Data-model hybrid driven ball screw thermal error prediction method

Abstract

The invention relates to the technical field of machine tool manufacturing, and discloses a data-model hybrid driven ball screw thermal error prediction method, which comprises the following steps: establishing a ball screw thermal error physical model, and obtaining sample data by adopting a Latin hypercube design LHD method; a ball screw pair thermal error test platform is built, an optimal temperature measurement point is analyzed through ANSYS Workbench, and state information such as temperature, current and power is collected; processing the data through grey correlation degree analysis, abnormal value processing and normalization operation; building a data-model hybrid-driven convolutional long short-term memory network model C-LSTM (long short-term memory); and inputting the currently collected related state information data into the trained model, and performing thermal error prediction. According to the invention, the interpretability of data is ensured, so that the accuracy of thermal error prediction of the ball screw is improved.

Description

Technical Field

[0001] The invention relates to the technical field of machine tool manufacturing, and in particular to a data-model hybrid driven ball screw thermal error prediction method. Background Art

[0002] CNC machine tool errors include geometric errors, errors caused by cutting forces, thermal deformation errors, tool wear errors, CNC monitoring system errors, etc. Among them, thermal errors account for a large proportion, about 40%, and even reach 40%-70% in precision machine tools. Thermal errors are highly nonlinear, have many influencing factors, and have a high degree of coupling.

[0003] In the field of CNC machine tools, the ball screw feed system is one of the core components for achieving precise machining. It is responsible for converting the rotary motion of the motor into precise linear motion of the tool or worktable, which directly determines the machining accuracy, surface quality and production efficiency of the CNC machine tool. However, under the long-term and high-load operation conditions of the CNC machine tool, the ball screw feed system will inevitably experience a decline in accuracy. This is mainly attributed to the continuous frictional heat generation, and the heat accumulation forms thermal expansion, which affects the accuracy of the ball screw feed system. Thermal errors will increase the pitch error of the screw and reduce the smoothness of movement, which will lead to excessive dimensional accuracy and increased surface roughness of the processed parts, seriously affecting product quality.

[0004] Among the existing thermal error prediction technologies, most are based on a single data-driven method with poor interpretability and lack of accurate modeling of the thermal characteristics of ball screws, ignoring some factors that may have intrinsic influence. Summary of the invention

[0005] In view of the shortcomings in the prior art, the purpose of the present invention is to provide a data-model hybrid driven ball screw thermal error prediction method, by establishing a ball screw thermal characteristic model, collecting temperature, voltage, vibration and other information of the CNC machine tool ball screw, coupling the thermal characteristic model with the state information, and constructing a data-model hybrid driven ball screw feed system thermal error prediction model, and finally realizing the prediction of the thermal error of the CNC machine tool ball screw feed system. Compared with a single data-driven method, the present invention can use the prior knowledge of the model to process and analyze the data more targetedly, reduce the influence of noise and outliers in the data on the prediction results, realize the real-time prediction of the thermal error of the CNC machine tool ball screw feed system, improve the interpretability and generalization of the thermal error prediction, and is of great significance to improving the performance and reliability of CNC machine tools.

[0006] In order to achieve the above-mentioned objects and advantages of the present invention, the present invention is implemented by the following technical solutions:

[0007] A data-model hybrid driven ball screw thermal error prediction method comprises the following steps:

[0008] Step 1, establish a physical model of thermal error of the ball screw of a CNC machine tool and obtain theoretical data;

[0009] Step 2, using ANSYS Workbench to analyze the steady-state thermal characteristics of the ball screw and determine the optimal layout position of the temperature sensor;

[0010] Step 3, build a CNC machine tool ball screw feed system test platform to obtain the temperature, power, speed, current and thermal error data of the ball screw feed system;

[0011] Step 4: preprocess the experimental data and theoretical data, including grey relational analysis, outlier processing and normalization, and use the processed data as a sample data set for data-model hybrid driven model training;

[0012] Step 5: Build a data-model hybrid driven model, namely, a convolutional long short-term memory network (C-LSTMN), and train the sample data set of the data-model hybrid driven model in step 4;

[0013] Step 6, based on the prediction model driven by the data-model hybrid in step 5, the thermal error of the ball screw feed system of the CNC machine tool is calculated in real time.

