KR102680841B1 - Electric motor failure online diagnosis system by Machine learning based magneto register current sensor - Google Patents

Abstract

The present invention relates to a method in which an artificial intelligence device, which has learned the pattern of the current flowing in a defective motor, detects the failure of the motor in real time online. While operating a motor used in a power plant or industrial site, various defects are created inside the motor. While reproducing the fault, a current sensor is installed in the motor power supply circuit, the sensor is connected to a learning artificial intelligence device, the fault is experienced through the information received from the current sensor, and the learned artificial intelligence device is stored as data based on the learning. This is about a method of diagnosing electric motor failures online in real time.

Description

Electric motor failure online diagnosis system by Machine learning based magneto register current sensor}

The present invention relates to a method in which an artificial intelligence device, which has learned the pattern of the current flowing in a defective motor, detects the failure of the motor in real time online. During the operation of the motor used in power plants or industrial sites, various defects are provided and reported inside the motor. This is about a method of diagnosing motor failures online in real time through experience and learning of the accumulated data by acquiring and storing field reproduction through a current sensor installed on the electric motor power line.

Motor failures are diverse and complex. While there are some failures where the fault current is large enough for the protection relay to operate, the failure phenomenon is obvious, and there are also very subtle failures that do not even know that the failure has occurred. In particular, this minor malfunction does not have any significant effect on the operation of the motor, so there are cases where it does not have any particular effect on the operation of the motor even if it runs for several months or years. However, such a subtle and slowly progressing failure gradually weakens the insulation of the motor, eventually destroying the motor and leading to an accident, resulting in many economic and social losses. If you can detect micro-failures such as partial discharge or inter-layer short circuit in an electric motor, you can prevent the breakdown before it progresses into a major accident and significantly reduce recovery costs, so it is important to detect micro-failures and prevent large-scale failures. do.

In order to detect the above microfailure, the present invention uses a current sensor based on a magneto resistor. The magneto resistor is installed in the magnetic field of the current flowing in the bus bar, and based on the characteristic that the electrical resistance changes as the magnetic field changes, it mainly uses a magnetic disk and It has been used for reading using the magneto resistance phenomenon in the same data storage technology, and has been applied to various magnetic sensors because its magnetic field resolution is superior to that of Hall elements.

As described above, the conventional magneto resistor technology is related to the manufacturing method of a magneto resistance element with improved giant magneto impedance characteristics of Korean Patent Publication No. 10-2011-0115774 and the magneto resistance element manufactured accordingly, which involves a simple process of irradiating ions to a ferromagnetic material. Since it is manufactured using , the magnetic sensitivity and magneto impedance characteristics are improved, so it can be usefully used in magnetic heads for high-sensitivity information recording media and magnetic sensors for detecting fine magnetic fields. At Aichi Seiko Co., Ltd. in Japan, magneto impedance elements and magneto By applying impedance sensors, we have been developing technologies that mainly apply technology in which electromagnetic properties change due to magnetic fields acting from the outside.

However, in order for a magneto resistor-based current sensor to measure current and estimate it as a valid value, it must have various signal processing processes. In particular, noise processing is important for online real-time diagnosis of rotating bodies such as electric motors, and noise can be measured at maximum. It increases as the current value increases, but in the case of a conventional Hall sensor-based current sensor, the noise value is as large as about 1 to 3% of the maximum measurable current value, so devices that consume a lot of current and devices that use a little amount of current are tracked and measured at the same time. There was a problem that it became very difficult to analyze or measure very small current changes in devices that use a lot of current, and as a result, monitoring and prediction of devices using machine learning (artificial intelligence learning) became impossible.

Republic of Korea Patent Publication No. 10-2011-0115774 (published on October 24, 2011) Republic of Korea Patent Publication No. 10-1492825 (registration date February 6, 2015)

The present invention was created to solve the above problems, and the failure of an electric motor is determined by combining the unique pattern of the current flowing in the motor, the rise of the specific frequency, phase characteristics, harmonic characteristics, waveform distortion, etc. for each failure. Based on learning about the motor, it is classified and recognized, and if a once-recognized failure is experienced again during motor operation, it can be judged as a failure as already learned. By continuously learning the motor's normal state during motor operation, it is not normal but different from usual. When a condition is detected, it is recognized as defective and notified. When a learned failure situation occurs, it is notified as a failure as named during the learning process. Even if a micro failure such as partial discharge or interlayer short circuit occurs in the motor, real-time online detection is possible. In addition, it was invented for the purpose of preventing damage to electric motors and accidents.

