WO2023185801A1 - 振动信号的识别方法及装置 - Google Patents

振动信号的识别方法及装置 Download PDF

Info

Publication number
WO2023185801A1
WO2023185801A1 PCT/CN2023/084285 CN2023084285W WO2023185801A1 WO 2023185801 A1 WO2023185801 A1 WO 2023185801A1 CN 2023084285 W CN2023084285 W CN 2023084285W WO 2023185801 A1 WO2023185801 A1 WO 2023185801A1
Authority
WO
WIPO (PCT)
Prior art keywords
signal
convolution
sub
signals
neural network
Prior art date
Application number
PCT/CN2023/084285
Other languages
English (en)
French (fr)
Inventor
陈曦
葛成
王明
Original Assignee
阿里云计算有限公司
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 阿里云计算有限公司 filed Critical 阿里云计算有限公司
Publication of WO2023185801A1 publication Critical patent/WO2023185801A1/zh

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01H—MEASUREMENT OF MECHANICAL VIBRATIONS OR ULTRASONIC, SONIC OR INFRASONIC WAVES
    • G01H17/00—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves, not provided for in the preceding groups
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the present invention relates to the field of fault identification technology, and in particular to a vibration signal identification method and device.
  • vibration signals can reflect the operating status of industrial equipment more intuitively, quickly and accurately, and are an ideal way to detect abnormalities in industrial equipment. An important means of detection and fault diagnosis.
  • Embodiments of the present application provide a method and device for identifying vibration signals to solve the problem of low accuracy in identifying fault types in the prior art.
  • inventions of the present application provide a method for identifying vibration signals.
  • the method includes:
  • the vibration signal is divided into multiple sub-signals
  • the acoustic characteristics of the multi-segment sub-signals are input into the neural network model, and the recognition results output by the neural network model are obtained.
  • the recognition results are used to characterize the occurrence probability of each fault type of the target component, so
  • the neural network model is used to perform secondary feature learning on the acoustic features.
  • the acoustic features are two-dimensional features, the first dimension of the acoustic features is used to characterize the number of frames of the sub-signal, and the second dimension of the acoustic features is used to characterize each frame. Mel spectrum cepstrum coefficients of the signal.
  • determining the acoustic characteristics of the multiple sub-signals respectively includes:
  • Discrete Fourier transform is used to convert logarithmic energies corresponding to the multi-segment sub-signals into Mel spectrum cepstral coefficients of the multi-segment sub-signals respectively.
  • the method before using discrete Fourier transform to convert the multi-segment sub-signal from a time domain signal to a frequency domain signal, the method further includes:
  • the multi-frame signals corresponding to each sub-signal are separately windowed.
  • the neural network model includes a one-dimensional dual convolutional neural network model.
  • the one-dimensional double convolutional neural network model includes a first convolution layer and a second convolution layer with the same structure.
  • the first convolution layer and the second convolution layer Each layer includes two convolution units, and the convolution parameters between each convolution unit are different.
  • the convolution parameters include the number of channels and the size of the convolution kernel, and the number of channels of the convolution unit of the first convolution layer is smaller than that of the convolution unit of the second convolution layer.
  • the number of channels, the convolution kernel size of the convolution unit of the first convolution layer is larger than the convolution kernel size of the convolution unit of the second convolution layer.
  • the channel of the convolution unit is used to receive Mel spectrum cepstrum coefficients of different frames of the same sub-signal.
  • the first convolution layer is provided before the second convolution layer, and a first pooling layer is provided between the first convolution layer and the second convolution layer, A second pooling layer is provided after the second convolution layer;
  • the first pooling layer is used to perform maximum pooling processing on the features extracted by the first convolution layer
  • the second pooling layer is used to perform max pooling processing on the features extracted by the second convolution layer. Perform adaptive average pooling.
  • inventions of the present application provide a device for identifying vibration signals.
  • the device method includes:
  • the acquisition module is used to acquire the vibration signal of the target component of the equipment to be detected
  • a segmentation module used to segment the vibration signal into multiple sub-signals according to the set sample length
  • Determining module used to determine the acoustic characteristics of the multiple sub-signals respectively
  • An identification module configured to input the acoustic characteristics of the multi-segment sub-signals into a neural network model and obtain the identification results output by the neural network model.
  • the identification results are used to characterize the occurrence probability of each fault type of the target component.
  • the neural network model is used to perform secondary feature learning on the acoustic features.
  • the acoustic features are two-dimensional features, the first dimension of the acoustic features is used to characterize the number of frames of the sub-signal, and the second dimension of the acoustic features is used to characterize each frame. Mel spectrum cepstrum coefficients of the signal.
  • the determination module is specifically configured to use discrete Fourier transform to convert the multi-segment sub-signals from time domain signals to frequency domain signals; respectively determine the frequency domain corresponding to the multi-segment sub-signals.
  • the power spectral rate of the signal use a triangular filter to perform Mel filtering on the power spectral rate of the frequency domain signal corresponding to the multi-segment sub-signal, and determine the logarithmic energy corresponding to the multi-segment sub-signal respectively; use discrete Fourier transform to The logarithmic energies corresponding to the multi-segment sub-signals are respectively converted into Mel spectrum cepstrum coefficients of the multi-segment sub-signals.
  • the determination module is further configured to perform frame processing on each sub-signal according to a preset number of signal sampling points to obtain a multi-frame signal corresponding to each sub-signal;
  • the multi-frame signals corresponding to each sub-signal are windowed separately.
  • the neural network model includes a one-dimensional dual convolutional neural network model.
  • the one-dimensional double convolutional neural network model includes a first convolution layer and a second convolution layer with the same structure, and the first convolution layer and The second convolution layer includes two convolution units, and the convolution parameters between each convolution unit are different.
  • the convolution parameters include the number of channels and the size of the convolution kernel, and the number of channels of the convolution unit of the first convolution layer is smaller than that of the convolution unit of the second convolution layer.
  • the number of channels, the convolution kernel size of the convolution unit of the first convolution layer is larger than the convolution kernel size of the convolution unit of the second convolution layer.
  • the channel of the convolution unit is used to receive Mel spectrum cepstrum coefficients of different frames of the same sub-signal.
  • the first convolution layer is provided before the second convolution layer, and a first pooling layer is provided between the first convolution layer and the second convolution layer, A second pooling layer is provided after the second convolution layer;
  • the first pooling layer is used to perform maximum pooling processing on the features extracted by the first convolution layer
  • the second pooling layer is used to perform max pooling processing on the features extracted by the second convolution layer. Perform adaptive average pooling.
  • the present application also provides an electronic device, including: a processor, and a memory; the memory is used to store a computer program of the processor; the processor is configured to implement by executing the computer program Any possible method in the first aspect.
  • the present invention also provides a computer storage medium that stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing any one of the possible methods in the first aspect.
  • embodiments of the present disclosure provide a computer program product, including a computer program that, when executed by a processor, implements the method described in the above first aspect and various possible designs of the first aspect.
  • the embodiments of the present application provide a method and device for identifying vibration signals.
  • the vibration signals of the target components of the equipment to be detected are obtained.
  • the vibration signal is divided into multiple sub-signals according to the set sample length. Again, the acoustic characteristics of the multiple sub-signals are determined respectively.
  • the acoustic features of the multi-segment sub-signals are input into the neural network model, and the recognition results output by the neural network model are obtained.
  • the recognition results are used to characterize the occurrence probability of each fault type of the target component, and the neural network model is used to classify the acoustic features. Secondary feature learning.
  • the acoustic features are extracted as the characteristics of the vibration signal, and the acoustic features are identified through the neural network model of secondary feature learning to determine the probability of occurrence of each fault type of the target component, thereby improving the accuracy of fault type identification.
  • Figure 1 is a schematic diagram of an application scenario of a vibration signal identification method provided by an embodiment of the present application
  • Figure 2 is a schematic flow chart of a vibration signal identification method provided by an embodiment of the present application.
  • FIG. 3 is a schematic flow chart of another vibration signal identification method provided by an embodiment of the present application.
  • Figure 4 is a schematic structural diagram of a neural network model provided by an embodiment of the present application.
  • Figure 5 is a schematic structural diagram of a vibration signal identification device provided by an embodiment of the present application.
  • FIG. 6 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.
  • the term “include” and its variations are open-ended, ie, “including but not limited to.”
  • the term “based on” means “based at least in part on.”
  • the term “one embodiment” means “at least one embodiment”; the term “another embodiment” means “at least one additional embodiment”; and the term “some embodiments” means “at least some embodiments”. Relevant definitions of other terms will be given in the description below.
  • Vibration signals usually include two types of vibration sources with different properties: one type of vibration source is caused by the unbalanced mass of mechanical moving parts, misalignment of geometric axes, poor gear kneading, improper coordination of transmission parts, excessive journal bearing clearance, etc.
  • Forced vibration such as periodic vibration, impact vibration, random vibration, etc.
  • another type of vibration source is vibration sound caused by structural response, self-excited vibration or environmental vibration. Response, such as: surge vibration of fluid, oil film vibration of bearings, response vibration of the component itself, local vibration of the structure, etc.
  • the vibration response will also change due to external load changes.
  • Monitoring and identifying vibration signals can provide insight into the different working conditions and health status of the equipment, which plays an important role in improving the stable operation of the equipment. Through the processing and analysis of vibration signals, early signs of equipment failure can be discovered in time, thereby predicting possible equipment failures, providing scientific basis for preventing accidents and scientifically arranging maintenance, saving maintenance costs, and improving equipment reliability and safety.