[0014] Furthermore, the step 1 establishes a physical model of thermal error of a ball screw of a CNC machine tool, and the process is as follows:

[0015] According to the theory of heat transfer, the heat transfer process of the linear feed axis mainly includes: heat conduction, heat convection and radiation heat transfer. According to Fourier's law of heat conduction, the expression of the heat flux density of the CNC machine tool ball screw parts in the heat conduction process is:

[0016]

[0017] Among them, q t is the heat flux density of heat conduction; λ is the thermal conductivity, T p is the temperature of the component, is the temperature gradient of the component. The negative sign indicates that the direction of heat transfer is opposite to the direction of the temperature gradient.

[0018] The heat flux density of convective heat transfer of CNC machine tool ball screw components can be calculated according to Newton's cooling formula, and its expression is:

[0019] q c =h c ·(T p -T f )

[0020] Among them, q c is the heat flux density of the convective heat transfer process, hc is the convective heat transfer coefficient, T f It is the temperature of the air in contact with the surface of the component.

[0021] The heat flux density radiated outward by the ball screw components of CNC machine tools can be calculated using the modified form of the Stefan-Boltzmann law, and its expression is:

[0022]

[0023] Among them, q ro is the heat flux density of the radiation heat transfer process, ε is the emissivity of the component, and its value is between 0 and 1, c r is the blackbody radiation coefficient;

[0024] Then, the temperature field T of the screw due to the bearing and screw nut is calculated by heat conduction, heat convection and radiation heat transfer formulas. st (x,t).

[0025] Finally, the thermal error of the screw at any time t is:

[0026]

[0027] Where c1 is the thermal expansion coefficient of the screw and c2 is the thermal drift correction factor.

[0028] Furthermore, the step 2 performs steady-state thermal characteristic analysis on the ball screw through ANSYS Workbench, and the steps are as follows:

[0029] First, the ball screw feed system model is simplified, exported into a .x_t format file, and input into Workbench; then, the ball screw model is meshed; then, parameters are set according to the material properties of each part; finally, the steady-state thermal characteristics of the ball screw are solved. Through finite element analysis, the important temperature measurement points in the ball screw are determined.

[0030] Furthermore, the step 3 builds a CNC machine tool ball screw feed system test platform, and the process is as follows:

[0031] By placing temperature sensors at the temperature measurement points in step 3 above, the temperature of key temperature points in the ball screw feed system and the room temperature are measured; the axial thermal error of the ball screw is measured by a laser interferometer; the power and speed of the ball screw feed system are measured by the module provided by the ball screw test bench control system. The data matrix of the ball screw can be expressed as X t =(T t 1 ,T t 2 ,T t 3 ,Tt 4 ,T t 5 ,E t ,P t ,I t ,n), where X t Represents the state data matrix of the ball screw at time t, T t i represents the temperature of the i-th temperature measurement point at time t, I t represents the motor current at time t, E t represents the thermal error measured by the laser interferometer at time t, P t It represents the power of the ball screw feed system at time t, and n represents the rotation speed of the ball screw.

[0032] Furthermore, the step 4 pre-processes the experimental data and theoretical data, including grey relational analysis, outlier processing and normalization, and uses the processed data as a sample data set for data-model hybrid driven model training. The process is as follows:

[0033] First, the collected experimental and theoretical data are processed for outliers, and the abnormal data are preliminarily screened out based on the physical model and prior data;

[0034] Then, through grey correlation analysis, the correlation degree of different factors affecting the thermal error of the ball screw is determined, and the influencing factors α of different factors are given. i ;

[0035] Finally, the maximum and minimum normalization is used to normalize the theoretical data of the ball screw thermal error. The processing method is as follows:

[0036]

[0037] Where E is the theoretical value obtained by the ball screw thermal error physical model, E min is the minimum theoretical value obtained by the ball screw thermal error physical model, E max It is the maximum theoretical value obtained by the physical model of ball screw thermal error.

