近藤 稔 氏 - 公益財団法人鉄道総合技術研究所
矢野 達哉 氏 - 株式会社 テス
The Railway Technical Research Institute ('RTRI') is an organization that began as the research division of the former Japanese National Railways (JNR). It was established following the division and privatization of JNR, when the JR Group was founded. The RTRI currently deals with a wide range of themes, including technical problems presented by railway operators belonging to the JR Group. RTRI actively carries out research and development (R&D) into technology that will further advance the railway field overall.
The RTRI Drive Systems Laboratory Vehicle Control Technology Division is involved in research activities relating to the drive components fitted to trains and diesel vehicles (hereafter, 'drive systems'). Specifically, in the case of trains, this refers to components such as the motor, coupling, and gear equipment, while in the case of diesel vehicles, this includes engines, the transmission, propeller shaft, and gear equipment. Currently, the Laboratory is involved in the status monitoring of drive systems fitted in diesel vehicles.
For example, research targeting condition monitoring for wheels, axle bearings, ground equipment such as rails and overhead catenary is already being conducted. The objective of this research is to continuously monitor each respective state to detect faults and facilitate accurate repair work. Alternatively, there is also a need for the early detection of signs that may lead to faults so that, based on this information, the equipment lifespan can be predicted and effective maintenance carried out accordingly. This is the concept known as 'predictive maintenance.' However, condition monitoring R&D in the field of drive systems is not yet very advanced. Trains usually comprise multiple carriages to form 10- or 12-car configurations. In these cases, a single train will be equipped with many motors overall. In actual fact, the train will continue to be able to run without problems even if a minor fault occurs in one of these motors, because this redundancy has been built into the design. Furthermore, based on past results, it would seem that drive systems do not break easily. However, there are some cases, such as on local railway lines, where a single-carriage diesel vehicle may be operating. In such cases, the train may be prevented from running if even a single engine suffers a fault, which could lead to stoppage of all services on that line. In order to prevent this kind of situation, and as a general consideration for enhancing safety and reliability, there is an existing need to be able to monitor the status of drive systems.
The RTRI was actually approached by the Maintenance Department of a certain railway operator to conduct research and development into status monitoring for diesel vehicles. This resulted in the decision by RTRI to pursue R&D in this field.
The objects being monitored under this proposal were drive systems such as diesel engines and transmissions. However, these drive systems are rarely destroyed unexpectedly. Rather, they tend to deteriorate over time, breaking only once a certain threshold has been reached. Accordingly, the condition monitoring system would need to be able to grasp this gradual deterioration. The railway industry operator who had approached the RTRI to conduct R&D wanted to be able to remotely monitor the status of all diesel vehicles under their jurisdiction. Being a request from the Maintenance Department, one definite objective could have been to achieve 'Predictive Maintenance.' Predictive maintenance aims to track the level of deterioration of components and units that are known to degrade over time, in order to predict their remaining lifespan and implement maintenance at the appropriate timing. One benefit of this concept is the reduction in maintenance costs. However, our actual research objective was not predictive maintenance, but rather, the detection of abnormalities. The reason for this was that drive systems, which were the focus of this project, are not typically prone to deterioration, nor are they known to be easily destroyed. Regardless, it is still possible that a fault could occur due to some unforeseen circumstance, so status monitoring is performed in order to detect abnormalities and facilitate prompt action accordingly in the event that a fault is detected. Note that predictive maintenance only becomes possible once the three-pronged process of error detection, diagnosis and lifespan prediction is achieved. This particular project was focused on establishing technology for error detection, with a view to the possible development of predictive maintenance in the future.
Figure 1 illustrates an example of what a condition monitoring system might look like in the future. This system firstly detects vibration being emitted from the drive system using vibration sensors. The acquired vibration signals are sent to the condition monitoring device and analog/digital (A/D) conversion is performed. The condition monitoring device then applies octave band analysis to this digital data. Octave band analysis is a technique that is often used to evaluate noise and vibration. The octave band is the frequency bandwidth at which the frequency ratio between the upper and lower limits is one octave (double). In other words, after analog/digital conversion, data is divided up into each octave band, and the power of the vibration in each respective band is calculated (Figure 2). The results obtained in this manner are then forwarded to the server. This data is acquired by accessing the server via a PC, after which it is analyzed for abnormalities using a diagnostic program. This diagnostic program is equipped with machine learning algorithms, and error detection is carried out based on comparisons with regular status operations (Figure 3). The development of this diagnostic program was at the heart of our research. The program was written in Python and was achieved with the use of the machine learning "scikit-learn" library.
