JP2022025861A - Water leakage detector and water leakage detection method - Google Patents

Abstract

To provide technology with which it is possible to detect water leakage with higher accuracy.SOLUTION: Provided is a water leakage detector comprising: a data storage unit for storing the time-series data of flow rate and pressure of the water supply and distribution process to be detected; a training data acquisition unit for acquiring training object data from the time-series data; an assessment object data acquisition unit for acquiring the object data to be assessed for water leakage in the water supply and distribution process; a data conversion unit for converting the training object data so as to be trainable by a convolutional neutral network and converting the object data to be assessed so as to be assessable; a training parameter update unit for updating model parameters by training these using the converted training object data; a training parameter storage unit for storing the updated training parameters having been updated; and a water leakage assessment unit for assessing the presence of water leakage by a difference between the estimate values estimated by the convolutional neural network using the converted assessment object data and updated training parameters and the measured values.SELECTED DRAWING: Figure 1

Description

The present invention relates to a leak detection device and a leak detection method.

Conventionally, there has been a demand for a technique for detecting water leakage generated in a pipe in a water distribution process.
In Patent Document 1, which is an example of the prior art, each of the plurality of prediction determination methods is used, and the determination time is set based on the measured values of the sensors of the monitored process in the data range corresponding to the plurality of prediction determination methods. The predicted value of the first measured value of the sensor is predicted, the difference between the predicted value and the first measured value at the judgment time, and the first measured value outside the range between the upper limit value and the lower limit value of the predicted value. The abnormality of the monitored process is judged based on at least one of the above, and the range including the second measured value affected by the abnormality included in the abnormality judgment result of the multiple prediction judgment methods as the influence data is affected. Select as the data range, reduce the reliability of the abnormality judgment result based on the influence data included in the influence data range among the abnormality judgment results, integrate the abnormality judgment results of multiple prediction judgment methods, and integrate the abnormality judgment result. The technology to output is disclosed.

Japanese Patent No. 6605357

However, the above-mentioned prior art has room for improvement in prediction based on data.

The present invention has been made in view of the above, and an object of the present invention is to provide a technique capable of detecting water leakage with higher accuracy.

The present invention, which solves the above-mentioned problems and achieves the object, has a data storage unit that stores time-series data of the flow rate and pressure of the water distribution process to be detected, and training data that acquires learning target data from the time-series data. The acquisition unit, the determination target data acquisition unit that acquires the determination target data of water leakage in the water distribution process, and the data that converts the learning target data into learnable by a convolutional neural network and converts the determination target data into judgmentable data. The conversion unit, the learning parameter update unit that updates the model parameters by learning using the converted learning target data converted by the data conversion unit, and the updated learning parameters updated by the learning parameter update unit are stored. The difference between the estimated value and the measured value estimated by the convolution neural network using the learning parameter storage unit, the converted determination target data converted by the data conversion unit, and the updated learning parameter is calculated, and the difference is used. It is a water leakage detection device including a water leakage determination unit that determines an abnormality and, if it is an abnormality, determines that there is a water leakage in the water supply distribution process.

In the present invention having the above configuration, it is preferable that the data conversion unit converts the Fourier-transformed data obtained by Fourier-transforming the time-series data into two-dimensional data of time and frequency.

Alternatively, in the present invention having the above configuration, it is preferable that the data conversion unit converts the Wavelet-converted data obtained by wavelet-converting the time-series data into three-dimensional data of the number of days, the time, and the number of basis functions.

Alternatively, the present invention enables the accumulation of time-series data of the flow rate and pressure of the water distribution process to be detected, the acquisition of training target data from the time-series data, and the convolutional neural network of the training target data. Conversion, updating model parameters by learning using the converted converted learning target data, storing the updated updated learning parameters, and determining water leakage in the water distribution process. The difference between the estimated value estimated by the convolutional neural network and the measured value using the converted determination target data and the updated learning parameter. It is a leak detection method including calculation and determination of an abnormality based on the difference, and if it is an abnormality, it is determined that there is a leak in the water distribution process.

