An adaptive fault diagnosis method and system for Roots-type compressor

Abstract

The present invention provides an adaptive fault diagnosis method for a Roots-type compressor, comprising obtaining a vibration signal to be measured transmitted from the Roots-type compressor; performing noise elimination and transformation processing on the vibration signal to be measured; and importing the processed vibration signal to be measured into a trained Roots-type compressor fault diagnosis model to obtain a corresponding diagnostic result. The Roots-type compressor fault diagnosis model is constructed based on an attention domain adaptive network and is trained by simultaneously inputting source domain samples with complete fault categories and target domain samples with incomplete fault categories, wherein the source domain samples and the target domain samples are derived from vibration signals of the Roots-type compressor under different operating conditions. Implementing the present invention can solve the problem of incorrect diagnostic results caused by missing training sample categories in existing transfer learning-based methods, thereby improving diagnostic accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment fault diagnosis, and in particular to an adaptive fault diagnosis method and system for a Roots-type compressor. Background Art

[0002] Roots compressors are rotary machines that rely on the relative rotation of two impellers to continuously compress and transport gas. Due to their stable operating characteristics, high volumetric efficiency, and strong adaptability, they are widely used in evaporation processes in industries such as electronics, petroleum, steel, building materials, chemicals, food, medical treatment, and papermaking.

[0003] As core equipment in the evaporation process, Roots compressors operate under complex operating conditions such as high speed and high temperature for long periods of time, making them extremely prone to damage and failure. Common Roots compressor failure modes include: damage to key and vulnerable parts such as the impeller, bearings, gears, and drive shaft; loose anchor bolts; motor winding short circuits, winding grounding or phase short circuits; broken rotor bars; and abnormal leakage. Among them, the impeller, bearings, and gears are the core components of the Roots compressor's power transmission and are also vulnerable parts. To prevent sudden failures of these core components, which could cause economic losses or even safety accidents, it is necessary to monitor and diagnose the working condition and performance of the Roots compressor's impeller, bearings, and gears. This is of great value and significance in ensuring normal and safe production operations and improving the company's economic benefits.

[0004] Since vibration signals are easy to measure and contain rich information, many fault detection methods based on vibration signals have been proposed to help improve the reliability and safety of rotating machinery. These fault detection methods based on vibration signals can be divided into fault feature extraction methods and machine learning methods.

[0005] Currently, traditional fault feature extraction methods rely on manual experience for fault diagnosis and are gradually being replaced by machine learning methods. However, most machine learning methods require large amounts of labeled data for training. Once the operating conditions change, data under the changed conditions must be collected and retrained, which greatly reduces the practical application of machine learning methods. To this end, many transfer learning-based methods have been proposed, which learn from a small number of labeled samples and then transfer them to a different operating condition. However, in real-world application scenarios with variable operating conditions, training sample categories may be missing, resulting in significant differences in the distribution of training and test data. This reduces the performance of these transfer learning-based methods and can lead to inaccurate fault diagnosis results.

[0006] Therefore, it is necessary to use a new method in the fault diagnosis of Roots compressors, which can solve the problem of incorrect diagnosis results caused by the missing training sample categories in the existing transfer learning-based method and improve the diagnostic accuracy. Summary of the Invention

[0007] The technical problem to be solved by the embodiments of the present invention is to provide an adaptive fault diagnosis method and system for Roots-type compressors, which can solve the problem of incorrect diagnosis results caused by the lack of training sample categories in existing transfer learning-based methods, realize accurate diagnosis of Roots-type compressor faults, and improve diagnostic accuracy.

[0008] In order to solve the above technical problems, an embodiment of the present invention provides an adaptive fault diagnosis method for a Roots-type compressor, the method comprising the following steps:

[0009] Obtain the vibration signal to be measured transmitted from the Roots compressor;

[0010] Based on a preset adaptive symplectic geometric mode decomposition algorithm and a preset wavelet scattering transform method, the vibration signal to be measured is subjected to denoising and transformation processing;

[0011] The vibration signal to be measured after denoising and transformation is introduced into a trained Roots compressor fault diagnosis model to obtain a diagnosis result of the vibration signal to be measured; wherein, the diagnosis result is one of normal, bearing inner ring fault, bearing outer ring fault, bearing rolling element fault, gear fault and impeller fault; the Roots compressor fault diagnosis model is constructed based on an attention domain adaptive network, and is trained by simultaneously inputting original domain samples with complete fault categories and target domain samples with incomplete fault categories; the original domain samples are formed by denoising and transforming multiple vibration signals with labels and complete fault categories under Roots compressor working condition A through the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method; the target domain samples are formed by denoising and transforming multiple vibration signals with labels but incomplete fault categories under Roots compressor working condition B through the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method.