  • vibration signals can reflect the operating status of industrial equipment more intuitively, quickly and accurately, and are an ideal tool for industrial equipment. An important means for anomaly detection and fault diagnosis.
  • manual feature extraction and feature screening can be performed on the vibration signal based on domain knowledge and actual engineering needs, and then a machine learning classifier model is established, using the filtered features as input and the state that needs to be identified as Output to achieve classification of different vibration signals to determine the fault status of the target component of the equipment.
  • the first several layers of the multi-layer network can carry out machine-independent feature extraction. Each layer can obtain different representations of the input data, and the last layer can achieve Status classification to determine the fault status of the target component of the equipment.
  • embodiments of the present application provide a method and device for identifying vibration signals.
  • a neural network model is obtained.
  • the output recognition results are used to determine the probability of occurrence of each failure type of the target component. Since the neural network model is more suitable for temporal feature learning, it can better extract high-order features from acoustic features, with higher robustness and generalization, thereby improving the accuracy of fault type identification.
  • FIG. 1 is a schematic diagram of an application scenario of a vibration signal identification method provided by an embodiment of the present application.
  • a vibration detection sensor is provided on a specific component of the equipment to be detected 101 .
  • the vibration detection sensor is used to detect the vibration signal of the specific component of the equipment to be detected 101 in real time, and send the detected vibration signal to the server 102 .
  • the server 102 is used to process the vibration signal, extract acoustic features therefrom, and input the acoustic features into the trained neural network model, thereby obtaining the recognition result output by the neural network model. Subsequently, the server 102 can send the identification result to the user's terminal device 103 to inform the user of the occurrence probability of each fault type.
  • the equipment to be detected 101 can be any type of mechanical equipment, such as a crane, a tractor, a hydraulic press, etc. wait.
  • the server 102 may be, but is not limited to, a single network server, a server group composed of multiple network servers, or a cloud composed of a large number of computers or network servers based on cloud computing.
  • cloud computing is a type of distributed computing, a super virtual computer composed of a group of loosely coupled computers.
  • the terminal device 103 may be a tablet computer (pad), a virtual reality (VR) terminal device, an augmented reality (AR) terminal device, a wireless terminal in self-driving, or remote medical surgery. Wireless terminals in surgery, wireless terminals in smart grids, wireless terminals in smart homes, etc.
  • VR virtual reality
  • AR augmented reality
  • vibration signal identification method can be implemented by the vibration signal identification device provided in the embodiment of the present application.
  • the vibration signal identification device can be part or all of a certain device, such as the above-mentioned server.
  • FIG. 2 is a schematic flowchart of a method for identifying vibration signals provided by an embodiment of the present application.
  • the embodiment of the present application relates to a specific process of how to identify vibration signals.
  • the identification method of the vibration signal includes:
  • the embodiments of the present application do not limit the type of equipment to be detected.
  • the equipment to be detected can be any type of mechanical equipment, such as a crane, a tractor, a hydraulic press, etc.
  • the embodiment of the present application does not limit the type of the target component, which may be a rotating component, a connecting component, etc.
  • a vibration sensor may be installed on the target component of the device to be detected, thereby collecting vibration signals of the target component of the device to be detected in real time. At the same time, the vibration sensor can send the vibration signal of the target component of the device to be detected to the server, so that the server can identify the vibration signal.
  • the vibration signal can be divided into multiple sub-signals in order to obtain the sub-signals of the sample length to be set.
  • the embodiment of the present application does not limit the set sample length, and the setting can be determined according to the sampling frequency of the vibration signal.
  • the vibration signal in the 1-20kHz sampling frequency range 2048 or 3072 can be used as the standard sample. length.
  • the vibration signal by dividing the vibration signal into multiple sub-signals with a set sample length, the problems of data redundancy and insufficient number of sample structures caused by too long sample lengths can be avoided. At the same time, it can also avoid the problem of insufficient information for sufficient feature extraction due to too short sample length.
  • the segmented multi-segment sub-signals may also be standardized, thereby preventing the original data from drifting and causing data instability.
  • the embodiments of the present application do not limit the standardization method.
  • the standard score (z-score) algorithm or the maximum and minimum (minmax) algorithm can be used for data standardization.
  • the server divides the vibration signal into multiple sub-signals
  • the acoustic characteristics of the multiple sub-signals can be determined respectively.
  • the acoustic features are two-dimensional features, the first dimension of the acoustic features is used to characterize the frame number of the sub-signal, and the second dimension of the acoustic features is used to characterize the Mel spectrum cepstrum coefficient of each frame signal.
  • the Mel spectrum cepstral coefficient is a feature determined through Mel frequency conversion and cepstrum analysis.
  • the Mel frequency describes the nonlinear characteristics of the human ear frequency, and the relationship between it and the frequency can be identified using formula (1):
  • Mel(f) is the Mel frequency and f is the frequency.
  • the server may first use discrete Fourier transform to convert the multi-segment sub-signals from time domain signals to frequency domain signals. Secondly, the server can separately determine the power spectral rate of the frequency domain signals corresponding to the multiple sub-signals. Thirdly, the server can use a triangular filter to perform Mel filtering on the power spectral rate of the frequency domain signal corresponding to the multi-segment sub-signal, and determine the logarithmic energy corresponding to the multi-segment sub-signal respectively. Finally, the server can use discrete Fourier transform to convert the logarithmic energy corresponding to the multi-segment sub-signal into the Mel spectrum cepstral coefficient of the multi-segment sub-signal.
  • the server may also perform framing and windowing processing on the multi-segment sub-signals. For example, the server can first perform frame processing on each sub-signal according to the preset number of signal sampling points to obtain multi-frame signals corresponding to each sub-signal. Subsequently, the multi-frame signals corresponding to each sub-signal are windowed separately.
  • the neural network model can be a one-dimensional dual convolutional neural network model, including a first convolution layer and a second convolution layer with the same structure.
  • Convolution layer, the first convolution layer and the second convolution layer each include two convolution units, and the convolution parameters between each convolution unit are different.
  • the first convolution layer is provided before the second convolution layer
  • the first pooling layer is provided between the first convolution layer and the second convolution layer
  • the third convolution layer is provided after the second convolution layer.
  • Second pool level After the second pooling layer, three layers of fully connected layers and normalized exponential function (Softmax) classification can also be set to obtain the probability value of each fault category.
  • Softmax normalized exponential function
  • the first pooling layer is used to perform maximum pooling on the features extracted by the first convolutional layer
  • the second pooling layer is used to perform adaptive average pooling on the features extracted by the second convolutional layer.
  • the embodiments of the present application provide a method and device for identifying vibration signals.
  • the vibration signals of the target components of the equipment to be detected are obtained.
  • the vibration signal is divided into multiple sub-signals according to the set sample length. again, Determine the acoustic characteristics of multiple sub-signals respectively.
  • the acoustic features of the multi-segment sub-signals are input into the neural network model, and the recognition results output by the neural network model are obtained.
  • the recognition results are used to characterize the occurrence probability of each fault type of the target component, and the neural network model is used to classify the acoustic features. Secondary feature learning.
  • the acoustic features are extracted as the features of the vibration signal, and the acoustic features are identified through the neural network model of secondary feature learning to determine the probability of occurrence of each fault type of the target component. Since the neural network model is more suitable for temporal feature learning, it can better extract high-order features from acoustic features, with higher robustness and generalization, thereby improving the accuracy of fault type identification.
  • FIG 3 is a schematic flow chart of another vibration signal identification method provided by an embodiment of the present application. As shown in Figure 3, the vibration signal identification method includes:
  • a set of several consecutive sampling points in a sub-signal can be used as an observation unit, called a frame.
  • the frame is usually an exponential multiple of 2, such as 256.
  • the embodiment of the present application does not limit the number of preset sampling points.
  • a set of 256 sampling points can be used as one frame.
  • the embodiment of the present application does not limit the method of adding windows.
  • it can be a Hamming window.
  • This application adds windows to each frame of signal to smooth the signal and reduce the side lobe size and spectrum leakage after Fourier transform.
  • S305 Use discrete Fourier transform to convert the multi-segment sub-signals from time domain signals to frequency domain signals.
  • formula (2) can be used to convert the time domain signal of each frame into the frequency domain signal Si (k):
  • N is the length of the vibration signal
  • k is the period of the vibration signal
  • K is the maximum period, 1 ⁇ k ⁇ K.
  • time domain signals into frequency domain signals
  • signal characteristics can be better reflected, which is beneficial to improving subsequent identification accuracy.
  • formula (3) can be used to calculate the power spectral rate P i (k) of the frequency domain signal:
  • N is the length of the vibration signal
  • Si (k) is the frequency domain signal.
  • formula (4) can be used to determine the frequency response H m (k) of the triangular filter:
  • m is the sequence number of the triangular filter
  • f(m) is the center frequency of the m-th triangular filter
  • k is the period of the vibration signal.
  • formula (5) can be used to calculate the logarithmic energy output by each filter bank:
  • m is the sequence number of the triangular filter
  • M is the number of triangular filters
  • k is the period of the vibration signal
  • K is the maximum period.
  • formula (6) can be used to calculate the Mel spectrum cepstral coefficient C(i):
  • m is the sequence number of the triangular filter
  • M is the number of triangular filters
  • i is the order of the Mel spectrum cepstral coefficient
  • i 1,2,...,I
  • I is the Mel spectrum cepstral coefficient.
