Data-Driven Intelligent Monitoring of Die-Casting Machine Injection System

Abstract

The quality and productivity of die castings are directly influenced by the injection system performance of the die-casting machine, making advanced performance monitoring of paramount importance. However, with the present technology, it is impossible to discriminate between the hydraulic components that influence the operation of a pressured injection system due to their sheer number and complexity. On the other hand, it is challenging to pinpoint the pressured injection system while it is in the poor performance stage due to the complexity and variety of the working conditions in actual production as well as the lack of data. In this paper, the hydraulic principle of the pressure injection system is examined, and a simulation model of the pressure injection system is built by adjusting the values of various components and running simulation experiments to produce a sample set. The sample set is fed into an intelligent evaluation approach that combines BP neural networks, convolutional neural networks (CNN), and long short-term memory networks (LSTM). The above intelligent algorithm is used to obtain both the performance index of the pressurized injection system and the components that lead to the low-performance index. The Dempster-Shafer (DS) theory is used to perform information fusion on the component classification results, and a new neural network is designed to perform information fusion on the performance metric evaluation results. The combined results are the final classification and regression results. Later, simulation tests are used to compare and validate the method. The findings demonstrate that the proposed intelligent algorithm outperforms previous algorithms in terms of accuracy and stability. In terms of component classification, the average accuracy for BP-LSTM is 87.83%, CNN-LSTM is 90.63%, after stacking it is 93.31%, and the proposed method is 95.67%. For performance evaluation, the average R2 of BP-LSTM is 0.88 and the average MAE is 3.09; the average R2 of CNN-LSTM is 0.908 and the average MAE is 2.64; and the average R2 of the proposed method is 0.947 and the average MAE is 1.86.

1. Introduction

Die casting is an advanced molding technology that allows the production of a variety of geometrically complex castings with high production efficiency and parts with high unit strength. Due to this, it is widely utilized in many sectors to manufacture equipment and product parts [1]. Die-casting machine performance significantly affects the quality of die castings because it is the primary machine used in die-casting production [2,3,4]. As a result, press injection system performance monitoring has a significant positive impact on both productivity and product quality.

However, the traditional monitoring method by virtue of manual experience frequently falls short of good results due to the complex structure of the hydraulic system of the pressured injection mechanism, the substantial number of hydraulic components, and the difficulty of observing the performance data during the production process. Due to this, the performance of the manufactured die-castings may suffer significantly, seriously affecting the efficiency of the company. Additionally, without proper guidance, die-casting machine maintenance costs are high and have a long cycle, which has a negative impact on the business’s ability to produce goods efficiently and reap financial rewards. Therefore, a superior monitoring method is urgently required to ensure the effectiveness and efficiency of die-casting machine maintenance. As Industry 4.0 advances and informationization levels rise, more and more factories begin to switch from traditional production modes to intelligent production modes [5]. Artificial intelligence (AI) ideas like machine learning and neural networks are constantly being improved, and their use in a variety of industries has produced positive results [6,7]. These techniques are frequently employed in the field of industrial production for the testing of product quality and the problem diagnosis of machine parts [8,9], but rarely for the monitoring of machine performance. According to Lv et al. [10], a machine learning-based failure level prediction and maintenance choice model was proposed, and it produced positive results for rolling bearings, where the proposed notion of failure level prediction is extremely valuable.

Nevertheless, there is a shortage of data in the actual die-casting machine manufacturing process, making it difficult to develop an efficient monitoring model. Migration learning has a significant advantage in the field of fault diagnosis in terms of addressing the issue of data scarcity [11], but the majority of migration learning techniques still call for fault data from the target domain, which adversely affects the performance of some applications that are unable to supply the target fault data. Hou et al. [12] proposed a theory of transfer learning based on simulation data. FM et al. [13] trained the model with data from hydraulic valve simulation and achieved great accuracy in valve fault identification. Li et al. [14] developed a simulation model based on AMESim and used support vector machines (SVM) to determine the fault and diagnose hydraulic system faults in combined harvesters, and operational efficiency was significantly increased.

Additionally, while assuring the correctness of the results, the size of the data collection should be taken into consideration due to the timeliness of the monitoring results and the expense of establishing the supporting equipment. FM et al. [15] developed a way of choosing data features using decision tree classifiers and made use of the benefits of decision tree white-box modeling, which offers some advice for post-maintenance.

