WO2021077567A1 - 设备故障检测方法、设备故障检测装置及计算机存储介质 - Google Patents

设备故障检测方法、设备故障检测装置及计算机存储介质 Download PDF

Info

Publication number
WO2021077567A1
WO2021077567A1 PCT/CN2019/124118 CN2019124118W WO2021077567A1 WO 2021077567 A1 WO2021077567 A1 WO 2021077567A1 CN 2019124118 W CN2019124118 W CN 2019124118W WO 2021077567 A1 WO2021077567 A1 WO 2021077567A1
Authority
WO
WIPO (PCT)
Prior art keywords
fault
spectrogram
learning model
sound signal
target device
Prior art date
Application number
PCT/CN2019/124118
Other languages
English (en)
French (fr)
Inventor
徐小峰
梁伟彬
Original Assignee
广东美的白色家电技术创新中心有限公司
美的集团股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 广东美的白色家电技术创新中心有限公司, 美的集团股份有限公司 filed Critical 广东美的白色家电技术创新中心有限公司
Publication of WO2021077567A1 publication Critical patent/WO2021077567A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M99/00—Subject matter not provided for in other groups of this subclass
    • G01M99/005—Testing of complete machines, e.g. washing-machines or mobile phones
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01H—MEASUREMENT OF MECHANICAL VIBRATIONS OR ULTRASONIC, SONIC OR INFRASONIC WAVES
    • G01H9/00—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves by using radiation-sensitive means, e.g. optical means
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M99/00—Subject matter not provided for in other groups of this subclass