  • the maximum order can usually be 12-16.
  • steps S301 to S309 can be understood with reference to steps S201 to S204 shown in FIG. 2 , and repeated information will not be repeated here.
  • Figure 4 is a schematic structural diagram of a neural network model provided by an embodiment of the present application.
  • the neural network model is a one-dimensional dual convolution neural network model, including: a first convolution layer, a second convolution layer layer, first pooling layer, second pooling layer, three fully connected layers and Softmax classification.
  • the first convolution layer is arranged before the second convolution layer
  • a first pooling layer is arranged between the first convolution layer and the second convolution layer
  • a second pooling layer is arranged after the second convolution layer.
  • Softmax classification is installed after the three fully connected layers.
  • the first convolutional layer and the second convolutional layer have the same structure.
  • both the first convolution layer and the second convolution layer include two convolution units, each convolution unit includes one-dimensional convolution Conv1d, batch normalization BN and excitation Living function ReLU, the input features can be extracted separately through two convolution units.
  • the convolution parameters between each convolution unit are different.
  • the convolution parameters include the number of channels and the size of the convolution kernel.
  • the number of channels of the convolution unit of the first convolution layer is smaller than the number of channels of the convolution unit of the second convolution layer, and the convolution kernel size of the convolution unit of the first convolution layer is larger than that of the second convolution unit.
  • the convolution kernel size of the convolution unit of the second convolution layer is larger.
  • the number of channels C of the first two convolution units can be 16 and 32, and the convolution kernel size K can be 15 and 9.
  • the channel numbers C of the latter two convolution units can be 64 and 128, and the convolution kernel sizes K are 7 and 5.
  • the first pooling layer is used to perform maximum pooling processing on the features extracted by the first convolution layer.
  • the second pooling layer is used to perform adaptive average pooling processing on the features extracted by the second convolution layer. After four convolutional units, an adaptive sliding window is used with the second pooling layer to obtain several window averages.
  • the neural network model involved in this application uses a multi-channel one-dimensional convolution kernel to scan acoustic features window by window to extract high-order features. At the same time, multiple channels are used to receive different frame characteristics of the same sub-signal, thereby reflecting multiple aspects of changes in the data segment within a short period of time.
  • the embodiments of the present application provide a method and device for identifying vibration signals.
  • the vibration signals of the target components of the equipment to be detected are obtained.
  • the vibration signal is divided into multiple sub-signals according to the set sample length. Again, the acoustic characteristics of the multiple sub-signals are determined respectively.
  • the acoustic features of the multi-segment sub-signals are input into the neural network model, and the recognition results output by the neural network model are obtained.
  • the recognition results are used to characterize the occurrence probability of each fault type of the target component, and the neural network model is used to classify the acoustic features. Secondary feature learning.
  • the acoustic features are extracted as the features of the vibration signal, and the acoustic features are identified through the neural network model of secondary feature learning to determine the probability of occurrence of each fault type of the target component. Since the neural network model is more suitable for temporal feature learning, it can better extract high-order features from acoustic features, with higher robustness and generalization, thereby improving the accuracy of fault type identification.
  • the aforementioned program can be stored in a computer-readable storage medium.
  • the program When the program is executed, It includes the steps of the above method embodiment; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk and other various media that can store program codes.
  • FIG. 5 is a schematic structural diagram of a vibration signal identification device provided by an embodiment of the present application.
  • the vibration signal identification device can be implemented by software, hardware, or a combination of both, to perform the vibration signal identification method in the above embodiment.
  • the vibration signal identification device 500 includes: an acquisition module 501 , a segmentation module 502 , a determination module 503 and an identification module 504 .
  • the acquisition module 501 is used to acquire the vibration signal of the target component of the device to be detected
  • the segmentation module 502 is used to segment the vibration signal into multiple sub-signals according to the set sample length
  • Determination module 503 used to determine the acoustic characteristics of multiple sub-signals respectively;
  • the identification module 504 is used to input the acoustic characteristics of the multi-segment sub-signals into the neural network model, and obtain the identification results output by the neural network model.
  • the identification results are used to characterize the occurrence probability of each fault type of the target component, and the neural network model is used to Acoustic features are used for secondary feature learning.
  • the acoustic features are two-dimensional features, the first dimension of the acoustic features is used to characterize the frame number of the sub-signal, and the second dimension of the acoustic features is used to characterize the Mel spectrum cepstrum coefficient of each frame signal. .
  • the determination module 503 is specifically configured to use discrete Fourier transform to convert multi-segment sub-signals from time domain signals to frequency domain signals; respectively determine the power spectral ratio of the frequency domain signals corresponding to the multi-segment sub-signals. ; Use a triangular filter to perform Mel filtering on the power spectrum of the frequency domain signal corresponding to the multi-segment sub-signal, and determine the logarithmic energy corresponding to the multi-segment sub-signal respectively; use the discrete Fourier transform to separate the logarithmic energy corresponding to the multi-segment sub-signal. Convert to Mel spectrum cepstrum coefficients of multi-segment sub-signals.
  • the determination module 503 is also used to perform frame processing on each sub-signal according to the preset number of signal sampling points to obtain a multi-frame signal corresponding to each sub-signal; Multi-frame signals are windowed separately.
  • the neural network model includes a one-dimensional dual convolutional neural network model.
  • a one-dimensional double convolutional neural network model the neural network model includes a first convolution layer and a second convolution layer with the same structure, and both the first convolution layer and the second convolution layer It includes two convolution units, and the convolution parameters between each convolution unit are different.
  • the convolution parameters include the number of channels and the size of the convolution kernel.
  • the number of channels of the convolution unit of the first convolution layer is smaller than the number of channels of the convolution unit of the second convolution layer.
  • the convolution kernel size of the convolution unit of the convolution layer is larger than the convolution kernel size of the convolution unit of the second convolution layer.
  • the channel of the convolution unit is used to receive the Mel spectrum cepstrum coefficients of different frames of the same sub-signal.
  • the first convolution layer is provided before the second convolution layer, a first pooling layer is provided between the first convolution layer and the second convolution layer, and after the second convolution layer A second pooling layer is provided;
  • the first pooling layer is used to perform maximum pooling on the features extracted by the first convolutional layer
  • the second pooling layer is used to perform adaptive average pooling on the features extracted by the second convolutional layer.
  • FIG. 5 shows a vibration signal identification device provided by an embodiment, which can be used to perform the vibration signal identification method provided by any of the above embodiments.
  • the specific implementation methods and technical effects are similar, and will not be described again here.
  • FIG. 6 is a schematic structural diagram of an electronic device provided by an embodiment of the present application. As shown in FIG. 6 , the electronic device may include: multiple processors 601 and memories 602 . Figure 6 shows an electronic device using a processor as an example.
  • Memory 602 is used to store programs.
  • the program may include program code, which includes computer operating instructions.
  • the memory 602 may include high-speed RAM memory, and may also include non-volatile memory (non-volatile memory), such as multiple disk memories.
  • the processor 601 is used to execute the computer execution instructions stored in the memory 602 to implement the above vibration signal identification method
  • the processor 601 may be a processor (Central Processing Unit, referred to as CPU), or a specific integrated circuit (Application Specific Integrated Circuit, referred to as ASIC), or one or more processors configured to implement the embodiments of the present application. integrated circuit.
  • CPU Central Processing Unit
  • ASIC Application Specific Integrated Circuit
  • the communication interface, the memory 602 and the processor 601 can be connected to each other through a bus and complete communication with each other.
  • the bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc.
  • ISA Industry Standard Architecture
  • PCI Peripheral Component
  • EISA Extended Industry Standard Architecture
  • the bus can be divided into address bus, data bus, control bus, etc., but it does not mean that there is only one bus or one type of bus.
  • the communication interface, the memory 602 and the processor 601 are integrated on one chip, the communication interface, the memory 602 and the processor 601 can communicate through the internal interface.
  • An embodiment of the present application also provides a chip, including a processor and an interface.
  • the interface is used to input and output data or instructions processed by the processor.
  • the processor is configured to execute the vibration signal identification method provided in the above method embodiment.
  • This chip can be used in vibration signal identification devices.
  • the computer-readable storage medium may 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.
  • the computer-readable storage medium stores program information, and the program information is used for the above-mentioned vibration signal identification method.
  • This application also provides a computer program product, including a computer program, which implements the above vibration signal identification method when executed by a processor.
  • This application also provides a computer program, which causes the computer to execute the above-mentioned vibration signal identification method.
  • a computer program product includes one or more computer instructions.
  • Computer instructions may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, e.g., computer instructions may be transmitted from a website, computer, server or data center via a wired link (e.g.
  • Coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless means to transmit to another website site, computer, server or data center.
  • Computer-readable storage media can be any available media that can be accessed by the computer or contain one or more Media-integrated servers, data centers and other data storage devices are available. Available media may be magnetic media (eg, floppy disk, hard disk, magnetic tape), optical media (eg, DVD), or semiconductor media (eg, Solid State Disk (SSD)), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

