JP7442310B2 - Trained model generator, failure prediction device, failure prediction system, failure prediction program, and trained model - Google Patents

Description

The present invention relates to a trained model generation device used to predict ticket gate failures, a failure prediction device that uses a learned model to predict ticket gate failures, a failure prediction system, a failure prediction program, and a learned model. Regarding.

Conventionally, ticket gates have been maintained by repairing them every time a breakdown occurs, and by conducting maintenance and inspections on each ticket gate every 1 to 6 months. In addition, in order to prevent malfunctions during ticket gate use as much as possible and to carry out maintenance and inspection efficiently, maintenance and inspection cycles were sometimes reviewed from time to time. For example, as shown in Patent Document 1, the operating status of each ticket gate is artificially determined by continuously referring to the operating status and failure history such as the values of various sensors installed in the ticket gate. They identified ticket gates that were likely to malfunction and reviewed their maintenance and inspection cycles, such as bringing the timing of maintenance and inspection of those ticket gates earlier.

Japanese Patent Application Publication No. 2004-38747

However, with the above-mentioned method of artificially reconsidering the timing of maintenance and inspection of ticket gates based on the values of various sensors, etc., if one component breaks down independently, Although failures can be predicted, in reality it has been difficult to adequately determine when a ticket gate will fail. For example, in a ticket inspection machine, the operational accuracy of a plurality of parts decreases, resulting in a compounded reduction in the operational accuracy of the ticket inspection machine, which often leads to failure of the ticket inspection machine. In some cases, it is possible to predict a certain degree of complex failure by artificially organizing the values of various sensors using statistical methods, but in reality, there are a wide variety of factors that can lead to ticket gate failures. It has been virtually impossible to artificially predict failures in ticket gates that occur comprehensively due to the operations of such multiple parts.

An object of the present invention is to predict failures of ticket gates with high accuracy.

In order to achieve the above object, a trained model generation device according to an embodiment of the present invention is a trained model generation device that generates a trained model used for predicting failures of ticket gates. Receiving model operation information that is operation information of the ticket gate and model environment information that is environmental information of the ticket gate, which are acquired over time in a predetermined first period in each of the ticket gates, and A model data acquisition unit that receives failure information when a failure occurs in each of the ticket gates during a first period, and calculates the model operation information and the model environment information to calculate the IC card usage rate and A feature calculation unit that generates a first feature that includes a differential value of the number of jammed regular tickets in a predetermined period , and machine learning that uses the first feature as input data and the failure information as training data, and a model generation unit that generates a first trained model that outputs failure prediction information for each of the ticket gates when the operation information and the environment information are input.

With such a configuration, it is possible to generate the first learned model used for predicting the failure of the ticket gate based on the operating information and environmental information that affect the failure of the ticket gate, and the failure information. A ticket gate is made up of various parts, and the conditions of these parts combine in a complex manner to cause the ticket gate to malfunction. It is extremely difficult to comprehensively judge the status of these complex parts, but machine learning can recognize the relationship between the status of many parts and failures. Therefore, it is possible to recognize signs that the ticket gate is malfunctioning beyond human judgment, and the possibility of discovering signs that the ticket gate is malfunctioning is improved even in situations where it is difficult to judge manually. As a result, it is possible to generate a first trained model that can accurately predict failures. By performing failure prediction using this first trained model, it becomes possible to accurately predict failures of ticket gates.

Also, a feature extraction unit that analyzes the first trained model and extracts a second feature from the first feature under predetermined conditions that take into account the degree of influence for outputting the failure prediction information. The model generation unit is configured to calculate the failure of each of the ticket gates when the operation information and the environment information are input by machine learning using the second feature amount as input data and the failure information as training data. It is preferable that a second trained model is generated that outputs predictive information, and that failure prediction of the ticket gate is performed using the second trained model.

With such a configuration, it is possible to generate a second trained model with higher failure prediction accuracy by generating the second trained model using the second feature amount that has a large influence on failure prediction. By predicting failures using this second trained model, it becomes possible to predict failures of ticket gates with higher accuracy.

A failure prediction device according to an embodiment of the present invention is a failure prediction device that predicts a failure of a ticket gate using a trained model, and the failure prediction device predicts a failure of a ticket gate using a trained model, and the failure prediction device predicts a failure of a ticket gate using a trained model, and the failure prediction device predicts a failure of a ticket gate by using a learned model. The first feature quantity generated by calculating the model operation information that is the operation information of the ticket gate and the model environment information that is the environment information of the ticket gate is used as input data, and the first feature amount is set as input data. When the operation information and the environment information are inputted by machine learning using failure information when a failure occurs in each of the ticket gates as training data, failure prediction information including the failure probability of each of the ticket gates is generated. a model acquisition unit that acquires the learned model generated to output the predicted operation information that is the operation information that is acquired over time in a predetermined second period at each of the plurality of ticket gates; and a prediction data acquisition unit that receives the prediction environment information that is the environment information, and inputs the prediction operation information and the prediction environment information into the trained model acquired through the model acquisition unit. , and a failure prediction unit that outputs the failure prediction information of each of the ticket gates.

With this configuration, a trained model used for predicting failures of ticket gates is obtained, which is generated from operating information and environmental information that affect failures of ticket gates, and failure information, and this trained model is used to predict failures of ticket gates. Predictions can be made. Therefore, it is possible to recognize signs that the ticket gate is malfunctioning beyond human judgment, and the possibility of discovering signs that the ticket gate is malfunctioning is improved even in situations where it is difficult to judge manually. As a result, failures can be predicted using a learned model that can accurately predict failures, and ticket gate machine failures can be predicted with high accuracy.

In the failure prediction system according to an embodiment of the present invention, the operation information for prediction is acquired over time in a predetermined second period in the learned model generation device and each of the plurality of ticket gates. a prediction data acquisition unit that receives the prediction environment information that is the information and the environment information, and the prediction operation information and the prediction environment information to the first trained model generated by the learned model generation device. and a failure prediction unit that inputs and outputs the failure prediction information of each of the ticket gates.

