A health assessment method for diesel engine main bearings

Abstract

The present invention discloses a health assessment method for a diesel engine main bearing, which collects vibration signals of the main bearing under different faults, obtains a visible phase trajectory diagram through phase space reconstruction and main vector projection; builds a deep convolutional neural network model; labels various fault phase trajectory diagrams, creates a training set and inputs it into the model for learning; stops iteration when the cross entropy loss function is minimized, and saves the diagnostic model; on the other hand, collects vibration signals of various fault main bearings under different health states, calculates and counts their maximum Lyapunov exponent mean values, and uses them as health assessment standards. During the assessment, a health assessment is performed according to the principle of proximity based on the corresponding health assessment standards for the fault type. The present invention can realize online comprehensive monitoring of the service status of the main bearing, which is helpful to promote the main bearing maintenance method from regular replacement to condition-based maintenance, and can significantly improve economic benefits.

Description

Technical Field

[0001] The invention relates to a fault diagnosis method for a diesel engine, and in particular to a health assessment method for a main bearing of a diesel engine. Background Art

[0002] The main bearing of a diesel engine is subjected to alternating loads during operation, which causes uneven oil film distribution. A thin oil film is prone to induce bearing failure. Abnormal main bearing failure will seriously reduce diesel engine efficiency, increase fuel consumption, and threaten the reliable service of the diesel engine. Therefore, it is particularly important to perform fault diagnosis and health assessment on the main bearing. This will help promote condition-based maintenance of the main bearing, strengthen maintenance decision-making assistance, and significantly improve economic benefits.

[0003] The main bearing of a diesel engine belongs to a rotating machine. When working, it will generate state signals with highly nonlinear and non-stationary characteristics, such as vibration, friction, friction coefficient and temperature signals. In the patent with patent number CN201510080839.5, the vibration time domain diagram of the bearing fault is directly used as the training object of the deep learning model. However, it is difficult to effectively characterize the fault characteristics by relying on time domain diagrams or conventional time-frequency analysis methods. In recent years, nonlinear methods have been widely used in the diagnosis of rotating machinery, such as chaos analysis, fractal analysis and recursive analysis. In the patent with patent number CN202010387072.1, rolling bearing faults were identified based on convolutional neural networks and fuzzy C-means clustering algorithms, but the algorithm process is too cumbersome. In addition, the existing patents make further health assessments based on the diagnosis of bearing faults, and comprehensive monitoring of the service status of the main bearings is more conducive to promoting condition-based maintenance. The existing fault diagnosis of diesel engine main bearings is restricted by variable operating conditions and strong noise and high vibration working environment. Under the premise of balancing calculation efficiency and result accuracy, only the main fault types of the main bearings can be obtained in a relatively rough manner. The problem of determining detailed states such as the severity of the fault and the specific state of the main bearing under different fault types cannot be solved. Summary of the invention

[0004] Purpose of the invention: In view of the above problems, the present invention proposes a health assessment method for a diesel engine main bearing, which can quickly and accurately assess the specific health status of a diesel engine main bearing, such as the degree of damage of different fault locations of a diesel engine main bearing.

[0005] Technical solution: The technical solution adopted by the present invention is a health assessment method for a diesel engine main bearing, comprising the following steps:

[0006] (1) collecting vibration signals of a main bearing of a diesel engine, and converting the vibration signals into a phase trajectory diagram using a phase space reconstruction method; the phase space reconstruction method comprises the following steps:

[0007] (11) The one-dimensional fault vibration signal is reconstructed in phase space. The algorithm is as follows:

[0008] X i =(x i , x i+τ , x i+2τ , ..., x i+(m-1)τ )

[0009] Where x is the fault vibration signal, X i is the phase point matrix, i = 1, 2, ..., N, N is the number of phase points, m is the embedding dimension, τ is the delay time;

[0010] (12) The reconstructed high-dimensional space phase trajectory is projected by the main vector to obtain the visual phase trajectory diagram. The specific algorithm is as follows: Let the matrix A = X T X, after performing singular value decomposition on A, we get m eigenvalues ​​and eigenvectors, select the three largest eigenvalues ​​λ1, λ2, λ3, and use the corresponding eigenvectors V1, V2, V3 as the principal vectors of matrix A. Project the reconstructed phase space to the direction of the principal vector to get the N×3-order principal vector matrix B:

[0011] B=X[V1,V2,V3]

[0012] Plot the matrix B in a three-dimensional coordinate system to visualize the phase trajectory.

