A multi-scale characteristic fault diagnosis method for rotating machinery based on non-intrusive signals

Abstract

The present invention belongs to the technical field of fault detection of industrial manufacturing equipment, and discloses a multi-scale feature fault diagnosis method for rotating machinery based on non-invasive signals. A multi-label data set is constructed, and the multi-label data set includes current fundamental wave signals and current residual signals collected under different working conditions, the current fundamental wave signals are provided with working condition labels, and the current residual signals are provided with fault labels; multiple TimesBlock models with the same network structure are introduced and the residual signals and fault labels of the training set are used for model pre-training, a multi-channel attention mechanism based on working condition information is introduced to obtain a low-dimensional representation of multi-channel multi-scale fault signals, and an adaptive joint loss function composed of working condition classification loss and fault classification loss is introduced. The present invention overcomes the problems of fundamental wave masking fault characteristics and low signal-to-noise ratio of the original current signal when performing rotating machinery fault diagnosis based on current signals.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault detection of industrial manufacturing equipment, and in particular relates to a multi-scale characteristic fault diagnosis method for rotating machinery based on non-intrusive signals. Background Art

[0002] As a key component of large-scale manufacturing equipment, rotating machinery is widely used in industries such as wind power generation, aerospace, military industry and machinery manufacturing. Since rotating machinery has been in service for a long time under complex working conditions such as high load and high strength, vibration, impact, wear and harsh environment will cause key components such as bearings, gears and screws to malfunction or fail. Once its key components fail or fail, it will affect the normal operation of rotating machinery and equipment, causing inevitable economic losses and even causing serious safety accidents.

[0003] Fault prediction technology is of great significance for ensuring the stable and reliable operation of rotating machinery and equipment. It uses the operating data of key components of rotating machinery and equipment to predict the fault type and failure time of key components through data analysis, model building and other methods, which helps to reduce the occurrence rate of equipment failure accidents, identify mechanical equipment failures in advance and take corresponding measures in time to reduce the adverse effects caused by equipment failure or failure.

[0004] For the fault diagnosis of rotating machinery, analysis based on vibration signals is a common and effective method. However, due to the high cost of vibration sensors and the need for direct contact with rotating machinery, there are limitations on installing sensors on equipment that has been put into production. Acoustic-based analysis is a non-invasive detection method that collects signals through a microphone and has low cost. However, the sound signal may be affected by environmental noise, which may mask the fault information. The three-phase current signal can be measured by installing a current transformer in the circuit, which has low installation cost, low invasiveness to the rotating machinery itself, and good resistance to external noise. Therefore, fault detection of rotating machinery based on three-phase current signals has attracted increasing attention from academia and industry.

[0005] However, the impact of mechanical faults on current signals is modulated by current, resulting in a very small amplitude. The fault characteristics are easily masked by the fundamental current. In addition, there are some harmonics and non-Gaussian noise in the current signal with a low signal-to-noise ratio. It is difficult to achieve good results by directly using current signals for fault diagnosis. Summary of the invention

[0006] The purpose of the present invention is to provide a rotating machinery multi-scale characteristic fault diagnosis method based on non-intrusive signals, which overcomes the problems of fundamental wave masking fault characteristics and low signal-to-noise ratio of original current signal when rotating machinery fault diagnosis is performed based on current signal.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A rotating machinery multi-scale characteristic fault diagnosis method based on non-intrusive signals, the rotating machinery multi-scale characteristic fault diagnosis method based on non-intrusive signals comprising:

[0009] Constructing a multi-label data set, the multi-label data set includes a current fundamental signal and a current residual signal obtained after filtering the current signal collected under different working conditions, the current fundamental signal is provided with a working condition label, and the current residual signal is provided with a fault label;

[0010] For each working condition, the same network structure is used to establish a fault classification model. The current residual signal and fault label under the corresponding working condition are used to pre-train the fault classification model until convergence. The pre-trained fault classification model is used to output the corresponding multi-scale low-dimensional representation of the fault according to the current residual signal.

[0011] Establish a working condition classification model with the same network structure as the fault classification model, input the current fundamental wave signal into the working condition classification model, and obtain a multi-scale low-dimensional representation of the working condition;

[0012] Based on the multi-scale low-dimensional representation of working conditions and the multi-scale low-dimensional representation of faults, weights are constructed, and the multi-scale low-dimensional representation of faults is weighted by the weights to obtain weighted features under each working condition;

[0013] The weighted features under each working condition are concatenated as fault features, and the fault classification vector is output through the fully connected layer.

