KR101040926B1 - Method for monitoring machine condition - Google Patents

Abstract

PURPOSE: A damage diagnosis method for a machine is provided to allow easy recognition of damage of a machine by analyzing the probability of a diagnosis signal to a normal signal model and making a portion with high damage probability prominent. CONSTITUTION: A damage diagnosis method for a machine comprises steps of: measuring a diagnosis signal from a normal machine and applying hidden Markov model algorithm to the measured diagnosis signal to obtain a normal signal model; measuring a real-time diagnosis signal from a diagnosis target portion of an object(300) and applying hidden Markov model algorithm to the diagnosis signal to obtain the probability to the normal signal model; determining that a portion where the probability is nonuniform compared to the normal signal model is damaged; and adding the probability to the normal signal model when the determined damage portion is not actually damaged.

Description

Fault diagnosis method of machine {METHOD FOR MONITORING MACHINE CONDITION}

The present invention relates to a method for detecting the occurrence of a defect in a machine, in particular, by measuring the diagnostic signal at the steady state and the actual diagnosis and analyzing it using a Hidden Markov Model (HMM) method, for example The present invention relates to a method for diagnosing a defect of a machine, such as accurately detecting the occurrence and location of a weld defect.

As a method of detecting a defect of a machine, a non-destructive test method for detecting a defect by applying a diagnostic signal such as current, ultrasound, or the like to a structure or by installing various sensors and comparing the diagnostic signal with a normal signal measured.

Hereinafter, a welding defect diagnosis method will be described as an example of a method for detecting a defect of a machine according to the prior art.

Automatic welding systems using robots or automatic welding devices are widely used in the industry. In welding work, due to various reasons such as poor pretreatment of welding area, uneven material of welding rod, and poor supply of welding gas, insufficient penetration or lack of blow-hole occurs in welding area. Various welding defects are generated.

Due to the nature of the automatic welding operation, it is impossible for the operator to continuously monitor the welding operation, and it is sometimes difficult to visually identify welding defects such as pores generated inside the welded portion. Therefore, even if a welding defect occurs, it is difficult to detect it completely. For example, a door panel or a vehicle body frame of an automobile factory is press-molded, and parts of various sizes and shapes are automatically welded by a robot or the like, and then coated and assembled. It is difficult to substantially detect weld defects during this continuous fabrication process. In addition, when welding the thick steel plate of heavy construction equipment, the welding work is performed a plurality of times on the weld line, and even if an abnormality occurs in the first welding, it is often judged to be normal after the second and third welding. Therefore, it is difficult to detect weld defects. In order to detect weld defects that do not appear in appearance, non-destructive inspection such as ultrasound or X-ray examination is performed, but such non-destructive inspection is time-consuming and expensive and difficult to apply to complex weld shapes.

Currently, the arc welding monitoring method that is commercially available is used by measuring the supply speed, welding current, welding voltage of the welding wire during welding. For example, the welding current and the voltage signal during normal welding, the transmission speed of the welding wire are measured, and the welding current and the welding voltage when observable welding defects such as burn-through and pore occur are observed. Compare the current and voltage signals when welding defects occur, and if there is a significant difference, detect the abnormality. For example, pre-modeled rules for visually observable signal characteristics such as changes in the minimum and average values of current and voltage signals during actual welding, occurrence of abnormal peaks, number of abnormal peaks per unit time, fluctuations in the transmission speed of the welding wire, and the like. Compared to the detection of abnormality.

1 is a block diagram showing an example of the arc welding monitoring apparatus according to the prior art. As shown in FIG. 1, the conventional welding detection system includes a first controller 11, a power converter 12, a first voltage sensor 13, and a first current sensor 14 of the arc welding power supply device 10. A second controller 23, a second voltage sensor 21, and a second current sensor 22, which are mounted outside of the arc welding power supply device 10, are included. Since they are electrically connected in the arc welding execution unit 30, the second voltage sensor 21 and the second current sensor 22 can detect a current and voltage change during arc welding. The detection of welding defects is detected by comparing the change of current and voltage signal during actual welding with the data of normal welding.

