WO2023060542A1 - 轴承故障检测方法及存储介质 - Google Patents

轴承故障检测方法及存储介质 Download PDF

Info

Publication number
WO2023060542A1
WO2023060542A1 PCT/CN2021/124036 CN2021124036W WO2023060542A1 WO 2023060542 A1 WO2023060542 A1 WO 2023060542A1 CN 2021124036 W CN2021124036 W CN 2021124036W WO 2023060542 A1 WO2023060542 A1 WO 2023060542A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
fault
bearing
signal
training
Prior art date
Application number
PCT/CN2021/124036
Other languages
English (en)
French (fr)
Inventor
邱志
佟庆波
陈士玮
夏美金
程润
Original Assignee
舍弗勒技术股份两合公司
邱志
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 舍弗勒技术股份两合公司, 邱志 filed Critical 舍弗勒技术股份两合公司
Priority to PCT/CN2021/124036 priority Critical patent/WO2023060542A1/zh
Publication of WO2023060542A1 publication Critical patent/WO2023060542A1/zh

Links

Images

Definitions

  • the invention relates to the field of fault diagnosis of mechanical equipment based on an artificial intelligence method, in particular to a bearing fault detection method and a storage medium.
  • the fault detection method of mechanical equipment is usually to detect the size of parts, but the assembled mechanical equipment will vibrate during operation, which will lead to automatic fine-tuning of the connection gap of parts, which will lead to poor overall operation effect.
  • the fault detection method for bearings is generally through the scheme of collecting the vibration signal of the measuring point, and then the fault detection is carried out through the method of frequency spectrum analysis.
  • the process of measuring and analyzing spindle bearings by means of vibration sensors is cumbersome and requires the support of professional engineers.
  • the running sound of the spindle bearing is different under normal and faulty conditions.
  • Some experienced engineers can judge whether the bearing is faulty by collecting the running sound of the bearing in real time, but this method is affected by subjective factors. Larger, the judgment results of different people are quite different, and it is difficult to standardize and promote.
  • This vibration frequency is the shock frequency, and it is modulated in the vibration signal of the spindle system to generate an abnormal sound signal. Since the human ear can only hear sounds within a certain frequency range, humans are more sensitive to low-frequency modulation frequencies. Therefore, the effect of fault discrimination on rotating parts with low rotational speed is more obvious.
  • the signal modulation frequency is 20-200Hz or less At 20Hz, the psychoacoustic roughness and fluctuation parameters can objectively reflect the modulation characteristics of the human ear to the fault frequency.
  • the spindle bearings basically operate at high speeds, and the modulation frequency of the bearing fault at this time has exceeded the accurate perception range of the low-frequency modulation frequency by the human ear.
  • the object of the present invention is to provide a bearing fault detection method and a storage medium to solve the technical problem in the prior art that it is difficult to use sound signals to detect bearing faults during bearing operation.
  • the present invention provides a bearing fault detection method, comprising the following steps: a fault judgment model building step, collecting a plurality of sound signals of a bearing in a normal state and a fault state, and preprocessing each sound signal, Obtain the distribution characteristics of each sound signal on the Bark domain, divide a plurality of said distribution characteristics into two groups as training samples, input the two groups of training samples and their group labels into a neural network model for training, and obtain a fault judgment model; and a fault detection step of inputting the test sample corresponding to the measured sound signal into the fault judgment model, and the fault judgment model judges whether the bearing operation state corresponding to the measured sound signal is a fault state, and outputs a judgment result.
  • the step of establishing the fault judgment model includes: a signal collection step, during the operation of the mechanical equipment, collecting a plurality of sound signals of the main shaft bearing of the mechanical equipment in normal state and fault state; a signal preprocessing step, for each Each sound signal is processed by framing to form a plurality of framing signals; each framing signal is input into a psychoacoustic model, and the psychoacoustic model outputs the distribution characteristics of each framing signal on the Bark domain; sample grouping Step, divide a plurality of described distribution features into two groups of training samples, the same group of training samples is marked with the same group label, and the two group labels are normal and fault respectively; and the model training step, a plurality of training samples and The group labels are input to a neural network model for training, and a fault judgment model is obtained after training.
  • a detection signal collection step collecting the measured sound signal of the main shaft bearing
  • a detection signal preprocessing step for each measured The sound signal is divided into frames, each measured sound signal is decomposed into multiple measured framed signals, each measured framed signal is input into the psychoacoustic model, and the distribution of each measured framed signal in the Bark domain is calculated feature, the distribution feature is used as a test sample.
  • the training error is represented by mean square error MSE, and the training is stopped when the calculated mean square error MSE is smaller than the set target error.
  • the characteristic loudness and characteristic sharpness of each sub-frame signal on the Bark domain are merged and added to obtain multiple feature vectors of each sub-frame signal on the Bark domain, as The distribution characteristics of the framed signal.
  • the neural network model includes an input layer, and the number of input nodes of the input layer is the same as the number of parameters of the psychoacoustic model.
  • the output values of the fault judgment model are 0 and 1 respectively, the group label corresponding to the output value 0 is normal, and the group label corresponding to the output value 1 Don't label it as a bug.
  • the present invention also provides a storage medium, the storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the steps in any one of the bearing fault detection methods described above.
  • the beneficial effect of the present invention is to provide a bearing fault detection method and storage medium, based on the principle that the operating sound of the bearing is different under the normal state and the fault state, the original collected sound data set is entered into the psychoacoustic model to obtain parameters, according to the parameters
  • To calculate the eigenvalues corresponding to different sounds use the neural network algorithm to train a fault detection model, and input the measured sound signals into the model to judge in real time whether the bearing is in fault during operation.
  • the detection results are highly accurate and not subject to human subjective factors.
  • the invention adopts the non-contact measurement signal as the data source, the detection process does not need to collect the physical parameters and rotational speed of the bearing, and uses the sound feature as the only criterion for bearing fault judgment, and the operation is simple and convenient.
  • Fig. 1 is the flow chart that obtains 24 comprehensive sound quality indexes as distribution feature in the embodiment of the present invention
