A machine tool process parameter self-adaptive control method based on multi-neural network fusion

Abstract

The application discloses a machine tool process parameter self-adaptive control method based on multi-neural network fusion, and belongs to the technical field of intelligent manufacturing. Firstly, data acquisition and preprocessing are performed, spindle power data and spindle vibration data during machining of a numerical control machine tool are acquired; then, a radial basis neural network and a stacked auto-encoder neural network are trained by using the preprocessed spindle power data and the spindle vibration data, a constraint target prediction model and a real-time cutting depth prediction model are constructed; then, a random forest algorithm is trained, a feed speed prediction model and a spindle speed prediction model are constructed; finally, self-adaptive control of machine tool process parameters is performed. In the self-adaptive control process of machine tool process parameters, the method fuses the suppression of chatter, avoids a complex chatter identification method, improves machining efficiency, guarantees machining quality of parts, and meets the high-quality and high-efficiency machining demand of uneven machining allowance parts.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing technology and relates to an adaptive control method for machine tool process parameters using multi-neural network fusion. Background Technology

[0002] The principle of adaptive control of machine tool process parameters is to maximize machining efficiency while ensuring part quality, thus meeting the demands for high-quality and high-efficiency machining. In traditional CNC machining, machine tool process parameters are determined during the programming stage before machining. These parameters are often fixed and relatively conservative, based on operator experience or computer-aided manufacturing software. In reality, using fixed machine tool process parameters cannot adapt to actual machining conditions, leading to several adverse consequences: on the one hand, if the machine tool process parameters are too small, machining efficiency will be low, making it difficult to fully utilize the performance of the CNC machine tool; on the other hand, if the machine tool process parameters are too large or their combination is unreasonable, excessive stress, machining temperature, and machining power will result in large tool deformation, rapid wear rate, and decreased workpiece surface quality, and in severe cases, damage to the tool or even the spindle. Therefore, adaptive control of machine tool process parameters is a key issue in improving the application performance of CNC systems, part machining efficiency, and machining quality.

[0003] In the field of adaptive control of machine tool process parameters, scholars have already conducted relevant research. In the patent "An Adaptive Control Method for Machine Tools Considering Chatter Suppression" (application number: CN202010522482.2), feed rate and spindle speed are optimized and adaptively adjusted in real time based on neural network algorithms and weighted wavelet packet entropy methods, which can effectively improve machining efficiency and quality. In the patent "A Real-Time Adaptive Optimization Method for CNC Machining Parameters" (application number: 201410719430.9), a model corresponding to spindle motor current and actual cutting force is established, and fuzzy algorithms are used to optimize spindle speed and feed rate, effectively improving machining quality and efficiency, reducing tool and machine tool damage, and lowering production costs. In the patent "An Optimization Method for Adaptive Fuzzy Control Rules of CNC Machining Parameters" (application number: CN201310081486.1), the power bond graph method is used to optimize the fuzzy control rules of adaptive control of CNC machining, which improves the control performance and machining stability. In the paper "Online Monitoring of Surface Roughness of Composite Material Drilling and Adaptive Optimization of Machining Parameters" (Journal of Mechanical Engineering, 2020, 56(02): 27-34+42.), an online monitoring model of hole wall roughness based on support vector regression is established, and the simulated annealing algorithm is used to optimize the drilling parameters under the current monitoring state to ensure the drilling quality.

[0004] However, through the analysis of the above-mentioned existing technologies, we believe that the industry still has the following technical problems that need to be solved: (1) Although many scholars have studied the adaptive control method of machine tool process parameters, the chatter caused by unreasonable parameters can improve the machining efficiency, aggravate tool wear, and damage machine tool parts. Chatter identification methods are complex and cumbersome and not suitable for real-time machining. Therefore, it is necessary to provide a machine tool process parameter adaptive control method that can integrate chatter suppression. (2) At present, most machine tool process parameter adaptive control algorithms are based on fuzzy control and neural network theory. Although fuzzy control is very suitable for handling systems that are difficult to establish accurate mathematical models, fuzzy reasoning based on expert experience may not be accurate enough for complex machining conditions. As for control methods using neural network theory, due to the difficulty in obtaining the labels of the controlled objects, it is rare to achieve direct control of feed rate and spindle speed. In summary, the existing machine tool process parameter adaptive control methods do not consider the cutting chatter that may be caused during parameter adjustment. Chatter can affect machining quality, aggravate tool wear, and damage machine tool parts. To address the aforementioned industry problems, this invention proposes a multi-neural network fusion-based adaptive control method for machine tool process parameters. By constructing a multi-neural network fusion model, chatter during machine tool processing can be effectively avoided, and feed rate and spindle speed can be directly controlled, significantly improving the processing efficiency and quality of CNC machine tools. Summary of the Invention