However, in the actual welding operation, irregular and high intensity noise is frequently generated and the welding signal fluctuates even during the same operation. Therefore, the above-described welding defect diagnosis method has a very low success rate and frequently fails to detect a defect.

The present invention is to solve the above-mentioned problems, by comparing the probability value calculated by applying the steady state signal model modeled using the hidden Markov model and the actual diagnostic signal measured during the actual diagnosis to the normal signal model The purpose is to easily detect defects in the machine.

In addition, the present invention has another object of detecting a defect generated during the specific operation in real time by measuring the diagnostic signal in real time and converting in real time when a specific operation and diagnosis are performed at the same time, such as when detecting a welding defect.

In addition, if there is a detection error, such as in the case where the actual diagnosis section is detected as a defect but there is no defect, there is another purpose to gradually strengthen the detection of the defect of the machine by updating the normal signal model.

In order to achieve the above object, the present invention measures a diagnostic signal for diagnosing a machine state from a non-defective machine, and forms a normal signal model for a time domain by applying a hidden Markov model algorithm. In the normal signal modeling step, the diagnostic signal is measured at the diagnosis target site of the diagnosis target machine in real time, and the normal signal model is applied by applying a hidden Markov model algorithm applied in the normal signal modeling step. Probability value calculation step of calculating a probability value for and a machine state diagnostic method comprising a defect determination step of determining that a defect occurred in a section in which the probability value is not kept constant compared to the normal signal model. The method may further include a normal signal model updating step of updating the probability value by adding the probability value to the normal signal model when an actual defect does not occur in the section where the defect is determined in the defect determination step.

In addition, the present invention is an example of the machine state advancing method as described above, by measuring the welding signal from the welding line of the welding object over the time domain, and confirms whether the welding line of the welding object is defective by normal welding without defect The normal welding signal modeling step of extracting the welding signal as a normal signal only for a section, and forming a normal welding signal model for the time domain of the entire section by applying the hidden Markov model algorithm, while welding the welding object. Probability for the normal welding signal model by measuring a welding signal of a welding target part in real time and applying a hidden Markov model algorithm used to form the normal welding signal model of the welding signal of the welding target part measured in real time. A probability value calculating step of calculating a value, and the probability value being the normal welding signal. The present invention provides a welding defect diagnosis method including a defect determination step of determining that a defect has occurred in a section which is not kept constant compared to a model. The welding signal may be a voltage signal or a current signal.

In the normal welding signal modeling step, it is possible to determine whether a defect occurs in the weld line through at least one of visual inspection, nondestructive inspection, and cut surface inspection.

In the normal welding signal modeling step, it is preferable to further include a normal welding signal preprocessing step of preprocessing the data of the section determined to be normal with a time function, the normal welding signal preprocessing step may be performed through filtering or time weighting. have.

After the normal welding signal preprocessing step, it is preferable to further include a normal welding signal conversion step of converting the data into a characteristic vector sequence using a conversion equation, the characteristic vector sequence is an average, standard deviation, energy, power spectrum coefficient , Filter bank, capstrum coefficient, wavelet coefficient, coefficient of AR model, or a combination thereof.

The probability value for the normal welding signal model and the normal welding signal model of the welding signal of the welding target portion is obtained from any one of an initial state probability, a state transition probability, a symbol observation probability, a state average vector, and a state covariance matrix. The hidden Markov model algorithm may be any one of a forward algorithm, a backward algorithm, a rate forward algorithm, a rate backward algorithm, a K mean clustering technique, a modified K mean clustering technique, an EM algorithm, and a Baum-welch algorithm.

In the probability value calculating step, it is preferable to further include a welding signal pre-treatment step for pre-processing the welding signal measured in real time at the welding target site by the time function, the welding signal pre-processing step for the welding target site is filtering or It is preferably done via time weighting.

Preferably, the welding target site welding signal conversion step of converting the time function preprocessed in the welding signal preprocessing step into the characteristic vector streams using a conversion equation, the characteristic vector streams may include an average, standard deviation, A method for diagnosing weld defects comprising any one or a combination of energy, power spectral coefficients, filter banks, capstrum coefficients, wavelet coefficients, and AR model coefficients.

