德伯爾，孔慶福，祝 劍

?

船用蓄電池組健康管理系統(tǒng)設計

德伯爾，孔慶福，祝 劍

（海軍工程大學，中國武漢 430033 ）

蓄電池是一種重要的船用能量源，比如當潛艇處于水下時。因此，保持蓄電池組良好的技術狀態(tài)非常重要。蓄電池性能劣化的機理非常復雜，船員往往難以掌握。為實現(xiàn)對船用蓄電池組的科學使用和管理、保證其工作可靠性，本文設計了一套船用蓄電池組健康管理系統(tǒng)。系統(tǒng)由一個狀態(tài)監(jiān)測子系統(tǒng)和性能預測子系統(tǒng)組成。為實現(xiàn)蓄電池組性能預測功能，構建了基于支持向量機的智能預測模型，通過歷史監(jiān)測數(shù)據(jù)自動尋找電池組的性能劣化規(guī)律。測試結果表明了系統(tǒng)性能預測功能的有效性。

蓄電池 健康管理 監(jiān)測數(shù)據(jù) 支持向量機 性能預測

0 引言

在艦船上，通常將單個蓄電池組合成蓄電池組的方式來給各類船用設備供電。當潛艇處于水下工況時，蓄電池組作為唯一的能量源，是潛艇上最重要的設備之一。蓄電池組技術狀態(tài)的好壞對于潛艇各類任務的順利完成具有極為重要的影響。如果船員能夠實時掌握蓄電池組的實際技術狀態(tài)，并以此為依據(jù)實現(xiàn)對蓄電池組的科學使用和合理維護，則可有效保證蓄電池組的工作可靠性和安全性。然而，船用蓄電池組無論是內部電化學反應、還是外部的工作環(huán)境都非常復雜，使得基于人工的船用蓄電池組健康管理工作往往變得非常困難[1-3]。有鑒于此，本文開發(fā)了一套船用蓄電池組健康管理系統(tǒng)（HMS）。系統(tǒng)主要由一個在線監(jiān)測子系統(tǒng)和一個性能預測子系統(tǒng)組成。其中，在線監(jiān)測子系統(tǒng)負責在線監(jiān)測數(shù)據(jù)收集、信息顯示和超限報警等任務；性能預測子系統(tǒng)采用了基于最小平方根-支持向量機(LS-SVM)的預測模型。

1 船用蓄電池組健康管理系統(tǒng)設計

1.1 船用蓄電池性能劣化機理

作為艦船用蓄電池，通常要求有較大的功率輸出，鉛酸蓄電池因具有相對較大的功率輸出而在艦船上得到了廣泛的應用。鉛酸蓄電池通過一個可逆的電化學反應來實現(xiàn)電能的存儲和提供。在充電過程中，蓄電池的陽極被氧化而產(chǎn)生電子，陰極消耗電子；蓄電池放電過程則相反。蓄電池在每次的充放電過程中，其內部都會產(chǎn)生一定的不可逆化學反應，從而對蓄電池的性能造成不可避免的損害。如果在使用中還存在過充或過放等不正確操作，蓄電池性能的下降將會更顯著。即使使用中沒有不正確操作，由于部件老化等原因，蓄電池的容量也會隨著充放電次數(shù)的增加而下降。上述因素都會造成蓄電池的性能劣化，從而導致蓄電池組總體性能的退化[4-5]。