With such a configuration, a first trained model used for predicting a failure of a ticket gate generated from operating information and environmental information that influences a failure of a ticket gate, and failure information is obtained, and this first trained model is used to predict a failure of a ticket gate. Failure prediction can be performed using Therefore, it is possible to recognize signs that the ticket gate is malfunctioning beyond human judgment, and the possibility of discovering signs that the ticket gate is malfunctioning is improved even in situations where it is difficult to judge manually. As a result, failures can be predicted using the first learned model that can accurately predict failures, and failures of ticket gates can be predicted with high accuracy.

Further, in the failure prediction system according to an embodiment of the present invention, the prediction that is the operation information is acquired over time in a predetermined second period in each of the learned model generation device and the plurality of ticket gates. a prediction data acquisition unit that receives the prediction operation information and the prediction environment information, which is the environment information; and a failure prediction unit that inputs information and outputs the failure prediction information of each of the ticket gates.

With such a configuration, by acquiring the second trained model generated using the second feature that has a large influence on failure prediction, failure prediction is performed using the second trained model with higher accuracy. This makes it possible to more accurately predict ticket gate machine failures.

A failure prediction program according to an embodiment of the present invention is a failure prediction program that uses a trained model to predict failures of ticket gates, and is configured to predict failures of ticket gates over time in each of the plurality of ticket gates in a predetermined first period. The first feature quantity generated by calculating the model operation information that is the operation information of the ticket gate and the model environment information that is the environment information of the ticket gate is used as input data, and the first feature amount is set as input data. When the operation information and the environment information are inputted by machine learning using failure information when a failure occurs in each of the ticket gates as training data, failure prediction information including the failure probability of each of the ticket gates is generated. a step of acquiring the trained model generated to output a step of receiving predictive environmental information that is environmental information; and inputting the predictive operating information and the predictive environmental information into the acquired trained model, and outputting the failure predictive information of each of the ticket gates; to cause the computer to perform the steps to cause the computer to perform the steps.

With this configuration, a trained model used for predicting failures of ticket gates is obtained, which is generated from operating information and environmental information that affect failures of ticket gates, and failure information, and this trained model is used to predict failures of ticket gates. Predictions can be made. Therefore, it is possible to recognize signs that the ticket gate is malfunctioning beyond human judgment, and the possibility of discovering signs that the ticket gate is malfunctioning is improved even in situations where it is difficult to judge manually. As a result, failures can be predicted using a learned model that can accurately predict failures, and ticket gate machine failures can be predicted with high accuracy.

Further, a failure prediction program according to an embodiment of the present invention is a failure prediction program that predicts failures of ticket gates using a trained model, and is a failure prediction program that uses a learned model to predict failures of ticket gates, and that model operating information that is operating information of the ticket gate and model environment information that is environmental information of the ticket gate, which are acquired during the first period, and a failure occurs in each of the ticket gates during the first period. a step of receiving failure information in the case of a jam, and calculating the operating information for the model and the environment information for the model to obtain a first information including an IC card usage rate and a differential value of the number of times that regular tickets are jammed in a predetermined period. When the operating information and the environment information are inputted, a step of generating a feature amount and machine learning using the first feature amount as input data and the failure information as training data determine the failure of each of the ticket gates. a step of generating a first trained model that outputs predictive information, and the predictive operating information that is the operating information and the environmental information that are acquired over time in a predetermined second period at each of the plurality of ticket gates. a step of inputting the predictive operation information and the predictive environmental information into the first trained model and outputting the failure predictive information of each of the ticket gates; have the computer execute it.

With such a configuration, it is possible to generate the first learned model used for predicting the failure of the ticket gate based on the operating information and environmental information that affect the failure of the ticket gate, and the failure information. Therefore, it is possible to recognize signs that the ticket gate is malfunctioning beyond human judgment, and the possibility of discovering signs that the ticket gate is malfunctioning is improved even in situations where it is difficult to judge manually. As a result, it is possible to generate a first trained model that can accurately predict failures. By performing failure prediction using this first trained model, it is possible to accurately predict failure of the ticket gate.

Further, a failure prediction program according to an embodiment of the present invention is a failure prediction program that predicts failures of ticket gates using a trained model, and is a failure prediction program that uses a learned model to predict failures of ticket gates, and that model operating information that is operating information of the ticket gate and model environment information that is environmental information of the ticket gate, which are acquired during the first period, and a failure occurs in each of the ticket gates during the first period. a step of receiving failure information in the case of a jam, and calculating the operating information for the model and the environment information for the model to obtain a first information including an IC card usage rate and a differential value of the number of times that regular tickets are jammed in a predetermined period. When the operating information and the environment information are inputted, a step of generating a feature amount and machine learning using the first feature amount as input data and the failure information as training data determine the failure of each of the ticket gates. a step of generating a first trained model that outputs prediction information; and analyzing the first trained model to generate the first feature amount under predetermined conditions that take into account the degree of influence for outputting the failure prediction information. A step of extracting a second feature quantity from among the ticket gates, and machine learning using the second feature quantity as input data and the failure information as teacher data, when the operation information and the environment information are input, each of the ticket gates generating a second trained model that outputs the failure prediction information of the machine; and predictive operation information that is the operation information acquired over time in a predetermined second period at each of the plurality of ticket gates. and a step of receiving prediction environment information which is the environment information, and inputting the prediction operation information and the prediction environment information into the second trained model to obtain the failure prediction information of each of the ticket gates. The computer executes the step of outputting the output.

With such a configuration, it is possible to generate a second trained model with higher failure prediction accuracy by generating the second trained model using the second feature amount that has a large influence on failure prediction. By predicting failures using this second trained model, it becomes possible to predict failures of ticket gates with higher accuracy.