[0038] Furthermore, the step 5 constructs a data-model hybrid driven model, namely a convolutional long short-term memory network (C-LSTMN) model, including a convolutional neural network (CNN) module and a long short-term memory network (LSTMN) module.

[0039] The convolutional neural network (CNN) layer is convolved with two convolution kernels of size K1=1×1 and K2=3×3. The data after grayscale correlation analysis, outlier processing and normalization in step 4 above is convolved through the input layer to extract the features between variables, and different features are learned through different convolution kernels. The feature vector contains the combined information of temperature, current and power, and retains the variable features with the strongest correlation with thermal error for subsequent model training;

[0040] The long short-term memory network (LSTMN) consists of an input layer, a hidden layer and an output layer;

[0041] The input layer is the output of the convolutional neural network;

[0042] The basic unit of the hidden layer is a storage unit, which is composed of a self-connected storage unit with a storage time state. The storage unit can transfer information from the previous output to the current output. By constantly memorizing and past input data, this layer learns the overall change characteristics of temperature, current, power and speed. Each storage unit consists of four parts: an input gate, a forget gate, an autoregressive connected neuron and an output gate. The output expression of LSTMN is:

[0043] Y=f(wx t +b)

[0044] Among them, x t is the input matrix containing temperature, power, current, and speed, Y represents the output of the fully connected layer, that is, the thermal error E of the ball screw, w represents the weight matrix of the fully connected layer, and b represents the bias vector of the fully connected layer.

[0045] The prediction performance of the algorithm is evaluated by the mean absolute error (MAE), and its mathematical expressions are:

[0046]

[0047] Among them, h(x r ) is the predicted value, y r is the true value, r is the number of samples, r=1,2,…,q,x r The input of the model is the mean absolute error. The iterative training of the network model is determined by setting a target mean absolute error until the accuracy requirement is met.

[0048] Furthermore, in step 6, the temperature, power, current and speed acquired in real time are input into the prediction model based on the prediction model driven by the data-model hybrid in step 5, so as to predict the thermal error of the ball screw at the next moment.

[0049] Design ideas of the present invention:

[0050] A physical model of ball screw thermal error is established, and the Latin hypercube design (LHD) method is used to obtain sample data. A ball screw pair thermal error test platform is built, and the optimal temperature measurement point is analyzed through ANSYS Workbench to collect state information such as temperature, current, and power. The data is processed through grey correlation analysis, outlier processing, and normalization operations. A data-model hybrid-driven convolutional long short-term memory network model C-LSTM is built. The currently collected relevant state information data is input into the trained model for thermal error prediction.

[0051] Beneficial effects of the present invention:

[0052] (1) By establishing a physical model to obtain theoretical data and combining it with the experimental data obtained from the actual test platform to hybrid drive the training model, the correlation between various factors can be comprehensively considered to ensure the interpretability of the data, thereby improving the accuracy of ball screw thermal error prediction.

[0053] (2) ANSYS Workbench was used to perform steady-state thermal characteristics analysis on the ball screw to determine the optimal layout of the temperature sensor, reducing unnecessary data and ensuring that the collected temperature data is more representative and effective.

[0054] (3) Through the C-LSTMN prediction model, the thermal error of the ball screw at the next time can be predicted, which provides accurate input for the subsequent thermal error compensation, thereby improving the accuracy of the ball screw feed system. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 The schematic diagram of the thermal error prediction process of the ball screw driven by data-model hybrid;

[0056] Figure 2 It is a schematic diagram of the three-dimensional model of the ball screw feed system;

[0057] Figure 3 This is the ANSYS Workbench steady-state thermal error cloud diagram;

[0058] Figure 4 It is a data acquisition test bench for ball screw feed system;

[0059] Figure 5 Schematic diagram of a convolutional long short-term memory neural network. DETAILED DESCRIPTION

[0060] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in combination with the drawings and the drawings in the specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] A data-model hybrid driven ball screw thermal error prediction method, the specific flow chart is as follows Figure 1 As shown, the following steps are included:

[0062] Step 1, establish a physical model of thermal error of the ball screw of a CNC machine tool and obtain theoretical data;

