New fault detection methods for air-to-air warmth pumps – pv journal International

July 17, 2024 Valerie Thompson

A Portuguese analysis crew in contrast a number of utilized machine studying methods to carry out fault detection in air-to-air warmth pumps in cooling mode and recognized 4 algorithm that exhibits a excessive stage of efficiency based mostly on the metrics of accuracy, precision, recall, and F1 rating, which is a key quantity used to measure the efficiency of classifications.

The speedy development within the deployment of warmth pumps in constructing heating, air flow, and air-conditioning (HVAC) methods creates a necessity for ease of upkeep and reliability by means of early detection of system errors, in line with the analysis crew.

“The use of synthetic intelligence (AI) instruments is regularly growing, which makes it attainable to make potential enhancements in lots of areas,” the corresponding creator of the analysis, Pedro Barandier, stated. pv journalexplaining that the analysis is pushed by the “extraordinary development” of warmth pumps, and the promise of classification algorithms, neural networks, and different AI instruments to enhance the reliability of such methods, particularly that when used together with web of issues applied sciences.

Python is used to conduct comparative evaluation of supervised studying classification algorithms for fault detection. The algorithms studied embody Naïve Bayes, Support Vector Machine, Logistic Regression, and Okay-nearest Neighbors with a plan to guage them based mostly on accuracy, precision, recall, and F1 rating, a measure of harmonic imply precision and recall.

The crew recognized widespread faults within the cooling mode, similar to compressor and reversing valve leakage, improper condenser and evaporator fouling, liquid line restriction, refrigerant undercharge and overcharge, and the presence of non-condensable gases.

A dataset derived from prior analysis involving a residential 8.8 kW warmth pump (HP), with a scroll compressor and a thermostatic growth machine was used. It consists of 96 variables and is constructed from 7,374 assessments carried out in cooling mode, together with temperatures, strain, air and refrigerant mass circulate price, electrical energy, coefficients of efficiency, error stage and others. Many parameters of the warmth pump thermodynamic circuit are additionally accessible, similar to temperature, strain, mass circulate price and compressor electrical energy.

Several rounds of characteristic choice are carried out. In addition, a principal element evaluation was additionally carried out, and it was noticed that ten parts defined greater than 90% of the variance. To establish the variables which have essentially the most influence on the outcomes, an ablation research was additionally carried out, similar to a correlation matrix to additional scale back the variety of components.

A connection weighted technique by means of a single-layer synthetic neural community was developed and 40 options have been chosen. In addition, an artificial minority oversampling method (SMOTE) was developed to steadiness the wholesome circumstances within the coaching set.

The greatest outcomes have been achieved by Okay-nearest Neighbors, which confirmed the very best values ​​for every of the 4 metrics, above 99% after a cross-validation contemplating the discount of the variety of parts. But Naïve Bayes, Support Vector Machine, and Logistic Regression additionally obtain 90% efficiency metrics.

Some fashions have been superficially analyzed utilizing the machine studying instrument PyCaret, in line with the crew. “Ridge Regression is usually the quickest of the perfect algorithms with excessive values ​​for precision, accuracy, recall, and F1. This is a vital discovering, as there isn’t any earlier software of this algorithm that discovered within the literature for diagnosing HP faults,” it concluded.

Looking forward, the analysis crew will increase past error detection. “As the fault analysis course of depends on varied intrinsic facets, this work is taken into account first, fault detection. The subsequent purpose of the crew is to additional enhance the acquired data to develop an entire and environment friendly warmth pump fault analysis system that engineers and professionals around the globe can reap the benefits of,” stated Barandier.

The research is detailed in “Comparative evaluation of 4 classification algorithms for fault detection in warmth pumps,” printed in Energy and Buildings. The researchers are from the Polytechnic University of Coimbra in Portugal and CISE – Electromechatronic Systems Research Center, University of Beira Interior.