Further, the trained model according to an embodiment of the present invention functions to cause a computer to output failure prediction information of the ticket gate based on the predictive operation information and predictive environment information acquired from the ticket gate. The model is a trained model that consists of a decision tree consisting of a plurality of branch points arranged in a tree structure, and the predictive operating information is operating information that is acquired over time at the ticket gate over a predetermined period. The predictive environmental information is environmental information acquired over time in a predetermined period at the ticket gate, and the predictive operational information and the predictive environmental information are input, and the predictive operational information and the predictive A plurality of feature quantities including the IC card usage rate and the differential value of the number of jammed regular tickets in a predetermined period are calculated from the usage environment information, and the corresponding feature quantity is calculated at each branch point to calculate the feature quantity. The failure prediction information is calculated by repeating the process of determining the direction of travel and proceeding to the next branch point, and adding up the evaluation values calculated at the branch points that have proceeded. Make the computer function to calculate and output.

With such a configuration, the trained model can predict failures of ticket gates using operating information and environmental information that affect failures of ticket gates. Therefore, it is possible to recognize signs that the ticket gate is malfunctioning beyond human judgment, and the possibility of discovering signs that the ticket gate is malfunctioning is improved even in situations where it is difficult to judge manually. As a result, it becomes possible to predict failures with high accuracy.

Furthermore, a plurality of decision trees may be combined.

With such a configuration, it is possible to generate a trained model according to the complexly related situations of ticket gates, and it is possible to predict failures with higher accuracy.

Further, the failure prediction information may include a time when each of the ticket gates is predicted to fail.

With such a configuration, it is possible to accurately grasp the period during which the ticket gate is expected to malfunction, and maintenance and inspection of the ticket gate can be performed at an appropriate time.

Further, a plurality of first trained models may be generated, and the failure prediction information output by each of the first trained models may indicate a probability that the ticket gate will fail during different periods.

With such a configuration, it is possible to output a plurality of pieces of failure prediction information with different periods of time until the ticket gate breaks down, using a plurality of trained models. Therefore, it is possible to more accurately grasp the period during which the ticket gate is expected to malfunction, and maintenance and inspection of the ticket gate can be performed at a more appropriate time.

Further, a plurality of second trained models may be generated, and the failure prediction information output by each of the second trained models may indicate a probability that the ticket gate will fail during different periods.

With such a configuration, it is possible to output a plurality of pieces of failure prediction information with different periods until the ticket gate breaks down, using a plurality of trained models. Therefore, it is possible to more accurately grasp the period during which the ticket gate is expected to malfunction, and maintenance and inspection of the ticket gate can be performed at a more appropriate time.

Further, the failure prediction information may include a part of the ticket gate that is predicted to fail.

With such a configuration, maintenance and inspection of the ticket gate can be efficiently performed by preferentially inspecting parts of the ticket gate that are predicted to malfunction.

Further, it may include a maintenance planning section that generates a maintenance plan for each of the ticket gates and changes each of the maintenance plans based on the failure prediction information.

With such a configuration, maintenance and inspection of the ticket gate can be performed more appropriately and efficiently, and failure of the ticket gate during operation can be suppressed.

Furthermore, a trained model generation program according to an embodiment of the present invention is a trained model generation program that generates a trained model used for predicting failures of ticket gates using the trained model, and includes a plurality of trained models. Receiving model operation information that is operation information of the ticket gate and model environment information that is environmental information of the ticket gate, which are acquired over time in a predetermined first period in each of the ticket gates, and a step of receiving failure information when a failure occurs in each of the ticket gates during a first period; and a step of generating a first feature amount by calculating the model operation information and the model environment information; First learning that outputs failure prediction information for each of the ticket gates when the operation information and the environment information are input by machine learning using the first feature amount as input data and the failure information as teacher data. and causing the computer to execute the step of generating a completed model.

With such a configuration, it is possible to generate the first learned model used for predicting the failure of the ticket gate based on the operating information and environmental information that affect the failure of the ticket gate, and the failure information. A ticket gate is made up of various parts, and the conditions of these parts combine in a complex manner to cause the ticket gate to malfunction. It is extremely difficult to comprehensively judge the status of these complex parts, but machine learning can recognize the relationship between the status of many parts and failures. Therefore, by recognizing signs of failure of the ticket gate beyond human judgment, it is possible to generate a first trained model that is capable of highly accurate failure prediction. By performing failure prediction using this first trained model, it becomes possible to accurately predict failures of ticket gates.

1 is a schematic diagram illustrating the configuration of a failure prediction system for a ticket gate; FIG. FIG. 2 is a diagram schematically showing the flow of processing performed by the ticket gate failure prediction system. FIG. 2 is a diagram illustrating the configuration of a learned model generation device and a failure prediction execution unit. FIG. 2 is a diagram schematically showing the flow of processing in which trained model generation is performed. FIG. 3 is a diagram illustrating the configuration of feature amounts. It is a figure explaining the structure of on-call data. FIG. 2 is a diagram schematically showing the flow of processing in which failure prediction is performed.

First, with reference to FIG. 1, the ticket gate machine 1 and the maintenance and repair of the ticket gate machine 1 will be explained.

The ticket gate machine 1 is installed at a railway station, etc., and manages entry and exit from the station premises. The ticket gate machine 1 reads the information recorded on the ticket, determines whether entry/exit is appropriate according to the information, and manages entry/exit into the station premises. As the ticket, a regular ticket, commuter pass, SF card (Stored Fare Card), IC card, etc. are used. The ticket gate 1 includes a slot (not shown), a reader (not shown), an antenna part (not shown), an analysis part (not shown), an operation control part (not shown), and a storage part (not shown). (not shown), a conveying section (not shown), a discharge port (not shown), etc. Regular tickets, commuter passes, SF cards, etc. are inserted into the slot. The reading unit reads information recorded magnetically or the like. The antenna section exchanges information with an IC card and the like. The analysis section analyzes the information acquired by the reading section and the antenna section. The operation control section controls necessary operations according to the analysis result of the analysis section. The storage section stores regular tickets, etc. whose management has been completed at the time of leaving the venue, etc. The conveyance section conveys regular tickets and the like inserted from the input port to the discharge port and storage section via the reading section. The exit is used to eject regular tickets, etc., which will be needed when entering the venue and subsequently exiting the venue. Furthermore, each part of the ticket gate 1, such as the slot and the reader, is provided with an operation sensor (not shown) that acquires the operation information 6 (see FIG. 1). The ticket gate 1 also includes an environmental sensor (not shown) inside or around it that acquires environmental information 7 (see FIG. 1) such as temperature, humidity, illuminance, atmospheric pressure, and weather.