[0063] Schematic diagram of the physical model of the ball screw of a CNC machine tool Figure 2 As shown, it includes a motor 1, a coupling 2, a front bearing 3, a workbench 4, a screw nut 5, a screw 6, a rear bearing 7, and a bed 8. The process of step 1 is as follows:

[0064] According to the theory of heat transfer, the heat transfer process of the linear feed axis mainly includes: heat conduction, heat convection and radiation heat transfer. According to Fourier's law of heat conduction, the expression of the heat flux density of the CNC machine tool ball screw parts in the heat conduction process is:

[0065]

[0066] Among them, q t is the heat flux density of heat conduction; λ is the thermal conductivity, T p is the temperature of the component, is the temperature gradient of the component. The negative sign indicates that the direction of heat transfer is opposite to the direction of the temperature gradient.

[0067] The heat flux density of convective heat transfer of CNC machine tool ball screw components can be calculated according to Newton's cooling formula, and its expression is:

[0068] q c =h c ·(T p -T f )(2)

[0069] Among them, q c is the heat flux density of the convective heat transfer process, h c is the convective heat transfer coefficient, T f It is the temperature of the air in contact with the surface of the component.

[0070] The heat flux density radiated outward by the ball screw components of CNC machine tools can be calculated using the modified form of the Stefan-Boltzmann law, and its expression is:

[0071]

[0072] Among them, q ro is the heat flux density of the radiation heat transfer process, ε is the emissivity of the component, and its value is between 0 and 1, c r is the blackbody radiation coefficient.

[0073] By integrating the above formula, the heat of the screw is obtained and the temperature field is determined.

[0074] First, calculate the temperature field of the screw under the friction between the nut and the screw:

[0075] According to formula (1), the heat conduction heat of the ball screw in the time period (t-Δt, t) is:

[0076]

[0077] Where R is the equivalent diameter of the ball screw, is the temperature gradient value at position x at time t, is the temperature gradient value at the position x+dx at time t. The heat transfer of the combined heat transfer process of thermal radiation and thermal convection in the time period (t-Δt, t) during the operation of the ball screw is:

[0078]

[0079] Where k is the amplification factor of the ball screw convective heat transfer, T s (x,t) is the temperature at the x position of the ball screw at time t, T f (t) is the temperature of the air in contact with the ball screw surface at time t.

[0080] The friction heat generated by the ball screw micro-end dx and the nut at position x in the time period (t-Δt, t) is:

[0081]

[0082] Among them, Q is the heat generated by the nut rubbing against the ball screw micro-segment dx once, and N is the average number of frictions between the nut and the ball screw dx segment per unit time.

[0083] The heat increment of the ball screw micro segment dx during the time period (t-Δt, t) is:

[0084]

[0085] Where c is the specific heat capacity of the ball screw material and ρ is the ball screw material density.

[0086] When only the friction between the nut and the ball screw is considered, the temperature of the ball screw is T sf(x, t). The thermal balance formula of the ball screw micro-segment dx at time t is:

[0087] Q f (x,t)-Q t (x,t)+Q t (x+dx,t)-Q a (x, t) = ΔQ(x, t) (9)

[0088] Substituting equations (4) to (8) into equation (9), we can obtain:

[0089]

[0090] Among them, T f (t) can be approximately the same as the ambient temperature T near the ball screw b (t) instead, that is, T f (t) = T b (t). The temperature field of the screw under the friction between the nut and the screw is:

[0091] T sf (x,t)=T s (x,t)(11)

[0092] Secondly, calculate the temperature field of the screw under the action of single-point heat source at both ends of the bearing:

[0093] According to the principle of energy conservation, the energy equation of the dx segment is:

[0094] -Q t (x,t)+Q t (x+dx,t)-Q cr (x, t) = ΔQ(x, t) (12)

[0095] According to the temperature field distribution of the screw under the action of a single point heat source, the temperature gradient on both sides of the screw micro-segment dx has the following relationship:

[0096]

[0097] Substituting equations (4) to (6), (8), and (13) into equation (12), we obtain:

[0098]

[0099] Therefore, the temperature field of the screw under the action of single point heat source of bearing seat 1 and bearing seat 2 is:

[0100]

[0101] Where L is the available stroke length of the screw.

