KR20230167589A - Fault Diagnosis System based on Deep Learning Model for Smart IoT Valve - Google Patents

Abstract

The present invention relates to a deep learning model-based fault diagnosis system for smart IoT valves. Specifically, equipment data is extracted from DDC (Digital Direct Control) through sensors, and then time series data is converted through FFT (Fast Fourier Transform). This is about a deep learning model-based fault diagnosis system for smart IoT valves that converts to frequency data and uses signal similarity analysis and a graph neural network model to determine signal failure and classify the type.

Description

Fault Diagnosis System based on Deep Learning Model for Smart IoT Valve}

The present invention relates to a deep learning model-based fault diagnosis system for smart IoT valves. Specifically, equipment data is extracted from DDC (Digital Direct Control) through sensors, and then time series data is converted through FFT (Fast Fourier Transform). This is about a deep learning model-based fault diagnosis system for smart IoT valves that converts to frequency data and uses signal similarity analysis and a graph neural network model to determine signal failure and classify the type.

Recently, the sensor-based fault diagnosis system applicable to the IoT (Internet of Things) field is a technology that diagnoses equipment faults through real-time monitoring, and is attracting attention as an important technology in the IoT environment by preventing losses that may be caused by faults in advance. there is.

In particular, as deep learning algorithms are rapidly developing, deep learning-based classification models are being widely studied for fault diagnosis.

Meanwhile, deep learning models for fault diagnosis mainly use the convolutional neural network (CNN) structure.

However, this convolutional neural network model uses a kernel and has the characteristic of only focusing on the surrounding context, so it may not be appropriate for data that does not have a Euclidean distance structure.

In particular, general convolutional neural network structures may not be able to properly extract features of multivariate time series data. Therefore, there is a need to develop models with a neural network structure that can be applied to a data environment with a structure that does not have a Euclidean distance.

Registered Patent No. 10-2364215

The present invention was created to solve the above problems. The present invention converts collected data from smart building equipment into frequency axis signals based on FFT (Fast Fourier Transform) and analyzes the similarity. The purpose is to provide a deep learning model-based fault diagnosis system for smart IoT valves that generates a graph in the form of a matrix and conducts learning based on the graph matrix to determine whether the equipment is broken and facilitate the extraction of data features.

In addition, the present invention applies the graph generated through the conversion unit and the generation unit to the graph convolutional neural network model to infer the failure type of smart building equipment by considering the characteristics of the entire context and using the information of the interrelationship. The purpose is to provide a deep learning model-based fault diagnosis system for smart IoT valves that can improve performance.

This purpose of the present invention is to include a measuring unit that measures data from a smart IoT valve; A conversion unit that converts time series data into frequency; A generator that generates a graph matrix through similarity analysis using the converted data; A storage unit that stores a graph matrix; A learning unit that trains a graph convolutional neural network model based on the stored graph matrix; An inference unit that determines whether equipment/equipment is broken and classifies the type based on the learned graph convolutional neural network model and measured data; This is achieved by a deep learning model-based fault diagnosis system for smart IoT valves according to the present invention, which includes.

In addition, the measuring unit is characterized by measuring equipment information by detecting signals converted into digital/analog from the smart building DDC (Direct Digital Control).

Additionally, the conversion unit converts the time series data measured by the measurement unit into frequency data.

In addition, the generator is characterized in that it generates a graph matrix by analyzing the similarity between the converted frequency data.

In addition, the storage unit is characterized by storing the graph matrix constructed in the generation unit.

In addition, the learning unit is characterized in that it learns a graph convolutional neural network model based on the graph matrix constructed in the storage unit.

In addition, the inference unit determines whether the equipment is broken and classifies the type using the learned graph convolutional neural network model and converted frequency data.

The deep learning model-based fault diagnosis system for smart IoT valves according to the present invention includes a measuring unit that measures data of the smart IoT valve; A conversion unit that converts time series data into frequency; A generator that generates a graph matrix through similarity analysis using the converted data; A storage unit that stores a graph matrix; A learning unit that trains a graph convolutional neural network model based on the stored graph matrix; An inference unit that determines whether equipment/equipment is broken and classifies the type based on the learned graph convolutional neural network model and measured data; Including, the data from smart building equipment collected is converted into a frequency axis signal based on FFT (Fast Fourier Transform), the similarity is analyzed, a graph in the form of a matrix is created, and learning is performed based on the graph matrix. Applying a graph convolutional neural network to the failure type classification model has the effect of improving the inference accuracy of failure types, and by quickly diagnosing failures in smart IoT equipment, the occurrence of failures is significantly reduced, reducing financial losses that may be caused by failures. It has the effect of reducing.