Definitions

  • This application relates to the field of fault detection, and in particular to a device fault detection method, device fault detection device and computer storage medium.
  • the current methods of detecting equipment failures are very backward, such as the use of human listening methods to determine whether the equipment is malfunctioning at the time.
  • Another example is the quality sampling before the equipment leaves the factory. It is necessary to establish an independent soundproof room, and it is necessary to test a large number of equipment, which results in low detection efficiency and high implementation cost.
  • this application provides an equipment failure detection method, an equipment failure detection device and a computer storage medium.
  • a non-contact vibration signal measurement solution is adopted without the need for an external soundproof room to efficiently detect equipment failure problems.
  • time and labor costs are saved, and the efficiency of quality inspection is improved.
  • a technical solution adopted in this application is to provide a device failure detection method.
  • the method includes: emitting a measurement beam to a target device, and collecting a detection beam reflected by the measurement beam through the target device; extracting a vibration signal from the detection beam; The signal is input to the trained fault learning model to identify the fault type associated with the vibration signal, thereby determining the fault type of the target device.
  • the trained fault learning model is obtained after training based on vibration signal sample data and predetermined fault type labels.
  • inputting the vibration signal to the trained fault learning model to identify the type of fault associated with the vibration signal includes: converting the vibration signal into a sound signal; inputting the sound signal to the trained fault learning model to identify and The fault type associated with the sound signal is used to determine the fault type of the target device; among them, the trained fault learning model is obtained after training based on the sound signal sample data and the predetermined fault type label.
  • the sound signal is input to the trained fault learning model to identify the fault type associated with the sound signal, so as to determine the fault type of the target device, including: preprocessing the sound signal; inputting the preprocessed sound signal To the trained fault learning model to identify the fault type associated with the sound signal, so as to determine the fault type of the target device.
  • the pre-processing of the sound signal includes: pre-emphasizing the sound signal to compensate for the high-frequency components in the sound signal; using a preset window function to perform frame and window processing on the sound signal to obtain the pre-processing After the sound signal.
  • the vibration signal is input to the trained fault learning model to identify the type of fault associated with the vibration signal, including: converting the vibration signal into a sound signal; preprocessing the sound signal; converting the preprocessed sound signal Is a spectrogram; the spectrogram is input to the trained fault learning model to identify the fault type associated with the spectrogram, so as to determine the fault type of the target device; among them, the trained fault learning model is based on the sound Spectrogram sample data and pre-determined fault type labels are obtained after training.
  • inputting the spectrogram into the trained fault learning model to identify the fault type associated with the spectrogram includes: extracting the time information, frequency information and energy information of the spectrogram to obtain the feature information to be processed; Input the feature information to be processed into the trained fault learning model to identify the fault type associated with the spectrogram.
  • inputting the spectrogram to the trained fault learning model to identify the fault type associated with the spectrogram includes: inputting the spectrogram to the trained fault learning model; using the fault learning model to compare the spectrogram Perform region segmentation and assign corresponding weights to different regions to obtain the segmented spectrogram to be processed; use the fault learning model to weight the segmented spectrogram to be processed to obtain the weighted similarity comparison result; When the weighted similarity comparison result is greater than the set threshold, it is determined that the fault type associated with the sound signal is the fault type corresponding to the spectrogram.
  • the target device before transmitting the measuring beam to the target device and collecting the detection beam reflected by the target device, it also includes: dividing the laser beam into the measuring beam and the reference beam; extracting the vibration signal from the detection beam, including: combining the detection beam and The reference beam interferes to detect the vibration signal.
  • the target device includes a motor, and the target device is vibrated by the operation of the motor.
  • the target device before emitting the measurement beam to the target device and collecting the detection beam reflected by the measurement beam through the target device, it further includes: connecting the target device through an external driving device to drive the target device to vibrate.
  • an equipment failure detection device which includes: a laser transmitter for emitting a measurement beam to the target device; a laser receiver for collecting the measurement beam through the target device Reflected detection beam; processor, connected to the laser receiver, used to extract the vibration signal from the detection beam, and input the vibration signal to the trained fault learning model to identify the type of fault associated with the vibration signal to determine the target
  • the fault type of the equipment among them, the trained fault learning model is obtained after training based on the vibration signal sample data and the predetermined fault type label.
  • the processor is also used to convert the vibration signal into a sound signal, and input the sound signal to the trained fault learning model to identify the fault type associated with the sound signal, thereby determining the fault type of the target device;
  • the trained fault learning model is obtained after training based on sound signal sample data and predetermined fault type labels.
  • the processor is also used to preprocess the sound signal, and input the preprocessed sound signal to the trained fault learning model to identify the fault type associated with the sound signal, so as to determine the fault type of the target device.
  • the processor is also used to convert the vibration signal into a sound signal, preprocess the sound signal, and convert the preprocessed sound signal into a spectrogram, and input the spectrogram into the trained fault learning model. Identify the fault types associated with the spectrogram; among them, the trained fault learning model is obtained after training based on the spectrogram sample data and predetermined fault type labels.
  • the processor is also used to extract the time information, frequency information and energy information of the spectrogram to obtain the feature information to be processed; input the feature information to be processed into the trained fault learning model to identify the correlation with the spectrogram The type of failure.
  • the processor is also used to input the spectrogram into the trained fault learning model; use the fault learning model to partition the spectrogram, and assign corresponding weights to different regions to obtain the segmented spectrogram to be processed ; Use the fault learning model to perform weighting processing on the segmented spectrogram to be processed to obtain the weighted similarity comparison result; in response to the weighted similarity comparison result being greater than the set threshold, determine the type of fault associated with the sound signal It is the fault type corresponding to the spectrogram.
  • the processor is connected to the laser transmitter, and is used to divide the laser beam emitted by the laser transmitter into a measurement beam and a reference beam, and interfere with the detection beam and the reference beam to detect and obtain a vibration signal.
  • Another technical solution adopted by this application is to provide a computer storage medium, which is used to store program data.
  • the program data is executed by a processor, it is used to implement any of the methods provided in the above-mentioned solutions.
  • a device failure detection method of the present application includes: emitting a measurement beam to a target device, and collecting the detection beam reflected by the measurement beam through the target device; The vibration signal is extracted from the light beam; the vibration signal is input to the trained fault learning model to identify the type of fault associated with the vibration signal, so as to determine the fault type of the target device.
  • FIG. 1 is a schematic flowchart of a first embodiment of a device fault detection method provided by the present application
  • FIG. 2 is a schematic structural diagram of an embodiment of an equipment failure detection device provided by the present application.
  • FIG. 3 is a schematic structural diagram of the application of the device failure detection method provided by the present application.
  • FIG. 4 is a schematic flowchart of a second embodiment of a device failure detection method provided by the present application.
  • FIG. 5 is a schematic flowchart of a third embodiment of a device failure detection method provided by the present application.
  • FIG. 6 is a schematic flowchart of a fourth embodiment of a device failure detection method provided by the present application.
  • FIG. 7 is a schematic flowchart of a fifth embodiment of a device fault detection method provided by the present application.
  • FIG. 8 is a schematic diagram of weighting processing in the device fault detection method provided by this application.
  • FIG. 9 is a schematic structural diagram of an embodiment of an equipment failure detection apparatus provided by this application.
  • FIG. 10 is a schematic structural diagram of another embodiment of the device fault detection device provided by the present application.
  • FIG. 11 is a schematic structural diagram of an embodiment of a computer storage medium provided by the present application.
  • Fig. 1 is a schematic flowchart of a first embodiment of a device failure detection method provided by the present application, and the method includes:
  • Step 11 Transmit the measurement beam to the target device, and collect the detection beam reflected by the measurement beam through the target device.
  • the laser transmitter in the equipment failure detection device emits the measurement beam to the target device, and a laser reflector is installed on the target device to reflect the detection beam after receiving the measurement beam.
  • the detection beam is provided by the laser receiver. To receive.
  • the target device as an air conditioner as an example
  • the outdoor unit of the air conditioner is assembled, the quality inspection is performed, the outdoor unit of the air conditioner is powered on to make the outdoor unit work, and the laser reflector is installed at the predetermined position of the outdoor unit. This position is based on the user Different location changes can be made on demand.
  • the equipment failure detection device emits a measuring beam to the laser reflector, and is reflected by the laser reflector to generate a detection beam, which is received by the laser receiver of the equipment failure detection device.
  • the outdoor unit of the air conditioner is already in working condition.
  • the laser transmitter in the device failure detection device emits a measurement beam to the target device, the measurement beam is reflected by the surface of the target device to generate a detection beam, and the detection beam is received by the laser receiver.
  • the device failure detection device of FIG. 2 is used for failure detection.
  • the device failure detection device 20 includes a laser transmitter 21, a laser receiver 22, a beam splitter 23, a reference beam reflector 24, and a measurement beam reflector 25. .
  • the laser transmitter 21 emits a HeNe laser, which is divided into a reference beam and a measurement beam by a beam splitter 23.
  • the reference beam is reflected by the reference beam reflector 24 and then enters the laser receiver 22 by the beam splitter 23.
  • the measurement beam passes through the measurement beam.
  • the reflector 25 reflects the detection beam, and the detection beam passes through the beam splitter 23 into the laser receiver 22. Among them, the measuring beam reflector 25 is arranged on the object to be measured. After the laser receiver receives the detection beam, proceed to step 12.
  • the beam splitter of the equipment fault detection device may use the wave front method, the amplitude method, and the polarization method to perform laser beam splitting.
  • the wavefront splitting method is to divide the wavefront of a point light source into two parts, and make them pass through two optical instrument groups respectively, and overlap after reflection, refraction or diffraction, forming interference in a certain area. Since any part of the wavefront can be regarded as a new light source, and each part of the same wavefront has the same phase, these separated parts of the wavefront can be used as light sources with the same initial phase, regardless of the phase change of the point light source How fast, the initial phase difference of these light sources is constant.
  • the sub-amplitude method is a method in which when a beam of light is projected on the interface of two transparent media, part of the light energy is reflected and the other part is refracted.
  • the polarization splitting method uses a polarization beam splitter, which is made up of a pair of glass prisms glued together, and magnesium fluoride and zinc sulfide film layers are alternately evaporated on the glued surface of one of the prisms.
  • the incident light enters the dielectric layer at Brewster's angle, and obtains high polarization S component reflected light and P component transmitted light through multiple transmission and reflection.
  • the polarizing beam splitter can be composed of polarizing prisms with orthogonal crystal axes, such as Wollaston prisms.
  • the reference beam reflector or the measurement beam reflector in the device failure detection device may be a flat reflector, a corner cube reflector, a right-angle prism reflector, or a cat's eye reflector.
  • Step 12 Extract the vibration signal from the detection beam.
  • the device failure detection device includes an optical path unit, a demodulation unit, and a signal generation unit.
  • the signal generating unit includes a signal generator.
  • the output signal of the signal generator is divided into two channels.
  • the first channel outputs two orthogonal signals sin(Csin( ⁇ t)) and cos(Csin( ⁇ t)) after the signal processor.
  • the two channels output phase shift drive signal sin( ⁇ t) after the DA converter; C is the sine coefficient, and ⁇ is the phase shift frequency caused by the carrier wave modulated by the electro-optic modulator in the optical circuit unit.
  • the optical path unit includes a laser and an electro-optic modulator.
  • the light output by the laser is divided into a measurement beam and a reference beam by the beam splitter.
  • the reference beam is input to the signal input end of the electro-optic modulator, the measurement beam is input to the first beam splitting prism, and the phase output of the signal generation unit is
  • the shift drive signal sin( ⁇ t) is input to the electrical signal drive end of the electro-optic modulator, and the phase-shifted reference beam output by the electro-optic modulator under the action of the phase shift drive signal sin( ⁇ t) is input to the second dichroic prism; it passes through the first dichroic prism
  • the return measurement beam formed by the measurement beam reflected on the target surface to be measured enters the first beam splitter again, and the return measurement beam is reflected to the second beam splitter.
  • the return measurement beam and the phase-shifted reference beam are mixed and interfered in the second beam splitter.
  • the interference light is formed, and the interference signal is output by the processor after the interference light is input into the photodetector;
  • ⁇ is the phase shift frequency caused by the carrier wave modulated by the electro-optic modulator.
  • the demodulation unit is a hardware module implemented by digital logic in FPGA, including four multipliers, two low-pass filters, two differentiators, one subtractor and one integrator.
  • the input of the demodulation unit is connected to the signal The two orthogonal signals sin(Csin( ⁇ t)) and cos(Csin( ⁇ t)) of the generating unit and the interference signal from the optical circuit unit.
  • the interference signal is divided into two paths after being passed through the AD converter, one of which is connected to cos(Csin( ⁇ t)) is sent to the first multiplier I for multiplication together, and the other is sent to the second multiplier together with sin(Csin( ⁇ t)) for multiplication; the product output by the first multiplier is passed through a low-pass filter Divided into two paths, one path is sent to the third multiplier after a differentiator d/dt, the other is directly sent to the fourth multiplier, and the product output by the second multiplier is divided into two after passing through another low-pass filter One of them is sent to the fourth multiplier after another differentiator d/dt, and the other is sent directly to the third multiplier; the product of the output of the third multiplier and the product of the output of the fourth multiplier are sent to the subtractor, The difference output from the subtractor is sent to the integrator.
  • the quadrature phase lock + DCM demodulation method includes the following steps:
  • the vibration signal such as the amplitude, frequency, and phase of the vibration signal, is extracted.
  • the vibration speed of the device under test is obtained by calculating the reference beam and the detection beam, and then the vibration frequency of the device under test is calculated.
  • Step 13 Input the vibration signal to the trained fault learning model to identify the fault type associated with the vibration signal, thereby determining the fault type of the target device.
  • the trained fault learning model is obtained after training based on vibration signal sample data and predetermined fault type labels.
  • the vibration signal is input to the trained fault learning model, and feature extraction of the vibration signal is performed.
  • Time-frequency analysis processing may be used to extract the characteristic parameters of the signal processing.
  • the time-frequency analysis method its idea is to design the joint function of time and frequency, and use it to simultaneously describe the energy density or intensity of the signal at different times and frequencies. Using time-frequency distribution to analyze the signal, the instantaneous frequency and its amplitude at each moment can be given, and time-frequency filtering and time-varying signal research can be carried out.
  • Time-frequency analysis methods include time-domain waveform analysis method, probability density analysis method, autocorrelation analysis method, trend analysis method, amplitude spectrum analysis method, power spectrum analysis method, cepstrum analysis method, envelope spectrum analysis method, refinement Spectral analysis method, short-time Fourier analysis method, wavelet analysis method.
  • the above are just examples, and the acquisition of the vibration characteristic parameters is not limited to the above methods, and other means can also be used.
  • the extracted feature information is identified to identify the type of failure associated with the vibration signal, thereby determining the type of failure of the target device.
  • the above-mentioned fault learning model may be constructed by means of machine learning.
  • the vibration signal samples are input to the learning model to be trained to learn to be trained
  • the model is trained to form a fault learning model.
  • the corresponding mark may be the fault type corresponding to the vibration signal. It is also possible to perform feature extraction of the vibration signal, and set corresponding marks to these feature information, and use these feature information and marks as training materials to construct a fault learning model.
  • a corresponding mark is output, and this mark is the fault type corresponding to the vibration signal.
  • Gaussian mixture model hidden Markov model, K-nearest neighbor, neural network, support vector machine, etc. can be used for model training to train the fault learning model. After the training is completed, the model can be used for unknown vibration The signal is recognized.
  • the vibration signal is input to a trained fault learning model, preprocessing is performed in the fault learning model, the preprocessed vibration signal is identified, and the fault type associated with the preprocessed vibration signal is identified, thereby Feedback user results.
  • the target device includes a motor, and the motor generates vibration during operation. If the target device is faulty, the vibration signal generated by the motor vibration is also different. Based on big data, first collect the vibration signals of many faults, so that the vibration signals can be used for deep learning to train the deep learning model. After the deep learning model is established, it is used to identify the subsequent vibration signals to determine the type of fault .
  • the target device cannot generate vibration, and an external driving device needs to be connected.
  • the external driving device is used to drive the target device to generate vibration.
  • the target device may be some welded components. 3, taking the application in welding technology as an example, connect the welded sheet metal 31 to the external drive device 32, so that the external drive device 32 vibrates to drive the sheet metal 31 to vibrate, where P is the solder joint, and equipment failure detection
  • the device 33 collects vibration signals.
  • Common defects during welding of sheet metal 31 include undercuts, weld bumps, depressions, and welding deformations, and sometimes surface pores and surface cracks. The root part of single-sided welding is not penetrated, there are pores, slag inclusions, cracks, and lack of fusion.
  • a device failure detection method of the present application includes: emitting a measuring beam to a target device, and collecting a detection beam reflected by the measuring beam through the target device; extracting a vibration signal from the detection beam; and combining the vibration signal Input to the trained fault learning model to identify the fault type associated with the vibration signal, thereby determining the fault type of the target device.