  • Fig. 2 is a schematic diagram of a bearing intelligent fault diagnosis method based on auditory perception in an embodiment of the present invention
  • Fig. 3 is the main flowchart of the bearing fault detection method described in the embodiment of the present invention.
  • Fig. 4 is the flow chart of the establishment step of fault judgment model described in the embodiment of the present invention.
  • Fig. 5 is the schematic diagram of neural network model training based on auditory perception in the embodiment of the present invention.
  • FIG. 6 is a flow chart of spindle bearing fault detection based on the HAP-NN model in an embodiment of the present invention
  • Fig. 7 is a flow chart of the bearing fault detection method in the embodiment of the present invention.
  • This application proposes an intelligent bearing fault detection method based on the human ear's auditory perception, and establishes the bearing sound-human ear-human brain - Objective evaluation model of diagnostic results. It can be understood that this method is applicable to the judgment of the running state of all mechanical devices, and this application is preferably for judging the running state of the main shaft bearing.
  • the parameters of the psychoacoustic model may include loudness, sharpness, roughness, fluctuation, tone, Pitch strength, Annoyance. These parameters describe the human ear's subjective auditory perception characteristics such as sound intensity, harshness, modulation, and periodicity.
  • loudness and sharpness are the main influencing factors for the judgment of the running state of the spindle bearing. Therefore, according to the working characteristics of the spindle bearing, using the distribution characteristics of the sound quality characteristic loudness and characteristic sharpness in the Bark domain as the input of the neural network model can effectively reduce the number of input nodes and speed up the calculation.
  • the sound signal of the main shaft bearing of the mechanical equipment under each group of working conditions is first collected and processed in frames, and each frame of sound signal is input into the psychoacoustic model to calculate the characteristic loudness and characteristic sharpness in the Bark domain
  • the eigenvalues of the degree; on the Bark domain are the eigenvalues of the 24-point characteristic loudness and the 24-point characteristic sharpness, and the 24-point characteristic loudness and the 24-point characteristic sharpness on the Bark domain are combined and added to produce 24 loudness and sharpness
  • the eigenvector on the Bark domain of degree common feature is used as 24-point comprehensive sound quality index, and described 24-point comprehensive sound quality index is used as the distribution characteristic of this sub-frame signal; Then described 24-point comprehensive sound quality index is input in the described neural network model Carry out model training and obtain a fault judgment model after training, which can effectively reduce the number of input nodes.
  • the 24-point comprehensive sound quality index in the test sample corresponding to the measured sound signal of the bearing is input into the trained fault judgment model, and then the human auditory perception neural network model (HAP-NN model , that is, the fault judgment model trained in this application) will intelligently judge the running state of the bearing.
  • HAP-NN model that is, the fault judgment model trained in this application
  • FIG. 2 it is a schematic diagram of an intelligent fault diagnosis method for spindle bearings based on human auditory perception, mainly for the sound signals of the spindle bearings during operation through the neural network model.
  • Sharpness calculates the characteristic loudness and characteristic sharpness on the Bark domain, and then merges and adds the two to generate 24 eigenvectors on the Bark domain with common features of loudness and sharpness as 24-point comprehensive sound quality index, each The 24-point integrated sound quality index of the framed signal on the Bark domain is used as the distribution feature of the framed signal; then the 24-point integrated sound quality index obtained by these calculations is used as the input vector of the neural network model, and the neural network model preferably outputs normal and There are two kinds of fault judgment results.
  • the Bark domain (Bark domain) is a psychoacoustic scale of sound proposed earlier.
  • Critical band is a term in audiology and psychoacoustics, it was proposed by Harvey Fletcher in the 1840s.
  • the cochlea is the sensory organ of hearing in the inner ear, and the critical band refers to the frequency bandwidth of the auditory filter due to the structure of the cochlea.
  • the critical frequency band is the band of sound frequencies in which the perceptibility of a first tone is disturbed by the auditory masking of a second tone.
  • people use auditory filters to simulate different critical frequency bands.
  • the structure of the human ear would resonate roughly at 24 frequency points. According to this conclusion, Eberhard Zwicker proposed in 1961 for the special structure of the human ear: the signal also presents 24 critical frequency bands in the frequency band, from 1 to 24, respectively. This is the Bark domain.
  • the embodiment of the present invention provides a bearing fault detection method, including the following steps: S10, the step of establishing a fault judgment model, collecting a plurality of sound signals of a bearing in a normal state and a fault state, and for each sound The signal is preprocessed to obtain the distribution characteristics of each sound signal on the Bark domain, and a plurality of the distribution characteristics are divided into two groups as training samples, and each group of training samples has a group label respectively (normal state is recorded as 0, fault The state is recorded as 1), two sets of training samples and their group labels are input to a neural network model for training to obtain a fault judgment model; and S20, the fault detection step, the test samples corresponding to the measured sound signal are input to the fault judgment model wherein, the fault judgment model judges whether the bearing operating state corresponding to the measured sound signal is a fault state, and outputs a judgment result; wherein the judgment result output by the fault judgment model is the group label corresponding to the measured sound signal.
  • the fault judgment model establishment step S10 specifically includes steps S11-S14.
  • the signal collection step during the operation of the mechanical equipment, collect a plurality of sound signals of the main shaft bearing of the mechanical equipment in the normal state and the fault state, specifically through the acquisition system to sample the main shaft bearing of the mechanical equipment under each group of working conditions
  • the sound signals form the sound data set.
  • the sampling rate of the acquisition system is set to 44000Hz-45000Hz, preferably 44100Hz.
  • the acquisition system includes a microphone sensor; the time-domain sound signal measured by the microphone sensor is processed in frames, each frame is 40ms-60ms, preferably 50ms, and sound signals under various working conditions are collected simultaneously.
  • the mechanical equipment is divided into various groups of working conditions based on the combination of rotation speed and load, for example, sound signals under different rotation speeds and different loads.
