MTL-Net: A Unified Deep Learning Architecture for Predicting Production Efficiency, Defect Rate, and Speed in Industry 4.0 Systems

The paper proposes a multi-task deep learning model for smart manufacturing that allows to classify efficiency and predicting two quality indicators: defect and speed. To capture dependencies in IIoT sensor stream, the model combines parallel CNN layers, BiLSTM-GRU recurrent encoders, and multi-head attention. The multi-sensor production sequences, amounting to 80,000, were employed in the training of the system through stratified balancing, adaptive loss weighting, and regularization. The experiments show that this method can predict well with accuracy above 93.3% for the classification of efficiency and R ²  = 0.924 for the defect rate and R ²  = 0.981 for the production speed. The outcomes validate that learning representations for multiple tasks is more effective than independently training each task. A scalable framework for decision support in Industry 4.0 real-time predictive maintenance resource allocation production optimization is provided in the proposed work.