[0014] Based on the fault classification vector, an adaptive joint loss function including the working condition classification loss and the fault classification loss is constructed, and the working condition classification model is optimized and updated until convergence using the adaptive joint loss function;

[0015] The current fundamental signal and current residual signal to be diagnosed are obtained, the operating condition classification model and the fault classification model under each operating condition are used to obtain the fault characteristics to be diagnosed, and the fault diagnosis result is output through the fully connected layer.

[0016] Several optional methods are also provided below, but they are not intended to be additional limitations on the above-mentioned overall solution, but are merely further supplements or preferences. Under the premise that there are no technical or logical contradictions, each optional method can be combined with the above-mentioned overall solution separately, and multiple optional methods can also be combined.

[0017] Preferably, the step of constructing a multi-label dataset comprises:

[0018] The Butterworth filter is used to perform low-pass filtering on the original collected three-phase current signal with the AC working frequency as the cutoff frequency to obtain the current fundamental signal and current residual signal under various working conditions. The formula is as follows:

[0019] ;

[0020] In the formula, is the original collected three-phase current signal, is the current fundamental signal obtained by low-pass filtering, is the current residual signal;

[0021] The obtained current fundamental signal and current residual signal are subjected to sliding window operation, and are respectively labeled with operating condition labels and fault labels to construct a multi-label dataset. , represents the current residual signal, Represents the current fundamental signal, Indicates the fault tag, Indicates the operating condition label.

[0022] Preferably, the fault classification model and the operating condition classification model are both Model, The network structure of the model is as follows:

[0023] For an input time length of , the channel dimension is One-dimensional time series , the specific formula is as follows:

[0024] ;

[0025] in, Indicates Channel data, , Indicates data Perform fast Fourier transform to convert the time series signal into frequency domain representation, and the output data is a data with the same length of The complex sequence of , Indicates Channel input signal The amplitude and phase information of each frequency component, ;

[0026] Plural To perform amplitude operation, the specific formula is as follows:

[0027] ;

[0028] in, Indicates plural The amplitude of represents the real part of a complex number, It means taking the imaginary part of a complex number;

[0029] The obtained amplitude The averaging operation is performed in the channel dimension. The specific formula is as follows:

[0030] ;

[0031] in, Represents a one-dimensional time series The intensity of each frequency component in is obtained by the intensity sequence Before obtaining The frequency sequence with the largest amplitude And the corresponding period length sequence , the specific formula is as follows:

[0032] ;

[0033] ;

[0034] in, Indicates the maximum amplitude frequency, Represents frequency sequence The frequency, Indicates taking Frequency in the range , Indicates rounding up operation. Represents a period length sequence The The cycle length, ;

[0035] According to the cycle length The original one-dimensional time series Reconstruct into a two-dimensional time series signal sequence , the specific formula is as follows:

[0036] ;

[0037] in, Represents a two-dimensional time series signal The A two-dimensional time series signal, Indicates padding with zeros at the end of the sequence. and Represents the number of rows and columns of the input two-dimensional tensor, Indicates that the data is reorganized into rows size;

[0038] For the obtained two-dimensional time series signal sequence , use the Inception module to extract the two-dimensional temporal change representation. The specific formula is as follows:

[0039] ;

[0040] in, It represents the two-dimensional temporal change representation extracted by the Inception module. Represents different convolution kernel operations, Represents a splicing operation;

[0041] The extracted two-dimensional time series changes are represented Transformed back to one-dimensional space, the specific formula is as follows:

[0042] ;

[0043] in, Represented by two-dimensional time series changes The transformed one-dimensional representation is Indicates that the data is reorganized into rows size, Indicates that The 0 added by the operation is removed;

[0044] For the obtained one-dimensional representation According to the previous The frequency with the largest amplitude Perform weighted summation, the specific formula is as follows:

[0045] ;

[0046] ;

[0047] in, Indicates frequency The intensity of the corresponding frequency component, For strength The probability distribution of is the activation function, It is the multi-scale output after adaptive fusion;

[0048] Using the fully connected layer Dimensionality reduction is performed to obtain the low-dimensional features of the reconstructed signal. The specific formula is as follows:

[0049] ;

[0050] in, is a multi-scale low-dimensional representation, , are the weights and biases of the fully connected layer, and T represents the matrix transpose operation.

