 respectively represent the candidate value of the memory cell, the updated value and the output value of the neuron cell at time t-1, W f , W i , W o , W c respectively represent the forget gate, the input gate, the output gate and the updated weight, b f , b i , b o , b c respectively represent the forget gate, the input gate, the output gate and the updated bias, and σ represents the sigmoid function; h t and h t-1 respectively represent the final output value of the memory cell at the current time and the final output value of the memory cell at the previous time.

[0073] The attention mechanism with T shared weights and biases is added after the association timing characteristics between T time steps output by the long short-term memory network, so as to further mine the internal correlation of the key timing characteristics at different times to the motion precision. After training, the weights of the attention mechanism are summed and weighted. The attention size of the final output of the network is only a one-dimensional value, so an Attention network containing a number of hidden layers is constructed, and the number of neurons in each layer is set to decrease layer by layer. The number of hidden layers of the Attention network is set to 1, and the number of neurons in each layer is set to 16.

[0074] The BPNN neural network has at least a three-layer network structure, which is divided into an input layer, a hidden layer and an output layer, is responsible for mapping the parameter time sequence characteristic information extracted by the former two from a high-dimensional space to a low-dimensional space, estimating the error of the directly preceding layer of the output layer by using the error of the output layer, and then estimating the error of the more preceding layer by using the error, and thus the error estimation of all other layers is obtained, forming a process of transmitting the error shown at the output end to the input end of the network in the direction opposite to the transmission of the input signal; in terms of the loss function setting of the BPNN network, the residual square sum is performed on the model predicted value and the motion precision actually measured value Y i to obtain a loss function Loss, and the specific formula is as follows:

[0075]

[0076] Finally, the model is converged through continuous cyclic iteration of the network, and a stable model structure with small bias and variance values is obtained.

[0077] Step 5: The augmented data samples generated by the generator are standardized, and the training set and the test set are divided in a ratio of 8:2; the long short-term memory network is trained by inputting the training set sequence data; the training is ended when the loss function in the BPNN neural network stops optimizing or reaches the maximum number of training times, and a motion precision prediction model is obtained;

[0078] Step 6: Part of the actually measured data is divided into a validation set and input into the above trained prediction model; the prediction accuracy of the prediction model is verified by comparing the difference between the prediction result and the actually measured result and comparing with other time sequence prediction algorithms.

[0079] The experimental results of the winding parameter sample enhancement method of the spool shaft of the winding machine under different assembly processes proposed in the application are shown in Figure 8 , and the comparison of the actually measured vibration displacement and the augmented vibration displacement of the spool winding parameters under the coaxiality of 0.022 between the fixed sleeve of the spool shaft of the winding machine and the shaft and the parallelism of 0.189 between the fixed sleeve of the spool shaft of the winding machine and the base is shown in the figure. It is found that, due to the fact that the sample enhancement model takes the simulation data as input, the model fits the actually measured data better, the error value of the augmented data and the actually measured data is smaller, the correlation coefficient R 2 is 0.96, the average absolute error MAE is 0.126, and the root mean square error RMSE is 0.582. In order to further compare and analyze the prediction effects of the methods, the average absolute error MAE and the root mean square error RMSE are used to quantitatively evaluate the effects of the above methods in the motion precision prediction scene of the spool shaft of the winding machine, and the experimental results of the motion precision prediction of the spool shaft of the winding machine are shown in Figure 9Compared with other three kinds of algorithms, the algorithm of the application is improved according to the winding running characteristics of the winding machine spindle, and the motion precision prediction accuracy is significantly improved.

[0080] The above-mentioned embodiments only express several embodiments of the application, the description is more specific and detailed, but it cannot be understood as the limitation of the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the application, a number of modifications and improvements can be made, which belong to the protection scope of the application. Therefore, the protection scope of the patent of the application should be subject to the appended claims.