Meanwhile, other unspecified purposes of the present invention will be additionally considered within the scope that can be easily inferred from the detailed description and effects below.

In order to achieve this purpose, according to one embodiment of the present invention, diagnosis of electric motor abnormality based on artificial intelligence learning is performed by installing a magneto resistor-based current sensor on the power line that supplies power to the motor, and measuring the current sensor from the current sensor. By acquiring, storing, and learning various current information (signals), an artificial intelligence device that recognizes the information pattern of the load current flowing when a motor fault occurs can diagnose motor abnormalities based on the learned information.

In the present invention, when a micro-failure such as partial discharge or inter-layer short circuit occurs in an electric motor, not only can the failure be prevented before it progresses into a major accident by detecting the failure online in real time, but the recovery cost can also be significantly reduced through this detection. There is.

Figure 1. Schematic diagram of measurement of magneto resistive-based current transducer (current sensor)
Figure 2. Sensitivity waveform of sensor to be applied to pattern recognition of the present invention
Figure 3. Integrated signal processing block diagram of the present invention
Figure 4. Overall configuration diagram of the system used in the present invention
Figure 5. Signal-to-noise (SNR) ratio of the present invention
Figure 6. Functional block diagram of the sensor module used in the present invention
Figure 7. Conceptual diagram of the current sensor module to be applied to the present invention
Figure 8. Deep learning data processing (stacking) concept diagram of the present invention
Figure 9. Mathematical representation of Training of the present invention

Hereinafter, the present invention will be described in detail.

In order for a device equipped with artificial intelligence to diagnose motor problems based on learning, it begins by measuring current using a magneto resistor-based current sensor. The current sensor has a cylindrical ring structure with a height of less than 100 mm, which is the size of a finger, so that when installed on a power supply line, one side of the current sensor can be opened and easily installed on the wire.

The principle of current measurement is that the magnetic field formed around the busbar through which current flows has the characteristic that the strength of the magnetic field changes depending on the size of the current. By installing a magneto-resistive element within a certain range of this magnetic field, the resistance of the element depends on the strength of the magnetic field. Current is measured based on this changing principle. The above current sensor is installed on the busbar that supplies power to the motor, and the resistance value of the magneto resistive element inside the current sensor changes in proportion to the change in the magnetic field formed by the current flowing in the busbar, so this element receives 5~10V DC. By applying a certain voltage, the correlation between the resistance size and the primary current is formulated and calculated based on the applied voltage and the current information flowing through the device as shown in Figure 9, and the current flowing through the primary bus is mathematically inverted to 1. Measure the current value of the car bus bar. The appearance and installation of the current sensor are shown in Figure 1.

In addition, to diagnose motor abnormalities based on artificial intelligence learning, current information measured from a magneto resistor-based current sensor installed on the power line that supplies power to the motor is acquired and stored, and the stored data is used to determine the load current flowing in the motor when a fault occurs. After comparative analysis with the pattern, it is recognized and the artificial intelligence device diagnoses motor problems based on the learned information.

Microfaults in electric motors can be recognized by classifying them based on learning about each defect by combining the unique pattern of the current flowing through the motor, the rise of a specific frequency, phase characteristics, harmonic characteristics, and waveform distortion. When a device equipped with artificial intelligence experiences a once-recognized malfunction again during electric motor operation, it determines that it is a malfunction as it has already learned. In addition, since the normal state of the motor is continuously learned during motor operation, if a state other than normal is detected, it is recognized as defective and notified through a smart device. If a learned failure situation occurs, the failure occurs as named during the learning process. It tells you the results.

The motor abnormality diagnosis device based on machine learning consists of a cylindrical current sensor the size of a finger and a current sensor that is ring-shaped and has an open structure on one side with a size of several tens of millimeters, that is, about a quarter of the size of the palm. It is largely composed of a signal processing unit (microcontroller) that receives current signals from and processes the signals.

In order for a current sensor to measure current and estimate it as a valid value, it must have various signal processing processes and processing noise is important. Noise increases as the maximum measurable current value increases, but the Hall sensor-based current sensor used conventionally has a noise value of about 1 to 3% of the maximum measurable current value, so it is suitable for devices that consume a lot of current and devices that use only a small amount of current. It becomes very difficult to simultaneously track or measure even the slightest current changes in devices that draw a lot of current, and as a result, monitoring and predicting devices using machine learning becomes impossible.