The normal welding section may be obtained by collecting a plurality of normal signal extraction sections from the welding line, or may be obtained by collecting the normal welding sections obtained by repeatedly welding the welding line.

The method may further include a normal signal model updating step of updating the probability value by adding the probability value to the normal welding signal model when an actual defect does not occur in the section in which the defect is determined in the defect determination step.

As described above, according to the present invention, by analyzing the probability value for the normal signal model of the diagnostic signal of the diagnosis target region converted from the diagnostic signal to the hidden Markov model, a section having a high probability of occurrence of a defect can be distinguished. Can easily be detected.

In addition, the present invention can detect the occurrence of a defect during the specific operation in real time by measuring the diagnostic signal in real time and converting in real time when the diagnosis is performed at the same time as a specific operation, such as when detecting a welding defect, such an immediate determination The defective rate can be reduced.

In addition, if a detection error is detected in the diagnosis section but there is no defect in the diagnosis section, there is a detection error so that the probability value of the corresponding diagnosis signal can be updated to a normal signal model. You can get it.

1 is a block diagram schematically showing an example of a welding defect detection apparatus according to the prior art,
2 is a conceptual diagram illustrating an example of a welding apparatus to which a welding defect detecting method according to the present invention is applied;
3 is a flowchart illustrating an embodiment of modeling a normal welding model of a welding defect diagnosis method according to the present invention;
4 is a flowchart illustrating an embodiment of generating a welding signal of a welding target portion as a probability value for a normal signal model of the welding defect diagnosis method according to the present invention;
FIG. 5 is a graph showing a voltage graph of a welding target site and a probability value for a normal signal of a voltage signal of the welding target site obtained by the process of FIG. 4.

Hereinafter, a method for diagnosing a defect of a machine according to the present invention will be described in detail with reference to FIGS. 2 to 5.

In this embodiment, a process of detecting a defect of a welded part among defects of a machine will be described.

2 schematically illustrates a welding robot to which a welding defect diagnosis method according to the present invention is applied. As shown in FIG. 2, the welding robot 200 has a plurality of arms 220, 230, 240 extending from the base 210 and having a plurality of joints 211, 212, 221, 231, 241. To move the welding torch 250 to a desired position. The welding torch 250 is also rotatable with a welding torch joint 251. A welding wire 201 is supplied to the welding torch 250, which is guided to the welding robot 200 by the wire guide 202 and supplied to the welding torch 250 by the torch cable 260.

The welding robot 200 includes a controller 140 for controlling the posture of the welding robot 200 and controlling a supply speed of the welding wire; A welding machine 130 having two terminals 132 and 133 connected to the welding torch 250 and the welding object 300 fixed to the jig 310, respectively; Computer 110 having a user interface. Voltage and current can be measured at the two terminals 132 and 133 of the welder 130, and the measured voltage and current signals are digitized to the A / D converter 120 via the amplifier 131, respectively, and then the computer 110. Is displayed in real time.

In order to measure the voltage, the wire with clamp pin is connected to the two terminals. To measure the current, the current sensor is installed in the wire connected to the welder terminal. In addition, the computer 110 is also connected to the controller 140 to obtain information related to welding (welding start, welding end, object type, welding line information, etc.).

A device for measuring voltage and current is installed at two terminals 132 and 133 of the welder 130 to measure welding current and voltage during welding, and detect the occurrence of welding defects in real time or immediately after welding by using the same. Although the welding current and the welding voltage signal have different signal characteristics, the change in the characteristic between the normal welding signal and the defect signal is simultaneously shown in both the voltage and current signals. Therefore, the analysis results using the hidden Markov model (hereinafter referred to as HMM) are shown. Produces almost the same result. Therefore, the analysis is performed using either signal.

HMM is a method mainly used for voice signal analysis. Each model is composed of various voice signals. When a random voice signal is input, the HMM calculates the probability that the signal corresponds to each model and determines the model with the highest probability. Do it. Therefore, in speech recognition, a signal measured by a person is measured by a microphone, and a change in the phoneme model is estimated using a model showing a maximum probability value using various phoneme models and speech recognition is performed using the same. The present invention applies the HMM method to a mechanical defect detection technique, and determines a defect of a machine through a probability calculation process using an HMM algorithm.