  • the number of working conditions is defined as L, that is, the sound signals of L groups of spindle running are collected.
  • the sound signal of each group of working conditions in the sound data set is divided into frames, and each frame of data is input into the psychoacoustic model, and the characteristic loudness and characteristic sharpness on the 24 Bark domains are calculated according to the parameters of the psychoacoustic model degree characteristic value.
  • the signal preprocessing step is to carry out frame processing on each sound signal to form a plurality of frame signals; each frame signal is input into a psychoacoustic model, and the psychoacoustic model outputs each frame signal at Distribution features over Bark domains.
  • the main function of the psychoacoustic model is to convert the sound signal into parameters including loudness, sharpness, roughness, fluctuation, pitch, pitch intensity, and annoyance.
  • the distribution characteristics of each sub-frame signal on the Bark domain include characteristic loudness and characteristic sharpness, and these two factors are the main influencing factors for judging the running state of the spindle bearing.
  • S13, sample grouping step divide a plurality of described distribution characteristics into two groups of training samples, the same group of training samples are marked with the same group label, the two group labels are normal and fault respectively, normal state is recorded as 0, fault The status is recorded as 1. Setting the two group labels of normal and fault can reasonably divide the multiple sound signals of the spindle bearing in the normal state and fault state collected in the signal collection step into two groups, which is beneficial to the subsequent data classification input for model training.
  • the model training step a plurality of training samples and their group labels are input to a neural network model for training, and a fault judgment model is obtained after training; wherein the training error is represented by the mean square error MSE, when the calculated mean square error Stop training when the MSE is less than the set target error.
  • the structure of the fault judgment model is the same as that of the neural network model, and the difference is mainly that the fault judgment model is a trained model of the neural network model.
  • the neural network model includes an input layer, a hidden layer and an output layer, wherein the activation function f1 used by the hidden layer is a sigmoid function, and the activation function f2 of the output layer is a linear function (purelin);
  • the number of input nodes of the input layer is the same as the number of parameters of the psychoacoustic model, and the eigenvalues obtained by calculating the parameters corresponding to the psychoacoustic model are used as the input vector of the neural network model, and the number of the input vectors is The number is the same as the number of input nodes in the input layer, and the number of output nodes in the output layer is at least 2.
  • the number M of input nodes of the neural network is 24; the number N of output nodes of the neural network is 2, and the output values of the neural network are 0 and 1 respectively.
  • 0 means normal, 1 means failure; or 0 means failure, 1 means normal.
  • the model training step before inputting the eigenvalues of the characteristic loudness and characteristic sharpness on the Bark domain into the established neural network model for training, the characteristic loudness and characteristic sharpness on the Bark domain are combined and added to generate 24
  • the eigenvector on the Bark domain with the common features of loudness and sharpness is used as 24-point comprehensive sound quality index, and the 24-point comprehensive sound quality index is used as the distribution feature of the framed signal, and then the 24-point comprehensive sound quality index is input into the Training in the neural network model, as shown in Figure 5,
  • Figure 5 is a schematic diagram of the neural network model training based on auditory perception, the training error is represented by the mean square error MSE, when the calculated mean square error MSE is less than the set target error Then stop training.
  • mean square error MSE is calculated with the following expression: where L is the signal length.
  • the sound signal collected in real time by the mechanical equipment needs to be used as a test sample of the measured sound signal, and the test sample of the measured sound signal is input into the fault judgment model (trained neural network model), so The fault judgment model outputs the group label corresponding to the measured sound signal as a detection result.
  • Figure 6 shows the spindle bearing fault detection based on the HAP-NN model (that is, the fault judgment model), mainly for the sound signals of the spindle bearing during operation through the fault judgment model, and these sound signals are performed according to the auditory parameters of the psychoacoustic model Analysis and judgment, specifically calculating the loudness and sharpness in the auditory parameters of the psychoacoustic model to the characteristic values of the characteristic loudness and characteristic sharpness on the Bark domain; Eigenvalues such as loudness and characteristic sharpness are merged and added to generate 24 eigenvectors on the Bark domain with common characteristics of loudness and sharpness as 24-point comprehensive sound quality index, and the 24-point comprehensive sound quality index is used as the distribution feature of the framed signal , and then input the 24-point comprehensive sound quality index as the input vector of the neural network model into the fault judgment model for judgment, and the fault judgment model preferably outputs two group labels of normal and fault as judgment results.
  • the HAP-NN model that is, the fault judgment model
  • S15 detection signal collection step, collecting the measured sound signal of the main shaft bearing
  • S16 detection
  • the signal preprocessing step is to perform frame processing on each measured sound signal, and each measured sound signal is decomposed into a plurality of measured framed signals, and each measured framed signal is input into the psychoacoustic model, and each measured sound signal is calculated.
  • the distribution characteristics of the framed signal on the Bark domain, and the distribution characteristics are used as test samples.
  • the present invention also provides a storage medium, the storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the steps in any one of the bearing fault detection methods described above.
  • the beneficial effect of the present invention is to provide a bearing fault detection method and storage medium, based on the principle that the operating sound of the bearing is different under the normal state and the fault state, the original collected sound data set is entered into the psychoacoustic model to obtain parameters, according to the parameters
  • To calculate the eigenvalues corresponding to different sounds use the neural network algorithm to train a fault detection model, and input the measured sound signals into the model to judge in real time whether the bearing is in fault during operation.
  • the detection results are highly accurate and not subject to human subjective factors.
  • the invention adopts the non-contact measurement signal as the data source, the detection process does not need to collect the physical parameters and rotational speed of the bearing, and uses the sound feature as the only criterion for bearing fault judgment, and the operation is simple and convenient.

