Loader gearbox fault diagnosis method based on typical working conditions

Abstract

The invention provides a loader gearbox fault diagnosis method based on typical working conditions. The method comprises the following steps: acquiring an operation signal and a vibration signal in the operation process of a gearbox; the operation signal comprises a rotating speed signal and a torque signal. The optimized operation signal and the optimized vibration signal are obtained through data preprocessing. And working condition screening is carried out on the optimized operation signal and the optimized vibration signal to obtain an operation signal under a typical working condition and a vibration signal under the typical working condition. Respectively performing feature extraction on the operation signal under the typical working condition and the vibration signal under the typical working condition to obtain a first feature set and a second feature set; and reconstructing the first feature set and the second feature set to form a feature parameter set. And inputting the characteristic parameter set into the width learning model to obtain the fault type of the gearbox. According to the method, real-time monitoring of the gearbox of the engineering machinery can be realized, the fault type of the gearbox under the non-stable working condition can be accurately realized, and predictive maintenance of the engineering machinery can be realized.

Description

Technical Field

[0001] The present invention relates to the fields of engineering machinery, big data and artificial intelligence, and in particular to a loader gearbox fault diagnosis method based on typical working conditions. Background Art

[0002] Construction machinery is widely used in engineering and construction fields such as construction, water conservancy, electricity, transportation and mining. As an important part of the equipment industry, the development level of construction machinery can measure the degree of industrialization of a country. As a relatively mature construction machinery market, my country has a very complete range of products, and the output of many types of products ranks first in the world, and has formed large-scale industrial clusters in many places. By the end of 2022, the number of construction machinery in my country has exceeded 9 million units, and the industry's revenue has exceeded 800 billion yuan.

[0003] As a key power transmission component of the loader, the gearbox is prone to tooth breakage, wear and pitting at key locations due to its large transmission ratio and strong load-bearing capacity, which in turn causes local failures and affects equipment reliability and safety. Therefore, effective fault warning for engineering machinery gearboxes can significantly improve equipment operating efficiency and reduce losses caused by downtime.

[0004] However, since engineering machinery gearboxes are in complex working conditions of high speed and variable load for a long time, and have complex internal structures, strong coupling characteristics, and multiple complex faults, it is difficult to ensure the reliability and accuracy of fault diagnosis using traditional methods. Summary of the invention

[0005] The purpose of the present invention is to provide a loader gearbox fault diagnosis method based on typical working conditions, which can realize real-time monitoring of the engineering machinery gearbox, so as to accurately realize the fault diagnosis under non-stable working conditions of the gearbox, accurately locate the fault type, and realize predictive maintenance of the engineering machinery.

[0006] To achieve the above object, the present invention provides a loader gearbox fault diagnosis method based on typical working conditions, the loader gearbox fault diagnosis method comprising the following steps:

[0007] S1: Acquire an operation signal and a vibration signal during the operation of the gearbox; the operation signal includes a speed signal and a torque signal;

[0008] S2: obtaining optimized operation signal and optimized vibration signal through data preprocessing;

[0009] S3: screening the optimized operation signal under working conditions to obtain an operation signal under a typical working condition; screening the optimized vibration signal under working conditions to obtain a vibration signal under a typical working condition;

[0010] S4: extracting features from the operating signal under typical working conditions to obtain a first feature set; extracting features from the vibration signal under typical working conditions to obtain a second feature set;

[0011] S5: reconstructing the first feature set and the second feature set to form a feature parameter set;

[0012] S6: Inputting the characteristic parameter set into a width learning model to obtain the fault type of the gearbox.

[0013] Optionally, a plurality of sensors are arranged on the surface of the gear box, and the operation signal and the vibration signal during the operation of the gear box are acquired through the sensors.

[0014] Optionally, the operation signal also includes at least one of a pressure signal and a temperature signal.

[0015] Optionally, in S1, the step of obtaining an operation signal and a vibration signal during the operation of the loader gearbox includes:

[0016] The slowly varying time sequence signal of each of the operation signals is obtained respectively, and all the slowly varying time sequence signals are constructed into a first array; the vibration signal is obtained, and all the vibration signals are constructed into a second array.

[0017] Optionally, in S2, the step of obtaining an optimized operation signal through data preprocessing includes:

[0018] All the slowly varying time series signals are subjected to low-pass filtering respectively, and normalized respectively to obtain normalized slowly varying time series signals;

[0019] In S2, the step of obtaining the optimized vibration signal through data preprocessing includes:

[0020] The vibration signal is subjected to VMD decomposition to obtain multiple modal components; the center frequency of each modal component is calculated; and the modal components corresponding to all the center frequencies are reconstructed to obtain the denoised vibration signal.

[0021] Optionally, in S2, the step of obtaining an optimized operation signal and an optimized vibration signal through data preprocessing further includes:

[0022] The normalized slowly varying time series signal and the denoised vibration signal are time stamp aligned, and the normalized slowly varying time series signal and the denoised vibration signal are short-time-series cut at equal intervals to obtain the fragmented operation signal and the fragmented vibration signal.