[0102] According to the temperature field superposition principle, the temperature field of the screw is:

[0103]

[0104] Thermal error of the screw at any time t:

[0105]

[0106] Where c1 is the thermal expansion coefficient of the screw and c2 is the thermal drift correction factor.

[0107] Latin hypercube design (LHD) is used to obtain sample data of the physical model. The Latin hypercube design first determines the number of variables and the value range in the physical model of the ball screw and evenly divides them into n parts. Then a value is randomly selected from the n intervals to form n groups of data. Finally, the final result is obtained according to the physical model of the ball screw. The final sample data is an n×(number of variables+1)-dimensional matrix.

[0108] Step 2, using ANSYS Workbench to analyze the steady-state thermal characteristics of the ball screw and determine the optimal layout position of the temperature sensor;

[0109] First, the ball screw feed system model is simplified, exported into a .x_t format file, and input into Workbench. Then, the ball screw model is meshed. Next, parameters are set according to the material properties of each part. Finally, the steady-state thermal characteristics of the ball screw are solved. The results are as follows: Figure 3 As shown in the figure. Through finite element analysis, the important temperature measurement points in the ball screw are determined.

[0110] Step 3, build a CNC machine tool ball screw feed system test platform to obtain the temperature, power, speed, current and thermal error data of the ball screw feed system;

[0111] Step 3: Schematic diagram of the experimental platform Figure 4 As shown, the process is as follows:

[0112] By arranging temperature sensors at the temperature measuring points in the above step 3, the temperatures of key temperature points in the ball screw feed system and the room temperature are measured;

[0113] The axial thermal error of the ball screw is measured by laser interferometer;

[0114] The power and speed of the ball screw feed system are measured using the built-in module of the ball screw test bench control system.

[0115] The data matrix of the ball screw can be expressed as X t =(T t 1 ,T t 2 ,T t3 ,T t 4 ,T t 5 ,E t ,P t ,I t ,n), where X t Represents the state data matrix of the ball screw at time t, T t i represents the temperature of the i-th temperature measurement point at time t, I t represents the motor current at time t, E t represents the thermal error measured by the laser interferometer at time t, P t It represents the power of the ball screw feed system at time t, and n represents the rotation speed of the ball screw.

[0116] Step 4: preprocess the experimental data and theoretical data, including grey relational analysis, outlier processing and normalization, and use the processed data as a sample data set for data-model hybrid driven model training;

[0117] First, the collected experimental and theoretical data are processed for outliers, and the abnormal data are preliminarily screened out based on the physical model and prior data;

[0118] Then, through grey correlation analysis, the correlation degree of different factors affecting the thermal error of the ball screw is determined, and the influencing factors α of different factors are given. i ;

[0119] Finally, the maximum and minimum normalization is used to normalize the theoretical data of the ball screw thermal error. The processing method is as follows:

[0120]

[0121] Where E is the theoretical value obtained by the ball screw thermal error physical model, E min is the minimum theoretical value obtained by the ball screw thermal error physical model, E max It is the maximum theoretical value obtained by the physical model of ball screw thermal error.

[0122] Step 5: Build a data-model hybrid driven model, namely, a convolutional long short-term memory network (C-LSTMN), and train the sample data set of the data-model hybrid driven model in step 4;

[0123] The convolutional long short-term memory network (C-LSTMN) model includes a convolutional neural network (CNN) module and a long short-term memory network (LSTMN) module, such as Figure 5 shown.