[0051] Preferably, the pre-training of the fault classification model by using the current residual signal and the fault label under the corresponding working condition until convergence comprises:

[0052] ;

[0053] ;

[0054] in, Indicates Fault classification model under various working conditions, is the input current residual signal, It is Multi-scale low-dimensional representation of faults under various working conditions. is the fault classification vector output during pre-training, , For the The weights and biases of the fully connected layer under these conditions;

[0055] The network parameters of the fault classification model are iterated through the Adam optimizer to minimize the loss function, and the loss function adopts the cross entropy function.

[0056] Preferably, the inputting of the current fundamental wave signal into the operating condition classification model to obtain a multi-scale low-dimensional representation of the operating condition includes:

[0057] ;

[0058] ;

[0059] in, represents the working condition classification model, is the input current fundamental signal, is a multi-scale low-dimensional representation of the working condition, is the operating condition classification vector, , are the network parameters of the fully connected layer in working condition classification.

[0060] Preferably, the weights are constructed based on the multi-scale low-dimensional representation of the working condition and the multi-scale low-dimensional representation of the fault, and the multi-scale low-dimensional representation of the fault is weighted by the weights to obtain the weighted features under each working condition, including:

[0061] Take multi-scale low-dimensional representation of working conditions Multi-scale low-dimensional representation of faults under different working conditions Perform multi-channel attention mechanism weighting. The specific formula is as follows:

[0062] ;

[0063] ;

[0064] in, Indicates The weight matrix under the working condition is: is the activation function, T represents the matrix transpose operation, Indicates Multi-scale low-dimensional representation of faults under various working conditions. Represents the weighted multi-channel attention mechanism The weighted characteristics under different working conditions, , is the total number of working conditions.

[0065] Preferably, the method of constructing an adaptive joint loss function including a working condition classification loss and a fault classification loss based on the fault classification vector comprises:

[0066] ;

[0067] ;

[0068] ;

[0069] ;

[0070] ;

[0071] in, represents the adaptive joint loss function, is the weight of the working condition classification loss, is the weight of the fault classification loss, is the cross entropy loss for condition classification, is the cross entropy loss for fault classification, is the probability distribution of the true working condition label, It is predicted to belong to the working condition category The probability of is the total number of working conditions, is the probability distribution of true fault labels, is predicted to belong to the fault category The probability of is the total fault category, is the operating condition classification vector, is the fault classification vector during training.

[0072] Preferably, the weight and They are dynamic hyperparameters respectively, and the specific formula for iteration is as follows:

[0073] ;

[0074] in, , represents the adjustment step size, Indicates The value of the adaptive joint loss function obtained by round iterative calculation, Represents the average value of the adaptive joint loss function obtained from all current iterations, Indicates The weight of the round iteration , Indicates The weight of the round iteration , .

[0075] The present invention provides a rotating machinery multi-scale characteristic fault diagnosis method based on non-intrusive signals, which has the following beneficial effects compared with the prior art:

[0076] (1) Based on the current signal, the relationship between the three-phase data of U, V, and W is considered to realize non-intrusive fault diagnosis of rotating machinery, reducing the intrusiveness to the equipment.

[0077] (2) An adaptive joint loss function consisting of operating condition classification loss and fault classification loss is proposed to achieve adaptive joint optimization of the operating condition classification task and the fault classification task.

[0078] (3) A multi-channel attention mechanism is introduced to realize the fusion of working condition characteristics and fault characteristics, provide feature representation, and improve the accuracy of fault classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 This is a flow chart of a method for multi-scale characteristic fault diagnosis of rotating machinery based on non-intrusive signals of the present invention;

[0080] Figure 2 The loss function decline curve of the present invention in the joint optimization training process under the public data set;

[0081] Figure 3 The accuracy change curve of the present invention during the joint optimization training process under the public data set;

[0082] Figure 4 Confusion matrix of the experimental results of the present invention under the public data set. DETAILED DESCRIPTION

[0083] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0084] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0085] like Figure 1 As shown, this embodiment proposes a rotating machinery multi-scale feature fault diagnosis method based on non-intrusive signals, comprising the following steps:

[0086] Step S1, construct a multi-label data set. Under various speed and load conditions, the U, V, and W three-phase current signals of misalignment and bearing failure of rotating machinery are collected using current transformers, and low-pass filtering is performed using a Butterworth filter with a cutoff frequency set to the working frequency of the alternating current to obtain the current fundamental signal and current residual signal under different working conditions. The obtained signal is subjected to a sliding window operation, and the current fundamental signal and the current residual signal are respectively marked with working condition labels and fault labels to construct a multi-label data set. , and randomly divide the data set into training set and test set according to 7:3. is the current residual signal, is the current fundamental signal, is the fault label, The working condition label.