[0012] The steps of executing the adaptive symplectic geometric mode decomposition algorithm are as follows:

[0013] The first step is to assume that the vibration signal to be de-noised is a={a1,a2,…,a n}; where n represents the length of signal a;

[0014] Then, define the trajectory matrix A of signal a as shown in the following formula (1):

[0015]

[0016] Where k represents the embedding dimension; the embedding dimension k is obtained by calculating the power spectral density of signal a and estimating the frequency of the maximum peak; where k = n / 3 if the normalized frequency is less than a given threshold, otherwise k = 1.2×(f s / f max );f s Expressed as sampling frequency, f max It is represented by the frequency of the maximum peak of the power spectrum density of signal a; τ represents the delay time; m = n-(k-1)τ;

[0017] The second step is to construct the Hamiltonian matrix B according to formula (2):

[0018]

[0019] Where C = A T A;

[0020] Step 3: Let D = B 2 , according to the definition of Hamiltonian matrix, D and B are both Hamiltonian matrices, and then a symplectic orthogonal matrix E is constructed, as shown in the following formula (3):

[0021]

[0022] Wherein, the matrix E is represented as an orthogonal symplectic matrix with symplectic matrix properties; F is represented as an upper triangular matrix, and the eigenvalue of the upper triangular matrix F is represented as λ i (i=1,2,…,k);f ij =0(i>j+1);

[0023] Step 4. If the matrix C is a real symmetric matrix, then the eigenvalues ​​of C are equal to the eigenvalues ​​of the upper triangular matrix F. According to the properties of the Hamiltonian matrix, the eigenvalues ​​of C are The eigenvalues ​​of the matrix C can be expressed as β1>β2>…>β k ;

[0024] Let P i (i=1,2,…,k) is the eigenvector of matrix C, let P i (i=1,2,…,k) are the eigenvectors of matrix C, and let And Z i =P i S i , then the initial reconstruction trajectory matrix Z is defined as follows:

[0025] Z=Z1+Z2+…+Z k (4);

[0026] Step 5: Since the initial reconstructed trajectory matrix Z is an m×k matrix, it is necessary to convert it into k initial symplectic geometric components Y with a length of n through diagonal averaging. i (i = 1, 2, …, k):

[0027] Any element of Z is defined as z ij , where 1 ≤ i ≤ k, 1 ≤ j ≤ m, k * = min(m, k), m * = max(m, k), n = m + (k - 1)τ;

[0028] If m < k, let Otherwise, let Therefore, the process of diagonal averaging is shown in the following formula (5):

[0029]

[0030] By processing the initial reconstruction trajectory matrix Z, k initial symplectic geometric components Y are obtained i (y1, y2, …, y n ), as shown in the following formula (6):

[0031] Y = Y1 + Y2 + … + Y k (6); [[ID=3!]]

[0032] Step 6: Calculate the cyclic kurtosis entropy of the initial symplectic geometric components Y i to obtain CKE1, CKE2, …, CKE k , and construct a weight function to calculate the weight value L corresponding to each initial symplectic geometric component Y i through the following formula (7); i ;

[0033]

[0034] where CKE mean represents the average value of CKE; i = 1, 2, …, k;

[0035] Step 7: The final denoised signal M can be obtained through the following formula (8):

[0036] [[ID=!0]]

[0037] where the calculation process of the cyclic kurtosis entropy of the initial symplectic geometric component Y i is as follows:

[0038] Given a signal b = {b1, b2, …, b n} with a length of n, then the defined cyclic kurtosis (CK) can be expressed as shown in the following formula (9):

[0039]

[0040] Wherein, RSE represents the autocorrelation function of the square envelope signal SE(b), and SE(b) = |b| 2 =|b+jHilber(b)| 2 , Hilbert represents Hilbert transform; Ra(0) is the autocorrelation function value of signal b when the delay coefficient is 0, and R b (τ) = E[b(ni)b(ni+τ)]; hT (h = 1, 2, ..., n) represents the delay coefficient;

[0041] The CK value of any delay coefficient is calculated by the following formula (10), and the cyclic kurtosis entropy can be expressed as follows by the following formula (11):

[0042]

[0043]

[0044] Where h = {1, 2,…, n}.

[0045] The wavelet scattering transform method is implemented by inputting the obtained noise reduction signal M into the wavelet scattering transform, performing a two-layer cascade operation through a preset first-order wavelet scattering transform and a preset second-order wavelet scattering transform to obtain a second-order wavelet scattering coefficient;

[0046] The first-order wavelet scattering transform is expressed as By performing first-order wavelet transform on the input noise reduction signal M, the first-order wavelet scattering coefficient is obtained, and the formula is used to calculate the first-order wavelet scattering coefficient. The modulus of the first-order wavelet scattering coefficient is calculated, and the next layer of the wavelet is further used to calculate U 1X Decompose, perform modular operation and calculate the local average; It is represented as a convolution operation; ψ(M) is represented as a wavelet function, which corresponds to a low-pass filter;

[0047] The second-order wavelet scattering transform is expressed as The second-order wavelet scattering coefficient is obtained by performing a second-order wavelet transform on the modulus local average of the first-order wavelet scattering coefficient obtained by the first-order wavelet scattering transform operation.

[0048] Among them, during the training process of the attention domain adaptive network model, a maximum mean error loss is added after the attention integration block to optimize the similarity of the extracted source domain sample and target domain sample features, so that the network can learn a predetermined number of common features of the source domain samples and the target domain samples, and classify the common features extracted by the network through the final fully connected layer and softmax, and use the cross entropy loss to evaluate the quality of the classification, and further update the model parameters through the back propagation algorithm to train the optimal model.