Landscapes

  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)

Abstract

一种轴承故障检测方法及存储介质。轴承故障检测方法包括：故障判断模型建立步骤，采集一轴承在正常状态与故障状态下的多个声音信号，对每个声音信号进行预处理获取在Bark域上的分布特征并分成两组训练样本且每组具有一组别标签，将多个训练样本及其组别标签输入至一神经网络模型进行训练获得一故障判断模型；故障检测步骤，将实测声音信号的测试样本输入至故障判断模型中，输出所述实测声音信号对应的组别标签。本方法将原始采集的声音数据集录入心理声学模型获得参数，根据该参数来计算不同声音对应的特征值，并训练出一故障检测模型，将实测声音信号录入该模型就可以实时判断轴承在运行中是否处于故障，不受人的主观因素影响。

Description

轴承故障检测方法及存储介质 技术领域
本发明涉及基于人工智能方法实现机械设备故障诊断领域，具体地说，涉及一种轴承故障检测方法及存储介质。
背景技术
机械设备的故障检测方法通常是对零部件的尺寸进行检测，但是组装后的机械设备在运行过程中会产生震动，导致部件衔接间隙的自动微调，进而导致整体运行效果欠佳。
机床的加工精度以及装配后的成品的部件间隙都会影响主轴轴承的运行，因此对装配后的轴承的工作状态进行有效的检测显得十分必要。目前对于轴承的故障检测方法一般是通过采集测点振动信号的方案，然后通过频谱分析的方法进行故障检测。依靠振动传感器来对主轴轴承进行测量和分析的过程比较繁琐，且需要专业的工程师支持。
在实际应用中，主轴轴承在正常状态下和故障状态下的运行声音是不同的，一些经验丰富的工程师可以通过实时采集轴承的运行声音来判断轴承是否发生故障，但这种方法受主观因素影响较大，不同人的判断结果差异较大，难以标准化推广。
此外，轴承产生故障时会产生特殊的振动，这种振动频率为冲击频率，并调制在主轴系统振动信号之中，产生异常的声音信号。由于人耳只能听到频率在一定范围内的声音，人类对低频调制频率的声音比较敏感，因此，对于转速低的旋转部件的故障区分效果比较明显，当信号调制频率为20-200Hz或小于20Hz时，心理声学的粗糙度和波动度参数能够客观的反映人耳对故障频率的调制特征。然而，主轴轴承工作时基本都是以高转速为主，这时的轴承故障调制频率已经超出人耳对低频调制频率的准确感知范围。
发明内容
本发明的目的在于，提供一种轴承故障检测方法及存储介质，以解决现有技术中在轴承运行过程中难以利用声音信号进行轴承故障检测的技术问题。
为实现上述目的，本发明提供一种轴承故障检测方法，包括如下步骤：故障判断模型建立步骤，采集一轴承在正常状态与故障状态下的多个声音信号，对每个声音信号进行预处理，获取每个声音信号在Bark域上的分布特征，将多个所述分布特征作为训练样本分成两组，将两组训练样本及其组别标签输入至一神经网络模型进行训练，获得一故障判断模型；以及故障检测步骤，将实测声音信号对应的测试样本输入至故障判断模型中，所述故障判断模型判断所述实测声音信号对应的轴承运行状态是否为故障状态，并输出判断结果。
进一步地，所述故障判断模型建立步骤包括：信号采集步骤，在机械设备运行中，采集所述机械设备的主轴轴承在正常状态与故障状态下的多个声音信号；信号预处理步骤，对每个声音信号进行分帧处理，形成多个分帧信号；将每一分帧信号输入至一心理声学模型中，所述心理声学模型输出每一分帧信号在Bark域上的分布特征；样本分组步骤，将多个所述分布特征分成两组训练样本，同组训练样本被标识有相同的组别标签，两个组别标签分别为正常及故障；以及模型训练步骤，将多个训练样本及其组别标签输入至一神经网络模型进行训练，经训练后获得一故障判断模型。
进一步地，在所述模型训练步骤之后，在所述故障检测步骤之前，还包括如下步骤：检测信号采集步骤，采集所述主轴轴承的实测声音信号；以及检测信号预处理步骤，对每个实测声音信号进行分帧处理，每个实测声音信号被分解为多个实测分帧信号，将每一实测分帧信号输入至心理声学模型中，计算出每一实测分帧信号在Bark域上的分布特征，将该分布特征作为测试样本。
进一步地，在所述模型训练步骤中，训练误差用均方误差MSE表示，当计算的均方误差MSE小于设定的目标误差后停止训练。
进一步地，在所述信号预处理步骤中，将每一分帧信号在Bark域上的特征响度及特征尖锐度合并相加，获得每一分帧信号在Bark域上的多个特征向量，作为该分帧信号的分布特征。
进一步地，在所述模型训练步骤中，所述神经网络模型包括输入层，所述输入层的输入节点个数与所述心理声学模型的参数的个数相同。
进一步地，向所述神经网络模型输入的训练样本用向量表示为X＝[x 1,x 2,x 3,…,x j,…,x M]，其中M为神经网络的输入节点个数；所述M为24。
进一步地，在所述模型训练步骤中，所述故障判断模型包括输出层，所述输出层的输出节点个数至少为2；所述故障判断模型的输出向量为Y＝[y 1,y 2,y 3,…,y i,…,y N]，其中N为神经网络的输出节点个数。
进一步地，当所述故障判断模型的输出节点个数N为2时，所述故障判断模型的输出值分别为0和1，输出值0对应的组别标签为正常，输出值1对应的组别标签为故障。
本发明还提供一种存储介质，所述存储介质存储有多条指令，所述指令适于处理器进行加载，以执行前文任一项所述的轴承故障检测方法中的步骤。
本发明的有益效果在于，提供一种轴承故障检测方法及存储介质，基于轴承正常状态下和故障状态下运行声音不同的原理，将原始采集的声音数据集录入心理声学模型获得参数，根据该参数来计算不同声音对应的特征值，利用神经网络算法训练出一故障检测模型，将实测声音信号录入该模型就可以实时判断轴承在运行中是否处于故障，检测结果准确率高，不受人的主观因素影响。本发明采用非接触测量信号为数据源，检测过程不需要采集轴承物理参数及转速，利用声音特征作为轴承故障判断的唯一标准，操作简单方便。