[0005] To address the problems of common machine tool process parameter adaptive control methods failing to consider the impact of chatter caused by changes in machining parameters on machining results, and being unable to directly control feed rate and spindle speed, this invention designs a multi-neural network fusion machine tool process parameter adaptive control method, which is implemented through the following steps:

[0006] S1: Data acquisition and preprocessing, specifically, acquiring spindle power data and spindle vibration data during CNC machine tool processing, and sending the acquired spindle power data and spindle vibration data to the industrial control computer after preprocessing;

[0007] S2: Construct a constraint target prediction model and a real-time cutting depth prediction model using the industrial control computer. Specifically, use the spindle power data and spindle vibration data preprocessed in step S1 to train a radial basis function neural network (RBF neural network) and a stacked autoencoder neural network (SAE neural network) to generate a constraint target prediction model and a real-time cutting depth prediction model, and output the constraint target prediction value and the real-time cutting depth prediction value.

[0008] S3: Construct a feed rate prediction model and a spindle speed prediction model using the industrial control computer. Specifically, use the spindle power data and spindle vibration data obtained after preprocessing in step S1, as well as the constraint target prediction value and the real-time cutting depth prediction value output in step S2, to train the random forest algorithm (RF algorithm) and generate the feed rate prediction model and the spindle speed prediction model.

[0009] S4: Adaptive control of machine tool process parameters is performed using the industrial control computer. Specifically, the outputs of the constraint target prediction model and the real-time cutting depth prediction model described in step S2 are used as the inputs of the feed rate prediction model and the spindle speed prediction model described in step S3. A multi-neural network fusion model is constructed. The feed rate, spindle speed, and cutting depth set in the machining program are input into the trained network. At the same time, the spindle power data and spindle vibration data are collected in real time. The optimal feed rate and optimal spindle speed of the CNC machine tool are obtained through the multi-neural network fusion model, thereby realizing real-time control of the machine tool process parameters.

[0010] Furthermore, S1 specifically refers to,

[0011] S1.1: Install a power sensor and a three-axis accelerometer at appropriate locations on the CNC machine tool. Use the power sensor to collect spindle power data during the machining process, and use the formula... Calculate the average value of the collected spindle power data, where P i The spindle power data is collected in real time. Let n be the average value of the spindle power data. P The total number of spindle power data samples taken during the entire machining process of the CNC machine tool;

[0012] S1.2: The spindle vibration data in the X, Y, and Z directions during machining is collected using the triaxial accelerometer. The root mean square (RMS) values ​​of the spindle vibration data in the X, Y, and Z directions are calculated using the following formula:

[0013]

[0014] Among them, R X R is the root mean square value of the spindle vibration data in the X direction. Y R is the root mean square value of the spindle vibration data in the Y direction. Z n is the root mean square value of the spindle vibration data in the Z direction. R A represents the total number of spindle vibration data samples collected during the entire machining process of the CNC machine tool. j A represents the vibration data of the main shaft acquired in the X direction. kA represents the vibration data of the main shaft collected in the Y direction. l The vibration data of the main shaft collected in the Z direction;

[0015] S1.3: The collected spindle power data and spindle vibration data are sent to the industrial control computer.