[0087] The sensor used in this embodiment is a three-phase current transformer. The various working conditions include 0N·m, 2N·m, and 4N·m loads, and the different fault conditions include normal bearing state, bearing 0.1mm inner ring fault state, bearing 0.3mm inner ring fault state, and bearing 0.5mm misalignment fault state.

[0088] The Butterworth filter is used to low-pass filter the original three-phase current signal. The specific formula is as follows:

[0089] ;

[0090] in, is the transfer function of the filter, is a complex frequency domain variable, is the cut-off frequency, set to the operating frequency of the AC power, It is the order of the filter, which determines the steepness of the filter. The higher the order, the steeper the transition band of the filter. After filtering, the current fundamental signal and current residual signal under various working conditions are obtained. The specific formula is as follows:

[0091] ;

[0092] in, is the original three-phase current signal, is the current fundamental signal obtained by low-pass filtering. is the current residual signal, Bring in This indicates that the Butterworth low-pass filter cannot completely filter the input signal at the cutoff frequency, so consider multiplying it by a factor of 2 to enhance the filtering effect.

[0093] Step S2: For each working condition, use the same network structure to establish a model of the corresponding channel (i.e. the corresponding working condition) , Indicates In this working condition, the model transforms the input signal from a 1D space sequence to a 2D space, and then uses 2D convolution to extract the 2D space signal features. The extracted features are transformed into a 1D space by sequence transformation, and the reconstructed signal is reduced in dimension through a fully connected layer to obtain low-dimensional features. The model is pre-trained by using the current residual signal and fault label of the corresponding working condition to perform fault classification pre-training, so each model has the ability to extract multi-scale features of the fault signal under the corresponding specified working condition.

[0094] For ease of understanding, this embodiment uses the same network structure as The model is generally described in In the application of the model, the corresponding output is obtained according to different inputs. The general description of the network structure is as follows:

[0095] For a time length of , the channel dimension is One-dimensional time series , express OK The real number set of the column, the specific formula is as follows:

[0096] ;

[0097] in, Indicates Channel data, Indicates the signal Perform a fast Fourier transform (FFT) to convert the time series signal into a frequency domain representation. The output data is a frequency domain with the same length. The complex sequence of , Indicates Channel input signal The amplitude and phase information of each frequency component.

[0098] For the frequency domain components obtained above To perform amplitude operation, the specific formula is as follows:

[0099] ;

[0100] in Indicates plural The amplitude reflects the energy of each frequency component. The function represents the real part of a complex number. The function represents the imaginary part of a negative number.

[0101] The obtained amplitude The averaging operation is performed in the channel dimension. The specific formula is as follows:

[0102] ;

[0103] in, express The intensity of each frequency component in express dimensional vector space, It is obtained by taking the average value of the amplitude of each channel data. Before obtaining The frequency with the largest amplitude And the corresponding cycle length , the specific formula is as follows:

[0104] ;

[0105] ;

[0106] in Indicates selecting the largest value among the given data. frequency, Represents frequency sequence The frequencies. Due to the conjugation in the frequency domain, this paper only considers , that is, only take Frequency in the range , Indicates rounding up operation. Represents a period length sequence The The length of a cycle.

[0107] According to the cycle obtained above The original one-dimensional time series data Data reconstruction into two-dimensional time series signal The specific formula is as follows:

[0108] ;

[0109] in, Indicates that zero padding is performed at the end of the sequence to achieve data alignment, so that the length of the sequence can be divisible, and Represents the number of rows and columns of the input two-dimensional tensor, Indicates that the data is reorganized into rows size.

[0110] After the above operation, we can get a set of two-dimensional tensors , Corresponding cycle The two-dimensional temporal changes dominated by it.

