Sudharson K , Varsha S , Santhiya R , Rajalakshmi D
{"title":"用于工业物联网系统预测性维护的量子增强LSTM","authors":"Sudharson K , Varsha S , Santhiya R , Rajalakshmi D","doi":"10.1016/j.mex.2025.103653","DOIUrl":null,"url":null,"abstract":"<div><div>An innovative solution for predictive maintenance in IIoT systems combining quantum computing with the proficiency of LSTM neural networks is proposed by us. Our concept is guided by a hybrid quantum-classical architecture to facilitate quantum computing to exploit high-dimensional industrial sensor measurements while preserving crucial temporal relationships through particular quantum channels. Through the combination of the representational ingenuity of quantum circuits, along with the sequence-based modelling of classical LSTMs, QE-LSTM is uniquely positioned to handle complicated time series coming out of industrial sensors. At the heart of our methodology are the following unique elements:<ul><li><span>•</span><span><div>A collaborative framework integrating quantum and classical technologies allowing for the quantum computer to manage the complex analysis of high dimensional sensor data in the industry.</div></span></li><li><span>•</span><span><div>Quantum channel designs were aimed at minimizing temporal dependencies in temporal series industrial measurements, thereby maximizing the quality of sequential analysis.</div></span></li><li><span>•</span><span><div><div>Under ODS hindcasting, QE-LSTM improved F1 by 4–5 percentage points on SECOM and reduced RMSE and NASA Score on C-MAPSS; trends were consistent on IMMD (<span><span>Table 1</span></span>, <span><span>Table 2</span></span>).</div><div><span><span><p><span>Table 1</span>. <!-->Performance comparison across datasets.</p></span></span><div><table><thead><tr><th>Dataset</th><th>Model</th><th>Accuracy</th><th>Precision</th><th>Recall</th><th>F1</th><th>AUC</th></tr></thead><tbody><tr><td>SECOM</td><td>LSTM</td><td>0.864</td><td>0.842</td><td>0.809</td><td>0.825</td><td>0.902</td></tr><tr><td></td><td>CNN-LSTM</td><td>0.878</td><td>0.862</td><td>0.824</td><td>0.842</td><td>0.914</td></tr><tr><td></td><td><strong>QE-LSTM (sim)</strong></td><td><strong>0.904</strong></td><td><strong>0.892</strong></td><td><strong>0.861</strong></td><td><strong>0.876</strong></td><td><strong>0.938</strong></td></tr><tr><td></td><td><strong>QE-LSTM (hardware)</strong></td><td>0.896</td><td>0.881</td><td>0.850</td><td>0.865</td><td>0.930</td></tr><tr><td>IMMD</td><td>LSTM</td><td>0.906</td><td>0.883</td><td>0.862</td><td>0.872</td><td>0.943</td></tr><tr><td></td><td>CNN-LSTM</td><td>0.913</td><td>0.891</td><td>0.869</td><td>0.880</td><td>0.949</td></tr><tr><td></td><td><strong>QE-LSTM (sim)</strong></td><td><strong>0.928</strong></td><td><strong>0.908</strong></td><td><strong>0.888</strong></td><td><strong>0.898</strong></td><td><strong>0.960</strong></td></tr></tbody></table></div><div><div>QE-LSTM (sim) vs LSTM F1 deltas: SECOM <strong>+5.1 pp</strong>, IMMD <strong>+2.6 pp</strong>; paired <em>t</em>-test <em>p</em> < 0.01.</div></div></div><div><span><span><p><span>Table 2</span>. <!-->RUL prediction performance metrics.</p></span></span><div><table><thead><tr><th>Metric</th><th>Classical LSTM</th><th>CNN-LSTM</th><th>QE-LSTM (sim)</th><th>QE-LSTM (hardware)</th><th>Improvement vs LSTM</th></tr></thead><tbody><tr><td>RMSE ↓</td><td>20.6</td><td>19.4</td><td><strong>18.1</strong></td><td>18.7</td><td><strong>12.1</strong> <strong>%</strong></td></tr><tr><td>MAE ↓</td><td>15.4</td><td>14.6</td><td><strong>13.8</strong></td><td>14.3</td><td><strong>10.4</strong> <strong>%</strong></td></tr><tr><td>NASA Score ↓</td><td>692</td><td>648</td><td><strong>603</strong></td><td>621</td><td><strong>12.8</strong> <strong>%</strong></td></tr></tbody></table></div><div><div>* Evaluations are positive if all the metrics have smaller values. Score is a metric that stands out because of the tendency to heavily penalise late predictions (which is highly relevant to effective maintenance planning).