Ji et al. [16] suggested a defect diagnostic approach based on DS theory to address the issue of unstable results from a single algorithm and to increase the stability of the results. In conjunction with the die-casting machine’s real manufacturing procedure, this research suggests an intelligent monitoring program, as depicted in Figure 1. It permits the classification of the problematic components that cause poor performance as well as the performance evaluation of the die-casting machine.

The main job descriptions are as follows:

(1)

The simulation model is established using the hydraulics of the pressure injection system of the die-casting machine to address the issue of insufficient actual data. In order to gather sample data under various performance conditions for the pressure injection system, the model parameters are adjusted in accordance with potential performance degradation.

(2)

A decision tree classifier is used to select the features generated from the original time series, while the front layer of LSTM is utilized for feature extraction from the time series. Furthermore, a monitoring model based on BP-LSTM and CNN-LSTM is established to accomplish component diagnosis and performance evaluation.

(3)

Ultimately, DS theory is used to combine the component type results from the two algorithms, a new neural network is used to combine the performance assessment results, and the combined classification and regression results are used as the final results.

The rest of this paper is structured as follows: The theoretical method used is detailed in Section 2. In Section 3, the proposed intelligent monitoring method is elaborated. In Section 4, the proposed method is validated experimentally, and the results are discussed. In Section 5, the research work of this paper is summarized.

2. Theoretical Background

2.1. Single Intelligent Monitoring Approach

In order to enhance the complementarity between intelligent algorithms, BP, CNN, and LSTM are selected in this paper.

One of the most significant algorithms for neural networks is the BP neural network, which has a wide range of potential applications in the field of intelligent manufacturing [17]. It has strong nonlinear mapping capabilities and works well for dealing with classification and regression problems. The fundamental idea behind the BP neural network is based on the gradient descent approach, which updates the weights and thresholds in the direction where the error function lowers the quickest, hence reducing the error between the output value of the neural network and the target value. It has been characterized by the forward transmission of the signal and the backward propagation of the error. It benefits from being both powerful and less computationally demanding. Its composition structure is made up of the input, hidden, and output layers. This structure using a neural network with just one hidden layer is shown in Figure 2.

According to Figure 2, the input layer is represented by x_i, the hidden layer by h_m, the output layer by y_n, the activation function by f, the weight connecting the input layer to the hidden layer by w_mi, and the output layer to the hidden layer by w_nm.

The hidden layer is calculated as follows:

$$ne{t}_m={\displaystyle \sum _{m=1}^M{w}_{mi}{x}_i}$$

(1)

$$h_m=f(ne{t}_m+b_m),m=1,2,\cdots ,M$$

(2)

where b_m is the threshold (neuron bias)

The output layer is calculated as follows:

$$ne{t}_n={\displaystyle \sum _{n=1}^N{w}_{nm}{h}_m}$$

(3)

$$y_n=f(ne{t}_n),n=1,2,\cdots ,N$$

(4)

The output error E is the difference between the output y_n of the neural network and the target value Y_n, which is calculated as follows:

$$E=\frac{1}{2}{\displaystyle \sum _{n=1}^N}{(Y_n-y_n)}^2$$

(5)

CNN has produced promising results in fields including image processing and audio processing, where they were initially applied [18]. CNN has been extensively applied in the field of fault diagnosis as a result of the advancement of deep learning theory and the quick improvement in hardware performance in recent years [19]. CNN is mainly made up of an input layer, a hidden layer, and an output layer, and the hidden layer contains numerous convolutional and pooling layers. By using a group of convolutional kernels, the convolutional layer extracts features from the data, and the pooling layer reduces the feature size while keeping the important ones.

Hochreiter et al. [20] proposed the LSTM network in 1997 to address the issue of gradient vanishing in RNN when handling long-time series problems. The LSTM network has the advantage of being able to capture long-term relationships and learn temporal features, which makes it an improvement over RNN. The LSTM network neuron structure is shown in Figure 3, and its structure contains three control gates to control the state of the unit. Information input to the memory cell is controlled by the input gate i_t, information retention about the historical state of the previous memory cell is controlled by the forget gate f_t, and information output from the memory cell is controlled by the output gate o_t.