[0111] For the obtained two-dimensional tensor , because of its two-dimensional locality, the 2D convolutional network Inception can be used to extract the two-dimensional temporal change representation. The specific formula is as follows:

[0112] ;

[0113] in, Represents the two-dimensional tensor features extracted using the Inception module. Represents different convolution kernel operations, Indicates the concatenation operation. Each module of the Inception model contains multiple convolution kernels of different sizes, which work in parallel to extract features of different scales.

[0114] For the time series features extracted above, they need to be converted back to one-dimensional space for information aggregation. The aggregate formula is as follows:

[0115] ;

[0116] in, Represented by two-dimensional time series changes The transformed one-dimensional representation is Indicates that the data is reorganized into rows size, Indicates that the above steps The operation complements the 0 removal.

[0117] For the obtained one-dimensional representation, according to the above The frequency with the largest amplitude Perform weighted summation, the specific formula is as follows:

[0118] ;

[0119] ;

[0120] in, Indicates frequency The intensity of the corresponding frequency component, For strength The probability distribution of It is a commonly used activation function that can convert a set of values ​​into a probability distribution. It is the multi-scale output after adaptive fusion.

[0121] Using the fully connected layer Dimensionality reduction is performed to obtain the low-dimensional features of the reconstructed signal. The specific formula is as follows:

[0122] ;

[0123] in, is a multi-scale low-dimensional representation, , are the weights and biases of the fully connected layer, and T represents the matrix transpose operation.

[0124] Using the current residual signal under the corresponding working conditions in the training set and fault labels Pre-train each The model uses a fully connected layer to reduce the dimension of the model output. The specific formula is as follows:

[0125] ;

[0126] ;

[0127] in, Indicates channel The corresponding model is used to extract multi-scale low-dimensional features of the input signal. is the input residual signal, It is a channel The multi-scale low-dimensional representation of the fault (i.e., the multi-scale low-dimensional representation in the general description of the model ), is the fault classification vector, , For the The network parameters of the fully connected layer under the working conditions are weights and biases.

[0128] use Vector, the Adam optimizer iterates the network parameters to minimize the loss function, and the loss function uses the cross entropy function. The specific formula is as follows:

[0129] ;

[0130] in, is the cross entropy function during pre-training of the fault classification model, is the probability distribution of the true label of the fault type, It is the fault category predicted during model pre-training The probability of is the total number of fault categories.

[0131] The purpose of model pre-training is to improve the model's ability to extract multi-scale features of fault signals under corresponding working conditions. After pre-training, the pre-trained fault classification model is used to output a multi-scale low-dimensional representation of the fault under the corresponding working condition based on the current residual signal and fault label under the corresponding working condition.

[0132] Step S3: Establish a network structure consistent with the initially constructed fault classification model The model, as a working condition classification model, trains the current fundamental signal Input the working condition model and use the fully connected layer to reduce the dimension of the model output. The specific formula is as follows:

[0133] ;

[0134] ;

[0135] in, Indicated by The working condition classification model of model construction, is the input current fundamental signal, is a multi-scale low-dimensional representation of the working condition, is the operating condition classification vector, , are the network parameters of the fully connected layer in working condition classification.

[0136] Step S4: construct weights based on the multi-scale low-dimensional representation of working conditions and the multi-scale low-dimensional representation of faults, and use the weights to weight the multi-scale low-dimensional representation of faults to obtain weighted features under each working condition.

[0137] In order to realize the fusion of working condition information and fault information, the multi-scale low-dimensional representation of working condition is used. Multi-scale low-dimensional representation of fault classification models for different channels Perform multi-channel attention mechanism weighting. The specific formula is as follows:

[0138] ;

[0139] ;

[0140] in, is an activation function that converts the matrix into a probability distribution. T represents the matrix transpose operation. Indicates The multi-scale low-dimensional representation of the faults of channels, Indicates The weight matrix of the channels, Represents the weighted attention mechanism The weighted features of the channels, , is the total number of working conditions.

[0141] Step S5: concatenate the weighted features under each working condition as fault features, and output a fault classification vector for the fault features through a fully connected layer.

[0142] The obtained multi-channel weighted features are spliced ​​to obtain the fault features that integrate the multi-channel information and the working condition information. , and use the fully connected layer to reduce its dimension. The specific formula is as follows:

[0143] ;

[0144] in, represents the fault classification vector obtained based on fault features during training, , are the network parameters of the fully connected layer that processes fault features during training.

