## Die-casting machine injection system mechanism model based on time series data drive

Time：2024-10-29 09:11:55 / Popularity： / Source：

In recent years, my country's aerospace and defense industries have developed rapidly, but there are many complex components on aircraft, missiles, spacecraft and satellites, which have extremely high requirements for forming process and need to meet requirements such as lightweight. Resulting forming methods are increasingly receiving attention. Die-casting process has been widely used in the field of aerospace, and can realize manufacture of various complex shapes of castings that cannot be achieved by traditional casting processes. Castings have characteristics of precise size, clear contours, and superior mechanical properties. Injection speed of die-casting machine during injection has a huge impact on quality of castings. Different products require different injection speeds. Therefore, it is necessary to establish a model of injection speed during die-casting machine injection process to accurately control it. At present, many die-casting machines still use ordinary manually adjustable cartridge valves to adjust injection speed. This valve cannot control injection speed in real time. It can only adjust valve core displacement after each injection so that injection speed reaches expected value at next injection. Therefore, it is necessary to establish an accurate model to predict valve core displacement corresponding to expected injection speed.

Researchers derived mathematical formulas of hydraulic cylinder piston rod movement speed and valve core displacement of hydraulic system, established a theoretical formula model, laid foundation for subsequent control and simulation of hydraulic system. Injection system of die-casting machine was modeled, and an open-loop transfer function model of injection system was established. In modeling process, load was simplified, a differential equation mathematical model of valve input signal and injection cylinder injection speed was established based on force balance equation and flow continuity equation. Above studies have greatly simplified real model, have not fully considered large inertia of die-casting machine injection system and friction between injection piston and injection cylinder, punch and mold during injection process, as well as flow change of cartridge valve when cartridge valve core is opened. Therefore, established mathematical model will have a large error with actual system. In order to achieve precise control of injection speed, it is necessary to establish a model that can accurately reflect mechanism of injection system.

Considering that die-casting machine has relatively strong nonlinear and hysteresis characteristics, it is difficult to establish an accurate mathematical model of its injection system, and injection speed data has strong time series characteristics. LSTM neural network can effectively process time series, which provides a new method and idea for solving problem of difficulty in obtaining an accurate model of injection system of die-casting machine. This study proposes a modeling method based on time series data driven to model relationship between speed curve and valve core displacement of the entire injection process. Considering all information contained in speed curve, LSTM neural network automatically learns characteristic information contained in speed curve and established model can more accurately describe relationship between injection speed and valve core displacement, effectively solving problems of large errors and inaccurate models in theoretical modeling method, and providing a reference for speed control of injection system with manual speed adjustment.

Researchers derived mathematical formulas of hydraulic cylinder piston rod movement speed and valve core displacement of hydraulic system, established a theoretical formula model, laid foundation for subsequent control and simulation of hydraulic system. Injection system of die-casting machine was modeled, and an open-loop transfer function model of injection system was established. In modeling process, load was simplified, a differential equation mathematical model of valve input signal and injection cylinder injection speed was established based on force balance equation and flow continuity equation. Above studies have greatly simplified real model, have not fully considered large inertia of die-casting machine injection system and friction between injection piston and injection cylinder, punch and mold during injection process, as well as flow change of cartridge valve when cartridge valve core is opened. Therefore, established mathematical model will have a large error with actual system. In order to achieve precise control of injection speed, it is necessary to establish a model that can accurately reflect mechanism of injection system.

Considering that die-casting machine has relatively strong nonlinear and hysteresis characteristics, it is difficult to establish an accurate mathematical model of its injection system, and injection speed data has strong time series characteristics. LSTM neural network can effectively process time series, which provides a new method and idea for solving problem of difficulty in obtaining an accurate model of injection system of die-casting machine. This study proposes a modeling method based on time series data driven to model relationship between speed curve and valve core displacement of the entire injection process. Considering all information contained in speed curve, LSTM neural network automatically learns characteristic information contained in speed curve and established model can more accurately describe relationship between injection speed and valve core displacement, effectively solving problems of large errors and inaccurate models in theoretical modeling method, and providing a reference for speed control of injection system with manual speed adjustment.

**Graphical results**

Injection process of die-casting machine can be divided into a slow stage and a fast stage, in which more than 90% of flow in fast stage is provided by accumulator. Speed in rapid stage is difficult to control, and speed in this stage has a great impact on quality of product. This study only selects rapid stage as research object. At the same time, in order to simplify model, flow provided by oil pump in rapid stage is ignored. Principle of die-casting machine injection system is shown in Figure 1. System is mainly composed of an accumulator, a solenoid valve, a cartridge valve and an injection cylinder.

