JP7332185B2 - Manhole pump abnormality detection method - Google Patents

Description

TECHNICAL FIELD The present invention relates to a method for detecting an abnormality in a manhole pump in advance.

There are about 47,000 manhole pumps installed nationwide. While the management and maintenance costs of aging facilities and the costs of updating them are increasing year by year, the shortage of workers on-site has become a major issue.
Traditionally, sewage manhole pumps have an alert function that issues an alert in the event of pump overload, pump operation for a long time, or pump failure. I went and did some repairs.
However, while the cost of maintenance and renewal of aging facilities is increasing year by year, the heavy burden on the personnel working on the site is also a problem.

JP 2019-020264 JP 2016-018526 JP 2016-012158 JP 2012-038298

SUMMARY OF THE INVENTION Accordingly, it is an object of the present invention to provide an abnormality detection method for a manhole pump, which can improve the efficiency of maintenance and management costs, reduce costs, and reduce the burden on field workers. do.

In order to achieve the above object, a manhole pump abnormality detection method according to the present invention extracts current value time series data of a manhole pump for a predetermined period with a learning reference time as an end point, and predicts from the learning reference time. By inputting to the neural network as a set together with the state of whether or not an abnormality has occurred before the detection period has elapsed and making it learn, calculation of the abnormality probability within the prediction detection period starting from an arbitrary reference time for evaluation By creating a device and inputting the current value time series data for the predetermined period up to the arbitrary evaluation reference time into the abnormality accuracy calculator, the prediction detection period from the evaluation reference time to the manhole pump It is characterized in that it is possible to detect the occurrence of an abnormality in advance by estimating the probability of occurrence of the abnormality until the time has elapsed.

According to the manhole pump abnormality detection method according to the present invention, it is possible to detect an abnormality of the manhole pump that may occur in the near future, so that it is possible to secure time until the abnormality occurs. This makes it possible to efficiently prepare for repairs and review plans, so it is possible to reduce the need for emergency dispatch and nighttime standby, which in turn contributes to the reduction of the burden on site workers. . In addition, since the abnormality accuracy calculator in the manhole pump abnormality detection method according to the present invention can be constructed if there are three or more abnormal data for learning for the target pump, it is generally said that a large amount of data is required. In deep learning, which is widely used, it is very useful because it is easy to secure data even in situations where there is little abnormal data.

FIG. 10 is an explanatory diagram of teacher data for an anomaly accuracy calculator; It is a graph of the number of correct answers in each of the abnormal and normal evaluation data of the trained neural network. It is a graph of the operation example of the detection method.

An embodiment of the abnormality detection method of the present invention will be described below with reference to the drawings.
The main purpose of the anomaly detection method of this embodiment is to reduce the dispatch of site workers at night (assuming 15 hours from 18:00 to 9:00 the next day). For that purpose, it is necessary to catch the signs of the occurrence of pump abnormalities between now and 15 hours from now.
In this embodiment, by creating teacher data and inputting it to a neural network for learning, an abnormality probability calculator within the prediction detection period is created, and the reference time for evaluation R2 calculated by the abnormality accuracy calculator is the starting point. The pump abnormality is detected in advance with the probability A that the pump abnormality will occur within the predicted detection period E of . The predictive detection period E is a certain period starting from the evaluation reference time R2, which is the target for predicting the occurrence of pump abnormality, and the length of the predictive detection period E is determined when the teacher data is created. is. Note that hereinafter, the reference time for learning R1 and the reference time for evaluation R2 both represent various events, such as the current time, an arbitrary judgment time, and the time when an abnormality occurs in a pump. is the time used when creating teacher data, and the reference time for evaluation R2 is the time specified by the user when using the anomaly probability calculator. Further, in the creation of training data, the pump abnormality occurrence time refers to the time at which an abnormality warning is issued by the existing alert issuing function of the pump.