[0023] Optionally, in S3, the step of performing working condition screening on the optimized operation signal and the optimized vibration signal respectively to obtain the operation signal under typical working conditions and / or the vibration signal under typical working conditions comprises:

[0024] Extracting operation timing features from the fragmented operation signals; fusing the operation timing features with the gear information of the aligned timestamp to obtain operation characteristic information under different gears; manually marking the operating condition type to which each operation characteristic information belongs; substituting all the operation characteristic information and the operating condition type to which each operation characteristic information belongs into the Bayesian model for screening to obtain the operation signal under typical working conditions;

[0025] And / or, extracting vibration time series features from the fragmented vibration signal; fusing the vibration time series features with the gear information of the aligned timestamp to obtain vibration feature information under different gears; manually marking the working condition type to which each vibration feature information belongs; substituting all the vibration feature information and the working condition type to which each vibration feature information belongs into the Bayesian model for screening to obtain the vibration signal under typical working conditions;

[0026] The operation timing characteristics and / or the vibration timing characteristics include one or more of a mean value, a peak-to-peak value, an effective value, a variance, and a kurtosis value.

[0027] Optionally, in S4, the step of extracting features from the operating signal under the typical working condition to obtain the first feature set includes:

[0028] Inputting the operating signal under typical working conditions into a sparse autoencoder for feature extraction to obtain the first feature set;

[0029] In S4, the step of extracting features from the vibration signal under typical working conditions to obtain the second feature set includes:

[0030] Performing Fourier transform on the vibration signal under typical working conditions to obtain a time-frequency spectrum of the vibration signal; inputting the time-frequency spectrum into a convolutional neural network for feature extraction to obtain the second feature set.

[0031] Optionally, the step of extracting features through the convolutional neural network includes:

[0032] The convolutional neural network is set as a deep residual network; the time-frequency spectrum is split, and the split time-frequency spectrum is input into the deep residual network for feature extraction to obtain the second feature set.

[0033] Optionally, in S6, the step of inputting the characteristic parameter set into a width learning model to obtain the fault type of the gearbox includes:

[0034] The width learning model is trained by using multiple sets of labeled feature parameter sets to obtain the trained width learning model, which includes multiple fault types of the gearbox: the feature parameter set during the operation of the gearbox is obtained, the feature parameter set is input into the trained width learning model, and the fault type of the gearbox is obtained.

[0035] The present invention fully considers the complex internal structure, strong coupling characteristics, multiple complex faults, and long-term high speed and variable load complex working conditions of the engineering machinery gearbox, and proposes a loader gearbox fault diagnosis method based on typical working conditions. The method first obtains multi-source heterogeneous data from the non-stationary process of the engineering machinery, and the multi-source heterogeneous data can fully reflect the operating status of the gearbox. Then, the multi-source heterogeneous data is screened and classified for typical working conditions, and then the typical working condition data set of the gearbox (including torque signal, speed signal and vibration signal) is obtained. Then, for the typical working condition data, a multimodal data fusion algorithm is introduced to realize the feature fusion of the typical working condition data set, and the width learning algorithm is used to quickly and accurately perform precise fault diagnosis on the gearbox.

[0036] This method can realize effective fault diagnosis during the operation of engineering machinery, accurately locate the fault type of the gearbox, realize predictive maintenance of engineering machinery, reduce operation and maintenance costs, improve operation and maintenance efficiency, and avoid losses caused by gearbox downtime. Since this method fully considers the actual operating characteristics of engineering machinery, it has important research significance and further exploration value in the practical application of engineering machinery. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A step diagram of a method for diagnosing a loader gearbox fault under typical working conditions in a preferred embodiment of the present invention;

[0038] Figure 2 A time-frequency spectrum of a gearbox vibration signal in a preferred embodiment of the present invention;

[0039] Figure 3 A schematic diagram of a network infrastructure in a deep residual network in a preferred embodiment of the present invention;

[0040] Figure 4 A schematic diagram of the network architecture of a deep residual network ResNet18 in a preferred embodiment of the present invention;

[0041] Figure 5 It is a structural diagram of a width learning model in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0042] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The advantages and features of the present invention will become more apparent from the following description. It should be noted that the accompanying drawings are in very simplified form and in non-precise proportions, and are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention.

[0043] It should be understood that, unless otherwise specified or indicated, the terms "first", "second", "third", etc. in the specification are only used to distinguish the various components, elements, steps, etc. in the specification, rather than to indicate the logical relationship or sequential relationship between the various components, elements, steps, etc. Moreover, the terms "comprise", "include", or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article, or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article, or device including the element.

[0044] The present invention is described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the absence of conflict, the following embodiments and features in the embodiments may complement or be combined with each other.

[0045] In order to achieve real-time monitoring of the gearbox of a construction machinery loader, a preferred embodiment of the present invention provides a loader gearbox fault diagnosis method based on typical working conditions, which can perform accurate fault diagnosis on the loader gearbox and provide guidance for timely maintenance.