When the ticket gate 1 malfunctions, a station staff member or the like makes an on-call to notify a maintenance company or the like of the malfunction. During the on-call, information identifying the faulty ticket gate 1 and the situation of the fault are communicated. The person in charge of the maintenance company or the like who receives the on-call goes to the relevant ticket gate 1 and repairs it. In addition, each ticket gate 1 is periodically maintained and inspected based on a maintenance plan in order to prevent it from breaking down during operation.

Next, the overall configuration of the ticket gate failure prediction system will be outlined using FIGS. 1 and 2.

One or more ticket gates 1 are connected to an internet line 2. Further, a trained model generation device 3 , a failure prediction execution unit 4 , and a server 5 are connected to the Internet line 2 .

The server 5 acquires operating information 6 and environmental information 7 over time from each ticket gate 1 via the Internet line 2 and stores them. Further, the server 5 acquires and stores on-call data 8 of on-calls made at each ticket gate 1 over time via the Internet line 2. Specifically, the on-call data 8 may be directly transmitted from the ticket inspection machine 1 to the server 5, or may be temporarily collected at a call center and transmitted from the call center to the server 5 or the trained model generation device.

The trained model generation device 3 first generates operation information 6, which corresponds to each ticket gate 1, and which is obtained from the server 5 over the Internet line 2 during a predetermined period (corresponding to the first period). Obtain environmental information 7 and on-call data 8 (step #1 in FIG. 2).

Next, the trained model generation device 3 sets the operating information 6 to model operating information 20 (see FIG. 3) and the environment information 7 to model environment information 21 (see FIG. The operation information 20 and the model environment information 21 are operated to calculate a feature quantity (corresponding to a first feature quantity) for generating a trained model (step #2 in FIG. 2). Further, the trained model generating device 3 extracts failure information from the on-call data 8 for each ticket gate 1. Note that in the following explanation, the extracted failure information is also simply referred to as on-call data 8 (step #3 in FIG. 2).

Next, the trained model generation device 3 uses the calculated feature amounts as input data for machine learning, uses the extracted failure information as training data, and associates the feature amounts and failure information corresponding to the same ticket gate 1. By performing machine learning using AI (artificial intelligence) 9 and inputting predictive operating information 6 (equivalent to predictive operating information) and predictive environmental information 7 (equivalent to predictive environmental information) A trained model 10 that outputs failure prediction information of 1 is generated (step #4 in FIG. 2). Here, the AI 9 may be provided on the Internet line 2 or may be provided within the trained model generation device 3. Note that the specific configuration of the trained model generation device 3 will be described in detail later.

Thereafter, the server 5 acquires and stores operating information 6 and environmental information 7 used for failure prediction from each ticket gate 1 via the Internet line 2 (step #5 in FIG. 2).

Next, the failure prediction execution unit 4 acquires the operating information 6 and the environmental information 7 stored in the server 5 via the Internet line 2. Then, the failure prediction execution unit 4 uses the operation information 6 continuously acquired during a predetermined period (second period) for each ticket gate 1 as prediction operation information, and continues for a predetermined period (second period). Using the environmental information 7 acquired as predictive environmental information, the predictive operating information and predictive environmental information are input into the learned model 10, and the failure predictive information 11 for each ticket gate 1 is output (see Fig. 2). Step #6). Note that the specific configuration of the failure prediction execution unit 4 will be described in detail later.

Furthermore, a maintenance planning section 12 connected to the Internet line 2 may be provided. The maintenance planning unit 12 formulates a maintenance plan for each ticket gate 1 and, based on the failure prediction information 11, performs maintenance planning to increase the priority of maintenance of ticket gates 1 that are likely to fail, as necessary. Change the plan (Step #7 in Figure 2). Note that the specific configuration of the maintenance planning section 12 will be detailed later.

In this way, in this embodiment, the operating information 6 and environmental information 7 of each ticket gate 1 are collected for a predetermined period, and these are calculated to calculate information suitable for predicting a failure and use it as input data. Failure prediction is performed using a trained model 10 that has been subjected to machine learning by labeling the data with on-call data 8 that actually occurred and using it as teacher data. Therefore, failure prediction can be performed by comprehensively judging a large amount of information, and failure prediction of the ticket gate 1 can be performed with high accuracy. In particular, there are so many factors that cause the ticket gate 1 to malfunction, and they interact with each other in a complex manner. Since the trained model 10 is generated using learning, failure prediction can be performed by making a composite judgment based on more information. It is also possible to use a program to manually select necessary information and predict failures from the correlations between them, but it is not possible to select the necessary information and derive the correlations between them. In the end, this has to be done manually, which is not practical considering the complexity of the ticket gate 1. Therefore, by generating the trained model 10 and performing failure prediction as described above, it is possible to perform failure prediction more easily and with higher accuracy than when such a program is used. Further, in the ticket gate 1, the influence on failure changes depending on the environment such as weather conditions such as atmospheric pressure and humidity. For example, the ease with which tickets get clogged changes depending on the environment, and the impact on the operating accuracy of electronic equipment changes. Therefore, by including environmental information in input data and performing machine learning, it is possible to generate a trained model 10 that takes into consideration the complex interaction between operating information and environmental information. Since the maintenance plan can be reviewed based on highly accurate failure prediction, it is possible to reduce the possibility that the ticket gate 1 in actual operation will break down.

[Trained model generation device]
Next, a specific example of the configuration of the trained model generation device 3 will be described using FIGS. 3 to 6 while referring to FIGS. 1 and 2.

The trained model generation device 3 includes a data communication section 13 , a model data acquisition section 14 , a storage section 15 , a generation control section 16 , a model generation section 17 , a feature calculation section 18 , and a feature extraction section 19 .