1.2 船用蓄電池組的健康管理策略

盡管大多數(shù)情況下蓄電池組性能退化往往是一個連續(xù)的發(fā)展過程，但由于外部使用環(huán)境和內部電化學反應的復雜性，蓄電池組性能退化范圍通常有較大的差別。當蓄電池性能劣化時，可觀察或測量到如下的現(xiàn)象：電池電壓下降、充電時間變長、電池內部電阻變大、電池溫度變高。也就是說，上述信息可用來評估蓄電池性能劣化的趨勢和程度，只是鑒于船用蓄電池組工作參數(shù)與實際技術狀態(tài)之間存在的復雜映射關系，找到一個能正確評估蓄電池實際技術狀態(tài)和預報性能發(fā)展趨勢的有效方法具有較大的難度。實踐證明，單純依靠人工進行船用蓄電池健康管理具有較大的難度。人工智能技術是解決此類復雜問題的有力工具，例如，基于最小平方根-支持向量機(LS-SVM)的學習算法已被證明能自動地從機械設備歷史工作數(shù)據(jù)中學習、并具有對數(shù)據(jù)進行有效分類的能力[3, 6]。

基于前述討論，本文設計的船用蓄電池組健康管理系統(tǒng)的健康管理策略和信息處理流程如圖1所示。系統(tǒng)實時監(jiān)測的蓄電池工作數(shù)據(jù)有4類：單體蓄電池電壓、蓄電池組電壓、蓄電池組電阻、蓄電池組溫度[5, 7]。系統(tǒng)采集到蓄電池（組）工作數(shù)據(jù)后，首先執(zhí)行超限報警判別；隨后，相關工作數(shù)據(jù)被送至一個基于最小平方根-支持向量機(LS-SVM)算法的預測模型用來進行蓄電池組性能發(fā)展預測。

1.3 船用蓄電池組健康管理系統(tǒng)組成

船用蓄電池組健康管理系統(tǒng)的基本組成如圖2所示，主要由一個在線監(jiān)測子系統(tǒng)和一個故障預測子系統(tǒng)組成。

1.3.1 在線監(jiān)測子系統(tǒng)

潛艇上通常配置有許多的蓄電池組，各蓄電池組分別布置在不同的位置；基于這一實際情況，在設計中采用了一個分布式監(jiān)測子系統(tǒng)來對所有蓄電池組的工作參數(shù)進行采集。

圖1 健康管理系統(tǒng)的信息處理流程

實時監(jiān)測，在各蓄電池組內部分別配置了相應的傳感器。通過一套無線通信模塊，所有監(jiān)測數(shù)據(jù)都被送到監(jiān)測計算機處。在監(jiān)測計算機中，各數(shù)據(jù)首先分別與其設定的報警值進行比較，如果超出報警值，系統(tǒng)將給出相應的報警信號；隨后，數(shù)據(jù)以一定的存儲間隔和數(shù)據(jù)結構形式被存儲在監(jiān)測計算機硬盤之中，并以數(shù)據(jù)和圖表的形式在顯示器中顯示出來。

1.3.2 性能預測子系統(tǒng)

在系統(tǒng)中采用了一個內含基于最小平方根-支持向量機(LS-SVM)算法預測模型的管理上位機來執(zhí)行蓄電池性能發(fā)展預報功能。在系統(tǒng)進行性能預報之前，先要對預測模型進行有效的充分訓練。在進行蓄電池性能預測時，在線監(jiān)測的工作數(shù)據(jù)和存儲在計算機硬盤中的最近歷史工作數(shù)據(jù)首先被送往上位機中的性能預測模型，預測模型基于對歷史工作數(shù)據(jù)的學習和辨識，能夠自動給出蓄電池（組）電壓、電阻和溫度參數(shù)的變化趨勢，并提供給船員蓄電池（組）可能的下次故障發(fā)生時間點，以協(xié)助船員進行正確的使用管理和設備維護。

圖2 健康管理系統(tǒng)的基本組成

2 基于LS-SVM的預測算法

最小平方根-支持向量機(LS-SVM)算法是一種有監(jiān)督的自學習算法，在其使用前須進行有效和充分的訓練。假設訓練集為：

以及

為解決非線性分類問題，核空間非線性匹配函數(shù)的目標從原始空間挑選出合適特征量、與訓練數(shù)據(jù)進行匹配以形成一個高維特征向量空間。為得到優(yōu)化函數(shù)，可定義如下的Lagrangian函數(shù)：