The computational flow of LSTM is shown as follows:

$$\{\begin{array}{l}{f}_t=\sigma (W_f\left[h_{t-1},x_t\right]+b_f)\\ i_t=\sigma (W_i\left[h_{t-1},x_t\right]+b_i)\\ \tilde{C}=\tanh (W_C\left[h_{t-1},x_t\right]+b_C)\\ C_t=f_t\odot C_{t-1}+i_t\odot {\tilde{C}}_t\\ o_t=\sigma (W_o\left[h_{t-1},x_t\right]+b_o)\\ h_t=o_t\odot \tanh (C_t)\end{array}$$

(6)

where: h_t−1 is the output of the hidden layer of the previous unit; σ is the sigmoid activation function; W_f, W_i, W_C, and W_o are the weight matrices; b_f, b_i, b_C, and b_o are bias terms; tanh is the hyperbolic tangent activation function; $\odot$ is Hadamard Product.

2.2. Methods for Feature Extraction

With less training data and better process visibility, decision tree algorithms are popular in the fault diagnostic industry [21]. The procedure starts with the recursive selection of the best feature, and then the data is divided according to that feature so that each sub-dataset can be classified to the highest degree. Making disorganized data more orderly is the general idea behind splitting up a data set, while Shannon entropy, or just entropy, is a measurement of how much information is ordered. The information entropy of a system decreases with increased order. On the contrary, information entropy rises with system disorder. In this paper, the Gini coefficient (G), which has the same properties as the entropy property, is used to classify the optimal features. The formula is as follows:

$$G=1-{{\displaystyle \sum _{i=1}^n(\frac{\left|D_i\right|}{\left|D\right|})}}^2$$

(7)

where: D is the total number of samples, and D_i is the number of samples of a particular class after division by selected characteristics.

2.3. Methods of Information Fusion

In order to combine the information from the classification results produced by several intelligence algorithms, this study is based on the Dempster-Shafer theory. As an extension of Bayesian theory in probability theory and a popular method for handling uncertain data, DS theory is also known as the theory of confidence functions [22]. The framework of the integration problem is defined as Θ = {θ₁,θ₂,···θ_n}, θ₁, θ₂,···θ_n, which represents a set of elementary propositions that are mutually exclusive and can form a complete set. The set consisting of all subsets of Θ is called the power set, denoted 2^Θ. Basic Probability Assignment (BPA) is the process of figuring out what the fundamental likelihood is for each piece of evidence in the identification framework Θ. The mass function (m function) is used to carry out this operation. The m function reflects the magnitude of propositional confidence, denoted m(x), satisfying the following equation:

$$\{\begin{array}{l}m(\varnothing )=0\\ \\ {\displaystyle \sum _{A\in 2^{\mathsf{\Theta }}}m(A)}=1\end{array}$$

(8)

The Dempster synthesis rule, which is described as follows, can be used to combine the independent BPA functions m₁ and m₂.

$$m(A)=\left[m_1\oplus m_2\right](A)$$

(9)

$$m(A)=\{\begin{array}{ll}0,& A=\varnothing \\ \frac{{\displaystyle \sum _{\begin{array}{l}X,Y\in 2^{\mathsf{\Theta }}\\ X\cap Y=A\end{array}}{m}_1(X)m_2(Y)}}{1-{\displaystyle \sum _{\begin{array}{l}X,Y\in 2^{\mathsf{\Theta }}\\ X\cap Y=\varnothing \end{array}}{m}_1(X)m_2(Y)}},& A\ne \varnothing \end{array}$$

(10)

Stacking is widely used in various machine learning competitions and has achieved exciting results in integrating different models. So in this paper, the stacking algorithm is used to contrast with the DS theory. The individual learners integrated by the stacking algorithm are referred to as the base model, and the learners used for combination are referred to as the meta-model. The output of each base model is used as input to the meta-model for secondary learning. Typically, the stacking algorithm is often accompanied by cross-validation operations in order to avoid the risk of overfitting, the structure of which is shown in Figure 4.

According to Figure 4, the results of the per-fold validation set of all base models are stacked as the training set of the meta-model; the results of the test set of all base models are averaged and stacked as the test set of the meta-model.

