PaperlayerView exact revision on arXiv

Source-linked AI summary

A Novel Semi-Supervised Data-Driven Method for Chiller Fault Diagnosis with Unlabeled Data

Bingxu Li, Fanyong Cheng, Xin Zhang, Can Cui, Wenjian Cai

arXiv:2011.00187v1cs.LGcs.SYeess.SY

TL;DR

Chiller fault diagnosis methods depend on labeled training data, although practical datasets are often only partially labeled and acquiring labels is laborious. This paper uses a semi-generative adversarial network to learn from labeled and unlabeled data; with enough unlabeled samples, it reduces the minimum labeled-data requirement by about 60%.

  • Problem

    Existing supervised methods need enough labeled samples, while available chiller datasets are often partially labeled and collecting labels is laborious.

    Source: §1.2 · p. 5

  • Method

    The proposed semi-supervised chiller diagnosis method trains with labeled and unlabeled data and learns data-distribution information from unlabeled samples.

    Source: §1.3 · pp. 5–6, §3.3 · p. 14

  • Results

    About 60% fewer labeled samples are required to reach the desired diagnostic accuracy when enough unlabeled samples are available.

    Source: §6 · p. 25

  • Takeaways & Limitations

    The method is presented as applicable to practical settings with limited labeled data and plentiful unlabeled data.

    Source: §6 · p. 25

    The method cannot diagnose unknown faults absent from training data or two faults occurring simultaneously.

    Source: §6 · p. 25

Abstract

from arXiv · show

In practical chiller systems, applying efficient fault diagnosis techniques can significantly reduce energy consumption and improve energy efficiency of buildings. The success of the existing methods for fault diagnosis of chillers relies on the condition that sufficient labeled data are available for training. However, label acquisition is laborious and costly in practice. Usually, the number of labeled data is limited and most data available are unlabeled. The existing methods cannot exploit the information contained in unlabeled data, which significantly limits the improvement of fault diagnosis performance in chiller systems. To make effective use of unlabeled data to further improve fault diagnosis performance and reduce the dependency on labeled data, we proposed a novel semi-supervised data-driven fault diagnosis method for chiller systems based on the semi-generative adversarial network, which incorporates both unlabeled and labeled data into learning process. The semi-generative adversarial network can learn the information of data distribution from unlabeled data and this information can help to significantly improve the diagnostic performance. Experimental results demonstrate the effectiveness of the proposed method. Under the scenario that there are only 80 labeled samples and 16000 unlabeled samples, the proposed method can improve the diagnostic accuracy to 84%, while the supervised baseline methods only reach the accuracy of 65% at most. Besides, the minimal required number of labeled samples can be reduced by about 60% with the proposed method when there are enough unlabeled samples.

1. Introduction

Chiller faults can waste energy, but existing diagnosis methods face limits from complex physical models or scarce labels. The paper proposes semi-supervised diagnosis that learns from unlabeled data to reduce labeled-data needs and improve performance.

1.1 Background

  • Accurate chiller fault diagnosis matters because soft faults can go unnoticed, reduce COP, and waste energy.The paper notes that chillers are major HVAC energy consumers, especially in the tropics.Source: §1.1 · p. 3

1.2 Review of the existing data-driven fault diagnosis methods for chillers

  • Existing supervised methods need enough labeled samples, while semi-supervised approaches still depend on sufficient normal data, leaving fully unlabeled data underused.The authors identify this as a practical gap because labeling is laborious and collected datasets are often only partially labeled.Source: §1.2 · p. 5

1.3 Contributions

  • The proposed semi-GAN chiller diagnosis framework trains with labeled and unlabeled data, learns data-distribution information, and aims to improve performance with fewer labels.The paper also compares performance across training sets with different labeled and unlabeled sample counts.Source: §1.3 · pp. 5–6

2. Description of the chiller system

The section describes the chiller refrigeration cycle and the practical chiller dataset used to study seven faults at multiple severity levels and operating conditions.