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。
图1为本发明实施例中获取24点综合音质指数作为分布特征的流程图；
图2为本发明实施例中基于听觉感知的轴承智能故障诊断方法原理图；
图3为本发明实施例中所述轴承故障检测方法的主要流程图；
图4为本发明实施例中所述故障判断模型建立步骤的流程图；
图5为本发明实施例中基于听觉感知的神经网络模型训练的原理图；
图6为本发明实施例中基于HAP-NN模型的主轴轴承故障检测流程图；
图7为本发明实施例中所述轴承故障检测方法的流程图。
具体实施方式
以下参考说明书附图完整介绍本发明的优选实施例，使其技术内容更加清楚和便于理解。本发明可以通过许多不同形式的实施例来得以体现，其保护范围并非仅限于文中提到的实施例。
依据工程实际应用，对于工程师通过耳朵的听力来判断出轴承是否发生故障的原理做了总结，本申请提出了基于人耳听觉感知的智能轴承故障检测方法，建立了轴承声音-人耳-人脑-诊断结果的客观评价模型。可理解的是，这种方式适用于所有的机械设置的运行状态的判断，本申请优选为对主轴轴承的运行状态判断。
若人耳听觉感知特性用心理声学模型的参数来描述，心理声学模型的参数可包括响度(loudness)、尖锐度(Sharpness)、粗糙度(Roughness)、波动度(Fluctuation)、音调(Tonality)、音调强度(Pitch strength)、烦恼度(Annoyance)。这些参数描述了人耳对声音强弱、刺耳、调制、周期性等主观听觉感知特性。经研究发现其中的响度(loudness)和尖锐度(Sharpness)对主轴轴承的运行状态判断为主要影响因素。因此针对主轴轴承工作特点，利用声品质特征响度和特征尖锐度在Bark域上的分布特征作为神经网络模型的输入，可有效减少输入节点的数量，且能加快计算速度。
如图1所示，首先采集机械设备的主轴轴承在各组工况下的声音信号并进行分帧处理，将每一帧声音信号输入至心理声学模型中计算Bark域上的特征响度和特征尖锐度的特征值；在Bark域上均是24点特征响度和24点特征尖锐度的特征值，将Bark域上的24点特征响度和24点特征尖锐度合并相加产生24个具有响度和尖锐度共同特征的Bark域上的特征向量作为24点综合音质指数，所述24点综合音质指数作为该分帧信号的分布特征；再将所述24点综合音质指数输入至所述神经网络模型中进行模型训练，经训练后获得一故障判断模型，可有效减少输入节点数量。
在后续判断轴承是否为故障状态时，将轴承的实测声音信号对应的测试样本中的24点综合音质指数输入到训练好的故障判断模型中，然后人耳听觉感知神经网络模型(HAP-NN模型，也就是本申请中训练好的故障判断模型)将智能的判断出轴承的运行状态。
如图2所示，为基于人耳听觉感知的主轴轴承智能故障诊断方法原理图，主要为主轴轴承运行时声音信号通过神经网络模型对这些声音信号按照心理声学模型的参数响度(loudness)、尖锐度(Sharpness)计算在Bark域上的特征响度和特征尖锐度，然后将两者合并相加产生24个具有响度和尖锐度共同特征的Bark域上的特征向量作为24点综合音质指数，每一分帧信号在Bark域上的24点综合音质指数作为该分帧信号的分布特征；然后将这些计算获得的24点综合音质指数作为所述神经网络模型的输入向量，神经网络模型优选输出正常和故障两种判断结果。
其中Bark域(巴克域)是较早提出来的一种声音的心理声学尺度。临界频带是听觉学和心理声学的专业名词，它于19世纪40年代年被Harvey Fletcher提出。耳蜗是内耳中听觉的传感器官，临界频带指的是由于耳蜗构造产生的听觉滤波器的频率带宽。概况地说，临界频带是声音频率带，在临界频带中第一个单音感知性会被第二单音的听觉掩蔽所干扰。声学研究中，人们使用听觉滤波器来模拟不同的临界频带。后来研究者发现人耳结构大致会对24 个频率点产生共振，根据这个结论Eberhard Zwicker在1961年针对人耳特殊结构提出：信号在频带上也呈现出24个临界频带，分别从1到24，这就是Bark域。
如图3所示，本发明实施例提供一种轴承故障检测方法，包括如下步骤：S10、故障判断模型建立步骤，采集一轴承在正常状态与故障状态下的多个声音信号，对每个声音信号进行预处理，获取每个声音信号在Bark域上的分布特征，将多个所述分布特征作为训练样本分成两组，每组训练样本分别具有一组别标签(正常状态记为0，故障状态记为1)，将两组训练样本及其组别标签输入至一神经网络模型进行训练获得一故障判断模型；以及S20、故障检测步骤，将实测声音信号对应的测试样本输入至故障判断模型中，所述故障判断模型判断所述实测声音信号对应的轴承运行状态是否为故障状态，并输出判断结果；其中所述故障判断模型输出的判断结果为所述实测声音信号对应的组别标签。
具体的，如图4所示，所述故障判断模型建立步骤S10具体包括步骤S11-S14。
S11、信号采集步骤，在机械设备运行中，采集所述机械设备的主轴轴承在正常状态与故障状态下的多个声音信号，具体是通过采集系统采样机械设备的主轴轴承在各组工况下的声音信号形成声音数据集。为了覆盖心理声学分析所需频率范围，设置采集系统采样率为44000Hz-45000Hz，优选为44100Hz。所述采集系统包括麦克风传感器；将麦克风传感器测得的时域声音信号进行分帧处理，每帧为40ms-60ms，优选为50ms，同时采集各个工况下的声音信号。所述机械设备是基于转速和载荷的组合划分为各组工况，例如不同转速不同载荷下的声音信号。这里定义工况个数为L，即采集L组主轴运行的声音信号。将声音数据集中的每组工况的声音信号进行分帧处理，将每一帧数据输入至心理声学模型中，根据所述心理声学模型的参数来计算24个Bark域上的特征响度和特征尖锐度特征值。
