The Use of Self-Learning Systems to Solve the Problems of Finding Failures on the Railway Blagoveshchenskaya E.A., Sergey V. Bochkarev Emperor Alexander I St. Petersburg State LLC "GC IMSAT", Saint-Petersburg, Russia Transport University, bochkareffsv@yandex.ru kblag2002@yahoo.com Peter the Great St. Petersburg State Polytechnic University Nikolay V. Gruzdev Denis V. Zuev LLC SvyazStroyServis, Saint-Petersburg, Russia LLC "GC IMSAT", Saint-Petersburg, Russia Nik_gru@mail.ru zuevdv@gmail.com 1 Introduction The share of railway freight turnover in the transport system of Russia is 45%, while the growth of freight traffic of Russian Railways continues, in Abstract 2018 it amounted to 2596.9 billion ton-kilometers. By order of the government dated March 19, 2019 The most frequent reason for the violation No. 466-r, a long-term development program for of the train schedule is the failure of the Russian Railways until 2025 was approved, which technical facilities of the infrastructure provides for the transition to a digital railway. In complex. The number and duration of train accordance with this, the active development of downtime and, as a result, the economic technical diagnostic and monitoring systems losses of Russian Railways depends on the (STDM) continues. time of search and elimination of failures. Technical diagnostics is the determination of the Today, failure search is carried out in an technical condition of an object [Efa12]. The object intuitive way. In practice, such a path leads of railway automation and telemechanics can be in to unnecessary time costs.In the study of one of the following states [Efa12] (Figure 1): methods for constructing algorithms, they 1) Intact - this is the state of the facility, during were classified in terms of the possibility of which it meets all the requirements established in constructing a model in which each step is a the normative and technical documentation for it. function of all previous steps and the 2) Faulty - the state of the facility in which it functions cover the entire space of failures. does not meet at least one of the requirements We built these functions and described the established in the normative and technical generation model of such functions. Such a documentation for it. model for constructing functions can be 3) Workable - the state of an object in which the applied to any technical branch. For the values of all parameters characterizing the ability to railway infrastructure, 6 functions were perform specified functions meet the requirements obtained. An automatic self-learning system established in the regulatory and technical has been developed, which is a multilayer documentation for this object. neural network constructed according to a recurrence model. Based on this model, a hardware-software complex was implemented and tested. 1 Copyright c by the paper's authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). In: A. Khomonenko, B. Sokolov, K. Ivanova (eds.): Selected Papers of the Models and Methods of Information Systems Research Figure 1. The technical condition of the Workshop, St. Petersburg, Russia, 4-5 Dec. 2019, published at facilities http://ceur-ws.org 21 4) Inoperative - a state in which the value of at number of pre-failure conditions accumulate, since least one of the parameters characterizing the any minor changes in the diagnostic parameters (for ability to perform specified functions does not meet example, voltage) are noted by the system [Kal09, the requirements in the normative and technical Efa12] (Figure 2). The fixation of “false” pre- documentation for this object. failure conditions can lead either to a failure, which 5) Failure - the state of the object, characterized by can cause a violation of the safety and uninterrupted an increased risk of its failure [Efa12, Boc12]. operation of trains, or to a “false” response of Failure - an event consisting in the violation of the service personnel, which will lead to an increase in operational state of the object. In the absence of labor costs. logical analysis and analytical forecasting, a large Figure 2. Distribution of pre-cancers resulting from STDM deficiencies To improve quality and reduce troubleshooting With an increase in cargo turnover, the time, monitoring systems for railway automation throughput of railways should be increased, which and telemechanics are being actively developed. requires serious resources. In 2018 alone, about 300 Directions for the development of technical billion rubles were spent on the development of the diagnostics and monitoring of HEAT: railway infrastructure of Russian Railways. But all 1) A full range of measurements. investments are leveled in cases of technical 2) Analysis of the operation of devices. equipment failure. 3) Forecast of changes in the status of devices. The most common cause of train schedule 4) Issuing service recommendations devices. disruptions is the failure of infrastructure facilities. 