[0046] like Figure 1 As shown, the loader gearbox fault diagnosis method includes the following steps:

[0047] S1: Obtain the operation signal and vibration signal of the gearbox during operation. The operation signal includes the speed signal and the torque signal.

[0048] S2: Obtain optimized operation signals and optimized vibration signals through data preprocessing.

[0049] S3: Screening the optimized operation signal to obtain the operation signal under the typical working condition; Screening the optimized vibration signal to obtain the vibration signal under the typical working condition.

[0050] S4: extracting features of the operating signal under typical working conditions to obtain a first feature set X1; extracting features of the vibration signal under typical working conditions to obtain a second feature set X2.

[0051] S5: Reconstruct the first feature set X1 and the second feature set X2 to form a feature parameter set X.

[0052] S6: Inputting the characteristic parameter set into a width learning model to obtain the fault type of the gearbox.

[0053] The present invention fully considers the complex internal structure, strong coupling characteristics, multiple faults, and long-term high speed and variable load complex working conditions of the gearbox of the engineering machinery, and proposes a loader gearbox fault diagnosis method based on typical working conditions. This method can realize real-time monitoring of the gearbox of the engineering machinery, and make full use of the multimodal data of the gearbox to construct the all-round characteristic information of the gearbox, and then select typical working condition information from the non-steady operation data to realize effective fault diagnosis during the operation of the engineering machinery, and can accurately locate the fault type, realize predictive maintenance of the engineering machinery, reduce operation and maintenance costs, improve operation and maintenance efficiency, and avoid losses caused by gearbox downtime. Since this method fully considers the actual operating characteristics of the engineering machinery, it has important research significance and further exploration value in the practical application of engineering machinery.

[0054] Specifically, the method first obtains multi-source heterogeneous data from the non-stationary process of engineering machinery, which can fully reflect the operating status of the gearbox. Then, the multi-source heterogeneous data is screened and classified for typical working conditions, and then the typical working condition data set of the gearbox (including torque signal, speed signal and vibration signal) is obtained. Then, for the typical working condition data, a multimodal data fusion algorithm is introduced to realize the feature fusion of the typical working condition data set, and the width learning algorithm is used to quickly and accurately perform precise fault diagnosis on the gearbox.

[0055] In a preferred embodiment, considering that the internal structure of the loader gearbox is complex, the coupling characteristics are strong, the volume is large, and the faults are complex and frequent, in order to more comprehensively understand the operating status of the gearbox, a variety of sensors can be set on the surface of the gearbox, and the operating signals and vibration signals of the gearbox during operation are obtained by the sensors, so as to comprehensively characterize the dynamic characteristics of the gearbox. The operating signals preferably include torque signals, speed signals, pressure signals, and temperature signals.

[0056] As a specific example, a speed sensor, a torque sensor, an acceleration sensor, a temperature sensor and a pressure sensor may be arranged on the surface of the gearbox to obtain a speed signal, a torque signal, a vibration signal, a temperature signal and a pressure signal during the operation of the gearbox.

[0057] When collecting information from multiple sensors, it is impossible to directly construct the data for unified processing due to the differences in the time scale and dimension of each information.

[0058] To solve the above problem, in S1, the step of obtaining the operation signal of the loader gearbox during operation preferably includes: obtaining the slowly varying time sequence signal of each operation signal respectively, and constructing all the slowly varying time sequence signals into a first array D1.

[0059] As a preferred implementation, the slowly varying timing signal includes a pressure slowly varying timing signal P, a torque slowly varying timing signal N, a speed slowly varying timing signal S and a temperature slowly varying timing signal T.

[0060] In more detail, each sensor may include one or more sensors. During the operation of the engineering machinery, multiple sensor data representing the state of the gearbox of the engineering machinery are collected in real time to obtain the operating state of the gearbox.

[0061] In a preferred example, a torque signal N, a speed signal S, p pressure signals P and t temperature signals can be collected by sensors, and a first array D1 is constructed, where D1 = [P1, ..., P p ,B1,S1,T1,…,T t ], where the subscripts represent the number of corresponding signal types.

[0062] Furthermore, in S1, the step of acquiring the vibration signal of the loader gearbox during operation preferably includes: acquiring the vibration signal of the gearbox during operation, and constructing all the vibration signals into a second array D2.

[0063] In a specific example, m vibration signals V can be collected by sensors and constructed to form a second array D2, D2 = [V1, ..., V m ], where the subscript represents the number of vibration signal types.

[0064] Furthermore, in S2, the step of obtaining the optimized operation signal by data preprocessing includes: performing low-pass filtering on all the slowly varying time series signals respectively, and performing normalization processing on them respectively to obtain the normalized slowly varying time series signals D1', D'1 = [P'1, ..., P' p ,N′1,S′1,T′1,…,T′1]. Normalization can achieve de-parameterization and improve the convergence of the model.

