Abstract: Structure of connector shell is complicated. In order to explore influence of injection process on warpage of molded plastic part, a three-layer topology neural network model is established using orthogonal scheme as a sample. Take injection temperature, mold temperature, injection time, holding pressure and holding time as input layer neurons, and warpage as output layer neurons. Model is coded and trained by Matlab software. After multiple iterations of calculation, model error converges to target error. After comparing results of 5 sets of test samples, established model has a prediction error of 2.5% to 3.6%, and has good prediction performance. Trained model is used for actual production guidance, qualified plastic parts are obtained, which provides a reference for subsequent process improvement.

Electrical connectors are widely used in automotive and aerospace fields, plastic products are often used to manufacture electrical connector shells due to their advantages of light weight and low cost. In addition to good flame-retardant and insulating properties, connector housing also needs to have excellent dimensional stability to reduce effect of warpage and ensure that there is no error in connection and mating. In order to explore influencing factors of warpage, Li Haimei and others analyzed mechanism and influence of warpage in terms of comprehensive process parameters, mold structure and human factors; Gao Yuehua studied warpage influence factors of different parts for process, found that except for holding pressure, weights of other process parameters are different, indicating that relationship between process parameters and warpage is intricate, especially for plastic parts with complex structures. Based on existing CAE, traditional optimization regression method alone cannot accurately predict influence trend of parameters. Application of artificial neural networks has brought new ideas and methods to study of warpage effects. Neural networks are intelligent processing systems and are often used to solve complex multi-variable nonlinear problems. When Wang Weidong and Xiu Huiping studied injection process of molded plastic parts, they created a BP neural network model based on samples of orthogonal experiments and proved that it has good predictive ability. Deng Qigui used DPA-BP neural network model to analyze influence of process parameters on four injection defects, verified reliability of neural network model.
Injection process of connector shell is now studied, combined with orthogonal experiment and BP neural network model, nonlinear relationship between process parameters and warpage is analyzed, it provides ideas for subsequent parameter design of connector shell.

Connector housing structure is shown in Figure 1. Material is PA6, containing 30% glass fiber, brand is Ultramid B3GK24, density is 1.34 g/cm3. Its melt fluidity is good. Measured value of MFR (melt mass flow rate) is 81.5 g/min, with excellent dimensional stability.

 

Figure 1 Connector shell structure
It can be seen from Figure 1 that connector housing size is 145 mm * 120 mm * 20 mm, with a large number of corners, bosses and round holes, which are plastic parts with complex structures. In its injection molding process, internal melt flow process is difficult to predict, fiber orientation is very different, which easily leads to defects such as insufficient filling, warpage and high shrinkage.
For injection process of connector shell, firstly, according to MoldFlow software, optimal injection position of gate was analyzed, a single-gate injection scheme was designed, with gate size of φ1.5 mm. Mold adopts a 2-cavity structure. Simulation analysis is carried out according to default process parameters of system, runner and gate are set. Figure 2 shows structure design and filling analysis results of single gate solution.

 

Figure 2 Single gate injection scheme
It can be seen from Figure 2(b) that filling time of part of boss of plastic part to be formed is too late, which leads to gradual solidification of front melt and incomplete filling. In response to this problem, combined with software's gate filling analysis, double-gate and three-gate injection schemes were re-formulated, injection simulation of two schemes was also carried out by using 1-mold 2-cavity process. According to analysis results, maximum deformation of molded plastic part of dual-gate scheme is 0.779 6 mm, and maximum deformation of molded plastic part of three-gate scheme is 0.751 7 mm, as shown in Figure 3. Therefore, three-gate injection scheme can not only ensure complete filling of melt, but also warpage deformation of molded plastic part is small. Now based on three-gate scheme, continue to analyze process parameters.

 

Figure 3 Two-gate and three-gate scheme

First, establish finite element model of connector shell, perform simulation analysis in MoldFlow software. Due to large number of small chamfers and text marks in 3D model, it is difficult to divide mesh. CAD Doctor software is used to simplify and repair model, delete small chamfers and text, but it will not affect accuracy of analysis. Mesh matching rate has been increased from 75% to 93%. Double-layer mesh is selected. Number of meshes is 80 971. Material is Ultramid B3GK24 in database.
In connector shell injection process research, injection temperature A, mold temperature B, injection time C, holding pressure time D, and holding pressure E are selected as parameter analysis objects. Combining analysis of molding window and simulation results of default parameters, reasonable value range of each parameter is obtained, and factor level table is established based on this, as shown in Table 1.

 

Table 1 Factor level
According to Table 1, select L16(45) design table to establish an orthogonal experiment scheme, analyze influence weight of each parameter, and provide a reference sample for subsequent network model design. Experiment scheme and warpage results are shown in Table 2, a total of 16 sets of data sample.

 

Table 2 Test scheme and warpage results
Range analysis is shown in Table 3. For shell warpage, order of influence of injection parameters is E>A>B>D>C, that is, holding pressure has the largest influence weight, optimal combination is injection temperature of 300 ℃, mold temperature of 80 ℃, injection time of 1.0 s, holding time of 6 s, and holding pressure of 55 MPa.