已證明，Karush-Kuhn-Tucker (KKT)條件對于非線性優(yōu)化問題的尋優(yōu)解法是充分和必要條件[8,9]?；贙arush-Kuhn-Tucker (KKT)條件[8,9]可得到如下的方程：

有不少函數(shù)可以被當作核函數(shù)，其中，徑向基（Radial Basis Function - RBF）函數(shù)是支持向量機（SVM）算法中最為廣泛采用的。徑向基核可以用下式描述：

3 性能預報測試結果

性能預報功能是船用蓄電池健康管理系統(tǒng)的核心功能。為驗證其有效性，采用了一個測試來檢驗其預測結果的有效性。在系統(tǒng)測試中，選取一套已在潛艇上服役了一段時間的蓄電池組作為測試對象，并通過如下步驟完成測試過程：

1) 收集該蓄電池組包括電壓、溫度和電阻在內的歷史監(jiān)測數(shù)據(jù)。

2) 對所收集的歷史數(shù)據(jù)進行預處理，包括無效數(shù)據(jù)清除和數(shù)據(jù)標準化等工作，最后形成訓練數(shù)據(jù)集。

3) 基于訓練數(shù)據(jù)集，采用LM-SVM 算法對預測模型進行有效訓練。

4) 向預測模型提供不同于訓練數(shù)據(jù)集的其它輸入數(shù)據(jù)，獲得預測結果。

5) 在潛艇蓄電池組工作一段時間后獲得相關電池組的實際工作數(shù)據(jù)。

6) 將預測模型輸出預測數(shù)據(jù)與實際數(shù)據(jù)進行比較。

圖3和圖4分別展示了預測數(shù)據(jù)與實際數(shù)據(jù)的部分比較結果。其中，圖3展示了蓄電池組電壓的比較結果，圖4展示了蓄電池組溫度的比較結果。從圖示比較結果可知，基于LM-SVM 算法的預測模型具有令人滿意的預測能力。

圖3 蓄電池組電壓預測值與實際值比較結果

圖4 蓄電池組溫度預測值與實際值比較結果

4 結論

在實際使用過程中，船用蓄電池的性能會逐漸劣化，從而給蓄電池的工作可靠性和安全使用帶來消極影響。由于蓄電池性能劣化機理的復雜性，單純憑人工來預測蓄電池的性能發(fā)展趨勢比較困難。本文介紹了所開發(fā)的一套船用蓄電池健康管理系統(tǒng)，通過實時采集和無線傳送重要的蓄電池（組）工作參數(shù)到監(jiān)測計算機，實現(xiàn)工作參數(shù)的實時監(jiān)測、顯示和報警功能。為實現(xiàn)精準的蓄電池性能預報功能，構建了一個基于支持向量機的智能預報模型，使得系統(tǒng)能夠從蓄電池組的歷史工作數(shù)據(jù)中進行有效自學習，并能最終給出科學的性能趨勢預報。本系統(tǒng)的應用可以有效協(xié)助船員進行船用蓄電池科學維修決策。

0 Introduction

Storage batteries are usually combined as batteries to supply power to all kinds of devices equipped in a battleship. Batteriesare one of the most important devices in the submarine because it is usually the unique power resource while the submarine is under the water. Performance of the batteries plays an important role on the effectiveness of the submarine to finish its task. In consideration of working reliability and safety, it will be very helpful for the crew to master the actual health conditions of batteries so that decisions of scientific usages and rational maintenances can be made to the batteries. However, because of the complexity of both the external environment and the internal electro-chemistry reactions of the batteries, manual health management of the batteries has been proved to be a difficult work[1-3].

In order to solve the problem mentioned above, a Health Management System (HMS) for marine batteries is presented in the paper. The HMS consists mainly of two parts: an on-line monitoring subsystem and a performance prediction subsystem. The monitoring subsystem is designed to finish the task of on-line data collection, information display and alarm judgement. As to the key task of performance prediction, a prediction model based on Least Square-Support Vector Machine (LS-SVM) is built in the prediction subsystem.

