Mechanical pump fault diagnosis method based on multi-source information and model construction method thereof

Abstract

The invention provides a multi-source information-based mechanical pump fault diagnosis method and a model construction method thereof, and belongs to the field of vacuumizing mechanical pump fault diagnosis in a semiconductor process. By introducing a diversity feature and multi-source information strategy, vibration signals of multiple measuring points under different faults of a mechanical pump are collected, and the fault diagnosis model is constructed. And carrying out time domain analysis, frequency domain analysis and time-frequency domain analysis to obtain different parameters, adopting a multi-source information fusion method to obtain fusion features with high-dimensional mixing characteristics, and finally using the fused features to train a deep learning network to carry out fault classification. The problems of poor algorithm robustness, poor generalization ability, low result accuracy and the like in mechanical pump fault diagnosis are effectively solved, so that state parameters of the mechanical pump are monitored in real time, and long-term safe and stable operation of the device is ensured while manpower and material resource investment of mechanical pump fault diagnosis is reduced.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology for vacuum mechanical pumps in semiconductor processes, and particularly relates to a fault diagnosis method for mechanical pumps based on multi-source information and its model construction method. Background Technology

[0002] With the development of the semiconductor industry, the semiconductor manufacturing market is also gradually expanding. As a crucial auxiliary equipment in semiconductor manufacturing, vacuum pumps play an indispensable role in processes requiring a vacuum environment, such as chemical vapor deposition, plasma etching, cleaning, and heat treatment. Mechanical pumps, as a type of vacuum pump, utilize mechanical motion principles to extract gas and create a vacuum. However, when a component of a mechanical pump malfunctions, it can lead to the failure of the semiconductor manufacturing system, resulting in significant economic losses. Therefore, the urgent development and improvement of intelligent diagnostic technologies capable of effectively diagnosing mechanical pump faults has become an important component of the semiconductor manufacturing field.

[0003] Currently, intelligent fault diagnosis systems applied to mechanical pumps generally employ data-driven models, including machine learning algorithms such as gradient boosting trees, support vector machines, and logistic regression. The extracted mechanical pump fault features are input into the algorithm model, and the model is trained by parameter tuning to obtain the final fault diagnosis model.

[0004] In recent years, with the introduction of deep learning and neural networks as the main approach to fault diagnosis, the shortcomings of traditional machine learning algorithms have been gradually overcome, enabling intelligent fault diagnosis systems to better identify subtle faults. However, the establishment of most models relies on sufficient data. In actual manufacturing processes, fault data for mechanical pumps is often small-sample data with missing data, making it difficult for algorithms to obtain ideal results. The algorithms suffer from poor robustness and generalization ability, resulting in low accuracy, misdiagnosis, and missed diagnosis of faults. Summary of the Invention

[0005] To address the above problems, the first aspect of this invention provides a method for constructing a mechanical pump fault diagnosis model based on multi-source information, comprising the following steps:

[0006] Step 1: Collect continuous vibration signal data and corresponding fault category data of the mechanical pump under different conditions, including vibration signals of bearing inner ring fault, bearing outer ring fault, and bearing rolling element fault under no fault condition and motor load levels of 1hp, 2hp, and 3hp respectively.

[0007] Step 2: Decompose the original mechanical pump vibration signal data based on variational mode decomposition;

[0008] Step 3: Process the decomposed mechanical pump vibration signal based on time domain analysis and frequency domain analysis to obtain different characteristic parameters;

[0009] Step 4: The feature selection method based on intra-class and inter-class sensitivity fuses the different feature parameters obtained from multiple sources to select features with stronger representation capabilities and constructs a model dataset.

[0010] Step 5: Construct a mechanical pump fault diagnosis model based on ResNet and input multi-source information fusion features for model training and testing to obtain the final mechanical pump fault diagnosis model.

[0011] Preferably, the variational mode decomposition in step 2 decomposes the acquired original mechanical pump vibration signal f(t) into a discrete number of mode signals u. k (t);

[0012] First, the original mechanical pump vibration signal f(t) is decomposed into K intrinsic mode function components (IMFs) with different center frequencies. Each IMF can be defined as an AM-FM signal u. k (t), the calculation formula is:

[0013] u k (t)=A k (t)coS(φ k ′(t)) (1)

[0014] Among them, A k (t) represents the component u k The instantaneous amplitude of (t), φ k (t) represents the component u k The phase of (t), which is a non-decreasing function, i.e., φ′ k (t)>0, obtained from the following relationship:

[0015]

[0016] Where, ω k (t) represents u k The instantaneous frequency of (t);

[0017] Then, based on the constructed variational mode decomposition model, the optimal decomposition result u is found. k (t), the specific process is as follows:

[0018] The original mechanical pump vibration signal is decomposed into K modal components (IMFs), each with a finite bandwidth and concentrated around its center frequency during the decomposition process; that is, the frequency components contained in the mode are all around ω. k Near (t), the modal components ω of the original mechanical pump vibration signal f(t) are calculated using equations (3) to (10). k The bandwidth process of (t):

[0019] Each IMF is analyzed using Hilbert transform, and the one-sided spectrum of each IMF is obtained after analysis:

[0020]

[0021] Where δ(t) represents the Diclave function;

[0022] By combining each modal function with By multiplying, the center spectrum of the mode is modulated onto the corresponding baseband:

[0023]

[0024] in, This indicates exponential mixing with a predicted center frequency.

[0025] The bandwidth of the original mechanical pump vibration signal was estimated using H1 Gaussian smoothing and the square of the norm, and the bandwidth was adjusted according to the IMF setting for {u k By applying constraints, a constrained variational model is ultimately established that minimizes the sum of bandwidths for all IMFs.

