A high-frequency acoustic wave-based rotating machinery early fault monitoring system and method

Abstract

The application discloses a kind of rotating machinery early fault monitoring system and method based on high-frequency acoustic wave, belong to the field of fault detection, system includes high-frequency acoustic wave sensor, circuit system and host computer, high-frequency acoustic wave signal generated by rotating machinery early fault is obtained by high-frequency acoustic wave sensor, circuit system is made of multiple signal conditioning board and a piece of communication control board card, and host computer carries out data communication and instruction sending with circuit system by communication module.Utilize high-frequency acoustic wave signal decibel value to monitor fault, adopt high-frequency acoustic wave demodulation signal to identify fault, fuse multiple signal processing methods to obtain early fault sensitive optimization features, and diagnose by LSSVM parameter optimization method.The application has the advantages of high integration, low hardware requirement, channel number can be expanded, wide frequency detection range, saving data storage resources, etc., to realize preventive maintenance to rotating machinery early fault.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fault detection, and in particular to a rotating machinery early fault monitoring system and method based on high-frequency acoustic waves. BACKGROUND

[0002] Rolling bearings, gears, etc. are an integral part of rotating machinery, and because they work in harsh environments for a long time, rolling bearings, gears, etc. are also the most vulnerable components in rotating machinery. According to statistics, 80% of the causes of failure and early failure of rotating machinery are caused by lubrication. In order to ensure the normal operation of rotating machinery in industrial production and manufacturing, and to avoid serious economic losses caused by downtime due to failure, monitoring the real-time running state information of lubrication, wear, etc. of key components such as rolling bearings and gears in rotating machinery is of great significance to the evaluation of the health status of rotating machinery.

[0003] Current rotating machinery fault diagnosis usually uses the method of monitoring vibration signals. However, vibration signals are easily disturbed by industrial noise, and when the rotating machinery is detected to have failed by the method of monitoring vibration signals, the rotating machinery system has often reached the stage of needing maintenance, and preventive maintenance cannot be achieved. Research shows that high-frequency acoustic wave signals can represent the early failure of rotating machinery better than vibration signals, and high-frequency acoustic waves are not easily affected by industrial environmental noise. However, the frequency range of high-frequency acoustic wave signals is between 20 kHz-100 kHz, which requires high-quality hardware for collection, is not convenient for signal collection and consumes a lot of storage resources. The nonlinear and dynamic time-varying characteristics of high-frequency acoustic wave signals also make it difficult to extract and analyze the fault state information of rotating machinery. How to effectively collect high-frequency acoustic wave signals and extract early fault information of rotating machinery from them, obtain state characteristic indexes with clear physical meaning and sensitivity, and accurately evaluate the running state of rotating machinery is a problem that needs to be solved. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a rotating machinery early fault monitoring system and method based on high-frequency acoustic waves. The existing vibration signal monitoring method is lagging behind, high-frequency acoustic wave signals can represent the early failure of rotating machinery better than vibration signals, but the hardware requirements for the collection system are high and consume a lot of storage resources. The nonlinear and dynamic time-varying characteristics of high-frequency acoustic wave signals also make it difficult to extract and analyze the fault state information of rotating machinery, and a single feature extraction method or index cannot obtain comprehensive characteristic of the health running state of the equipment.

[0005] To solve the above technical problems, the technical scheme of the present application is: a rotating machinery early fault monitoring system based on high-frequency acoustic waves, comprising high-frequency acoustic sensors, circuit systems and host computers connected in sequence, the circuit systems adopt plug-in card structure design and are composed of multiple signal conditioning board cards with the same structure and a communication control board card, each signal conditioning board card is provided with a high-frequency acoustic sensor for acquiring high-frequency acoustic signals generated by rotating machinery early faults;

[0006] The signal conditioning board card comprises an amplitude control unit, a mixer module, a filter amplification module, a decibel conversion module and a state switching module.

[0007] The communication control board card comprises a master control module, an analog-digital conversion module, a digital-analog conversion module and a communication module.

[0008] The host computer communicates data and sends instructions between the communication module and the communication control board card.

[0009] The further improvement of the technical scheme of the present application is that the amplitude control unit comprises a primary amplification module, a voltage-controlled amplification module and a secondary amplification module connected in sequence.