  • FIG. 4 is a schematic flowchart of a second embodiment of a device failure detection method provided by the present application, and the method includes:
  • Step 41 Transmit the measurement beam to the target device, and collect the detection beam reflected by the measurement beam through the target device.
  • Step 42 Extract the vibration signal from the detection beam.
  • Steps 41-42 have the same or similar technical solutions as the above-mentioned embodiment, and will not be repeated here.
  • Step 43 Convert the vibration signal into a sound signal.
  • Step 44 preprocessing the sound signal.
  • the sound signal is pre-emphasized, the purpose of which is to emphasize the high-frequency part of the speech and increase the high-frequency resolution of the speech, so as to compensate for the high-frequency component in the sound signal.
  • Digital filters are usually used to achieve pre-emphasis.
  • Voice signal is a signal that changes with time, and is mainly divided into two categories: voiced and unvoiced.
  • the pitch period of voiced sounds, the amplitude of unvoiced and voiced signals, and vocal tract parameters all change slowly with time. Due to the inertial movement of the vocal organs, it can be considered that the voice signal is approximately unchanged in a short period of time (usually 10-30ms), that is, the voice signal has short-term stability. In this way, the speech signal can be divided into short segments (called analysis frames) for processing.
  • the framing of the speech signal is achieved by using a movable finite-length window for weighting.
  • the number of frames per second is 33-100 frames, depending on the actual situation.
  • Framing can adopt the method of continuous segmentation or the method of overlapping segmentation. This is to make the transition between frames smoothly and maintain its continuity.
  • the method of overlapping segmentation is adopted.
  • the overlapping part of the previous frame and the next frame is called the frame shift, and the ratio of the frame shift to the frame length is generally 0 to 1/2.
  • rectangular windows or Hamming windows can be used for windowing.
  • Step 45 Input the preprocessed sound signal to the trained fault learning model to identify the fault type associated with the sound signal, so as to determine the fault type of the target device.
  • the trained fault learning model is obtained after training based on sound signal sample data and predetermined fault type labels.
  • the preprocessed sound signal is input to the trained fault learning model, and the sound signal is feature extracted.
  • the extracted feature information may include: time domain features, frequency domain features, cepstrum features, time-frequency features, and so on.
  • the time domain features mainly include short-term energy, short-term average amplitude, and short-term zero-crossing rate.
  • the short-term average amplitude reflects the energy of a frame of speech signal.
  • the short-time zero-crossing rate indicates the number of times the waveform crosses the horizontal axis in a frame of audio signal.
  • the frequency domain feature is the feature obtained by transforming the signal into the frequency domain by Fourier transform, and then calculating it in the frequency domain.
  • Mainly include spectrum centroid, bandwidth, and spectrum roll-off coefficient.
  • the spectral centroid reflects the average value of the energy of each frame.
  • the bandwidth reflects the fluctuation degree of the energy of the sampling point near the mean value.
  • the spectrum roll-off describes the degree of inclination of the spectrum.
  • the cepstrum is the inverse Fourier transform of the logarithm of the power spectrum or energy spectrum.
  • the preprocessed sound signal is input to a deep learning network, and the deep learning network is preprocessed, the preprocessed sound is recognized, and the fault type associated with the preprocessed sound signal is identified, So as to feedback the user's results.
  • FIG. 5 is a schematic flowchart of a third embodiment of a device failure detection method provided by the present application, and the method includes:
  • Step 51 Transmit the measurement beam to the target device, and collect the detection beam reflected by the measurement beam through the target device.
  • Step 52 Extract the vibration signal from the detection beam.
  • Step 53 Convert the vibration signal into a sound signal.
  • Step 54 Preprocessing the sound signal.
  • Steps 51-54 have the same or similar technical solutions as the foregoing embodiment, and will not be repeated here.
  • Step 55 Convert the preprocessed sound signal into a spectrogram.
  • the spectrogram is a distribution diagram of time and frequency.
  • the spectrogram not only reflects the frequency domain and time domain characteristics of the acoustic signal, but also shows the relationship between the time domain and the frequency domain. From the spectrogram, it can be observed that some characteristics of the frequency domain change with the occurrence of the sound signal. Changes in the situation; you can also observe the changes in energy with the sound process. Therefore, the information of the sound signal carried by the spectrogram is far greater than the information carried by the pure time domain signal and the pure frequency domain signal.
  • the spectrogram combines the characteristics of the spectrogram and the time-domain waveform, and clearly shows the change of the sound spectrum over time, or the spectrogram is a dynamic spectrum.
  • the framing window length is 512 points
  • the window function is Hamming window
  • the frame overlap is 0.75 times the window length
  • specgram is called through the Matlab (Matrix Laboratory) third-party tool VoiceBox (voice processing tool box) The function draws the spectrogram.
  • Step 56 Input the spectrogram into the trained fault learning model to identify the type of fault associated with the spectrogram.
  • the trained fault learning model is obtained after training based on the spectrogram sample data and predetermined fault type labels.
  • the converted sound signal is pre-emphasized, and a specific high-pass filter is used to compensate the high-frequency part of the collected sound signal; then, endpoint detection is performed on the pre-emphasized sound signal to determine the sound The starting point of the effective signal in the signal.
  • this embodiment can determine the effective sound information in the sound signal (That is, the starting point of the effective signal), and then perform abnormal sound matching or detection on the effective signal; then, the characteristic parameters of the effective signal with the determined starting point within a certain range are framed and windowed to meet the statistical requirements. The characteristics are stable.
  • the preprocessed sound signal is converted into a spectrogram.
  • the spectrogram is a two-dimensional graph with time on the horizontal axis and frequency on the vertical axis.
  • the point corresponding to the coordinates (x, y) represents the sound intensity at time x and frequency y, and the sound intensity is represented by different colors.
  • the distribution and change of the sound intensity in the entire time-frequency range can be obtained; this is not present in the waveform diagram.
  • the sound signal is divided into short frames, and adjacent frames will overlap to a certain extent. Then the short-time Fourier transform is performed on each frame to obtain the corresponding spectrum information. Since the spectrogram is composed of three dimensional information of frequency, time, and sound intensity, it is necessary to calculate the value of the sound intensity. Finally, the spectrum information is connected into a complete spectrogram.
  • the effective signal segment is windowed and divided into several frames; short-time Fourier transform is performed on each frame to obtain the frequency spectrum information of the frame, and the frequency spectrum information is used to indicate the frequency and sound of the frame.
  • the spectrogram is composed of several points. The coordinates (x, y) of any point are used to indicate that the point is at time x, y The corresponding sound intensity on the frequency.
  • the short-time Fourier transform is performed on the time sequence signal of the sound, and the length of the Fourier transform is 2N points.
  • the signal of each frame can obtain the frequency spectrum of length N, and the sound pressure of each point
  • a spectrogram matrix is generated. Extract the identification feature matrix to be tested, which is used to characterize the sound intensity distribution of the spectrogram, from the spectrogram matrix.
  • the identification feature matrix to be tested is input into the trained fault learning model, and the fault type of the identification feature matrix to be tested is output through the internal identification of the model.
  • FIG. 6 is a schematic flowchart of a fourth embodiment of a device failure detection method provided by the present application, and the method includes:
  • Step 61 Transmit the measurement beam to the target device, and collect the detection beam reflected by the measurement beam through the target device.
  • Step 62 Extract the vibration signal from the detection beam.
  • Step 63 Convert the vibration signal into a sound signal.
  • Step 64 Preprocessing the sound signal.
  • Step 65 Convert the preprocessed sound signal into a spectrogram.
  • Steps 61-65 have the same or similar technical solutions as the foregoing embodiment, and will not be repeated here.
  • Step 66 Extract the time information, frequency information and energy information of the spectrogram to obtain the feature information to be processed.
  • the abscissa of the spectrogram represents time
  • the ordinate represents frequency
  • the value of the coordinate point of time and frequency represents the energy of the sound signal.
  • the size of the energy value is expressed by color, and the darker color indicates the stronger the energy of the point.
  • the Gabor transform method is used for feature extraction, and the features of the spectrogram can be extracted from multiple directions and multiple frequency scales.
  • a projection method is used for feature extraction.
  • the projection method is to project a spectrogram in four directions (0 degrees, 45 degrees, 90 degrees, 135 degrees), and project them in the direction of 0 degrees (abscissa), that is, the same time component but different frequency components Pixel brightness values are all accumulated; projected to the 90 degree direction (the ordinate), that is, the pixel brightness values of the same frequency component and different time components are accumulated; projected to the direction of 45 degrees and 135 degrees, and the pixel brightness values perpendicular to the direction are projected In order to ensure the consistency of the amount of data, after each accumulation, it must be divided by the number of pixels.
  • Step 67 Input the feature information to be processed into the trained fault learning model to identify the fault type associated with the spectrogram.
  • the trained fault learning model is obtained after training based on the spectrogram sample data and predetermined fault type labels.
  • the fault learning model is modeled by supervised learning in machine learning.
  • the content of the labeling can be the fault type.
  • the fault types are classified. Taking washing machines as an example, the screws are not fixed into one category, and the motor faults are classified into one category. When it is determined that the characteristic information of the current spectrogram is the category of unfixed screws, the current spectrogram is continuously compared in the category of unsecured screws.
  • the fault type associated with the sound signal is the fault type corresponding to the standard characteristic information.
  • the target device as a refrigerator as an example
  • the vibration signal obtained by the fault detection device is converted into a sound signal
  • the sound signal is converted into a spectrogram.
  • the spectrogram is extracted in the deep learning network, and the characteristic information is stored in advance.
  • the standard feature information is compared for similarity, and the similarity of the comparison is 85%, and the set threshold is 80%. If the similarity is greater than the set threshold, the fault type associated with the sound signal is determined It is the fault type corresponding to the standard feature information.
  • FIG. 7 is a schematic flowchart of a fifth embodiment of a voice interaction method for an electronic device provided by the present application. The method includes:
  • Step 71 Transmit the measurement beam to the target device, and collect the detection beam reflected by the measurement beam through the target device.
  • Step 72 Extract the vibration signal from the detection beam.
  • Step 73 Convert the vibration signal into a sound signal.
  • Step 74 Preprocessing the sound signal.
  • Step 75 Convert the preprocessed sound signal into a spectrogram.
  • Step 76 Input the spectrogram into the trained fault learning model.
  • Steps 71-76 have the same or similar technical solutions as the foregoing embodiment, and will not be repeated here.
  • Step 77 Use the fault learning model to partition the spectrogram into regions, and assign corresponding weights to different regions, so as to obtain the segmented spectrogram to be processed.
  • Step 78 Use the fault learning model to perform weighting processing on the segmented spectrogram to be processed to obtain a weighted similarity comparison result.
  • Step 79 In response to the weighted similarity comparison result being greater than the set threshold, it is determined that the fault type associated with the sound signal is the fault type corresponding to the spectrogram.
  • Fig. 8 The left side of Fig. 8 is the block spectrogram to be processed, and the right side is the standard block spectrogram.
  • A*a represents the similarity comparison result between A and a.
  • B*b represents the comparison result of similarity between B and b.
  • C*c represents the comparison result of similarity between C and c.
  • D*d represents the comparison result of the similarity between D and d.
  • the threshold is set to 70%
  • FIG. 9 is a schematic structural diagram of an embodiment of an equipment failure detection device provided by this application.
  • the equipment failure detection device 90 includes a laser transmitter 91, a laser receiver 92, and a processor 93.
  • the laser transmitter 91 is used to emit a measuring beam to the target device.
  • the laser receiver 92 is used to collect the detection beam reflected by the measurement beam through the target device.
  • the processor 93 connected to the laser receiver, is used to extract the vibration signal from the detection beam and input the vibration signal to the trained fault learning model to identify the type of fault associated with the vibration signal, thereby determining the fault type of the target device ; Among them, the trained fault learning model is obtained after training based on vibration signal sample data and predetermined fault type labels.
  • the processor 93 is also used to convert the vibration signal into a sound signal, and input the sound signal to the trained fault learning model to identify the fault type associated with the sound signal, thereby determining the fault type of the target device;
  • the fault learning model is obtained after training based on sound signal sample data and predetermined fault type labels.
  • the processor 93 is also used to preprocess the sound signal, and input the preprocessed sound signal to the trained fault learning model to identify the fault type associated with the sound signal, so as to determine the fault type of the target device.
  • the processor 93 is also used to convert the vibration signal into a sound signal, preprocess the sound signal, and convert the preprocessed sound signal into a spectrogram, and input the spectrogram into the trained fault learning model to identify The fault type associated with the spectrogram; among them, the trained fault learning model is obtained after training based on the sample data of the spectrogram and the predetermined fault type label.
  • the processor 93 is also used to extract the time information, frequency information and energy information of the spectrogram to obtain the feature information to be processed; input the feature information to be processed into the trained fault learning model to identify the spectrogram associated with the Fault type.
  • the processor 93 is also used to input the spectrogram into the trained fault learning model; use the fault learning model to partition the spectrogram into regions, and assign corresponding weights to different regions to obtain the segmented spectrogram to be processed; Use the fault learning model to perform weighting processing on the segmented spectrogram to be processed to obtain the weighted similarity comparison result; in response to the weighted similarity comparison result being greater than the set threshold, the fault type associated with the sound signal is determined to be The fault type corresponding to the spectrogram.
  • the processor 93 is connected to the laser transmitter 91, and is used to divide the laser beam emitted by the laser transmitter 91 into a measurement beam and a reference beam, and interfere with the detection beam and the reference beam to detect and obtain a vibration signal.
  • the equipment failure detection device 100 includes a laser transmitter 101, a beam splitter 102, a beam splitter 103, a beam splitter 104, a Bragg box 105, a fixed reflector 106, and a photodetector 107.
  • FIG. 10 only illustrates the transmission and reception process of the laser beam, and the partial structure diagram of the subsequent processing is not shown.
  • 200 is the target device.
  • the laser transmitter 101 emits a laser beam to the beam splitter 102, and the beam splitter 102 divides the laser beam into a measurement beam and a reference beam; as shown in the figure, the direction of the measurement beam is toward the beam splitter 103, the reference beam It is directed to the fixed reflector 106, the reference beam is reflected by the fixed reflector 106, enters the Bragg box 105, passes through the Bragg box 105 and then is directed to the beam splitter 104, and the reference beam is directed to the photodetector 107 through the beam splitter 104; measuring beam After the beam splitter 103, it is directed to the target device 200, and the reflected detection beam is collected.
  • the detection beam is deflected downward by the beam splitter 103 and directed to the beam splitter 104; the detection beam is directed to the photodetector 107 by the beam splitter 104; At this time, both the reference beam and the detection beam are incident on the photodetector 107, and the vibration signal of the target device is calculated by the photodetector 107 using the Doppler frequency shift proportional to the vibration speed of the target device 200.
  • the vibration of the target device 200 will generate bright/dark fringes on the photodetector 107, which is an interferometry used.
  • a complete light/dark periodic fringe on the photodetector 107 exactly corresponds to the displacement of half the wavelength of the laser used. This corresponds to a displacement of 316 nm in the case of the helium-neon laser frequently used in the laser transmitter 101.
  • the optical path change per unit time is manifested as the Doppler frequency shift of the measuring beam.
  • the Doppler frequency shift is directly proportional to the vibration speed of the sample. Since the bright and dark fringes (and modulation frequency) produced by the movement of the object away from the interferometer are the same as those produced by the movement of the object toward the interferometer, this setting cannot determine the direction of the object's movement.
  • an acousto-optic modulator with a typical optical frequency shift of 40 MHz is placed in the reference beam (for comparison purposes, the laser frequency is 4.74 ⁇ 1014 Hz). When the target device 200 is in a static state, a typical interferometric modulation frequency of 40 MHz will be generated.
  • the modulation frequency will increase; when the target device 200 moves away from the interferometer, the frequency received by the detector is less than 40 MHz. Not only can accurately detect the optical path length, but also detect the direction of movement.
  • the equipment failure detection device 100 can also directly measure the displacement.
  • the maximum amplitude of harmonic vibration can be expressed as follows:
  • v 2 ⁇ *f*s; where v represents the speed of the target device; f represents the vibration frequency of the target device; s represents the vibration displacement of the target device. As the frequency increases, the vibration speed increases, and the vibration displacement decreases.
  • FIG. 11 is a schematic structural diagram of an embodiment of a computer storage medium provided by the present application.
  • the computer storage medium 110 is used to store program data 111.
  • the program data 111 is executed by a processor, it is used to implement the following method steps:
  • Transmit the measurement beam to the target device and collect the detection beam reflected by the measurement beam through the target device; extract the vibration signal from the detection beam; input the vibration signal into the trained fault learning model to identify the type of fault associated with the vibration signal, In order to determine the type of failure of the target device.
  • program data 111 when executed by the processor, it is also used to implement the method in any one of the foregoing embodiments.
  • the disclosed method and device may be implemented in other ways.
  • the device implementation described above is merely illustrative.
  • the division of the modules or units is only a logical function division.
  • there may be other divisions for example, multiple units or components may be Combined or can be integrated into another system, or some features can be ignored or not implemented.
  • the units described above as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of this embodiment.
  • the functional units in the various embodiments of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit.
  • the above-mentioned integrated unit can be implemented in the form of hardware or software functional unit.
  • the integrated unit in the other embodiments described above is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium.
  • the technical solution of the present application essentially or the part that contributes to the existing technology or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , Including several instructions to enable a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in the various embodiments of the present application.
  • the aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disks or optical disks and other media that can store program codes. .