[0049] The mathematical form of the maximum mean error loss MMD is as shown in the following formula (12):

[0050]

[0051] Among them, D s and D t Represented as source domain sample x s and target domain samples x t distribution; k(.,.) is represented as the kernel function.

[0052] The mathematical form of the cross entropy loss is as shown in the following formula (13):

[0053]

[0054] in, Represents the predicted value of a sample; Y s Expressed as The true value of .

[0055] An embodiment of the present invention further provides an adaptive fault diagnosis system for a Roots-type compressor, comprising:

[0056] A signal acquisition unit is used to acquire the vibration signal to be measured transmitted from the Roots compressor;

[0057] a signal processing unit, configured to perform noise elimination and transformation processing on the vibration signal to be measured based on a preset adaptive symplectic geometric mode decomposition algorithm and a preset wavelet scattering transform method;

[0058] A signal diagnosis unit is used to import the vibration signal to be measured after noise reduction and transformation processing into a trained Roots compressor fault diagnosis model to obtain a diagnosis result of the vibration signal to be measured; wherein the diagnosis result is one of normal, bearing inner ring fault, bearing outer ring fault, bearing rolling element fault, gear fault and impeller fault; the Roots compressor fault diagnosis model is constructed based on an attention domain adaptive network and is trained by simultaneously inputting original domain samples with complete fault categories and target domain samples with incomplete fault categories; the original domain samples are formed by de-noising and transforming multiple vibration signals with labels and complete fault categories under Roots compressor working condition A through the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method; the target domain samples are formed by de-noising and transforming multiple vibration signals with labels but incomplete fault categories under Roots compressor working condition B through the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method.

[0059] The implementation of the embodiments of the present invention has the following beneficial effects:

[0060] 1. In the Roots compressor fault diagnosis model constructed based on the attention domain adaptive network, the present invention allows the simultaneous input of original domain samples with complete fault categories and target domain samples with incomplete fault categories under different working conditions, so that the incompletely labeled samples of the Roots compressor under different working conditions can be accurately classified and relabeled. This can effectively solve the problem of a large number of unlabeled samples collected in the existing Roots compressor fault diagnosis, thereby overcoming the problem of incorrect diagnosis results caused by the lack of training sample categories in the existing transfer learning-based methods, achieving accurate diagnosis of Roots compressor faults and improving diagnostic accuracy.

[0061] 2. The present invention applies the adaptive symplectic geometry decomposition algorithm and the wavelet scattering transform method to all signal samples (including source domain signals and target domain signals) to achieve noise removal preprocessing and extract stable features, thereby making the fault features more prominent and separable, which is beneficial to the subsequent network feature recognition and extraction;

[0062] 3. The Roots compressor fault diagnosis model of the present invention adopts an attention domain adaptive network, and in view of the strong global feature extraction capability of the transformer block, it is able to extract more comprehensive deep features and learn more common features between the target and source domains, so that it has higher classification accuracy in cross-working condition semi-supervised tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, without paying any creative work, other drawings obtained based on these drawings still fall within the scope of the present invention.

[0064] Figure 1 A flowchart of an adaptive fault diagnosis method for a Roots-type compressor provided by an embodiment of the present invention;

[0065] Figure 2 A block diagram of an attention domain adaptive network in an adaptive fault diagnosis method for a Roots-type compressor provided by an embodiment of the present invention;

[0066] Figure 3 A diagram showing the training results of an attention domain adaptive network in an application scenario of an adaptive fault diagnosis method for a Roots-type compressor provided by an embodiment of the present invention;

[0067] Figure 4 A schematic structural diagram of an adaptive fault diagnosis system for a Roots-type compressor provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be described in further detail below with reference to the accompanying drawings.

[0069] like Figure 1 As shown in the figure, an adaptive fault diagnosis method for a Roots-type compressor is provided in an embodiment of the present invention, and the method includes the following steps:

[0070] Step S1, obtaining a vibration signal to be measured transmitted from a Roots compressor;

[0071] Step S2: performing denoising and transformation processing on the vibration signal to be measured based on a preset adaptive symplectic geometric mode decomposition algorithm and a preset wavelet scattering transform method;

[0072] Step S3, importing the vibration signal to be measured after denoising and transformation processing into the trained Roots compressor fault diagnosis model to obtain a diagnosis result of the vibration signal to be measured; wherein, the diagnosis result is one of normal, bearing inner ring fault, bearing outer ring fault, bearing rolling element fault, gear fault and impeller fault; the Roots compressor fault diagnosis model is constructed based on an attention domain adaptive network, and is trained by simultaneously inputting original domain samples with complete fault categories and target domain samples with incomplete fault categories; the original domain samples are formed by denoising and transforming a plurality of vibration signals with labels and complete fault categories under the Roots compressor working condition A through the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method; the target domain samples are formed by denoising and transforming a plurality of vibration signals with labels but incomplete fault categories under the Roots compressor working condition B through the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method.

