利用服役數據和機器學習模型評估混合動力汽車鎳氫電池的健康狀態

Abstract

隨著全球氣侯變遷加劇及環保意識抬頭，油電混合車及電動車的銷量與日俱增，其中油電混合車於上世紀末商業化，至今已有逾千萬台的油電車於全球運行；而鎳氫電池作為其主要的儲電元件，未來全球將有大量的汰役電池需要被妥善處理。過往處理電池需要對電池做完全充放電，知道電池實際的充放電能力後，再依照量測結果判斷電池是否適合繼續服役，抑或是廢棄淘汰，但是完全充放電時間太長，不利於應付未來龐大的電池回收潮。此外也希望能夠預測電池性能未來的使用狀況和變化程度，依照電池老化狀況安排合適的使用方式，延長服役時間以節省製造電池的成本和資源，為此我們急需快速且準確的手段，檢測電池實際性能並進階地預測到未來的電池狀態變化。 由於科技的普及與資料科學之快速進展，人工智慧(artificial intelligence, AI)已經被應用於許多產業以增進生產效率和社會福祉。過去已有文獻將其應用於預測電池之健康狀態(state of health, SOH)，能準確地甄別電池狀態並取得優秀的成效，但是訓練樣本通常不大，僅限於幾百甚至幾十個，樣本數過小可能導致訓練出現偏差；再者，文獻中的電池大多是處於實驗室特定參數設置下規律地操作，與大多數上路電池經歷的操作歷程相當不同，訓練出的模型是否能處理使用歷程複雜的電池值得懷疑；最後，大部分的文獻關注探討鋰電池的性能和應用，近年少有針對鎳氫電池之健康狀態的預測和研究。但參考市場機構的調查，鎳氫電池以其突出的安全性依舊能在新能源交通工具中扮演重大的角色，因此發展出一套系統性判斷電池健康狀態，以利廠商分級回收再利用是很重要的事情。 我們也將應用人工智慧進行本研究。首先，我們透過詳盡的鎳氫電池文獻回顧，了解鎳氫電池的運作機制和衰退成因，以及其健康狀態的定義及量測方法，決定用恆流充放電(galvanostatic charge/discharge, GCD)檢測電池性能。我們認為充電曲線對電池性能高度相關，並據此設計並選擇了5個特徵。之後，歸納常見的機器學習模型並選擇貝葉斯山脊型迴歸(Bayesian ridge regression, BRR)、前饋神經網路(feedforward neural network, FNN)和隨機森林(Random forest, RF)進行訓練、測試，以很小的誤差準確預測電池之性能。為了確認特徵的泛化性能(generalizability)，我們執行了兩組老化實驗，驗證選擇的特徵在不同環境下產生的數據上依然適用，最後輔以遷移學習(transfer learning)，縮小不同數據中特徵的差異，進一步地縮小預測誤差。

Along with the aggravating climate change and rising environmental awareness, sales of hybrid electric vehicles (HEVs) and electric vehicles (EVs) are increasing. The HEVs in particular, are commercialized by the end of last century and more than 10 million of them are operating around the world. Since nickel metal hydride (NiMH) batteries are the core energy storage component on board of HEVs, a significant amount of retired NiMH batteries will flood the world, requiring proper treatment. In the past, battery testing required a complete charge and discharge process to measure the actual charging and discharging capacity of the battery, and determine theirs application, to be on service or be discarded afterward. However, the complete charge and discharge process is time-consuming, not suitable for dealing with the upcoming massive battery recycling wave. Additionally, there is a desire to predict the future usage conditions and changes in battery performance, based on the battery's aging status, to arrange appropriate usage methods and extend the service life to save the cost and resources of battery manufacturing. Therefore, a fast and accurate method to obtain the actual performance of batteries and predict their future state changes is desired. Due to the popularization of technology and rapid advancements in data science, artificial intelligence (AI) has been applied in various industries to improve production efficiency and social welfare. Previous studies have applied AI to predict the state of health (SOH) of batteries, achieving accurate identification of battery conditions and excellent results. However, the training samples are usually small, a few hundreds or even dozens. The small sample size may lead to bias in training. Moreover, the batteries in the literature are mostly operated under specific laboratory parameters, which are quite different from those on-road batteries, doubtful whether the trained models can handle batteries with complex usage history. Lastly, most of the literature focuses on the performance and application of lithium batteries, with limited research on the prediction of the state of health of nickel-hydrogen batteries, which is however, still play a significant role in new energy vehicles due to their outstanding safety. Therefore, it is important to develop a systematic approach to assess the health status of batteries, allowing manufacturers to classify and recycle them for reuse. We also applied AI techniques in this study. Firstly, through a comprehensive literature review on NiMH, we gained an understanding of their operating mechanisms, degradation factors, and the definition and measurement methods of their SOH. We used galvanostatic charge/discharge (GCD) to assess battery performance. Recognizing the high correlation between charge curves and battery performance, we designed and selected five features. Subsequently, we summarized common machine learning models and chose Bayesian ridge regression (BRR), feedforward neural network (FNN), and random forest (RF) for training and testing. These models accurately predict battery performance with minimal errors. To confirm the generalizability of features, we conducted two aging experiments, validating that the selected features remain applicable in different environments. Finally, employing transfer learning helped narrow feature differences across various datasets, further reducing prediction errors