1 Design on HMS for marine batteries

1.1 Performance deterioration of storage battery

It is usually required for the marine storage battery to supply relative larger output power, so lead-acid type rechargeable battery is widely used in the submarine. The battery stores or supplies electrical energy through a reversible electro-chemical reaction. During charging course of the battery, the positive active material of the battery is oxidized, producing electrons, and the negative material is reduced, consuming electrons. During discharging course of the battery, the process is reversed. However, during each charging or discharging course, there are usually undesirable and irreversible chemical reactions called cell reverse occurred inside the battery cell, which cause a permanent damage to the cell. If the battery is operated with mistreatment, such as over charged or over discharged, the situation will become worse. Even if the battery is used repeatedly without mistreatment, it loses capacity as the number of charging and discharging cycles increase because of aging effect of the components. All the factors mentioned above cause performance deterioration in the battery, which also results in a general degradation of the batteries[4-5].

1.2Health management strategy of batteries

Although the performance descending course of the battery is a continual developing course in most cases, but the descending tendency usually varies in a large range because of the complexity of the external usage environment and the internal electro-chemistry reactions of the storage battery. However, following phenomena can be observed and measured while the performance of the battery descends: voltage of the cell descends; charging time becomes longer; internal resistance will be enlarged; temperature of the battery becomes higher. That is to say, the information mentioned above can be used to judge the descending tendency of the battery performance. The problem is just how to find a way to give accurate performance judgement and prediction since the relationship between working parameters and performance conditions in a batteries is so complex. Manual health management for marine batteries has been proved to be a difficult work, but artificial intelligence is a powerful tool to solve this problem. For example, the LS-SVM algorithm has already been tested to have great ability to learn key knowledge from historical data[3, 6].

Based on the discussion mentioned above, health management strategy and information processing course of HMS for marine batteries are designed, as shown in figure 1. Four types of working parameters are measured on-line: voltage of the single battery, voltage of the batteries, electrical resistance of whole batteries, temperature of the batteries[5, 7]. Once the data are acquired, an alarm judgement is firstly completed. Furthermore, relative historical data are sent to a prediction model based on LS-SVM algorithm to give accurate prediction.

1.3 Construction of the HMS for marine batteries

Figure 2 shows the fundamental construction of the HMS for marine batteries. There are two primary parts: an on-line monitoring subsystem and a fault prediction subsystem.

1.3.1 On-line monitoring subsystem

There are many batteries equipped in a submarine, each batteries is placed in a suitable space. According to this factor, a distributed monitoring subsystem is designed to collect the working parameters of all batteries.

Fig. 1 Information processing course of the HMS

Sensors are equipped within the batteries so that the essential working parameters can be measured. By means of a set of wireless communication unit, all measured parameters are transferred to the monitoring computer. In the monitoring computer, the parameters are firstly compared with their alarming-setting values, if exceed the setting value, an alarm will be presented by the computer. Then, the data are saved with a certain sampling time and structure in the hard disk of the computer and then displayed by mean of data and diagram in the monitor.

1.3.2 Performance prediction subsystem

A managing computer is engaged in the HMS to perform the function of performance prediction. A model based on LS-SVM algorithm is built in the managing computer. The performance prediction model should be well trained before it is engaged to give performance prediction. While making prediction, the on-line monitored working parameters and recent historical data saved in the hard disk of the monitoring computer are firstly sent to the performance prediction model in the managing computer, and then developing tendencies of both voltage, resistance and temperature can be made automatically by the model based on the historical data. Estimated time points when faults maybe occur will be presented to the crew to help them to take correct maintenance operations.