[0016] Furthermore, step S2 consists of two steps:

[0017] S2.1: Construct the constraint target prediction model based on a radial basis function neural network. Let the feed rate set in the machining program be f0, the spindle speed set in the machining program be S0, and the depth of cut set in the machining program be d0. The constraint target prediction model outputs the constraint target prediction value based on the feed rate f0, the spindle speed S0, and the depth of cut d0. The constraint target prediction value includes the target cutting power value P0 and the root mean square amplitude R in the X direction. X0 Root mean square amplitude in the Y direction R Y0 Root mean square amplitude R in the Z direction Z0 ;

[0018] S2.2: Construct the real-time cutting depth prediction model based on a stacked autoencoder neural network. The stacked autoencoder neural network includes an encoder and a decoder. The stacked autoencoder neural network is divided into an input layer, a hidden layer, and a softmax classifier. By real-time monitoring of the spindle power data and the spindle vibration data, the stacked autoencoder neural network is used to perform dimensionality reduction and feature extraction on the spindle power data and the spindle vibration data. Finally, the softmax classifier is used for classification, and the real-time cutting depth prediction value during the machining process is output.

[0019] Furthermore, the specific steps of step S2.1 are as follows:

[0020] S2.1.1: First, a radial basis function neural network is established, which is divided into an RBF input layer, an RBF hidden layer, and an RBF output layer. The RBF output layer is a linear weighted sum of the outputs of the RBF hidden layer.

[0021] S2.1.2: Input signal source a to the radial basis neural network, wherein the signal source a consists of the spindle speed S0, the feed rate f0, and the depth of cut d0;

[0022] S2.1.3: The output of the radial basis function neural network is calculated using the activation function formula of the m-th neuron in the RBF hidden layer and the RBF output layer formula. The activation function formula of the m-th neuron in the RBF hidden layer is ρ(a, C). m )=exp(-β m ||aCm || 2 ), where β m C represents the input received by the m-th neuron. m For the m-th center point, the activation function ρ(a,C) of the m-th neuron in the RBF hidden layer m ) is for the center point C m A radially symmetric and decaying nonnegative linear function; the formula for the RBF output layer is: Where r is the number of nodes in the RBF hidden layer, w ki β represents the weights from the RBF hidden layer to the RBF output layer. k C represents the input received by the k-th neuron. k b is the kth center point; i The output of the radial basis function neural network is given, where output b1 is the target cutting power value P0, and outputs b2, b3, and b4 are the root mean square amplitude R in the X direction, respectively. X0 The root mean square amplitude R in the Y direction Y0 and the root mean square amplitude R in the Z direction Z0 .

[0023] Furthermore, the specific steps of step S2.2 are as follows:

[0024] S2.2.1: Establish a stacked self-encoder neural network, assuming the spindle speed is S0, the feed rate is f0, and the real-time acquired spindle power data P... i The main shaft vibration data A in the X, Y, and Z directions. j A k A l The matrix c formed includes the spindle speed S0, the feed rate f0, and the spindle power data P. i The matrix c is input into the stacked autoencoder neural network, and the output of the stacked autoencoder neural network is calculated using the SAE hidden layer formula and the softmax classifier formula.

[0025] S2.2.2: Let the number of hidden layers in the SAE be i, and the number of input layer vectors in the SAE be c, then the formula for the hidden layer h of the SAE is: h = f(w ij +b ij ), where w ij Let b be the weight vector from the input layer vector c to the SAE hidden layer. ij Let f(·) be the bias vector from the input layer vector c to the SAE hidden layer, and f(·) be the activation function from the input layer vector c to the SAE hidden layer.

[0026] S2.2.3: The formula for the softmax classifier is di =g(w ji +b ji ), where d i w is the predicted real-time cutting depth. ji Let b be the weight vector from the SAE hidden layer to the softmax classifier. ji Let g(·) be the bias vector from the SAE hidden layer to the softmax classifier, and g(·) be the activation function from the SAE hidden layer to the softmax classifier. The output of the stacked autoencoder neural network is the real-time cutting depth prediction value d. i .

[0027] Furthermore, step S3 involves setting the optimal feed rate as f. c The optimal spindle speed is s c Based on the target cutting power value P0 and the root mean square amplitude R in the X direction output in step S2 X0 The root mean square amplitude R in the Y direction Y0 The root mean square amplitude R in the Z direction Z0 The real-time cutting depth prediction value d i A feed rate prediction model and a spindle speed prediction model based on the random forest algorithm are constructed to obtain the optimal feed rate f. c and optimal spindle speed s c .