S12、信号预处理步骤，对每个声音信号进行分帧处理，形成多个分帧信号；将每一分帧信号输入至一心理声学模型中，所述心理声学模型输出每一分帧信号在Bark域上的分布特征。心理声学模型的主要作用是可以将声音信号转换为包括响度、尖锐度、粗糙度、波动度、音调、音调强度、烦恼度的参数。本实施例中每一分帧信号在Bark域上的分布特征包括特征响度和特征尖锐度，这两个因素为对主轴轴承的运行状态判断为主要影响因素。
S13、样本分组步骤，将多个所述分布特征分成两组训练样本，同组训练样本被标识有相同的组别标签，两个组别标签分别为正常及故障，正常状态记为0，故障状态记为1。设置正常及故障两个组别标签可合理划分信号采集步骤中所对应采集的主轴轴承在正常状态与故障状态下的多个声音信号为两组，有利于后续对模型训练的数据分类输入。
S14、模型训练步骤，将多个训练样本及其组别标签输入至一神经网络模型进行训练，经训练后获得一故障判断模型；其中训练误差用均方误差MSE表示，当计算的均方误差MSE小于设定的目标误差后停止训练。
值得注意的是，所述故障判断模型的结构与所述神经网络模型的结构相同，其差异主要是所述故障判断模型为所述神经网络模型进行训练后的模型。
所述神经网络模型包括输入层、隐含层和输出层，其中所述隐含层所用的激活函数f 1为sigmoid函数，所述输出层的激活函数f 2为线性函数(purelin)；所述输入层的输入节点个数与所述心理声学模型的参数的个数相同，将对应所述心理声学模型的参数计算获得的特征值作为所述神经网络模型的输入向量，所述输入向量的个数与所述输入层的输入节点个数相同，所述输出层的输出节点个数至少为2。
其中，向所述神经网络模型输入的训练样本可用向量表示为X＝[x 1,x 2,x 3,…,x j,…,x M]，其中M为神经网络的输入节点个数；所述神经网络模型的输出向量为Y＝[y 1,y 2,y 3,…,y i,…,y N]，其中N为神经网络的输出节点个数。
优选地，所述神经网络的输入节点个数M为24；所述神经网络的输出节点个数N为2，所述神经网络的输出值分别为0和1。其中，0代表正常，1代表失效；或者0代表失效，1代表正常。
所述隐含层的输出计算公式为：
Figure PCTCN2021124036-appb-000001
其中j＝1,2,3,…,l；k＝1,2,3,…,M；w (1)为对应输入向量x k的权重。
所述输出层的计算公式为：
Figure PCTCN2021124036-appb-000002
其中i＝1,2,3,…,N；w (2)为对应输出向量y j的权重。
在模型训练步骤中，在将Bark域上的特征响度和特征尖锐度特征值输入至所建立的神经网络模型进行训练之前，先将Bark域上的特征响度和特征尖锐度合并相加产生24个具有响度和尖锐度共同特征的Bark域上的特征向量作为24点综合音质指数，所述24点综合音质指数作为该分帧信号的分布特征，再将所述24点综合音质指数输入至所述神经网络模型中进行训练，如图5所示，图5为基于听觉感知的神经网络模型训练的原理图，训练误差用均方误差MSE表示，当计算的均方误差MSE小于设定的目标误差后停止训练。
其中所述均方误差MSE用如下表达式计算：
Figure PCTCN2021124036-appb-000003
其中L为信号长度。
在进行所述故障检测步骤S20之前，需要对机械设备实时采集的声音信号作为实测声音信号的测试样本，将实测声音信号的测试样本输入至故障判断模型(训练好的神经网络模型)中，所述故障判断模型输出所述实测声音信号对应的组别标签作为检测结果。
如图6所示，图6为基于HAP-NN模型(即故障判断模型)的主轴轴承故障检测，主要为主轴轴承运行时声音信号通过故障判断模型对这些声音信号按照心理声学模型的听觉参数进行分析判断，具体是将心理声学模型的听觉参数中的响度(loudness)、尖锐度(Sharpness)计算Bark域上的特征响度和特征尖锐度的特征值；然后将这些计算获得的Bark域上的特征响度和特征尖锐度等特征值 合并相加产生24个具有响度和尖锐度共同特征的Bark域上的特征向量作为24点综合音质指数，所述24点综合音质指数作为该分帧信号的分布特征，再将所述24点综合音质指数作为所述神经网络模型的输入向量输入至所述故障判断模型中进行判断，故障判断模型优选输出正常和故障两种组别标签作为判断结果。
如图7所示，在所述模型训练步骤S14之后，在所述故障检测步骤S20之前，还包括如下步骤：S15、检测信号采集步骤，采集所述主轴轴承的实测声音信号；以及S16、检测信号预处理步骤，对每个实测声音信号进行分帧处理，每个实测声音信号被分解为多个实测分帧信号，将每一实测分帧信号输入至心理声学模型中，计算出每一实测分帧信号在Bark域上的分布特征，将该分布特征作为测试样本。通过步骤S15和S16可保证测试样本与训练样本为在Bark域上具有相同数据格式的分布特征，利于检测信号数据分析和判断。
本发明还提供一种存储介质，所述存储介质存储有多条指令，所述指令适于处理器进行加载，以执行前文任一项所述的轴承故障检测方法中的步骤。
本发明的有益效果在于，提供一种轴承故障检测方法及存储介质，基于轴承正常状态下和故障状态下运行声音不同的原理，将原始采集的声音数据集录入心理声学模型获得参数，根据该参数来计算不同声音对应的特征值，利用神经网络算法训练出一故障检测模型，将实测声音信号录入该模型就可以实时判断轴承在运行中是否处于故障，检测结果准确率高，不受人的主观因素影响。本发明采用非接触测量信号为数据源，检测过程不需要采集轴承物理参数及转速，利用声音特征作为轴承故障判断的唯一标准，操作简单方便。
以上所述仅是本发明的优选实施方式，使本领域的技术人员更清楚地理解如何实践本发明，这些实施方案并不是限制本发明的范围。对于本技术领域的普通技术人员，在不脱离本发明原理的前提下，还可以做出若干改进和润饰，这些改进和润饰也应视为本发明的保护范围。