[0010] The primary amplification module performs primary amplification gain on the high-frequency acoustic signals output by the high-frequency acoustic sensor, the gain is controlled by the state switching module controlled by the master control module, the voltage-controlled amplification module adopts a voltage-controlled amplifier to realize the attenuation of the output signal of the primary amplification module, and the attenuation multiple is controlled by the level output by the digital-analog conversion module controlled by the master control module; the secondary amplification module is used for further amplifying the output signal of the voltage-controlled amplification module, and the gain is also controlled by the state switching module controlled by the master control module.

[0011] The further improvement of the technical scheme of the present application is that the master control module controls the two state switching modules to be synchronously switched through the IO port level of the master control module according to the attenuation multiple of the current voltage-controlled amplifier, when low-amplitude high-frequency acoustic signals are collected, the state switching module switches the primary amplification module to high gain, and the secondary amplification module is switched to low gain; when high-amplitude high-frequency acoustic signals are collected, the state switching module switches the primary amplification module to low gain, and the secondary amplification module is switched to high gain; the overall gain of the primary amplification module and the secondary amplification module remains dynamic balance.

[0012] 5、The further improvement of the technical scheme of the present application is that the mixer module realizes demodulation of the high-frequency acoustic signals processed by the amplitude control unit through the heterodyne frequency reduction principle, the demodulated high-frequency acoustic signals are filtered by the filter amplification module, the low-frequency components in the signals are retained, and the high-frequency acoustic signals are reduced to below 20 kHz.

[0013] The mixer module is built-in local oscillator, the oscillator adopts the voltage-controlled oscillator which can adjust the oscillation frequency according to the output level of the digital-to-analog conversion module, so as to realize the demodulation of the high-frequency acoustic signals with different center frequencies generated by the rotating machinery fault.

[0014] Further improvement of the technical scheme of the application is that the decibel conversion module adopts an RMS-DC converter to perform root mean square measurement on the low-frequency demodulation signal and output the logarithm, the host module collects the output voltage of the decibel conversion module through the analog-to-digital conversion module, obtains the decibel value intensity of the current high-frequency acoustic signal through calculation, records the change trend of the decibel value of the high-frequency acoustic signal generated by the rotating machinery, and pre-judges the running state of the rotating machinery.

[0015] Further improvement of the technical scheme of the application is that the high-frequency acoustic sensor adopts a piezoelectric ceramic bending vibration mode design, the communication module adopts a USB or WIFI wireless data transmission mode, the WIFI module adopts an ESP8266 or ESP32 module, and the host module adopts an ARM chip or an FPGA chip.

[0016] A rotating machinery early fault monitoring method based on high-frequency acoustic waves has the following steps:

[0017] Step one: extracting, fusing, training and constructing a fault classification diagnosis model through a fault database;

[0018] Step two: setting and calibrating the monitoring system by the user to ensure that the demodulation signal is not distorted and the center frequency is within 5 kHz;

[0019] Step three: recording the decibel value of the high-frequency acoustic signal for judging the running state of the current rotating machinery, and further analyzing the received high-frequency acoustic demodulation data when the decibel value exceeds the set threshold;

[0020] Step four: importing the current high-frequency acoustic demodulation signal data feature fusion result into the fault classification diagnosis model by the host computer to perform fault diagnosis, and issuing a fault alarm and prompting the fault type if the rotating machinery early fault is continuously identified.

[0021] Further improvement of the technical scheme of the application is that the fault feature extraction method in the step one of constructing the fault classification diagnosis model is:

[0022] The inherent feature representation of the original signal is obtained in a local feature learning manner, the high-frequency acoustic demodulation signal in the fault database is equally divided into frames, and the system identification parameters and sparse filtering feature learning are performed on each small sample segment in view of the nonlinear and dynamic time-varying characteristics of the high-frequency acoustic demodulation signal.

[0023] The echo state network in the recurrent neural network is used for system identification of acoustic induction signals, the output weight of the network is used as a characteristic variable, and the characteristic space dimension is consistent with the size of the reservoir network.

[0024] Sparse filtering feature learning is performed on each small sample segment, and for each small sample segment in the training set, corresponding features are obtained through feature learning of the sparse filtering network The jth feature component of the ith sample is represented, R is a vector space, N out The number of learned sparse features is N, and the sparse features of the training set X are obtained through sparse filtering feature learning on the whole training set X J is the number of sample segments, and the final feature expression is obtained by averaging the sparse features:

[0025]

[0026] In the formula, T represents matrix transposition.

[0027] The framed signal is decomposed by the CEEMDAN method, various scale IMF components, i.e., intrinsic mode functions, are obtained by autonomously adding various white noises and averaging, and the first L layers with high energy value information are selected for the next decomposition.