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)

Abstract

一种设备故障检测方法、对应的设备故障检测装置及计算机存储介质，该方法包括：向目标设备发射测量光束，并采集测量光束经由目标设备反射的检测光束（11）；从检测光束中提取振动信号（12）；将振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，从而确定目标设备的故障类型（13）。通过上述方式，一方面采用非接触式振动信号测量方案，无需外置隔音房，高效检测设备故障问题，另一方面在大批量设备质检时，节省时间和人力成本，提高质检效率。

Description

设备故障检测方法、设备故障检测装置及计算机存储介质
本申请要求于2019年10月24日提交的申请号为2019110197201，发明名称为“设备故障检测方法、设备故障检测装置及计算机存储介质”的中国专利申请的优先权，其通过引用方式全部并入本申请。
【技术领域】
本申请涉及故障检测领域，特别是涉及一种设备故障检测方法、设备故障检测装置及计算机存储介质。
【背景技术】
设备在运行过程中会发出各种各样的声音，一部分是设备在正常运行状态下发出的声音，一部分是设备在故障等情况下发出的异常声音。当出现异常声音时，通常表示设备出现了故障，需要进行维护。
但是目前检测设备故障的方式很落后，如采用人为听音辨别的方式来判断当时设备是否故障。又例如在设备出厂前进行质量抽检，需要建立独立的隔音房，且需要对大量的设备进行检测，检测效率低，实施成本高。在一些情况下，出厂前需要对设备进行破坏性质量抽检，如对焊点的检测，需要对抽样的样本进行敲击破坏，以检查焊点处质量是否合格，此种方法同样检测效率低且实施成本高。
【发明内容】
为了解决上述问题，本申请提供一种设备故障检测方法、设备故障检测装置及计算机存储介质，一方面采用非接触式振动信号测量方案，无需外置隔音房，高效检测设备故障问题，另一方面在大批量设备质检时，节省时间和人力成本，提高质检效率。
本申请采用的一种技术方案是提供一种设备故障检测方法，该方法包括：向目标设备发射测量光束，并采集测量光束经由目标设备反射的检测光束；从检测光束中提取振动信号；将振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，从而确定目标设备的故障类型。
其中，已训练的故障学习模型，是基于振动信号样本数据以及预先确定的故障类型标签进行训练后得到的。
其中，将振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，包括：将振动信号转化为声音信号；将声音信号输入至已训练的故障学习模型，以识别与声音信号相关联的故障类型，从而确定目标设备的故障类型；其中，已训练的故障学习模型，是基于声音信号样本数据以及预先确定的故障类型标签进行训练后得到的。
其中，将声音信号输入至已训练的故障学习模型，以识别与声音信号相关联的故障类型，从而确定目标设备的故障类型，包括：对声音信 号进行预处理；将预处理之后的声音信号输入至已训练的故障学习模型，以识别与声音信号相关联的故障类型，从而确定目标设备的故障类型。
其中，对声音信号进行预处理，包括：对声音信号进行预加重，以对声音信号中的高频分量进行补偿；采用预先设置的窗函数对声音信号进行分帧加窗处理，以得到预处理之后的声音信号。
其中，将振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，包括：将振动信号转化为声音信号；对声音信号进行预处理；将预处理之后的声音信号转化为声谱图；将声谱图输入至已训练的故障学习模型，以识别与声谱图相关联的故障类型，从而确定目标设备的故障类型；其中，已训练的故障学习模型，是基于声谱图样本数据以及预先确定的故障类型标签进行训练后得到的。
其中，将声谱图输入至已训练的故障学习模型，以识别与声谱图相关联的故障类型，包括：提取声谱图的时间信息、频率信息和能量信息，以得到待处理特征信息；将待处理特征信息输入至已训练的故障学习模型，以识别与声谱图相关联的故障类型。
其中，将声谱图输入至已训练的故障学习模型，以识别与声谱图相关联的故障类型，包括：将声谱图输入至已训练的故障学习模型；利用故障学习模型对声谱图进行区域分块，并将不同区域分配对应权重，以得到待处理分块声谱图；利用故障学习模型对待处理分块声谱图进行加权处理，以得到加权后的相似度比对结果；响应于加权后的相似度比对结果大于设定阈值，确定声音信号相关联的故障类型为声谱图对应的故障类型。
其中，向目标设备发射测量光束，并采集测量光束经由目标设备反射的检测光束之前，还包括：将激光束分为测量光束和参考光束；从检测光束中提取振动信号，包括：将检测光束和参考光束进行干涉，以检测得到振动信号。
其中，目标设备包括电机，目标设备由电机工作而产生振动。
其中，向目标设备发射测量光束，并采集测量光束经由目标设备反射的检测光束之前，还包括：通过外部驱动设备连接目标设备，以驱动目标设备产生振动。
本申请采用的另一种技术方案是提供一种设备故障检测装置，该设备故障检测装置包括：激光发射器，用于向目标设备发射测量光束；激光接收器，用于采集测量光束经由目标设备反射的检测光束；处理器，连接激光接收器，用于从检测光束中提取振动信号，并将振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，从而确定目标设备的故障类型；其中，已训练的故障学习模型，是基于振动信号样本数据以及预先确定的故障类型标签进行训练后得到的。
其中，处理器还用于将振动信号转化为声音信号，并将声音信号输入至已训练的故障学习模型，以识别与声音信号相关联的故障类型，从 而确定目标设备的故障类型；其中，已训练的故障学习模型，是基于声音信号样本数据以及预先确定的故障类型标签进行训练后得到的。
其中，处理器还用于对声音信号进行预处理，并将预处理之后的声音信号输入至已训练的故障学习模型，以识别与声音信号相关联的故障类型，从而确定目标设备的故障类型。
其中，处理器还用于将振动信号转化为声音信号，对声音信号进行预处理，并将预处理之后的声音信号转化为声谱图，将声谱图输入至已训练的故障学习模型，以识别与声谱图相关联的故障类型；其中，已训练的故障学习模型，是基于声谱图样本数据以及预先确定的故障类型标签进行训练后得到的。
其中，处理器还用于提取声谱图的时间信息、频率信息和能量信息，以得到待处理特征信息；将待处理特征信息输入至已训练的故障学习模型，以识别与声谱图相关联的故障类型。
其中，处理器还用于将声谱图输入至已训练的故障学习模型；利用故障学习模型对声谱图进行区域分块，并将不同区域分配对应权重，以得到待处理分块声谱图；利用故障学习模型对待处理分块声谱图进行加权处理，以得到加权后的相似度比对结果；响应于加权后的相似度比对结果大于设定阈值，确定声音信号相关联的故障类型为声谱图对应的故障类型。
其中，处理器连接激光发射器，用于将激光发射器发射的激光束分为测量光束和参考光束，并将检测光束和参考光束进行干涉，以检测得到振动信号。
本申请采用的另一种技术方案是提供一种计算机存储介质，该计算机存储介质用于存储程序数据，程序数据在被处理器执行时，用于实现如上述方案中提供的任一方法。
本申请的有益效果是：区别于现有技术的情况，本申请的一种设备故障检测方法，该方法包括：向目标设备发射测量光束，并采集测量光束经由目标设备反射的检测光束；从检测光束中提取振动信号；将振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，从而确定目标设备的故障类型。通过上述方式，一方面采用非接触式振动信号测量方案，无需外置隔音房，高效检测设备故障问题，另一方面在大批量设备质检时，节省时间和人力成本，提高质检效率。
【附图说明】
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。其中：
图1是本申请提供的设备故障检测方法第一实施例的流程示意图；
图2是本申请提供的设备故障检测装置一实施例的结构示意图；
图3是本申请提供的设备故障检测方法的应用的结构示意图；
图4是本申请提供的设备故障检测方法第二实施例的流程示意图；
图5是本申请提供的设备故障检测方法第三实施例的流程示意图；
图6是本申请提供的设备故障检测方法第四实施例的流程示意图；
图7是本申请提供的设备故障检测方法第五实施例的流程示意图；
图8为本申请提供的设备故障检测方法中加权处理的示意图；
图9为本申请提供的设备故障检测装置一实施例的结构示意图；
图10是本申请提供的设备故障检测装置另一实施例的结构示意图；
图11是本申请提供的计算机存储介质一实施例的结构示意图。
【具体实施方式】
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述。可以理解的是，此处所描述的具体实施例仅用于解释本申请，而非对本申请的限定。另外还需要说明的是，为了便于描述，附图中仅示出了与本申请相关的部分而非全部结构。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
在本文中提及“实施例”意味着，结合实施例描述的特定特征、结构或特性可以包含在本申请的至少一个实施例中。在说明书中的各个位置出现该短语并不一定均是指相同的实施例，也不是与其它实施例互斥的独立的或备选的实施例。本领域技术人员显式地和隐式地理解的是，本文所描述的实施例可以与其它实施例相结合。
参阅图1，图1是本申请提供的设备故障检测方法第一实施例的流程示意图，该方法包括：
步骤11：向目标设备发射测量光束，并采集测量光束经由目标设备反射的检测光束。
在一些实施例中，由设备故障检测装置中的激光发射器向目标设备发射测量光束，在目标设备上安装有激光反射器，用于接收到测量光束后反射检测光束，检测光束由激光接收器进行接收。
以目标设备为空调为例，当空调的室外机在组装完成时，进行质量检测，给空调室外机上电，以使室外机工作，将激光反射器安装于室外机的预定位置，此位置根据用户需求，可进行不同的位置更改。设备故障检测装置发射测量光束到激光反射器，经激光反射器反射生成检测光束，由设备故障检测装置的激光接收器接收。此时空调室外机已处于工作状态。
在一些实施例中，设备故障检测装置中的激光发射器向目标设备发射测量光束，测量光束由目标设备表面反射，生成检测光束，检测光束由激光接收器进行接收。
在一些实施例中，采用图2的设备故障检测装置进行故障检测，设备故障检测装置20包括激光发射器21、激光接收器22、分束器23、参 考光束反射器24、测量光束反射器25。
激光发射器21发射氦氖激光，经由分束器23分为参考光束和测量光束，参考光束经由参考光束反射器24反射再通过分束器23射入激光接收器22中，测量光束经由测量光束反射器25反射得到检测光束，检测光束再通过分束器23射入激光接收器22中。其中，测量光束反射器25安置在被测物体上。激光接收器接收到检测光束后进行步骤12。
在一些实施例中，设备故障检测装置的分束器可以采用分波阵面法、分振幅法、分偏振法进行激光分束。
分波阵面法是将点光源的波阵面分割为两部分，使之分别通过两个光具组，经反射、折射或衍射后交迭起来，在一定区域形成干涉。由于波阵面上任一部分都可看作新光源，而且同一波阵面的各个部分有相同的位相，所以这些被分离出来的部分波阵面可作为初相位相同的光源，不论点光源的位相改变得如何快，这些光源的初相位差却是恒定的。
分振幅法是当一束光投射到两种透明媒质的分界面上，光能一部分反射，另一部分折射的方法。
分偏振法采用偏振分束器，它由一对玻璃棱镜相胶合而成，在其中一块棱镜的胶合面上交替蒸镀氟化镁和硫化锌膜层。入射光以布儒斯特角进入介质层，经多次透射和反射得到高偏振度的S分量反射光和P分量透射光。偏振分束器可以由晶轴正交的偏光棱镜组成，如渥拉斯顿棱镜。