2.1 Basic Working Principle of the Chiller System

  • The chiller cycle transfers building heat to refrigerant in the evaporator, compresses the vapor, rejects heat in the condenser, and expands the liquid through the valve.The schematic identifies the compressor, condenser, expansion valve, and evaporator as the four main components.Source: §2.1 · p. 8, Figure 2 · p. 8

2.2 Experimental Data of a Practical Chiller System

  • The experiments use data from a McQuay PEH048J 90-t (316 kW) centrifugal chiller collected through ASHRAE Project 1043-rp.The system has a centrifugal compressor, shell-and-tube evaporator and condenser, and pilot-driven expansion valve.Source: §2.2 · p. 8
  • The dataset covers seven typical chiller faults, each manually introduced at four severity levels.The faults include reduced condenser water flow, non-condensable gas, condenser fouling, reduced evaporator water flow, excess oil, refrigerant overcharge, and refrigerant leak.Source: §2.2 · pp. 8–9
  • The study uses 61 of the 65 recorded features, sampled every 10 seconds, across 27 predefined operating conditions.Each fault-severity condition was collected for one day with about 5,000 samples; normal operation was recorded for several days with more than 30,000 samples.Source: §2.2 · p. 9

3. Semi-Generative Adversarial Network

The section develops Semi-GAN from the adversarial generator–discriminator framework into a classifier trained with labeled, unlabeled, and generated data. Its training refinements and two-ring example explain how unlabeled data can inform classification when labels are scarce.

3.1 Generative adversarial network

  • The conventional GAN pits a generator producing samples from random noise against a discriminator distinguishing generated samples from real samples.The discriminator learns to classify real and fake samples; the generator aims to produce samples that fool it.Source: §3.1 · pp. 10–11

3.2 Semi-Generative Adversarial Network

  • Semi-GAN trains a discriminator-classifier with labeled samples, unlabeled samples, and generated samples, using different targets for each type.Labeled samples receive their class labels, unlabeled samples are assigned to real, and generated samples are assigned to fake.Source: §3.2 · pp. 11–12
  • Semi-GAN extends GAN to semi-supervised classification: its discriminator predicts among N real classes and an additional N+1 fake class.For labeled samples the discriminator learns the correct class; unlabeled samples are trained as real without assigning them a class.Source: §3.2 · pp. 11–12
  • Alternating discriminator and generator training lets the classifier learn patterns and distributions from unlabeled data, which the section says can improve classification with few labeled samples.The generator is trained to produce samples the discriminator regards as real, while discriminator updates use the three data types.Source: §3.2 · p. 13

3.3 Improving technique for training Semi-GAN

  • The training refinement removes the extra fake-class output by setting its logit to zero and changing the probability calculation, leaving an N-class classifier.The procedure otherwise follows the training described for Semi-GAN.Source: §3.3 · p. 14
  • With three labeled samples per class, the example contrasts a supervised boundary that fails to separate two rings with Semi-GAN’s boundary using their unlabeled distribution.The example assigns the inner ring to class 1 and the outer ring to class 2.Source: §3.3 · pp. 14–15

4. Proposed Fault Diagnosis Framework based on Semi-GAN

The framework preprocesses chiller data, then trains a semi-GAN fault classifier using labeled and unlabeled operating data. Its trained discriminator classifies real-time measurements, while the generator and discriminator use different network capacities.

4.1 Data preprocessing: remove the outliers and standardize

  • Outliers are removed before feature standardization to [-1,1].The method filters values beyond seven standard deviations and removes points whose adjacent changing rates both exceed that threshold.Source: §4.1 · p. 16