[0028] The first L layers of the decomposed components are calculated for nonlinear complexity, the system identification parameters, sparse filtering, and nonlinear complexity multi-domain fault feature extraction results are fused to construct a high-dimensional fault feature set.

[0029] The further improvement of the technical scheme of the application is that the fault feature fusion method in the step one of constructing the fault classification diagnosis model is:

[0030] In order to reveal the complementary properties between different features and eliminate the redundancy between features, the deep auto-encoding network structure is used to perform layer-by-layer encoding and decoding reconstruction learning on the multi-feature input, obtain the common and complementary features between different modal features, obtain the optimized feature space through the intermediate coding layer, and obtain the low-dimensional, compact and sensitive optimized features for early fault of the equipment.

[0031] The further improvement of the technical scheme of the application is that the fault feature training method in the step one of constructing the fault classification diagnosis model is: the learning method based on a support vector machine is used to establish a training model in an HPO parameter global optimization LSSVM mode, and the low-dimensional fault feature set is input into the training model.

[0032] Initialize the population; input the population size, the maximum iteration times and the adjustment parameter; set the range value of the regularization parameter γ (gam) and the square bandwidth δ2 (sig2) in the Gaussian RBF kernel, and the position of each member in the initial population is randomly generated in the search space by the following formula;

[0033] x i = rand(1, d) * *(ub - Ib) + Ib

[0034] In the formula, x i is the position of the hunter or the prey, lb is the minimum value of the solving variable, ub is the maximum value of the solving variable, and d is the dimension of the problem variable; rand(1, d) represents creating a 1xd vector composed of a random array.

[0035] By adjusting the two parameters γ and δ2, the LSSVM prediction model with the Gaussian RBF kernel is established, and the samples are trained;

[0036] Model = trainlssvm ({x train , y train , 'c', γ, δ2, 'RBF-kernel'})

[0037] In the formula, x train is the training set data, y train is the training set label, c represents classification by LSSVM, and RBF-kernel is the Gaussian RBF kernel.

[0038] The test set is used as the input of the model, in each iteration, the predicted label y of the test set is obtained, and then the fitness function value of each search agent is calculated by the least square method.

[0039]

[0040] In the formula, m is the sample number of the test set, y is the predicted label of the test set, is the actual label of the test set;

[0041] According to the fitness function, the optimal solution of the population is updated to obtain the minimum training loss value of the HPO optimization model, and then the optimal prediction value of the parameters γ and δ2 is determined;

[0042] The optimal γ and δ2 are brought into the LSSVM model for final training to obtain the early fault classification and diagnosis model of the rotating machinery.

[0043] With the technical scheme, the technical progress is achieved as follows: the card type structure design is adopted, the number of channels can be expanded for different rotating machinery systems and different monitoring requirements, and multiple sites of the rotating machinery system can be monitored simultaneously; the system can adjust the gain and carrier frequency according to different high-frequency acoustic signal strength and frequency range generated by early faults of different rotating machinery systems, and switch the working state.

[0044] The amplitude control unit is designed to have a step-by-step amplification structure, so that the signal amplification multiple is high, the frequency response is flat, and the distortion is small; the high-frequency acoustic signal is amplified and then attenuated, which can effectively reduce the circuit board noise and improve the signal signal-to-noise ratio.

[0045] The high-frequency acoustic signal is demodulated and down-converted by using the heterodyne down-conversion principle, the rotating machinery early fault is monitored and identified by using the high-frequency acoustic demodulation signal, the requirement for the system hardware is reduced, the signal signal-to-noise ratio is improved, and the data storage resource is saved.

[0046] In view of the nonlinear and dynamic time-varying characteristics of the high-frequency acoustic demodulation signal, multiple signal processing methods are fused to obtain multiple features representing the equipment health state from different analysis angles, meanwhile, in order to eliminate the redundancy among the features, the deep network is used to obtain low-dimensional and compact optimized features sensitive to the early fault of the equipment through the powerful feature learning and modeling capability, and finally the LSSVM parameter optimization method is used for fault diagnosis, so that the rotating machinery fault monitoring precision is greatly improved under the industrial strong noise and complex environment, and the whole system can be flexibly applied to various industrial rotating machinery fault monitoring environments, and has wide practicability. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor;

[0048] Figure 1 is a structural schematic diagram of the monitoring system of the present application;

[0049] Figure 2 is a working flow chart of the monitoring system of the present application;

[0050] Figure 3 is a monitoring method flow chart of the present application;

[0051] Figure 4 is a rotating machinery early fault diagnosis model training flow chart of the present application;

[0052] Figure 5This is a schematic diagram of the early fault warning process for rotating machinery according to the present invention;

[0053] Figure 6 This is a schematic diagram of the industrial field application of the early fault monitoring system for rotating machinery of the present invention. Detailed Implementation

[0054] The present invention will be further described in detail below with reference to embodiments:

[0055] Example 1:

[0056] like Figure 1 As shown, the present invention provides an early fault monitoring system for rotating machinery based on high-frequency sound waves, including a high-frequency sound wave sensor, a circuit system, and a host computer connected in sequence.