[0026]

[0027] Wherein, the preset number of modes to be decomposed is K (K∈Z), {u k} = {u1, u2, ..., u k} represents the set of modal components obtained from the variational mode decomposition of the original mechanical pump vibration signal, {ω k}={ω1,ω2,…,ω k} represents the set of center frequencies of the modal components, and f represents the input signal;

[0028] By introducing the Lagrange operator λ and the quadratic penalty factor α, the constrained variational problem is transformed into an unconstrained problem, ensuring the absolute integrability of the decomposed signal in the above equation. The optimized calculation formula is as follows:

[0030]

[0031] Where L() represents the Laplace transform, and α represents the fidelity of the reconstructed signal;

[0032] The alternating multiplier algorithm is used to solve the unconstrained variational problem. Two variables are fixed, and the third variable is updated, thus achieving the update of modal components and center frequency, and the frequency domain component u. k The updated expression is:

[0033]

[0034] Among them, uk (ω) represents u k The corresponding Fourier transform;

[0035] Center frequency ω k The update expression is:

[0036]

[0037] When ω≥0, double lifting is applied to the Lagrange operator:

[0038] λ(ω)=λ(ω)+γ(f(ω)-∑ K u k (ω)) (10)

[0039] Where γ represents the noise tolerance function;

[0040] Repeat the update iterations until the preset maximum number of iterations is met;

[0041] The original mechanical pump vibration signal f(t) can be denoised by variational mode decomposition to obtain the preprocessed signal result F(t).

[0042] Preferably, step 3 involves data processing based on time-domain and frequency-domain analysis. After obtaining multiple time-domain and frequency-domain features of the denoised mechanical pump vibration signal F(t), the column vectors with different feature vectors in the time and frequency domains are concatenated to obtain the feature-extracted signal.

[0043] F(X) = (s1, s2, ..., s m p1, p2, ..., p n (11)

[0044] Among them, s m p represents the column vector of the m-th time-domain eigenvector. n This represents the column vector of the nth frequency domain eigenvector.

[0045] Preferably, the specific process of the time-domain analysis is as follows:

[0046] To obtain different eigenvalues, let X = {x1, x2, ..., x...} n Let F(t) be the discrete-time sequence of the denoised mechanical pump vibration signal, with n sampling points. Then, time-domain features are extracted from this sequence; multiple time-domain features are used:

[0047] The formula for calculating the time-domain mean characteristic is:

[0048]

[0049] Where s1 represents the column vector of time-domain mean features used in subsequent feature extraction output signal channel splicing;

[0050] The formula for calculating the time-domain variance characteristic is:

[0051]

[0052] Where s2 represents the column vector of temporal variance features used in subsequent feature extraction output signal channel splicing;

[0053] The formula for calculating the peak characteristics in the time domain is:

[0054]

[0055] Where s3 represents the column vector of time-domain peak features used in subsequent feature extraction output signal channel splicing;

[0056] The formula for calculating temporal kurtosis features is:

[0057]

[0058] Where s4 represents the column vector of temporal kurtosis features used in subsequent feature extraction output signal channel splicing;

[0059] The formula for calculating the root mean square feature of temporal kurtosis is:

[0060]

[0061] Where s5 represents the column vector of the root mean square feature of temporal kurtosis used in subsequent feature extraction output signal channel splicing;

[0062] The formula for calculating the time-domain impulse factor characteristic is:

[0063]

[0064] Where s6 represents the column vector of time-domain pulse factor features used in subsequent feature extraction output signal channel splicing;

[0065] The formula for calculating the time-domain skewness characteristic is:

[0066]

[0067] Where s7 represents the column vector of the time-domain skewness feature used in the subsequent concatenation of the output signal channels for feature extraction.

[0068] Preferably, the frequency domain analysis process is as follows:

[0069] The noise-reduced mechanical pump vibration signal F(t) is transformed from the time domain to the frequency domain using the Discrete Fourier Transform, and its calculation formula is as follows:

[0070]

[0071] The above equation is a discrete Fourier transform pair, where x(kT) represents the sampled value, n represents the number of sampling points in the sampling sequence, T represents the sampling interval, k represents the index of the discrete value in the time domain, and i represents the index of the discrete value in the frequency domain.

[0072] For the frequency domain analysis of the noise-reduced mechanical pump vibration signal, multiple frequency domain features are also sampled:

[0073] The formula for calculating the average frequency characteristic in the frequency domain is:

[0074]

[0075] Where p1 represents the column vector of the frequency domain average frequency features used in the subsequent feature extraction output signal channel splicing;

[0076] The formula for calculating the center frequency characteristic in the frequency domain is:

[0077]

[0078] Where p2 represents the column vector of the frequency domain center frequency feature used in the subsequent concatenation of the output signal channels for feature extraction;

[0079] The formula for calculating the root mean square frequency characteristic in the frequency domain is:

[0080]

[0081] Where p3 represents the column vector of time-domain mean features used in subsequent feature extraction output signal channel splicing;

[0082] Where P(k) represents the discrete-time sequence X = {x1, x2, ..., xk} n The spectrum obtained after discrete Fourier transform, where K represents the number of spectral lines, f k This represents the frequency value of the k-th modal line.

[0083] Preferably, step 4 specifically includes:

[0084] The feature distance between and within classes is used as the key to sensitivity discrimination; that is, the smaller the feature distance within a class, the more similar the classes are, and the larger the feature distance between classes, the greater the difference between classes. The average Euclidean distance of the feature values ​​of different mechanical pump vibration signals after feature extraction is:

[0085]

[0086] Among them, H i Let X represent the total distance between the i-th column vector of the feature extraction signal F(X) and other signals.n，l Let X represent the l-th feature corresponding to the nth column vector, and X... i，l Let represent the l-th feature corresponding to the i-th column vector, and N represent the total feature dimension of the column vectors. Based on the sum of distances between each column vector and other feature column vectors, to obtain features with stronger representational power, the column vector combination with the largest sum is selected as the filtered result G(X), and this is used as the output of the feature fusion and filtering process for model training.

[0087] G(X) = (S1, S2, ..., S...) α P1, P2, ..., P β (26)

[0088] Among them, S α P represents the column vector of the α-th temporal feature vector after fusion and filtering. β This represents the column vector of the β-th frequency domain feature vector after fusion and filtering. After feature fusion and filtering, α < m and β < n.