[0124] The convolutional neural network (CNN) layer is convolved with two convolution kernels of size K1 = 1 × 1 and K2 = 3 × 3. The data after grayscale correlation analysis, outlier processing and normalization in step 4 above is convolved through the input layer to extract the features between variables, and different features are learned through different convolution kernels. The feature vector contains the combined information of temperature, current and power, and retains the variable features with the strongest correlation with thermal error for subsequent model training;

[0125] The Long Short-Term Memory Network (LSTMN) consists of an input layer, a hidden layer, and an output layer;

[0126] The input layer is the output of the above convolutional neural network;

[0127] The basic unit of the hidden layer is the storage unit, which is composed of self-connected storage units with storage time states. The storage unit can transfer information from the previous output to the current output. By constantly memorizing and past input data, this layer learns the overall change characteristics of temperature, current, power and speed. Each storage unit consists of four parts: input gate, forget gate, autoregressive connected neuron and output gate. The output expression of LSTMN is:

[0128] Y=f(wx t +b)

[0129] Among them, x t is the input matrix containing temperature, power, current, and speed, Y represents the output of the fully connected layer, that is, the thermal error E of the ball screw, w represents the weight matrix of the fully connected layer, and b represents the bias vector of the fully connected layer.

[0130] The prediction performance of the algorithm is evaluated by the mean absolute error (MAE), and its mathematical expressions are:

[0131]

[0132] Among them, h(x r ) is the predicted value, y r is the true value, r is the number of samples, r=1,2,…,q,x r The input of the model is the mean absolute error. The iterative training of the network model is determined by setting a target mean absolute error until the accuracy requirement is met.

[0133] Step 6: Based on the prediction model of the data-model hybrid drive in step 5, the temperature, power, current and speed collected in real time are input into the prediction model to predict the thermal error of the ball screw at the next moment.

Claims

1. A data-model hybrid driven ball screw thermal error prediction method, characterized in that: The steps include: Step 1, establish a physical model of thermal error of the ball screw of a CNC machine tool and obtain theoretical data; Step 2, using ANSYS Workbench to analyze the steady-state thermal characteristics of the ball screw and determine the optimal layout position of the temperature sensor, i.e., the temperature measurement point; Step 3, build a CNC machine tool ball screw feed system test platform to obtain the temperature, power, speed, current and thermal error data of the ball screw feed system; Step 4: preprocess the experimental data and theoretical data, including grey relational analysis, outlier processing and normalization, and use the processed data as a sample data set for data-model hybrid driven model training; Step 5: Build a data-model hybrid driving model, i.e., a convolutional long short-term memory network C-LSTMN model, and train the sample data set of the data-model hybrid driving model in step 4; Step 6, according to the data-model hybrid drive model of step 5, the thermal error of the ball screw feed system of the CNC machine tool is calculated in real time.

2. A data-model hybrid driven ball screw thermal error prediction method according to claim 1, characterized in that: The process of establishing the physical model of thermal error of the ball screw of a CNC machine tool in step 1 is as follows: First, according to heat transfer theory, the heat transfer process includes: heat conduction, heat convection and radiation heat transfer; according to Fourier's law of heat conduction, the expression of the heat flux density of the CNC machine tool ball screw components in the heat conduction process is: Among them, q t is the heat flux density of heat conduction; λ is the thermal conductivity, T p is the temperature of the component, is the temperature gradient of the component, and the negative sign indicates that the direction of heat transfer is opposite to the direction of the temperature gradient; The heat flux density of the convective heat transfer of the ball screw parts of CNC machine tools is calculated according to the Newton cooling formula, and its expression is: q c =h c ·(T p -T f ) Among them, q c is the heat flux density of the convective heat transfer process, h c is the convective heat transfer coefficient, T f is the temperature of the air in contact with the surface of the component; The heat flux density radiated outward by the ball screw components of CNC machine tools is calculated using the modified form of the Stefan-Boltzmann law, and its expression is: Among them, q ro is the heat flux density of the radiation heat transfer process, ε is the emissivity of the component, and its value is between 0 and 1, c r is the blackbody radiation coefficient; Then, the temperature field T of the screw due to the bearing and screw nut is calculated by heat conduction, heat convection and radiation heat transfer formulas. st (x,t); Finally, the thermal error of the screw at any time t is: Where c1 is the thermal expansion coefficient of the screw and c2 is the thermal drift correction factor.