The data communication unit 13 transmits and receives data to and from the ticket gate 1, AI 9, etc. via the Internet line 2. The model data acquisition unit 14 acquires the model operation information 20 and the model environment information 21 of the ticket gate 1 received by the data communication unit 13, and the on-call data 8, and stores them in the storage unit 15 (steps in FIG. 4). #1). The model operation information 20 and the model environment information 21 are information obtained by collecting operation information 6 and environment information 7 acquired from the ticket gate 1 over time over a predetermined period (first period). Although the predetermined period can be arbitrarily set, it can be set to, for example, one month, and may be one month before the month in which failure prediction is performed, or one month immediately before. The model operation information 20 and the model environment information 21 are information acquired from each ticket gate 1 and linked to the ticket gate 1, and include, for example, a total of 2000 pieces of information (features). In addition to the model operating information 20, the model environment information 21, and the on-call data 8, the storage unit 15 also stores learning input data 22, a first trained model 23, a second trained model 24, etc., which will be described later. Stores various data. The generation control unit 16 includes a processor such as a CPU or a GPU, and controls the operation of the learned model generation device 3 such as the model generation unit 17 and the feature amount calculation unit 18.

The feature calculation section 18 calculates at least a part of the model operation information 20 of the model operation information 20 and the model environment information 21 to generate the first feature amount 25 under the control of the generation control section 16 . (Step #2 in FIG. 4). For example, as shown in FIG. 5, the first feature amount 25 is the usage rate of the IC card per day calculated from the number of times each of the IC card, ordinary ticket, commuter pass, and SF card is used in the operation information 6. be done. Further, as the first feature amount 25, the total or average value of the usage rate of the IC card over three days, or the total or average value over seven days may be calculated. In addition, a differential value of the number of times regular tickets are jammed in a predetermined period, etc. is calculated as the first feature amount 25. In this way, the first feature amount 25 is configured by adding the calculated items and their values to the model operation information 20 and the model environment information 21. As a result, the first feature amount 25 consists of, for example, 20,000 pieces of information (feature amount) in total. The first feature amount 25 is information that may affect the failure of the ticket gate 1, and by generating a large amount of this information and making a composite judgment, it is possible to accurately predict the failure of the ticket gate 1. It becomes like this.

The model generation unit 17 includes a processor such as a CPU or a GPU, and based on the control of the generation control unit 16, the input data is the first feature amount 25 and the training data is the on-call data 8 (fault information). 22 is input to the AI 9 and subjected to machine learning to generate the first learned model 23 (step #3 in FIG. 4). The on-call data 8 is information including information on on-call performed when the ticket gate 1 breaks down and the results of repairs performed in conjunction with the on-call, and is accumulated for each on-call. Therefore, the on-call data 8 includes various information, and as shown in FIG. It is composed of the date and time, on-call details, the failure part found as a result of repair, and the cause of the failure.

The feature extractor 19 inputs the first feature 25 to the first learned model 23 and analyzes it under the control of the generation controller 16, and determines the degree of influence of each feature on failure prediction. Then, the feature extraction unit 19 extracts, for example, 100 features from the 20,000 features in descending order of the degree of influence on failure prediction based on the control of the generation control unit 16. The feature amount is set as the second feature amount 26 (step #4 in FIG. 4).

Then, under the control of the generation control unit 16, the model generation unit 17 inputs the learning input data 27, in which the input data is the second feature quantity 26 and the teacher data is the on-call data 8, to the AI 9 to perform machine learning. , generates the second trained model 24 (step #5 in FIG. 4).

Note that the first trained model 23 and the second trained model 24 are models used to predict whether or not the ticket gate 1 will malfunction within a predetermined prediction period, for example, within 7 days. . Further, the first trained model 23 and the second trained model 24 are, for example, decision trees. The first trained model 23 and the second trained model 24 having such a decision tree structure evaluate a predetermined feature amount (first feature amount 25 or second feature amount 26) at each branch point of the decision tree. , the structure is such that an evaluation value (failure probability) is assigned to each branch point according to the evaluation result. Then, the evaluation values are added up along the branches of the decision tree to obtain failure prediction information. Note that the first trained model 23 and the second trained model 24 may be ensemble models such as XGBoost, Random Forest, LightGBM, CatBoost, and AdaBoost, in which a plurality of decision trees are provided in association with each other.

In this way, from the operation information 6 and the environment information 7, feature quantities that may affect the failure of the ticket gate 1 are selected, and these are calculated to make the degree of influence on the failure clear. is added to generate the first feature amount 25. Then, a first trained model 23 is generated using these first feature quantities 25 that are estimated to have a large influence on failure as input data. Therefore, it is possible to perform failure prediction that appropriately reflects the interaction of each feature amount, and it is possible to perform failure prediction of the ticket gate 1 with high accuracy.

Furthermore, the first feature quantity 25 is analyzed using the first trained model 23, the feature quantities that have a large influence on failure are extracted as second feature quantities 26, and these second feature quantities 26 are used as input data. A second trained model 24 is generated as follows. As a result, failure prediction can be performed using the second trained model 24 with higher failure prediction accuracy, and failure prediction of the ticket gate 1 can be performed with higher accuracy.

[Failure prediction execution unit]
Next, a specific configuration example of the failure prediction execution unit 4 will be described using FIGS. 3 and 7 while referring to FIGS. 1 and 2.

The failure prediction execution unit 4 includes a data communication unit 28, a model acquisition unit 29, a prediction data acquisition unit 30, a storage unit 31, a prediction control unit 32, and a failure prediction unit 33.