According to the present invention, it is possible to provide a technique capable of detecting a leak with higher accuracy.

FIG. 1 is a block diagram showing a configuration of a leak detection device 100 according to the first embodiment. FIG. 2 is a flowchart showing the processing of the leak detection device according to the first embodiment. FIG. 3 is a diagram showing a data conversion result in the first embodiment. FIG. 4 is a diagram showing an image of the learning model in the first embodiment. FIG. 5 is a diagram showing a data conversion result in the second embodiment. FIG. 6 is a diagram showing an image of the learning model in the second embodiment. FIG. 7 is a diagram showing a data conversion result in the third embodiment. FIG. 8 is a diagram showing an image of three-dimensional input data in the third embodiment. FIG. 9 is a diagram showing an image of the learning model in the third embodiment.

Hereinafter, embodiments for carrying out the present invention will be described with reference to the accompanying drawings.
However, the present invention is not limited to the description of the following embodiments.

(Embodiment 1)
FIG. 1 is a block diagram showing a configuration of a leak detection device according to the present embodiment.
The leak detection device 100 shown in FIG. 1 includes a data storage unit 101, a learning data acquisition unit 102, a determination target data acquisition unit 103, a data conversion unit 104, a learning parameter updating unit 105, and a learning parameter storage unit 106. , A water leakage determination unit 107.

FIG. 2 is a flowchart showing the processing of the leak detection device 100 according to the present embodiment.
FIG. 2A shows a learning processing flow, and FIG. 2B shows a determination processing flow.

First, the learning process flow process will be described.
In the learning processing flow shown in FIG. 2A, when the processing is started, the data storage unit 101 accumulates time-series data of the flow rate and pressure of the water distribution process to be detected (S11).
The learning data acquisition unit 102 acquires learning target data from the time-series data stored in the data storage unit 101 (S12).

The data conversion unit 104 converts the learning target data so that it can be learned (S13).
Here, the conversion of the training target data is performed so that learning by the convolutional neural network is possible.

FIG. 3 is a diagram showing the data conversion result of S13 in the present embodiment.
The data conversion unit 104 in the present embodiment simply converts the learning target data, which is time-series data, into two-dimensional data having a size (D × n) for the latest D days and the latest n times.
In FIG. 3, for example, when the data for the last 24 hours is input with the resolution in increments of 15 minutes for the data for the last 30 days, the data is converted into two-dimensional data of size (30 × 96). ..
This two-dimensional data becomes the input of the convolutional neural network.
The data conversion unit 104 creates this input data as many times as the number of determinations to be performed in one day.
That is, when the leak detection device 100 makes a determination every 15 minutes, 96 input data are required for learning model generation.

The learning parameter updating unit 105 learns using the converted learning target data converted by the data conversion unit 104 as an input (S14).
This training is performed by a convolutional neural network, and this training updates the model parameters.

FIG. 4 is a diagram showing an image of the learning model of S14 in the present embodiment.
The learning parameter update unit 105 in the present embodiment generates a model that estimates the time series data for the latest n times from the determination target time by the convolutional neural network.
The learning parameter update unit 105 inputs two-dimensional data of the size (D × n) of the latest D days and the latest n times generated by the data conversion unit 104, and uses the time series data of the latest n times as the objective variable. learn.
The number of learning models generated by the learning parameter update unit 105 is the number of times to be determined in one day.
The learning parameter update unit 105 generates, for example, 24 learning models when making a determination every hour, and 96 learning models when making a determination every 15 minutes.

The learning parameter storage unit 106 stores the updated learning parameters updated by the learning parameter update unit 105 (S15).

First, the learning process flow will be described.
In the determination processing flow shown in FIG. 2B, when the processing is started, the determination target data acquisition unit 103 acquires the determination target data for water leakage in the water distribution process to be detected (S21).
The data conversion unit 104 converts the determination target data so that it can be determined (S22).

