طرّاحي يك حسگر نرم‌افزاري به‌صورت سبك‌وزن جهت استفاده در كاربردهاي كنترل كيفي

In modern industrial processes, direct measurement of quality variables faces serious challenges due to technical limitations, high maintenance costs of hardware sensors, an‎d delays caused by laboratory analyses. Software sensors, as an efficient alternative, enable real-time estimation of these variables by utilizing data-driven models such as machine learning. With the emergence of edge computing, transferring these sensors from centralized servers to locations closer to the data source (onto edge hardware) has become necessary in order to reduce latency an‎d enhance data security. The main challenge lies in coping with the limited computational an‎d memory resources of edge hardware such as STM32 series microcontrollers, which constitutes a major obstacle to the deployment of software sensors based on machine learning models. In this research, a lightweight software sensor based on the machine learning model Broad Learning System (BLS) has been designed an‎d implemented. The selec‎tion of the BLS model, due to its flat architecture, small number of parameters, an‎d lack of need for long an‎d iterative training procedures, provides high potential as a lightweight model suitable for execution in resource-constrained environments. To further lightweight this model an‎d better address the limitations of edge hardware, a custom quantization pipeline is proposed in this work to convert BLS weights from 32-bit floating-point precision to 8-bit integer representation. This approach achieves up to a 75% reduction in model parameter storage size with an acceptable accuracy degradation of 0.0683%. This method optimizes the utilization of STM32 hardware computational units an‎d significantly reduces memory consumption. In simulation on the STM32F746NG, the model achieved an average inference time of 6.79 milliseconds per sample. Furthermore, to address the phenomenon of data drift in multimode industrial processes, which leads to a degradation of sensor accuracy over time, the main innovation of this research is the proposal of a comprehensive workflow for managing multimode processes through the definition of a multimodel approach an‎d the design of a novel drift detection index. This index redefines an‎d lightweightens the Kolmogorov–Smirnov (KS) statistical test through computational approximation an‎d demonstrates a 90.98% to 97.25% improvement in drift detection speed compared to the original KS test, while being executable on edge devices with minimal computational overhead. In the proposed software sensor, the selec‎tion of the appropriate model for each operating mode is also performed using this same index, thereby avoiding additional code overhead for defining a separate selec‎tion mechanism. eva‎luations conducted on the multimode industrial chemical process dataset of Tennessee Eastman (TE) show that the proposed drift index achieves an average improvement of 66.18% to 96.93% in drift detection speed compared to conventional methods. Finally, the performance of the proposed software sensor, implemented in the C language for execution on microcontrollers as part of a control system, was eva‎luated under online scenarios in the presence of drift an‎d transitions between operating modes, an‎d the effectiveness of the sensor in these scenarios was assessed.

6

160744