[0089] Preferably, the mechanical pump fault diagnosis model constructed based on ResNet in step 5 has the following specific structure:

[0090] It includes convolutional layers, pooling layers, fully connected layers, and residual calculation modules. The multi-source information fusion features first pass through 7*7 convolutional layers and max pooling layers to iteratively extract more complex features and reduce the number of parameters. Then, 59 residual calculation modules are repeatedly stacked to ensure that the learned features reach the optimal level. Finally, the fault diagnosis results are output through average pooling layers, fully connected layers, and Softmax.

[0091] The overall structure of the residual calculation module is as follows: the output value of the shallow layer passes through a 1*1 convolutional layer, a 3*3 convolutional layer, and a 1*1 convolutional layer, and ReLU is used as the activation function between each layer. Finally, the number of layers of the ResNet network is increased to 182, and it is used for the mechanical pump fault diagnosis model.

[0092] Preferably, the loss function used for model training in step 5 is the cross-entropy loss function, and the expression of the cross-entropy loss function is:

[0093]

[0094] Where loss(r, class) represents the cross-entropy loss function, r represents the predicted classification result, class represents the sample label of the fault data, e represents the exponential constant, v represents the sample class, and r class r represents the classification result of the real sample. v This indicates the classification result of the sample label belonging to category v.

[0095] A second aspect of the present invention provides a method for diagnosing mechanical pump faults based on multi-source information, which involves deploying a mechanical pump fault diagnosis model constructed by the method described in the first aspect into a detection system; and includes the following process:

[0096] S1, real-time acquisition of mechanical pump vibration signal data;

[0097] S2, the vibration signal data is decomposed through variational mode decomposition;

[0098] S3 uses time-domain and frequency-domain analysis to obtain different characteristic parameters of the decomposed vibration signal data;

[0099] S4, employing a feature selection method based on intra- and inter-class sensitivity to fuse multi-source information for different feature parameters;

[0100] S5, input the multi-source information fusion features of the decomposed vibration signal data into the constructed mechanical pump fault diagnosis model;

[0101] S6 outputs the detection results and issues a warning if a fault is detected.

[0102] A third aspect of the present invention provides a mechanical pump fault diagnosis device based on multi-source information. The device includes at least one processor and at least one memory, the processor and the memory being coupled together. The memory stores a computer-executable program of a mechanical pump fault diagnosis model constructed by the construction method described in the first aspect. When the processor executes the computer-executable program stored in the memory, it causes the processor to execute a mechanical pump fault diagnosis method based on multi-source information.

[0103] A fourth aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program or instructions for a mechanical pump fault diagnosis model constructed by the construction method described in the first aspect, wherein when the program or instructions are executed by a processor, the processor executes a mechanical pump fault diagnosis method based on multi-source information.

[0104] Compared with the prior art, the present invention has the following beneficial effects:

[0105] This invention aims to propose a fault diagnosis method for mechanical pumps based on diverse features and multi-source information, addressing the challenges of current intelligent fault diagnosis methods for mechanical pumps. This method introduces a strategy of diverse features and multi-source information. Vibration signals from multiple measuring points under different fault conditions of the mechanical pump are collected and subjected to time-domain, frequency-domain, and time-frequency-domain analyses to obtain different parameters. A multi-source information fusion method is employed to obtain fused features with high-dimensional hybrid characteristics. Finally, the fused features are used to train a deep learning network for fault classification. This effectively solves the problems of poor algorithm robustness, poor generalization ability, and low accuracy in mechanical pump fault diagnosis, thereby enabling real-time monitoring of the mechanical pump's state parameters. This reduces the manpower and material resources required for mechanical pump fault diagnosis while ensuring the long-term safe and stable operation of the device. This will bring more efficient and intelligent application solutions for mechanical pumps in the semiconductor manufacturing field. Attached Figure Description

[0106] Figure 1 This is a flowchart of the overall solution of the present invention.

[0107] Figure 2 This is a structural diagram of the mechanical pump fault diagnosis model based on ResNet according to the present invention.

[0108] Figure 3 This is a schematic diagram of the traditional residual calculation module.

[0109] Figure 4 This is a schematic diagram of the residual calculation module designed in this invention.

[0110] Figure 5 This is a simplified structural diagram of the mechanical pump fault diagnosis device in Example 2. Detailed Implementation

[0111] The present invention will be further described below with reference to embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0112] Example 1:

[0113] This invention provides a mechanical pump fault diagnosis method based on diverse characteristics and multi-source information, the overall idea of ​​which is as follows: Figure 1As shown, firstly, continuous vibration signals of a mechanical pump under different conditions and corresponding fault categories are collected and generated, including vibration signals of bearing inner ring faults, bearing outer ring faults, and bearing rolling element faults under no-fault conditions and motor load levels of 1, 2, and 3 hp (horsepower), respectively. Then, the original mechanical pump vibration signals are decomposed based on variational mode decomposition. The denoised mechanical pump vibration signals are processed based on time-domain and frequency-domain analysis to obtain different feature parameters. Based on a feature selection method using intra- and inter-class sensitivity, the different feature parameters are fused from multiple sources to select features with stronger representational capabilities. A mechanical pump fault diagnosis model is constructed based on ResNet, and the multi-source information fusion features are input for network training. Finally, based on the trained ResNet mechanical pump fault diagnosis model, fault warnings and diagnoses are performed according to the real-time collected mechanical pump data.

[0114] 1. Dataset Collection and Creation

[0115] In this embodiment, the collection and fabrication of mechanical pump fault diagnosis data primarily relies on the SKF6205 bearing at the drive end of the mechanical pump. Data is collected from mechanical pump bearings that were fault-free during the full lifespan experiment, bearings with inner ring faults, bearings with outer ring faults, and bearings with rolling element faults. At a sampling frequency of 12kHz and load levels of 1, 2, and 3 hp, the aforementioned mechanical pump bearings were replaced, resulting in 10 categories of mechanical pump vibration data. The final collected vibration signal data is f(t), corresponding to the state category r. class The dataset includes: bearing inner ring failures under no-fault conditions and load levels of 1, 2, and 3 hp; bearing outer ring failures under load levels of 1, 2, and 3 hp; and bearing rolling element failures under load levels of 1, 2, and 3 hp. 120,000 data points were collected for each type of failure. 1200 data points were selected as one sample for each failure type, resulting in 100 samples per failure type. The dataset was then divided into a 3:1 training set to test set ratio.