[0065] In more detail, the normalization process includes:

[0066] All the slow-changing time series signals after low-pass filtering are subjected to Z-score standardization respectively, so that all the slow-changing time series signals are normalized to between 0 and 1. The calculation formula is:

[0067]

[0068] Where x' is the normalized value of the slow-varying time series signal, u is the mean of the original slow-varying time series signal, σ is the standard deviation of the original slow-varying time series signal, and x is the input value of the original slow-varying time series signal.

[0069] Furthermore, in S2, the step of obtaining an optimized vibration signal through data preprocessing includes: performing VMD decomposition on the vibration signal to obtain multiple modal components, calculating the center frequency of each modal component, and reconstructing all modal components corresponding to all center frequencies to obtain a denoised vibration signal.

[0070] In more detail, the steps of performing VMD decomposition on the vibration signal include:

[0071] Get the original vibration signal f of the gearbox and set the VMD decomposition parameter range. Perform VMD decomposition on the vibration signal f of the gearbox to obtain multiple component signals. All component signals have a central frequency ω k The limited bandwidth modal component u k .

[0072] Construct a variational expression so that each modal component u k The sum of the estimated bandwidths of is minimized, subject to the constraint that all modal components u k The sum is equal to the original vibration signal f.

[0073] The expression of constrained variation is:

[0074]

[0075] In the formula, f is the original input vibration signal, t represents time, k is the number of modal components to be decomposed (a positive integer), and u k ={u1,u2,…,u k} is the function of each modal component, w k ={w1,w2,…,w k} is the center frequency of each modal component, δ(t) is the Dirac function, j is the imaginary number symbol, and * is the convolution operator.

[0076] By solving the above formula, each modal component u can be adaptively matched k The optimal center frequency ω k and finite bandwidth. Then it will have the optimal center frequency ω k The denoised vibration signal can be obtained by reconstructing the modal component function.

[0077] In addition, in S2, the step of obtaining the optimized operation signal and the optimized vibration signal through data preprocessing also includes: aligning the timestamps of the normalized slowly varying time series signal and the denoised vibration signal, and cutting the normalized slowly varying time series signal and the denoised vibration signal into short time series at equal intervals to obtain fragmented operation signals and fragmented vibration signals, thereby providing data support for the subsequent screening of typical working conditions in non-stationary signals.

[0078] Since the operating conditions of the loader switch frequently and the data characteristics under different operating conditions vary significantly, when diagnosing gearbox faults, data sets under typical operating conditions can be obtained through operating condition screening to provide high-quality data for gearbox fault diagnosis.

[0079] Furthermore, the step of screening the optimized operating signal for operating conditions to obtain the operating signal under typical operating conditions includes:

[0080] Extract the operation timing features from the fragmented operation signals. Fuse the operation timing features with the gear information of the aligned timestamp to obtain the operation characteristic information under different gears. Manually mark the operating condition type to which each operation characteristic information belongs; substitute all the operation characteristic information and the operating condition type to which each operation characteristic information belongs into the Bayesian model for screening to obtain the operation signal under typical working conditions.

[0081] Similarly, the steps of screening the optimized vibration signal under working conditions and obtaining the vibration signal under typical working conditions include:

[0082] Extract vibration time series features from fragmented vibration signals. Fuse vibration time series features with gear information of aligned timestamps to obtain vibration feature information under different gears. Manually mark the type of working condition to which each vibration feature information belongs. Substitute all vibration feature information and the type of working condition to which each vibration feature information belongs into the Bayesian model for screening to obtain vibration signals under typical working conditions.

[0083] It should be noted that the operation timing characteristics and / or vibration timing characteristics include one or more of a mean value, a peak-to-peak value, an effective value, a variance, and a kurtosis value.

[0084] This application uses the collected fragmented data and introduces the Bayesian model to filter the characteristic information for typical working conditions, which can effectively classify the working conditions of the characteristic information and filter out the typical working conditions. Filtering the characteristic information can reduce data interference, improve the quality of data screening, and simplify the calculation amount during fault diagnosis.

[0085] In more detail, the step of obtaining the operation signal under typical working conditions and / or the vibration signal under typical working conditions by screening through the Bayesian model includes:

[0086] The probability that the operating characteristic information at each specific gear position A is manually marked as the typical operating condition B is obtained through the Bayesian model. The probability that the vibration characteristic information at each specific gear position A is manually marked as the typical operating condition B is obtained through the Bayesian model.

[0087] The expression of the Bayesian model is as follows:

[0088]

[0089] Wherein, P(B / A) is the probability that the operation characteristic information or vibration characteristic information under a specific gear position A is manually marked as a typical operating condition B; P(A / B) is the probability that the operation characteristic information or vibration characteristic information under a specific gear position A appears in the manually marked typical operating condition B; P(B) / P(A) is the ratio of the operation characteristic information or vibration characteristic information in the manually marked typical operating condition B to all the operation characteristic information or all the vibration characteristic information;

[0090] Further, the step after obtaining the probability that the operating characteristic information under each specific gear position A is manually marked as the typical working condition B includes: comparing the probabilities of all working condition types obtained by the operating characteristic information, and taking the working condition type with the greatest probability as the typical working condition to which the operating characteristic information belongs. Through the above method, all operating characteristic information under typical working conditions can be obtained, and the operating characteristic information includes all operating characteristic information under different gear positions, and then the operating signal under the typical working condition can be obtained.