The water leakage determination unit 107 determines water leakage in the water supply distribution process to be detected by using the converted determination target data converted by the data conversion unit 104 and the updated learning parameters (S23).
In the determination of water leakage in S23, the water leakage determination unit 107 first estimates the time series data for the latest n hours from the determination target time by the convolutional neural network.
The leak determination unit 107 calculates the difference between the estimated value by the convolutional neural network and the actually measured value for n times, determines an abnormality based on this difference, and determines that there is a leak if it is abnormal.
Here, the difference between the estimated value and the actually measured value may be quantified by the maximum value of n differences, the average error, the average squared error, or the like.

In the conventional technique, the feature amount according to the data is not extracted, and there is room for improvement in the prediction according to the data.
According to this embodiment, by performing deep learning using a convolutional neural network, it is possible to naturally extract features according to the data, and it is possible to detect water leakage with high accuracy according to the data. Become.
Further, according to the prior art, since a plurality of prediction determination methods are used, there is a problem that the calculation load increases according to the number of prediction determination methods to be used. However, according to the present embodiment, a plurality of prediction methods are used. Since the judgment method is not used, the calculation load can be suppressed.

(Embodiment 2)
In the first embodiment, the learning target data, which is time-series data, is simply converted into two-dimensional data having a size (D × n) for the latest D days and the latest n times, and is used as input data for the convolutional neural network. The present invention is not limited to this.
In this embodiment, an embodiment capable of suppressing the influence of noise will be described.

The water leakage detection device according to the present embodiment includes a data conversion unit 104a instead of the data conversion unit 104 of the first embodiment, and includes a learning parameter updating unit 105a instead of the learning parameter updating unit 105 of the first embodiment. Since only the difference is different and the others are the same as those of the first embodiment, the illustration is omitted and the description of the first embodiment is incorporated.

The data conversion unit 104a converts the learning target data so that it can be learned (S13), and converts the determination target data so that it can be determined (S22).
Here, the conversion of the training target data is performed so that learning by the convolutional neural network is possible in the same manner as in the data conversion unit 104.

FIG. 5 is a diagram showing the data conversion result of S13 in the present embodiment.
In FIG. 5, for example, frequency decomposition by Fourier transform is performed on time-series data (4 × 7 × 24 hours = 672 records) for the latest 4 weeks in 1-hour increments every 24 hours going back from the present to the past 4 weeks. When is applied, the length of the spectrogram display shown on the right side of FIG. 5 in the time direction, that is, in the vertical axis direction is 672 ÷ 24 = 28.
Further, when the amplitude when the frequency is 1 / 24, 1/48, 1/60, 1/648, 1/672 is extracted as the frequency component on the horizontal axis, the length in the horizontal axis direction is also 28, which is 28 × horizontal. Twenty-eight spectrogram displays are obtained.

The learning parameter updating unit 105a learns using the converted learning target data converted by the data conversion unit 104a (S14).
This learning is performed by the convolutional neural network as in the learning parameter updating unit 105, and the model parameters are updated by this learning.

FIG. 6 is a diagram showing an image of the learning model of S14 in the present embodiment.
In FIG. 6, only the input data is different from FIG.

The data conversion unit 104a of the leak detection device of the present embodiment converts the Fourier-transformed data obtained by Fourier-transforming the time-series data into two-dimensional data of time and frequency.
As described above, according to the present embodiment, in addition to the effect of the first embodiment, the weight coefficient of the non-dominant frequency component can be reduced and the influence of noise can be suppressed by Fourier transforming the input data.

(Embodiment 3)
In the first embodiment, the learning target data, which is time-series data, is simply converted into two-dimensional data having a size (D × n) for the latest D days and the latest n times, and used as input data for a convolutional neural network. In 2, the Fourier transform is applied, but the present invention is not limited thereto.
In the present embodiment, a mode in which data disappearing at the time of Fourier transform can be left in the second embodiment will be described.

The water leakage detection device according to the present embodiment includes a data conversion unit 104b instead of the data conversion unit 104 of the first embodiment, and includes a learning parameter updating unit 105b instead of the learning parameter updating unit 105 of the first embodiment. Since only the difference is different and the others are the same as those of the first embodiment, the illustration is omitted and the description of the first embodiment is incorporated.