[0116] 2. Variational Mode Decomposition

[0117] The goal of variational mode decomposition is to decompose the acquired raw mechanical pump vibration signal f(t) into a discrete number of mode signals u. k (t). First, the original mechanical pump vibration signal f(t) is decomposed into K intrinsic mode function components (IMFs) with different center frequencies. Each IMF can be defined as an AM-FM signal u. k (t), the calculation formula is:

[0118] u k (t)=A k (t)coS(φ k ′(t)) (1)

[0119] Among them, A k (t) represents the component u k The instantaneous amplitude of (t), φ k (t) represents the component u k The phase of (t), which is a non-decreasing function, i.e., φ′ k If (t) > 0, it can be obtained from the following relationship:

[0120]

[0121] Where, ω k (t) represents u k The instantaneous frequency of (t).

[0122] Then, based on the constructed variational mode decomposition model, the optimal decomposition result u is found. k (t).

[0123] The original mechanical pump vibration signal is decomposed into K modal components (IMFs), each with a finite bandwidth and concentrated around its center frequency during the decomposition process; that is, the frequency components contained in the mode are all around ω. k Near (t), the modal components ω of the original mechanical pump vibration signal f(t) are calculated using equations (3) to (10). k The bandwidth process of (t):

[0124] Each IMF is analyzed using Hilbert transform, and the one-sided spectrum of each IMF is obtained after analysis:

[0125]

[0126] Here, δ(t) represents the Diclave function.

[0127] By combining each modal function with By multiplying, the center spectrum of the mode is modulated onto the corresponding baseband:

[0128]

[0129] in, This indicates an exponential mixer with a predicted center frequency.

[0130] The bandwidth of the original mechanical pump vibration signal was estimated using H1Gaussian smoothness and the square of the norm, and the bandwidth was further calculated based on the IMF setting for {u k By applying constraints, a constrained variational model is ultimately established that minimizes the sum of bandwidths for all IMFs.

[0131]

[0132] Wherein, the preset number of modes to be decomposed is K (K∈Z), {u k} = {u1, u2, ..., u k} represents the set of modal components obtained from the variational mode decomposition of the original mechanical pump vibration signal, {ω k}={ω1,ω2,…,ω k} represents the set of center frequencies of the modal components, and f represents the input signal.

[0133] By introducing the Lagrange operator λ and the quadratic penalty factor α, the constrained variational problem is transformed into an unconstrained problem, ensuring the absolute integrability of the decomposed signal in the above equation. The optimized calculation formula is as follows:

[0134]

[0135] Where L() represents the Laplace transform, and α represents the fidelity of the reconstructed signal.

[0136] The alternating multiplier algorithm is used to solve the unconstrained variational problem. Two variables are fixed, and the third variable is updated, thus achieving the update of modal components and center frequency, and the frequency domain component u. k The updated expression is:

[0137]

[0138] Among them, u k (ω) represents u k The corresponding Fourier transform.

[0139] Center frequency ω k The update expression is:

[0140]

[0141] When ω≥0, double lifting is applied to the Lagrange operator:

[0142] λ(ω)=λ(ω)+γ(f(ω)-∑ K u k (ω)) (10)

[0143] Where γ represents the noise tolerance function.

[0144] Repeat the update iterations until the preset maximum number of iterations is met.

[0145] The variational mode decomposition is performed on the original mechanical pump vibration signal. An optimization objective function for the variational mode decomposition features is constructed, and the optimal decomposition result u is found through optimization. kThe original mechanical pump vibration signal f(t) is used as the preprocessed data signal. Noise reduction can be achieved by performing variational mode decomposition on the original mechanical pump vibration signal f(t), yielding the preprocessed signal result F(t).

[0146] 3. Data processing based on time-domain and frequency-domain analysis

[0147] The noise-reduced mechanical pump vibration signal F(t) was analyzed in the time domain and frequency domain to obtain different characteristic parameters.

[0148] Time-domain analysis is a direct and effective signal processing method. To obtain different eigenvalues, let X = {x1, x2, ..., x...} n Let F(t) be the discrete-time sequence of the denoised mechanical pump vibration signal, with n sampling points. Then, time-domain features are extracted from this sequence. In mechanical pump vibration signals containing a lot of interference, using only a single time-domain feature is prone to errors, so multiple time-domain features are used:

[0149] The formula for calculating the time-domain mean characteristic is:

[0150]

[0151] Where s1 represents the column vector of time-domain mean features used in subsequent feature extraction output signal channel splicing;

[0152] The formula for calculating the time-domain variance characteristic is:

[0153]

[0154] Where s2 represents the column vector of temporal variance features used in subsequent feature extraction output signal channel splicing;

[0155] The formula for calculating the peak characteristics in the time domain is:

[0156]

[0157] Where s3 represents the column vector of time-domain peak features used in subsequent feature extraction output signal channel splicing;

[0158] The formula for calculating temporal kurtosis features is:

[0159]

[0160] Where s4 represents the column vector of temporal kurtosis features used in subsequent feature extraction output signal channel splicing;

[0161] The formula for calculating the root mean square feature of temporal kurtosis is:

[0162]

[0163] Where s5 represents the column vector of the root mean square feature of temporal kurtosis used in subsequent feature extraction output signal channel splicing;

[0164] The formula for calculating the time-domain impulse factor characteristic is:

[0165]

[0166] Where s6 represents the column vector of time-domain pulse factor features used in subsequent feature extraction output signal channel splicing;

[0167] The formula for calculating the time-domain skewness characteristic is:

[0168]

[0169] Wherein, s7 represents the column vector of the time-domain skewness feature used in the subsequent concatenation of the output signal channels for feature extraction;

[0170] Frequency domain analysis transforms the noise-reduced mechanical pump vibration signal F(t) from the time domain to the frequency domain using the Discrete Fourier Transform (DFT). The calculation formula is as follows:

[0171]

[0172] The above equation is a discrete Fourier transform pair, where x(kT) represents the sampled value, n represents the number of sampling points in the sequence, T represents the sampling interval, k represents the index of the discrete value in the time domain, and i represents the index of the discrete value in the frequency domain.