[0091] Similarly, the step after obtaining the probability that the vibration characteristic information under each specific gear position A is manually marked as the typical working condition B includes: comparing the probabilities of all working condition types obtained by the vibration characteristic information, and taking the working condition type with the highest probability as the typical working condition to which the vibration characteristic information belongs. Through the above method, all vibration characteristic information under typical working conditions can be obtained, and the vibration characteristic information includes all operating characteristic information under different gear positions, and then the vibration signal under the typical working condition can be obtained.

[0092] The following is a specific example to illustrate the specific steps of obtaining the operation signal under typical working conditions or the vibration signal under typical working conditions through the Bayesian model:

[0093] Assume that the gearbox has n gear information, and the characteristic information under n gears is represented by A1, A2…A n The gearbox has m types of working conditions, which are represented by B1, B2...B m .

[0094] The probability P(B1 / A1) that the characteristic information A1 (including the timing characteristic information in gear 1 or the vibration characteristic information in gear 1) belongs to the operating condition type B1 is obtained through the expression of the Bayesian model.

[0095]

[0096] P(A)=P(A1)+P(A2)+…+P(An)

[0097] In the formula, P(A1 / B1), P(B1) and P(A) are all known quantities. Similarly, the probability P(B2 / A1) that feature information A1 belongs to operating condition type B2 can be obtained through the Bayesian model… m The probability P(B m / A1).

[0098] By comparison, we can get P(B1 / A1), P(B2 / A1)…P(B m / A1). Assume that P(B x / A1) has the highest probability, then the operating condition type B x Set to the typical working condition corresponding to characteristic information A1. At this time, characteristic information A1 can be added to typical working condition B x According to the above method, all feature information A (including A1, A2...A n ) to obtain the typical working condition B x All characteristic information A in the typical working condition B can be obtained x The data set (including the operation information data set or the vibration information data set) is used to complete the working condition screening of the fragmented operation signal and the fragmented vibration signal by the Bayesian model.

[0099] Furthermore, in S4, the step of extracting features from the operating signal under typical working conditions to obtain a first feature set X1 includes:

[0100] The operating signal under typical working conditions is input into the sparse autoencoder for feature extraction to obtain the first feature set X1.

[0101] It should be noted that the sparse autoencoder needs to be trained in advance with a marked operation signal. After the sparse autoencoder is trained, the output data of the encoder is extracted to obtain the first feature set X1.

[0102] Specifically, the steps of extracting features using the sparse autoencoder algorithm include: inputting the slowly varying time series signal D1' under typical working conditions with labels into the encoder, D'1 = [P'1, ..., P' p ,N′1,S′1,T′1,…,T′1]. Among them, the encoder of sparse autoencoding is:

[0103] h=f(x):=s f (Wx+p)

[0104] In the formula, x represents the input data of the encoder, W x represents the weight matrix from the input layer to the hidden layer, p represents the bias matrix from the input layer to the hidden layer, s f represents the activation function of the encoder, and h represents the output data of the hidden layer.

[0105] s f Preferably, the Sigmod function is selected, and the Sigmod function is:

[0106]

[0107] In the formula, z represents the input data of the Sigmod function, s f (z) represents the output data of the Sigmod function. The output data of the Sigmod function s f (z) is the output data of the encoder.

[0108] Furthermore, the steps of training the sparse autoencoder include:

[0109] The labeled slowly varying time series signal is input into the encoder to obtain the encoder output data h. Then the encoder output data h is input into the decoder to obtain the output data of the output layer. The decoder used in the decoding stage is:

[0110] r=g(h):=s g (W T h+q)

[0111] Where W T is the weight matrix from the hidden layer to the output layer, W T W x The transposed matrix of , q is the bias matrix from the hidden layer to the output layer, s g is the activation function of the decoder. g Preferably, a Sigmod function is also selected.

[0112] Then, the output data of the output layer is compared with the loss through the cross entropy function and the weight matrix W is updated through the BP algorithm. x , to complete the training of the sparse autoencoder.

[0113] The cross entropy function is constructed as:

[0114]

[0115] In the formula, n is the amount of data during training, y i is the output value of the decoder output layer, x i is the tag value. In the sparse autoencoding algorithm, the tag value x i Equal to the encoder input value x.

[0116] As a preferred implementation, in S4, the step of extracting features from the vibration signal under typical working conditions to obtain the second feature set X2 includes:

[0117] The vibration signal under typical working conditions (i.e., time-domain vibration signal) is subjected to Fourier transform to obtain the time-frequency spectrum of the vibration signal. The time-frequency spectrum is input into the convolutional neural network for feature extraction to obtain the second feature set X2.

