Research on Intelligent Computing Methods for Multistep Process Capability

In the context of the deep integration of artificial intelligence and industrial fields, precise calculation of multi-step process capability has become a current research hotspot. In most industrial fields, the estimation of multi-step process capability is mostly in the rough estimation stage, and there are relatively few precise quantitative calculation methods. This paper focuses on multi-step processes in the industrial field, using deep learning models to learn the features of each step step step by step, and then comprehensively estimating the weights between each step, ultimately achieving accurate prediction of multi-step process capabilities. This paper conducts in-depth analysis of the performance and efficiency of different models on such problems by designing a large number of validation experiments, and also provides ideas and suggestions for subsequent research in this field.


Introduction
Multi step process capability calculation refers to a class of comprehensive multi-step problems, which are usually executed in sequence by multiple steps.The learning process of each step is different, and the process capability of each step will affect the overall capability of the multi-step problem.Therefore, in the process of solving comprehensive multi-step problems, it is not only necessary to calculate the capability value of each step, but also to learn the impact of different steps on the overall capability value, and achieving comprehensive learning of local and global features.In practical applications, multi-step problems are prevalent in various industries, especially in industrial processes.For example, in the petrochemical industry, the oil or natural gas extraction process mainly includes the following processes: exploration and drilling, pipeline construction, oil and gas production, and oil and gas transmission.When calculating the energy efficiency of a certain oil and gas development platform, it is necessary to comprehensively consider the capabilities of these four processes and the different technical means in these four processes; In the field of automobile manufacturing, a car needs to go through four workshop processes: stamping, welding, painting, and final assembly.The ability of different workshop processes and even a certain technical detail ability can affect the quality level of the entire vehicle; In the field of network security, network security investigation generally includes preliminary research, security scanning and vulnerability investigation, security configuration investigation, traffic monitoring and other steps.To calculate the network security capability of a platform, the influencing factors and capabilities of the above steps need to be comprehensively considered.It can be seen that the research on multi-step process capability calculation will have strong research needs and large application space in various industries and fields in the future.
Multi step process capability calculation is not simply adding and solving the capabilities of each step, because in a comprehensive multi-step problem, the impact of different step capabilities on the overall capability is different.For example, the capability formed in a key step is weak or 0, while the capability values formed in other steps are high, but the overall capability value is still weak or even 0.So in multi-step problems, it is not only necessary to consider the influencing factors of each step, but also to learn the weights of each step's impact on the overall ability and the mutual influence between different steps.The difficulty of this study is how to solve the ability calculation function for each factor among multiple non independently distributed influencing factors.The rapid development of artificial intelligence technology has brought certain ideas for solving such problems, such as deep learning models such as EM algorithms [1,2], Attention mechanisms [3,4], etc.This article explores the application of deep learning models in multi-step process capability calculation.Based on the simplest fully connected neural network design, the modified model is applied to experimental data for verification and analysis, proposing feasible methods for intelligent calculation of multi-step process capability.

Related Work
The research on multi-step process capability calculation based on artificial intelligence methods mainly focuses on deep learning models, but existing algorithm models focus on predicting or analyzing a single problem, and there is not much research on capability estimation algorithms for multi-step problems.Reference [5] proposed using deep neural networks to predict aviation ammunition consumption, and the DNN model showed significant performance advantages in experiments compared to traditional algorithms.Reference [6] studied the processing of sensor data using a single ultrasonic sensor to classify different object categories and the traversability of obstacles, and proposed a scale graph based signal processing chain and convolutional neural network, which outperformed methods such as LeNet-5.Reference [7] proposed the first end-to-end CCI learning method based on convolutional neural networks, forming a string representation of chemical structures.Reference [8] proposes the use of Fourier transform for trend analysis and long short-term memory neural network prediction of financial time series, and selects three typical stock market indices in the real world and their closing prices within 30 trading days to test their performance and prediction accuracy.Reference [9] used feature frames of different sizes to input into the LSTM network for efficient transportation mode detection, achieving a classification accuracy of up to 98%.Reference [10] studied two stochastic models, Conditional Constrained Boltzmann Machine (CRBM) and Factor Conditional Constrained Boltzmann Machine (FCRBM), based on deep neural network models for energy consumption time series prediction.Reference [11] proposes using a supervised convolutional neural network (CNN) to improve the performance of retinal vessel segmentation.This model has a series of upsampling and downsampling processes for feature extraction.Reference [12] proposes a hybrid deep learning framework based on quantum computing for power system fault diagnosis, which combines the feature extraction ability of conditional Boltzmann machines with the effective classification of deep networks.The training method based on quantum computing effectively utilizes the complementary advantages of quantum assisted learning and classical training techniques, overcoming the computational challenges brought by the complexity of this deep learning model.