[0057] The circuit system adopts a plug-in card structure design, consisting of multiple signal conditioning boards and one communication control board. The signal conditioning boards include: an amplitude control unit, a state switching module, a mixer module, a filter amplification module, and a decibel conversion module; the communication control board includes an analog-to-digital conversion module, a main control module, a digital-to-analog conversion module, and a communication module; the communication control board communicates with the host computer and receives commands through the communication module.

[0058] The high-frequency acoustic wave sensor is used to acquire high-frequency acoustic wave signals generated by early faults in rotating machinery. It is designed using the bending vibration mode of piezoelectric ceramics. The sensor is equipped with threaded holes and accessories such as a fixed base and a magnetic base. Non-invasive installation methods such as magnetic attraction, screws, and adhesives can be selected.

[0059] For the signal conditioning board, the amplitude control unit receives a gain control signal from the main control module to control the amplitude of the output signal from the high-frequency acoustic wave sensor. The state switching module adjusts the gain of the primary and secondary amplification modules according to the strength of the acquired signal to avoid signal distortion. The mixer module demodulates the high-frequency acoustic wave signal; the filtering amplification module filters the signal output from the mixer module, retaining only the low-frequency signal components; and the decibel conversion module quantizes the demodulated analog signal from the high-frequency acoustic wave output from the filtering amplification module, converting it into a corresponding decibel value.

[0060] For the communication control board, the main control module, as the main processor, receives parameter setting commands from the host computer and transmits them downwards. On the other hand, it encapsulates the multi-channel high-frequency acoustic demodulation signals and decibel signals collected by the analog-to-digital conversion module into data frames and transmits them upwards. The digital-to-analog conversion module is used to convert gain control commands into corresponding voltages to control the gain of the amplitude control unit. The communication module is used to realize communication between the main control module and the host computer, and to transmit data and commands.

[0061] The amplitude control unit comprises a primary amplification module, a voltage-controlled amplification module and a secondary amplification module.

[0062] The primary amplification module performs primary amplification on the high-frequency acoustic wave signal output by the high-frequency acoustic wave sensor, and the gain is controlled by the state switching module; the voltage-controlled amplification module adopts a voltage-controlled amplifier to realize attenuation of the output signal of the primary amplification module, and the attenuation multiple is controlled by the level output by the digital-analog conversion module of the main control module, so that attenuation of 0-70 decibels can be realized; the secondary amplification module is used for further amplifying the output signal of the voltage-controlled amplification module, and the gain is also controlled by the state switching module.

[0063] The amplitude control unit adopts a step-by-step amplification structure, so that the high-frequency acoustic wave sensing signal has high amplification multiple, flat frequency response and small distortion; the high-frequency acoustic wave signal is amplified first and then attenuated, so as to reduce the noise of the circuit board and improve the signal-to-noise ratio.

[0064] The state switching module adopts an analog switch to control the resistance value of the voltage amplifier feedback loop of the primary amplification module and the secondary amplification module to realize gain control.

[0065] The state switching module is controlled by the IO port level of the main control module, and the two state switching modules are synchronously switched; when collecting a low-amplitude high-frequency acoustic wave signal, the state switching module switches the primary amplification module to high gain, and the secondary amplification module is switched to low gain; when collecting a high-amplitude high-frequency acoustic wave signal, the state switching module switches the primary amplification module to low gain, and the secondary amplification module is switched to high gain; the overall gain of the primary amplification module and the secondary amplification module remains dynamic balance, which better copes with different application scenarios and prevents signal distortion.