5) Coverage of all devices with diagnostic tools. The number and duration of train downtime and, as To date, the disadvantages of technical diagnosis a consequence, the economic losses of Russian include: - “manual” processing of diagnostic Railways depend on the search and elimination of information, which leads to an increase in the time failures. for its analysis and decision making; - lack of To date, the search for failures is carried out identification of the reasons for the failure; - lack of intuitively. In practice, this way leads to evidence-based methods for fixing pre-failure unnecessary time costs. conditions; - lack of troubleshooting algorithms built into STDM; - lack of an optimal set of 2 Failure search process controlled diagnostic parameters, which leads to low reliability of determining the technical The analysis of the revealed failures shows that condition. their main reason is a violation of the technological Therefore, expanding the functionality of systems process of operation (operational failures). for identifying the causes of failure, determining the Operational failures account for up to 86% of all optimal algorithm for troubleshooting and failures of determining the optimal set of monitored railway AT devices [Ana12, Ana09, Sap02]. parameters is an urgent task. According to statistics, most of the operational failures cause train delays. 22 Despite a number of measures to increase the improve its quality, and this is only possible by reliability of railway AT devices, which include increasing the level of its (technological process) scheduled preventive maintenance, the organization automation. of new maintenance methods [Aks09], and training The troubleshooting process takes place in the of maintenance personnel rules and troubleshooting following sequence: After the failure information methods, the time to search for localization and appears, preparatory steps for troubleshooting are troubleshooting remains relatively large. started: identification of the failed device, The long time to find and eliminate the failure is collection and analysis of additional information, explained by a number of objective and subjective study of technical documentation, analysis of the factors. Objective factors include territorial train situation to localize the location of the failure dispersal, difficult access to some outdoor signaling and its nature, then collection of necessary tools devices, and sometimes lack of complete technical and materials. Delivery of the employee to the documentation. Subjective factors include the lack place. Additional checks and elimination of the of experience and qualifications of the service detected malfunction are carried out on site (Figure personnel of the distance, the inability to read the 3). schematic and wiring diagrams. Reducing the influence of the human factor on the technological process is a necessary measure to Informing Preparing to Arrival Search Troubleshooting search - from - afoot; chipboard; - determination of the place and - transport. - measurements, - from nature of the failure; - replacement of faulty inspections and their SHCHD. - determination of the failed elements of the device, analysis. device; adjustment; - collection and analysis of - check the working diagnostic information, tools and good condition. materials; - study of technical documentation. Figure 3. The process of recovering the operability of the device But there are situations that an electrician has t’av.p - troubleshooting time taking into account difficulty finding a malfunction, for example, he STDM (14 min). ttz - technical delay (5 min). incorrectly localized the location of a malfunction It is worth noting that the preparatory period is the search: he searches for a malfunction in the field longest from the beginning of the failure to its when the malfunction is in the relay room. . In such elimination. The safety and security of the railway situations, the role of automated algorithms is transport depend on the knowledge and skills of the increasing, which systematize and structure employee. Therefore, we have directed efforts to troubleshooting. minimize the human factor. Let us compare the recovery times of the function The ideal model for eliminating the problem of of operability of devices without and with Heinrich Saulovich Altshuller suggests that the automated algorithms for troubleshooting. problem should be eliminated by itself, Average recovery time of TV function of automatically. To solve the problem in this way, operability of devices of railway automation and automatic troubleshooting is implemented. telemechanics systems in 1 region on the Moscow Railway: 3 Failure search algorithm Tv = topv + tav.p + tav.u + ttz = 40 min. where topv is the time of notification of a There are 2 ways to compile a failure search