在一些实施例中，设备故障检测装置中的参考光束反射器或测量光束反射器可以是平面反射器、角锥棱镜反射器、直角棱镜反射器、猫眼反射器。
步骤12：从检测光束中提取振动信号。
在一些实施例中，设备故障检测装置包括光路单元、解调单元、信号发生单元。
信号发生单元包括信号发生器，信号发生器的输出信号分为两路，第一路经信号处理器后输出两路正交信号sin(Csin(ωt))和cos(Csin(ωt))，第二路经DA转换器后输出相移驱动信号sin(ωt)；C为正弦系数，ω为光路单元中电光调制器调制的载波引起的移相频率。
光路单元包括激光器和电光调制器，激光器输出的光经分束器分为测量光束和参考光束，参考光束输入电光调制器的信号输入端，测量光束输入第一分光棱镜，信号发生单元输出的相移驱动信号sin(ωt)输入电光调制器的电信号驱动端，电光调制器在相移驱动信号sin(ωt)的作用下输出的相移参考光束输入到第二分光棱镜；通过第一分光棱镜的测量光束在待测目标表面反射后形成的返回测量光束再次进入第一分光棱镜，返回测量光束再反射到第二分光棱镜，返回测量光束与相移参考光束在第二分光棱镜中混合并干涉形成干涉光，干涉光输入光电探测器后经处理器输出干涉信号；ω为电光调制器调制的载波引起的移相频率。
解调单元为在FPGA中通过数字逻辑实现的硬件模块，包括四个乘法器、两个低通滤波器、两个微分器、一个减法器和一个积分器，解调单元的输入端连接来自信号发生单元的两路正交信号sin(Csin(ωt))及cos(Csin(ωt))和来自光路单元的干涉信号，干涉信号经AD转换器后分为两路，其中一路与cos(Csin(ωt))一并送入第一乘法器Ⅰ相乘，另一路与sin(Csin(ωt))一并送入第二乘法器相乘；第一乘法器输出的乘积经一低通滤波器后分为两路，其中一路经一微分器d/dt后送入第三乘法器，另一路直接送入第四乘法器，第二乘法器输出的乘积经另一低通滤波器后分为两路，其中一路经另一微分器d/dt后送入第四乘法器，另一路直接送入第三乘法器；第三乘法器输出的乘积和第四乘法器输出的乘积送入减法器，减法器输出的差送入积分器。
采用正交锁相+DCM解调方法，包括如下步骤：
(1)选择两路正交信号sin(Csin(ωt))及cos(Csin(ωt))和一路模拟干涉信号，模拟干涉信号经AD转换器后分为两路数字干涉信号作为解调输入参数。
(2)其中一路数字干涉信号与cos(Csin(ωt))一并送入第一乘法器相乘，另一路数字干涉信号与sin(Csin(ωt))一并送入第二乘法器相乘；
(3)将第一乘法器输出的乘积经低通滤波器后输出的信号分为两路，其中一路经微分后送入第三乘法器，另一路直接送入第四乘法器；将第二乘法器输出的乘积经低通滤波器后输出的信号分为两路，其中一路经微分后送入第四乘法器，另一路直接送入第三乘法器；
(4)第三乘法器输出的乘积和第四乘法器输出的乘积相减得到差值；
(5)对上述差值进行积分得到积分值，解调出待测目标的振动信号。
通过对检测光束进行解调，提取出振动信号，如振动信号的振幅、频率、相位。
在一些实施例中，通过对参考光束和检测光束进行计算，得到被测量设备的振动速度，然后计算出被测设备的振动频率等。
步骤13：将振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，从而确定目标设备的故障类型。
其中，已训练的故障学习模型，是基于振动信号样本数据以及预先确定的故障类型标签进行训练后得到的。
在一些实施例中，将振动信号输入至已训练的故障学习模型，对振动信号进行特征提取，可以采用时频分析处理，提取信号处理的特征参数。其中，时频分析的方法，它的思路是设计时间和频率的联合函数，用它同时描述信号在不同时间和频率的能量密度或强度。利用时频分布来分析信号，能给出各个时刻的瞬时频率及其幅值，并且能够进行时频滤波和时变信号研究。时频分析的方法包括时域波形分析方法、概率密度分析方法、自相关分析方法、趋势分析方法、幅值谱分析方法、功率谱分析方法、倒频谱分析方法、包络谱分析方法、细化谱分析方法、短 时傅里叶分析方法、小波分析方法。以上只是举例说明，对于振动特征参数的获得不局限于上述方法，还可以采用其他的手段实现。
将提取到的特征信息进行识别，以识别与振动信号相关联的故障类型，从而确定目标设备的故障类型。
在一些实施例中，上述故障学习模型可以通过机器学习的方式来构建。利用监督式学习的方法，通过人为的输入不同的振动信号样本及根据振动信号样本对应的故障类型对振动信号样本进行标记处理，将振动信号样本输入至待训练的学习模型，以对待训练的学习模型进行训练，形成故障学习模型。相应标记可以是振动信号对应的故障类型。还可以将振动信号进行特征提取，并给这些特征信息设置对应的标记，将这些特征信息及标记作为训练资料，构建故障学习模型。当未知的振动信号输入该故障学习模型，则输出对应的标记，此标记即为振动信号对应的故障类型。
在一些实施例中，可以采用高斯混合模型、隐马尔科夫模型、K近邻、神经网络、支持向量机等进行模型训练，以训练故障学习模型，训练完成后，可采用该模型对未知的振动信号进行识别。
在一些实施例中，将振动信号输入已训练的故障学习模型，在故障学习模型中进行预处理，将预处理的振动信号进行识别，识别出与预处理的振动信号相关联的故障类型，从而反馈用户结果。
在一些实施例中，目标设备包括电机，电机工作产生振动，若目标设备有故障，电机振动所产生的振动信号也并不相同。基于大数据，先收集许多的故障的振动信号，以使振动信号来进行深度学习，以训练深度学习模型，当深度学习模型建立完成后，用于对后续的振动信号进行识别，以判断故障类型。
在一些实施例中，目标设备并不能产生振动，需要连接外部驱动设备，通过外部驱动设备来驱动目标设备产生振动，目标设备可以是一些焊接的部件。参考图3，以在焊接技术中的应用为例，将已焊接的钣金31与外部驱动设备32连接，以使外部驱动设备32振动带动钣金31振动，其中P为焊点，设备故障检测装置33进行振动信号的收集。钣金31焊接时存在常见缺陷有咬边、焊瘤、凹陷及焊接变形等，有时还有表面气孔和表面裂纹。单面焊的根部未焊透等、还有气孔、夹渣、裂纹、未熔合等。另外在焊接完成后，也存在焊点松动等故障。可以理解，每个缺陷均会导致在钣金31后续的应用中不良影响，所以均是不合格的焊接。并且每个缺陷所产生的振动信号均有区别，通过将采集到的振动信号进行特征提取输入已训练的故障学习模型中，以识别出相关联的故障类型。
区别于现有技术的情况，本申请的一种设备故障检测方法，包括：向目标设备发射测量光束，并采集测量光束经由目标设备反射的检测光束；从检测光束中提取振动信号；将振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，从而确定目标设备的故障 类型。通过上述方式，一方面采用非接触式振动信号测量方案，无需外置隔音房，高效检测设备故障问题，另一方面在大批量设备质检时，节省时间和人力成本，提高质检效率。
参阅图4，图4是本申请提供的设备故障检测方法第二实施例的流程示意图，该方法包括：
步骤41：向目标设备发射测量光束，并采集测量光束经由目标设备反射的检测光束。
步骤42：从检测光束中提取振动信号。
步骤41-42与上述实施例有相同或相似的技术方案，这里不做赘述
步骤43：将振动信号转化为声音信号。
步骤44：对声音信号进行预处理。
在一些实施例中，将声音信号进行预加重，其目的为了对语音的高频部分进行加重，增加语音的高频分辨率，以对声音信号中的高频分量进行补偿。通常采用数字滤波器来实现预加重。
进行预加重数字滤波处理后，接下来进行加窗分帧处理。语音信号是一种随时间而变化的信号，主要分为浊音和清音两大类。浊音的基音周期、清浊音信号幅度和声道参数等都随时间而缓慢变化。由于发声器官的惯性运动，可以认为在一小段时间里(一般为10～30ms)语音信号近似不变，即语音信号具有短时平稳性。这样，可以把语音信号分为一些短段(称为分析帧)来进行处理。语音信号的分帧是采用可移动的有限长度窗口进行加权的方法来实现的。一般每秒的帧数为33～100帧，视实际情况而定。分帧可以采用连续分段的方法，也可以采用交叠分段的方法，这是为了使帧与帧之间平滑过渡，保持其连续性，这里采用交叠分段的方法。前一帧和后一帧的交叠部分称为帧移，帧移与帧长的比值一般取0～1/2。这里可以采用矩形窗或者汉明窗来进行加窗处理。
步骤45：将预处理之后的声音信号输入至已训练的故障学习模型，以识别与声音信号相关联的故障类型，从而确定目标设备的故障类型。
其中，已训练的故障学习模型，是基于声音信号样本数据以及预先确定的故障类型标签进行训练后得到的。
在一些实施例中，将预处理之后的声音信号输入已训练的故障学习模型，对于声音信号进行特征提取。提取的特征信息可以包括：时域特征、频域特征、倒谱特征、时频特征等。
其中，时域特征主要有短时能量、短时平均幅度和短时过零率等。短时平均幅度反应一帧语音信号能量的大小。短时过零率表示一帧音频信号中波形穿过横轴的次数。
频域特征是将信号做傅里叶变换转换到频域，然后在频域上计算得到的特征。主要有频谱质心、带宽和频谱滚降系数等。频谱质心反映了每帧能量的均值。带宽反映了采样点能量在均值附近的波动程度。频谱滚降描述的是频谱的倾斜程度。
倒谱是功率谱或能量谱的对数值的逆傅里叶变换。
在一些实施例中，将预处理之后的声音信号输入到深度学习网络，在深度学习网络中进行预处理，将预处理的声音进行识别，识别出与预处理的声音信号相关联的故障类型，从而反馈用户结果。
参阅图5，图5是本申请提供的设备故障检测方法第三实施例的流程示意图，该方法包括：
步骤51：向目标设备发射测量光束，并采集测量光束经由目标设备反射的检测光束。
步骤52：从检测光束中提取振动信号。
步骤53：将振动信号转化为声音信号。
步骤54：对声音信号进行预处理。
步骤51-54与上述实施例有相同或相似的技术方案，这里不做赘述。
步骤55：将预处理之后的声音信号转化为声谱图。
声谱图是时间和频率的分布图。声谱图不仅体现了声信号的频域和时域特征，还同时展现出时域和频域两者的相互关系，从声谱图上可以观察到频域的一些特征随声音信号的发生而变化的情况；还可以观察到能量随声音过程的变化情况。所以声谱图所承载的声音信号的信息远大于单纯时域信号和单纯频域信号承载的信息。声谱图综合了频谱图和时域波形的特点，明显地显示出了声音频谱随时间的变化情况，或者说声谱图是一种动态的频谱。
在一些实施例中，分帧窗长选用512点，窗函数选用汉明窗，帧叠选0.75倍的窗长，然后通过Matlab(矩阵实验室)第三方工具VoiceBox(语音处理工具箱)调用specgram函数画出声谱图。
步骤56：将声谱图输入至已训练的故障学习模型，以识别与声谱图相关联的故障类型。
其中，已训练的故障学习模型，是基于声谱图样本数据以及预先确定的故障类型标签进行训练后得到的。
在一些实施例中，对转换的声音信号进行预加重，利用特定的高通滤波器对采集到的声音信号的高频部分进行补偿；然后，对预加重后的声音信号进行端点检测，确定该声音信号中有效信号的起始点，一般来说，由于采集到的声音信号中可能存在一段时间的静音或空白，为了提高异常声音的检测效率，本实施例可以通过确定声音信号中的有效声音信息(即，有效信号)的起始点，然后再对有效信号进行异常声音匹配或检测；再然后，对已确定起始点的有效信号在一定范围内的特征参数进行分帧加窗处理，使其满足统计特性平稳。
将预处理之后的声音信号转化为声谱图，声谱图是二维图，其横轴是时间，纵轴是频率。坐标(x，y)对应的点表示在时刻x，频率y上的声音强度，声音强度通过不同的颜色来表现。从声音信号的声谱图中，可以得出整个时间-频率范围内声音强度的分布和变化情况；而这是波形图中无法呈现的。为获得声谱图，声音信号被分割成很短的帧，相邻帧会有一定的重叠。然后对每个帧做短时傅里叶变换得到对应的频谱信息， 由于声谱图由频率、时间、声音强度三个维度信息构成，因此需要对声音强度的取值进行计算。最终将频谱信息连接成完整的声谱图。