In this embodiment, current value data obtained in real time using an ammeter having an IoT function (hereinafter referred to as an IoT ammeter) is used.
The IoT ammeter in this embodiment transmits 1200 pieces of current value data sampled every 0.05 seconds every minute.
In addition, hereinafter, the current value time-series data is a data string in which the current value data is arranged for each time. Collection is made with an arbitrary learning reference time R1 or evaluation reference time R2 of the shortest interval of current value data transfer by the ammeter as an end point. Then, one current value time-series data is obtained every minute or every minute for a predetermined period P (24 hours in this embodiment) ending at the learning reference time R1 or an arbitrary evaluation reference time R2. It is configured as a string of current value data of the pump operating time average for each second (in this embodiment, a total of 24 hours of 1438-minute intervals and 120-second intervals).
Here, the current value data of the average operating time of the pump at every minute or every second is the one minute measured by the IoT ammeter (in this embodiment, from 0 second to 59.95 seconds of the minute) 1200 pieces of current value data for 1 second or 20 pieces of current value data for 1 second (in this embodiment, from 0 to 0.95 seconds of the second interval), the time when the pump was operated (minute or second interval) Current value data only, i.e. current value data with a value other than zero, is the current value data averaged by dividing the total of the current values by the total number of data, and every minute or every second It means representative current value data.
Then, the column of current value data (current value time series data) averaged over the operating time of the pump every minute or every second for the predetermined period P is the pump is a data string in which the average current value data for the operation time of is continued for a predetermined time P. (In this embodiment, a continuous 24-hour data string of 1438 average operating times per minute and then 120 average operating times per second was used.) In other words, the length of the predetermined period P is the data length of the current value time-series data, which is the input data to the anomaly accuracy calculator.

The abnormality accuracy calculator is the probability that the pump will be in an abnormal state within the prediction detection period E starting from the evaluation reference time R2 from the current value time series data before the time when the pump abnormality occurs. It outputs A.
The construction of the anomaly accuracy calculator is performed as follows using a neural network. First, create training data for the neural network to learn. FIG. 1 is an explanatory diagram of teacher data.

The teacher data consists of input data, which is a set of current value time-series data, and the state of the pump at that time (normal or abnormal). The input data consists of normal data 5 and abnormal data 6. . Both the normal data 5 and the abnormal data 6 are selected from the current value time-series data collected with the learning reference time R1 as the end point of the predetermined period P.
The abnormality time data 6 is current value time-series data in which the end point (learning reference time R1) of the predetermined period P is included in the prediction detection period E (15 hours in this embodiment) until the pump abnormality occurs. It's about. Therefore, it is possible to collect the abnormality time data 6 with 15 hours×60 minutes=900 times as the learning reference time R1 for each case of actual pump abnormality. However, if all of these 900 pieces of data are used, the contribution of the relevant abnormal cases to the judged abnormalities is considered to be too high. adopt.
Normal time data 5 is data other than abnormal time data 6 among all collected current value time-series data, that is, data from the end point of predetermined period P (learning reference time R1) to predicted detection period E (15 in this embodiment). It is current value time-series data when no pump abnormality occurs during the elapse of a period of time. Regarding the number of normal data 5, a large amount can be collected because the number of abnormal occurrences is usually small. In the example, 450 pieces of normal data are adopted per abnormal data.
In other words, the teacher data includes the state of whether or not an abnormality has occurred during the period from the end of the predetermined period P (learning reference time R1) to the time when the prediction detection period E has elapsed, and the extracted data for the predetermined period P. It is a data matrix in which multiple sets of current value time-series data are collected.

The three pump facilities, Shimamotogo, Tajima, and Eirakucho, were targeted for the construction of this anomaly accuracy calculator. is. In addition, the time of occurrence of the pump abnormality obtained from the repair record at the time of emergency dispatch as shown in Tables 1 and 2, and the current value obtained in real time using an ammeter with IoT function that matches that time By comparing it with the data, it can be known that the real-time current value at the time when the abnormality occurred is the current value data at the time when the abnormality occurred.

The training data is learned by the neural network as follows.
First, training is started using a neural network that is easy to acquire multi-scale features with a multi-layered SE structure with multiple stages of squeeze-and-excitation blocks as a base network. Create automatically. Learning by AutoML/NAS means that in neural network learning, parameters set by engineers, generally called hyperparameters, are automatically set by software (AutoML), or the network structure is automatically changed (NAS: network It is a general term for functions that perform automatic structure search) and automatically proceed with tuning work.
Input items to the neural network are learning data including three or more abnormalities of the same pump station, pump rated current, and alert setting value. The alert setting value is a coefficient indicating how many times the rated current of the pump is to trigger the alert in the existing system. The abnormality accuracy calculator of the present invention can be constructed if there are three or more abnormal data for learning for the target pump, so deep learning, which is generally said to require a large amount of data, can However, it is very useful because it is an amount that is easy to secure data.