一种振动信号的识别方法，包括：获取待检测设备的目标部件的振动信号（S201）；根据设定的样本长度，将振动信号分割为多段子信号（S202）；分别确定多段子信号的声学特征（S203）；将多段子信号的声学特征输入神经网络模型中，并获取神经网络模型输出的识别结果，识别结果用于表征目标部件的各个故障类型的发生概率，神经网络模型用于对声学特征进行二次特征学习（S204）。通过该方式，将声学特征作为振动信号的特征进行提取，并通过二次特征学习的神经网络模型对声学特征进行识别，来确定目标部件的各个故障类型的发生概率，从而提高了故障类型的识别准确性。还公开了一种振动信号的识别装置、一种电子设备和一种计算机存储介质。

Description

振动信号的识别方法及装置
本申请要求于2022年03月29日提交中国专利局、申请号为202210324019.6、申请名称为“振动信号的识别方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及故障识别技术领域，尤其涉及一种振动信号的识别方法及装置。
背景技术
工业设备中存在大量旋转部件，例如轴承、齿轮等。工业设备在运行过程中由于受到不同载荷、环境以及退化程度等影响，会产生多种不同的振动信号。相较工业设备的润滑油等内部流体的温度、压力、流量等状态参数，或者电机的转态参数，振动信号可以更加直观、快速、准确地反映工业设备的运行状态，是对工业设备进行异常检测和故障诊断的重要手段。
相关技术中，在通过振动信号对设备进行故障诊断时，可以先基于领域知识和实际工程需求对振动信号进行人工特征提取和特征筛选择，再建立机器学习的分类器模型，以筛选出的特征作为输入，需要识别的状态作为输出，实现不同振动信号的分类，从而判断设备的旋转部件的故障状态。
然而，振动信号的有效特征往往会随着不同的设备、不同的工况或不同的故障改变，从而导致基于领域知识和实际工程需求对振动信号的筛选可能会遗漏有利于分类的有效特征，进而导致故障类型的识别准确性不高。
发明内容
本申请实施例提供一种振动信号的识别方法及装置，以解决现有技术中故障类型的识别准确性不高的问题。
第一方面，本申请实施例提供一种振动信号的识别方法，所述方法包括：
获取待检测设备的目标部件的振动信号；
根据设定的样本长度，将所述振动信号分割为多段子信号；
分别确定所述多段子信号的声学特征；
将所述多段子信号的声学特征输入神经网络模型中，并获取所述神经网络模型输出的识别结果，所述识别结果用于表征所述目标部件的各个故障类型的发生概率，所 述神经网络模型用于对所述声学特征进行二次特征学习。
一种可选的实施方式中，所述声学特征为二维特征，所述声学特征的第一维用于表征所述子信号的帧数，所述声学特征的第二维用于表征每帧信号的梅尔谱倒谱系数。
一种可选的实施方式中，所述分别确定所述多段子信号的声学特征，包括：
使用离散傅里叶变换将所述多段子信号由时域信号转换为频域信号；
分别确定所述多段子信号对应的频域信号的功谱率；
使用三角滤波器对所述多段子信号对应的频域信号的功谱率进行梅尔滤波，分别确定所述多段子信号对应的对数能量；
使用离散傅里叶变换将所述多段子信号对应的对数能量分别转换为多段子信号的梅尔谱倒谱系数。
一种可选的实施方式中，在所述使用离散傅里叶变换将所述多段子信号由时域信号转换为频域信号之前，所述方法还包括：
根据预设的信号采样点数量，对每段子信号进行分帧处理，得到所述每段子信号对应的多帧信号；
将所述每段子信号对应的多帧信号分别进行加窗处理。
一种可选的实施方式中，所述神经网络模型包括一维双卷积神经网络模型。
一种可选的实施方式中，所述一维双卷积神经网络模型包括具有相同结构的第一卷积层和第二卷积层，所述第一卷积层和所述第二卷积层均包括两个卷积单元，每个卷积单元之间的卷积参数均不相同。
一种可选的实施方式中，所述卷积参数包括通道数和卷积核尺寸，所述第一卷积层的卷积单元的通道数均小于所述第二卷积层的卷积单元的通道数，所述第一卷积层的卷积单元的卷积核尺寸均大于所述第二卷积层的卷积单元的卷积核尺寸。
一种可选的实施方式中，所述卷积单元的通道用于接收同一段子信号的不同帧的梅尔谱倒谱系数。
一种可选的实施方式中，所述第一卷积层设置在所述第二卷积层之前，所述第一卷积层和第二卷积层之间设置有第一池化层，所述第二卷积层之后设置有第二池化层；
其中，所述第一池化层用于对所述第一卷积层提取到的特征进行最大池化处理，所述第二池化层用于对所述第二卷积层提取到的特征进行自适应平均池化处理。
第二方面，本申请实施例提供一种振动信号的识别装置，所述装置法包括：
获取模块，用于获取待检测设备的目标部件的振动信号；
分割模块，用于根据设定的样本长度，将所述振动信号分割为多段子信号；
确定模块，用于分别确定所述多段子信号的声学特征；
识别模块，用于将所述多段子信号的声学特征输入神经网络模型中，并获取所述神经网络模型输出的识别结果，所述识别结果用于表征所述目标部件的各个故障类型的发生概率，所述神经网络模型用于对所述声学特征进行二次特征学习。
一种可选的实施方式中，所述声学特征为二维特征，所述声学特征的第一维用于表征所述子信号的帧数，所述声学特征的第二维用于表征每帧信号的梅尔谱倒谱系数。
一种可选的实施方式中，所述确定模块，具体用于使用离散傅里叶变换将所述多段子信号由时域信号转换为频域信号；分别确定所述多段子信号对应的频域信号的功谱率；使用三角滤波器对所述多段子信号对应的频域信号的功谱率进行梅尔滤波，分别确定所述多段子信号对应的对数能量；使用离散傅里叶变换将所述多段子信号对应的对数能量分别转换为多段子信号的梅尔谱倒谱系数。
一种可选的实施方式中，所述确定模块，还用于根据预设的信号采样点数量，对每段子信号进行分帧处理，得到所述每段子信号对应的多帧信号；将所述每段子信号对应的多帧信号分别进行加窗处理。
一种可选的实施方式中，所述神经网络模型包括一维双卷积神经网络模型。
一种可选的实施方式中，所述一维双卷积神经网络模型，所述神经网络模型包括具有相同结构的第一卷积层和第二卷积层，所述第一卷积层和所述第二卷积层均包括两个卷积单元，每个卷积单元之间的卷积参数均不相同。
一种可选的实施方式中，所述卷积参数包括通道数和卷积核尺寸，所述第一卷积层的卷积单元的通道数均小于所述第二卷积层的卷积单元的通道数，所述第一卷积层的卷积单元的卷积核尺寸均大于所述第二卷积层的卷积单元的卷积核尺寸。
一种可选的实施方式中，所述卷积单元的通道用于接收同一段子信号的不同帧的梅尔谱倒谱系数。
一种可选的实施方式中，所述第一卷积层设置在所述第二卷积层之前，所述第一卷积层和第二卷积层之间设置有第一池化层，所述第二卷积层之后设置有第二池化层；
其中，所述第一池化层用于对所述第一卷积层提取到的特征进行最大池化处理，所述第二池化层用于对所述第二卷积层提取到的特征进行自适应平均池化处理。
第三方面，本申请还提供一种电子设备，包括：处理器，以及存储器；所述存储器用于存储所述处理器的计算机程序；所述处理器被配置为通过执行所述计算机程序来实现第一方面中任意一种可能的方法。
第四方面，本发明还提供一种计算机存储介质，所述计算机存储介质存储有多条指令，所述指令适于由处理器加载并执行第一方面中任意一种可能的方法。
第五方面，本公开实施例提供一种计算机程序产品，包括计算机程序，该计算机程序被处理器执行时实现如上第一方面以及第一方面各种可能的设计中所述的方法。
本申请实施例提供的一种振动信号的识别方法及装置，首先获取待检测设备的目标部件的振动信号。其次，根据设定的样本长度，将振动信号分割为多段子信号。再次，分别确定多段子信号的声学特征。最后，将多段子信号的声学特征输入神经网络模型中，并获取神经网络模型输出的识别结果，识别结果用于表征目标部件的各个故障类型的发生概率，神经网络模型用于对声学特征进行二次特征学习。通过该方式，将声 学特征作为振动信号的特征进行提取，并通过二次特征学习的神经网络模型对声学特征进行识别，来确定目标部件的各个故障类型的发生概率，从而提高了故障类型的识别准确性。