3. The method for predicting ball screw thermal errors driven by data-model hybrid according to claim 1, characterized in that: The steps of performing steady-state thermal characteristic analysis on the ball screw by ANSYS Workbench in step 2 are as follows: First, the 3D model of the ball screw feed system is simplified, exported into a .x_t file, and input into Workbench. Then, the 3D model of the ball screw feed system is meshed. Next, parameters are set according to the material properties of each part. Finally, the steady-state thermal characteristics of the ball screw are solved. The temperature measurement points in the ball screw are determined through finite element analysis.

4. The method for predicting ball screw thermal error driven by data-model hybrid according to claim 1, characterized in that: The process of building a CNC machine tool ball screw feed system test platform in step 3 is as follows: By arranging temperature sensors at the temperature measuring points in the above step 2, measuring the temperature of the temperature measuring points in the ball screw feed system and the room temperature; measuring the axial thermal error of the ball screw; and measuring the power and speed of the ball screw feed system; the data matrix of the ball screw is obtained, which is expressed as: X t =(T t 1 ,T t 2 ,T t 3 ,T t 4 ,T t 5 ,E t ,P t ,I t ,n) Among them, X t Represents the state data matrix of the ball screw at time t, T t i represents the temperature of the i-th temperature measurement point at time t, I t represents the motor current at time t, E t represents the thermal error measured by laser interferometer at time t, P t It represents the power of the ball screw feed system at time t, and n represents the rotation speed of the ball screw.

5. The method for predicting ball screw thermal error driven by data-model hybrid according to claim 1, characterized in that: The process of step 4 is as follows: First, the collected experimental and theoretical data are processed for outliers, and the abnormal data are preliminarily screened out based on the physical model and prior data; Then, through grey correlation analysis, the correlation degree of different factors affecting the thermal error of the ball screw is determined, and the influencing factors α of different factors are given. i ; Finally, the maximum and minimum normalization is used to normalize the theoretical data of the ball screw thermal error. The processing method is as follows: Where E is the theoretical value obtained by the ball screw thermal error physical model, E min is the minimum theoretical value obtained by the ball screw thermal error physical model, E max It is the maximum theoretical value obtained by the physical model of ball screw thermal error.

6. The method for predicting ball screw thermal error driven by data-model hybrid according to claim 1, characterized in that: The convolutional long short-term memory network C-LSTMN model in step 5 includes a convolutional neural network CNN module and a long short-term memory network LSTMN module; The convolutional neural network (CNN) module includes a single convolutional layer, i.e., data input-convolution-output; the convolutional neural network (CNN) module performs convolution through two convolution kernels of size K1=1×1 and K2=3×3. After the data after grayscale correlation analysis, outlier processing and normalization in step 4 is convolved through the input layer, the features between the variables are extracted, and different features are learned through different convolution kernels; The feature vector contains the combined information of temperature, current, and power, and retains the variable features with the strongest correlation with thermal error for subsequent model training; The long short-term memory network LSTMN module consists of an input layer, a hidden layer and an output layer; the input layer is the output of the above-mentioned convolutional neural network CNN module; the basic unit of the hidden layer is a storage unit, which is composed of a self-connected storage unit with a storage time state, and the storage unit transfers information from the previous output to the current output; by constantly memorizing and past input data, this layer learns the overall change characteristics of temperature, current, power and speed; each storage unit consists of four parts: an input gate, a forget gate, an autoregressive connected neuron and an output gate; the output expression of LSTMN is: Y=f(wx t +b) Among them, x t is the input matrix containing temperature, power, current, and speed, Y represents the output of the fully connected layer, that is, the thermal error E of the ball screw, w represents the weight matrix of the fully connected layer, and b represents the bias vector of the fully connected layer; The prediction performance of the algorithm is evaluated by the mean absolute error MAE, and its mathematical expressions are: Among them, h(x r ) is the predicted value, y r is the true value, r is the number of samples, r=1,2,…,q,x r is the input of the model; a target mean absolute error is set to determine the iterative training of the network model until the accuracy requirement is met.

7. The method for predicting ball screw thermal errors driven by data-model hybrid according to claim 1, characterized in that: The step 6 inputs the temperature, power, current and speed collected in real time into the model according to the data-model hybrid drive model in step 5, and predicts the thermal error of the ball screw at the next moment.

Images