[0145] Step S6: Based on the fault classification vector, construct an adaptive joint loss function including the operating condition classification loss and the fault classification loss, and use the adaptive joint loss function to optimize and update the fault classification model and the operating condition classification model until convergence.

[0146] The adaptive joint loss function constructed is as follows:

[0147] ;

[0148] ;

[0149] ;

[0150] ;

[0151] ;

[0152] in is the cross entropy loss for condition classification, is the cross entropy of fault classification, is the probability distribution of the true working condition label, The model predicts that the The probability of Represents the total number of working conditions, In terms of numerical value, same, In terms of numerical value, Same, here for the convenience of description, an equivalent replacement is made, that is, it can also be expressed as , is the probability distribution of true fault labels, It is the fault category predicted by the model The probability of is the total fault category, which is numerically the same as that in step S2. Consistent. and They are dynamic hyperparameters respectively, and the specific formula for iteration is as follows:

[0153] ;

[0154] in, , , The training strategy is based on the loss of the previous round of training The average loss over all rounds The difference is one step (for example, set to 0.05), The initial value of is set to 0.5. When , the weight is taken as the initial value. In subsequent iterations, the weight is dynamically adjusted before calculating the adaptive joint loss function each time.

[0155] The Adam optimizer is used to jointly optimize the two classification tasks, and the adaptive joint loss function is used to optimize and update the working condition classification model until convergence.

[0156] After the training, the current fundamental wave signal and current residual signal to be diagnosed are obtained, and the fault features to be diagnosed are obtained by using the working condition classification model and the fault classification model under each working condition, and the fault diagnosis result is output through the fully connected layer. Specifically, the fault classification model under each working condition after training outputs the multi-scale low-dimensional representation of the fault under the corresponding working condition according to the current residual signal to be diagnosed; the current fundamental wave signal to be diagnosed is input into the trained working condition classification model to obtain the multi-scale low-dimensional representation of the working condition; weights are constructed based on the multi-scale low-dimensional representation of the working condition and the multi-scale low-dimensional representation of the fault, and the multi-scale low-dimensional representation of the fault is weighted by the weights to obtain the weighted features under each working condition; the weighted features under each working condition are spliced ​​as the fault features to be diagnosed, and the fault classification vector is output as the fault diagnosis result for the fault features to be diagnosed through the fully connected layer.

[0157] In a specific embodiment, the present invention is verified based on a public dataset of bearing faults in the document “Vibration, acoustic, temperature, and motor current dataset of rotating machine under varying operating conditions for fault diagnosis”.

[0158] In the data set, the three-phase current data of normal, bearing 0.1mm inner ring fault (fault one), bearing 0.3mm inner ring fault (fault two), and bearing 0.5mm misalignment fault (fault three) under the conditions of 0Nm, 2Nm, and 4Nm loads and a constant speed of 3010r / min are selected. The fundamental signal and residual signal of the three-phase current signal are obtained by low-pass filtering with a Butterworth filter, and the data is divided into multiple signals of the same length by sliding window technology, and the current residual signals under different working conditions are used for pre-training. The multi-scale fault signal extraction capability of the model is to input the fundamental signal with working condition label to Model, obtain multi-scale low-dimensional representation of working condition information , the multi-channel attention mechanism is used to realize the weighted fusion of multi-channel fault information and working condition information, and the weighted results of different channels are spliced ​​to output the final fault features containing working condition information . Using linear layers Dimensionality reduction is performed, and an adaptive joint loss function consisting of working condition classification loss and fault classification loss is constructed. The Adam optimizer is used to realize the joint optimization of fault classification and working condition classification tasks.

[0159] The data training set and test set are randomly divided into two parts according to the ratio of 7:3. The training set data is used to train the model, and the test set data is classified by the trained model. The experimental results of the method of the present invention under three working conditions and four fault conditions are as follows: Figure 2~Figure 4 As shown in . The loss function changes with the number of training rounds as shown in Figure 2 As shown in the figure, the model loss decreases significantly in the first few rounds of training, and then slowly fluctuates and decreases in the later stages of training. The classification accuracy changes with the number of training rounds as shown in Figure 3 As shown in the figure, the accuracy rate increases significantly in the first few rounds of training, and then slowly fluctuates and increases in the later stages of training. Figure 4 As shown in the figure, the model is more accurate in classifying faults 1 and 3, but slightly inferior in classifying normal and fault 1. The classification accuracy reaches 96.11%, and the fault detection accuracy is high. Based on the above experiments, the method of the present invention has excellent fault diagnosis performance.