Figure 1 is a block diagram of a deep learning model-based fault diagnosis system for smart IoT valves according to an embodiment of the present invention.
Figure 2 is a flowchart of the steps of learning a graph convolutional neural network through a learning algorithm of a deep learning model-based fault diagnosis system for smart IoT valves and generating and classifying a fault type classification model according to an embodiment of the present invention.
Figure 3 is a graph showing loss images of training data and verification data according to the epoch performance of the GCN (Graph Convolutional Network) learning model.
Figure 4 is a graph showing the correlation between flow rate and differential pressure

The above-described purpose, features and advantages will become clearer through a detailed description of the deep learning model-based fault diagnosis system for smart IoT valves according to the present invention, which is described later with reference to the attached drawings, and accordingly, the technical field to which the present invention belongs. A person skilled in the art will be able to easily implement the technical idea of the present invention.

Additionally, in describing the present invention, if it is determined that a detailed description of known technologies related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description will be omitted. Hereinafter, various embodiments of the present invention will be described in detail with reference to the attached drawings.

Figure 1 is a block diagram of a deep learning model-based fault diagnosis system for smart IoT valves according to an embodiment of the present invention.

Referring to Figure 1, the deep learning model-based fault diagnosis system (A) for smart IoT valves according to an embodiment of the present invention includes a Direct Digital Control (DDC) 100 and a graph convolutional neural network model generation module 140. and a network 130 connecting them.

The DDC (Direct Digital Control) 100 includes a measurement unit 105, a conversion unit 110, and an inference unit 115.

In addition, the DDC 100 measures equipment signals received through integrated sensors, converts the measured time axis signal into a frequency axis signal, and types the failure type through a failure type classification model generated through the converted data. Classifies and outputs the classified results.

The graph convolutional neural network model generation module 140 includes a creation unit 145, a storage unit 150, and a learning unit 155.

In addition, the graph convolutional neural network model generation module 150 uses the time series data converted from the DDC 100 to convert it to frequency, builds a graph matrix, and uses a convolution neural network learning-based graph neural network model. A failure type classification model is generated, the generated failure type classification model is learned, and transmitted to the inference unit of the DDC 100 through the network 130.

Next, each configuration of the DDC (100) and the graph convolutional neural network model generation module (140) will be described in detail.

The measuring unit 105 of the DDC (100) measures the signal coming through the pipe/line to the DDC (Direct Digital Control) (100) attached in the smart building through the IoT valve (105a) and measures the signal thus measured. The signal is transmitted to the conversion unit 110, and the conversion unit 110 converts the time axis signal measured in the measurement unit 105 into a frequency axis signal, and the converted signals are converted into a graph convolutional neural network model generation module. It is transmitted to the generating unit 145 of (140).

The data of the IoT valve 105a is set and measured as data items for flow rate, pressure, and temperature, and values (data) outside each operating range are set as failures and notified through the inference unit 115. .

The generation unit 145 generates a graph matrix using signals obtained from the conversion unit 110 through the network 130 and transmits it to the storage unit 150, and the storage unit 150 generates the graph matrix 145. The graphs created are saved.

When necessary, the learning unit 155 learns a failure type classification model by loading and using data in the graph convolutional neural network model of the storage unit 150 stored in the storage unit 150. The failure type classification models learned in this way are transmitted back to the inference unit 160 of the DDC (100) through the network 130.

After receiving the failure type classification model through the network 130, the inference unit 160 learns the input abnormal signal when an abnormal signal that falls outside the range of the normal signal is input from the conversion unit 110. The failure type classification result of the equipment is output using the failure type classification model.

Additionally, the converter 110 can convert a time axis signal into a frequency axis signal using a frequency transform (Fourier Transform).

At this time, the Fourier transform may be implemented with a Fast Fourier Transform (FFT) algorithm, and is not necessarily limited thereto, but hereinafter, the frequency transform algorithm is referred to as the FFT algorithm.

The FFT algorithm uses equation 1 below.

Here, f _i is a value (numerical value) determined according to Equation 1, k is an integer assumed from 0 to n-1, x is a discrete number and is assumed to be a complex number, n is the number of data, Time series data can be entered into x with a certain length and calculated as a signal on the frequency axis with a complexity of O(nlogn), and 2πf is the frequency.

And, the generator 145 of the graph convolutional neural network model generation module 140 of the deep learning model-based fault diagnosis system (A) for smart IoT valves according to an embodiment of the present invention uses the received data to create a neighbor matrix. An Adjacency Matrix and a Node Feature Matrix are constructed, and the constructed graph is delivered to the storage unit 150.