[0173] For the frequency domain analysis of the noise-reduced mechanical pump vibration signal, multiple frequency domain features were also sampled:

[0174] The formula for calculating the average frequency characteristic in the frequency domain is:

[0175]

[0176] Where p1 represents the column vector of the frequency domain average frequency features used in the subsequent feature extraction output signal channel splicing;

[0177] The formula for calculating the center frequency characteristic in the frequency domain is:

[0178]

[0179] Where p2 represents the column vector of the frequency domain center frequency feature used in the subsequent concatenation of the output signal channels for feature extraction;

[0180] The formula for calculating the root mean square frequency characteristic in the frequency domain is:

[0181]

[0182] Where p3 represents the column vector of time-domain mean features used in subsequent feature extraction output signal channel splicing;

[0183] Where P(k) represents the discrete-time sequence X = {x1, x2, ..., xk} n The spectrum obtained after discrete Fourier transform, where K represents the number of spectral lines, f k This represents the frequency value of the k-th modal line.

[0184] After obtaining multiple time-domain and frequency-domain features of the denoised mechanical pump vibration signal F(t) through data processing based on time-domain and frequency-domain analysis, the column vectors with different feature vectors in the time and frequency domains can be concatenated to obtain the feature-extracted signal.

[0185] F(X) = (s1, s2, ..., s m p1, p2, ..., p n ) (twenty three)

[0186] Among them, s m p represents the column vector of the m-th time-domain eigenvector. n Let m be the column vector representing the nth frequency domain eigenvector. In this invention, m = 7 and n = 3.

[0187] 4. Multi-source information fusion

[0188] The feature selection criterion based on intra-class and inter-class sensitivity is a feature fusion and selection method that uses the feature extraction signal F(X) obtained by splicing multiple time-domain and frequency-domain feature vector channels to select features with stronger representation capabilities.

[0189] Feature selection criteria based on intra- and inter-class sensitivity differ from simply superimposing and fusing multiple feature parameters. Its main principles are:

[0190] Similarity means that the features in the time domain and the features in the frequency domain must be similar to each other and have a certain relationship.

[0191] Difference, that is, there must be a difference between the time domain features and the frequency domain features.

[0192] The feature distance between intra-class and inter-class features is used as the key to sensitivity discrimination. That is, the smaller the intra-class feature distance, the more similar the intra-class features are, and the larger the inter-class feature distance, the greater the difference between the inter-class features are. The features selected in this way are the most sensitive.

[0193] The average Euclidean distance of the feature values ​​of different mechanical pump vibration signals after feature extraction is:

[0194]

[0195] Among them, H iLet X represent the total distance between the i-th column vector of the feature extraction signal F(X) and other signals. n，l Let X represent the l-th feature corresponding to the nth column vector, and X... i，l Let represent the l-th feature corresponding to the i-th column vector, and N represent the total feature dimension of the column vectors. Based on calculating the sum of distances between each column vector and other feature column vectors, to obtain features with stronger representational power, the column vector combination with the largest sum is selected as the filtered result G(X), and this is used as the output of the feature fusion and filtering process for model training.

[0196] G(X) = (S1, S2, ..., S...) α P1, P2, ..., P β (25)

[0197] Among them, S α P represents the column vector of the α-th temporal feature vector after fusion and filtering. β Let represent the column vector of the β-th frequency domain feature vector after fusion and filtering. In this invention, after feature fusion and filtering, α < m and β < n.

[0198] 5. Construction of Mechanical Pump Fault Diagnosis Model

[0199] The mechanical pump fault diagnosis model is built based on ResNet, such as... Figure 2 As shown, the model structure includes: convolutional layer, pooling layer, fully connected layer and residual calculation module. The multi-source information fusion features first pass through 7*7 convolutional layer and max pooling layer to iteratively extract more complex features and reduce the number of parameters. Then, 59 residual calculation modules are repeatedly stacked to ensure that the learned features reach the optimal level. Finally, the fault diagnosis results are output through average pooling layer, fully connected layer and Softmax.

[0200] The ResNet-based mechanical pump fault diagnosis model optimizes the output of each convolutional layer using the residual calculation module. A schematic diagram of the traditional residual calculation module is shown below. Figure 2 The residual calculation module structure used in this invention is as follows: Figure 3 As shown, the residual calculation module used in this invention is as follows: Figure 3 As shown on the right, its main structure is that the output value of the shallow layer passes through a 1*1 convolutional layer, a 3*3 convolutional layer, and a 1*1 convolutional layer, and ReLU is used as the activation function between each layer. Finally, the number of layers of the ResNet network is increased to 182 layers, and it is used in the mechanical pump fault diagnosis model to obtain better network training effect and more accurate fault diagnosis results.

[0201] The residual calculation includes an identity mapping for each of the convolutional layers in the mechanical pump fault diagnosis model;

[0202] Identity mapping occurs when, after a network has been stacked and deepened, it has reached an optimal level of learning from the input features. The remaining stacked layers of the deep network then stop learning and modifying those features, simply outputting the optimal value sequentially, thus completing the identity mapping. This ensures that deeper networks do not perform worse than shallower networks, effectively preventing network degradation.

[0203] The main implementation of the identity mapping concept is through residual modules that enable the network to perform identity mapping. z represents the output value of the shallow layer, H(z) represents the output value of the deep layer, and D(z) = H(z) - z is the nonlinear transformation output value of the two- or three-layer structure between them. When the features learned by the shallow layer's output value z have reached their optimal level, D(z) will automatically approach 0, allowing z to be passed down along the identity mapping path. This achieves the goal of ensuring identity mapping in the remaining layers of the deep network, preventing the network training results from deteriorating, provided the shallow layer output is sufficiently good.