The data communication unit 28 transmits and receives data to and from the ticket gate 1, the trained model generation device 3, and the like via the Internet line 2. The model acquisition unit 29 acquires the first trained model 23 or the second trained model 24, which is a trained model used for failure prediction, from the trained model generation device 3 via the data communication unit 13, and stores it. 31 (step #1 in FIG. 7). The prediction data acquisition unit 30 acquires the prediction operation information 34 and the prediction environment information 35 of the ticket gate 1 received by the data communication unit 13, and stores them in the storage unit 31 (step #2 in FIG. 7). The predictive operating information 34 and the predictive environmental information 35 are information obtained by collecting operating information 6 and environmental information 7 acquired from the ticket gate 1 over time over a predetermined period (second period). Although the predetermined period can be arbitrarily set, for example, it is the day before failure prediction is to be performed. The prediction control unit 32 includes a processor such as a CPU, and controls the operation of the failure prediction unit 33 and the like.

The failure prediction unit 33 performs failure prediction by inputting the prediction operation information 34 and the prediction environment information 35 into the learned model based on the control of the prediction control unit 32 (Step #3 in FIG. 7). After storing the information 36 in the storage unit 31, it is output (step #4 in FIG. 7). Although the trained model is the second trained model 24 in this embodiment, it is also possible to use the first trained model 23.

The failure prediction information 36 indicates the probability that each ticket gate 1 will fail within the prediction period. If the prediction period is 7 days, the probability that the ticket gate 1 will malfunction within 7 days is indicated by a value of 0 or more and 1 or less for each ticket gate 1. This value indicates that the closer it gets to 1, the higher the probability of failure. When the second trained model 24 is composed of a decision tree, the failure prediction information 36 calculates the second feature amount 26 from the predictive operating information 34 and the predictive environment information 35, and calculates the second feature amount at each branch point. It is derived by performing 26 evaluations and adding up the evaluation values each time a branch point is passed.

Note that the failure prediction performed using the second learned model 24 was verified under the following conditions. The second trained model 24 was generated with a first period of 4 months and 19303 first feature quantities 25, and then failure prediction was performed for one month. As a result, the model accuracy was AUC (Area under an ROC curve): 75.31%, Precision (rate of not missing): 10%, and Recall (rate of not missing): 70%. Therefore, according to the present embodiment, failures can be predicted with higher accuracy than when maintenance planning is performed based on conventional human judgment, and when the failure prediction of this embodiment is performed, the ticket gate 1 is It can be seen that the probability of failure during operation is dramatically reduced compared to the conventional method.

Further, as the failure prediction information 36, a failure part predicted to fail in each ticket gate 1 may be outputted together with its failure probability. For example, for each ticket gate 1, the top three parts with a high probability of failure are output.

In failure prediction, an evaluation value of each second feature amount 26 is calculated. The failure probability for each part of each ticket gate 1 is determined by adding up the evaluation values of the second feature quantities 26 that are predetermined to have an influence on the failure of each part, among the evaluation values of the second feature quantities 26. It can be found by As failure prediction information 36, for each ticket gate 1, parts with a high probability of failure are output together with the probability of failure.

In this way, the trained models (first trained model 23, second trained model 23, Since failure prediction is performed using the model 24), failure prediction of the ticket gate 1 can be performed with high accuracy. In particular, by performing failure prediction using the second learned model 24 that has been machine learned using the second feature quantity 26 extracted in consideration of the magnitude of the influence on failure prediction, failures of the ticket gate 1 can be predicted. Predictions can be made more accurately. As a result, based on the failure prediction information 36, the deadline for maintaining the ticket gate 1 that has a high probability of failure becomes clear, and the cycle and timing of the maintenance plan can be reviewed to perform maintenance and inspection more appropriately. It becomes possible.

In addition, by outputting the parts with a high probability of failure for each ticket gate 1, the parts that should be maintained and inspected with priority become clear, and maintenance and inspection can be carried out efficiently.

[Maintenance plan]
The failure prediction execution unit 4 may further include a maintenance planning unit 37. The maintenance planning section 37 first determines a maintenance plan for periodically maintaining and inspecting each ticket gate 1. Then, based on the failure prediction information 36, the maintenance planning unit 37 changes the maintenance plan so that maintenance and inspection are performed for each ticket gate 1 within the expected failure period. do.

The maintenance planning section 37 may be configured to be connected to a display device (not shown). The maintenance planning section 37 displays the maintenance plan and failure prediction information 36 on a display device. A person in charge of maintenance, a manager, etc. can further review the maintenance plan while referring to the maintenance plan and failure prediction information 36 displayed on the display device.

In this way, by changing the maintenance plan based on the failure prediction information 36, maintenance and inspection can be carried out efficiently, and the ticket gate 1 can be maintained and inspected before failure, and the ticket gate 1 in operation can be prevented from breaking down. can be restrained from doing so.

Note that the maintenance planning section 37 is not limited to the configuration provided in the failure prediction execution section 4, and may be provided at any location. For example, the maintenance planning section 37 may be configured to be connected to the Internet line 2 like the maintenance planning section 12 shown in FIG.

[Another embodiment]
(1) In the above embodiment, the configuration of the learned model generation device 3 and the failure prediction execution unit 4 is not limited to the case where the learned model generation device 3 and the failure prediction execution unit 4 are divided into functional blocks. The configuration of the functional blocks is arbitrary as long as it can be realized. For example, the functional blocks of the generation control unit 16, model generation unit 17, feature amount calculation unit 18, and feature amount extraction unit 19 are not limited to such a configuration, but may be formed by functional blocks that integrate some or all of their functions. Alternatively, each functional block may be further subdivided into functional blocks. Further, in the embodiment described above, part or all of the functions of the trained model generation device 3 and the failure prediction execution unit 4 may be configured in an integrated manner.

(2) In each of the above embodiments, the learned model generation device 3 may be configured to generate and complete the learned model by itself, and not transmit the learned model to the failure prediction execution unit 4. Conversely, the failure prediction execution unit 4 may perform failure prediction using a separately generated trained model, without using the trained model generated by the trained model generation device 3.

(3) In each of the above embodiments, part or all of the learned model generation method realized by the learned model generation device 3 and the failure prediction method realized by the failure prediction execution unit 4 are not limited to the above device configuration, but are arbitrary. This can be realized with the following configuration. Further, part or all of the learned model generation method realized by the learned model generation device 3 and the failure prediction method realized by the failure prediction execution unit 4 may be realized by a program. The program is stored in an arbitrary recording device such as the storage section 15 or the storage section 31, and an arbitrary processor (equivalent to a computer) such as a CPU or a GPU included in the generation control section 16 executes this program.