[0073] The specific process is as follows: before step S1, first, a Roots compressor fault diagnosis model is constructed based on the attention domain adaptive network.

[0074] Secondly, obtain the training samples. The process is as follows:

[0075] (1) The acceleration sensor is fixed on the Roots compressor and connected to the multi-channel data acquisition and analysis instrument, so that the multi-channel data acquisition and analysis instrument can obtain the vibration signal transmitted by the Roots compressor in real time.

[0076] In this test environment, vibration signals under different operating conditions were separated into source and target domain signals. The source domain signal was a labeled vibration signal with complete fault categories from Roots compressor operating condition A; the target domain signal was a labeled vibration signal from Roots compressor operating condition B, but with incomplete fault categories. Fault categories included normal, bearing outer race fault, bearing inner race fault, bearing rolling element fault, gear fault, and impeller fault.

[0077] In one example, the source domain signals include the normal vibration signal under a 60 Hz shaft rotation frequency (Condition A), the vibration signal when the bearing outer race is faulty, the vibration signal when the bearing inner race is faulty, the vibration signal when the bearing rolling element is faulty, the vibration signal when the gear is faulty, and the vibration signal when the impeller is faulty. The target domain signals only include the normal vibration signal and the vibration signal when the gear is faulty under a 100 Hz shaft rotation frequency (Condition B).

[0078] (2) Using the preset adaptive symplectic geometric mode decomposition algorithm and the preset wavelet scattering transform method, the source domain signal and the target domain signal are denoised and transformed to obtain the original domain samples and the target domain samples respectively. It can be seen that the original domain samples are formed by the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method to remove the noise and transform the multiple vibration signals with labels and complete fault categories under the working condition of the Roots compressor A; the target domain samples are formed by the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method to remove the noise and transform the multiple vibration signals with labels but incomplete fault categories under the working condition of the Roots compressor B.

[0079] (21) The execution steps of the adaptive symplectic geometric mode decomposition algorithm are as follows:

[0080] The first step is to assume that the vibration signal to be de-noised is a={a1,a2,…,a n}; where n represents the length of signal a;

[0081] Then, define the trajectory matrix A of signal a as shown in the following formula (1):

[0082]

[0083] Where k represents the embedding dimension; the embedding dimension k is obtained by calculating the power spectral density of signal a and estimating the frequency of the maximum peak; where k = n / 3 if the normalized frequency is less than a given threshold, otherwise k = 1.2×(f s / f max );f s Expressed as sampling frequency, f max It is represented by the frequency of the maximum peak of the power spectrum density of signal a; τ represents the delay time; m = n-(k-1)τ;

[0084] The second step is to construct the Hamiltonian matrix B according to formula (2):

[0085]

[0086] Where C = A T A;

[0087] Step 3: Let D = B 2 , according to the definition of Hamiltonian matrix, D and B are both Hamiltonian matrices, and then a symplectic orthogonal matrix E is constructed, as shown in the following formula (3):

[0088]

[0089] Wherein, the matrix E is represented as an orthogonal symplectic matrix with symplectic matrix properties; F is represented as an upper triangular matrix, and the eigenvalue of the upper triangular matrix F is represented as λ i(i = 1, 2, …, k); f ij = 0 (i > j + 1);

[0090] Step 4. If matrix C is a real symmetric matrix, then the eigenvalues of C are equal to the eigenvalues of the upper triangular matrix F; According to the properties of the Hamiltonian matrix, the eigenvalues of C are The eigenvalues of matrix C can be expressed as β1 > β2 > … > β k ;

[0091] Let P i (i = 1, 2, …, k) be the eigenvectors of matrix C, let P i (i = 1, 2, …, k) be the eigenvectors of matrix C, and let S i = P i T A T And Z i = P i S i , then the initial reconstructed trajectory matrix Z is defined as follows:

[0092] Z = Z1 + Z2 + … + Z k (4);

[0093] Step 5. Since the initial reconstructed trajectory matrix Z is an m×k matrix, it is necessary to transform it into k initial symplectic geometric components Y of length n through diagonal averaging i (i = 1, 2, …, k):

[0094] Any element of Z is defined as z ij , where, 1 ≤ i ≤ k, 1 ≤ j ≤ m, k * = min(m, k), m * = max(m, k), n = m + (k - 1)τ;

[0095] If m < k, let Otherwise, let Therefore, the process of diagonal averaging is shown in the following formula (5):

[0096]

[0097] By processing the initial reconstructed trajectory matrix Z, k initial symplectic geometric components Y i (y1, y2, …, y n ) are obtained, as shown in the following formula (6):

[0098] Y = Y1 + Y2 + … + Y k (6);

[0099] Step 6. Calculate the initial symplectic geometric components Y iThe cyclic kurtosis entropy is obtained by CKE1, CKE2,…, CKE k , and construct the weight function to calculate each initial symplectic geometric component Y by the following formula (7) i The corresponding weight value L i ;

[0100]

[0101] Among them, CKE mean Expressed as the average value of CKE; i = 1, 2, ..., k;

[0102] It should be noted that any initial symplectic geometric component Y i The calculation process of the cyclic kurtosis entropy is as follows:

[0103] Given a signal b={b1,b2,…,b n}, then the cycle kurtosis (CK) can be expressed as shown in the following formula (9):

[0104]

[0105] Wherein, RSE represents the autocorrelation function of the square envelope signal SE(b), and SE(b) = |b| 2 =|b+jHilber(b)| 2 , Hilbert represents Hilbert transform; Ra(0) is the autocorrelation function value of signal b when the delay coefficient is 0, and R b (τ) = E[b(ni)b(ni+τ)]; hT (h = 1, 2, ..., n) represents the delay coefficient;

[0106] The CK value of any delay coefficient is calculated by the following formula (10), and the cyclic kurtosis entropy can be expressed as follows by the following formula (11):

[0107]

[0108]

[0109] Where h = {1, 2,…, n}.

[0110] Step 7: The final noise reduction signal M can be obtained by the following formula (8):

[0111]

[0112] (22) The execution steps of the wavelet scattering transform method are:

[0113] The noise reduction signal M obtained by formula (8) is input into the wavelet scattering transform, and a two-layer cascade operation is performed through a preset first-order wavelet scattering transform and a preset second-order wavelet scattering transform to obtain the second-order wavelet scattering coefficient.

[0114] Among them, the first-order wavelet scattering transform is expressed as By performing first-order wavelet transform on the input noise reduction signal M, the first-order wavelet scattering coefficient is obtained, and the formula is used to calculate the first-order wavelet scattering coefficient. The modulus of the first-order wavelet scattering coefficient is calculated, and the next layer of the wavelet is further used to calculate U 1X Decompose, perform modular operation and calculate the local average; It is represented as a convolution operation; ψ(M) is represented as a wavelet function, which corresponds to a low-pass filter;

[0115] Among them, the second-order wavelet scattering transform is expressed as The second-order wavelet scattering coefficient is obtained by performing a second-order wavelet transform on the modulus local average of the first-order wavelet scattering coefficient obtained by the first-order wavelet scattering transform operation.

[0116] Finally, the source domain samples and target domain samples obtained above are used as the input of the attention domain adaptive network to train the Roots compressor fault diagnosis model, and the test samples are used to test it to obtain the trained Roots compressor fault diagnosis model, as shown in Figure 2. Figure 2 As shown. Figure 2 In the training process of the attention domain adaptation network, a maximum mean error loss is added after the attention integration block to optimize the similarity of the features of the extracted source domain samples and the target domain samples, so that the network can learn a predetermined number of common features of the source domain samples and the target domain samples, and classify the common features extracted by the network through the final fully connected layer and softmax, and use the cross entropy loss to evaluate the quality of the classification, and further use the back propagation algorithm to update the model parameters to train the optimal model.

[0117] At this time, the mathematical form of the maximum mean error loss MMD is expressed as shown in the following formula (12):

[0118]

[0119] Among them, D s and D t Represented as source domain sample x s and target domain samples x t distribution; k(.,.) is represented as the kernel function.

[0120] At this time, the mathematical form of cross entropy loss is expressed as shown in the following formula (13):

[0121]

[0122] in, Represents the predicted value of a sample; Y s Expressed as The true value of .

[0123] In step S1, an acceleration sensor is fixed on a Roots-type compressor and connected to a multi-channel data acquisition and analysis instrument, so that the multi-channel data acquisition and analysis instrument can acquire the vibration signal to be measured transmitted from the Roots-type compressor in real time.

[0124] In step S2, based on the above-mentioned adaptive symplectic geometric mode decomposition algorithm and wavelet scattering transform method, the vibration signal to be measured is denoised and transformed. For the specific process, please refer to the processing process of the original domain signal or the target domain signal, which will not be repeated here.

[0125] In step S3, the vibration signal to be measured after the noise elimination and transformation processing in step S2 is introduced into the trained Roots compressor fault diagnosis model, and the diagnosis result of the vibration signal to be measured is obtained as one of normal, bearing inner ring fault, bearing outer ring fault, bearing rolling element fault, gear fault and impeller fault.

[0126] like Figure 3 As shown, the application scenario of the adaptive fault diagnosis method for a Roots-type compressor provided in an embodiment of the present invention is further described. The specific process is as follows:

[0127] Experimental data was extracted, including the vibration signals under a 60Hz shaft rotation frequency (Condition A), normal vibration signals, bearing outer race faults, bearing inner race faults, bearing rolling element faults, gear faults, and impeller faults; and normal vibration signals under a 100Hz shaft rotation frequency (Condition B). The vibration signals under Condition B, including those under the outer race faults, inner race faults, bearing rolling element faults, and gear faults, were not included in model training and were used only for testing the Roots compressor fault diagnosis model.

[0128] The number of training samples for each category is set to 1000 and the number of test samples is set to 200. The training samples are first pre-processed by the adaptive symplectic geometric mode decomposition algorithm ASGMD, and then the denoised signals are converted into image samples by the wavelet scattering transform method WST to extract stable features. The image samples after denoising and transformation of the original domain samples and the target domain samples are used as the input of the attention domain adaptive network to train the Roots compressor fault diagnosis model. The number of iterations is 200 and the learning rate of Adam is 0.05. The training result diagram (as shown in the figure) Figure 3As shown in the figure, the classification accuracy of the final test data is 99.9%, so the method introduced in this case has a very high accuracy in the fault data of Roots compressors across working conditions.