4.2 Chiller fault diagnosis based on Semi-GAN

  • The semi-GAN trains a fault classifier from labeled and unlabeled operating data, using its discriminator for online classification.The classifier takes real-time chiller measurements as input and reports diagnosed fault types.Source: §4.2 · p. 16, §4.3 · p. 18
  • The generator uses a more complex network than the discriminator, which the authors associate with generated samples closer to real data and improved discriminator performance.Their selected network sizes are 8-64-64-61 for the generator and 61-32-16-8 for the discriminator.Source: §4.2 · pp. 16–17
  • Training updates the discriminator with unlabeled, generated, and labeled samples, then updates the generator while holding the discriminator fixed.The procedure repeats over minibatches and training iterations and outputs the discriminator parameters.Source: §4.2 · p. 17
  • The real-system workflow obtains mostly unlabeled data, trains the classifier with the labeled and unlabeled sets, then applies it to online measurements.The paper notes that labeling is laborious and costly, so only a small portion of operating data may be labeled manually.Source: §4.3 · p. 18

5. Validation Using Experimental Data

Across experimental comparisons, the proposed method improves diagnostic accuracy particularly when labeled training data are scarce. Its gains also increase with unlabeled-data availability, and it can reduce the labeled-data requirement under the tested conditions.

5. Validation Using Experimental Data

  • The dataset contained 162379 samples, and experiments varied labeled-set sizes while fixing the unlabeled training set at 16000 samples for the supervised comparison.The methods were evaluated on a fixed 16000-sample test set; classifiers were trained ten times per scenario, with results averaged across ten dataset splits.Source: §5 · pp. 19–20
  • 84% accuracy with 80 labeled samples exceeded the supervised methods’ maximum of 65%; with 40 labeled samples, accuracy was 71.01% versus below 45% for each supervised method.Table 3 reports average testing accuracy over ten experiments, with standard deviations in parentheses.Source: §5 · p. 20, Table 3 · p. 20
  • For 80 labeled samples, accuracy increased with fault severity, while the proposed method’s relative improvement was greater for lower-severity faults.The authors report that semi-GAN learns distribution information from unlabeled samples, helping the classifier identify fault patterns.Source: §5 · pp. 20–21, Figure 7 · p. 21, Figure 8 · p. 21
  • With fewer than 100 labeled samples, the proposed method improved diagnostic performance by more than 20%, compared with 2%–5% for semi-SVM.The comparison is against supervised learning results.Source: §5 · p. 22, Table 4 · p. 22

5.3 Effect of the unlabeled data size and the labeled data size

  • With 40 labeled samples, 800 unlabeled samples raised testing accuracy from 45% to 61%; across tested cases, more unlabeled data improved performance, especially with fewer labels.All reported proposed-method accuracy values were above the supervised baseline, and gains tended to level off as unlabeled data increased.Source: §5.3 · pp. 22–23, Figure 9 · p. 23

5.4 Minimal required labeled data size for different desired diagnostic accuracy

  • When unlabeled samples exceeded 4800, the minimum labeled-sample requirement was reduced by about 60% versus the supervised baseline.The minimum required labeled-data size increased with the desired diagnostic accuracy.Source: §5.4 · pp. 23–24, Figure 10 · p. 23, Figure 11 · p. 24

6. Conclusion

The semi-generative adversarial network uses labeled and unlabeled data for chiller fault diagnosis. The conclusion reports improved performance, reduced labeled-data requirements, and remaining limits for unknown and simultaneous faults.

  • With 80 labeled samples, the proposed method achieves 84% diagnostic accuracy, while supervised baseline methods reach 65% at most.Source: §6 · p. 25
  • The method incorporates labeled and unlabeled data and uses information from unlabeled-data distributions to improve chiller fault diagnosis.Source: §6 · p. 25
  • Performance improvement is greater for lower-severity faults, supporting accurate diagnosis at the incipient stage before faults become severe.Source: §6 · p. 25
  • Increasing the unlabeled training-set size improves diagnostic performance, with greater gains when fewer labeled samples are available.Source: §6 · p. 25
  • With enough unlabeled samples, the method reduces the minimal required number of labeled samples by about 60%.The passage attributes this reduction to the network exploiting information in unlabeled data.Source: §6 · p. 25
  • The method cannot diagnose faults absent from the training data or two faults occurring simultaneously.Source: §6 · p. 25

Related papers

Loading 2011.00187v1…