3.2.3. Classification

Finally, the concatenated feature vector obtained in the feature learning stage is directly fed to the classification module, which is composed of a convolution and global average pooling layer, a fully connected layer, and a softmax layer. As a result, it outputs conditional probability for each class. The softmax function rescales the *n*-dimensional vector of the FC layer output so that the output value is in the range [0, 1] and the sum is 1, which is defined by the following:

$$s(v\_i) = \frac{e^{v\_i}}{\sum\_{j=1}^{n} e^{v\_j}} \tag{15}$$

The full convolution module and LSTM module process the same time series input in two different fields of view. The full convolution is a fixed-size perception field to extract local features of time series. On the contrary, LSTM effectively captures time dependencies. The method of combining with convolutional and recurrent neural networks is crucial to enhance the performance of the proposed framework.

#### *3.3. Quality Prediction System Based on MCRNN*

Based on a large amount of process information collected by sensors, a quality prediction and control system is established for intelligent decision-making and control. To elab-

orate on the infrastructure of an established system, the framework of the system based on MCRNN is described in Figure 5. It mainly consists of three parts: data acquisition, quality prediction, and dynamic control. Data acquisition module based on various sensor networks collects massive real-time production data about the continuous casting process, such as temperature, water volume, and casting speed. The real-time collected process data will be sent to the quality prediction module and stored as historical data for visualizing the display and training of the model. Moreover, the quality information of each rolled slab is collected to label continuous casting data.

**Figure 5.** The framework of quality prediction system based on MCRNN.

With production process parameters and slab labels, a quality prediction model based on the proposed MCRNN is built. In the real-time production process, the original time series data are entered into the model and transformed with different time scales and frequencies. The output of the model is the quality label of CCS. Once the slab in producing is judged to be abnormal by the prediction model, the knowledge of domain experts will be employed to dynamically adjust the production process. The dynamic control module adjusts the process and equipment parameters in time through the programmable logic controller to avoid affecting the next rolling process and causing waste. Abnormal CCS produced will be sorted into the cleaning process of the machine to eliminate defects. The workflow improves efficiency, reduces costs, and enhances yield greatly.

#### **4. Experiments and Results**

In this section, we first describe the dataset and the evaluation metrics. Then, the effects of the RUS method and multiscale transformations are discussed in our studies. Finally, the proposed MCRNN model compares with different baseline models.