[0129] like Figure 4 FIG. 1 is an example of an adaptive fault diagnosis system for a Roots-type compressor provided in an embodiment of the present invention, comprising:

[0130] The signal acquisition unit 110 is used to acquire the vibration signal to be measured transmitted from the Roots compressor;

[0131] The signal processing unit 120 is configured to perform noise elimination and transformation processing on the vibration signal to be measured based on a preset adaptive symplectic geometric mode decomposition algorithm and a preset wavelet scattering transform method;

[0132] The signal diagnosis unit 130 is used to import the vibration signal to be measured after noise reduction and transformation processing into the trained Roots compressor fault diagnosis model to obtain a diagnosis result of the vibration signal to be measured; wherein the diagnosis result is one of normal, bearing inner race fault, bearing outer race fault, bearing rolling element fault, gear fault and impeller fault; the Roots compressor fault diagnosis model is constructed based on an attention domain adaptive network and is trained by simultaneously inputting original domain samples with complete fault categories and target domain samples with incomplete fault categories; the original domain samples are formed by de-noising and transforming multiple vibration signals with labels and complete fault categories under Roots compressor working condition A through the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method; the target domain samples are formed by de-noising and transforming multiple vibration signals with labels but incomplete fault categories under Roots compressor working condition B through the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method.

[0133] The implementation of the embodiments of the present invention has the following beneficial effects:

[0134] 1. In the Roots compressor fault diagnosis model constructed based on the attention domain adaptive network, the present invention allows the simultaneous input of original domain samples with complete fault categories and target domain samples with incomplete fault categories under different working conditions, so that the incompletely labeled samples of the Roots compressor under different working conditions can be accurately classified and relabeled. This can effectively solve the problem of a large number of unlabeled samples collected in the existing Roots compressor fault diagnosis, thereby overcoming the problem of incorrect diagnosis results caused by the lack of training sample categories in the existing transfer learning-based methods, achieving accurate diagnosis of Roots compressor faults and improving diagnostic accuracy.

[0135] 2. The present invention applies the adaptive symplectic geometry decomposition algorithm and the wavelet scattering transform method to all signal samples (including source domain signals and target domain signals) to achieve noise removal preprocessing and extract stable features, thereby making the fault features more prominent and separable, which is beneficial to the subsequent network feature recognition and extraction;

[0136] 3. The Roots compressor fault diagnosis model of the present invention adopts an attention domain adaptive network, and in view of the strong global feature extraction capability of the transformer block, it is able to extract more comprehensive deep features and learn more common features between the target and source domains, so that it has higher classification accuracy in cross-working condition semi-supervised tasks.

[0137] It is worth noting that in the above system embodiment, the various units included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

[0138] Those skilled in the art will understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc.