In this paper, the evaluation of performance using different algorithms is a regression-type problem. Thus, the obtained results are not used for information fusion utilizing DS theory since its output structure differs from that of the classification issue; that is, there is no set of fundamental propositions that are mutually exclusive and may form a full set. The generated regression results of various intelligence algorithms are utilized as inputs to a neural network, which then yields the final regression results since neural networks are very fault-tolerant and self-learning.

3. Intelligent Monitoring Methods

The specific approach of the method suggested in this research is to simultaneously evaluate the performance and categorize the problematic components of the pressure injection system using various intelligence algorithms, and the results acquired are defined as initial results. In terms of the classification of components, each algorithm may have different classification results for the problem components, a situation that is known as the information source conflict problem in this study. In other words, it is similar to how different human experts with different experiences perceive different components causing low-performance situations. The classification results are determined in the end using the DS theory since it is effective at resolving conflict or uncertainty problems that are frequently encountered in the information outcome process. This study defines index parameters for performance evaluation, and it applies neural networks to combine the data of the evaluation parameters produced from various intelligence methods.

3.1. Feature Generation and Selection

Since the original data are all-time series, the following 12 sets of time-domain features with low computational cost are chosen by human-defined procedures, and the computational formulas are presented in

Table 1. Time-domain characteristic formulas.

Features	Formulas	Features	Formulas
Mean	$\overline{x}=\frac{1}{n}{\displaystyle \sum _{i=1}^n{x}_i}$	Peak-peak	$Vpp=x_{\max }-x_{\min }$
Standard deviation	$\sigma =\sqrt{\frac{1}{n}{\displaystyle \sum _{i=1}^n{(x_i-\overline{x})}^2}}$	Energy	$m={\displaystyle \sum _{i=1}^n{x}_i^2}$
Skewness	$S_K=\frac{1}{n}{\displaystyle {\sum }_{i=1}^n\frac{{(x_i-\overline{x})}^3}{{\sigma }^3}}$	Root mean square	$M=\sqrt{\frac{m}{n}}$
Kurtosis	$K_{}=\frac{1}{n}{\displaystyle {\sum }_{i=1}^n\frac{{(x_i-\overline{x})}^4}{{\sigma }^4}}$	Absolute mean	$p=\frac{1}{n}{\displaystyle \sum _{i=1}^n\left|x_i\right|}$
Maximum	$x_{\max }$	Minimum	$x_{\min }$
Waveform factor	$F=M/p$	Amplitude factor	$c=x_{\max }/M$

. The amount of processing required for the classification process has been reduced.

Decision trees are used to feature extract the various characteristics from the data collected by various sensors. On one hand, it may eliminate unnecessary and pointless features. On the other hand, as it is a white-box model, it can boost the credibility of the component categorization algorithm, which can aid the staff in the ensuing equipment maintenance.

3.2. Baseline

Lv et al. [10] proposed a parallel model for simultaneous fault diagnosis and fault severity prediction, as shown in Figure 5.

In this method, the front layers of LSTM extract the time series features. They are combined with the extracted time-frequency features as the input of the BP neural network, which are used as a supplement to the fault type features to further improve the accuracy of fault type prediction. At the same time, the extracted time series features are input to the subsequent LSTM network to predict the fault severity. Thus, the fault type and the fault severity are simultaneously predicted. The model achieves a diagnostic accuracy of 99.74% for bearing fault types and an R2 value of 0.93 for fault severity prediction under 4910 samples. Considering that there is very little data available in the actual production, this will greatly reduce the stability and accuracy of the results. Therefore, in this paper, based on this model, the results are fused with information to solve the problem.

3.3. Intelligent Monitoring Model

The parallel monitoring model and the resulting information fusion are two components of the intelligent monitoring model suggested in this paper. A simultaneous monitoring model based on the BP model, CNN model, and LSTM model is created in order to complete the classification and regression tasks concurrently. The resulting information fusion is carried out by neural networks and DS theory, respectively, and the proposed method’s structure is depicted in Figure 6.

The specific procedure of the intelligent algorithm is made up of the following steps:

Step 1: Time-series features are extracted from the original data by the front layer of LSTM, and these features are combined with human-defined time-domain features that have been filtered by a decision tree to serve as inputs to the CNN and BPNN, further enhancing the component classification accuracy. At the same time, the extracted time series features are fed into the subsequent LSTM network for further performance evaluation.