附图说明
为了更清楚地说明本发明或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作一简单地介绍，显而易见地，下面描述中的附图是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1为本申请实施例提供的一种振动信号的识别方法的应用场景示意图；
图2为本申请实施例提供的一种振动信号的识别方法的流程示意图；
图3为本申请实施例提供的另一种振动信号的识别方法的流程示意图；
图4为本申请实施例提供的一种神经网络模型的结构示意图；
图5为本申请实施例提供的一种振动信号的识别装置的结构示意图；
图6为本申请实施例提供的一种电子设备的结构示意图。
具体实施方式
为使本申请实施例的目的、技术方案和优点更加清楚，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，显然，所描述的实施例是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有作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。
应当理解，本公开的方法实施方式中记载的各个步骤可以按照不同的顺序执行，和/或并行执行。此外，方法实施方式可以包括附加的步骤和/或省略执行示出的步骤。本公开的范围在此方面不受限制。
本文使用的术语“包括”及其变形是开放性包括，即“包括但不限于”。术语“基于”是“至少部分地基于”。术语“一个实施例”表示“至少一个实施例”；术语“另一实施例”表示“至少一个另外的实施例”；术语“一些实施例”表示“至少一些实施例”。其他术语的相关定义将在下文描述中给出。
需要注意，本公开中提及的“一个”、“多个”的修饰是示意性而非限制性的，本领域技术人员应当理解，除非在上下文另有明确指出，否则应该理解为“一个或多个”。
工业设备中存在大量目标部件，例如轴承、齿轮等。工业设备在运行过程中由于受到不同载荷、环境以及退化程度等影响，会产生多种不同的振动信号。振动信号通常包括两类性质不同的振源：一类振源是由于机械运动件的质量不平衡、几何轴线不对中、齿轮捏合差、传动件配合失当、轴颈轴承间隙过大等引起的机械强迫振动，例如，周期振动、冲击振动、随机振动等；另一类振源是由于结构响应、自激励振动或环境振动引起的振动响 应，比如：流体的喘激振动、轴承的油膜振动、部件本身的响应振动、结构的局部振动等。此外，由于外部的载荷变化，振动响应也会发生变化。
对振动信号进行监测和识别可以洞悉设备的不同工况以及健康状态，对提高设备的稳定运行起到很重要作用。通过对振动信号的处理与分析，可以及时发现设备早期的故障征兆，从而预测设备可能的故障，为预防事故、科学安排检修提供科学的依据，节约维修成本，提高设备的可靠性和安全性。
因此，相较工业设备的润滑油等内部流体的温度、压力、流量等状态参数，或者电机的转态参数，振动信号可以更加直观、快速、准确地反映工业设备的运行状态，是对工业设备进行异常检测和故障诊断的重要手段。
相关技术中，存在两种通过振动信号对设备进行故障诊断的方式。
在第一种方式中，可以先基于领域知识和实际工程需求对振动信号进行人工特征提取和特征筛选择，再建立机器学习的分类器模型，以筛选出的特征作为输入，需要识别的状态作为输出，实现不同振动信号的分类，从而判断设备的目标部件的故障状态。
然而，振动信号的有效特征往往会随着不同的设备、不同的工况或不同的故障改变，从而导致基于领域知识和实际工程需求对振动信号的筛选可能会遗漏有利于分类的有效特征，从而导致故障类型的识别准确性不高。
在第二种方式中，可以通过一个多层网络结构进行深度学习，该多层网络的前若干层可以进行机器自主的特征提起，每一层都能获得输入数据的不同表征，最后一层实现状态分类，从而确定设备的目标部件的故障状态。
然而，多层网络结构的深度学习的适配性较差，往往仅能适用于某个特定工况下的场景，并且，随着网络层数的增多和各种网络结构的发展，通常也无法为每层网络都设定合适的结构参数，从而同样导致故障类型的识别准确性不高。
为解决上述问题，本申请实施例提供一种振动信号的识别方法与装置，通过从振动信号中提取出声学特征，再将声学特征输入二次特征学习的神经网络模型，从而得到神经网络模型输出的识别结果，来确定目标部件的各个故障类型的发生概率。由于神经网络模型更适合时序特征学习，从而可以更好地从声学特征中提取高阶特征，鲁棒性和泛化性更高，从而提高了故障类型的识别准确性。
下面对本申请实施例涉及的振动信号的识别方法的应用场景进行说明。
图1为本申请实施例提供的一种振动信号的识别方法的应用场景示意图。如图1所示，待检测设备101的特定部件上设置有振动检测传感器，该振动检测传感器用于实时检测待检测设备101的特定部件的振动信号，并将检测到的振动信号发送给服务器102。服务器102用于对振动信号进行处理，从中提取出声学特征，并将声学特征输入训练好的神经网络模型中，从而得到神经网络模型输出的识别结果。随后，服务器102可以将识别结果发送到用户的终端设备103上，以告知用户各个故障类型的发生概率。
其中，待检测设备101可以为任意类型的机械设备，例如，起重机、拖拉机、液压机 等。
服务器102可以但不限于单个网络服务器、多个网络服务器组成的服务器组或基于云计算的由大量计算机或网络服务器构成的云。其中，云计算是分布式计算的一种，由一群松散耦合的计算机组成的一个超级虚拟计算机。
终端设备103可以为平板电脑(pad)、虚拟现实(virtual reality，VR)终端设备、增强现实(augmented reality，AR)终端设备、无人驾驶(self driving)中的无线终端、远程手术(remote medical surgery)中的无线终端、智能电网(smart grid)中的无线终端、智慧家庭(smart home)中的无线终端等。
可以理解，上述振动信号的识别方法可以通过本申请实施例提供的振动信号的识别装置实现，振动信号的识别装置可以是某个设备的部分或全部，例如为上述服务器。
下面以集成或安装有相关执行代码的服务器为例，以具体地实施例对本申请实施例的技术方案进行详细说明。下面这几个具体的实施例可以相互结合，对于相同或相似的概念或过程可能在某些实施例不再赘述。
图2为本申请实施例提供的一种振动信号的识别方法的流程示意图，本申请实施例涉及的是如何识别振动信号的具体过程。如图2所示，该振动信号的识别方法包括：
S201、获取待检测设备的目标部件的振动信号。
应理解，本申请实施例对于待检测设备的类型不作限制，在一些实施例中，待检测设备可以为任意类型的机械设备，例如，起重机、拖拉机、液压机等。相应的，本申请实施例对于目标部件的类型也不作限制，可以为旋转部件、连接部件等。
在一些实施例中，待检测设备的目标部件上可以安装有振动传感器，从而实时收集待检测设备的目标部件的振动信号。同时，振动传感器可以将待检测设备的目标部件的振动信号发送给服务器，以便服务器进行振动信号的识别。
S202、根据设定的样本长度，将振动信号分割为多段子信号。
在本步骤中，当服务器获取到待检测设备的目标部件的振动信号后，可以将振动信号分割为多段子信号，以便获取待设定的样本长度的子信号。
应理解，本申请实施例对于设定的样本长度不作限制，可以根据振动信号的采样频率确定设置，示例性的，对于1-20kHz采样频率范围内的振动信号，可以使用2048或者3072作为标准样本长度。
在本申请中，通过将振动信号分割为设定的样本长度的多段子信号，可以避免因样本长度过长造成的数据冗余和样本构造数量不足的问题。同时，也可以避免因样本长度过短造成的信息量不够进行充分的特征提取的问题。