[0118] Figure 2 The time-frequency spectrum of the vibration signal in a preferred embodiment of the present invention. Since simply extracting features from the time-domain vibration signal will significantly reduce the feature information, thereby reducing the diagnostic effect. Therefore, in the application, the time-domain vibration signal can be subjected to a short-time Fourier transform to convert the time-domain vibration signal into a time-frequency spectrum, and the time-frequency spectrum can basically retain all characteristic signals in the time-domain vibration signal.

[0119] Furthermore, the step of extracting features from the time-frequency spectrum through a convolutional neural network includes:

[0120] The convolutional neural network is set to the deep residual network ResNet18. The time-frequency spectrum is split, and the split time-frequency spectrum is input into the deep residual network for feature extraction to obtain the second feature set X2.

[0121] It should be noted that in the time-frequency spectrum under typical working conditions, the significant peak information accounts for a small proportion in the time-frequency spectrum. Therefore, in a preferred case, the deep residual network can be trained first with the labeled time-frequency spectrum training data. Then, the split time-frequency spectrum is input into the trained deep residual network for feature extraction, and then the second feature set X2 is obtained.

[0122] The present invention builds a time-series-based deep residual network (ResNet18) when acquiring the second feature set X2. The main body of the network model is composed of a convolutional neural network with residual blocks, which can obtain all feature information in the image through multiple convolutional layers and output the second feature set X2. The network model can avoid the loss of feature information and improve the effect of feature extraction. In addition, since the network model includes at least one residual block, it can effectively alleviate the error in the convolution operation, ensure the effective transmission of features, improve the accuracy of acquiring feature vectors, and improve the accuracy of gearbox fault diagnosis.

[0123] For more details, see Figure 3As shown, the network infrastructure of the deep residual network uses a parallel structure containing two routes to realize feature extraction of the time-frequency spectrum. The network infrastructure has a first route L1 and a second route L2, wherein the first route L1 performs a convolution operation with a convolution kernel of 1*1; the second route L2 sequentially performs two convolution operations with a convolution kernel of 3*3. After the data of the time-frequency spectrum is input into the network infrastructure, convolution operations can be performed in the first route L1 and the second route L2 respectively, wherein the input data in the second route L2 needs to be circulated through two convolution layers, batch normalization layers and activation functions to obtain output data. Since this application only performs feature extraction, the convolutional neural network does not include pooling layers, normalization layers and fully connected layers. Then the output data of the first route L1 and the output data of the second route L2 are summed and output through the activation function.

[0124] In the entire deep residual network, the output data of one network infrastructure can be used as the input data or output data of other network infrastructures, thereby realizing the calculation process of the time-frequency graph in the deep residual network.

[0125] Preferably, the activation function of the deep residual network can be set to a ReLU function, and the expression of the ReLU function is as follows:

[0126] f(x)=max(0,x)

[0127] In the formula, x is the input value of the activation function, and f(x) is the output value of the activation function.

[0128] Reference Figure 4 As shown, in a non-limiting embodiment, the process of extracting features of the time-frequency spectrum in the deep residual network is as follows: first, the marked time-frequency spectrum is input into the deep residual network for training to update the weights in the deep residual network. After the deep residual network training is completed, the time-frequency spectrum of the vibration signal during the operation of the gearbox can be obtained, and then the time-frequency spectrum is split, and the split time-frequency spectrum is input into the trained deep residual network. The time-frequency spectrum first undergoes a convolution operation with a convolution kernel of 7*7, and then is sequentially input into a residual block having at least one network infrastructure for convolution operation, and after the last convolution operation is completed, a second feature set X2 containing feature information of the time-frequency spectrum is output.

[0129] The present invention introduces a multimodal fault diagnosis algorithm, adopts a sparse autoencoder algorithm to realize the feature extraction of slowly varying time series signals, and adopts a convolutional neural network ResNet18 to realize the feature extraction of vibration signals, and then uses width learning to realize high-precision fault diagnosis, so as to realize the feature fusion and fault diagnosis of multi-source sensor information.

[0130] Further, in S6, the characteristic parameter set X is input into the width learning model, and the step of obtaining the fault type of the gearbox includes: training the width learning model through multiple sets of labeled characteristic parameter sets X to obtain a trained width learning model. The trained width learning model includes multiple fault types of the gearbox. After the width learning model is trained, the classification of the gearbox fault state can be achieved.

[0131] The first feature set X1 and the second feature set X2 are obtained during the operation of the gearbox, and the first feature set X1 and the second feature set X2 are reconstructed to form a feature parameter set X. The feature parameter set X is input into the trained width learning model to obtain the fault type of the gearbox.

[0132] Specifically, the width learning model is a multi-classification model, which includes various types of gearbox fault conditions. The gearbox fault conditions include but are not limited to gear pitting faults, gear tooth breakage faults, bearing wear faults, bearing eccentric rotation faults, bearing fracture faults and other types.

[0133] This application diagnoses the fault type of the gearbox by establishing a width learning model. Compared with the convolutional neural network, the model has a simple network structure, a greatly reduced number of parameters, a fast network weight calculation speed, and a quick establishment of the entire model, which can better meet data calculation requirements.