Step 2: The CNN and BPNN in the component classification portion learn the probabilities of low performance for each component by spontaneous learning, and each probability is treated as an independent BPA function, m_CNN(X_i) and m_BP(X_i). The BPA functions are then fused to get the final categorization results using the Dempster synthesis rule.

Step 3: The performance indicators are evaluated by the CNN and BPNN in the performance evaluation section using spontaneous learning. This research offers an intelligent approach, that is, the neural network method, to carry out the information fusion operation in comparison to fusing them using the weighted average [23] and the Kalman filter [24]. The two assessment outcomes are fed into a fully connected neural network, which combines them using its ability to learn by itself to produce the final assessment results.

4. Experimental Verification

A simulation model based on AMESim is used to experimentally test the proposed intelligent algorithm due to the dearth of data in real production. The suggested approach is implemented on MATALB R2022b using a computer with the device configuration of an Intel Core i7-13700K processor, an NVIDIA GeForce RTX 3070 graphics processor, and 32 GB of RAM.

4.1. Pressure Injection System Modeling

The simulation model of the pressure injection mechanism is constructed using AMESim and is depicted in Figure 7 in accordance with the operation of the hydraulic system of the pressure injection mechanism and the real component parameters.

The specific action process is as follows:

Slow phase: The solenoids of valves 420 and 414a are both turned on at the same time. Oil from the pressure injection accumulator enters the rodless chamber of the pressure injection cylinder through valve 420, flows out of the rodded chamber through valve 414a, and the cylinder slowly moves forward. The speed of the slow phase is controlled by valves v414a and v420. In this experimental model, the duration of this phase is 3 s, and the preset velocity is 0.5 m/s.

Fast phase: When the displacement sensor on the pressure injection cylinder determines that the cylinder has reached the position of rapid pressure injection, the solenoids of valves v414 and v419 are simultaneously turned on. At this time, oil from the pressure injection accumulator primarily enters the rodless cavity of the cylinder through valve v419 and exits the rodded cavity of the cylinder through valve v414. The speed of the fast phase is controlled by valves v419 and valve v414. In this experimental model, the duration of this phase is 0.14 s, and the preset velocity is 4 m/s.

Pressurization phase: The solenoid of valve v428 is activated as the punch quickly pushes the liquid metal to fill the cavity. Through valve v428, the oil in the pressurizing accumulator flows into the right chamber of the pressurizing cylinder, causing the piston of the pressurizing cylinder to move to the left, and valve v428 completes the control of the pressurizing phase. In this experimental model, the duration of this phase is 1.86 s, and the pressurized accumulator is preset at 160 bar.

This study chooses the five performance indicator groups of fast speed accuracy, slow speed accuracy, acceleration performance, braking performance, and pressure accuracy for evaluation in accordance with the real production needs. The performance index is defined as q due to the various production constraints. For instance, the acceleration performance index q_a is defined as follows:

$$q_a=100-100\cdot k\frac{{\left|a_i-a_c\right|}_{\max }}{a_c}$$

(11)

where a_c is the desired target acceleration, a_i is the measured actual acceleration, and k is a scaling factor to meet different production requirements. The maximum value of q is 100, which represents the ideal situation in which the actual performance and the desired performance are exactly the same. The lower the value of q, the poorer the real performance turns out to be.

The performance deterioration of valves v414, v414a, v419, v420, and v428 is the element impacting the performance index, according to workflow and maintenance experience in actual production; therefore, these five components are selected as the aim of component classification.

4.2. Creation of the Sample Set

The following eight sets of signal data for the chosen stages are derived by AMESim simulation based on the signals that can be recorded in the actual pressure injection mechanism, as illustrated in Figure 8.

The displacement signal of the pressure injection cylinder, the pressure signal of the rodless chamber and the rodded chamber, and the current signals from the valves v414, v414a, v420, v419, and v428 are the signals used, respectively. The above three processes take a total of 5 s, and the signal is sampled at a frequency of 1000 Hz with a time interval of 1 ms.