在另一些实施例中，在对振动信号进行分割后，还可以对分割出的多段子信号进行标准化，从而防止原始数据发生零飘而导致数据的不稳定。
应理解，本申请实施例对于标准化的方式不作限制，示例性的，可以采用标准分数(z-score)算法或者极大极小(minmax)算法的方式进行数据标准化。
S203、分别确定多段子信号的声学特征。
在本步骤中，当服务器获将振动信号分割为多段子信号，可以分别确定多段子信号的声学特征。
其中，声学特征为二维特征，声学特征的第一维用于表征子信号的帧数，声学特征的第二维用于表征每帧信号的梅尔谱倒谱系数。
需要说明的是，梅尔谱倒谱系数为经过梅尔频率转换和倒谱分析确定出的特征。
其中，梅尔频率描述了人耳频率的非线性特征，可以采用公式(1)标识与频率的关系：
其中，Mel(f)为梅尔频率，f为频率。
应理解，本申请实施例对于如何分别确定多段子信号的声学特征不作限制，在一些实施例中，服务器可以首先使用离散傅里叶变换将多段子信号由时域信号转换为频域信号。其次，服务器可以分别确定多段子信号对应的频域信号的功谱率。再次，服务器可以使用三角滤波器对多段子信号对应的频域信号的功谱率进行梅尔滤波，分别确定多段子信号对应的对数能量。最后，服务器可以使用离散傅里叶变换将多段子信号对应的对数能量分别转换为多段子信号的梅尔谱倒谱系数。
在一些实施例中，对多段子信号进行频域转换和进行倒谱分析之前，服务器还可以对多段子信号进行分帧和加窗处理。示例性的，服务器可以先根据预设的信号采样点数量，对每段子信号进行分帧处理，得到每段子信号对应的多帧信号。随后，将每段子信号对应的多帧信号分别进行加窗处理。
S204、将多段子信号的声学特征输入神经网络模型中，并获取神经网络模型输出的识别结果，识别结果用于表征目标部件的各个故障类型的发生概率，神经网络模型用于对声学特征进行二次特征学习。
应理解，本申请实施例对于神经网络模型的结构不作限制，在一些实施例中，神经网络模型可以为一维双卷积神经网络模型，包括具有相同结构的第一卷积层和第二卷积层，第一卷积层和第二卷积层均包括两个卷积单元，每个卷积单元之间的卷积参数均不相同。
此外，下面对于神经网络模型的各个层之间的顺序进行说明。在一些实施例中，第一卷积层设置在第二卷积层之前，第一卷积层和第二卷积层之间设置有第一池化层，第二卷积层之后设置有第二池化层。在第二池化层之后，还可以设置三层全联接层和归一化指数函数(Softmax)分类，以获得各个故障类别的概率值。
其中，第一池化层用于对第一卷积层提取到的特征进行最大池化处理，第二池化层用于对第二卷积层提取到的特征进行自适应平均池化处理。
本申请实施例提供的一种振动信号的识别方法及装置，首先获取待检测设备的目标部件的振动信号。其次，根据设定的样本长度，将振动信号分割为多段子信号。再次， 分别确定多段子信号的声学特征。最后，将多段子信号的声学特征输入神经网络模型中，并获取神经网络模型输出的识别结果，识别结果用于表征目标部件的各个故障类型的发生概率，神经网络模型用于对声学特征进行二次特征学习。通过该方式，将声学特征作为振动信号的特征进行提取，并通过二次特征学习的神经网络模型对声学特征进行识别，来确定目标部件的各个故障类型的发生概率。由于神经网络模型更适合时序特征学习，从而可以更好地从声学特征中提取高阶特征，鲁棒性和泛化性更高，从而提高了故障类型的识别准确性。
在上述实施例的基础上，下面对于如何从振动信号中提取梅尔谱倒谱系数进行说明。图3为本申请实施例提供的另一种振动信号的识别方法的流程示意图，如图3所示，该振动信号的识别方法包括：
S301、获取待检测设备的目标部件的振动信号。
S302、根据设定的样本长度，将振动信号分割为多段子信号。
S303、根据预设的信号采样点数量，对每段子信号进行分帧处理，得到每段子信号对应的多帧信号。
在申请中，一个子信号中若干个连续采样点集合可以作为一个观测单位，称为帧。需要说明的是，帧通常为2的指数倍，如256。
应理解，本申请实施例对于预设的采样点数量不作限制，示例性的，可以将256个采样点集合作为一个帧。
S304、将每段子信号对应的多帧信号分别进行加窗处理。
应理解，本申请实施例对于加窗的方式不作限制，示例性的，可以为汉明窗。
本申请通过对每一帧信号进行加窗，从而平滑信号，减弱傅立叶变换后旁瓣大小及频谱泄露。
S305、使用离散傅里叶变换将多段子信号由时域信号转换为频域信号。
示例性的，可以采用公式(2)将每帧时域信号转化为频域信号Si(k)：
其中，N为振动信号的长度，k为振动信号的周期，K为最大周期，1≤k≤K。
应理解，通过将时域信号转换为频域信号，可以更好的反映信号特征，有利于提高后续的识别准确性。
S306、分别确定多段子信号对应的频域信号的功谱率。
示例性的，可以采用公式(3)计算频域信号的功谱率Pi(k)：
其中，N为振动信号的长度，Si(k)为频域信号。
S307、使用三角滤波器对多段子信号对应的频域信号的功谱率进行梅尔滤波，分别确定多段子信号对应的对数能量。
应理解，由于频域信号有较多冗余，使用一组梅尔尺度的三角滤波器组对获得的功率谱进行平滑并消除谐波作用。其中，三角滤波器的数量可以根据实际情况具体设置，例如，20-40。
示例性的，可以采用公式(4)确定三角滤波器的频率响应Hm(k)：
其中，m为三角滤波器的序号，f(m)为第m个三角滤波器的中心频率，k为振动信号的周期。
示例性的，可以采用公式(5)计算每个滤波器组输出的对数能量：
其中，m为三角滤波器的序号，M为三角滤波器的个数，k为振动信号的周期，K为最大周期。
S308、使用离散傅里叶变换将多段子信号对应的对数能量分别转换为多段子信号的梅尔谱倒谱系数。
示例性的，可以采用公式(6)计算梅尔谱倒谱系数C(i)：
其中，m为三角滤波器的序号，M为三角滤波器的个数，i为梅尔谱倒谱系数的阶数，i＝1,2,…,I，I为梅尔谱倒谱系数的最大阶数，通常可以取12-16。
S309、将多段子信号的梅尔谱倒谱系数输入神经网络模型中，并获取神经网络模型输出的识别结果，识别结果用于表征目标部件的各个故障类型的发生概率，神经网络模型用于对声学特征进行二次特征学习。
步骤S301至步骤S309的技术名词、技术效果、技术特征，以及可选实施方式，可参照图2所示的步骤S201至S204理解，对于重复的信息，在此不再累述。
在上述实施例的基础上，下面对于神经网络模型进行说明。图4为本申请实施例提供的一种神经网络模型的结构示意图，如图4所示，该神经网络模型为一维双卷积神经网络模型，包括：第一卷积层、第二卷积层、第一池化层、第二池化层、三层全联接层和Softmax分类。
其中，第一卷积层设置在第二卷积层之前，第一卷积层和第二卷积层之间设置有第一池化层，第二卷积层之后设置有第二池化层，第二池化层之后设置有三层全联接层，三层全联接层之后设置有Softmax分类。
其中，第一卷积层和第二卷积层具有相同的结构。示例性的，第一卷积层和第二卷积层均包含两个卷积单元，每个卷积单元包括一维卷积Conv1d、批标准化BN和激 活函数ReLU，输入的特征可以通过两个卷积单元分别进行特征提取。
需要说明的是，每个卷积单元之间的卷积参数均不相同。其中，卷积参数包括通道数和卷积核尺寸。在一些实施例中，第一卷积层的卷积单元的通道数均小于第二卷积层的卷积单元的通道数，第一卷积层的卷积单元的卷积核尺寸均大于第二卷积层的卷积单元的卷积核尺寸。
应理解，在前两个卷积单元中设置较大的卷积核，有利于从较长的特征点中捕捉大范围有效特征。在后两个卷积单元中设置较小的卷积核，有利于进一步提取高级特征。
示例性的，前两个卷积单元的通道数量C可以为16和32，卷积核尺寸K为15和9。相应的，后两个卷积单元的通道数量C可以为64和128，卷积核尺寸K为7和5。
此外，第一池化层用于对第一卷积层提取到的特征进行最大池化处理。第二池化层用于对第二卷积层提取到的特征进行自适应平均池化处理。在四个卷积单元后，同第二池化层使用自适应滑窗获得若干窗口平均值。