[0204] 6. Model Training

[0205] The expression for extracting key features from the training set by the convolutional layer based on the feature fusion and selection result G(X) is as follows:

[0206] h1=CONV(G(X)) (26)

[0207] Where h1 represents the key features of the training set extracted by the convolutional layer; CONV represents the convolutional layer; and G(X) represents the mechanical pump vibration signal features of the training set after feature fusion and filtering.

[0208] The expression for the characteristics of the mechanical pump vibration signal after dimensionality reduction of the training set is as follows:

[0209] h2 = MAX_POOLING(h1) (27)

[0210] Where h2 represents the mechanical pump vibration signal features after dimensionality reduction of the training set, and MAX_POOLING represents the maximized pooling layer;

[0211] The expression for further extraction of the vibration signal features of the mechanical pump after dimensionality reduction of the training set is as follows:

[0212] F(h2)=D(h2)-S(h2) (28)

[0213] Where F(h2) represents the mechanical pump vibration signal features further extracted from the mechanical pump vibration signal features after dimensionality reduction of the training set, S(h2) represents the output value of the shallow convolutional layer, and D(h2) represents the output value of the deep convolutional layer.

[0214] The extracted features are summarized by the fully connected layer, resulting in the following expression summarizing the characteristics of the mechanical pump vibration signal from the training set:

[0215] h3=FC(F(h2)) (29) where h3 represents the characteristics of mechanical pump vibration signal summarized in the training set, and FC represents the fully connected layer;

[0216] The expression for outputting the classification result of the mechanical pump vibration signal features in the training set is:

[0217] h4 = softmax(h3) (30)

[0218] Where h4 represents the classification result of the mechanical pump vibration signal features in the training set, and softmax represents the softmax classifier.

[0219] The loss function in the ResNet-based mechanical pump fault diagnosis model is the cross-entropy loss function, and the expression of the cross-entropy loss function is as follows:

[0220]

[0221] Where loss(r, class) represents the cross-entropy loss function, r represents the predicted classification result, class represents the sample label of the fault data, e represents the exponential constant, v represents the sample class, and r class r represents the classification result of the real sample. v This indicates the classification result of the sample label belonging to category v.

[0222] 7. Model Deployment and Application

[0223] The mechanical pump fault diagnosis model constructed using the above method is deployed in the detection system; and includes the following process:

[0224] S1, real-time acquisition of mechanical pump vibration signal data;

[0225] S2, the vibration signal data is decomposed through variational mode decomposition;

[0226] S3 uses time-domain and frequency-domain analysis to obtain different characteristic parameters of the decomposed vibration signal data;

[0227] S4, employing a feature selection method based on intra- and inter-class sensitivity to fuse multi-source information for different feature parameters;

[0228] S5, input the multi-source information fusion features of the decomposed vibration signal data into the constructed mechanical pump fault diagnosis model;

[0229] S6 outputs the detection results and issues a warning if a fault is detected.

[0230] Example 2:

[0231] like Figure 5 As shown, this invention also provides a mechanical pump fault diagnosis device based on multi-source information. The device includes at least one processor and at least one memory, as well as a communication interface and an internal bus. The memory stores a computer-executable program of a mechanical pump fault diagnosis model constructed by the construction method described in Embodiment 1. When the processor executes the computer-executable program stored in the memory, it can execute a mechanical pump fault diagnosis method based on multi-source information. The internal bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, the bus in the accompanying drawings is not limited to only one bus or one type of bus. The memory may include high-speed RAM memory, and may also include non-volatile memory (NVM), such as at least one disk storage device, and may also be a USB flash drive, portable hard drive, read-only memory, disk, or optical disk, etc.

[0232] The device can be provided as a terminal, server, or other form of device.

[0233] Figure 5 This is a block diagram illustrating an illustrative device. The device may include one or more of the following components: a processing component, a memory, a power supply component, a multimedia component, an audio component, an input / output (I / O) interface, a sensor component, and a communication component. The processing component typically controls the overall operation of the electronic device, such as operations associated with display, telephone calls, data communication, camera operation, and recording operations. The processing component may include one or more processors to execute instructions to perform all or part of the steps of the methods described above. Furthermore, the processing component may include one or more modules to facilitate interaction between the processing component and other components. For example, the processing component may include a multimedia module to facilitate interaction between the multimedia component and the processing component.

[0234] Memory is configured to store various types of data to support the operation of electronic devices. Examples of this data include instructions for any application or method used to operate on an electronic device, contact data, phonebook data, messages, pictures, videos, etc. Memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0235] A power supply component provides power to various components of an electronic device. The power supply component may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the electronic device. A multimedia component includes a screen that provides an output interface between the electronic device and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors may sense not only the boundaries of the touch or swipe action but also the duration and pressure associated with the touch or swipe operation. In some embodiments, the multimedia component includes a front-facing camera and / or a rear-facing camera. When the electronic device is in an operating mode, such as a shooting mode or a video mode, the front-facing camera and / or the rear-facing camera may receive external multimedia data. Each front-facing and rear-facing camera may be a fixed optical lens system or have focal length and optical zoom capabilities.

[0236] The audio component is configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the electronic device is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals may be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals. The I / O interface provides an interface between the processing component and peripheral interface modules, which may be a keyboard, click wheel, buttons, etc. These buttons may include, but are not limited to, a home button, volume buttons, a power button, and a lock button.

[0237] The sensor assembly includes one or more sensors for providing state assessments of various aspects of the electronic device. For example, the sensor assembly can detect the on / off state of the electronic device, the relative positioning of components such as the display and keypad of the electronic device, changes in the position of the electronic device or a component of the electronic device, the presence or absence of user contact with the electronic device, the orientation or acceleration / deceleration of the electronic device, and temperature changes of the electronic device. The sensor assembly may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. The sensor assembly may also include an optical sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly may also include an accelerometer, a gyroscope, a magnetometer, a pressure sensor, or a temperature sensor.