malfunction (1 min); algorithm [Bla19]: tav.p - average time for troubleshooting (20 1) to paint all possible algorithms and min); depending on the input data to give the optimal tav.u - average time for troubleshooting (9 min); search path. Although there are a finite number of ttz - technical delay (10 min). such algorithms, from a practical point of view it is The development and implementation of automated large and such a path is not reasonable. troubleshooting algorithms minimizes the value of 2) The second way is to build a dynamic, tav.p. when this time tends to a minimum, the ratio recurrent model, where each step will be a function will take the following form: of all previous steps. But to build such an algorithm Tv = topv + t’s av.p + tav.u + ttz = 29 min. for all cases is also not optimal. Therefore, we 23 classified all types of algorithms from the point of classification problem in the case of many classes, view: the possibility of constructing such functions it is recommended to use nonlinear classifiers and covering the functions of the entire failure [Kru01], i.e. classifiers that use a nonlinear surface space. We built these functions and described a to separate classes. An example of such classifiers model for generating such functions. For the is a neural network. railway infrastructure were obtained 6 functions. A neural network is a distributed parallel processor Such a model for constructing functions can be consisting of interconnected elementary applied to any technical industry. information processing units (neurons) that In general, classification methods [Dmi86, accumulate experimental knowledge for their Zor07] can be attributed to the following types: subsequent processing [Kru07, Cal01]. Neurons are 1. The linear classifier method. implemented by a nonlinear function of one 2. The method of nonlinear classifier. argument-the weighted sum of all input signals. 3. The method of constructing decision trees. This function is called the activation function. The The linear classifier [Kal09] allows one to set of interconnected neurons determines the determine the linear dividing surface. In the case of structure of the network and the tasks that the two classes, such a surface is a hyperplane dividing neural network is able to solve. The weights that the space of attributes into two half-spaces. Linear characterize the strength of the connection between classifiers include the support vector method, two neurons are called SYNOPTIC coefficients. Bayesian classifier, and other methods. The process of selecting SYNOPTIC coefficients is In the method of support vectors [Kly87, Par13], a called network learning [Cal01, Hai06]. In the set of training examples is proposed for each state, process of training, an array of input values (class given as points in multidimensional space. These attributes) is mapped to each class. Neural networks points form regions in space corresponding to are able to generalize information obtained during different classes. The extreme points of a class are training. Also, the advantage of using neural called reference points, and the distance between networks is the absence of the need to adjust the the two reference points is the length of the algorithms when changing the number or reference vector. It is required to find such a characteristics of classes. The disadvantage of using hyperplane that the length of the support vectors is neural networks can be a large computational maximal. For the application of this method complexity [Sim94] when using complex network requires that the classes were linearly separable structures (for example, convolutional neural among themselves. The disadvantage of the method networks [Cal01, Hai06, Sim94]). is that it is suitable for solving the classification The method of constructing decision trees allows to problem with only two linearly separable classes. construct a visual algorithm of object classification. To solve a problem with a large number of classes, The decision tree consists of nodes (also called the division of the problem into subtasks of vertices) and branches connecting the nodes. The classification according to the scheme "one-against- very first node is called the root of the tree, and the the-others" is used. It is necessary to solve the extreme nodes are called leaves. Each vertex is problem of combining the results. mapped to some characteristic that describes the The Bayesian classifier [Sap04] is a method based object, and the branches-the value area of this on the theorem stating that if the density of the characteristic. The procedure for constructing a distribution of each class is known, then the decision tree is an iterative process in which the required algorithm can be written out in an explicit sign that best satisfies a certain branching criterion analytic form. For each of the classes, likelihood is selected for the next vertex of the tree [Hai06]. functions are defined, by which the a posteriori The branching criterion is selected depending on probabilities of the classes are calculated. The the algorithm used. Popular algorithms for building object belongs to the class for which the a posteriori a decision tree are ID3 [Hai06, Sim94] (or its probability is maximal. As a rule, in practice, the improved version C4. 