Displacement of cartridge valve core is adjustable, which is reflected in die-casting machine injection system as adjustment of valve opening. At the same time, on and off of cartridge valve can be controlled by solenoid valve installed on it. Control of injection speed of die-casting machine is achieved by adjusting displacement of cartridge valve core. Process flow of injection process is shown in Figure 2.

Displacement of cartridge valve core is adjustable, which is reflected in die-casting machine injection system as adjustment of valve opening. At the same time, on and off of cartridge valve can be controlled by solenoid valve installed on it. Control of injection speed of die-casting machine is achieved by adjusting displacement of cartridge valve core. Process flow of injection process is shown in Figure 2.

Figure 1 Principle of die-casting machine injection system

1. Mold 2. Punch 3. Magnetic grating displacement sensor 4. Cartridge valve 5. Solenoid valve 6. Accumulator 7. Booster cylinder 8. Injection cylinder

1. Mold 2. Punch 3. Magnetic grating displacement sensor 4. Cartridge valve 5. Solenoid valve 6. Accumulator 7. Booster cylinder 8. Injection cylinder

Figure 2 Injection process flow chart

Injection system mechanism model driven by time series data consists of two parts: data preprocessing part and LSTM neural network model. LSTM neural network model includes sequence input layer, LSTM layer, Relu layer, fully connected layer and regression layer. Its structure is shown in Figure 3. During injection process of die-casting machine, displacement-time information of piston rod is collected by displacement sensor and further processed into speed-time information. Injection speed curve is a curve that changes with time and has a strong time series characteristic. Therefore, LSTM neural network that is good at solving time series data is used to find relationship between injection speed and valve core displacement.

Due to friction between injection piston and injection cylinder, punch and mold of injection system, change in pressure difference between two ends of valve port when cartridge valve core is opened and reaches a steady state, speed curve collected by actual measurement will have noise. It needs to be denoised before analyzing speed curve. First, data preprocessing decomposes speed information into multiple high-frequency and low-frequency components through wavelet decomposition, then corresponding high-frequency part is denoised by wavelet threshold.

Injection system mechanism model driven by time series data consists of two parts: data preprocessing part and LSTM neural network model. LSTM neural network model includes sequence input layer, LSTM layer, Relu layer, fully connected layer and regression layer. Its structure is shown in Figure 3. During injection process of die-casting machine, displacement-time information of piston rod is collected by displacement sensor and further processed into speed-time information. Injection speed curve is a curve that changes with time and has a strong time series characteristic. Therefore, LSTM neural network that is good at solving time series data is used to find relationship between injection speed and valve core displacement.

Due to friction between injection piston and injection cylinder, punch and mold of injection system, change in pressure difference between two ends of valve port when cartridge valve core is opened and reaches a steady state, speed curve collected by actual measurement will have noise. It needs to be denoised before analyzing speed curve. First, data preprocessing decomposes speed information into multiple high-frequency and low-frequency components through wavelet decomposition, then corresponding high-frequency part is denoised by wavelet threshold.

Figure 3 Injection system mechanism model framework

Figure 4 Schematic diagram of wavelet decomposition process

Figure 5 LSTM neural network unit

LSTM neural network is a special recurrent neural network (RNN). On the basis of traditional RNN, memory cell units are added, which can be used to save time state. Each memory cell unit contains a forget gate, an input gate, and an output gate. These three gates can jointly control forgetting and memory of input information, solve problems of RNN gradient disappearance and gradient explosion, can effectively deal with long-term and distance dependency problems, making it better at processing time series data.

Research data comes from a die-casting machine with a clamping force of 2800 kN. Physical diagram of its structure is shown in Figure 6. Magnetic grating sensor installed on piston rod collects speed data corresponding to different opening percentages of fast cartridge valve of die-casting machine in fast stage when accumulator supplies oil. Maximum displacement of cartridge valve core is ymax, and valve opening percentage P takes a more representative value of 20%~52% (a value is taken every 2%). Pressure of accumulator when storing energy is 14 MPa. 30 sets of speed data are collected for each valve opening percentage, a total of 510 sets of data, and some of collected speed curves are shown in Figure 7.

LSTM neural network is a special recurrent neural network (RNN). On the basis of traditional RNN, memory cell units are added, which can be used to save time state. Each memory cell unit contains a forget gate, an input gate, and an output gate. These three gates can jointly control forgetting and memory of input information, solve problems of RNN gradient disappearance and gradient explosion, can effectively deal with long-term and distance dependency problems, making it better at processing time series data.