举例来说，本申请实施例将有效信号片段加窗划分为若干个帧；对每一帧进行短时傅里叶变换，得到该帧的频谱信息，频谱信息用于表示该帧的频率与声音强度之间的关系；连接所有帧的频谱信息，得到有效信号片段的声谱图，声谱图由若干个点组成，任一点的坐标(x，y)用于表示该点在x时刻，y频率上对应的声音强度。在本申请实施例中，对声音的时序信号进行短时傅里叶变化，傅里叶变换的长度为2N点，这样每一帧的信号都可以得到长度为N的频谱，每一点的声压值表示为：P＝20*log10|x(1/N)|；其中，P为该点的声压值，x为该帧信号的频谱值。
根据声谱图区别异常声音和背景噪声。根据声谱图中的每个点所处于的帧、频率和声音强度，生成声谱图矩阵。从声谱图矩阵中提取用于表征声谱图的声音强度分布情况的待测试识别特征矩阵。将待测试识别特征矩阵输入至已训练的故障学习模型中，通过模型的内部识别，输出待测试识别特征矩阵的故障类型。
参阅图6，图6是本申请提供的设备故障检测方法第四实施例的流程示意图，该方法包括：
步骤61：向目标设备发射测量光束，并采集测量光束经由目标设备反射的检测光束。
步骤62：从检测光束中提取振动信号。
步骤63：将振动信号转化为声音信号。
步骤64：对声音信号进行预处理。
步骤65：将预处理之后的声音信号转化为声谱图。
步骤61-65与上述实施例有相同或相似的技术方案，这里不做赘述。
步骤66：提取声谱图的时间信息、频率信息和能量信息，以得到待处理特征信息。
在本实施例中，声谱图横坐标表示时间，纵坐标表示频率，时间与频率的坐标点的值表示声音信号能量。通常能量值的大小通过颜色来表示，颜色深表示该点的能量越强。
在一些实施例中，采用Gabor变换法进行特征提取，可以从多方向多频率尺度提取声谱图的特征。
在一些实施例中，采用投影法进行特征提取。投影法就是把一幅声谱图分别往四个方向上(0度,45度,90度,135度)做投影，往0度方向(横坐标)投影，即同一时间分量，不同频率分量的像素亮度值都累加起来；往90度方向(纵坐标)投影，即同一频率分量，不同时间分量的像素亮度值都累加起来；往45度、135度方向投影，把垂直该方向的像素亮度值都累加起来，为了保证数据量的一致性，每次累加后都要除以像素点的个数。
步骤67：将待处理特征信息输入至已训练的故障学习模型，以识别与声谱图相关联的故障类型。
其中，已训练的故障学习模型，是基于声谱图样本数据以及预先确定的故障类型标签进行训练后得到的。
在本实施例中，故障学习模型通过机器学习的中的监督式学习来进行建模。预先将故障类型与特征信息对应，且对特征信息进行标注，标注的内容就可以是故障类型，在故障学习模型建立后，输入待处理的特征信息，故障学习模型则输出对应的标注信息，即故障类型。
在一些实施例中，在训练故障学习模型时，将故障类型进行分类。以洗衣机为例，将螺丝未固定分为一类，将电机故障分为一类。当确定当前声谱图的特征信息为螺丝未固定一类，则将当前声谱图继续在螺丝未固定一类中进行比对。
在一些实施例中，当目标设备的声音信息的待处理特征信息和标准特征信息的相似度大于设定阈值，确定声音信号相关联的故障类型为标准特征信息对应的故障类型。以目标设备为冰箱为例，将通过故障检测设备获取到的振动信号转换为声音信号，将声音信号转换为声谱图，在深度学习网络中对声谱图进行特征提取，将特征信息和预存的标准特征信息进行相似度比对，比对的相似度为百分之八十五，而设定阈值为百分之八十，相似度大于设定阈值，则确定声音信号相关联的故障类型为标准特征信息对应的故障类型。
参阅图7，图7是本申请提供的电子设备的语音交互方法第五实施例的流程示意图，该方法包括：
步骤71：向目标设备发射测量光束，并采集测量光束经由目标设备反射的检测光束。
步骤72：从检测光束中提取振动信号。
步骤73：将振动信号转化为声音信号。
步骤74：对声音信号进行预处理。
步骤75：将预处理之后的声音信号转化为声谱图。
步骤76：将声谱图输入至已训练的故障学习模型。
步骤71-76与上述实施例有相同或相似的技术方案，这里不做赘述。
步骤77：利用故障学习模型对声谱图进行区域分块，并将不同区域分配对应权重，以得到待处理分块声谱图。
步骤78：利用故障学习模型对待处理分块声谱图进行加权处理，以得到加权后的相似度比对结果。
步骤79：响应于加权后的相似度比对结果大于设定阈值，确定声音信号相关联的故障类型为声谱图对应的故障类型。
参考图8，对步骤77-79进行说明，图8的左侧为待处理分块声谱图，右侧为标准的分块声谱图。将左侧的待处理分块声谱图分为ABCD四块，并对ABCD四块分配权重，并对A与a、B与b、C与c、D与d进行加权处理。举例说明，将A、B、C、D分别分配权重30％、20％、40％、10％；其中，A、B、C、D的权重值相加为1。相似度比对结果的计算公式为S＝A*a*30％+B*b*20％+C*c*40％+D*d*10％。
其中，A*a表示A与a的相似度比对结果。B*b表示B与b的相似度比对结果。C*c表示C与c的相似度比对结果。D*d表示D与d的相似度比对结果。
当S的值大于设定阈值，则确定声音信号相关联的故障类型为设定阈值对应的故障类型。
例如设定阈值为70％，而
S＝90％*30％+80％*20％+70％*40％+60％*10％＝77％，因此，S>70％；则确定声音信号相关联的故障类型为声谱图对应的故障类型。
参阅图9，图9为本申请提供的设备故障检测装置一实施例的结构示意图，设备故障检测装置90包括激光发射器91、激光接收器92和处理器93。
激光发射器91，用于向目标设备发射测量光束。
激光接收器92，用于采集测量光束经由目标设备反射的检测光束。
处理器93，连接激光接收器，用于从检测光束中提取振动信号，并将振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，从而确定目标设备的故障类型；其中，已训练的故障学习模型，是基于振动信号样本数据以及预先确定的故障类型标签进行训练后得到的。
处理器93还用于将振动信号转化为声音信号，并将声音信号输入至已训练的故障学习模型，以识别与声音信号相关联的故障类型，从而确定目标设备的故障类型；其中，已训练的故障学习模型，是基于声音信号样本数据以及预先确定的故障类型标签进行训练后得到的。
处理器93还用于对声音信号进行预处理，并将预处理之后的声音信号输入至已训练的故障学习模型，以识别与声音信号相关联的故障类型，从而确定目标设备的故障类型。
处理器93还用于将振动信号转化为声音信号，对声音信号进行预处理，并将预处理之后的声音信号转化为声谱图，将声谱图输入至已训练的故障学习模型，以识别与声谱图相关联的故障类型；其中，已训练的故障学习模型，是基于声谱图样本数据以及预先确定的故障类型标签进行训练后得到的。
处理器93还用于提取声谱图的时间信息、频率信息和能量信息，以得到待处理特征信息；将待处理特征信息输入至已训练的故障学习模型，以识别与声谱图相关联的故障类型。
处理器93还用于将声谱图输入至已训练的故障学习模型；利用故障学习模型对声谱图进行区域分块，并将不同区域分配对应权重，以得到待处理分块声谱图；利用故障学习模型对待处理分块声谱图进行加权处理，以得到加权后的相似度比对结果；响应于加权后的相似度比对结果大于设定阈值，确定声音信号相关联的故障类型为声谱图对应的故障类型。
处理器93连接激光发射器91，用于将激光发射器91发射的激光束 分为测量光束和参考光束，并将检测光束和参考光束进行干涉，以检测得到振动信号。
可以理解，采用设备故障检测装置90可以实现上述任一实施例的技术方案。
在一些实施例中，参考图10，设备故障检测装置100包括激光发射器101、分束器102、分束器103、分束器104、布拉格盒105、固定反射器106、光电检测器107。可以理解，图10中仅示意了激光束的发射和接收过程，后续处理的部分结构图未示。图中200为目标设备。
下面介绍下具体工作流程，激光发射器101发射激光束到分束器102，分束器102将激光束将分为测量光束和参考光束；如图测量光束的方向往分束器103，参考光束射向固定反射器106，参考光束经由固定反射器106反射，射入布拉格盒105，通过布拉格盒105后射向分束器104，参考光束经分束器104射向光电检测器107；测量光束经分束器103后射向目标设备200，并收集反射的检测光束，检测光束经分束器103向下偏转射向分束器104；检测光束经分束器104射向光电检测器107；此时参考光束和检测光束均射入光电检测器107，通过光电检测器107利用多普勒频移与目标设备200的振动速度成正比的关系，计算得到目标设备的振动信号。
具体地，由于参考光束的光路为常数，目标设备200的振动会在光电检测器107上产生亮/暗条纹，这是利用的干涉法。光电检测器107上的一个完整的亮/暗周期条纹正好与所用激光的半个波长的位移量相对应。这在激光发射器101经常使用的氦氖激光的情况下，对应于316nm的位移。
每单位时间的光程改变表现为测量光束的多普勒频移。在计量方面，意味着多普勒频移直接与样本振动速度成正比。由于远离干涉仪的物体运动所产生的明暗条纹(和调制频率)与物体朝向干涉仪移动所产生的相同，因此这种设置无法明确物体移动的方向。鉴于此，将光频移典型值为40MHz的声光调制器放置在参考光束中(出于比较目的，激光频率为4.74·1014Hz)。当目标设备200处于静态时，将产生40MHz的典型干涉调制频率。因此，当目标设备200朝设备故障检测装置移动时，调制频率会增加；当目标设备200远离干涉仪移动时，则检测器接收到的频率则小于40MHz。不仅能精确检测光程长度，还能检测出运动方向。
具体地，除可以直接测量出振动速度外，设备故障检测装置100还可直接测量出位移量。谐波振动的最大振幅可以表示如下：
v＝2π*f*s；其中v代表目标设备的速度；f代表目标设备的振动频率；s代表目标设备的振动位移。随着频率的增加，振动速度增加，振动位移则减小。
参阅图11，图11是本申请提供的计算机存储介质一实施例的结构示意图，计算机存储介质110用于存储程序数据111，程序数据111在 被处理器执行时，用于实现以下的方法步骤：
向目标设备发射测量光束，并采集测量光束经由目标设备反射的检测光束；从检测光束中提取振动信号；将振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，从而确定目标设备的故障类型。
可以理解，程序数据111在被处理器执行时，还用于实现上述任一实施例方法。
在本申请所提供的几个实施方式中，应该理解到，所揭露的方法以及设备，可以通过其它的方式实现。例如，以上所描述的设备实施方式仅仅是示意性的，例如，所述模块或单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。
上述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施方式方案的目的。
另外，在本申请各个实施方式中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
上述其他实施方式中的集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)或处理器(processor)执行本申请各个实施方式所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、磁碟或者光盘等各种可以存储程序代码的介质。
以上所述仅为本申请的实施方式，并非因此限制本申请的专利范围，凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本申请的专利保护范围内。