</div></div></div></div></span></li></ul>In the application of failure detection of bearing, QE-LSTM improves F1 over classical baselines on SECOM by 4–5 pp, with similar gains on IMMD; results on C-MAPSS (RUL) show consistent reductions in RMSE and NASA score.</div></div>","PeriodicalId":18446,"journal":{"name":"MethodsX","volume":"15 ","pages":"Article 103653"},"PeriodicalIF":1.9000,"publicationDate":"2025-09-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Quantum-enhanced LSTM for predictive maintenance in industrial IoT systems\",\"authors\":\"Sudharson K , Varsha S , Santhiya R , Rajalakshmi D\",\"doi\":\"10.1016/j.mex.2025.103653\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>An innovative solution for predictive maintenance in IIoT systems combining quantum computing with the proficiency of LSTM neural networks is proposed by us. Our concept is guided by a hybrid quantum-classical architecture to facilitate quantum computing to exploit high-dimensional industrial sensor measurements while preserving crucial temporal relationships through particular quantum channels. Through the combination of the representational ingenuity of quantum circuits, along with the sequence-based modelling of classical LSTMs, QE-LSTM is uniquely positioned to handle complicated time series coming out of industrial sensors. At the heart of our methodology are the following unique elements:<ul><li><span>•</span><span><div>A collaborative framework integrating quantum and classical technologies allowing for the quantum computer to manage the complex analysis of high dimensional sensor data in the industry.</div></span></li><li><span>•</span><span><div>Quantum channel designs were aimed at minimizing temporal dependencies in temporal series industrial measurements, thereby maximizing the quality of sequential analysis.</div></span></li><li><span>•</span><span><div><div>Under ODS hindcasting, QE-LSTM improved F1 by 4–5 percentage points on SECOM and reduced RMSE and NASA Score on C-MAPSS; trends were consistent on IMMD (<span><span>Table 1</span></span>, <span><span>Table 2</span></span>).</div><div><span><span><p><span>Table 1</span>. <!-->Performance comparison across datasets.</p></span></span><div><table><thead><tr><th>Dataset</th><th>Model</th><th>Accuracy</th><th>Precision</th><th>Recall</th><th>F1</th><th>AUC</th></tr></thead><tbody><tr><td>SECOM</td><td>LSTM</td><td>0.864</td><td>0.842</td><td>0.809</td><td>0.825</td><td>0.902</td></tr><tr><td></td><td>CNN-LSTM</td><td>0.878</td><td>0.862</td><td>0.824</td><td>0.842</td><td>0.914</td></tr><tr><td></td><td><strong>QE-LSTM (sim)</strong></td><td><strong>0.904</strong></td><td><strong>0.892</strong></td><td><strong>0.861</strong></td><td><strong>0.876</strong></td><td><strong>0.938</strong></td></tr><tr><td></td><td><strong>QE-LSTM (hardware)</strong></td><td>0.896</td><td>0.881</td><td>0.850</td><td>0.865</td><td>0.930</td></tr><tr><td>IMMD</td><td>LSTM</td><td>0.906</td><td>0.883</td><td>0.862</td><td>0.872</td><td>0.943</td></tr><tr><td></td><td>CNN-LSTM</td><td>0.913</td><td>0.891</td><td>0.869</td><td>0.880</td><td>0.949</td></tr><tr><td></td><td><strong>QE-LSTM (sim)</strong></td><td><strong>0.928</strong></td><td><strong>0.908</strong></td><td><strong>0.888</strong></td><td><strong>0.898</strong></td><td><strong>0.960</strong></td></tr></tbody></table></div><div><div>QE-LSTM (sim) vs LSTM F1 deltas: SECOM <strong>+5.1 pp</strong>, IMMD <strong>+2.6 pp</strong>; paired <em>t</em>-test <em>p</em> < 0.01.</div></div></div><div><span><span><p><span>Table 2</span>. <!-->RUL prediction performance metrics.