[0028] Furthermore, step S4 is as follows:

[0029] S4.1: Let the feed rate multiplier of the CNC machine tool be M. f The calculation formula is: Among them, f c For the optimal feed rate, f o The feed rate set for the machining program;

[0030] S4.2: Let the spindle speed multiplier of the CNC machine tool be M. s The calculation formula is: Among them, s c For the optimal spindle speed, s o The spindle speed set for the machining program;

[0031] S4.3: The industrial computer and the CNC system of the CNC machine tool communicate in real time;

[0032] S4.4: Adjust the feed rate multiplier control value M f and the spindle speed multiplier control value M sThe data is sent to the CNC system to achieve adaptive control of the machine tool process parameters.

[0033] The beneficial effects of this invention are as follows:

[0034] (1) The machine tool process parameter adaptive control method of the present invention, which integrates multiple neural networks, fully considers the problem of chattering that may be caused during the machining process. On the basis of conventional adjustment of feed speed, the function of adjusting spindle speed is added. Through the method of integration of multiple neural networks, the coordinated adjustment of feed speed and spindle speed is realized. While improving machining efficiency, chattering is suppressed, which effectively improves the machining quality of CNC machine tool parts.

[0035] (2) The machine tool process parameter adaptive control method of multi-neural network fusion of the present invention effectively improves the processing efficiency of parts with uneven processing allowance and saves production time and cost.

[0036] (3) The machine tool process parameter adaptive control method of the present invention, which integrates multiple neural networks, can adaptively adjust the machine tool process parameters during the machining process, thereby improving the intelligence level of CNC machine tools and reducing the technical requirements for operators. Attached Figure Description

[0037] Figure 1 This is a schematic diagram showing the arrangement of the power sensor and the triaxial accelerometer in an embodiment.

[0038] Figure 2 This is a diagram of the radial basis neural network structure used for constrained target prediction in an example.

[0039] Figure 3 This is a diagram of a stacked autoencoder neural network structure used for real-time cutting depth prediction in an embodiment.

[0040] Figure 4 This is a network structure diagram of the random forest algorithm used for feed rate prediction and spindle speed prediction in an example embodiment;

[0041] Figure 5 The flowchart below shows the network model of the machine tool process parameter adaptive control method using multi-neural network fusion as an example.

[0042] Figure 6 The feed rate multiplier control value M f and spindle speed ratio control value M s Graph showing changes over time;

[0043] In the diagram: 1 is a vertical CNC milling machine; 2 is the spindle servo motor; 3 is the worktable; 4 is the power sensor; 5 is the three-axis accelerometer; 6 is the spindle. Detailed Implementation

[0044] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0045] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", "front", "back", etc., which indicate the direction or positional relationship, are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0046] This embodiment uses a vertical CNC milling machine; the control method for other CNC machine tools is the same as in this embodiment. Embodiment 1: A multi-neural network fusion-based adaptive control method for machine tool process parameters is implemented through the following steps:

[0047] S1: Data acquisition and preprocessing, specifically, acquiring spindle power data and spindle vibration data during CNC machine tool machining, and preprocessing the acquired spindle power data and spindle vibration data before sending them to the industrial control computer; S1 is specifically divided into three steps.

[0048] S1.1: A power sensor 4 and a three-axis accelerometer 5 are installed at appropriate positions on the CNC machine tool. In this embodiment, a spindle servo motor 2 is installed on the vertical CNC milling machine 1, located above the worktable 3. The power sensor 4 is arranged on the spindle servo motor 2, and its three current sensing coils are respectively wrapped around the three power lines U, V, and W of the spindle motor. The voltage is connected to the three power lines U, V, and W in parallel. The three-axis accelerometer 5 is magnetically attached to the side of the spindle 6 near the worktable 3. The power sensor 4 is used to collect the spindle power data during the cutting process, and the formula is used to calculate the power data. Calculate the average value of the collected spindle power data, where P i The spindle power data is collected in real time. Let n be the average value of the spindle power data. P The total number of spindle power data samples taken during the entire machining process of the CNC machine tool;

[0049] S1.2: The spindle vibration data in the X, Y, and Z directions during the cutting process are collected using a triaxial accelerometer 5. The root mean square (RMS) values ​​of the spindle vibration data in the X, Y, and Z directions are calculated using the following formula:

[0050]

[0051] Among them, R XR is the root mean square value of the spindle vibration data in the X direction. Y R is the root mean square value of the spindle vibration data in the Y direction. Z n is the root mean square value of the spindle vibration data in the Z direction. R A represents the total number of spindle vibration data samples collected during the entire machining process of the CNC machine tool. j A represents the vibration data of the main shaft acquired in the X direction. k A represents the vibration data of the main shaft collected in the Y direction. l The vibration data of the main shaft collected in the Z direction;

[0052] S1.3: The collected spindle power data and spindle vibration data are sent to the industrial control computer.

[0053] S2: Construct a constraint target prediction model and a real-time cutting depth prediction model using the industrial control computer. Specifically, use the spindle power data and spindle vibration data preprocessed in step S1 to train a radial basis neural network and a stacked autoencoder neural network to generate a constraint target prediction model and a real-time cutting depth prediction model, and output the constraint target prediction value and the real-time cutting depth prediction value.

[0054] S2 specifically consists of two steps:

[0055] S2.1: Radial basis function neural networks are a type of single-hidden-layer feedforward neural network, and their topology is as follows: Figure 2 As shown, a constraint target prediction model based on a radial basis function neural network is constructed. Let the feed rate set in the machining program be f0, the spindle speed set in the machining program be S0, and the depth of cut set in the machining program be d0. The constraint target prediction model outputs the predicted constraint target value based on the feed rate f0, the spindle speed S0, and the depth of cut d0. The predicted constraint target value includes the target cutting power value P0 and the root mean square amplitude R in the X direction. X0 Root mean square amplitude in the Y direction R Y0 Root mean square amplitude R in the Z direction Z0 Specifically:

[0056] S2.1.1: First, a radial basis function neural network is established, which is divided into an RBF input layer, an RBF hidden layer, and an RBF output layer. The RBF output layer is a linear weighted sum of the outputs of the RBF hidden layer.

[0057] S2.1.2: Input signal source a to the radial basis neural network, wherein the signal source a consists of the spindle speed S0, the feed rate f0, and the depth of cut d0;

[0058] S2.1.3: The output of the radial basis function neural network is calculated using the activation function formula of the m-th neuron in the RBF hidden layer and the RBF output layer formula. The activation function formula of the m-th neuron in the RBF hidden layer is ρ(a, C). m )=exp(-β m ||aC m || 2 ), where β m C represents the input received by the m-th neuron. m For the m-th center point, the activation function ρ(a,C) of the m-th neuron in the RBF hidden layer m ) is for the center point C m A radially symmetric and decaying nonnegative linear function; the formula for the RBF output layer is: Where r is the number of nodes in the RBF hidden layer, w ki β represents the weights from the RBF hidden layer to the RBF output layer. k C represents the input received by the k-th neuron. k b is the kth center point; i The output of the radial basis function neural network is given, where output b1 is the target cutting power value P0, and outputs b2, b3, and b4 are the root mean square amplitude R in the X direction, respectively. X0 The root mean square amplitude R in the Y direction Y0 and the root mean square amplitude R in the Z direction Z0 .

[0059] S2.2: Stacked autoencoder neural networks are an unsupervised learning technique. Their function is to learn representations of the input information by using the input information as the learning target. Their topology is as follows: Figure 3 As shown, a real-time cutting depth prediction model based on a stacked autoencoder neural network is constructed. The stacked autoencoder neural network includes an encoder and a decoder, and consists of an input layer, a hidden layer, and a softmax classifier. By real-time monitoring of the spindle power data and spindle vibration data, the stacked autoencoder neural network is used to perform dimensionality reduction and feature extraction on the spindle power data and spindle vibration data. Finally, the softmax classifier is used for classification, outputting the real-time cutting depth prediction value during the machining process. Specifically:

[0060] S2.2.1: Establish a stacked self-encoder neural network, assuming the spindle speed is S0, the feed rate is f0, and the real-time acquired spindle power data P... i The main shaft vibration data A in the X, Y, and Z directions. j A k A lThe matrix c formed includes the spindle speed S0, the feed rate f0, and the spindle power data P. i The matrix c is input into the stacked autoencoder neural network, and the output of the stacked autoencoder neural network is calculated using the SAE hidden layer formula and the softmax classifier formula.

