Predicting process quality in multi-stage manufacturing using AE-BilA: an autoencoder-BiLSTM with attention mechanism

Abstract As industrial processes are becoming increasingly complex and data-driven, the need for accurate quality predictions in manufacturing systems is regarded as critical. To address this challenge, AE-BiLA (Autoencoder-Bidirectional Long Short-Term Memory with Attention mechanism) has been proposed as a framework in which a stacked Long Short-Term Memory (LSTM) autoencoder is combined with a bidirectional LSTM enhanced by an Attention mechanism for predicting quality in multi-stage manufacturing processes (MMP). First, high-dimensional, noisy data are reduced by employing the stacked LSTM autoencoder, with essential information being preserved. Next, the compressed features are fed into the bidirectional LSTM, where significant temporal patterns are highlighted by the Attention mechanism. The method was validated on a real-world MMP dataset. An R 2 (coefficient of determination) of 0.9452 was obtained in Stage 1, demonstrating that upstream process dynamics were effectively captured. In contrast, an R 2 of 0.7329 was produced in Stage 2, reflecting increased complexity and variability in downstream operations. Moreover, the Symmetric Mean Absolute Percentage Error (SMAPE) in Stage 2 was reduced to 1.9319 from 19.3583 in Stage 1, thereby underscoring that outliers and noise were successfully managed. Overall, the AE-BiLA framework outperforms existing methods by effectively integrating denoising with a bidirectional recurrent structure. Despite the increased computational overhead, it is expected that the framework will yield substantial gains in productivity, lower waste levels, and reduce operational costs.