The biggest issues we encountered when building this kind of system were the sensors used and the robustness of the equipment. In other words, we needed components that would be capable of tolerating extreme conditions, such as the vibration and temperature of the carriage during actual operations. Train motors and diesel vehicle engines rarely break down. They are tough enough to withstand collisions with ballast (gravel scattered on the train tracks) as well as ice and snow. It's clear that components such as the equipment, sensors, and wiring used in the status monitoring systems are more fragile than the drive systems themselves. Although you would be unlikely to install a condition monitoring device in a location that could be subjected to flying pieces of gravel, there is no doubt that this would be an undesirable place to install measuring equipment. Furthermore, the sensors and wiring are exposed to fairly harsh conditions. Accordingly, the components chosen must be as robust as possible. From a cost perspective, this is disadvantageous.
NI's CompactRIO was a great choice as a product able to withstand this type of harsh environment. In other words, it was highly rated as a robust product capable of tolerating vibrations and temperature changes when used in a condition monitoring device. By choosing CompactRIO, we were also able to use NI's graphical development platform, LabVIEW, to program all required functionality. Ultimately, we chose this combination for all the reasons described below.
Firstly, if the system is installed in an actual operating carriage, we wanted to be able to avoid any need for tasks such as complex switch operations. In other words, we wanted to build a mechanism that would enable measurement to be started and stopped automatically. CompactRIO would enable us to use LabVIEW to easily achieve this kind of process, with excellent flexibility. It could be said that NI products offer a streamlined solution that integrates the measurement devices with the PC. In other words, the PC can be used to operate the measurement devices in any way the user wishes. It is unlikely that anything capable of rivalling NI products exists. In fact, although we did find an industrial PC-based automated measurement product from another manufacturer, the connection to the vibration sensors and data acquisition components seemed troublesome. Conversely, CompactRIO makes it extremely simple to connect the sensors and import data. Furthermore, a field-programmable gate array (FPGA) can also be used to achieve extremely high-speed processing. For all these reasons, we chose to use NI products.
Figure 1 shows this system in practical use. There is no need for communications functionality with the server at this R&D stage - it is sufficient to be able to acquire data using the status monitoring device. We also connected a vibration sensor to CompactRIO to collect data from the sensor, and used CompactRIO to apply octave band analysis. The analysis results were then saved to an SD card. Once a certain volume of data was accumulated, the SD card was retrieved and the diagnostic program on the PC was used to determine whether there were any indications that could lead to faults occurring. Note that although CompactRIO is highly resistant to fluctuations in vibration and temperature, the range of voltage fluctuations in the carriage's power source environment was extremely large, to the extent that not even CompactRIO was able to tolerate it. For this reason, we built the condition monitoring device by combining it with a power module specifically for railway carriages (Figure 4).
TESS Co., Ltd. was tasked with installing this system monitoring device. This corporation was established to support the research and development carried out by RTRI. Mr. Yano, the TESS staff member in charge of this development, had not previously been tasked with any programming duties during his career to that point. Over the course of approximately one year, he used e-learning to study programming techniques using LabVIEW, and once completed, he was placed in charge of the development of the condition monitoring device. In other words, based on our design, Mr. Yano set about performing tasks such as the wiring, and then packaged the other components including CompactRIO and the power module into a single unit. He also achieved all required functionality by using LabVIEW for the programming. As a result, he was able to complete the first model status monitoring device in just six weeks. There is no doubt that the ability to use LabVIEW's graphical operations for programming contributed to the short development time. For example, the textbook used in the e-learning was extremely useful for the installation tasks, and plenty of sample programs were provided that were able to be used subsequently without any further modification. Furthermore, LabVIEW is also equipped with a library for octave band analysis, making it extremely easy to achieve the required functionality. In fact, a wide range of tasks can be performed simply by combining the sample programs with the library. When LabVIEW is used, there is no need for text-based coding for every single function, so tasks can be developed with great peace of mind.
Since the development of the first model was completed, various improvements and modifications have been made to enable data such as the revolution speed or the load to be obtained, in addition to being able to measure vibrations. As a result, the condition monitoring device was fitted to an actual carriage to obtain vibration data, and the results of the octave band analysis were able to be accumulated. We were also able to confirm that abnormalities could be determined using the diagnostic program as per our original objective. Specifically, we are now able to use a single algorithm to detect abnormalities in the various machines included in the drive system. There is no need to perform tasks such as tuning for each individual machine or fault. If the fault is of the type that displays changes in vibration, it should be able to be detected in any form. However, simply detecting the presence of an error will not inform the department in charge of maintenance about where the problem has occurred, so they will not know how it should be dealt with. To counteract this, we were able to demonstrate that the changes appear in different frequency bandwidths according to the content of the error. This should now make it much easier to deal with the faults.