[0061] S2.2.2: Let the number of SAE hidden layers be i. In this embodiment, the number of SAE hidden layers i is 3, that is, the first layer is the SAE input layer, the second to fourth layers are SAE hidden layer 1, SAE hidden layer 2 and SAE hidden layer 3 respectively, and the fifth layer is a softmax classifier. The SAE input layer vector is c, and the number of vectors is j. Then the formula for the SAE hidden layer h is: h = f(w ij +b ij ), where w ij Let b be the weight vector from the input layer vector c to the SAE hidden layer. ij Let f(·) be the bias vector from the input layer vector c to the SAE hidden layer, and f(·) be the activation function from the input layer vector c to the SAE hidden layer.

[0062] S2.2.3: The formula for the softmax classifier is d i =g(w ji +b ji ), where d i w is the predicted real-time cutting depth. ji Let b be the weight vector from the SAE hidden layer to the softmax classifier. ji Let g(·) be the bias vector from the SAE hidden layer to the softmax classifier, and g(·) be the activation function from the SAE hidden layer to the softmax classifier. The output of the stacked autoencoder neural network is the real-time cutting depth prediction value d. i .

[0063] S3: Construct a feed rate prediction model and a spindle speed prediction model using the industrial control computer. Specifically, using the spindle power data and spindle vibration data obtained after preprocessing in step S1, as well as the constraint target prediction value and the real-time cutting depth prediction value output in step S2, train a random forest algorithm to generate the feed rate prediction model and the spindle speed prediction model. S3 specifically involves:

[0064] like Figure 4 As shown, let the optimal feed rate be f. c The optimal spindle speed is s c Based on the target cutting power value P0 and the root mean square amplitude R in the X direction output in step S2 X0The root mean square amplitude R in the Y direction Y0 The root mean square amplitude R in the Z direction Z0 The real-time cutting depth prediction value d i A feed rate prediction model and a spindle speed prediction model based on the random forest algorithm are constructed to obtain the optimal feed rate f. c and optimal spindle speed s c .

[0065] S4: Adaptive control of machine tool process parameters is performed using the industrial control computer. Specifically, the outputs of the constraint target prediction model and the real-time cutting depth prediction model described in step S2 are used as the inputs of the feed rate prediction model and the spindle speed prediction model described in step S3. A multi-neural network fusion model is constructed. The feed rate, spindle speed, and cutting depth set by the machining program are input into the trained network. At the same time, the spindle power data and spindle vibration data are collected in real time. The optimal feed rate and optimal spindle speed of the CNC machine tool are obtained through the multi-neural network fusion model, thereby realizing real-time control of the machine tool process parameters.

[0066] S4 consists of four steps:

[0067] S4.1: Let the feed rate multiplier of the CNC machine tool be M. f The calculation formula is: Among them, f c For the optimal feed rate, f o The feed rate set for the machining program;

[0068] S4.2: Let the spindle speed multiplier of the CNC machine tool be M. s The calculation formula is: Among them, s c For the optimal spindle speed, s o The spindle speed set for the machining program;

[0069] S4.3: The industrial computer and the CNC system of the CNC machine tool communicate in real time;

[0070] S4.4: Adjust the feed rate multiplier control value M f and the spindle speed multiplier control value M s The data is sent to the CNC system, which then uses the built-in PLC module of the CNC machine tool to achieve adaptive control of the machine tool's process parameters.

[0071] The network model in this embodiment is as follows: Figure 5 As shown.

[0072] like Figure 6As shown, in this embodiment, when the spindle speed is set to 2000 rpm and the feed rate is set to 100 mm / min in the machining program, the resulting feed rate multiplier control value M is... f and spindle speed ratio control value M s The technical effects of this embodiment are: the part processing efficiency is improved by 27%, production time and costs are saved, and no cutting chatter occurs during the part processing.