The Statistics of Dataset
This paper studies the capability estimation of multi-step processes, which requires learning the features of each step to ultimately achieve comprehensive capability estimation of multiple steps.To validate and analyze the model studied in this article, a large amount of experimental data is required.Due to limited research in this field and a lack of comprehensive data collection tools in industrial processes, research in this field lacks open and universal experimental data.Therefore, this article can only simulate the data of multi-step processes, forming a batch of simulation data as the experimental dataset for this article.It should be noted that although simulation data was used in this article, the simulation process conforms to the practical logic of industrial processes, and random interference data is added on this basis.This processing makes the experimental data as consistent as possible with the real process in the industrial field, but also increases the learning difficulty of the model, ensuring that the model can produce the same results and performance on real data.To provide an objective explanation of the experimental data in this article, Table 1 presents a detailed statistical analysis of the dataset.This paper simulates data from 2000 multi-step processes.On the premise of not affecting the essence of the research problem, in order to simplify the calculation, this article sets the number of steps in the multi-step process problem to 2. It is equivalent to dividing a comprehensive problem into two steps, where different operational processes lead to different dimensional process characteristics in the two steps.In this dataset, set the feature dimension of the first step process to 10 and the feature dimension of the second step process to 6.Each step of the operation process has a competency value, and after completing two steps, there is a comprehensive competency value.In this paper, it is not yet possible to accurately estimate the ability value, and only the interval of the ability value can be estimated.Therefore, this article divides the ability into intervals.The specific method is to correspond the ability values within {[0,0.2),[0.2,0.4),[0.4,0.6), [0.6,0.8), and [0.8,1]} to {0.2, 0.4, 0.6, 0.8, 1}, respectively.

The Prediction Models
The research approach of this article is to learn the characteristics of each step and then comprehensively estimate the process capability of multiple steps, essentially solving multi classification problems.The calculation process in this article can be expressed in the following form: Where e and u represent the features of the two steps, W represents the parameter matrix, F represents the objective function, and A represents the capability value of the multi-step process.According to the above ideas, the objective function F can be expressed in the following form: This paper is based on problem solving logic and deep neural network models, and has modified the neural network structure to form the following four neural network models. DNN-3: In this article, DNN-3 is designed as a 3-layer fully connected neural network structure, with 16, 8, and 1 neurons in the input layer, hidden layer, and output layer, respectively.The activation function for each layer is selected as the "sigmoid".To prevent overfitting, batch normalization and random freezing of neurons were added to each layer. DNN-3-r: In this article, DNN-3-r is designed as a 3-layer fully connected neural network structure, with 16, 8, and 1 neurons in the input layer, hidden layer, and output layer, respectively.The activation function of each layer is selected as the "relu".To prevent overfitting, batch normalization and random freezing of neurons were added to each layer. DNN-4: In this article, DNN-4 is designed as a 4-layer fully connected neural network structure, with 16, 8, 4, and 1 neurons in the input layer, hidden layer, and output layer, respectively.The activation function for each layer is selected as the "sigmoid".To prevent overfitting, batch normalization and random freezing of neurons were added to each layer. DNN-4-r: In this article, DNN-4-r is designed as a 4-layer fully connected neural network structure, with 16, 8, 4, and 1 neurons in the input layer, hidden layer, and output layer, respectively.The activation function of each layer is selected as the "relu".It should be noted that in order to prevent overfitting, batch normalization and random freezing of neurons have been added to each layer.

Experiments
The essence of the problem studied in this article is a multi classification prediction.To verify the estimation performance of different models for multi-step process capability, accuracy is selected as the evaluation index.This paper randomly sorts 2000 data points in the dataset and divides them into training and testing sets.50, 100, 200, 500, 1000, and 1500 pieces of data were randomly selected from the dataset as training data, and the remaining data was used as testing data.
Before the experiment, this paper will normalize the features of each step, using the feature data as input data for the deep model, and the interval category of the ability value is labeled data.The experimental comparison results of four neural network models on the same dataset are shown in

Conclusion
This article investigates the ability estimation of multi-step processes, designs four deep learning models based on the simplest fully connected neural network, and verifies the performance of different models on experimental data.Although fully connected neural networks have relatively weaker feature extraction capabilities compared to complex deep network models such as convolutional neural networks and recurrent neural networks, due to their relatively fewer network layers and simpler structure, they have relatively good interpretability in intelligent computing of multi-step process capabilities in the industrial field, achieving a good balance between algorithm performance and reliability.In future research, more attention will be paid to real data collection and mining in different industrial fields, and artificial intelligence technology will be applied in more industrial scenarios.

Figure 1 .
According to Figure1, it can be seen that the four deep neural network models showed the best performance in multi-step process capability estimation, with a prediction accuracy of 0.1945.DNN-3-r has the worst prediction performance, with a 0.1935 lower prediction accuracy compared to DNN-4.Based on the above experiments, this article also found that the four layer neural network model performs better than the three layer neural network in this problem.Meanwhile, in this experiment, the performance difference between the deep neural network model with "relu" activation function and the neural network model with "sigmoid" activation function is not significant.In order to explore the impact of different training data on model performance, this paper will set the training data items to {50, 100, 200, 500, 1000, 1500} respectively.The experimental results of the four neural network models are shown in Figure2.The neural network model designed in this article is the simplest backpropagation fully connected neural network structure.The training of this deep neural network model is a supervised learning process, and as the training data increases, the model learns more prior knowledge, thereby promoting the enhancement of the model's learning ability.However, according to Figure2, as the training data increases, the predictive ability of the neural network model for multi-step processes in this paper also deteriorates.This is because the model structure in this article is simple, and as the training data increases, overfitting occurs to a certain extent, leading to a decrease in predictive performance.To verify the stability of the performance of neural network models in multi-step process capability estimation, this paper statistically analyzed the changes in loss values of four neural network models as shown in Figure3.According to Figure3, it can be seen that the performance of the neural network model gradually stabilizes as the number of iterations increases during the training process.

Table 1 .
The statistics of dataset