Claims (19)

  1. 一种设备故障检测方法，其特征在于，所述方法包括：
    向目标设备发射测量光束，并采集所述测量光束经由所述目标设备反射的检测光束；
    从所述检测光束中提取振动信号；
    将所述振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，从而确定所述目标设备的故障类型。
  2. 根据权利要求1所述的方法，其特征在于，
    所述已训练的故障学习模型，是基于振动信号样本数据以及预先确定的故障类型标签进行训练后得到的。
  3. 根据权利要求1所述的方法，其特征在于，
    所述将所述振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，包括：
    将所述振动信号转化为声音信号；
    将所述声音信号输入至已训练的故障学习模型，以识别与声音信号相关联的故障类型，从而确定所述目标设备的故障类型；
    其中，所述已训练的故障学习模型，是基于声音信号样本数据以及预先确定的故障类型标签进行训练后得到的。
  4. 根据权利要求3所述的方法，其特征在于，
    所述将所述声音信号输入至已训练的故障学习模型，以识别与声音信号相关联的故障类型，从而确定所述目标设备的故障类型，包括：
    对所述声音信号进行预处理；
    将预处理之后的所述声音信号输入至已训练的故障学习模型，以识别与声音信号相关联的故障类型，从而确定所述目标设备的故障类型。
  5. 根据权利要求3所述的方法，其特征在于，
    所述对所述声音信号进行预处理，包括：
    对所述声音信号进行预加重，以对所述声音信号中的高频分量进行补偿；
    采用预先设置的窗函数对所述声音信号进行分帧加窗处理，以得到预处理之后的所述声音信号。
  6. 根据权利要求1所述的方法，其特征在于，
    所述将所述振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，包括：
    将所述振动信号转化为声音信号；
    对所述声音信号进行预处理；
    将预处理之后的所述声音信号转化为声谱图；
    将所述声谱图输入至已训练的故障学习模型，以识别与所述声谱图相关联的故障类型，从而确定所述目标设备的故障类型；
    其中，所述已训练的故障学习模型，是基于声谱图样本数据以及预 先确定的故障类型标签进行训练后得到的。
  7. 根据权利要求6所述的方法，其特征在于，
    所述将所述声谱图输入至已训练的故障学习模型，以识别与所述声谱图相关联的故障类型，包括：
    提取所述声谱图的时间信息、频率信息和能量信息，以得到待处理特征信息；
    将所述待处理特征信息输入至已训练的故障学习模型，以识别与所述声谱图相关联的故障类型。
  8. 根据权利要求6所述的方法，其特征在于，
    所述将所述声谱图输入至已训练的故障学习模型，以识别与所述声谱图相关联的故障类型，包括：
    将所述声谱图输入至已训练的故障学习模型；
    利用所述故障学习模型对所述声谱图进行区域分块，并将不同区域分配对应权重，以得到待处理分块声谱图；
    利用所述故障学习模型对所述待处理分块声谱图进行加权处理，以得到加权后的相似度比对结果；
    响应于所述加权后的相似度比对结果大于设定阈值，确定所述声音信号相关联的故障类型为所述声谱图对应的故障类型。
  9. 根据权利要求1所述的方法，其特征在于，
    所述向目标设备发射测量光束，并采集所述测量光束经由所述目标设备反射的检测光束之前，还包括：
    将激光束分为测量光束和参考光束；
    所述从所述检测光束中提取振动信号，包括：
    将所述检测光束和所述参考光束进行干涉，以检测得到振动信号。
  10. 根据权利要求1所述的方法，其特征在于，
    所述目标设备包括电机，所述目标设备由所述电机工作而产生振动。
  11. 根据权利要求1所述的方法，其特征在于，
    所述向目标设备发射测量光束，并采集所述测量光束经由所述目标设备反射的检测光束之前，还包括：
    通过外部驱动设备连接所述目标设备，以驱动所述目标设备产生振动。
  12. 一种设备故障检测装置，其特征在于，所述设备故障检测装置包括：
    激光发射器，用于向目标设备发射测量光束；
    激光接收器，用于采集所述测量光束经由所述目标设备反射的检测光束；
    处理器，连接所述激光接收器，用于从所述检测光束中提取振动信号，并将所述振动信号输入至已训练的故障学习模型，以识别与振动信号相关联的故障类型，从而确定所述目标设备的故障类型；
    其中，所述已训练的故障学习模型，是基于振动信号样本数据以及预先确定的故障类型标签进行训练后得到的。
  13. 根据权利要求12所述的装置，其特征在于，
    所述处理器还用于将所述振动信号转化为声音信号，并将所述声音信号输入至已训练的故障学习模型，以识别与声音信号相关联的故障类型，从而确定所述目标设备的故障类型；
    其中，所述已训练的故障学习模型，是基于声音信号样本数据以及预先确定的故障类型标签进行训练后得到的。
  14. 根据权利要求13所述的装置，其特征在于，
    所述处理器还用于对所述声音信号进行预处理，并将所述预处理之后的所述声音信号输入至已训练的故障学习模型，以识别与所述声音信号相关联的故障类型，从而确定所述目标设备的故障类型。
  15. 根据权利要求12所述的装置，其特征在于，
    所述处理器还用于将所述振动信号转化为声音信号，对所述声音信号进行预处理，并将预处理之后的所述声音信号转化为声谱图，将所述声谱图输入至已训练的故障学习模型，以识别与所述声谱图相关联的故障类型；
    其中，所述已训练的故障学习模型，是基于声谱图样本数据以及预先确定的故障类型标签进行训练后得到的。
  16. 根据权利要求15所述的装置，其特征在于，
    所述处理器还用于提取所述声谱图的时间信息、频率信息和能量信息，以得到待处理特征信息；将所述待处理特征信息输入至已训练的故障学习模型，以识别与所述声谱图相关联的故障类型。
  17. 根据权利要求15所述的装置，其特征在于，
    所述处理器还用于将所述声谱图输入至已训练的故障学习模型；利用所述故障学习模型对所述声谱图进行区域分块，并将不同区域分配对应权重，以得到待处理分块声谱图；利用所述故障学习模型对所述待处理分块声谱图进行加权处理，以得到加权后的相似度比对结果；响应于所述加权后的相似度比对结果大于设定阈值，确定所述声音信号相关联的故障类型为所述声谱图对应的故障类型。
  18. 根据权利要求12所述的装置，其特征在于，
    所述处理器连接所述激光发射器，用于将所述激光发射器发射的激光束分为测量光束和参考光束，并将所述检测光束和所述参考光束进行干涉，以检测得到振动信号。
  19. 一种计算机存储介质，其特征在于，所述计算机存储介质用于存储程序数据，所述程序数据在被处理器执行时，用于实现如权利要求1-10任一项所述的方法。
PCT/CN2019/124118 2019-10-24 2019-12-09 设备故障检测方法、设备故障检测装置及计算机存储介质 WO2021077567A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911019720.1 2019-10-24
CN201911019720.1A CN112710486B (zh) 2019-10-24 2019-10-24 设备故障检测方法、设备故障检测装置及计算机存储介质