最后，通过三层全联接层和Softmax分类，可以获得各个故障类别的概率值。
本申请中涉及的神经网络模型，使用多通道一维卷积核对声学特征逐窗口扫描提取高阶特征。同时，多通道用于接收同一子信号的不同帧特征，从而可以反映该数据片段短时间内变化的多个方面。
本申请实施例提供的一种振动信号的识别方法及装置，首先获取待检测设备的目标部件的振动信号。其次，根据设定的样本长度，将振动信号分割为多段子信号。再次，分别确定多段子信号的声学特征。最后，将多段子信号的声学特征输入神经网络模型中，并获取神经网络模型输出的识别结果，识别结果用于表征目标部件的各个故障类型的发生概率，神经网络模型用于对声学特征进行二次特征学习。通过该方式，将声学特征作为振动信号的特征进行提取，并通过二次特征学习的神经网络模型对声学特征进行识别，来确定目标部件的各个故障类型的发生概率。由于神经网络模型更适合时序特征学习，从而可以更好地从声学特征中提取高阶特征，鲁棒性和泛化性更高，从而提高了故障类型的识别准确性。
本领域普通技术人员可以理解：实现上述方法实施例的全部或部分步骤可以通过程序指令相关的硬件来完成，前述的程序可以存储于一计算机可读取存储介质中，该程序在执行时，执行包括上述方法实施例的步骤；而前述的存储介质包括：ROM、RAM、磁碟或者光盘等各种可以存储程序代码的介质。
图5为本申请实施例提供的一种振动信号的识别装置的结构示意图。该振动信号的识别装置可以通过软件、硬件或者两者的结合实现，以执行上述实施例中振动信号的识别方法。如图5所示，该振动信号的识别装置500包括：获取模块501、分割模块502、确定模块503和识别模块504。
获取模块501，用于获取待检测设备的目标部件的振动信号；
分割模块502，用于根据设定的样本长度，将振动信号分割为多段子信号；
确定模块503，用于分别确定多段子信号的声学特征；
识别模块504，用于将多段子信号的声学特征输入神经网络模型中，并获取神经网络模型输出的识别结果，识别结果用于表征目标部件的各个故障类型的发生概率，神经网络模型用于对声学特征进行二次特征学习。
一种可选的实施方式中，声学特征为二维特征，声学特征的第一维用于表征子信号的帧数，声学特征的第二维用于表征每帧信号的梅尔谱倒谱系数。
一种可选的实施方式中，确定模块503，具体用于使用离散傅里叶变换将多段子信号由时域信号转换为频域信号；分别确定多段子信号对应的频域信号的功谱率；使用三角滤波器对多段子信号对应的频域信号的功谱率进行梅尔滤波，分别确定多段子信号对应的对数能量；使用离散傅里叶变换将多段子信号对应的对数能量分别转换为多段子信号的梅尔谱倒谱系数。
一种可选的实施方式中，确定模块503，还用于根据预设的信号采样点数量，对每段子信号进行分帧处理，得到每段子信号对应的多帧信号；将每段子信号对应的多帧信号分别进行加窗处理。
一种可选的实施方式中，神经网络模型包括一维双卷积神经网络模型。
一种可选的实施方式中，一维双卷积神经网络模型，神经网络模型包括具有相同结构的第一卷积层和第二卷积层，第一卷积层和第二卷积层均包括两个卷积单元，每个卷积单元之间的卷积参数均不相同。
一种可选的实施方式中，卷积参数包括通道数和卷积核尺寸，第一卷积层的卷积单元的通道数均小于第二卷积层的卷积单元的通道数，第一卷积层的卷积单元的卷积核尺寸均大于第二卷积层的卷积单元的卷积核尺寸。
一种可选的实施方式中，卷积单元的通道用于接收同一段子信号的不同帧的梅尔谱倒谱系数。
一种可选的实施方式中，第一卷积层设置在第二卷积层之前，第一卷积层和第二卷积层之间设置有第一池化层，第二卷积层之后设置有第二池化层；
其中，第一池化层用于对第一卷积层提取到的特征进行最大池化处理，第二池化层用于对第二卷积层提取到的特征进行自适应平均池化处理。
需要说明的，图5示实施例提供的振动信号的识别装置，可用于执行上述任意实施例所提供的振动信号的识别方法，具体实现方式和技术效果类似，这里不再进行赘述。
图6为本申请实施例提供的一种电子设备的结构示意图。如图6示，该电子设备可以包括：多个处理器601和存储器602。图6的是以一个处理器为例的电子设备。
存储器602，用于存放程序。具体地，程序可以包括程序代码，程序代码包括计算机操作指令。
存储器602可能包含高速RAM存储器，也可能还包括非易失性存储器(non-volatile memory)，例如多个磁盘存储器。
处理器601用于执行存储器602存储的计算机执行指令，以实现上述振动信号的识别方法；
其中，处理器601可能是一个处理器(Central Processing Unit，简称为CPU)，或者是特定集成电路(Application Specific Integrated Circuit，简称为ASIC)，或者是被配置成实施本申请实施例的一个或多个集成电路。
可选的，在具体实现上，如果通信接口、存储器602和处理器601独立实现，则通信接口、存储器602和处理器601可以通过总线相互连接并完成相互间的通信。总线可以是工业标准体系结构(Industry Standard Architecture，简称为ISA)总线、外部设备互连(Peripheral Component，简称为PCI)总线或扩展工业标准体系结构(Extended Industry Standard Architecture，简称为EISA)总线等。总线可以分为地址总线、数据总线、控制总线等，但并不表示仅有一根总线或一种类型的总线。
可选的，在具体实现上，如果通信接口、存储器602和处理器601集成在一块芯片上实现，则通信接口、存储器602和处理器601可以通过内部接口完成通信。
本申请实施例还提供了一种芯片，包括处理器和接口。其中接口用于输入输出处理器所处理的数据或指令。处理器用于执行以上方法实施例中提供的振动信号的识别方法。该芯片可以应用于振动信号的识别装置中。
本申请还提供了一种计算机可读存储介质，该计算机可读存储介质可以包括：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)、磁盘或者光盘等各种可以存储程序代码的介质，具体的，该计算机可读存储介质中存储有程序信息，程序信息用于上述振动信号的识别方法。
本申请还提供了一种计算机程序产品，包括计算机程序，该计算机程序被处理器执行时实现如上述的振动信号的识别方法。
本申请还提供了一种计算机程序，计算机程序使得计算机执行上述的振动信号的识别方法。
在上述实施例中，可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行计算机程序指令时，全部或部分地产生按照本发明实施例的流程或功能。计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，计算机指令可以从一个网站站点、计算机、服务器或数据中心通过有线(例如同轴电缆、光纤、数字用户线(DSL))或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务器或数据中心进行传输。计算机可读存储介质可以是计算机能够存取的任何可用介质或者是包含一个或多个 可用介质集成的服务器、数据中心等数据存储设备。可用介质可以是磁性介质，(例如，软盘、硬盘、磁带)、光介质(例如，DVD)、或者半导体介质(例如固态硬盘Solid State Disk(SSD))等。
最后应说明的是：以上各实施例仅用以说明本发明的技术方案，而非对其限制；尽管参照前述各实施例对本发明进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分或者全部技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本发明各实施例技术方案的范围。