The speed of the simulation model is adjusted to 4 m/s in the rapid phase and 0.5 m/s in the slow phase, and Figure 9 depicts the velocity curve of the pressure injection cylinder.

In this paper, the total number of samples obtained through simulation is 480 (80 × 6) based on the division of low-performance component types, of which the number of samples in the training set is 384 (64 × 6) and the number of samples in the test set is 96 (16 × 6).

Based on the factory production requirements, the threshold value for component classification is calculated by Equation (11) as shown in

Table 2. Thresholds of component categorization.

Performance Indicator	Threshold
Fast speed accuracy	81.25
Slow speed accuracy	76
Pressure accuracy	79.3
Accelerated performance	70.6
Braking performance	72.5

, and if the performance index is lower than this value, it is recognized as the presence of low-performance components.

4.3. Training of the Model and Analysis of Results

The fewer samples used, taking into consideration computation costs and calculation times, the better due to actual production needs. Therefore, the selection of sample size is made by using various sample sizes for testing and training. The average accuracy for classifying low-performance components after ten 5-fold cross-validation passes is shown in

Table 3. Classification accuracy of different algorithms with different number of samples.

Sample Size	Accuracy (%)
	BP-LSTM	CNN-LSTM	The Proposed Approach	Stacking
10	70.16	77.87	84.63	80.29
20	76.75	84.17	90.04	86.93
30	84.44	87.78	92.67	90.16
40	87.83	90.63	95.67	93.31
50	90.07	92.58	96.17	95.14
60	92.31	93.88	96.97	96.92
70	93.57	94.93	97.65	98.21
80	94.53	95.87	98.26	99.13

.

The R2 coefficient of determination is used to examine the effect of performance evaluation. It measures how well the anticipated value matches the observed value and is determined as follows:

$$R2=1-\frac{{\displaystyle {\sum }_{i=1}^n{\left({\widehat{r}}_i-r_i\right)}^2}}{{\displaystyle {\sum }_{i=1}^n{\left({\overline{r}}_{}-r_i\right)}^2}}$$

(12)

where n is the number of samples, ${\widehat{r}}_i$ is the performance evaluation index of the ith sample, r_i is the actual performance index of the ith sample, and $\overline{r}$ is the average performance index of the actual sample. The closer the value of R2 is to 1, the closer the predicted value is to the actual value. The average R2 values of the performance metrics determined after 10 5-fold cross-validations are presented in

Table 4. R2 of performance evaluation of different algorithms with different number of samples.

Sample Size	R2 Score
	BP-LSTM	CNN-LSTM	The Proposed Approach
10	0.433	0.601	0.703
20	0.754	0.81	0.896
30	0.834	0.886	0.936
40	0.88	0.908	0.947
50	0.909	0.922	0.956
60	0.914	0.936	0.965
70	0.926	0.944	0.972
80	0.935	0.951	0.978

.

The line graphs associated with the above two tables are shown in Figure 10 and Figure 11.

From the above tables and figures, it can be seen that as the number of samples increases, the classification and performance evaluation of all the intelligent algorithms are getting better. The method proposed in this paper has a much higher accuracy and R2 value than the other two single methods when the sample size is small. Furthermore, as the number of samples increases, it can quickly reach a high accuracy and R2 value. Compared with stacking ensemble methods, DS methods have higher accuracy when the sample size is small. However, as the sample size increases, the accuracy of stacking will be higher than that of the DS method. This has been analyzed as a result of the poor performance of the base model with a small sample size and the small number of base models used in this paper.

The proposed method achieved a classification accuracy of 95.67% under a sample size of 40 × 6, which is comparable to that of CNN-LSTM with a sample size of 80 × 6, which achieved 95.87% accuracy. Similarly, the R2 value of the proposed method under the sample size of 40 × 6 is 0.947, which is also comparable to that of CNN-LSTM under the sample size of 80 × 6, which achieved a value of 0.951. The average MAE values for performance evaluation at the 40 × 6 sample size are shown in Figure 12.

As seen above, the method proposed in this paper has produced excellent results in component classification and performance evaluation when the number of samples is more than 40 × 6. The performance of BP-LSTM and CNN-LSTM improves significantly as the number of samples rises, and while the proposed method still performs better than the other two methods, the performance gap between them is gradually closing. This paper employs a sample set of 40 × 6 for additional testing, taking into account the practical use and computing cost.