Research data comes from a die-casting machine with a clamping force of 2800 kN. Physical diagram of its structure is shown in Figure 6. Magnetic grating sensor installed on piston rod collects speed data corresponding to different opening percentages of fast cartridge valve of die-casting machine in fast stage when accumulator supplies oil. Maximum displacement of cartridge valve core is ymax, and valve opening percentage P takes a more representative value of 20%~52% (a value is taken every 2%). Pressure of accumulator when storing energy is 14 MPa. 30 sets of speed data are collected for each valve opening percentage, a total of 510 sets of data, and some of collected speed curves are shown in Figure 7.

Figure 6 Physical picture of injection machine structure

1. Punch 2. Piston rod 3. Accumulator 4. Magnetic grating displacement sensor 5. Injection cylinder

1. Punch 2. Piston rod 3. Accumulator 4. Magnetic grating displacement sensor 5. Injection cylinder

Figure 7 Injection velocity curve

Figure 8 sym8 wavelet scale function and wavelet function

High and low frequency information after wavelet decomposition is shown in Figure 9, where u2 is low frequency approximation coefficient of second layer, w1 and w2 are high frequency detail coefficients of the first and second layers respectively. High frequency information after decomposition is subjected to threshold denoising, results after denoising and reconstruction are shown in Figure 10. By comparing initial velocity signal with denoised velocity signal, it can be seen that after wavelet threshold denoising, noise caused by friction between injection piston and injection cylinder, punch and mold, pressure difference change at both ends of cartridge valve port can be effectively removed, and velocity curve becomes smoother. Denoised velocity signal can minimize interference to automatic feature extraction when it is input into model.

High and low frequency information after wavelet decomposition is shown in Figure 9, where u2 is low frequency approximation coefficient of second layer, w1 and w2 are high frequency detail coefficients of the first and second layers respectively. High frequency information after decomposition is subjected to threshold denoising, results after denoising and reconstruction are shown in Figure 10. By comparing initial velocity signal with denoised velocity signal, it can be seen that after wavelet threshold denoising, noise caused by friction between injection piston and injection cylinder, punch and mold, pressure difference change at both ends of cartridge valve port can be effectively removed, and velocity curve becomes smoother. Denoised velocity signal can minimize interference to automatic feature extraction when it is input into model.

Figure 9 Wavelet decomposition result

Figure 10 Data preprocessing result

Figure 11 Model prediction result

Figure 12 Speed signal

Speed/(m*s-1) | Actual valve opening/% | Injection system mechanism model | Theoretical formula model | ||

Opening/% | Error/% | Opening/% | Error/% | ||

1.7 | 20 | 22.27 | 11.35 | 18.49 | -7.55 |

2.1 | 24 | 25.67 | 6.96 | 20.62 | -14.08 |

2.6 | 28 | 29.62 | 5.79 | 23.29 | -16.82 |

3.0 | 32 | 32.40 | 1.25 | 25.42 | -20.56 |

3.6 | 36 | 35.98 | -0.05 | 28.62 | -20.50 |

4.3 | 40 | 39.45 | -1.38 | 32.35 | -19.13 |

4.9 | 44 | 41.98 | -4.59 | 35.55 | -19.20 |

5.6 | 48 | 44.55 | -7.19 | 39.28 | -18.17 |

6.1 | 52 | 46.18 | -11.19 | 41.94 | -20.21 |

Table 1 Comparison of prediction results

**Conclusion**

(1) For noise in injection speed curve, wavelet threshold denoising method is used to reduce noise in speed data. Mechanism model of die-casting machine injection system driven by time series data is established, which fully utilizes ability of LSTM neural network to process time series. Root mean square error RMSE of established model is 0.336%, correlation coefficient R2 is 0.998, and model accuracy is high.

(2) By comparing with prediction results of theoretical formula model, average error of theoretical formula model is -17.36%, and average error of injection system mechanism model is 0.11%, which shows that injection system mechanism model is better than theoretical formula model and can better predict valve core displacement corresponding to required injection speed.

(2) By comparing with prediction results of theoretical formula model, average error of theoretical formula model is -17.36%, and average error of injection system mechanism model is 0.11%, which shows that injection system mechanism model is better than theoretical formula model and can better predict valve core displacement corresponding to required injection speed.

#### Recommended

#### Related

- Macro-segregation of Fe and Si elements in cast aluminum alloy die-castings11-08
- Do you know what issues should be paid attention to when designing water transport for molds?11-08
- Dynamic Analysis of Packing Force of Die Castings11-07
- Plasticizing parameter setting and optimization method11-07
- Structural characteristics and formation mechanism of defect zones in die-cast AM60B magnesium alloy11-06