As an example, the motor load current reaches 300A, but the amplitude of the discharge current signal generated during partial discharge also exists in a very small current pattern of several mA, so existing current converters can produce currents of several tens of A and several mA. It is difficult to detect change. This fact is why the sensitivity of the current sensor is so important for load learning.

Figure 2 shows two types of current sensors that recognize mA current patterns in a large current load of 300A, and the sensitivity characteristics of the magneto resistor-based current sensor to be applied in the present invention are shown in (2) of Figure 2. When a load is in operation, the sensitivity of the current sensor, which can detect minute changes in the size of 1/1000 of the load, is a successful factor in recognizing the current information flowing in the load in detail. The magneto resistive used in the system of the present invention is The base current sensor shows characteristics suitable for recognizing fine patterns of current changes, as shown in (2) of FIG. 2.

The pattern recognition technology of the present invention applies technology to diagnose faults by learning, storing, and naming the current waveform pattern for each motor defect and storing it in a neural network circuit. One of the collective blocks of deep learning is artificial neural networks (ANNs). ANNs originate from the information processing of biological neurons. The "deep" part of deep learning is for stacking the layers of ANNs to process data. The basic concept diagram is shown in Figure 8.

In addition, in order to predict and diagnose the condition of the electric motor, a training process must be performed, and the deep learning model shows a mathematical function that learns how to classify data in Figure 9. The goal of Machine Learning, or Training, is to identify a set of parameters for this function. Training uses a parameter search algorithm (e.g., Adam, Adagrad, stochastic gradient reduction), and Back propagation, a parameter gradient identification algorithm. ), STPD (Spike Timing Dependent Plasticity), etc., the above algorithms work together to identify the best parameters and train patterns based on motor failure.

The present invention recognizes patterns using a Time Frequency Representation (TFR) method that effectively simultaneously expresses the acquired current information in time, space, and frequency space. TFR has a complex computational technique, so it is essential to develop optimization techniques to effectively embed it in a microcontroller. It provides the advantages of edge computing such as low-latency, low-power, fast response, resilience to network disturbance, etc. only when it is effectively embedded in a microcontroller. can do. In the case of TFR, the amount of information provided is the largest compared to the amount of data given, so it can minimize the amount of computation required and the data storage space that must be used in the cloud, while simultaneously increasing the accuracy of machine learning. The smaller the noise of the current signal, the more precise the TFR pattern is obtained. When performing the task of recognizing the original current pattern using the most widely used neural network-based reverse image search algorithm using the two TFR patterns, in the case of a signal with less noise It shows a recognition probability of over 99%, and in the case of TFR patterns with a lot of noise, the recognition probability drops to around 80%. It has been confirmed that the performance of the TFR expression technique can be optimized when used with the magneto resistor-based current sensor applied in the present invention. did.

Additionally, by using an algorithm that recognizes the current pattern that occurs at the start of the device's operation, it is possible to accurately calculate the time when sequential convolution begins for each device. Based on this unique technology, we apply a technique to accurately calculate the starting point of sequential convolutional operation for each fault, solving the problem of having to start sequential convolutional operation at multiple points due to synchronization uncertainty, drastically reducing the amount of computation required for parallel computation. Defects can be distinguished and recognized

The signal processing unit, which receives the current measured by the current sensor through wired and wireless communication and is installed within 10m from the sensor, uses an analog/digital conversion circuit that converts the analog signal of the current sensor into a digital value and frequency analysis of the digitally converted current information. It consists of a pattern recognition (signal processing) circuit that recognizes fine patterns, a load recognition neural network circuit equipped with machine learning technology, and a communication circuit that transmits processing information to an external smart device.

The overall system configuration diagram showing a block diagram of the sensing-signal processing module to be used in the present invention is shown in Figure 3. The microcontroller in charge of digital processing performs matched filtering for noise reduction and offset reduction using feedback. It is responsible for supporting various data communication protocols and has a built-in neural network processing function responsible for NILM (non-intrusive load monitoring). It has a built-in function that automatically transmits data to a database in the cloud using various built-in communication support functions or provides results by visualizing and imaging them with various smart devices.

Each module includes a magneto resistor-based current sensor module, an analog input terminal that receives the analog input signal received from the sensor, a signal processing module including a microcontroller consisting of a power circuit, amplification and noise removal filtering circuit, and data consisting of Bluetooth, Wi-Fi, and Ethernet. It consists of a communication circuit and a remote monitoring system, and the overall system configuration is shown in Figure 4.