[0139] The above disclosure is merely a preferred embodiment of the present invention and certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. An adaptive fault diagnosis method for a Roots compressor, characterized in that: The method comprises the following steps: Obtain the vibration signal to be measured transmitted from the Roots compressor; Based on a preset adaptive symplectic geometric mode decomposition algorithm and a preset wavelet scattering transform method, the vibration signal to be measured is subjected to denoising and transformation processing; The vibration signal to be measured after noise elimination and transformation processing is introduced into the trained Roots compressor fault diagnosis model to obtain the diagnosis result of the vibration signal to be measured; wherein, the diagnosis result is one of normal, bearing inner ring fault, bearing outer ring fault, bearing rolling element fault, gear fault and impeller fault; the Roots compressor fault diagnosis model is constructed based on the attention domain adaptive network, and is trained by simultaneously inputting original domain samples with complete fault categories and target domain samples with incomplete fault categories; the original domain samples are Roots compressor A The target domain samples are formed by denoising and transforming multiple vibration signals with labels and complete fault categories under working conditions through the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method; the target domain samples are formed by Roots compressor B The multiple vibration signals with labels but incomplete fault categories under working conditions are formed after being denoised and transformed by the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method; The steps of executing the adaptive symplectic geometric mode decomposition algorithm are as follows: The first step is to assume that the vibration signal to be denoised is a ={ a 1, a 2,···, a n };in, n Represented as a signal a length; Then, define the signal a The trajectory matrix A , as shown in the following formula (1): （1）； in, k Represented as embedding dimension; embedding dimension k By calculating the signal a The power spectral density of the power spectrum is obtained by estimating the frequency of the maximum peak; if the normalized frequency is less than a given threshold, then k=n / 3 ,otherwise, k=1.2×(f s / f max ) ; f s Expressed as sampling frequency, f max Represented as a signal a The frequency of the maximum peak of the power spectral density; τ Expressed as delay time; m = n -( k -1) τ ; The second step is to construct the Hamiltonian matrix according to formula (2). B : （2）； in, C=A T A ; Step 3: Order D=B 2 , according to the definition of Hamiltonian matrix, D and B Are all Hamiltonian matrices, and then construct a symplectic orthogonal matrix E , as shown in the following formula (3): （3）； Among them, the matrix E It is represented as an orthogonal symplectic matrix with symplectic matrix properties; F is represented as an upper triangular matrix, and the upper triangular matrix F The eigenvalues ​​of λ i ( i =1,2,···, k ); f ij =0( i > j +1); Step 4: If the matrix C is a real symmetric matrix, then C The eigenvalues ​​of the upper triangular matrix F The eigenvalue of ; According to the properties of the Hamiltonian matrix, C The characteristic value of β i = ( i =1,2,···, k ),matrix C The eigenvalues ​​of β 1> β 2>···> β k ; make P i ( i =1,2,···, k ) is a matrix C The characteristic vector of P i ( i =1,2,···, k ) is a matrix C The eigenvectors of S i = P i T A T and Z i = P i S i , then the initial reconstruction trajectory matrix Z The definition is as follows: （4）； Step 5: Due to the initial reconstruction trajectory matrix Z for m×k Matrix, so it needs to be converted into a matrix of length by diagonal averaging. n of k Initial symplectic geometry components Y i ( i =1,2,···, k ): Any Z The elements are defined as z ij , where 1≤ i ≤ k , 1≤ j ≤ m , k * =min( m , k ), m * =max( m , k ), n = m +( k- 1) τ ; if m < k ,make z ij * = z ij Otherwise, let z ij * = z ji ; Therefore, the diagonal averaging process is shown in the following formula (5): （5）； By processing the initial reconstruction trajectory matrix Z ,get k Initial symplectic geometry components Y i ( y 1, y 2 , ···, y n ), as shown in the following formula (6): （6）； Step 6. Calculate the initial symplectic geometry components Y i The cyclic kurtosis entropy of CKE 1 ,CKE 2 ,···,CKE k , and construct the weight function to calculate each initial symplectic geometric component through the following formula (7): Y i Corresponding weight value L i ; （7）； in, CKE mean Expressed as CKE Average value ; i=1,2,···,k ; Step 7. Final noise reduction signal M The following formula (8) is used to obtain: （8）； The execution steps of the wavelet scattering transform method are as follows: M Input into the wavelet scattering transform, perform two-layer cascade operation through a preset first-order wavelet scattering transform and a preset second-order wavelet scattering transform to obtain the second-order wavelet scattering coefficient; Among them, the first-order wavelet scattering transform is expressed as , by inputting the noise reduction signal M Perform first-order wavelet transform to obtain the first-order wavelet scattering coefficient, and use the formula , calculate the modulus of the first-order wavelet scattering coefficient, and further use the next layer of wavelet to U 1X Decompose, perform modular operation and calculate the local average; Represented as a convolution operation; Expressed as a wavelet function, it corresponds to a low-pass filter; The second-order wavelet scattering transform is expressed as , performing a second-order wavelet transform on the modulus local average of the first-order wavelet scattering coefficient obtained by the first-order wavelet scattering transform operation to obtain the second-order wavelet scattering coefficient.

2. The adaptive fault diagnosis method for a Roots-type compressor according to claim 1, wherein: The initial symplectic geometry components Y i The calculation process of the cyclic kurtosis entropy is as follows: Given a length of n signal b ={ b 1 , b 2 ,···, b n }, then the cycle kurtosis (CK) is defined as shown in the following formula (9): （9）； in, R SE Represented as a square envelope signal SE(b) The autocorrelation function of , Hilbert Expressed as Hilbert transform; Ra(0) The signal when the delay coefficient is 0 b The autocorrelation function value of ; hT ( h =1,2,···, n ) is expressed as the delay coefficient; The CK value of any delay coefficient is calculated by the following formula (10), and the cyclic kurtosis entropy is expressed as follows by the following formula (11): （10）； （11）； in, h={1,2,···,n} .

3. The adaptive fault diagnosis method for a Roots-type compressor according to claim 1, wherein: During the training process, the attention domain adaptive network model optimizes the similarity of the extracted source domain samples and target domain sample features by adding a maximum mean error loss after the attention integrated block, so that the network learns a predetermined number of common features of the source domain samples and the target domain samples, and then passes the final fully connected layer to the target domain sample. softmax The common features extracted by the network are classified, and the quality of the classification is evaluated using cross entropy loss. The model parameters are further updated through the back propagation algorithm to train the optimal model.

4. The adaptive fault diagnosis method for a Roots-type compressor according to claim 3, wherein: The maximum mean error loss MMD The mathematical form of is as shown in the following formula (12): （12）； in, D s and D t Represented as source domain samples x s and target domain samples x t distribution of k (.,.) is represented as the kernel function.

5. The adaptive fault diagnosis method for a Roots-type compressor according to claim 3, wherein: The mathematical form of the cross entropy loss is as shown in the following formula (13): （13）； in, Expressed as the predicted value of a sample; Y s Expressed as The true value of .