(4) In each of the embodiments described above, the on-call data 8 may include only on-call information and not include repair results. In this case, the on-call data 8 (failure information) includes an ID for identifying the target ticket gate 1, the date and time of the on-call, and the content of the on-call. Note that the on-call content may be only information indicating whether or not an on-call has been performed, and in this case, the on-call data 8 (failure information) includes the ID that identifies the target ticket gate 1 and the on-call It consists of the date and time of the call and whether or not the user is on call.

(5) In each of the embodiments described above, the trained model generation device 3 may be configured without the feature amount extraction unit 19 and without generating the second feature amount 26 and the second learned model 24. In this case, the trained model generation device 3 generates only the first feature amount 25 and the first trained model 23.

(6) In each of the above embodiments, the configuration is not limited to the configuration in which failure prediction is performed only when the prediction period is 7 days, but failure prediction may be performed that outputs the probability of failure during a plurality of different prediction periods. . For example, the prediction period is set to 1 month, 14 days, 7 days, and 3 days, and the failure predictions are the probability that each ticket gate 1 will break down within one month, the probability that it will break down within 14 days, and the probability that it will break down within 7 days. Outputs the probability of failure within 3 days and the probability of failure within 3 days. In this case, the first trained model 23 and the second trained model 24 are a trained model for determining the probability of failure within one month, a trained model for determining the probability of failure within 14 days, and a trained model for determining the probability of failure within 1 month. Four types of trained models are generated: a trained model for determining the probability of failure within 3 days, and a trained model for determining the probability of failure within 3 days. The failure prediction execution unit 4 then performs failure prediction using the four types of first trained models 23 or the four types of second learned models 24, and outputs four types of failure prediction information 36.

This shows the probability of failure during a plurality of periods, making it possible to more accurately grasp the timing at which maintenance and inspection should be performed. Accordingly, maintenance plans can be changed more accurately.

(7) In each of the above embodiments, the first trained model 23 and the second trained model 24 may be updated as failure prediction is performed. By inputting the predictive operating information 34 and predictive environmental information 35 used in failure prediction performed during a predetermined period into the first trained model 23 or the second trained model 24 and performing machine learning. , the first trained model 23 or the second trained model 24 is updated. At this time, the results of failure prediction using the predictive operating information 34 and the predictive environmental information 35 and the results of maintenance and repair may be added to the teacher data.

In this way, by updating the learned model in conjunction with failure prediction, the accuracy of the learned model improves, making it possible to perform more accurate failure prediction.

(8) In each of the above embodiments, the ticket gate 1 may be configured without an environmental sensor. In this case, a learned model 10 is generated from the operating information 6 and on-call data 8 (failure information), and the operating information 6 is input to the learned model 10 when predicting a failure. Further, other information may be used to generate the trained model 10, and any of the operating information 6, the environment information 7, and other information may be combined and used to generate the trained model 10. It's okay.

The environmental information 7 may not have a large effect on failure prediction. In such a case, the ticket gate 1 can have a simpler configuration by not having an environmental sensor.

(9) The ticket gate 1, the trained model generation device 3, the failure prediction execution unit 4, and the maintenance planning unit 12 are configured to be able to exchange data with each other in any configuration, not just the configuration connected to the Internet line 2. That's fine. For example, at least a portion of these devices may be connected to a network line other than the Internet line 2, such as an intranet, or may be configured to allow transmission and reception over a dedicated line such as a LAN, or may be configured to allow transmission and reception via a storage medium. It may also be configured to exchange data.

(10) The failure information used to generate the trained model 10 may be extracted from the on-call data 8, but may also be obtained from sources other than the on-call data 8. For example, the administrator of the ticket gate 1 may decide to create failure information and input it to the trained model generation device 3 via the Internet line 2, or the ticket gate 1 may determine the failure by itself. Failure information may be generated and input to the trained model generation device 3 via the Internet line 2 or the like.

The present invention provides a trained model generation device, a failure prediction device, and a failure prediction device for use not only in ticket gates installed in railway stations and other transportation stations, but also in ticket gates installed in various facilities such as theme parks. It can be applied to systems, failure prediction programs, and trained models.

1 Ticket gate 2 Internet line 3 Trained model generation device 4 Failure prediction execution unit 5 Server 6 Operation information 7 Environmental information 8 On-call data 9 AI
10 Learned model 11 Failure prediction information 12 Maintenance planning section 13 Data communication section 14 Model data acquisition section 15 Storage section 16 Generation control section 17 Model generation section 18 Feature amount calculation section 19 Feature amount extraction section 20 Model operation information 21 Model environment information 22 input data for learning 23 first trained model (trained model)
24 Second trained model (trained model)
25 First feature amount (feature amount)
26 Second feature amount (feature amount)
27 Learning input data 28 Data communication section 29 Model acquisition section 30 Prediction data acquisition section 31 Storage section 32 Prediction control section 33 Failure prediction section 34 Prediction operation information 35 Prediction environment information 36 Failure prediction information 37 Maintenance planning section

Claims (10)