[0073] It should be noted that the specific embodiments described above are merely illustrative of the principles and processes of the present invention and do not constitute a limitation thereof. Therefore, any modifications and equivalent substitutions made without departing from the spirit and scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for adaptive control of machine tool process parameters using multi-neural network fusion, characterized in that, This can be achieved through the following steps: S1: Data acquisition and preprocessing, specifically, acquiring spindle power data and spindle vibration data during CNC machine tool processing, and sending the acquired spindle power data and spindle vibration data to the industrial control computer after preprocessing; S2: Construct a constraint target prediction model and a real-time cutting depth prediction model using the industrial control computer. Specifically, use the spindle power data and spindle vibration data preprocessed in step S1 to train a radial basis neural network and a stacked autoencoder neural network to generate a constraint target prediction model and a real-time cutting depth prediction model, and output the constraint target prediction value and the real-time cutting depth prediction value. S3: Construct a feed rate prediction model and a spindle speed prediction model using the industrial control computer. Specifically, use the spindle power data and spindle vibration data obtained after preprocessing in step S1, as well as the constraint target prediction value and the real-time cutting depth prediction value output in step S2, to train a random forest algorithm and generate a feed rate prediction model and a spindle speed prediction model. S4: Adaptive control of machine tool process parameters is performed using the industrial control computer. Specifically, the outputs of the constraint target prediction model and the real-time cutting depth prediction model described in step S2 are used as the inputs of the feed rate prediction model and the spindle speed prediction model described in step S3. A multi-neural network fusion model is constructed. The feed rate, spindle speed, and cutting depth set in the machining program are input into the trained network. At the same time, the spindle power data and spindle vibration data are collected in real time. The optimal feed rate and optimal spindle speed of the CNC machine tool are obtained through the multi-neural network fusion model, thereby realizing real-time control of the machine tool process parameters.

2. The machine tool process parameter adaptive control method based on multi-neural network fusion as described in claim 1, characterized in that, The specific steps of step S1 are as follows: S1.1: Install a power sensor and a three-axis accelerometer at appropriate locations on the CNC machine tool. Use the power sensor to collect spindle power data during the machining process, and use the formula... Calculate the average value of the collected spindle power data, where P i The spindle power data is collected in real time. Let n be the average value of the spindle power data. P The total number of spindle power data samples taken during the entire machining process of the CNC machine tool; S1.2: The spindle vibration data in the X, Y, and Z directions during machining is collected using the triaxial accelerometer. The root mean square (RMS) values ​​of the spindle vibration data in the X, Y, and Z directions are calculated using the following formula: wherein R X is the root mean square value of the spindle vibration data in the X direction, R Y is the root mean square value of the spindle vibration data in the Y direction, R Z is the root mean square value of the spindle vibration data in the Z direction, n R is the total number of samples of the spindle vibration data in the entire machining process of the numerical control machine tool, A j is the spindle vibration data collected in the X direction, A k is the spindle vibration data collected in the Y direction, and A l is the spindle vibration data collected in the Z direction. S1.3: The collected spindle power data and spindle vibration data are sent to the industrial control computer.

3. The machine tool process parameter adaptive control method based on multi-neural network fusion as described in claim 2, characterized in that, S2 consists of two steps: S2.1: Construct the constraint target prediction model based on a radial basis function neural network. Let the feed rate set in the machining program be f0, the spindle speed set in the machining program be S0, and the depth of cut set in the machining program be d0. The constraint target prediction model outputs the constraint target prediction value based on the feed rate f0, the spindle speed S0, and the depth of cut d0. The constraint target prediction value includes the target cutting power value P0 and the root mean square amplitude R in the X direction. X0 Root mean square amplitude in the Y direction R Y0 Root mean square amplitude R in the Z direction Z0 ; S2.2: Construct the real-time cutting depth prediction model based on a stacked autoencoder neural network. The stacked autoencoder neural network includes an encoder and a decoder. The stacked autoencoder neural network is divided into an input layer, a hidden layer, and a softmax classifier. By real-time monitoring of the spindle power data and the spindle vibration data, the stacked autoencoder neural network is used to perform dimensionality reduction and feature extraction on the spindle power data and the spindle vibration data. Finally, the softmax classifier is used for classification, and the real-time cutting depth prediction value during the machining process is output.