[0134] Reference Figure 5 As shown, in a non-limiting embodiment, the width learning model is trained by multiple sets of labeled feature parameter sets X, and the steps of obtaining the trained width learning model include:

[0135] Create a labeled feature parameter set X to feature node z i The mapping of the marked feature parameter set X has n sets of feature data, each set of feature data contains m features, X∈R N×M .

[0136] Feature node z i The output formula is:

[0137] z i =Φ(XW ei +β ei )

[0138] In the formula, X is the input data in the feature parameter set, z i is the output data of the i-th feature node, W ei is the weight matrix of the i-th activation function randomly generated by the feature mapping layer, β ei is the bias matrix of the i-th excitation function randomly generated by the feature mapping layer, and Ф is the activation function.

[0139] The activation function Φ is preferably a Relu function.

[0140] All generated feature nodes z i Construct the characteristic node vector Z n , z n =[z1,…,z n ].

[0141] Then establish the characteristic node vector Z n To the enhanced node H m The mapping of the enhanced node H m The output formula is:

[0142] H m =ξ(z n W hm +β hm )

[0143] In the formula, H m is the output data of the mth enhancement node, W hm is the weight matrix randomly generated by the mth enhancement node, β hm is the bias matrix randomly generated by the mth enhancement node, and ξ is the activation function.

[0144] The activation function ξ is preferably a tanh function.

[0145] All generated enhancement nodes H m Constructing the enhanced node vector H m , H m =[H1,…,H m ].

[0146] Furthermore, we obtain the characteristic node vector Z n and the enhanced node vector H m The next steps include:

[0147] The characteristic node vector Z n and the enhanced node vector H m Perform matrix concatenation to obtain the feature matrix A, A = [z n |H m ].

[0148] The feature matrix A is used as the input layer of the width learning system to train the width learning model. Calculate and update the network weights W of the width learning model m , and obtain the trained width learning model.

[0149] The model for width learning is:

[0150] Y=[z1,…,z n |ξ(z nW h1 +β h1 ),…,ξ(z n W hm +β hm )]W m =[z n |H m ]W m

[0151] Where Y is the label value of the width learning model, W m The network weights of the model are learned for width.

[0152] Specifically, update the network weights W of the width learning model m The steps include:

[0153] W m The update formula is:

[0154] W m =[z n |H m ] + Y

[0155] In the formula, [z n |H m ] + is the Moore-Penrose pseudo-inverse, and defines A = [z n |H m ], then the Moore-Penrose pseudo-inverse of A is as follows:

[0156]

[0157] When diagnosing a gearbox fault, the operation signal and vibration signal of the gearbox during operation can be obtained through the above process, and then the characteristic node vector Z can be obtained. n and the enhanced node vector H m . So when solving the network weight W m When we need to first calculate the characteristic node vector Z n and the enhanced node vector H m Get [z n |H m ], and the feature matrix A is obtained. Then the feature matrix A is input into the width learning model for training. During the training process, the Moore-Penrose pseudo-inverse A of A can be obtained through the feature matrix A. + . Because at this time A + and the label value Y are known, so we can + And the label value Y is solved to get the network weight W m . Network weight W mAfter updating, a trained width learning model can be obtained, and the trained width learning model includes various gearbox fault types.

[0158] After the width learning model is trained, the feature matrix A is obtained according to the operating signals and vibration signals obtained during the operation of the gearbox. The feature matrix A is input into the trained width learning model to obtain the fault type corresponding to the gearbox, thereby achieving high-precision fault diagnosis of the gearbox.

[0159] In summary, the present invention provides a loader gearbox fault diagnosis method based on typical working conditions. The method first obtains multi-source heterogeneous data from the non-stationary process of engineering machinery, and the multi-source heterogeneous data can fully reflect the operating status of the gearbox. Subsequently, the Bayesian model is used to screen and classify typical working conditions, and then the typical working condition data set of the gearbox (including torque signal, speed signal and vibration signal) is obtained. Then, for the typical working condition data, a multimodal data fusion algorithm is introduced to realize the feature fusion of the typical working condition data set, and the width learning algorithm is used to quickly and accurately perform precise fault diagnosis on the gearbox.

[0160] This method can realize effective fault diagnosis during the operation of engineering machinery, accurately locate the fault type of the gearbox, realize predictive maintenance of engineering machinery, reduce operation and maintenance costs, and improve operation and maintenance efficiency, avoiding losses caused by gearbox downtime. This method fully considers the actual operating characteristics of engineering machinery and has important research significance and further exploration value in the practical application of engineering machinery.

[0161] The above description is only a description of the preferred embodiments of the present invention, and is not intended to limit the scope of the present invention. Any changes or modifications made by a person skilled in the art in the field of the present invention based on the above disclosure shall fall within the protection scope of the present invention.