[0238] The communication component is configured to facilitate wired or wireless communication between electronic devices and other devices. The electronic device can access wireless networks based on communication standards, such as WiFi, 2G, or 3G, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, the communication component further includes a Near Field Communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on Radio Frequency Identification (RFID) technology, Infrared Data Association (IrDA) technology, Ultra-Wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0239] In an exemplary embodiment, the electronic device may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the methods described above.

[0240] Example 3:

[0241] The present invention also provides a computer-readable storage medium storing a computer program or instructions for a mechanical pump fault diagnosis model constructed by the construction method described in Embodiment 1. When the program or instructions are executed by a processor, the processor can perform a mechanical pump fault diagnosis method based on multi-source information.

[0242] Specifically, a system, apparatus, or device may be provided equipped with a readable storage medium on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system, apparatus, or device reads and executes the instructions stored in the readable storage medium. In this case, the program code read from the readable medium itself can implement the functions of any of the embodiments described above, therefore, the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of the present invention.

[0243] The aforementioned storage media can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disks or optical disks (such as CD-ROM, CD-R, CD-RW, DVD-20ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tape, etc. The storage media can be any available medium accessible to general-purpose or special-purpose computers.

[0244] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0245] It should be understood that the storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and storage medium can reside in application-specific integrated circuits (ASICs). Alternatively, the processor and storage medium can exist as discrete components in a terminal or server.

[0246] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0247] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0248] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0249] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for constructing a mechanical pump fault diagnosis model based on multi-source information, characterized in that, The method comprises the following steps: Step 1, collecting mechanical pump continuous vibration signal data and corresponding fault category data in different states, including no-fault state and bearing inner ring fault, bearing outer ring fault and bearing rolling element fault under the conditions of 1 hp, 2 hp and 3 hp motor load; Step 2, decomposing the original mechanical pump vibration signal data based on variational mode decomposition; Step 3, processing the decomposed mechanical pump vibration signal based on time domain analysis and frequency domain analysis to obtain different characteristic parameters; Step 4, performing multi-source information fusion on the obtained different characteristic parameters based on the feature selection method of intra-class and inter-class sensitivity to screen features with stronger representation ability and construct a model data set; Step 5, constructing a mechanical pump fault diagnosis model based on ResNet, inputting the multi-source information fusion features for model training and testing, and obtaining the final mechanical pump fault diagnosis model.

2. The mechanical pump fault diagnosis model construction method based on multi-source information according to claim 1, wherein: The variational modal decomposition in step 2 is to decompose the collected original mechanical pump vibration signal f(t) into a discrete number of modal signals u k (t); Firstly, the original mechanical pump vibration signal f(t) is decomposed into K intrinsic mode function components IMF with different central frequencies, each IMF can be defined as an AM-FM signal u k (t), and the calculation formula is: u k (t) = A k (t) cos(φ k ′(t)) (1) where A k (t) denotes the instantaneous amplitude of the component u k (t), and φ k (t) denotes the phase of the component u k (t), the function being non-decreasing, i.e. φ' k (t) > 0, is obtained from the relationship where ω k (t) represents the instantaneous frequency of u k (t). Then based on the constructed variational modal decomposition model, the optimal decomposition result u is found k The specific process is as follows: The original mechanical pump vibration signal is decomposed into K modal components IMF, each of which has a limited bandwidth and is concentrated around a respective center frequency during the decomposition, i.e. the frequency components contained in the mode are around ω k (t). Then the modal components ω k (t) of the original mechanical pump vibration signal f(t) are calculated by equations (3)-(10) and the frequency band bandwidth process of ω k (t) is as follows: each IMF is analyzed by Hilbert transform to obtain the one-sided spectrum of each IMF after analysis: wherein, δ(t) represents Dirac function; The center frequency spectrum of each mode is modulated to the corresponding baseband by multiplying each mode function with ​ wherein, represents an exponential mixing of the estimated center frequency; The original mechanical pump vibration signal bandwidth is estimated by H1 Gaussian smoothing and square of norm, and the constraint is made on {u k} according to the setting of IMF, and finally the constraint variation model capable of predicting the sum of all IMF bandwidths is established: Wherein, the preset number of modes to be decomposed is K (K is an integer), {u k}={u1, u2, …, u k} represents a set of modal components obtained by variational modal decomposition of the original mechanical pump vibration signal, {ω k}={ω1, ω2, …, ω k} represents a set of center frequencies of the modal components, and f represents an input signal. Lagrange operator λ and quadratic term penalty factor α are introduced to convert the constrained variational problem into an unconstrained problem to ensure the absolute integrability of the decomposed signal, and the optimized calculation formula is as follows: wherein, L() represents Lagrange transform, and α represents the authenticity of the enhanced signal reconstruction; The non-constrained variation problem is solved by using an alternating multiplier algorithm, two variables are fixed, and another variable is updated, so as to realize the updating of modal components and central frequency, and the frequency domain component u k The expression for updating is: where u k (ω) denotes the Fourier transform of u k corresponding Fourier transform; Center frequency ω k The update expression is when ω≥0, the Lagrange operator is doubly promoted: λ(ω) = λ(ω) + γ(f(ω) - ∑ K u k (ω)) (10) wherein, γ represents noise tolerance function; the update iteration is repeated until the preset maximum iteration number is met; the variational mode decomposition of the original mechanical pump vibration signal f(t) can complete signal denoising to obtain the preprocessed signal result F(t).

3. The method of claim 1, wherein the method comprises: After obtaining the multiple time domain features and frequency domain features of the denoised mechanical pump vibration signal F(t) through the data processing based on time domain analysis and frequency domain analysis in the step 3, the column vectors of the feature vectors in the time domain and the frequency domain are spliced to obtain the signal extracted by the feature: F(X) = (s1, s2,..., s m , p1, p2,..., p n )(11) where s m represents a column vector of the mth time-domain feature vector, p n represents a column vector of the nth frequency-domain feature vector.