The evaluation data shown in Table 2 are input to the 300 to 400 trained neural networks created in this way, and the neural network with the optimum number of normal/abnormal correct answers is selected. When inputting the evaluation data, a period including a date and time of occurrence of an abnormality, which is not included in the teacher data, is input as the reference time for evaluation R2. FIG. 2 is a graph showing the number of correct answers for each of the abnormal and normal evaluation data of the trained neural network. Since it is generally impossible to find a perfect neural network that is correct in all cases, a realistic neural network is selected based on the Pareto optimal solution. Considering the characteristics of the pump facility and the needs of the site, a neural network that emphasizes oversight reduction (inside the one-dot chain line circle in Figure 2) or a neural network that emphasizes miss-swing reduction (inside the two-dot chain line circle in Figure 2). Choose one of the networks. In selecting a neural network, when visualizing the characteristics of each neural network in a graph, display on a logarithmic axis may eliminate the difficulty of viewing due to the difference in the number of normal/abnormal data. An abnormality probability calculator is constructed as described above.
In addition, since the selected neural network is a Pareto optimal solution, when it is actually incorporated into an anomaly notification application, it is possible to detect anomalies in advance due to variations in the accuracy A over time caused by noise contained in the input data. In order to suppress the problem of (1), it is necessary to stabilize the accuracy A by removing noise by a method such as moving average when operating the anomaly notification application. The anomaly notification application has functions such as calling the accuracy A calculated by the anomaly accuracy calculator, displaying it, passing it to other sections, and issuing an anomaly detection information.

The selected trained neural network (abnormality accuracy calculator) is read as a so-called weight file into the abnormality notification application together with current value time series data for a predetermined period P up to the evaluation reference time R2, which is input data. Then, the probability A that an abnormality will occur in the manhole pump from the reference time for evaluation R2 to the elapse of the predicted detection period E is calculated. Final anomaly accuracy B is obtained by removing noise from this accuracy A by an anomaly notification application using a technique such as moving average as described above. In this embodiment, the final abnormality probability B is obtained by taking a 15-hour moving average from the accuracy A. FIG. When the final abnormality probability B exceeds the threshold value T (60% in this embodiment), the abnormality warning application issues the abnormality pre-detection information. In this embodiment, real-time measurement results by an ammeter having an IoT function are used as input data, and the final abnormality accuracy B is calculated every 30 minutes.

Next, the hit rate of the constructed abnormality probability calculator will be described. The definition of the accuracy calculation is as follows. If there is an anomaly detection (forecast notification) with a final anomaly probability B of 60% or more in the time zone before the alert is issued by the existing system, it is determined to be correct. The hit rate is defined as the ratio of the number of cases where the anomaly detection by the anomaly probability calculator hits against the alert issued by the existing system.
Table 3 shows the hit rate of the abnormality probability calculator for the verification data in Table 2 above. Of the 8 existing alerts issued, 7 were hit, and the hit rate was 87.5%.

Next, the result of actually operating the manhole pump abnormality detection method of the present invention will be shown. This result is an example in which the maintenance management contractor 6 was notified of the abnormality detection information by the abnormality notification application, and the maintenance management contractor took action. FIG. 3 is a graph showing the relationship between the abnormality accuracy of the No. 1 manhole pump at the Taya north sewage relay pump station and the evaluation reference time R2 calculated and notified by the abnormality accuracy calculator and the abnormality notification application. At 5:30 on July 8, 2021, the final anomaly accuracy B (15-hour moving average of anomaly accuracy A) reached 60%. was done. After receiving this notification, maintenance contractors responded to the site at 14:00 on July 8, 2021. During this on-site response, it was discovered that foreign matter (screen residue) was present inside the casing. As a result, the final abnormality probability B fell below 60% at 6:30 on July 9, 2021, and the standby system for maintenance contractors was lifted. In this way, by notifying the maintenance contractor of the abnormality detection information in advance by the detection method of the present invention during the daytime, it was possible to secure the time for repair and repair before the abnormality occurred during the daytime. As a result, maintenance contractors were able to complete the work without dispatching at night (18:00 to 9:00 the next day). This greatly contributed to reducing the burden on field workers.

The invention is not limited to the embodiments described above. For example, the predicted detection period E is not limited to 15 hours, and may be divided into a plurality of periods, such as 1 hour units until 15 hours later. In addition, the pump abnormality occurrence time in the creation of training data does not have to be the time when an abnormality alarm is issued by the existing alert issuing function of the pump. It's okay.

A Accuracy E Prediction detection period P Predetermined period R1 Reference time for learning R2 Reference time for evaluation

Claims (1)

Extracting the manhole pump current value time series data for a predetermined period with the learning reference time as the end point,
By inputting to the neural network as a set together with the state of whether or not an abnormality has occurred from the learning reference time to the time when the prediction detection period has elapsed, and making it learn, the prediction with an arbitrary evaluation reference time as a starting point Create an anomaly accuracy calculator within the detection period,
By inputting the current value time-series data for the predetermined period up to the arbitrary evaluation reference time into the abnormality accuracy calculator, the manhole pump can detect A method for detecting abnormality in advance of a manhole pump, characterized in that the occurrence of abnormality can be detected in advance by estimating the probability of occurrence of abnormality in the manhole pump.

Images