5 [Hai06, Sim94]) and CART density of the distribution of classes are unknown, [Hai06, Sim94]. The difference between these and they have to be collected from the training algorithms is in the way the branching feature is sample. Recovery is possible only with some selected. The advantage of decision trees is the sinfulness, and the smaller the training sample, the visibility of the resulting model and the simplicity higher the probability of a retraining effect, when of its interpretation by a person. The disadvantage the method loses its generalizing properties and of the method is the problem of retraining, i.e. the correctly classifies only the examples from the possibility of constructing an excessively large tree training sample. Also, the efficiency of the method that will not fully represent the data. There is also a drops sharply when there is an error in the need to build a tree from scratch (i.e. a complete hypotheses about the density of the class change in the diagnostic algorithm) when changing distribution. the number of classes and the description of the Linear classifiers are effective for classification input data. problems with two classes. To solve the 24 Neural networks provide multi-class classification problem, teacher training is used. The most popular regardless of the linear time-separability of classes. algorithm for learning with a teacher is The In addition, neural networks are able to determine backpropaganation algorithm [Sim94], based on the the presence of the analyzed example features of gradient descent method on the hyperplane of the several classes. The effectiveness of neural error function, and its modified version RProp networks for classifying the technical condition of [Cal01, Hai06, Sim94], which is one of the best devices is shown in [Zue12, Zue13]. Therefore, it first-order learning algorithms. was decided to apply the theory of neural networks The process of recovering the operability of the for the development of methods and algorithms for device with an automatic self-learning system, fault detection. which is a multi-layer neural network based on a In order to solve the classification problem with the recurrent model is presented in Figure 4) help of a neural network, the network needs to be As a result, we changed the Troubleshooting taught examples of different images. The training process. There was a unification of information sample should include measures that fully describe flows, automated process of preparation for the image. In practical problems, to achieve the Troubleshooting, calculation and issuance of the most complete description of possible images in the optimal algorithm of actions. training sample, it is necessary to collect a In the course of elimination of failure there is an sufficient number of examples. Examples can be adjustment of the optimal algorithm at change of located in different parts of the corresponding input data (by results of measurements, change of a image area in DK space. train situation, the carried-out tests), To solve the problem of neural network training, recommendations are given, after elimination check there are many algorithms [Cal01, Hai06]. To train of operability is carried out. neural networks designed to solve the classification . Inform Preparation Arrival Search Troubleshooting - fromDSP; to search - from - afoot; SHCHD; - the collection of tools - transport. - replacement of faulty - carrying out - from STDM. and materials. elements of the device, additional adjustment. measurements (absent in STDM), checks. Preparation Search Troubleshooting to search - carrying out - determination of the location additional - check the working and nature of the failure; measurements that are good condition. - determination of the failed not present in the device; STDM, and their - collection and analysis of analysis diagnostic information in STDM; - train situation analysis; - analysis of technical APK-PN documentation. Figure 4. The process of recovering the operability of the device with an automatic self-learning system 4 Description of APK-PN system - reduced operating costs; - development of information exchange between The APK-PN system with a mobile measuring and adjacent offices. software complex is designed for automatization of The main objectives of creation of APK-PN are: fault finding of JAT devices, logging of the fault - reduced Troubleshooting time; finding process, checking the device operability - reducing the number of train delays and