In order to test the stability of the proposed model, a one-time 5-fold cross-validation of the three methods mentioned above is performed. The accuracy and R2 values obtained for each of the low-performance component classifications and performance evaluations are shown in

Table 5. Per-fold classification accuracy of different algorithms.

Fold	Accuracy(%)
	BP-LSTM	CNN-LSTM	The Proposed Approach
1	89.58	93.75	95.83
2	87.5	91.67	95.83
3	89.58	89.58	95.83
4	83.33	89.58	93.57
5	85.42	91.67	95.83

and

Table 6. Per-fold R2 scores of different algorithms.

Fold	R2 Score
	BP-LSTM	CNN-LSTM	The Proposed Approach
1	0.83	0.92	0.93
2	0.89	0.91	0.94
3	0.84	0.89	0.94
4	0.92	0.88	0.94
5	0.91	0.92	0.95

, and the corresponding line graphs are shown in Figure 13 and Figure 14.

The two figures above show that the performance of the two approaches without information fusion fluctuates, and the suggested method beats the other two ways in terms of stability in addition to having a better accuracy and R2 value.

For the classification accuracy of a specific class of components, the confusion matrices of the four methods are shown in Figure 15, Figure 16, Figure 17 and Figure 18.

The confusion matrix demonstrates that a single intelligent algorithm has significant accuracy swings when categorizing a particular class of components, although it has a respectable average accuracy. The following tags provide examples: Tag 2 (Valve v420), Tag 5 (Valve v428), and Tag 6 (Normal Performance). The method proposed in this study effectively resolves the issue of conflicting information sources through the information fusion of algorithm results, not only in the classification of tags 2 and 6 with 100% accuracy but also in tag 5 with increased accuracy. In addition, it can be seen that the single algorithm is not considered ideal for the classification of tag 5, and the stack algorithm does not get a good performance boost for the classification of tag 5 because it is affected by the performance of the base model. As a result, the method not only has high accuracy but also good stability in categorizing a specific class of components.

As regards the degree of fitting for the performance metrics of specific samples, take the acceleration performance metrics as an example. Regarding the degree of fit for a particular sample of performance metrics, the result of comparing the performance evaluation metrics with the actual performance metrics, using acceleration performance metrics as an example, is shown in Figure 19.

The graph shows that the method suggested in this study on some samples, where the single intelligent algorithm cannot achieve good evaluation results, through the neural network to the outcomes of the information fusion and therefore get comparatively good evaluation results.

5. Conclusions

This research proposes an information fusion-based intelligent monitoring technique for the injection system of die-casting machines that is capable of concurrent component categorization and performance index evaluation and is validated using a simulation model. The method has excellent classification accuracy and performance, even with small samples.

In terms of component classification, the method outperforms a single intelligent algorithm not only in terms of average component classification accuracy but also in terms of accuracy and stability in one classification. As for the classification of a specific component, the method can solve the problem of conflicting information sources, which makes the results more accurate and stable.

The method not only has a higher average R2 value when evaluating performance measures but also a higher R2 value and stability within a single performance evaluation. For a specific sample, the results of the method also have a higher degree of similarity, and the method for some single algorithm that cannot achieve a good fit on the sample has a relatively good effect.

Finally, when the number of samples for a single class of components is 20, 40, 60, and 80, the average classification accuracies are 90.04%, 95.67%, 96.97%, and 98.26%, respectively, and the average R2 values for performance evaluation are 0.896, 0.947, 0.965, and 0.978. The results in the sample size of 40 conditions have met the actual engineering needs to provide support and guidance for the maintenance of die-casting machines.

The method should be applied in practice with attention to the following points:

• Because the simulation data are partially different from the actual data, the number of samples should be flexibly selected in the actual use so that the model can achieve satisfactory results.
• The actual data used for training should cover all stages of the problem components, and the data of the components at a certain stage should not be selected only for training.
• The categorization of components in this paper is that of a single component, and it remains to be investigated in cases where multiple problem components appear to be coupled. The number of tags for multiple problematic components increases exponentially with the number of problematic components, and in practice, it is recommended that a single component stack be used to resolve the coupling situation.