At this time, the node feature matrix is constructed using frequency axis data as a node of the matrix, and Equation 2 below is used to construct a neighborhood matrix.

here means the adjacency matrix and Each refers to a signal expressed as a node in a graph. Through Equation 2 above, the distance between nodes can be obtained to obtain similarity, and the similarity between all nodes can be obtained to obtain an adjacency matrix between nodes.

To verify the performance of the Graph Convolutional Network (GCN) learning model, the accuracy can be checked using Equation 3.

TP (True Positive): When predicted to be the correct answer, the actual value is the correct answer.

TN (True Negatives): When the actual value is incorrect when predicted to be incorrect.

FP (False Positives): When the correct answer is predicted to be correct, the actual value is incorrect.

FN (False Negatives): When the actual value is the correct answer when predicted to be incorrect

Figure 2 is a flowchart of the steps of learning a graph convolutional neural network and generating and classifying a failure type classification model through a learning algorithm according to an embodiment of the present invention.

Referring to FIG. 2, the measurement unit 105 of the DDC 100 can acquire the signal of the facility/device (S105), and the conversion unit 110 can convert the time axis signal into a frequency axis signal ( S110), the generation unit 145 may generate a graph based on the converted signal (S145), and the learning unit 155 may learn a failure type classification model through a graph convolutional neural network (S155). The inference unit 115 can classify the failure type using the learned model (115) and outputs the type when a failure occurs (S116).

The operation of the deep learning model-based fault diagnosis system for smart IoT valves according to an embodiment of the present invention having such a configuration will be described in detail below.

(data collection)

The flow rate within the IoT valve (105a) installed in the measuring unit 105 of the DDC (100) was measured to measure the reference instantaneous flow rate value within the flow test device. A 100% opening command for the IoT valve (105a) was executed through the control device (not shown) of the DDC (100) and the standard instantaneous flow rate value was measured in the flow test device.

This was measured by measuring the pressure within the IoT valve (105a) installed in the measuring unit 105 of the DDC (100). The pressure within the IoT valve (105a) connected to the DDC (100) was measured and the pressure value was transmitted to the display unit connected to the IoT valve (105a) and measured.

The temperature inside the IoT valve (105a) installed in the measuring unit 105 of the DDC (100) was measured and measured. This was measured by monitoring the temperature value in the chamber related to temperature monitoring in the IoT valve (105a) linked to the DDC (100).

In this way, the data of the IoT valve 105a installed in the measuring unit 105 of the DDC 100 was set to data items for flow rate, pressure, and temperature.

The IoT valve 105a measured in the example had a diameter of 15 mm, a flow rate of 25 l/min, a pressure of 25 kPa, and a valve surface temperature of 20°C.

(data settings)

To conduct an experiment applicable to situations similar to the IoT valve failure above, a simulation was conducted based on CWRU (Case Western Reserve University), a public dataset.

At this time, the CWRU dataset was divided into training, validation, and test datasets at a ratio of 7:2:1. To preprocess the sensor data sequence obtained from the CWRU dataset, preprocessing was performed using the Python libraries Pandas and Numpy. The CWRU sequences obtained through this process each have a length of 2048 and consist of two channels.

Time series data for learning to detect failures have a lot of variability due to noise. In order to increase learning efficiency by alleviating data variability, the data is converted to the frequency axis, and this conversion to the frequency axis is performed in the conversion unit 110 of the DDC 100.

FFT (Fast Fourier Transform) is used as a technique for frequency axis conversion. The frequency axis signal is obtained by applying Fourier transform to the time series sequence using fft from Scipy, a Python library used for engineering. By using FFT, the obtained frequency data becomes half of the original sequence by removing the negative frequency side. Therefore, the sequences have a length of 1024, which is half of the existing length of 2048, but the number of channels is maintained.

To analyze data failure types, the degree of similarity between data is analyzed. To analyze the similarity between data, we use the distance analysis method in spatial of Scipy, a Python library. The Euclidean distance between data is analyzed and additional weights are given according to the degree of similarity to convert the neighbor matrix into a weight matrix.

The function of converting this neighborhood matrix into a weight matrix is performed in the generation unit 145 of the graph convolutional neural network model generation module 140.

(Artificial neural network structure)

The storage unit 150 stores the graph matrix constructed in the generation unit 145, and the learning unit 155 learns a graph convolutional neural network model based on the graph matrix constructed in the storage unit 150, When necessary, the learning unit 155 learns a failure type classification model by loading and using data in the graph convolutional neural network model of the storage unit 150 stored in the storage unit 150. The failure type classification models learned in this way are transmitted back to the inference unit 160 of the DDC (100) through the network 130.