Claims

1. A loader gearbox fault diagnosis method based on typical working conditions, characterized in that: The loader gearbox fault diagnosis method comprises the following steps: S1: Acquire an operation signal and a vibration signal during the operation of the gearbox; the operation signal includes a speed signal and a torque signal; S2: obtaining optimized operation signal and optimized vibration signal through data preprocessing; S3: screening the optimized operation signal under working conditions to obtain an operation signal under a typical working condition; screening the optimized vibration signal under working conditions to obtain a vibration signal under a typical working condition; S4: extracting features from the operating signal under typical working conditions to obtain a first feature set; extracting features from the vibration signal under typical working conditions to obtain a second feature set; S5: reconstructing the first feature set and the second feature set to form a feature parameter set; S6: Inputting the characteristic parameter set into a width learning model to obtain the fault type of the gearbox.

2. The loader gearbox fault diagnosis method based on typical working conditions according to claim 1, characterized in that: A variety of sensors are arranged on the surface of the gear box, and the operation signal and the vibration signal during the operation of the gear box are obtained through the sensors.

3. The loader gearbox fault diagnosis method based on typical working conditions according to claim 1, characterized in that: The operation signal further includes at least one of a pressure signal and a temperature signal.

4. The loader gearbox fault diagnosis method based on typical working conditions according to claim 1, characterized in that: In S1, the step of obtaining the operation signal and the vibration signal of the loader gearbox during operation includes: The slowly varying time sequence signal of each of the operation signals is obtained respectively, and all the slowly varying time sequence signals are constructed into a first array; the vibration signal is obtained, and all the vibration signals are constructed into a second array.

5. The loader gearbox fault diagnosis method based on typical working conditions as claimed in claim 4, characterized in that: In S2, the step of obtaining an optimized operation signal through data preprocessing includes: All the slowly varying time series signals are subjected to low-pass filtering respectively, and normalized respectively to obtain normalized slowly varying time series signals; In S2, the step of obtaining an optimized vibration signal through data preprocessing includes: The vibration signal is subjected to VMD decomposition to obtain multiple modal components; the center frequency of each modal component is calculated; and the modal components corresponding to all the center frequencies are reconstructed to obtain the denoised vibration signal.

6. The method for loader gearbox fault diagnosis based on typical working conditions as claimed in claim 5, characterized in that: In S2, the step of obtaining the optimized operation signal and the optimized vibration signal through data preprocessing also includes: The normalized slowly varying time series signal and the denoised vibration signal are time stamp aligned, and the normalized slowly varying time series signal and the denoised vibration signal are short-time-series cut at equal intervals to obtain the fragmented operation signal and the fragmented vibration signal.

7. The method for loader gearbox fault diagnosis based on typical working conditions as claimed in claim 6, characterized in that: In S3, the step of performing working condition screening on the optimized operation signal and / or the optimized vibration signal to obtain the operation signal under typical working conditions and the vibration signal under typical working conditions includes: Extracting operation timing features from the fragmented operation signals; fusing the operation timing features with the gear information of the aligned timestamp to obtain operation characteristic information under different gears; manually marking the operating condition type to which each operation characteristic information belongs; substituting all the operation characteristic information and the operating condition type to which each operation characteristic information belongs into the Bayesian model for screening to obtain the operation signal under typical working conditions; And / or, extracting vibration time series features from the fragmented vibration signal; fusing the vibration time series features with the gear information of the aligned timestamp to obtain vibration feature information under different gears; manually marking the working condition type to which each vibration feature information belongs; substituting all the vibration feature information and the working condition type to which each vibration feature information belongs into the Bayesian model for screening to obtain the vibration signal under typical working conditions; The operation timing characteristics and / or the vibration timing characteristics include one or more of a mean value, a peak-to-peak value, an effective value, a variance, and a kurtosis value.

8. The method for loader gearbox fault diagnosis based on typical working conditions as claimed in claim 1, characterized in that: In S4, the step of extracting features from the operating signal under the typical working condition to obtain the first feature set includes: Inputting the operating signal under typical working conditions into a sparse autoencoder for feature extraction to obtain the first feature set; In S4, the step of extracting features from the vibration signal under typical working conditions to obtain the second feature set includes: Performing Fourier transform on the vibration signal under typical working conditions to obtain a time-frequency spectrum of the vibration signal; inputting the time-frequency spectrum into a convolutional neural network for feature extraction to obtain the second feature set.

9. The method for loader gearbox fault diagnosis based on typical working conditions as claimed in claim 8, characterized in that: The step of extracting features through the convolutional neural network includes: The convolutional neural network is set as a deep residual network; the time-frequency spectrum is split, and the split time-frequency spectrum is input into the deep residual network for feature extraction to obtain the second feature set.

10. The method for loader gearbox fault diagnosis based on typical working conditions according to claim 1, characterized in that: In S6, the step of inputting the characteristic parameter set into the width learning model to obtain the fault type of the gearbox includes: The width learning model is trained by using multiple sets of labeled feature parameter sets to obtain the trained width learning model, which includes multiple fault types of the gearbox: the feature parameter set during the operation of the gearbox is obtained, the feature parameter set is input into the trained width learning model, and the fault type of the gearbox is obtained.

Images