Industrial sensing is a core technology for intelligent manufacturing. In recent years, utilizing artificial neural networks (ANNs) to improve industrial sensing accuracy has become a hot research topic. However, when dealing with industrial field data with small sample sizes, high noise, and strong nonlinearity, randomly initialized neural networks show poor generalization performance and are hard to deploy in practical industrial scenarios. To tackle this problem, a research team led by Professor Lanxiang Sun from the Shenyang Institute of Automation, Chinese Academy of Sciences, published a research paper in Engineering and proposed a partial least squares (PLS)-assisted optimization network (PLSaoNET). This model eliminates random initialization in neural networks and provides a clear path for network training.
PLSaoNET involves the following two core innovations:
1. PLS-based initialization mechanism. PLS is a classic statistical method. It has a clear solution path and strong interpretability, but can only fit linear relationships. The research team used the solution of PLS as the initial values for the neural network: The number of PLS latent variables determines the number of neurons in the hidden layer, and the PLS weight matrix is used to initialize the hidden layer. This gives the neural network a starting point with clear statistical meaning. The subsequent training is no longer a blind search, but a re-optimization of the PLS result, which introduces nonlinear modeling ability while maintaining interpretability.
2. Stratified sampling-based training strategy. Sample labels from industrial fields are often unevenly distributed, which causes large loss fluctuations when training the network with mini-batch gradient descent. The research team designed a stratified sampling method: The training samples are divided into several strata based on their label values, and samples are drawn proportionally from each stratum during each mini-batch training iteration. This effectively smooths the gradient updates, enhances the model's robustness to unevenly distributed samples, and further improves the stability of model training.
Professor Lanxiang Sun, the corresponding author, said: "Labeled data from industrial sites is hard to obtain. What we need is not blindly pursuing complex network structures, but enabling the model to output stable and reliable results even with limited data. The original intention of PLSaoNET is to let statistical mechanism provide a trustworthy starting point for data-driven learning."
The research team thoroughly validated the proposed method in two typical industrial scenarios: online monitoring of iron ore concentrate slurry grade using laser-induced breakdown spectroscopy (LIBS) technology, and diesel fuel quality assessment using near-infrared (NIR) spectroscopy. The results showed that compared with the PLS regression model and the backpropagation neural network (BPNN) with Xavier initialization, PLSaoNET exhibited the best modeling accuracy and generalization performance.
Notably, by visualizing the hidden layer weights of the LIBS spectrum–iron grade model, the research team revealed the essence of PLSaoNET's reliability:
- Although the PLS model lacked precision, the dominant latent variables it extracted clearly focused on the characteristic emission lines of mineral elements such as Si, Mg, and Fe, with clear physical correspondence.
- In the randomly initialized BPNN, the neuron weights in the hidden layer could hardly distinguish background noise from valid spectral signals. The feature extraction process resulted in an uninterpretable combination of data.
- PLSaoNET inherited the PLS's high attention to key spectral lines, while moderately enhancing the learning of auxiliary information such as Ca and Al through retraining, fundamentally suppressing overfitting.
PLSaoNET has been deployed in the LIBS slurry analyzer at a mineral processing plant for real-time iron ore concentrate grade monitoring. As a general method, PLSaoNET can effectively lower the threshold to industrial ANN application and provide technical support for intelligent sensing in complex industrial settings.