[0066] The state switching module switches according to the attenuation multiple of the voltage-controlled amplifier. When the attenuation multiple of the voltage-controlled amplifier is small, such as 0-60 decibels, it indicates that the current high-frequency acoustic wave signal amplitude is weak, so the primary amplification module needs to have high gain, and the state switching module switches the primary amplification module to high gain and the secondary amplification module to low gain; when the attenuation multiple of the voltage-controlled amplifier is large, such as 60-70 decibels, it indicates that the current high-frequency acoustic wave signal amplitude is strong, and if the primary amplification module is still in high gain, the high-frequency acoustic wave sensing signal will be distorted after amplification, so the state switching module needs to switch the primary amplification module to low gain and the secondary amplification module to high gain.

[0067] The mixer module realizes demodulation of the high-frequency acoustic signal through the heterodyne frequency reduction principle, and the low-frequency components in the signal are reserved through the filter and amplifier module. In the embodiment, the filter and amplifier module adopts a Sallen-Key low-pass filter, and the cutoff frequency is set to 5 kHz. Compared with the vibration signal, the signal after demodulation has a higher signal-to-noise ratio and is more sensitive to the lubrication state and early wear of the equipment.

[0068] The mixer module is built-in with a local oscillator, and the oscillator adopts a voltage-controlled oscillator, which can adjust the oscillation frequency according to the output level of the digital-to-analog conversion module, so as to realize demodulation of high-frequency acoustic signals with different center frequencies generated by rotating machinery faults.

[0069] The filter and amplifier module has a voltage amplification effect, which compensates for the signal attenuation after demodulation of the high-frequency acoustic signal.

[0070] The decibel value conversion module adopts an RMS-DC converter to perform root mean square measurement on the low-frequency demodulated signal and output a logarithm. The main control module collects the output voltage of the decibel value conversion module through the analog-to-digital conversion module, and can calculate the decibel value intensity of the current high-frequency acoustic signal. By recording the change trend of the decibel value of the high-frequency acoustic signal generated by the rotating machinery under the current working condition, the running state of the rotating machinery can be determined.

[0071] The communication module can adopt USB or WIFI wireless data transmission mode.

[0072] When the upper computer adopts WIFI to communicate with the main control module, the WIFI communication module adopts ESP8266 or ESP32 module.

[0073] The main control module can adopt an ARM chip or an FPGA chip.

[0074] In the embodiment, the upper computer unit is a desktop computer, a notebook computer or other computers that can run a graphical operating system.

[0075] In the interface of the upper computer module in the embodiment, the communication mode, the gain of each channel and the carrier frequency are settable items. The current high-frequency acoustic demodulation signal time domain and frequency domain waveform diagram, high-frequency acoustic signal decibel value and other information can be displayed in real time. If the decibel value exceeds the set threshold, the signal time domain index, frequency domain index, nonlinear complexity and fault classification diagnosis results can also be displayed.

[0076] As shown in Figure 2 The working process of the rotating machinery early fault monitoring system is shown in the schematic diagram. The working process includes the following steps:

[0077] S1: Start.

[0078] S2: set the system data communication mode and the decibel value threshold.

[0079] S3: the system collects, amplifies, demodulates, converts and packages the high-frequency sound wave demodulation signal into a data frame and uploads it to the upper computer.

[0080] S4: the upper computer judges the signal collection quality. If the signal collection quality (amplitude and frequency) is not good, the system gain and carrier frequency are adjusted according to the corresponding situation of the collected signal to ensure that the demodulation signal is not distorted and the center frequency is within 5 kHz.

[0081] S5: record the current high-frequency sound wave decibel value, compare it with the historical high-frequency sound wave decibel value, judge whether it exceeds the set threshold, if the high-frequency sound wave decibel value exceeds the set threshold, further fault diagnosis analysis is performed on the uploaded high-frequency sound wave demodulation signal data, according to the fault diagnosis classification result, whether early fault occurs is judged, if early fault occurs during continuous diagnosis, the fault type is indicated and alarm is issued; otherwise, the system continues to monitor.

[0082] S6: end.

[0083] As shown in Figure 3 , the present application provides a kind of based on high-frequency sound wave's rotating machinery early fault monitoring method, comprising the following steps:

[0084] Step one: extract, fuse, train and build fault diagnosis model by fault database.

[0085] Step two: user sets and calibrates the monitoring system to ensure that the demodulation signal is not distorted and the center frequency is within 5 kHz.

[0086] Step three: record the high-frequency sound wave signal decibel value of the current rotating machinery running state, if it exceeds the set threshold, further analyze the received high-frequency sound wave demodulation data.

[0087] Step four: the upper computer imports the current high-frequency sound wave demodulation signal data feature fusion result into the fault classification diagnosis model for fault diagnosis, and if early fault of rotating machinery is continuously identified, fault alarm is issued and fault type is prompted.