[0013] (2) calculating the maximum Lyapunov exponent of the vibration signal according to the phase trajectory diagram; inputting the phase trajectory diagram into the diagnosis model to complete the diagnosis of the fault type of the collected vibration signal, and inputting the result into the health assessment model; the establishment of the health assessment model includes the following steps:

[0014] (21) Collect vibration signals of various fault types of diesel engine main bearings in different health states; the bearing fault types include fault states and normal states; the fault states include inner ring fault, outer ring fault, and ball fault.

[0015] (22) The phase space reconstruction method is used to convert the vibration signals under different health states into phase trajectory diagrams;

[0016] (23) calculating the maximum Lyapunov exponent according to the phase trajectory diagram obtained in step (22);

[0017] (24) For each fault type of the bearing, the maximum Lyapunov index mean of each fault type under different health conditions is calculated and used as the health assessment standard;

[0018] The maximum Lyapunov exponent solution comprises the following steps:

[0019] 1) For each phase point X in the reconstructed phase space a and its neighboring point Yb , after time t, the distance between the two phase points is:

[0020] d a (t) = C a e γ(t×Δt)

[0021] Where t = 1, 2, ..., min (Na, Nb), γ is X a and Y b The divergence rate, C a =d a (0); Δt represents the sampling interval of the time series;

[0022] 2) Taking the logarithm of both sides:

[0023] ln d a (t) = ln C a +γ(t×Δt)

[0024] 3) Linearize the above equation into:

[0025]

[0026] Where q represents non-zero d a (t) quantity;

[0027] 4) Perform a linear fit on y(t) and the slope is the maximum Lyapunov exponent.

[0028] The establishment of the diagnostic model comprises the following steps:

[0029] (31) Collect fault signals of various faults of diesel engine main bearings;

[0030] (32) The phase space reconstruction method is used to convert the vibration signals under different faults into phase trajectory diagrams;

[0031] (33) The phase trajectory diagrams under different main bearing faults are labeled and divided into training set and test set; the training set and test set are divided according to the ratio of 8:2.

[0032] (34) Establishing an image learning model of a deep convolutional neural network, and setting parameters for the input layer, convolutional layer, pooling layer, and fully connected layer; the image learning model of the deep convolutional neural network is established, and the Softmax algorithm is used for prediction and recognition, and the diagnosis result is output. The model structure includes:

[0033] 1) Input layer, the phase trajectory image size is uniformly preprocessed to 227×227;

[0034] 2) C1 convolutional layer, the number of kernels is 24, the size is 11×11, and the stride is 4;

[0035] 3) P1 pooling layer, the pooling window size is 3×3, the step size is 2, and the maximum pooling method is used;

[0036] 4) C2 convolution layer, the number of kernels is 24, the size is 5×5, and the stride is 1;

[0037] 5) P2 pooling layer, the pooling window size is 3×3, the step size is 2, and the maximum pooling method is used;

[0038] 6) C3 convolutional layer, the number of kernels is 48, the size is 5×5, and the stride is 1;

[0039] 7) P3 pooling layer, the pooling window size is 3×3, the step size is 2, and the maximum pooling method is used;

[0040] 8) F1 fully connected layer;

[0041] 9) F2 fully connected layer, in which a random dropout layer is provided.

[0042] (35) The training set is input into the deep convolutional neural network model for learning. The training process includes forward propagation and reverse weight update. After finding the minimum cross entropy loss function, the iteration is stopped to obtain the diagnosis model. In the training process of step (5), the misdiagnosed samples are input into the model for further training. The trained model parameters are input into the previous diagnosis model through transfer learning to continuously correct the model.

[0043] (3) The health assessment model performs health assessment on the fault type diagnosed in step (2) according to the health assessment standard under the fault type and the maximum Lyapunov exponent of the vibration signal obtained in step (2). The health assessment includes fine assessment and rough assessment. The fine assessment refers to calculating the Lyapunov exponent value under normal conditions, and taking the Lyapunov exponent value under fault conditions and its degree of deviation as the health degree; the rough assessment refers to calculating the Lyapunov exponent under four stages of normal conditions and fault degradation, and comparing the maximum Lyapunov exponent of the signal to be evaluated with the health assessment standard. The health state corresponding to the health assessment standard that is closest in value is the assessment result.