Publications (1)

Publication Number Publication Date
WO2021077567A1 true WO2021077567A1 (zh) 2021-04-29

Family

ID=75540471

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/124118 WO2021077567A1 (zh) 2019-10-24 2019-12-09 设备故障检测方法、设备故障检测装置及计算机存储介质

Country Status (2)

Country Link
CN (1) CN112710486B (zh)
WO (1) WO2021077567A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113268924A (zh) * 2021-05-18 2021-08-17 国网福建省电力有限公司电力科学研究院 基于时频特征的变压器有载分接开关故障识别方法
CN113820624A (zh) * 2021-09-30 2021-12-21 南方电网科学研究院有限责任公司 一种配电网高阻接地故障识别装置
CN113820109A (zh) * 2021-08-23 2021-12-21 广东电力发展股份有限公司 一种电厂辅机旋转设备巡检装置、方法、设备及介质
CN115629930A (zh) * 2022-12-23 2023-01-20 北京东远润兴科技有限公司 基于dsp系统的故障检测方法、装置、设备及存储介质
CN115951002A (zh) * 2023-03-10 2023-04-11 山东省计量科学研究院 一种气质联用仪故障检测装置
CN116933170A (zh) * 2023-09-18 2023-10-24 福建福清核电有限公司 一种机械密封故障分类算法

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114689161B (zh) * 2022-03-30 2023-05-09 珠海格力电器股份有限公司 检测方法、非易失性存储介质以及检测系统
CN114754413B (zh) * 2022-04-11 2023-10-27 青岛海信日立空调系统有限公司 一种多联机空调系统及故障定位方法
CN115691537B (zh) * 2022-12-28 2023-06-23 江苏米笛声学科技有限公司 一种耳机音频信号的分析与处理系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070175283A1 (en) * 2006-02-01 2007-08-02 General Electric Company Systems and Methods for Remote Monitoring of Vibrations in Machines
CN203241143U (zh) * 2013-05-14 2013-10-16 西安邮电大学 一种基于激光多普勒效应的电机振动在线监测装置
CN104215320A (zh) * 2014-09-22 2014-12-17 珠海格力电器股份有限公司 用于电机设备的测振装置和系统
CN105737965A (zh) * 2016-02-29 2016-07-06 莆田学院 一种风力发电机振动检测装置及分析方法
CN105890740A (zh) * 2016-06-17 2016-08-24 佛山科学技术学院 一种空调六维振动测试系统及其方法
CN107024331A (zh) * 2017-03-31 2017-08-08 中车工业研究院有限公司 一种神经网络对列车电机振动在线检测方法

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
RU2090853C1 (ru) * 1993-08-06 1997-09-20 Павел Анатольевич Давыдов Способ виброакустической диагностики машинного оборудования
US5679899A (en) * 1995-03-06 1997-10-21 Holographics Inc. Method and apparatus for non-destructive testing of structures
CN105810213A (zh) * 2014-12-30 2016-07-27 浙江大华技术股份有限公司 一种典型异常声音检测方法及装置
CN107103270A (zh) * 2016-02-23 2017-08-29 云智视像科技（上海）有限公司 一种基于idf的动态计算分块加权系数的人脸识别系统
CN106404388B (zh) * 2016-09-13 2018-10-19 西安科技大学 一种刮板输送机飘链故障诊断方法
CN106546892A (zh) * 2016-11-10 2017-03-29 华乘电气科技（上海）股份有限公司 基于深度学习的局部放电超声音频识别方法及系统
CN106599185B (zh) * 2016-12-14 2020-10-23 北京微智信业科技有限公司 基于hsv的图像相似度识别方法
CN109697501A (zh) * 2017-10-20 2019-04-30 丹东东方测控技术股份有限公司 基于深度学习的磨机声音声谱分析设备
CN108710756A (zh) * 2018-05-18 2018-10-26 上海电力学院 基于谱聚类分析下多特征信息加权融合的故障诊断方法
CN110322896A (zh) * 2019-06-26 2019-10-11 上海交通大学 一种基于卷积神经网络的变压器故障声音识别方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070175283A1 (en) * 2006-02-01 2007-08-02 General Electric Company Systems and Methods for Remote Monitoring of Vibrations in Machines
CN203241143U (zh) * 2013-05-14 2013-10-16 西安邮电大学 一种基于激光多普勒效应的电机振动在线监测装置
CN104215320A (zh) * 2014-09-22 2014-12-17 珠海格力电器股份有限公司 用于电机设备的测振装置和系统
CN105737965A (zh) * 2016-02-29 2016-07-06 莆田学院 一种风力发电机振动检测装置及分析方法
CN105890740A (zh) * 2016-06-17 2016-08-24 佛山科学技术学院 一种空调六维振动测试系统及其方法
CN107024331A (zh) * 2017-03-31 2017-08-08 中车工业研究院有限公司 一种神经网络对列车电机振动在线检测方法

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113268924A (zh) * 2021-05-18 2021-08-17 国网福建省电力有限公司电力科学研究院 基于时频特征的变压器有载分接开关故障识别方法
CN113268924B (zh) * 2021-05-18 2023-01-31 国网福建省电力有限公司电力科学研究院 基于时频特征的变压器有载分接开关故障识别方法
CN113820109A (zh) * 2021-08-23 2021-12-21 广东电力发展股份有限公司 一种电厂辅机旋转设备巡检装置、方法、设备及介质
CN113820624A (zh) * 2021-09-30 2021-12-21 南方电网科学研究院有限责任公司 一种配电网高阻接地故障识别装置
CN113820624B (zh) * 2021-09-30 2024-04-16 南方电网科学研究院有限责任公司 一种配电网高阻接地故障识别装置
CN115629930A (zh) * 2022-12-23 2023-01-20 北京东远润兴科技有限公司 基于dsp系统的故障检测方法、装置、设备及存储介质
CN115629930B (zh) * 2022-12-23 2023-08-08 北京东远润兴科技有限公司 基于dsp系统的故障检测方法、装置、设备及存储介质
CN115951002A (zh) * 2023-03-10 2023-04-11 山东省计量科学研究院 一种气质联用仪故障检测装置
CN116933170A (zh) * 2023-09-18 2023-10-24 福建福清核电有限公司 一种机械密封故障分类算法
CN116933170B (zh) * 2023-09-18 2024-01-02 福建福清核电有限公司 一种机械密封故障分类方法

Also Published As

Publication number Publication date
CN112710486B (zh) 2022-01-25
CN112710486A (zh) 2021-04-27

Similar Documents

Publication Publication Date Title
WO2021077567A1 (zh) 设备故障检测方法、设备故障检测装置及计算机存储介质
Mohd Ghazali et al. Vibration analysis for machine monitoring and diagnosis: a systematic review
Dai et al. Structure damage localization with ultrasonic guided waves based on a time–frequency method
Li et al. A novel ensemble deep learning model for cutting tool wear monitoring using audio sensors
CN104483009B (zh) 一种低频随机扰动下中高频振动的纳米级振幅测量方法
CN110333285A (zh) 基于变分模态分解的超声兰姆波缺陷信号识别方法
Lv et al. The effect of speckles noise on the Laser Doppler Vibrometry for remote speech detection
Xu et al. Remote eavesdropping at 200 meters distance based on laser feedback interferometry with single-photon sensitivity
Scislo Quality assurance and control of steel blade production using full non-contact frequency response analysis and 3d laser doppler scanning vibrometry system
TWI789645B (zh) 沖壓品質檢測系統及沖壓品質檢測方法
CN104061998A (zh) 漫反射式零差正交激光测振仪
Chiariotti et al. Delamination detection by multi-level wavelet processing of continuous scanning laser Doppler vibrometry data
Hu et al. Adaptive instantaneous frequency ridge extraction based on target tracking for frequency-modulated signals
Singh et al. Countermeasures to replay attacks: A review
Becker et al. Evaluation of an autoregressive spectral estimator for modal wave number estimation in range-dependent shallow water waveguides
CN116243111A (zh) 一种电缆束故障检测定位方法及系统
Sracic et al. Experimental investigation of the effect of speckle noise on continuous scan laser Doppler vibrometer measurements
CN105737965A (zh) 一种风力发电机振动检测装置及分析方法
CN104991245A (zh) 一种无人飞行器预警装置及其预警方法
CN114994185A (zh) 一种激励数据驱动下能量分布转异的疲劳损伤检测方法
Kereliuk et al. Improved hidden Markov model partial tracking through time-frequency analysis
Salvino et al. EMD and instantaneous phase detection of structural damage
WO2020029236A1 (zh) 振动监控方法和系统
CN111141830A (zh) 基于微纳耦合光纤传感器的线性定位系统及方法
Xiao et al. Full-field laser heterodyne imaging vibrometry using a CMOS–DVR system

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19950079

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 29/09/2022)

122 Ep: pct application non-entry in european phase

Ref document number: 19950079

Country of ref document: EP

Kind code of ref document: A1