A trained model generation device that generates a trained model used to predict failures of ticket gates,
Receiving model operation information that is operation information of the ticket gate and model environment information that is environmental information of the ticket gate, which are acquired over time in a predetermined first period in each of the plurality of ticket gates; , a model data acquisition unit that receives failure information when a failure occurs in each of the ticket gates during the first period;
a feature amount calculation unit that generates a first feature amount including an IC card usage rate and a differential value of the number of jammed regular tickets in a predetermined period by calculating the model operation information and the model environment information;
First learning that outputs failure prediction information for each of the ticket gates when the operation information and the environment information are input by machine learning using the first feature amount as input data and the failure information as teacher data. A trained model generation device comprising a model generation unit that generates a trained model. a feature extraction unit that analyzes the first learned model and extracts a second feature from the first feature under predetermined conditions that take into account the degree of influence for outputting the failure prediction information; ,
The model generation unit generates the failure prediction information of each of the ticket gates when the operation information and the environment information are input by machine learning using the second feature amount as input data and the failure information as training data. 2. The learned model generation device according to claim 1, wherein a second learned model is generated to be output, and failure prediction of the ticket gate is performed using the second learned model. A failure prediction device that predicts failures of ticket gates using a trained model,
Calculating model operation information that is operation information of the ticket gate and model environment information that is environmental information of the ticket gate, which are acquired over time in a predetermined first period in each of the plurality of ticket gates. The operation information and the environment information are input by machine learning using the first feature generated by the above as input data and failure information when a failure occurs in each of the ticket gates during the first period as training data. a model acquisition unit that acquires the trained model generated to output failure prediction information including a failure probability of each of the ticket gates;
a prediction data acquisition unit that receives prediction operation information that is the operation information and prediction environment information that is the environment information, which are acquired over time in a predetermined second period at each of the plurality of ticket gates;
a failure prediction unit that inputs the prediction operation information and the prediction environment information into the trained model acquired via the model acquisition unit and outputs the failure prediction information of each ticket gate. Failure prediction device. The trained model generation device according to claim 1;
a prediction data acquisition unit that receives prediction operation information that is the operation information and prediction environment information that is the environment information, which are acquired over time in a predetermined second period at each of the plurality of ticket gates;
a failure prediction unit that inputs the predictive operation information and the predictive environment information to the first trained model generated by the trained model generation device and outputs the failure predictive information of each ticket gate; A failure prediction system equipped with The trained model generation device according to claim 2;
a prediction data acquisition unit that receives prediction operation information that is the operation information and prediction environment information that is the environment information, which are acquired over time in a predetermined second period at each of the plurality of ticket gates;
a failure prediction unit that inputs the predictive operation information and the predictive environment information to the second trained model generated by the trained model generation device and outputs the failure predictive information of each ticket gate; A failure prediction system equipped with A failure prediction program that predicts failures of ticket gates using a trained model,
Calculating model operation information that is operation information of the ticket gate and model environment information that is environmental information of the ticket gate, which are acquired over time in a predetermined first period in each of the plurality of ticket gates. The operation information and the environment information are input by machine learning using the first feature generated by the above as input data and failure information when a failure occurs in each of the ticket gates during the first period as training data. then, obtaining the trained model generated to output failure prediction information including the failure probability of each of the ticket gates;
a step of receiving predictive operating information that is the operating information and predictive environmental information that is the environmental information, which are acquired over time in a predetermined second period at each of the plurality of ticket gates;
A failure prediction program that causes a computer to execute the step of inputting the prediction operation information and the prediction environment information into the acquired trained model and outputting the failure prediction information of each of the ticket gates. A failure prediction program that predicts failures of ticket gates using a trained model,
Receiving model operation information that is operation information of the ticket gate and model environment information that is environmental information of the ticket gate, which are acquired over time in a predetermined first period in each of the plurality of ticket gates; , receiving failure information when a failure occurs in each of the ticket gates during the first period;
generating a first feature amount including an IC card usage rate and a differential value of the number of jammed regular tickets in a predetermined period by calculating the model operation information and the model environment information;
First learning that outputs failure prediction information for each of the ticket gates when the operation information and the environment information are input by machine learning using the first feature amount as input data and the failure information as teacher data. a step of generating a completed model;
a step of receiving predictive operating information that is the operating information and predictive environmental information that is the environmental information, which are acquired over time in a predetermined second period at each of the plurality of ticket gates;
A failure prediction program that causes a computer to execute the step of inputting the predictive operation information and the predictive environment information into the first trained model and causing the computer to output the failure predictive information of each of the ticket gates. A failure prediction program that predicts failures of ticket gates using a trained model,
Receiving model operation information that is operation information of the ticket gate and model environment information that is environmental information of the ticket gate, which are acquired over time in a predetermined first period in each of the plurality of ticket gates; , receiving failure information when a failure occurs in each of the ticket gates during the first period;
generating a first feature amount including an IC card usage rate and a differential value of the number of jammed regular tickets in a predetermined period by calculating the model operation information and the model environment information;
First learning that outputs failure prediction information for each of the ticket gates when the operation information and the environment information are input by machine learning using the first feature amount as input data and the failure information as teacher data. a step of generating a completed model;
analyzing the first trained model and extracting a second feature from among the first features under predetermined conditions that take into account the degree of influence for outputting the failure prediction information;
A second learned machine that outputs the failure prediction information of each of the ticket gates when the operation information and the environment information are input by machine learning using the second feature amount as input data and the failure information as teacher data. a step of generating a model;
a step of receiving predictive operating information that is the operating information and predictive environmental information that is the environmental information, which are acquired over time in a predetermined second period at each of the plurality of ticket gates;
A failure prediction program that causes a computer to execute a step of inputting the predictive operation information and the predictive environment information into the second trained model and causing the computer to output the failure predictive information of each of the ticket gates. A learned model for operating a computer to output failure prediction information of the ticket gate based on predictive operation information and predictive environment information acquired from the ticket gate,
It consists of a decision tree consisting of multiple branch points arranged in a tree structure,
The predictive operating information is operational information that is acquired over time at the ticket gate over a predetermined period, and the predictive environmental information is environmental information that is acquired over time over a predetermined period at the ticket gate.
The predictive operating information and the predictive environmental information are input, and a plurality of features including an IC card usage rate and a differential value of the number of times regular tickets are jammed in a predetermined period are obtained from the predictive operational information and the predictive environmental information. The amount is calculated, the corresponding feature amount is calculated at each of the branch points, an evaluation value corresponding to the feature amount is calculated, and a traveling direction is determined to determine whether to proceed to the next branch point. A learned model for causing a computer to function so as to calculate and output the failure prediction information by repeatedly summing up the evaluation values calculated at the progressed branch points. The trained model according to claim 9, which is configured by combining a plurality of decision trees.

Images