Claims (12)

  1. 一种振动信号的识别方法，其特征在于，所述方法包括：
    获取待检测设备的目标部件的振动信号；
    根据设定的样本长度，将所述振动信号分割为多段子信号；
    分别确定所述多段子信号的声学特征；
    将所述多段子信号的声学特征输入神经网络模型中，并获取所述神经网络模型输出的识别结果，所述识别结果用于表征所述目标部件的各个故障类型的发生概率，所述神经网络模型用于对所述声学特征进行二次特征学习。
  2. 根据权利要求1所述的方法，其特征在于，所述声学特征为二维特征，所述声学特征的第一维用于表征所述子信号的帧数，所述声学特征的第二维用于表征每帧信号的梅尔谱倒谱系数。
  3. 根据权利要求2所述的方法，其特征在于，所述分别确定所述多段子信号的声学特征，包括：
    使用离散傅里叶变换将所述多段子信号由时域信号转换为频域信号；
    分别确定所述多段子信号对应的频域信号的功谱率；
    使用三角滤波器对所述多段子信号对应的频域信号的功谱率进行梅尔滤波，分别确定所述多段子信号对应的对数能量；
    使用离散傅里叶变换将所述多段子信号对应的对数能量分别转换为多段子信号的梅尔谱倒谱系数。
  4. 根据权利要求3所述的方法，其特征在于，在所述使用离散傅里叶变换将所述多段子信号由时域信号转换为频域信号之前，所述方法还包括：
    根据预设的信号采样点数量，对每段子信号进行分帧处理，得到所述每段子信号对应的多帧信号；
    将所述每段子信号对应的多帧信号分别进行加窗处理。
  5. 根据权利要求1所述的方法，其特征在于，所述神经网络模型包括一维双卷积神经网络模型。
  6. 根据权利要求5所述的方法，其特征在于，所述一维双卷积神经网络模型包括具有相同结构的第一卷积层和第二卷积层，所述第一卷积层和所述第二卷积层均包括两个卷积单元，每个卷积单元之间的卷积参数均不相同。
  7. 根据权利要求6所述的方法，其特征在于，所述卷积参数包括通道数和卷积核尺寸，所述第一卷积层的卷积单元的通道数均小于所述第二卷积层的卷积单元的通道数，所述第一卷积层的卷积单元的卷积核尺寸均大于所述第二卷积层的卷积单元的卷积核尺寸。
  8. 根据权利要求7所述的方法，其特征在于，所述卷积单元的通道用于接收同一段子信号的不同帧的梅尔谱倒谱系数。
  9. 根据权利要求6-8任一项所述的方法，其特征在于，所述第一卷积层设置在所述第二卷积层之前，所述第一卷积层和第二卷积层之间设置有第一池化层，所述第二卷积层之后设置有第二池化层；
    其中，所述第一池化层用于对所述第一卷积层提取到的特征进行最大池化处理，所述第二池化层用于对所述第二卷积层提取到的特征进行自适应平均池化处理。
  10. 一种振动信号的识别装置，其特征在于，所述装置包括：
    获取模块，用于获取待检测设备的目标部件的振动信号；
    分割模块，用于根据设定的样本长度，将所述振动信号分割为多段子信号；
    确定模块，用于分别确定所述多段子信号的声学特征；
    识别模块，用于将所述多段子信号的声学特征输入神经网络模型中，并获取所述神经网络模型输出的识别结果，所述识别结果用于表征所述目标部件的各个故障类型的发生概率，所述神经网络模型用于对所述声学特征进行二次特征学习。
  11. 一种电子设备，其特征在于，包括：至少一个处理器和存储器；
    所述存储器存储计算机执行指令；
    所述至少一个处理器执行所述存储器存储的计算机执行指令，使得所述至少一个处理器执行如权利要求1至9任一项所述的方法。
  12. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质中存储有计算机执行指令，当处理器执行所述计算机执行指令时，实现如权利要求1至9任一项所述的方法。
PCT/CN2023/084285 2022-03-29 2023-03-28 振动信号的识别方法及装置 WO2023185801A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210324019.6 2022-03-29
CN202210324019.6A CN114543983A (zh) 2022-03-29 2022-03-29 振动信号的识别方法及装置

Publications (1)

Publication Number Publication Date
WO2023185801A1 true WO2023185801A1 (zh) 2023-10-05

Family

ID=81664988

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/084285 WO2023185801A1 (zh) 2022-03-29 2023-03-28 振动信号的识别方法及装置

Country Status (2)

Country Link
CN (1) CN114543983A (zh)
WO (1) WO2023185801A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114543983A (zh) * 2022-03-29 2022-05-27 阿里云计算有限公司 振动信号的识别方法及装置

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108896296A (zh) * 2018-04-18 2018-11-27 北京信息科技大学 一种基于卷积神经网络的风电齿轮箱故障诊断方法
KR101967301B1 (ko) * 2017-12-07 2019-04-09 한국생산기술연구원 학습 데이터 융합을 이용한 회전체의 고장 진단 시스템
KR20190067441A (ko) * 2017-12-07 2019-06-17 한국생산기술연구원 딥 러닝과 웨이블렛 변환을 이용한 회전체의 고장 진단 시스템
CN110867196A (zh) * 2019-12-03 2020-03-06 桂林理工大学 一种基于深度学习及声音识别的机器设备状态监测系统
CN112820321A (zh) * 2021-03-05 2021-05-18 河北雄安友平科技有限公司 一种抽油机远程智能音频诊断系统、方法、设备及介质
CN113125135A (zh) * 2021-03-31 2021-07-16 中石化石油工程技术服务有限公司 旋转机械的故障诊断方法、存储介质及电子设备
CN114543983A (zh) * 2022-03-29 2022-05-27 阿里云计算有限公司 振动信号的识别方法及装置

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR102235568B1 (ko) * 2018-03-21 2021-04-05 한국과학기술원 합성곱 신경망 기반 환경음 인식 방법 및 시스템
CN112364779B (zh) * 2020-11-12 2022-10-21 中国电子科技集团公司第五十四研究所 信号处理与深-浅网络多模型融合的水声目标识别方法
CN112508901B (zh) * 2020-12-01 2024-04-05 广州大学 一种水下结构病害识别方法、系统、装置及存储介质
CN112541552B (zh) * 2020-12-16 2022-04-19 中国计量大学上虞高等研究院有限公司 Dccnn和lgbm相结合的空气处理机组故障检测与诊断方法
CN113707176B (zh) * 2021-09-02 2022-09-09 国网安徽省电力有限公司铜陵供电公司 一种基于声信号及深度学习技术的变压器故障检测方法
CN114237976A (zh) * 2021-11-02 2022-03-25 阿里巴巴（中国）有限公司 一种数据获取方法及装置

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101967301B1 (ko) * 2017-12-07 2019-04-09 한국생산기술연구원 학습 데이터 융합을 이용한 회전체의 고장 진단 시스템
KR20190067441A (ko) * 2017-12-07 2019-06-17 한국생산기술연구원 딥 러닝과 웨이블렛 변환을 이용한 회전체의 고장 진단 시스템
CN108896296A (zh) * 2018-04-18 2018-11-27 北京信息科技大学 一种基于卷积神经网络的风电齿轮箱故障诊断方法
CN110867196A (zh) * 2019-12-03 2020-03-06 桂林理工大学 一种基于深度学习及声音识别的机器设备状态监测系统
CN112820321A (zh) * 2021-03-05 2021-05-18 河北雄安友平科技有限公司 一种抽油机远程智能音频诊断系统、方法、设备及介质
CN113125135A (zh) * 2021-03-31 2021-07-16 中石化石油工程技术服务有限公司 旋转机械的故障诊断方法、存储介质及电子设备
CN114543983A (zh) * 2022-03-29 2022-05-27 阿里云计算有限公司 振动信号的识别方法及装置

Also Published As

Publication number Publication date
CN114543983A (zh) 2022-05-27

Similar Documents

Publication Publication Date Title
WO2023185801A1 (zh) 振动信号的识别方法及装置
CN111401136B (zh) 一种柱塞泵空化程度检测方法、装置及终端
CN113125135A (zh) 旋转机械的故障诊断方法、存储介质及电子设备
WO2021232320A1 (zh) 超声图像处理方法、系统及计算机可读存储介质
CN113008583A (zh) 一种旋转机械状态监测和异常自动报警的方法及装置
CN110808068A (zh) 一种声音检测方法、装置、设备和存储介质
Wang et al. Construction of the efficient attention prototypical net based on the time–frequency characterization of vibration signals under noisy small sample
CN114034481A (zh) 一种轧机齿轮箱故障诊断系统及方法
US11589760B2 (en) System and method for physiological monitoring and feature set optimization for classification of physiological signal
CN112052712B (zh) 一种电力设备状态监测与故障识别方法及系统
CN114048787B (zh) 一种基于Attention CNN模型的轴承故障实时智能诊断方法与系统
Mobtahej et al. Deep learning-based anomaly detection for compressors using audio data
CN115310490B (zh) 基于多域特征与敏感特征选择的旋转设备故障分析方法
CN117576632A (zh) 基于多模态ai大模型的电网监控火灾预警系统及方法
Gao et al. Weak fault detection with a two-stage key frequency focusing model
Wang et al. An approach to fault diagnosis for gearbox based on image processing
Li et al. Fault diagnosis for machinery based on feature extraction and general regression neural network
CN115993503A (zh) 一种变压器的运行检测方法、装置、设备及存储介质
CN114639391A (zh) 机械故障提示方法、装置、电子设备及存储介质
CN115526882A (zh) 一种医学图像的分类方法、装置、设备及存储介质
TWI438416B (zh) 使用連續遞移轉換之信號分析系統及方法
CN116304633A (zh) 一种机械部件在线智能故障诊断方法及故障诊断系统
CN117457029A (zh) 设备故障检测方法、装置、计算机设备及存储介质
CN117373487B (zh) 基于音频的设备故障检测方法、装置及相关设备
Jiang et al. Rolling Bearing Fault Diagnosis

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: 23778166

Country of ref document: EP

Kind code of ref document: A1