In order to detect weld defects, only one model of normal welding is modeled, and then the welding signal during welding of the welding object is transformed using HMM algorithm, and then the probability value for the normal welding signal model is continuously calculated. In other words, if the welding signal is normal during actual welding, the calculated probability is relatively large, and the calculated probability is very stable over time. If any abnormality occurs, the calculated probability for the normal welding model decreases rapidly. . Therefore, it is determined that the welding defect has occurred in this section because the probability that the welding defect has occurred is very high in the section where the calculated probability value falls large.

That is, the present embodiment observes the continuous probability change of the utility plate with respect to the normal model using only the normal welding interval data, and determines the section in which the probability value becomes smaller as the abnormality occurrence section of the welding defect.

Hereinafter, a modeling process of a normal signal and a process of detecting whether a defect occurs in actual welding will be described in detail with reference to FIGS. 3 and 4.

3 illustrates a modeling process of a normal welding signal. As shown in FIG. 3, in order to model the normal welding signal, first, a welding operation is performed by using the welding robot 200, and a welding current or welding voltage signal is measured at a terminal of the welding machine 130 over time. (S1). The weld line (weld line) is completed by visual inspection, non-destructive inspection or inspection of the cut through the weld line to determine whether the weld line defects, if there is a defect and pinpoint the defect location (S2). Using either the welding current or the welding voltage signal in the section corresponding to the normal welding for one of the welding current or the welding voltage signal, the feature is expanded as a time function through preprocessing (filtering, time weighting, etc.) (S3). ). The preprocessed time function uses the characteristic vector streams [average, standard deviation, energy, power spectral coefficient, filter bank, cepstrum coefficient, wavelet] using a transformation that can extract the characteristic contained in the signal. ) Coefficients, AR model coefficients, etc.] (S4). Algorithms for obtaining HMM parameters (initial state probability, state transition probability, symbol observation probability, state average vector, state covariance matrix, etc.) using characteristic vector sequences, such as forward / backward algorithms and rate forwards Scaled forward / backward algorithm, k-means clustering method, modified k-mean clustering method, expansion-maximize algorithm, EM-baum-welch algorithm (Baum-Welch algorithm) and the like] is applied (S5). The result of applying the selected algorithm is completed in the HMM model in the normal welding (S6).

 4 illustrates a process of obtaining a welding signal probability value of a welding target region. As shown in Figure 4, the welding voltage or current at the time of welding the welding target site is measured (S11). For example, in the case of obtaining data sampled at 10 KHz, data for time may be acquired as 0.2 seconds of a time interval corresponding to 2,000 sampled data. The data are preprocessed into a time function through filtering, time weighting, and the like (S12), and the preprocessed time function is converted into a characteristic vector sequence using a transform equation used when modeling a normal welding model (S13). The HMM algorithm used to model the normal welding model is applied to the transformed characteristic vector sequence (S14). The result of applying the algorithm is completed by the welding signal probability value of the welding target region with respect to the normal welding signal (S15). When the time interval of the normal signal is shorter than the time interval of the actual welding signal, the probability value of the entire interval of the actual welding signal may be obtained by repeatedly using the normal signal interval.

In general, HMM uses the algebraic function of the calculated probability as the final probability because the difference between the minimum and maximum values of the calculated probability is very large. If this operation is repeated until the end of welding can be calculated the probability value of the welding signal for the entire time interval of the actual welding.

5 is a graph showing the probability value of the voltage signal during the actual welding and the welding signal obtained by the above process. If there is no defect in the weld line, the probability of the weld signal is almost constant except for the initial and final sections. As shown in FIG. 5, when there is a welding defect, a large peak downward occurs as the probability drops rapidly. The portion where such a large peak value occurs can be determined as the defect occurrence position.

Although a large amount of noise is mixed in the welding signal, the peak fluctuation range is remarkable in the section where the probability of defect is high in the probability value converted through the HMM algorithm as described above.