[0160] Different from the existing fault diagnosis methods based on vibration signals and other signals, the present invention realizes higher-precision bearing fault diagnosis from the perspective of three-phase current which is less affected by the external environment, taking into account the working environment with high noise, and enhances the robustness of the system bearing fault diagnosis while reducing the intrusion into the system. It realizes multi-task dynamic joint optimization, and overcomes the problems of fundamental wave masking fault characteristics and low signal-to-noise ratio of the original current signal when performing rotating machinery fault diagnosis based on current signals.

[0161] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0162] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A multi-scale characteristic fault diagnosis method for rotating machinery based on non-intrusive signals, characterized in that: The rotating machinery multi-scale characteristic fault diagnosis method based on non-intrusive signals comprises: Constructing a multi-label data set, the multi-label data set includes a current fundamental signal and a current residual signal obtained after filtering the current signal collected under different working conditions, the current fundamental signal is provided with a working condition label, and the current residual signal is provided with a fault label; For each working condition, the same network structure is used to establish a fault classification model. The current residual signal and fault label under the corresponding working condition are used to pre-train the fault classification model until convergence. The pre-trained fault classification model is used to output the corresponding multi-scale low-dimensional representation of the fault according to the current residual signal. Establish a working condition classification model with the same network structure as the fault classification model, input the current fundamental wave signal into the working condition classification model, and obtain a multi-scale low-dimensional representation of the working condition; Based on the multi-scale low-dimensional representation of working conditions and the multi-scale low-dimensional representation of faults, weights are constructed, and the multi-scale low-dimensional representation of faults is weighted by the weights to obtain weighted features under each working condition; The weighted features under each working condition are concatenated as fault features, and the fault classification vector is output through the fully connected layer. Based on the fault classification vector, an adaptive joint loss function including the working condition classification loss and the fault classification loss is constructed, and the working condition classification model is optimized and updated until convergence using the adaptive joint loss function; The current fundamental signal and current residual signal to be diagnosed are obtained, the operating condition classification model and the fault classification model under each operating condition are used to obtain the fault characteristics to be diagnosed, and the fault diagnosis result is output through the fully connected layer.

2. The rotating machinery multi-scale characteristic fault diagnosis method based on non-intrusive signals according to claim 1 is characterized in that: The constructing of a multi-label dataset includes: The Butterworth filter is used to perform low-pass filtering on the original collected three-phase current signal with the AC working frequency as the cutoff frequency to obtain the current fundamental signal and current residual signal under various working conditions. The formula is as follows: ； In the formula, is the original collected three-phase current signal, is the current fundamental signal obtained by low-pass filtering. is the current residual signal; The obtained current fundamental signal and current residual signal are subjected to sliding window operation, and are respectively labeled with operating condition labels and fault labels to construct a multi-label dataset. , represents the current residual signal, Represents the current fundamental signal, Indicates the fault tag, Indicates the operating condition label.

3. The rotating machinery multi-scale characteristic fault diagnosis method based on non-intrusive signals according to claim 1 is characterized in that: The fault classification model and the operating condition classification model are both based on Model, The network structure of the model is as follows: For an input time length of , the channel dimension is One-dimensional time series , the specific formula is as follows: ； in, Indicates Channel data, , Indicates data Perform fast Fourier transform to convert the time series signal into frequency domain representation, and the output data is a data with the same length of The complex sequence of , Indicates Channel input signal The amplitude and phase information of each frequency component, ; Plural To perform amplitude operation, the specific formula is as follows: ； in, Indicates plural The amplitude of represents the real part of a complex number, It means taking the imaginary part of a complex number; The obtained amplitude The averaging operation is performed in the channel dimension. The specific formula is as follows: ； in, Represents a one-dimensional time series The intensity of each frequency component in is obtained by the intensity sequence Before obtaining The frequency sequence with the largest amplitude And the corresponding period length sequence , the specific formula is as follows: ； ； in, Indicates the maximum amplitude frequency, Represents frequency sequence The frequency, Indicates taking Frequency in the range , Indicates rounding up operation. Represents a period length sequence The The cycle length, ; According to the cycle length The original one-dimensional time series Reconstruct into a two-dimensional time series signal sequence , the specific formula is as follows: ； in, Represents a two-dimensional time series signal The A two-dimensional time series signal, Indicates padding with zeros at the end of the sequence. and Represents the number of rows and columns of the input two-dimensional tensor, Indicates that the data is reorganized into rows size; For the obtained two-dimensional time series signal sequence , use the Inception module to extract the two-dimensional temporal change representation. The specific formula is as follows: ； in, It represents the two-dimensional temporal change representation extracted by the Inception module. Represents different convolution kernel operations, Represents a splicing operation; The extracted two-dimensional time series changes are represented Transformed back to one-dimensional space, the specific formula is as follows: ； in, Represented by two-dimensional time series changes The transformed one-dimensional representation is Indicates that the data is reorganized into rows size, Indicates that The 0 added by the operation is removed; For the obtained one-dimensional representation According to the previous The frequency with the largest amplitude Perform weighted summation, the specific formula is as follows: ； ； in, Indicates frequency The intensity of the corresponding frequency component, For strength The probability distribution of is the activation function, It is the multi-scale output after adaptive fusion; Using the fully connected layer Dimensionality reduction is performed to obtain the low-dimensional features of the reconstructed signal. The specific formula is as follows: ； in, is a multi-scale low-dimensional representation, , are the weights and biases of the fully connected layer, and T represents the matrix transpose operation.