Claims

1. A data-augmented spooler motion accuracy prediction method, characterized by, Specifically comprising the following steps: 1) Establishing a spindle dynamics model: solving the mass eccentricity of assembly error in the spindle assembly process, and introducing the unbalanced force caused by it into the rotor dynamics equation; defining initial data, time variable and time step, solving the intermediate parameters of the spindle dynamics model and the spindle vibration response at each time point, forming a one-to-one correspondence to form simulation data for sample expansion; The initial data includes: the final assembly accuracy of the spindle, the initial vibration displacement, speed, acceleration, running time and angular acceleration of the spindle in the winding stage, which considers the error transmission of the spindle assembly parameters and the influence of the spindle running parameters on the vibration displacement; wherein the fixed sleeve in the sleeve joint of the winding machine spindle assembly directly participates in the winding process of the filament cake, so the assembly accuracy of the fixed sleeve can be regarded as the final assembly accuracy of the spindle, and the final assembly accuracy of the spindle includes the coaxiality of the spindle fixed sleeve and the slender shaft, and the parallelism of the spindle fixed sleeve and the base; The intermediate parameters of the spindle dynamics model include: the mass matrix, the stiffness matrix and the damping matrix of the spindle system, which are used to finally calculate the spindle vibration response at each time point, i.e. to obtain the winding process simulation data of the winding machine spindle under different assembly processes; 2) Collecting measured data: collecting the final assembly accuracy parameters of the spindle under different assembly processes, including positioning error and directional error, and the winding running parameters of the spindle, including spindle speed, winding filament cake mass, spindle deformation and spindle vibration displacement, forming a one-to-one correspondence to form measured data samples for prediction and verification; 3) Generating expanded data: based on the conditional generative adversarial network, a data expansion model is established, the simulation data of step 1) and the vibration influencing factors corresponding thereto are input into the generator of the adversarial network to generate expanded data, and then the expanded data, the measured data of step 2) and the vibration influencing factors corresponding thereto are input into the discriminator of the adversarial network for true and false data discrimination; according to the adversarial training principle of the conditional generative adversarial network, the generator and the discriminator constantly update the network parameters, and finally the generator generates a sufficient number of expanded data closer to the real data; The method for generating expanded data: considering the influence of the time sequence characteristics of the spindle running parameters, a long short-term memory network is added to the generator and discriminator based on the conditional generative adversarial network, which is used to extract time sequence characteristic parameters; The labels input into the generator and discriminator indicate the coaxiality of the spindle fixed sleeve and the spindle shaft, the parallelism of the spindle fixed sleeve and the base, the spindle speed, the winding filament cake mass and the spindle deformation; 4) Establishing a prediction model: a spindle motion precision prediction model combining long short-term memory network, attention mechanism and BPNN neural network is established, wherein the long short-term memory network extracts time sequence feature information, the attention mechanism provides the importance of different features, and the BPNN neural network fits the complex nonlinear relationship between variables and output, the connection mode of the three is horizontal connection, and the winding machine spindle motion precision is predicted. 5) Training the prediction model: the augmented data samples generated by the generator are standardized, and the training set and the test set are divided in the ratio of 8:

2. The long short-term memory network is trained by inputting the training set sequence data. When the loss function in the BPNN neural network stops optimizing or reaches the maximum training times, the training is ended, and the final winding machine spindle motion precision prediction model is obtained. The winding machine spindle motion precision prediction model is verified by the test set.

2. The data-augmented coiler-spindle motion accuracy prediction method of claim 1, wherein, The spindle speed in step 2 refers to the angular velocity of the rotating shaft on the spindle; the winding bobbin mass refers to the mass of the chemical fiber filament wound into a bobbin during the operation of the spindle; and the spindle deformation refers to the deformation of the spindle caused by the mass change during the winding process.

3. The data enhanced winding machine spindle motion precision prediction method according to claim 2, characterized in that, 4) The spindle motion precision prediction model is established by combining the long short-term memory network, the attention mechanism and the BPNN neural network. The long short-term memory network is composed of LSTM units for transmitting cyclic information, and the subsequent lateral connection attention mechanism. Both of them play a role in feature extraction, and the output layer is composed of a BPNN network. The former is connected to the latter in a horizontal direction, and finally the spindle motion precision is predicted; The specific implementation method is as follows: the sequence length of the augmented data generated by the conditional generative adversarial network is set as L, the sliding window length is set as S, the augmented data is cut into a series of continuous sub-sequences containing S data by the sliding window method, the time step of the long short-term memory network is set as T, and the batch size is set as A. The input of the long short-term memory network is a three-dimensional tensor of (A, T, S). The long short-term memory network is trained to process the complex correlation between data and extract time series feature information of the operating parameters. The correlation between the T time steps output by the long short-term memory network is increased by T attention mechanisms with shared weights and biases, which further dig deeper into the internal correlation of key time series features at different times on the motion precision. After training, the attention mechanism weights are summed and weighted; the BPNN neural network has at least 3 network structures, which are divided into input layer, hidden layer and output layer, responsible for mapping the parameter time series feature information extracted by the former two from high-dimensional space to low-dimensional space, using the error of the output layer to estimate the error of the directly preceding layer of the output layer, and then using this error estimate to estimate the error of the more previous layer. In this way, the error estimates of all other layers are obtained, forming a process of transmitting the error shown at the output end to the input end of the network in the opposite direction of the input signal transmission.

Images