[0088] As shown in Figure 4 , in step one, the monitoring system collects high-frequency sound wave demodulation data under different fault state conditions of rotating machinery bearing at a sampling frequency of 24 kHz, labels each data sample with fault label, and builds a database.

[0089] The high-frequency acoustic demodulation data of each label of the rotating machinery, such as normal, under-lubrication, over-lubrication, lubricating oil pollution, and micro-abrasion, collected above is equalized and framed, and the internal feature representation of the original signal is acquired by using the local feature learning mode. In this embodiment, 40s of one type of state data are divided into 320 frames.

[0090] The high-frequency acoustic demodulation signal has obvious nonlinearity and dynamic time-varying characteristics, and the system identification parameters, sparse filtering, and multi-scale decomposition of each small sample segment are performed.

[0091] The traditional autoregressive-based identification method is based on a linear autoregressive model to perform system identification on the signal, and cannot accurately describe the dynamic time-varying non-linear high-frequency acoustic demodulation signal.

[0092] The echo state network in the recurrent neural network is used to perform system identification on the acoustic induction signal, the output weight of the network is used as a feature variable, and the feature space dimension is consistent with the size of the reservoir network.

[0093] The learning process conforms to the following convex optimization process:

[0094]

[0095] In the formula, W x , W in , and w are the internal weight matrix, input weight matrix, and output weight matrix of the reservoir, respectively, tanh(·) is an activation function, y is a reservoir state variable, and u is a reservoir input variable, that is, a feature parameter related to the state variable y in the feature information space MOR.

[0096] In addition, sparse filtering feature learning is performed on each small sample segment. For each small sample segment in the training set, the corresponding feature is obtained through the sparse filtering network feature learning, where x out is the jth feature component of the ith sample, R is a vector space, N is the number of learned sparse features, and X

[0097]

[0098] In the formula, T represents matrix transposition.

[0099] Through averaging, the features will be more representative and discriminative.

[0100] The signals after framing are decomposed by the CEEMDAN method, and various scale IMF (Intrinsic Mode Function) components are obtained by adding various white noises autonomously and averaging.

[0101] In this embodiment, the first 6 layers with higher energy value information are selected for the next step of decomposition; the permutation entropy, wavelet entropy, sample entropy and other nonlinear complexity of the first 6 layers after decomposition are calculated.

[0102] The system identification parameters, sparse filtering and nonlinear complexity fault feature extraction results are normalized and a high-dimensional fault feature set is constructed;

[0103] In order to reveal the complementary properties between different features and eliminate the redundancy between features, the deep self-encoding network structure is preferred to be used for encoding and decoding reconstruction learning of multi-feature input layer by layer, to obtain common and complementary features between different modal features, and to obtain low-dimensional and compact optimized features sensitive to early equipment faults through the intermediate coding layer.

[0104] Based on the learning method of support vector machine, the HPO parameter global optimization LSSVM is used to establish the training model, and the low-dimensional fault feature set is input into the training model; the predator algorithm is used to optimize the LSSVM model structure, and the specific steps are as follows:

[0105] Initialize the population. Input the population number, the maximum number of iterations and the adjustment parameter; set the range value of the regularization parameter γ (gam) and the square bandwidth δ2 (sig2) in the Gaussian RBF kernel. The position of each member in the initial population is randomly generated in the search space by the following formula.

[0106] x i =r a n d(1,d)·*(ub-Ib)+Ib

[0107] In the formula, x i is the position of the hunter or the prey, lb is the minimum value of the solution variable, ub is the maximum value of the solution variable, d is the dimension of the problem variable, and rand(1,d) represents creating a 1xd vector composed of random arrays.

[0108] By adjusting the two parameters γ and δ2, an LSSVM prediction model with a Gaussian RBF kernel is established to train the samples.

[0109] Model=trainlssvm({x train ,y train ,'c',γ,δ2,'RBF-kernel'})

[0110] In the formula, x train is the training set data, y train is the training set label, c represents classification by LSSVM, and RBF-kernel is a Gaussian RBF kernel.

[0111] The test set is used as the input of the model, and in each iteration, the prediction set y of the test sample is obtained, and then the fitness function value of each search agent is calculated by the least square method.

[0112]

[0113] In the formula, m is the number of samples of the test set, y is the predicted label of the test set, is the actual label of the test set;

[0114] The HPO parameters are adjusted according to the fitness function, the optimal solution of the population is updated, the minimum training loss value of the HPO optimization model is obtained, and then the optimal prediction values of the gamma and delta2 parameters are determined.