4. The rotating machinery multi-scale characteristic fault diagnosis method based on non-intrusive signals according to claim 1 is characterized in that: The method of pre-training the fault classification model by using the current residual signal and the fault label under the corresponding working condition until convergence includes: ； ； in, Indicates Fault classification model under various working conditions, is the input current residual signal, It is Multi-scale low-dimensional representation of faults under various working conditions. is the fault classification vector output during pre-training, , For the The weights and biases of the fully connected layer under these conditions; The network parameters of the fault classification model are iterated through the Adam optimizer to minimize the loss function, and the loss function adopts the cross entropy function.

5. The rotating machinery multi-scale characteristic fault diagnosis method based on non-intrusive signals according to claim 1 is characterized in that: The current fundamental wave signal is input into the working condition classification model to obtain a multi-scale low-dimensional representation of the working condition, including: ； ； in, represents the working condition classification model, is the input current fundamental signal, is a multi-scale low-dimensional representation of the working condition, is the operating condition classification vector, , are the network parameters of the fully connected layer in working condition classification.

6. The rotating machinery multi-scale characteristic fault diagnosis method based on non-intrusive signals according to claim 1 is characterized in that: The weights are constructed based on the multi-scale low-dimensional representation of the working condition and the multi-scale low-dimensional representation of the fault, and the multi-scale low-dimensional representation of the fault is weighted by the weights to obtain the weighted features under each working condition, including: Take multi-scale low-dimensional representation of working conditions Multi-scale low-dimensional representation of faults under different working conditions Perform multi-channel attention mechanism weighting. The specific formula is as follows: ； ； in, Indicates The weight matrix under the working condition is: is the activation function, T represents the matrix transpose operation, Indicates Multi-scale low-dimensional representation of faults under various working conditions. Represents the weighted multi-channel attention mechanism The weighted characteristics under different working conditions, , is the total number of working conditions.

7. The rotating machinery multi-scale characteristic fault diagnosis method based on non-intrusive signals according to claim 1 is characterized in that: The method of constructing an adaptive joint loss function including a working condition classification loss and a fault classification loss based on the fault classification vector includes: ； ； ； ； ； in, represents the adaptive joint loss function, is the weight of the working condition classification loss, is the weight of the fault classification loss, is the cross entropy loss for condition classification, is the cross entropy loss for fault classification, is the probability distribution of the true working condition label, It is predicted to belong to the working condition category The probability of is the total number of working conditions, is the probability distribution of true fault labels, is predicted to belong to the fault category The probability of is the total fault category, is the operating condition classification vector, is the fault classification vector during training.

8. The rotating machinery multi-scale characteristic fault diagnosis method based on non-intrusive signals according to claim 7 is characterized in that: The weight and They are dynamic hyperparameters respectively, and the specific formula for iteration is as follows: ； in, , represents the adjustment step size, Indicates The value of the adaptive joint loss function obtained by round iterative calculation, Represents the average value of the adaptive joint loss function obtained from all current iterations, Indicates The weight of the round iteration , Indicates The weight of the round iteration , .

Images