4. The method of claim 3, wherein the method further comprises: The specific process of the time domain analysis is as follows: To get different feature values, let X = {x1, x2, …, x n} be the discrete-time sequence of the de-noised mechanical pump vibration signal F(t) with n sampling points, and then extract the time-domain features of the sequence; adopt multiple time-domain features: The calculation formula of the time domain mean feature is as follows: wherein, s1 represents the column vector of the time domain mean feature used for subsequent signal channel splicing of feature extraction output; The calculation formula of the time domain variance feature is as follows: wherein, s2 represents the column vector of the time domain variance feature used for subsequent signal channel splicing of feature extraction output; The calculation formula of the time domain peak value feature is as follows: wherein, s3 represents the column vector of the time domain peak value feature used for subsequent signal channel splicing of feature extraction output; The calculation formula of the time domain kurtosis feature is as follows: wherein, s4 represents the column vector of the time domain kurtosis feature used for subsequent signal channel splicing of feature extraction output; The calculation formula of the time domain kurtosis root mean square feature is as follows: wherein, s5 represents the column vector of the time domain kurtosis root mean square feature used for subsequent signal channel splicing of feature extraction output; The calculation formula of the time domain pulse factor feature is as follows: wherein, s6 represents the column vector of the time domain pulse factor feature used for subsequent signal channel splicing of feature extraction output; The calculation formula of the time domain skewness feature is as follows: Wherein, s7 represents the time domain skewness feature in the subsequent column vector used for feature extraction output signal channel splicing.

5. The method of claim 3, wherein the method further comprises: The specific process of the frequency domain analysis is: The noise-reduced mechanical pump vibration signal F(t) is transformed from the time domain state to the frequency domain state, and the discrete Fourier transform is used, and the calculation formula is: The above formula is the discrete Fourier transform, wherein x(kT) represents the sampling value, n represents the sampling sequence point number, T represents the sampling interval, k represents the serial number of the time domain discrete value, and i represents the serial number of the frequency domain discrete value; For the frequency domain analysis of the noise-reduced mechanical pump vibration signal, a plurality of frequency domain features are also sampled: The calculation formula of the frequency domain average frequency feature is: Wherein, p1 represents the frequency domain average frequency feature in the subsequent column vector used for feature extraction output signal channel splicing; The calculation formula of the frequency domain center frequency feature is: Wherein, p2 represents the frequency domain center frequency feature in the subsequent column vector used for feature extraction output signal channel splicing; The calculation formula of the frequency domain root mean square frequency feature is: Wherein, p3 represents the time domain mean feature in the subsequent column vector used for feature extraction output signal channel splicing; Where P(k) represents the discrete-time sequence X = {x1, x2, ..., xk} n The spectrum obtained after discrete Fourier transform, where K represents the number of spectral lines, f k This represents the frequency value of the k-th modal line.

6. The method of claim 1, wherein the method further comprises: The step 4 is specifically: The feature distance value between the classes is taken as the key to sensitivity discrimination, that is, the smaller the feature distance within the class, the more similar the class is, and the larger the feature distance between the classes, the greater the difference between the classes; the average Euclidean distance of the feature values of different mechanical pump vibration signals after feature extraction is: wherein H i represents the total distance between the ith column vector signal of the feature extraction signal F(X) and other signals, X n，l represents the lth dimension feature corresponding to the nth column vector, and X i，l represents the lth dimension feature corresponding to the ith column vector, and N represents the total feature dimension of the column vector; according to the calculation of the sum of the distances between each column vector and other feature column vectors, in order to obtain features with stronger representation ability, the column vector combination with the maximum sum is selected as the screened result G(X), and is used as the output after feature fusion screening, for model training: G(X) = (S1, S2,..., S α , P1, P2,..., P β ) (26) wherein S α represents a column vector of the a-th time domain feature vector after fusion screening, P β represents a column vector of the β-th frequency domain feature vector after fusion screening, after feature fusion screening, α < m, β < n.

7. The method of claim 1, wherein the method further comprises: The mechanical pump fault diagnosis model is constructed based on ResNet in the step 5, and the specific structure is: It includes a convolution layer, a pooling layer, a full connection layer and a residual calculation module, the multi-source information fusion feature first passes through a 7*7 convolution layer and a maximum pooling layer to complete iterative extraction of more complex features and reduce the parameter order of magnitude, then passes through 59 residual calculation modules stacked repeatedly to ensure that the learned features are optimal, and finally passes through an average pooling layer, a full connection layer and Softmax to complete the output of the fault diagnosis result. The overall structure of the residual calculation module is that the output value of the shallow layer passes through a 1*1 convolution layer, a 3*3 convolution layer and a 1*1 convolution layer, and ReLU is used as the activation function between the layers, finally the number of layers of the ResNet network is deepened to 182 layers, and it is used for the mechanical pump fault diagnosis model.

8. The method of claim 1, wherein the method further comprises: The loss function used in the model training in the step 5 is a cross-entropy loss function, and the expression of the cross-entropy loss function is: wherein loss(r, class) represents a cross-entropy loss function, r represents a predicted classification result, class represents a sample label of the fault data, e represents an exponential constant, v represents a sample category, r class represents a true sample classification result, r v represents a classification result that the sample label belongs to the category v.

9. A mechanical pump fault diagnosis method based on multi-source information, characterized in that, The mechanical pump fault diagnosis model constructed by the construction method of any one of claims 1 to 8 is deployed in a detection system. And It includes the following processes: S1, real-time acquisition of mechanical pump vibration signal data; S2, the vibration signal data is decomposed by variational mode decomposition for decomposition processing; S3, time domain and frequency domain analysis are used to obtain different characteristic parameters of the decomposed vibration signal data; S4, the feature selection method of intra-class and inter-class sensitivity is used to perform multi-source information fusion on different characteristic parameters; S5, the multi-source information fusion feature of the decomposed vibration signal data is input into the constructed mechanical pump fault diagnosis model; S6, output the detection result, and if a fault occurs, perform early warning.

10. A mechanical pump fault diagnosis device based on multi-source information, characterized by: The device comprises at least one processor and at least one memory, which are coupled; the memory stores a computer-executed program of the mechanical pump fault diagnosis model constructed by the construction method in any one of claims 1 to 8; and the processor executes the computer-executed program stored in the memory, so that the processor executes a mechanical pump fault diagnosis method based on multi-source information.

Images