[0115] The optimal gamma and delta2 are brought into the LSSVM model for final training, and the early fault monitoring model of the rotating machinery is obtained.

[0116] As Figure 5 shown, the fault early warning refers to the decibel value characteristics of the high-frequency acoustic signal on one hand, and introduces the high-frequency acoustic demodulation signal into the fault diagnosis model to diagnose the early fault type on the other hand.

[0117] As an example, Figure 6 shown is an application schematic diagram of the early fault monitoring system of the rotating machinery in an industrial field, the early fault monitoring system of the rotating machinery can collect signals in real time through the high-frequency acoustic sensor deployed in the industrial field environment for the key components of the rotating machinery and feed back to the monitoring room; the monitoring room can be in communication connection with the early fault monitoring system of the rotating machinery deployed in the industrial field environment; on one hand, the early fault of the rotating machinery is monitored and early warned, and on the other hand, the monitoring data is regularly uploaded to the server; the server generates a decibel value record table and expands the data set, trains the model according to the rotating machinery fault diagnosis model training method, and migrates the trained model to the monitoring room, and regularly updates the fault diagnosis model in the early fault monitoring system of the rotating machinery.

[0118] The application has the advantages of high integration, low hardware requirement, expandable channel number, wide frequency detection range, saving of data storage resources, etc., can improve the fault diagnosis precision in an industrial strong noise and complex environment, and realizes preventive maintenance of the early fault of the rotating machinery.

[0119] The above-described embodiments are merely intended to describe the preferred embodiments of the present application, and are not intended to limit the scope of the present application. Various modifications and improvements to the present application made by those skilled in the art are intended to fall within the scope of the present application defined in the claims.

Claims

1. A high-frequency acoustic wave-based rotating machinery early fault monitoring system, comprising a high-frequency acoustic wave sensor, a circuit system and an upper computer connected in sequence, characterized in that: The circuit system adopts a plug-in card structure design and is composed of multiple signal conditioning board cards with the same structure and a communication control board card. The signal conditioning board card comprises an amplitude control unit, a mixer module, a filter amplification module, a decibel conversion module and a state switching module. The communication control board card comprises a main control module, an analog-digital conversion module, a digital-analog conversion module and a communication module. The host computer communicates data and sends instructions between the communication module and the communication control board card. The amplitude control unit comprises a primary amplification module, a voltage-controlled amplification module and a secondary amplification module connected in sequence. The primary amplification module performs primary amplification gain on the high-frequency acoustic wave signal output by the high-frequency acoustic wave sensor, and the gain is controlled by the state switching module controlled by the main control module. The voltage-controlled amplification module adopts a voltage-controlled amplifier to realize attenuation of the output signal of the primary amplification module, and the attenuation multiple is controlled by the level output by the digital-analog conversion module controlled by the main control module.

2. The system for early fault monitoring of rotating machinery based on high frequency acoustic waves as claimed in claim 1 wherein: The secondary amplification module is used for further amplifying the output signal of the voltage-controlled amplification module, and the gain is also controlled by the state switching module controlled by the main control module.