[0044] Beneficial effects: Compared with the prior art, the present invention has the following advantages: (1) The specific health status of the main bearing of the diesel engine is evaluated by combining the phase space reconstruction technology, the deep convolutional neural network model and the maximum Lyapunov index evaluation, which can improve the efficiency and accuracy of the health status of the main bearing of the diesel engine; (2) The phase trajectory diagram is easily interfered by noise when characterizing the fault characteristics, resulting in serious changes in the evolution trajectory of the phase trajectory. In order to avoid the use of cumbersome denoising algorithms to affect the rapidity of diagnosis, it is proposed to combine the convolutional neural network with it to extract the invariant features of the phase trajectory diagram. (3) In the recognition of various fault phase trajectory diagrams, a convolutional neural network model with good image classification effect is introduced, and the misdiagnosed samples are input into the model for further training, and the trained model parameters are input into the previous diagnosis model through transfer learning to improve the diagnosis rate. The above algorithm is simple and easy to implement, with a small amount of calculation, and can meet the real-time calculation requirements of online bearing fault monitoring. The present invention can realize comprehensive monitoring of the service status of the main bearing, which is helpful to promote the main bearing maintenance method from regular replacement to condition-based maintenance, and can significantly improve economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the method for health assessment of a main bearing of a diesel engine according to the present invention;

[0046] Figure 2 It is the phase trajectory diagram of the bearing under different faults;

[0047] Figure 3 is the training accuracy and loss of the deep convolutional neural network model described in the present invention;

[0048] Figure 4 is the health assessment standard for inner race faults according to the present invention;

[0049] Figure 5 It is the confusion matrix of the fault diagnosis result described in the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0051] Based on the Case Western Reserve University bearing database, the present invention takes the drive end vibration signal as an example to describe the specific implementation of fault diagnosis and health assessment in detail. In this example, different fault locations of the bearing represent different fault types of the bearing, and the damage size represents the health status.

[0052] The bearing fault types include inner ring fault, outer ring fault, ball fault and normal state. Taking the inner ring fault as an example, the damage sizes are 0, 0.007 and 0.021 mm respectively. During the collection of the above vibration signals, the motor speed is 1797rpm and the sampling frequency is 12kHz.

[0053] The health assessment method of a diesel engine main bearing of the present invention comprises the following steps:

[0054] (1) Collect the vibration signal of the diesel engine main bearing and convert it into a phase trajectory diagram using the phase space reconstruction method;

[0055] (2) calculating the maximum Lyapunov exponent of the vibration signal according to the phase trajectory diagram; inputting the phase trajectory diagram into the diagnosis model to complete the diagnosis of the fault type of the collected vibration signal, and inputting the result into the health assessment model;

[0056] (3) The health assessment model performs a health assessment on the fault type diagnosed in step (2) according to the health assessment standard under the fault type and the maximum Lyapunov exponent of the vibration signal obtained in step (2).

[0057] like Figure 1 As shown, the detailed implementation steps of the diagnosis model establishment, health assessment model establishment and health assessment process described in the present invention are as follows:

[0058] (1) Collect two types of bearing fault signals. One is the vibration signal of the bearing under different faults, including inner ring fault, outer ring fault, ball fault and normal state. The other is the vibration signal of various faulty bearings under different health states. Here, the vibration signal of the inner ring fault with damage size of 0, 0.007 and 0.021 mm is taken as an example.

[0059] (2) In order to form a certain number of training sets, the vibration signal of the bearing under each fault is divided into windows. The window length is 1000 points and the number of windows is 120. The vibration signals under different faults are first converted into phase trajectory diagrams using phase space reconstruction technology. The phase trajectory diagrams under different faults are shown in Figure 2. Figure 2 The specific process includes:

[0060] 1) Reconstruct the one-dimensional fault vibration signal in phase space. The algorithm is as follows:

[0061] X i =(x i , x i+τ , x i+2τ , ..., x i+(m-1)τ )

[0062] Where x is the fault vibration signal, X i is the phase point matrix. i = 1, 2, ..., N, where N is the number of phase points. m is the embedding dimension, and τ is the delay time.