The data conversion unit 104b converts the learning target data so that it can be learned (S13), and converts the determination target data so that it can be determined (S22).
Here, the conversion of the learning target data is performed so that learning by the convolutional neural network is possible in the same manner as in the data conversion units 104 and 104a.

FIG. 7 is a diagram showing the data conversion result of S13 in the present embodiment.
The data conversion unit 104b in the present embodiment performs data conversion using the Wavelet transform.
In the Fourier transform used in the second embodiment, the transformation result is one-dimensional vector data, but in the Wavelet transformation used in the present embodiment, the transformation result is obtained as two-dimensional data of the number of basis functions and the time.
This is because the coefficient value corresponding to each time can be obtained for each basis function.

When inputting and using the latest D days, for example, when wavelet transforming the data for the latest n times with k basis functions, it becomes three-dimensional data of size (n × k × D).
FIG. 8 is a diagram showing an image of three-dimensional input data in the present embodiment.

The learning parameter updating unit 105b learns using the converted learning target data converted by the data conversion unit 104b as an input (S14).
This training is performed by a convolutional neural network, and this training updates the model parameters.

FIG. 9 is a diagram showing an image of the learning model of S14 in the present embodiment.
As shown in FIG. 9, in the present embodiment, the wavelet-converted input data is input, and the change with time of the flow rate data for one day is output.

The data conversion unit 104b of the leak detection device of the present embodiment converts the Wavelet-converted data obtained by wavelet-converting the time-series data into three-dimensional data of the number of days, the time, and the number of basis functions.
As described above, according to the present embodiment, it is possible to leave data that disappears during the Fourier transform, and it is possible to utilize these.

The present invention is not limited to the above-described embodiment, and includes various modifications in which components are added, deleted, or converted from the above-mentioned configuration.

100 Leakage detection device 101 Data storage unit 102 Learning target data acquisition unit 103 Judgment target data acquisition unit 104 Data conversion unit 105 Learning parameter update unit 106 Learning parameter storage unit 107 Water leakage determination unit

Claims (4)

A data storage unit that stores time-series data of the flow rate and pressure of the water distribution process to be detected,
A learning data acquisition unit that acquires learning target data from the time-series data,
The judgment target data acquisition unit that acquires the judgment target data of water leakage in the water distribution process, and the judgment target data acquisition unit.
A data conversion unit that converts the learning target data so that it can be learned by a convolutional neural network, and converts the judgment target data so that it can be judged.
A learning parameter update unit that updates model parameters by learning using the converted learning target data converted by the data conversion unit, and a learning parameter update unit.
A learning parameter storage unit that stores updated learning parameters updated by the learning parameter update unit, and a learning parameter storage unit.
The difference between the estimated value estimated by the convolutional neural network and the measured value is calculated using the converted determination target data converted by the data conversion unit and the updated learning parameter, and the abnormality is determined based on this difference. A leak detection device including, in some cases, a leak determination unit for determining that there is a leak in the water distribution process. The water leakage detection device according to claim 1, wherein the data conversion unit converts the Fourier-transformed data obtained by Fourier-transforming the time-series data into two-dimensional data of time and frequency. The water leakage detection device according to claim 1, wherein the data conversion unit converts the Wavelet-converted data obtained by wavelet-converting the time-series data into three-dimensional data of the number of days, the time, and the number of basis functions. Accumulating time-series data on the flow rate and pressure of the water distribution process to be detected,
Acquiring learning target data from the time series data,
Converting the training target data so that it can be learned by a convolutional neural network,
Updating the model parameters by training using the converted converted training target data,
Remembering the updated updated learning parameters,
Acquiring data to be determined for leaks in the water distribution process,
Converting the determination target data so that it can be determined,
When the difference between the estimated value estimated by the convolutional neural network and the measured value is calculated using the converted converted determination target data and the updated learning parameter, and the abnormality is determined based on the difference, and the abnormality is found. A leak detection method comprising determining that there is a leak in the water distribution process.

Images