The artificial neural network structure executed in the learning unit 155 uses Specktral, a graph module using tensorflow, an open source deep learning framework, to obtain a convolution using a matrix between the feature matrix of the graph and the neighboring structure, and uses GCN (Graph Convolutional Network). Network) was built. The output of each layer is a set of feature graphs and neighbor graphs with calculated weights. The weight calculation process is similar to the convolution calculation process, and by proceeding with the calculation process, information between neighbors can be calculated. It is possible to perform the role of a readout layer by adding a global connection layer to ensure invariance in permutation before final output. The final output is the probability that the node belongs to each class.

(Artificial neural network learning)

Learning of the GCN model learned in the learning unit 155 was performed for 400 epochs. For appropriate gradient convergence, the batch size was set to 64 and the popular Adam optimizer was applied. The learning rate was set to 1e-3, and other parameters used the initial settings provided by tensorflow. Learning was conducted based on CPU, and the CPU specifications of the PC where learning was conducted were i5-8400.

(Learning model performance verification)

Accuracy is used to verify the performance of the GCN learning model, which is as shown in Equation 3 above. The results of the GCN learning model accuracy were confirmed through a total of 5 tests, and the results are shown in Table 1. The loss image of the training data and verification data according to the epoch is shown in Figure 3. It can be seen that as the number of learning times according to the epoch increases, the loss value almost converges to O, so the GCN learning model is properly learned through the image. You can confirm that this is happening.

Test 1 Test 2 Test 3 Test 4 Test 5 average accuracy(%) 99.8 100 99.5 98.9 99.7 99.58

The setting range of the flow rate used in this example is shown in <Table 2> below.

The Cv value indicated in <Table 2> below refers to the flow rate value used when selecting the valve.

In the present invention, the pressure measurement range is 0 to 1.0 MPa ^*3 and the temperature measurement range is 0 to 80°C.

The temperature of the IoT valve (105a) was measured by measuring the surface temperature of the IoT valve (105a). The measurement range is 0 to 80°C, and the precision of the temperature difference is -25 to +. It is ±1.0℃ ^*5 at 40℃.

Calorie calculations applied in the present invention are shown in <Table 3> below.

As described above, the deep learning model-based fault diagnosis system for smart IoT valves according to the present invention converts the collected data of smart building equipment into a frequency axis signal based on FFT (Fast Fourier Transform) and analyzes the similarity. By creating a graph in the form of a matrix and learning based on the graph matrix, the graph convolutional neural network is applied to the failure type classification model, which has the effect of improving the inference accuracy of the failure type and quickly diagnosing failures in smart IoT equipment. By doing so, the occurrence of breakdowns is significantly reduced, thereby reducing financial losses that may be caused by breakdowns.

The deep learning model-based fault diagnosis system for smart IoT valves according to the present invention can be said to be an invention with industrial applicability because it can be said that it is possible to repeatedly manufacture the same product in the industrial field that manufactures general smart IoT valves.

100: DDC (Direct Digital Control)
105: measuring unit
110: conversion unit
115: reasoning unit
130: network
140: Graph convolutional neural network model generation module
145: Generation unit
150: storage unit
155: Learning Department

Claims (7)

In a deep learning model-based fault diagnosis system for smart IoT valves,
A measuring unit that measures the equipment status;
A conversion unit that converts the measured signal from time data to frequency data;
A generation unit that generates a graph matrix using the converted data;
A storage unit that stores the constructed graph matrix;
A learning unit that trains a failure classification model using a graph matrix;
an inference unit that infers the failure state of equipment using the gasped model;
Deep learning model-based fault diagnosis system for smart IoT valves, comprising:
According to clause 1,
The measuring unit,
A deep learning model-based fault diagnosis system for smart IoT valves, characterized by detecting and measuring signals from smart IoT valves connected through DDC.
According to clause 1,
The conversion unit,
A deep learning model-based fault diagnosis system for smart IoT valves, characterized by converting the time axis signal of the measured equipment into a frequency signal.
According to clause 1,
The generating unit,
A deep learning model-based fault diagnosis system for smart IoT valves that generates a graph matrix by analyzing the similarity of the converted frequency signal.
According to clause 1,
The storage unit,
Deep learning model-based fault diagnosis system for smart IoT valves, characterized by storing constructed graph matrix data According to clause 1,
The learning department,
A deep learning model-based fault diagnosis system for smart IoT valves, characterized by learning a graph convolutional neural network model using stored graph matrix data.
According to clause 1,
The inference unit,
A deep learning model-based fault diagnosis system for smart IoT valves, characterized by determining and classifying device failures through facility data measured as input to a learned graph convolutional neural network model.