Even in the case of welding determined as normal welding, some characteristic difference may occur due to various causes such as the state of the base metal, the pretreatment difference, and the unevenness of the welding wire. Therefore, one modeling section may be used in the modeling process of obtaining a normal welding model, but the normal welding sections may be collected from various measurement data on the same welding line obtained by taking multiple sections from one welding line data or by repeated welding measurement. At the same time, the normal model can be obtained.

When the present invention is applied at the actual welding site, the determination of defects can be robustly improved. For example, in the early stage of operation after obtaining a suitable normal model, it may occur that the probability graph of the actual welding signal shows a smaller level of variation than the probability value of the welding defect level over the entire welding line or in some sections. Typically, this is not a welding defect, but the characteristics of the welding current and the welding voltage are usually changed due to various reasons such as the weld surface treatment condition or the material unevenness of the welding rod. Therefore, if it is determined that welding is normally performed in these areas by visual inspection or non-destructive inspection, normal modeling can be performed by adding data of this part to existing normal model data. After that, this section is judged to be normal, so that the normal model makes a more robust decision.

Although only the method of detecting weld defects has been described in the above-described embodiment, the present invention is not limited thereto, and may be applied to detecting a defect of a machine.

The present invention described above is not limited to the above-described embodiment and the accompanying drawings, and various substitutions, modifications, and changes are possible within the scope without departing from the spirit of the present invention. It will be apparent to those who have knowledge.

110: computer
120: A / D converter
130: welding machine
131: amplifier
132, 133: terminal
140: controller
200: welding robot
210: base
220, 230, 240: arm
211, 212, 221, 231, 241: joint
251: welding torch joint
260: torch cable
300: welding target

Claims (20)