3. A method for early fault monitoring of rotating machinery based on high frequency acoustic waves realized by the monitoring system according to claim 1 or 2, characterized in that, The main control module controls the two state switching modules to be synchronously switched through the IO port level according to the current attenuation multiple of the voltage-controlled amplifier. When collecting low-amplitude high-frequency acoustic wave signals, the state switching module switches the primary amplification module to high gain, and the secondary amplification module is switched to low gain. When collecting high-amplitude high-frequency acoustic wave signals, the state switching module switches the primary amplification module to low gain, and the secondary amplification module is switched to high gain. The overall gain of the primary amplification module and the secondary amplification module remains dynamic balance. The mixer module realizes demodulation of the high-frequency acoustic wave signal processed by the amplitude control unit through heterodyne frequency reduction principle. The demodulated high-frequency acoustic wave signal is filtered by the filter amplification module to retain the low-frequency components in the signal, and the high-frequency acoustic wave signal is reduced to below 20 kHz. The mixer module has a built-in local oscillator. The oscillator adopts a voltage-controlled oscillator that can adjust the oscillation frequency according to the level output by the digital-analog conversion module to realize demodulation of high-frequency acoustic wave signals with different center frequencies generated by rotating machinery faults. The decibel conversion module adopts an RMS-DC converter to measure the root mean square of the low-frequency demodulated signal and output the logarithm. The main control module collects the output voltage of the decibel conversion module through the analog-digital conversion module, calculates the decibel value intensity of the current high-frequency acoustic wave signal, and records the change trend of the decibel value of the high-frequency acoustic wave signal generated by the rotating machinery to predict the operating state of the rotating machinery. The high-frequency acoustic wave sensor adopts a piezoelectric ceramic bending vibration mode design. The communication module adopts USB or WIFI wireless data transmission mode. The WIFI module adopts ESP8266 or ESP32 module. The main control module adopts ARM chip or FPGA chip. The steps are as follows: Step one: extract, fuse, train and build a fault classification diagnosis model through a fault database. In step one, the fault feature extraction method in the construction of the fault classification diagnosis model is: The internal feature representation of the original signal is obtained by local feature learning, and the high-frequency acoustic demodulation signals in the fault database are equalized and framed. In view of the nonlinear and dynamic time-varying characteristics of the high-frequency acoustic demodulation signals, the system identification parameters and sparse filtering feature learning are performed on each small sample segment; The echo state network in the recurrent neural network is used to perform system identification on the acoustic induction signal, and the output weight of the network is used as a feature variable. The dimension of the feature space is consistent with the size of the reservoir network; Sparse filter feature learning is performed on each small sample segment, and for each small sample segment in the training set, corresponding features are obtained through feature learning of the sparse filter network , represents the jth feature component of the ith sample, R is a vector space, is the number of learned sparse features, and the entire training set is subjected to sparse filter feature learning to obtain sparse features of the training set , J is the number of sample segments, and the final feature expression is obtained by averaging the sparse features. ； In the formula, T represents the matrix transpose; The framed signals are decomposed by the CEEMDAN method, and various scale IMF components, i.e. intrinsic mode functions, are obtained by adding various white noises autonomously and averaging; The first L layers with high energy value information are selected for the next decomposition. The nonlinear complexity of the decomposed first L layers is calculated, and the system identification parameters, sparse filtering and nonlinear complexity multi-domain fault feature extraction results are fused to construct a high-dimensional fault feature set; The fault feature fusion method in the fault classification and diagnosis model constructed in step one is: In order to reveal the complementary properties between different features and eliminate the redundancy between features, the deep auto-encoding network structure is used to perform layer-by-layer encoding and decoding reconstruction learning on multi-feature input, obtain the common and complementary features between different modal features, and obtain the low-dimensional and compact optimized features sensitive to early equipment faults through the intermediate coding layer. The fault feature training method in the fault classification and diagnosis model constructed in step one is: based on the learning method of support vector machine, the HPO parameter global optimization LSSVM is used to establish the training model, and the low-dimensional fault feature set is input into the training model; The predator algorithm is used to optimize the LSSVM model structure, including: Initialize population; input population quantity, maximum iteration times and adjustment parameter; set regular parameter range value and square bandwidth in Gaussian RBF kernel range value, the position of each member in initial population is randomly generated in search space by the following formula; ； wherein is the position of the hunter or prey, lb is the minimum value of the solution variable, ub is the maximum value of the solution variable, d is the dimension of the problem variable, and rand(l,d) represents creating a 1 x d vector consisting of a random array; By adjusting and Two parameters, the establishment of LSSVM prediction model with Gaussian RBF kernel, training samples; ； wherein is the training set data, is the training set label, denotes classification with LSSVM, is the Gaussian RBF kernel; The test set is used as input to the model, and in each iteration, the predicted labels for the test set are obtained The fitness function value of each search agent is then calculated using the least squares method; ； In the formula, is the number of samples in the test set, is the predicted label for the test set, is the actual label for the test set; According to the fitness function, the HPO parameters are adjusted, the optimal solution of the population is updated, the minimum training loss value of the HPO optimization model is obtained, and then the optimal prediction value of the parameters is determined and parameters. The optimal and into the LSSVM model for final training to obtain the rotating machinery early fault classification diagnosis model; Step 2: The user sets and calibrates the monitoring system to ensure that the demodulation signal is not distorted and the center frequency is within 5 kHz; Step 3: Record the decibel value of the high-frequency acoustic signal for judging the running state of the rotating machinery. When the value exceeds the set threshold, further analyze the received high-frequency acoustic demodulation data; Step 4: The host computer imports the feature fusion results of the current high-frequency acoustic demodulation signal into the fault classification and diagnosis model for fault diagnosis. If the early fault of the rotating machinery is continuously identified, a fault alarm is issued and the fault type is prompted.

Images