 

Table 3 Range analysis

Artificial neural network is an optimized processing technology developed in recent years. It mainly refers to information processing method of biological brain. It creates many neural units to connect to each other to form a set of intelligent information processing system that can self-feedback. It is suitable for non-linear problems with complex interaction of reasoning factors. So far, a large number of improved models and optimization algorithms have been proposed one after another. Among them, the most common application is BP neural network, which has classic "Back Propagation" idea, that is, first forward, data information starts from input layer and is processed by intermediate layer to be passed to output layer. If there is a large error, reverse information transmission is carried out, connection weight between each layer is modified in process, and purpose of reducing error is achieved through this repetition.
In connector shell injection molding, relationship between process parameters and warpage has obvious nonlinearity. Combining nonlinear processing mechanism of BP neural network, it can more closely reflect relationship trend of various parameters and warpage. On the basis of orthogonal test samples, establishment of a three-layer topology neural network structure can better explain most of nonlinear problems. Too many layers may lead to overfitting. Five selected injection parameters are used as input layer neurons, and amount of warpage is used as output layer. Network structure and information transmission mode are shown in Figure 4.

 

Figure 4 Neural network structure

Relying on Matlab software coding to construct a network model and train established model, choice of training function will also affect training speed of model. In model construction, newff function is used to achieve back propagation. According to Kolmogorov theorem, 10 hidden layer processing elements are selected, and 70% of data set is extracted as training data. Learning rate is 0.02, maximum number of training times is 10,000, expected minimum deviation is 10-5. Trainlm function is used for training, sim is used to calculate simulation. In order to make model iterate more stably, mapminmax function is used to normalize and de-normalize data. After multiple iterations of calculation, training error gradually converges to target error.

After established neural network model is trained, its accuracy still needs to be tested to judge predictive ability and accuracy of model. Randomly select 5 untrained process plan combinations as samples, use MoldFlow software to re-simulate, compare prediction results of network model. Error between network prediction value of warpage and software simulation value is shown in Table 4.

 

Table 4 Comparison of network predicted value and simulated value of warpage
It can be seen from Table 4 that in verification test of 5 sets of samples, error between output result of neural network and simulation result is 2.5%~3.6%, results are basically consistent. Comparison of sample test results is shown in Figure 5, which can more intuitively observe error between network training value and actual simulation value. Synthesizing Table 4 and Figure 5 shows that model has high prediction accuracy.

 

Figure 5 Comparison of sample test results
Traditional mold test method is time-consuming and laborious. Using CAE finite element analysis can save a lot of time. However, every time injection parameters are changed in connector shell injection simulation, simulation analysis needs to be re-analyzed, and it takes 4 to 5 hours. Experimental design of simulated DOE takes 5 to 6 days. In order to further save time and speed up product development, trained neural network model is used to directly input corresponding process parameters, corresponding predicted warpage can be obtained. Due to high accuracy of network model, there is no need to repeat simulation analysis for each group of parameters, which reduces time cost.
It can be seen from above orthogonal test results that injection time C and holding pressure time D have little effect on warpage of connector shell, so optimal level can be kept unchanged, that is, injection time is 1.0 s and pressure holding time is 6 s. Secondly optimize influencing injection temperature A, mold temperature B and holding pressure D, fine-tune parameters based on orthogonal optimal level, then use established network model to predict warpage of adjusted parameter combination, parameter combination after fine-tuning and warpage prediction results are shown in Table 5.

 

Table 5 Fine-tuning parameter combinations and warpage prediction results
It can be seen from Table 5 that fifth group of parameters has the smallest warpage deformation, namely, injection temperature is 305 ℃, mold temperature is 85 ℃, injection time is 1.0 s, holding time is 6 s, and holding pressure is 60 MPa. At this time, warpage prediction value is 0.664 mm. Using MoldFlow software to simulate and verify this group of parameters, simulated deformation result is 0.657 mm, which is close to predicted value, and Z-direction deformation (flatness deviation) of shell is 0.48 mm. Using this parameter combination to guide sample production, equipment used is a 200T-LZ vertical injection molding machine, material is PA6 produced by BASF, and brand is Ultramid B3GK24. Trial-produced sample is shown in Figure 6. Sample is measured in three-dimensional dimensions. Flatness deviation of sample is 0.46 mm, which is close to simulation value. Maximum deviation of flatness required by drawing does not exceed 0.6 mm, and produced sample meets quality requirements.

 

Figure 6 Connector shell

Aiming at problem that connector shell structure is complex and injection process is difficult to predict, a neural network model with a three-layer topology is established to analyze and predict influence trend of process parameters on warpage. Orthogonal experiment was established, range analysis showed that holding pressure had the greatest impact on warpage. A BP neural network model was established on the basis of samples of orthogonal experiment, model was trained with Matlab software. Through 5 sets of sample tests, it is found that prediction error of neural network optimization model is between 2.5% and 3.6%, which proves that established network model has good predictive ability. Using trained model to continue second optimization of process parameters, warpage deformation amount corresponding to each parameter can be directly obtained, which saves software analysis time. Optimal combination of process parameters obtained was injection temperature of 305 ℃, mold temperature of 85 ℃, injection time of 1.0 s, holding time of 6 s, and holding pressure of 60 MPa. Using this parameter to guide production, samples with qualified quality were obtained. Therefore, model can be used to guide process parameter setting of connector housing, and provide application value for subsequent production.