Claims (10)

  1. 一种轴承故障检测方法，其特征在于，包括步骤：
    故障判断模型建立步骤，采集一轴承在正常状态与故障状态下的多个声音信号，对每个声音信号进行预处理，获取每个声音信号在Bark域上的分布特征，将多个所述分布特征作为训练样本分成两组，将两组训练样本及其组别标签输入至一神经网络模型进行训练，获得一故障判断模型；以及
    故障检测步骤，将实测声音信号对应的测试样本输入至故障判断模型中，所述故障判断模型判断所述实测声音信号对应的轴承运行状态是否为故障状态，并输出判断结果。
  2. 如权利要求1所述的轴承故障检测方法，其特征在于，所述故障判断模型建立步骤包括：
    信号采集步骤，在机械设备运行中，采集所述机械设备的主轴轴承在正常状态与故障状态下的多个声音信号；
    信号预处理步骤，对每个声音信号进行分帧处理，形成多个分帧信号；将每一分帧信号输入至一心理声学模型中，所述心理声学模型输出每一分帧信号在Bark域上的分布特征；
    样本分组步骤，将多个所述分布特征分成两组训练样本，同组训练样本被标识有相同的组别标签，两个组别标签分别为正常及故障；
    模型训练步骤，将多个训练样本及其组别标签输入至一神经网络模型进行训练，经训练后获得一故障判断模型。
  3. 如权利要求2所述的轴承故障检测方法，其特征在于，在所述模型训练步骤之后，在所述故障检测步骤之前，还包括如下步骤：
    检测信号采集步骤，采集所述主轴轴承的实测声音信号；以及
    检测信号预处理步骤，对每个实测声音信号进行分帧处理，每个实测声音信号被分解为多个实测分帧信号，将每一实测分帧信号输入至心理声学模型中，计算出每一实测分帧信号在Bark域上的分布特征，将该分布特征作为测试样本。
  4. 如权利要求2所述的轴承故障检测方法，其特征在于，在所述模型训练步骤中，训练误差用均方误差MSE表示，当计算的均方误差MSE小于设定的目标误差后停止训练。
  5. 如权利要求2所述的轴承故障检测方法，其特征在于，在所述信号预处理步骤中，将每一分帧信号在Bark域上的特征响度及特征尖锐度合并相加，获得每一分帧信号在Bark域上的多个特征向量，作为该分帧信号的分布特征。
  6. 如权利要求2所述的轴承故障检测方法，其特征在于，在所述模型训练步骤中，所述神经网络模型包括输入层，所述输入层的输入节点个数与所述心理声学模型的参数的个数相同。
  7. 如权利要求6所述的轴承故障检测方法，其特征在于，向所述神经网络模型输入的训练样本用向量表示为X＝[x 1,x 2,x 3,…,x j,…,x M]，其中M为神经网络的输入节点个数；所述M为24。
  8. 如权利要求2所述的轴承故障检测方法，其特征在于，在模型训练步骤中，所述故障判断模型包括输出层，所述输出层的输出节点个数至少为2；所述故障判断模型的输出向量为Y＝[y 1,y 2,y 3,…,y i,…,y N]，其中N为神经网络的输出节点个数。
  9. 如权利要求2所述的轴承故障检测方法，其特征在于，当所述故障判断模型的输出节点个数N为2时，所述故障判断模型的输出值分别为0和1，输出值0对应的组别标签为正常，输出值1对应的组别标签为故障。
  10. 一种存储介质，其特征在于，所述存储介质存储有多条指令，所述指令适于处理器进行加载，以执行权利要求1至9任一项所述的轴承故障检测方法中的步骤。
PCT/CN2021/124036 2021-10-15 2021-10-15 轴承故障检测方法及存储介质 WO2023060542A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/124036 WO2023060542A1 (zh) 2021-10-15 2021-10-15 轴承故障检测方法及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/124036 WO2023060542A1 (zh) 2021-10-15 2021-10-15 轴承故障检测方法及存储介质