</p></span></span><div><table><thead><tr><th>Metric</th><th>Classical LSTM</th><th>CNN-LSTM</th><th>QE-LSTM (sim)</th><th>QE-LSTM (hardware)</th><th>Improvement vs LSTM</th></tr></thead><tbody><tr><td>RMSE ↓</td><td>20.6</td><td>19.4</td><td><strong>18.1</strong></td><td>18.7</td><td><strong>12.1</strong> <strong>%</strong></td></tr><tr><td>MAE ↓</td><td>15.4</td><td>14.6</td><td><strong>13.8</strong></td><td>14.3</td><td><strong>10.4</strong> <strong>%</strong></td></tr><tr><td>NASA Score ↓</td><td>692</td><td>648</td><td><strong>603</strong></td><td>621</td><td><strong>12.8</strong> <strong>%</strong></td></tr></tbody></table></div><div><div>* Evaluations are positive if all the metrics have smaller values. Score is a metric that stands out because of the tendency to heavily penalise late predictions (which is highly relevant to effective maintenance planning).</div></div></div></div></span></li></ul>In the application of failure detection of bearing, QE-LSTM improves F1 over classical baselines on SECOM by 4–5 pp, with similar gains on IMMD; results on C-MAPSS (RUL) show consistent reductions in RMSE and NASA score.</div></div>\",\"PeriodicalId\":18446,\"journal\":{\"name\":\"MethodsX\",\"volume\":\"15 \",\"pages\":\"Article 103653\"},\"PeriodicalIF\":1.9000,\"publicationDate\":\"2025-09-28\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"MethodsX\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S2215016125004972\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q2\",\"JCRName\":\"MULTIDISCIPLINARY SCIENCES\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"MethodsX","FirstCategoryId":"1085","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2215016125004972","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"MULTIDISCIPLINARY SCIENCES","Score":null,"Total":0}
Quantum-enhanced LSTM for predictive maintenance in industrial IoT systems
An innovative solution for predictive maintenance in IIoT systems combining quantum computing with the proficiency of LSTM neural networks is proposed by us. Our concept is guided by a hybrid quantum-classical architecture to facilitate quantum computing to exploit high-dimensional industrial sensor measurements while preserving crucial temporal relationships through particular quantum channels. Through the combination of the representational ingenuity of quantum circuits, along with the sequence-based modelling of classical LSTMs, QE-LSTM is uniquely positioned to handle complicated time series coming out of industrial sensors. At the heart of our methodology are the following unique elements:
•
A collaborative framework integrating quantum and classical technologies allowing for the quantum computer to manage the complex analysis of high dimensional sensor data in the industry.
•
Quantum channel designs were aimed at minimizing temporal dependencies in temporal series industrial measurements, thereby maximizing the quality of sequential analysis.
•
Under ODS hindcasting, QE-LSTM improved F1 by 4–5 percentage points on SECOM and reduced RMSE and NASA Score on C-MAPSS; trends were consistent on IMMD (Table 1, Table 2).
Table 1. Performance comparison across datasets.
Dataset
Model
Accuracy
Precision
Recall
F1
AUC
SECOM
LSTM
0.864
0.842
0.809
0.825
0.902
CNN-LSTM
0.878
0.862
0.824
0.842
0.914
QE-LSTM (sim)
0.904
0.892
0.861
0.876
0.938
QE-LSTM (hardware)
0.896
0.881
0.850
0.865
0.930
IMMD
LSTM
0.906
0.883
0.862
0.872
0.943
CNN-LSTM
0.913
0.891
0.869
0.880
0.949
QE-LSTM (sim)
0.928
0.908
0.888
0.898
0.960
QE-LSTM (sim) vs LSTM F1 deltas: SECOM +5.1 pp, IMMD +2.6 pp; paired t-test p < 0.01.
Table 2. RUL prediction performance metrics.
Metric
Classical LSTM
CNN-LSTM
QE-LSTM (sim)
QE-LSTM (hardware)
Improvement vs LSTM
RMSE ↓
20.6
19.4
18.1
18.7
12.1%
MAE ↓
15.4
14.6
13.8
14.3
10.4%
NASA Score ↓
692
648
603
621
12.8%
* Evaluations are positive if all the metrics have smaller values. Score is a metric that stands out because of the tendency to heavily penalise late predictions (which is highly relevant to effective maintenance planning).
In the application of failure detection of bearing, QE-LSTM improves F1 over classical baselines on SECOM by 4–5 pp, with similar gains on IMMD; results on C-MAPSS (RUL) show consistent reductions in RMSE and NASA score.