A normal signal modeling step of measuring a diagnostic signal for diagnosing a machine state from a machine without defects and forming a steady signal model for a time domain by applying a hidden Markov model algorithm;
A diagnostic signal is measured in real time at a diagnosis target part of a machine to be diagnosed, and a probability value for the normal signal model is calculated by applying a hidden Markov model algorithm applied in the normal signal modeling step. Probability calculation step,
A defect determination step of determining that a defect has occurred in a section in which the probability value is not kept constant compared to the normal signal model;
The normal signal model update step of updating the probability value by adding the probability value to the normal signal model when the actual defect does not occur in the section where the defect is determined in the defect determination step.
Machine condition diagnostic method comprising a. delete The welding signal from the welding line of the object to be welded is measured for a time domain, and the welding signal of the welding object of the welding object is checked for defects, and the welding signal is extracted as the normal signal only for the normal welding section without defects. The normal welding signal modeling step of forming a normal welding signal model for the time domain of the entire interval by applying a hidden Markov model algorithm,
The normal welding is performed by measuring a welding signal of a welding target part in real time while welding a welding object, and applying a hidden Markov model algorithm used to form the normal welding signal model of the welding signal of the welding target part measured in real time. A probability value calculating step of calculating probability values for the signal model;
A defect determination step of determining that a defect has occurred in a section in which the probability value is not kept constant compared to the normal welding signal model
Welding defect diagnosis method comprising a. The method of claim 3,
And a welding signal is a voltage signal. The method of claim 3,
And welding signal is a current signal. The method of claim 3,
In the normal welding signal modeling step, the welding defect diagnosis method for determining whether a defect has occurred in the weld line through at least one of visual inspection, non-destructive inspection and cut surface inspection. The method of claim 3,
In the normal welding signal modeling step, the welding defect diagnosis method further comprises a normal welding signal preprocessing step of preprocessing the data of the section determined to be normal with a time function. The method of claim 7, wherein
The normal welding signal preprocessing step is a welding defect diagnosis method performed by filtering or time weighting. The method of claim 7, wherein
And a normal welding signal converting step of converting the data into characteristic vector strings using a conversion equation after the normal welding signal preprocessing step. 10. The method of claim 9,
The characteristic vector sequence is any one of, or a combination of, average, standard deviation, energy, power spectral coefficient, filter bank, capstrum coefficient, wavelet coefficient, and AR model coefficient. The method of claim 3,
The probability value for the normal welding signal model and the normal welding signal model of the welding signal of the welding target portion is obtained from any one of an initial state probability, a state transition probability, a symbol observation probability, a state average vector, and a state covariance matrix.
The hidden Markov model algorithm is any one of a forward algorithm, a backward algorithm, a rate forward algorithm, a rate backward algorithm, a K mean clustering technique, a modified K mean clustering technique, an EM algorithm, and a Baum-welch algorithm.
How to diagnose weld defects. The method of claim 3,
In the probability value calculation step, the welding defect diagnosis method further comprises a welding signal pre-treatment step for pre-processing the welding signal measured in real time at the welding target site by a time function. The method of claim 12,
The welding signal preprocessing step of the welding target region is a welding defect diagnosis method performed by filtering or time weighting. The method of claim 12,
Weld defect diagnosis method further comprises a welding signal conversion step of the welding target site to convert the time function pre-processed in the welding signal pre-processing step of the welding target site into a characteristic vector sequence using a conversion equation. The method of claim 14,
The characteristic vector sequence is any one of, or a combination of, average, standard deviation, energy, power spectral coefficient, filter bank, capstrum coefficient, wavelet coefficient, and AR model coefficient. The method of claim 3,
The normal welding section is weld defect diagnosis method obtained by combining a plurality of normal signal extraction section from the welding line. The method of claim 3,
The normal welding section is weld defect diagnosis method obtained by collecting the normal welding section obtained by repeatedly welding the welding line. The method of claim 3,
And a normal signal model updating step of updating the probability value by adding the probability value to the normal welding signal model when an actual defect does not occur in the section in which the defect is determined in the defect determination step.
How to diagnose weld defects. The welding voltage signal from the welding line of the object to be welded is measured in the time domain, and the welding line signal of the welding object is checked for defects, and the welding voltage signal is extracted as the normal welding voltage signal only for the normal welding section without defects. Preprocessing the normal welding voltage signal of the normal welding section with a time function for the entire section, converting the preprocessed time function of the normal welding voltage signal into characteristic vector sequences of the normal voltage signal using a conversion equation, and A normal welding voltage signal modeling step of forming a normal welding voltage signal model for a time domain by applying a hidden Markov model algorithm to characteristic vector strings of the normal welding voltage signal;
While welding the welding object, the welding voltage signal of the welding target site is measured in real time, and the welding voltage signal of the welding target site measured in real time is preprocessed as a time function, and the time function of the pre-processed welding voltage signal of the welding target site is determined. Convert the characteristic vector strings of the welding voltage signal of the welding target region to the characteristic vector strings using the conversion equation, and apply the hidden Markov model algorithm used in the normal welding voltage signal modeling step. A probability value calculating step of calculating probability values for the normal welding voltage signal model;
Defect determination step of determining that a defect occurred in a section in which the probability value of the welding voltage signal of the welding target region is not kept constant compared to the normal voltage signal model
Welding defect diagnosis method comprising a. Measure the welding current signal from the welding line of the object to be welded in the time domain, check whether the welding line of the welding object is defective, and extract the welding current signal as a normal welding current signal only for the normal welding section without defects. Preprocessing the normal welding current signal of the normal welding section with a time function for the entire section, converting the time function of the preprocessed normal welding current signal into characteristic vector sequences of the normal current signal using a conversion equation, and A normal welding current signal modeling step of forming a normal welding current signal model for a time domain by applying a hidden Markov model algorithm to characteristic vector streams of the normal welding current signal;
While welding a welding object, the welding current signal of the welding target part is measured in real time, and the welding current signal of the welding target part measured in real time is preprocessed as a time function, and the time function of the pre-processed welding current signal of the welding target part is determined. Convert the characteristic vector strings of the welding current signal of the welding target region to the characteristic vector strings using the conversion equation, and apply the hidden Markov model algorithm used in the normal welding current signal modeling step. A probability value calculating step of calculating probability values for the normal welding current signal model;
Defect determination step of determining that a defect occurred in a section in which the probability value of the welding current signal of the welding target region is not kept constant compared to the normal current signal model
Welding defect diagnosis method comprising a.

Images