6. An adaptive fault diagnosis system for a Roots compressor, characterized in that: include: A signal acquisition unit is used to acquire the vibration signal to be measured transmitted from the Roots compressor; a signal processing unit, configured to perform noise elimination and transformation processing on the vibration signal to be measured based on a preset adaptive symplectic geometric mode decomposition algorithm and a preset wavelet scattering transform method; The signal diagnosis unit is used to import the vibration signal to be measured after noise elimination and transformation processing into the trained Roots compressor fault diagnosis model to obtain the diagnosis result of the vibration signal to be measured; wherein, the diagnosis result is one of normal, bearing inner ring fault, bearing outer ring fault, bearing rolling element fault, gear fault and impeller fault; the Roots compressor fault diagnosis model is constructed based on the attention domain adaptive network, and is trained by simultaneously inputting original domain samples with complete fault categories and target domain samples with incomplete fault categories; the original domain samples are Roots compressor A The target domain samples are formed by denoising and transforming multiple vibration signals with labels and complete fault categories under working conditions through the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method; the target domain samples are formed by Roots compressor B Multiple vibration signals with labels but incomplete fault categories under working conditions are formed after being denoised and transformed by the adaptive symplectic geometric mode decomposition algorithm and the wavelet scattering transform method; the execution steps of the adaptive symplectic geometric mode decomposition algorithm are as follows: The first step is to assume that the vibration signal to be denoised is a ={ a 1, a 2,···, a n };in, n Represented as a signal a length; Then, define the signal a The trajectory matrix A , as shown in the following formula (1): （1）； in, k Represented as embedding dimension; embedding dimension k By calculating the signal a The power spectral density of the power spectrum is obtained by estimating the frequency of the maximum peak; if the normalized frequency is less than a given threshold, then k=n / 3 ,otherwise, k=1.2×(f s / f max ) ; f s Expressed as sampling frequency, f max Represented as a signal a The frequency of the maximum peak of the power spectral density; τ Expressed as delay time; m = n -( k -1) τ ; The second step is to construct the Hamiltonian matrix according to formula (2). B : （2）； in, C=A T A ; Step 3: Order D=B 2 , according to the definition of Hamiltonian matrix, D and B Are all Hamiltonian matrices, and then construct a symplectic orthogonal matrix E , as shown in the following formula (3): （3）； Among them, the matrix E It is represented as an orthogonal symplectic matrix with symplectic matrix properties; F is represented as an upper triangular matrix, and the upper triangular matrix F The eigenvalues ​​of λ i ( i =1,2,···, k ); f ij =0( i > j +1); Step 4: If the matrix C is a real symmetric matrix, then C The eigenvalues ​​of the upper triangular matrix F The eigenvalue of ; According to the properties of the Hamiltonian matrix, C The characteristic value of β i = ( i =1,2,···, k ),matrix C The eigenvalues ​​of β 1> β 2>···> β k ; make P i ( i =1,2,···, k ) is a matrix C The characteristic vector of P i ( i =1,2,···, k ) is a matrix C The eigenvector of , and let S i = P i T A T and Z i = P i S i , then the initial reconstruction trajectory matrix Z The definition is as follows: （4）； Step 5: Due to the initial reconstruction trajectory matrix Z for m×k Matrix, so it needs to be converted into a matrix of length by diagonal averaging. n of k Initial symplectic geometry components Y i ( i =1,2,···, k ): Any Z The elements are defined as z ij , where 1≤ i ≤ k , 1≤ j ≤ m , k * =min( m , k ), m * =max( m , k ), n = m +( k- 1) τ ; if m < k ,make z ij * = z ij Otherwise, let z ij * = z ji ; Therefore, the diagonal averaging process is shown in the following formula (5): （5）； By processing the initial reconstruction trajectory matrix Z ,get k Initial symplectic geometry components Y i ( y 1, y 2 , ···, y n ), as shown in the following formula (6): （6）； Step 6. Calculate the initial symplectic geometry components Y i The cyclic kurtosis entropy of CKE 1 ,CKE 2 ,···,CKE k , and construct the weight function to calculate each initial symplectic geometric component through the following formula (7): Y i Corresponding weight value L i ; （7）； in, CKE mean Expressed as CKE Average value ; i=1,2,···,k ; Step 7. Final noise reduction signal M The following formula (8) is used to obtain: （8）； The execution steps of the wavelet scattering transform method are as follows: M Input into the wavelet scattering transform, perform two-layer cascade operation through a preset first-order wavelet scattering transform and a preset second-order wavelet scattering transform to obtain the second-order wavelet scattering coefficient; Among them, the first-order wavelet scattering transform is expressed as , by inputting the noise reduction signal M Perform first-order wavelet transform to obtain the first-order wavelet scattering coefficient, and use the formula , calculate the modulus of the first-order wavelet scattering coefficient, and further use the next layer of wavelet to U 1X Decompose, perform modular operation and calculate the local average; Represented as a convolution operation; Expressed as a wavelet function, it corresponds to a low-pass filter; The second-order wavelet scattering transform is expressed as , performing a second-order wavelet transform on the modulus local average of the first-order wavelet scattering coefficient obtained by the first-order wavelet scattering transform operation to obtain the second-order wavelet scattering coefficient.

Images