their after Troubleshooting and providing information to duration; operational personnel. Implementation of APK-PN - reducing the impact of human factors on the assumes achievement of the following results: Troubleshooting process; - reducing the recovery time of the device's - reducing the complexity of work to restore the working capacity; operability of railway automation and - improvement of safety and continuity of trains; telemechanics devices; 25 - reduction of "post-preventive" failures. Devices for which the APC-PN system implements To automate the above operations, the APK-PN the construction of fault finding algorithms: system provides: - centralization of arrows and signals; - collection of information from STDM; - track blocking; - download technical documentation from - moving and barrage centralization; ARM-VTD; - the formation and transmission of the signals - display the search algorithm of fault; of ALSN; - selection of possible faulty elements on circuit - other devices had been controlled by the diagrams and indication of additional measurement STDM on the basis of the APK-DK. points; On the basis of the General formulated - logging of the process of finding non- requirements to system APK-PN the hierarchical rightness; principle of construction is chosen and two levels - integration with existing control systems, are allocated: interaction with ASU-SH2 databases. - linear data collection point (LPS), which is a On the basis of the above tasks and goals, we mobile measurement and software system (level 1); present the structure of the APK-PN system. - the Central point of construction of algorithms System APK-PN includes subsystems: for fault claim in the distances of signaling, - collection of information in STDM; centralization and blocking (level 2). - downloads of technical documentation for the The hierarchical structure of the APC-PN railway AT devices; system is shown in Figure 5. - analysis of diagnostic information from STDM At level 1, there are linear points for collecting and technical documentation from ARM-VTD; additional diagnostic information (LPS), - construction and display of the fault finding performing additional measurements of diagnostic algorithm of the railway AT device, as well as the parameters of the control object and receiving image of connection points for additional functions from the Central point of the fault finding measurements and selection of the checked algorithm. elements on the technical documentation In case of absence of technical documentation on (schematic diagram). the failed device, the standard scheme for the - check the operation of the railway AT device corresponding EC or AB system in which the after troubleshooting; device is operated is loaded. - measurement of diagnostic parameters and . their analysis. CTDM Level 2 ASU-SH2 ARM-VTD ARM-PN, ARM-SHCHD (Central point) Level 1 Level 1 (LPD) Mobil ARM-SHN MPK ARM-SHNS (LPS) Concentrator LPD Devices Railway АТ Figure 5. Block diagram of APK-PN After receiving the full amount of information, the necessary to carry out checks in order to reduce the APK-PN begins synthesis of the algorithm for search time is issued. In case of insufficient amount searching for faults on the basis of its knowledge of diagnostic information to form a list of the base and a list of the minimum number of possible minimum number of possible faulty elements, the irregular elements in the sequence in which it is APK-PN synthesizes an algorithm for additional 26 measurements of diagnostic parameters of the device to reduce the search area and identify the faulty element. It also provides for the selection of elements in the schematic diagrams of the device References obtained from the ARM-VTD, which must be [Efa12] Efanov D. V. Bases of construction and checked and the points of additional measurements principles of functioning of systems of are indicated. technical diagnostics and monitoring of Further, the generated list of the minimum number devices of railway automatics and tele- of possible faulty elements or the synthesized mechanics./ Efanov D. V., Lykov A. A. / algorithm of additional measurements of diagnostic / - SPb.: St. Petersburg state University of parameters of the device is transmitted to the linear Railways, 2012. - 59c. point of STDM-ARM-SHN and to the linear point [Boc12] Bochkarev S. V. Identification of pre- of information collection-mobile measuring and failure States of railway automatics and software complex. Additional measurements of telemechanics devices. / Bochkarev S. V., diagnostic parameters of the device are carried out Lykov A. A. / / Intellectual technologies with the help of mobile measuring and software at the TRANS-port: materials of the II complex, and the analysis of measurement results is international scientific and practical carried out. Based on the analysis of the conference "Intellect TRANS-2012". – measurement results, a list of possible faulty SPb. St. Petersburg state University of elements is formed. The results of the means of communication, 2012 – p. 82- Troubleshooting are transmitted to the network 88. printer, where the final Protocol is printed. Also, [Kal09] Kalyavin V. p. Reliability and diagnostics the information is transferred back to the workplace of elements of electrical installations / of the SHCHD, after which the information is Kalyavin V. P. Ry-Bakov L. M.: entered into the database of ASU-SH2. Textbook. //Mar.state UN-T.