Generally speaking, the following point of concern is expected to arise when monitoring vibration in vehicles. The vibration status will differ according to the location, so there is a concern that it might be difficult to make accurate diagnoses based on the data acquired. In actual fact, for this reason, there are some cases that use a method of monitoring only vibrations that occur when operating in a particular section of track in order to determine whether there are any abnormalities. However, the condition monitoring device that we developed does not do anything like this. There are several reasons for this, but the most important one is that machine learning is used. Even when the vibration status differs depending on the section of track where the vehicle is running, if the equipment can be taught to absorb sufficient learning data between each respective section of track, it should be possible to create a mechanism that can detect the occurrence of abnormalities whenever there are any vibrations that differ to normal operations. Accordingly, location-dependent vibrations are not currently a major issue.
The current approach is to identify any signs of a fault as quickly as possible before the machine breaks down completely. We are proud to say that we are generally already achieving this from a technology point of view. However, it is also true that we still have a long way to go before we reach our ideal situation. Accordingly, we hope to pursue the following list of goals in the future.
Firstly, a huge volume of data will be accumulated if actual vehicle data continues to be measured on a daily basis. Also, changes to vibration will occur based on any changes to conditions such as the temperature or the wear status of the wheels. In order to perform machine learning across such a great variety of conditions, an extremely large volume of data must be input and learned. No matter how powerful the computer is, this volume of data will still cause problems, so we are currently researching algorithms capable of extracting only the necessary characteristics from within all the data.
Once an abnormality is detected, we also need to find out at what stage that error occurred. In other words, we need to know whether the train service will need to be stopped urgently so that the relevant component can be replaced, or if it is acceptable to continue using that component for a while longer without issue. We are beginning to understand that we might be able to identify the degree of deterioration of a certain component by evaluating the ratio of abnormal vibrations per day.
In addition, we are trying to detect abnormalities by measuring elements other than vibration. Vibration sensors are expensive, so one motivating factor is the desire to use other less-expensive sensors. For example, one method we are considering is focusing on the voltage or current of an object and then analyzing the noise emitted to determine if an error has occurred. For the analysis method, we think that we can probably use a similar algorithm to the one used to analyze vibration. Naturally, we hope to expand the scope of application to include drive systems for trains, and not just for diesel vehicles.
Furthermore, the condition monitoring system performs to the extent of octave band analysis, while the diagnostic program used to detect the presence of abnormalities is operated on the PC. However, if this system is put into practical use, it will be operated by staff from maintenance departments, so we hope to be able to detect abnormalities on the vehicles in close-to-real time. We would like to be able to detect abnormalities at the condition monitoring system-side and obtain an indicator of the degree of severity of the error as a result. This indicator can then be sent on its own to the server. In other words, we want to equip the status monitoring system with machine learning algorithms to calculate the degree of abnormalities. Note that the massive computational load created by machine learning occurs only at the learning stage. In the actual application, the model used would be obtained post-learning, so the computational load would be minimal. Accordingly, execution using CompactRIO should also be possible.
As a final note, the biggest obstacle to putting this system into practical use is the cost. As previously explained, drive systems are not easily broken, and in most cases, the train can continue to run even if one drive system breaks. Even if a single-carriage diesel vehicle is suddenly no longer able to operate, it will not be that great a problem in monetary terms. Based on these conditions, placing sensor and status monitoring devices on every drive system in every carriage is probably not going to offer the greatest cost-benefit performance. Drive systems are not initially designed to deteriorate over time, so highly frequent maintenance is unlikely to be performed in the future. Accordingly, the cost reduction benefits typically associated with predictive maintenance will be limited. Regardless, status monitoring continues to garner attention probably because of the fact that the cost of systems needed for information processing has been dramatically decreasing. For example, it used to be necessary to spend several million yen just to purchase equipment capable of performing FFT (Fast Fourier Transform) algorithms. However, these days, a sensor with built-in FFT functionality can be purchased for as little as approximately 50,000 yen. Status monitoring systems definitely enable the user to cope with unforeseen circumstances, as well as improve safety and reliability, so I sincerely hope that the price of these compositional elements will further decrease in the future.
近藤 稔 氏