4. The machine tool process parameter adaptive control method based on multi-neural network fusion as described in claim 3, characterized in that, Step S2.1 is as follows: S2.1.1: First, a radial basis function neural network is established, which is divided into an RBF input layer, an RBF hidden layer, and an RBF output layer. The RBF output layer is a linear weighted sum of the outputs of the RBF hidden layer. S2.1.2: Input signal source a to the radial basis neural network, wherein the signal source a consists of the spindle speed S0, the feed rate f0, and the depth of cut d0; S2.1.3: The output of the radial basis function neural network is calculated using the activation function formula of the m-th neuron in the RBF hidden layer and the RBF output layer formula. The activation function formula of the m-th neuron in the RBF hidden layer is ρ(a, C). m )=exp(-β m ||aC m || 2 ), where β m C represents the input received by the m-th neuron. m For the m-th center point, the activation function ρ(a,C) of the m-th neuron in the RBF hidden layer m ) is for the center point C m A radially symmetric and decaying nonnegative linear function; the formula for the RBF output layer is: Where r is the number of nodes in the RBF hidden layer, w ki β represents the weights from the RBF hidden layer to the RBF output layer. k C represents the input received by the k-th neuron. k b is the kth center point; i The output of the radial basis function neural network is given, where output b1 is the target cutting power value P0, and outputs b2, b3, and b4 are the root mean square amplitude R in the X direction, respectively. X0 The root mean square amplitude R in the Y direction Y0 and the root mean square amplitude R in the Z direction Z0 .

5. The machine tool process parameter adaptive control method based on multi-neural network fusion as described in claim 4, characterized in that, The specific steps of step S2.2 are as follows: S2.2.1: Establish a stacked self-encoder neural network, assuming the spindle speed is S0, the feed rate is f0, and the real-time acquired spindle power data P... i The main shaft vibration data A in the X, Y, and Z directions. j A k A l The matrix c formed includes the spindle speed S0, the feed rate f0, and the spindle power data P. i The matrix c is input into the stacked autoencoder neural network, and the output of the stacked autoencoder neural network is calculated using the SAE hidden layer formula and the softmax classifier formula. S2.2.2: Let the number of hidden layers in the SAE be i, and the number of input layer vectors in the SAE be c, then the formula for the hidden layer h of the SAE is: h = f(w ij +b ij ), where w ij Let b be the weight vector from the input layer vector c to the SAE hidden layer. ij Let f(·) be the bias vector from the input layer vector c to the SAE hidden layer, and f(·) be the activation function from the input layer vector c to the SAE hidden layer. S2.2.3: The formula for the softmax classifier is d i =g(w ji +b ji ), where d i w is the predicted real-time cutting depth. ji Let b be the weight vector from the SAE hidden layer to the softmax classifier. ji Let g(·) be the bias vector from the SAE hidden layer to the softmax classifier, and g(·) be the activation function from the SAE hidden layer to the softmax classifier. The output of the stacked autoencoder neural network is the real-time cutting depth prediction value d. i .

6. The machine tool process parameter adaptive control method based on multi-neural network fusion as described in claim 5, characterized in that, Step S3 is as follows: Let the optimal feed rate be f. c The optimal spindle speed is s c Based on the target cutting power value P0 and the root mean square amplitude R in the X direction output in step S2 X0 The root mean square amplitude R in the Y direction Y0 The root mean square amplitude R in the Z direction Z0 The real-time cutting depth prediction value d i A feed rate prediction model and a spindle speed prediction model based on the random forest algorithm are constructed to obtain the optimal feed rate f. c and optimal spindle speed s c .

7. The machine tool process parameter adaptive control method based on multi-neural network fusion as described in claim 6, characterized in that, The specific steps of step S4 are as follows: S4.1: Let the feed rate multiplier of the CNC machine tool be M. f The calculation formula is: Among them, f c For the optimal feed rate, f o The feed rate set for the machining program; S4.2: Let the spindle speed multiplier of the CNC machine tool be M. s The calculation formula is: Among them, s c For the optimal spindle speed, s o The spindle speed set for the machining program; S4.3: The industrial computer and the CNC system of the CNC machine tool communicate in real time; S4.4: Adjust the feed rate multiplier control value M f and the spindle speed multiplier control value M s The data is sent to the CNC system to achieve adaptive control of the machine tool process parameters.

Images