The current measurement value of the magneto resistor-based current sensor is transmitted in analog form to the microcontroller, which is a signal processing unit, and the microcontroller, which includes an A/D conversion unit that converts the analog signal received from the current sensor into digital, uses a matched filter to reduce noise. It performs the task of reducing noise and realizes ultra-precision resolution through this noise reduction technology. In addition, in order to minimize resolution degradation caused by noise generated from the analog amplifier built into the signal processing unit including the microcontroller, a signal up-conversion function is used at the analog amplifier stage. As shown in FIG. 5, the signal generated from the sensor is By raising the analog amplifier stage to a high frequency band where flicker noise does not exist, the signal-to-noise ratio (SNR) is significantly improved, achieving high resolution with little effect from noise.

In the signal processing unit, the signal is amplified at the analog amplifier stage to solve the noise problem, and at the same time, noise shaping technology is applied to shape most of the noise in the frequency space so that it is included in a specific frequency band. Afterwards, noise is minimized through optimal detection, which removes shaped noise from the digital signal processing circuit. Additionally, in the case of offset, the offset value is accurately predicted using technology that measures and predicts it in real time, and then the offset is compensated in real time using a feedback circuit. Since the offset value of the sensor continues to change depending on the external environment such as temperature and the aging of the motor itself, the present invention applies the real-time compensation technology developed and applied as described above. The functional block diagram of the sensor module used in the present invention is shown in Figure 6, and the conceptual diagram of the current sensor module is shown in Figure 7.

The signal processing unit, which includes a microcontroller installed within a certain distance, for example, 10 m, from the current sensor, has not only a current data signal processing function, but also a machine learning function using a neural network, and a wired, It has a built-in communication chip that can transmit signals wirelessly, so it can transmit signals to external smart devices or monitoring devices.

The processing method of the input electrical signal to be used in the neural network responsible for AI and non-intrusive load monitoring (NILM) using machine learning is to use data patterns in space and time, but when only space-time data is used, the data Since the amount of useful information that can be extracted is limited compared to the amount of , a method of using signals in the frequency space is applied. By applying the time frequency representation (TFR) method, which effectively expresses signals simultaneously in time and space and frequency space, edge computing such as reducing delay, realizing fast response speed, low power, and effective response to emergencies such as communication network down It can provide advantages.

In the case of TFR, the amount of information provided is the largest compared to the amount of data given, so it is a very important technology for minimizing the amount of computation required and the data storage space that must be used in the cloud while simultaneously increasing the accuracy of machine learning. If the signal has less noise, a much more precise TFR pattern is obtained. When recognizing the original current pattern of two TFR patterns using the most widely used convolutional neural network (CNN)-based reverse image search algorithm, a recognition probability of over 99% can be secured for signals with low noise. , In the case of TFR patterns with a lot of noise, the recognition probability decreases to around 80%. Performance is optimized when the developed TFR expression technique is used with a current sensor.

An external smart device notifies the operator of motor defect occurrence information based on signals transmitted from the on-site signal processing unit.

current sensor
current carrying busbar

Claims (7)

In the online fault diagnosis system based on magneto resistor current sensor-based motor fault learning,
By applying a voltage of 5 to 10 V DC based on a magneto resistor installed on the power supply line, the resistance value of the internal magneto resistive element is changed. The current pattern is imaged based on the applied voltage and the current information flowing through the element, and the current pattern is imaged within 100 mm. A current sensor having a cylindrical ring structure with a height;
An A/D conversion unit that converts the analog signal measured from the current sensor into digital,
Through frequency analysis of the converted digital signal, fine patterns are recognized and motor partial discharge and interlayer short circuits, including specific frequency rise, phase characteristics, harmonic characteristics, and waveform distortion, are diagnosed and occur at the start of the motor operation. A pattern recognition circuit that recognizes the current pattern and calculates the start time for each fault, a filtering circuit that amplifies the signal, removes noise, and compensates for offset using noise shaping and feedback, a neural network circuit equipped with machine learning technology, and the above circuit. A signal processing unit including a microcontroller and communication circuit that processes signals;
An online fault diagnosis system based on magneto resistor current sensor-based motor fault learning, characterized in that it consists of a smart device that displays the processed signal.

delete According to clause 1,
The communication circuit is an online fault diagnosis system based on magneto resistor current sensor-based motor fault learning, characterized in that it consists of Bluetooth, Wi-Fi, and Ethernet,
According to clause 1,
A neural network circuit equipped with machine learning technology is an online fault diagnosis system based on motor fault learning based on a magneto resistor current sensor, which is characterized by applying the time frequency representation (TFR) method.
delete delete delete