Publications (1)

Publication Number Publication Date
WO2023060542A1 true WO2023060542A1 (zh) 2023-04-20

Family

ID=85987918

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/124036 WO2023060542A1 (zh) 2021-10-15 2021-10-15 轴承故障检测方法及存储介质

Country Status (1)

Country Link
WO (1) WO2023060542A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117571321A (zh) * 2023-11-24 2024-02-20 浙江大学 轴承故障检测方法、装置、设备及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110567720A (zh) * 2019-08-07 2019-12-13 东北电力大学 非平衡小样本场景下风机轴承故障深度对抗诊断方法
CN111275024A (zh) * 2020-03-22 2020-06-12 东北电力大学 含噪声标签有限数据驱动的变工况风机滚动轴承故障诊断方法
CN111413091A (zh) * 2020-04-02 2020-07-14 天津大学 一种基于数据驱动的强噪声干扰下齿轮箱故障诊断方法
CN112686110A (zh) * 2020-12-23 2021-04-20 武汉理工大学 一种离心泵故障诊断方法、装置、设备及存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110567720A (zh) * 2019-08-07 2019-12-13 东北电力大学 非平衡小样本场景下风机轴承故障深度对抗诊断方法
CN111275024A (zh) * 2020-03-22 2020-06-12 东北电力大学 含噪声标签有限数据驱动的变工况风机滚动轴承故障诊断方法
CN111413091A (zh) * 2020-04-02 2020-07-14 天津大学 一种基于数据驱动的强噪声干扰下齿轮箱故障诊断方法
CN112686110A (zh) * 2020-12-23 2021-04-20 武汉理工大学 一种离心泵故障诊断方法、装置、设备及存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117571321A (zh) * 2023-11-24 2024-02-20 浙江大学 轴承故障检测方法、装置、设备及存储介质
CN117571321B (zh) * 2023-11-24 2024-04-30 浙江大学 轴承故障检测方法、装置、设备及存储介质

Similar Documents

Publication Publication Date Title
TWI422460B (zh) Tool nose detection method for cutting machine tool
CN110570880B (zh) 一种鼾声信号识别方法
US11984135B2 (en) System and method for offline embedded abnormal sound fault detection
EP2085028A1 (en) Processing low signal-to-noise ratio signals
CN112052712B (zh) 一种电力设备状态监测与故障识别方法及系统
KR0145146B1 (ko) 신경회로망을 이용한 회전기기의 이상유무 진단장치 및 그 진단방법
WO2023060542A1 (zh) 轴承故障检测方法及存储介质
CN111508517A (zh) 一种基于噪声特征的微电机智能品控方法
CN208520476U (zh) 一种高速插秧机噪声品质评价系统
DE102019004239A1 (de) Musikklassifizierer und verwandte verfahren
CN109241838A (zh) 基于心理声学客观参数的变速器质量评价方法
CN108919962B (zh) 基于脑机数据集中处理的辅助钢琴训练方法
CN103983345A (zh) 一种基于人耳听觉特性的单频有调声信号自动监测方法
CN108344503A (zh) 一种高速插秧机噪声品质评价系统
CN117292713A (zh) 基于端边云协同的变压器声纹监测方法及系统
CN112735468A (zh) 一种基于mfcc的汽车座椅电机异常噪声检测方法
JP2002323371A (ja) 音響診断装置及び音響診断方法
CN114358321A (zh) 一种电机异响机器学习检测方法
JPH1123411A (ja) 異音判定装置及び異音判定方法
JP3608676B2 (ja) 自己回帰モデルを利用した時系列データの収集、処理方法
CN112992182B (zh) 车辆风噪水平测试系统及其测试方法
CN115002642A (zh) 一种基于听觉掩蔽结合svd-mrmr的扬声器异常声的特征提取方法
CN105628380B (zh) 一种可调式中智模型轴承故障分类器的故障分类方法
CN114566185A (zh) 一种异响自动检测方法、设备及计算机可读存储介质
JP3611291B2 (ja) 非定常騒音のラウドネス評価装置