-Yoshkar- Ola-2009-p. 336. 5 Conclusion [Efa12] Efanov D. V. Continuous diagnostics of SCB devices / Efanov D. V., Plekhanov This APK is successfully used in the distance of P. A. / / Automation, communication, the SCB on the October railway. Informatics-2012-No. 7-p. 18-20. We will calculate the economic effect by [Ana13] Analysis of the state of safety of trains, reducing unplanned breaks in the movement of reliability of systems and REAPER trains. devices in the economy of automation Savings by reducing downtime.: and mechanics in 2012. Moscow: JSC EP = Tskr G. x Spp g + Tskr p. x Spp p + tskr "Russian Railways", 2013-156 p. PR. x Spp PR. =290,764 rubles. [Ana10] Analysis of the state of train safety, Calculation of savings by reducing the reliability of Reaper systems and downtime of trains while reducing the time to DEVICES in the automation and tele- search for failures in SCB devices mechanics sector of JSC "Russian About 41.9% of all failures led to train delays, Railways" in 2009 to meet the the number of failures Notk / ZP = 255. requirements of the quality management The above costs associated with one stop of the system.Moscow: JSC "Russian train SOP = 191 rubles. Railways", 2010-156 p. Saving train hours by reducing the time of [Sap02] Sapozhnikov V. V. Reliability of systems elimination of failure by 27.5% ∆ TP = 0.64. of self-road automation, telemechanics Reducing the number of delayed trains while and communication / Sapozhnikov V. V., reducing the time of elimination of failure by Sapozhnikov VL.V., Shamanov V. I. / / 27.5% ∆ NP = 165 Textbook for universities railway Savings by reducing the downtime of trains transport. First edition. Edited by VL.V. while reducing the time to search for failures in the Sapozhnikov. – M., UMK Ministry of STB devices will be (one failure): Railways of the Russian Federation, EPO (1)= (Sppp+Sppg+Spppr) x Т TP + SOP x 2002. - pp. 285. ∆ NP = 36 516 RUB [Aks09] Aksamentov N. N. The use of special- For this calculation at Notk/ZP = 255 savings ized vehicles in the distance. Automation, will be: communication and Informatics. 2009- EPO = Notk/ZP x EPO(1) = 9 311 570 RUB. No. 1-pp. 48-50. Annual economic benefit: [Dmi12] Dmitrienko I. E. Technical diagnostics Eg = EP + EPO= 290 764 +9 311 570=9 602 and auto-control of railway automation 333 RUB. and telemechanics systems. 2-ed., Rev. and extra M – Transport, 1986 – 144. 27 [Zor07] Zorich V. A. "Mathematical analysis". Ed. Mtsnmo 2007. [Kal09] Kalyavin V. P. Reliability and diagnostics of elements of electrical installations/ Kalyavin V. P., Ry-Bakov L. M. / Textbook. //Mar.state UN-T.-Yoshkar- Ola-2009-p. 336. [Kly87] Klyueva V. V. Technical diagnostics. Volume 9 / Klyueva V. V., Parkhomenko P. P., / / ed. - M.: Mashinostroenie, 1987.- 352c. [Par81] Parkhomenko P. p. Fundamentals of technical diagnostics: optimization of diagnosis algorithms, hardware / P. p. Parkhomenko, E. S. Soghomonyan. - Moscow: Energo-Atomizdat, 1981. - 320 PP. [Bla19] Blagoveshchenskaya E. A. Synthesis of models of automatic Troubleshooting of railway infrastructure./ Blagoveshchenskaya E. A., Bulavsky P. E., Gruzdev N. V. / Proceedings of the XXI International conference on computational mechanics and modern applied software systems’vmspps ' 2019), may 24-31, 2019, Alushta. - Moscow: MAI Publishing house, 2019. — 816 p.: Il. [Sap04] Sapozhnikov, V. V. Fundamentals of technical diagnostics/ V. V. Sapozhnikov, Vol. V. Sapozhnikov. – M. : The Route, 2004. - 316 p. - ISBN 5-89035-123-0. [Kru01] Kruglov V. V. "Fuzzy logic and artificial neural networks" / Kruglov V. V., DLI M. I., Golunov R. Yu. / / Ed. FIZMATLIT 2001. [Cal01] Callan R. Basic concepts of neural networks. - M.: Williams, 2001. - 288c. [Hai06] Haikin S. Neural networks. Full course. - M.: Williams, 2006. - 1104 PP. [Hai94] Simon Haykin "Neural Networks: a Com- prehensive Foundation". 2-nd Edition. Ed. Mac-millan Coll Div, 1994. [Zue12] Zuev D. Solution of the problem of nonin-variance of using connectionist method for image recognition./ Zuev D., Bochkarev S. / / Materials of the II international research and practice conference, Vol. I, Munich, December 18-19, 2012; Germany, 2012-650p 257- 259 pp. [Zue13]. Zuev D. V. Analysis of diagnostic information/ Zuev D. V., Bochkarev S. V., Dmitriev V. V. / / Automation, communication, Informatics. - 2013. - No. 9. - pp. 16-17. 28