Fig. 2 Fundamental Construction of the System

2 Prediction algorithm of LS-SVM

The LS-SVM is a supervised learning algorithm and should be well trained before it works. Suppose that the training dataset is:

Subject to:

In order to solve the optimal function, the Lagrangian function may be defined as:

Following equations can be obtained based on the Karush-Kuhn-Tucker (KKT) condition[8,9], which is a necessary and sufficient condition for the optimal solution of the object function in the problem of non-linear optimization.

Many functions can be adopted as the kernel function, the radial basis function (RBF) is one of the most popular kernel function for SVM. The RBF kernel can be described as the following formula:

3 Testing results of performance prediction

Performance prediction is the key function of the HMS for marine batteries. In order to test its validity, an experiment is completed to check the accuracy of the prediction results. A set of marine batteries is chosen as testing object, who has served in a submarine for a certain period. The test is completed with following steps:

1) Collecting the historical monitored voltage, temperature and resistance data of the batteries.

2) Pre-treating the collected data to build the training data set, including invalid data eliminating, data standardization and so on.

3) Training the prediction model with LM-SVM algorithm using the training data set.

4) Presenting input data different from the training data set to the prediction model and acquire the prediction results.

5) Acquiring actual working data from the submarine after a period of time.

6) Making a comparison between the prediction results and the actual work data.

Figure 3 and 4 show partial comparison results between the prediction results and the actual work data. Figure 3 shows the comparison result of cell voltage; and figure 4 shows the comparison result of cell temperature. It can be seen from the comparison results that the model based on LM-SVM algorithm has an excellent prediction performance.

Fig. 3Comparison result of cell voltage

4 Conclusion

Performance of storage battery usually deteriorates during usage, which brings negative effect on the using reliability and safety. It is difficult to master the performance developing tendency manually because of the complex deterioration mechanism. A health management system for marine batteries is introduced in the paper. Important parameters are collected and sent wirelessly to a computer to complete on-line monitoring function including parameters display and alarm judgement. In order to realize performance prediction, an intelligent forecasting model based on support vector machine is built, with which the system can learn knowledge from the historical data and make accurate prediction. With help of the system, maintenance decisions for marine batteries can be made scientifically.

Fig. 4 Comparison result of batteries temperature

//Reference：

[1] Wagner R. Failure modes of valve-regulated lead-acid batteries different applications. Power Sources, 2009, 53: 153-162.

[2] Lin Z C. The design of intelligent monitoring unit of aviation battery charge and discharge system. Chengdu: University of Electronic Science and Technology of China, 2010.

[3] Ji Y X, Du H J, Sun H. A survey of state of charge estimation methods. Electrical Measurement＆ Instrumentation, 2014, 51: 18-22.

[4] Li J H, Yan G G, Zhu Y. Design of VRLA battery-state monitoring system based on DSP2812. Electrical Measurement＆Instrumentation, 2012, 49: 85-88.

[5] Zhao Y B, He X N, Ding J S. Study on the storage battery real and dynamic intelligent monitor and control system. Electric Engineering, 2005, 12: 21-23.

[6] Doerffel D, Sharkh S A. A critical review of using the Peukert equation for determining the remaining capacity of lead-acid and lithiumion batteries. Journal of Power Sources, 2006, 155: 395-400.

[7] Ye D. Design of precise charge and discharge coulometer apply for storage battery. Electric Engineering, 2004, (4): 16-19.

[8] Nie Z Q. Research and Implementation of battery monitoring and diagnosis expert system. Beijing: Tsinghua University, 1998.Date of receipt：2017-11-20About the author：Aymen Derbel(1990-), Male, postgraduate, Professional direction: Marine Engineering.

Design on A Health Management System for Marine Batteries

Aymen Derbel, Qingfu Kong, Jian Zhu

(Naval University of Engineering, Wuhan 430033, Hubei, China)

TM912.1

A

1003-4862（2018）02-0001-07

2017-11-20

德伯爾（1990-），男，外國留學生。研究方向：輪機工程。Email: aymenamine10@yahoo.com