[0063] 2) The embedding dimension m of the rotating machinery vibration signal is generally greater than 3, and the reconstructed high-dimensional space phase trajectory must be projected by the main vector to obtain the visible phase trajectory diagram. The specific algorithm is as follows: Let the matrix A = X T X, after performing singular value decomposition on A, we get m eigenvalues ​​and eigenvectors. Select the three largest eigenvalues ​​λ1, λ2, λ3, and use the corresponding eigenvectors V1, V2, V3 as the principal vectors of matrix A. Project the reconstructed phase space to the principal vector direction to get the N×3-order principal vector matrix B:

[0064] B=X[V1,V2,V3]

[0065] Plot the matrix B in a three-dimensional coordinate system to visualize the phase trajectory.

[0066] (3) The phase trajectory diagrams of the inner race fault, outer race fault, ball fault and normal state of the bearing are labeled 1, 2, 3 and 4 respectively. The training set and the test set are divided into 100 and 20 training sets and 20 test sets respectively for each fault. The total number of training sets and test sets for the four faults is 400 and 80 respectively.

[0067] (4) Build a deep convolutional neural network image learning model:

[0068] 1) Input layer, the phase trajectory image size is uniformly preprocessed to 227×227;

[0069] 2) C1 convolutional layer, the number of kernels is 24, the size is 11×11, and the stride is 4;

[0070] 3) P1 pooling layer, the pooling window size is 3×3, the step size is 2, and the maximum pooling method is used;

[0071] 4) C2 convolution layer, the number of kernels is 24, the size is 5×5, and the stride is 1;

[0072] 5) P2 pooling layer, the pooling window size is 3×3, the step size is 2, and the maximum pooling method is used;

[0073] 6) C3 convolutional layer, the number of kernels is 48, the size is 5×5, and the stride is 1;

[0074] 7) P3 pooling layer, the pooling window size is 3×3, the step size is 2, and the maximum pooling method is used;

[0075] 8) F1 fully connected layer;

[0076] 9) F2 fully connected layer, in which a random dropout layer is provided;

[0077] 10) Use the Softmax algorithm for prediction and recognition and output the diagnosis results.

[0078] (5) Input the training set into the deep convolutional neural network model for learning. The process mainly includes two processes: forward propagation and reverse weight update. After finding the minimum cross entropy loss function, stop the iteration and save the diagnostic model. The training accuracy and loss are as follows Figure 3 As shown in the figure, the training accuracy after the iteration is 100%, the loss is 0.000058, and the training effect is good.

[0079] (6) To improve the reliability of the evaluation criteria, the vibration signals of the inner race faults with damage sizes of 0, 0.007 and 0.021 mm are divided into windows. The window length is 1000 points and the number of windows is 100. For the three damage sizes of the inner race faults, the maximum Lyapunov exponent of each window is calculated.

[0080] The maximum Lyapunov exponent is solved using the Rosenstein algorithm:

[0081] 1) For each phase point X in the reconstructed phase space a and its neighboring point Y b , after time t, the distance between the two phase points is:

[0082]

[0083] Where t = 1, 2, ..., min (Na, Nb), γ is X a and Y b The divergence rate, C a =d a (0).

[0084] 2) Taking the logarithm of both sides:

[0085] ln d a (t) = ln C a +γ(t×Δt)

[0086] Where Δt represents the sampling interval of the time series.

[0087] 3) Curve lnd a (t)~t have a linear relationship within a certain period of time, and the above equation can be linearized into:

[0088]

[0089] Where q represents non-zero d a (t) quantity.

[0090] 4) Perform a linear fit on y(t) and the slope is the maximum Lyapunov exponent.

[0091] (7) According to statistics, the maximum Lyapunov index averages of the inner ring fault under the damage sizes of 0, 0.007 and 0.021 mm are 0.18, 0.09 and 0.12, which are used as the health assessment criteria under the inner ring fault, such as Figure 4 shown.

[0092] (8) The test set in step (3) is used as monitoring data and input into the diagnosis model for classification to complete the diagnosis of the main bearing fault type. The confusion matrix of the fault diagnosis result is as follows: Figure 5 As shown in the figure, only one inner ring fault sample was misdiagnosed as an outer ring fault, with a diagnosis rate of 98.8%. After the specific fault is diagnosed, if the diagnosis result is an inner ring fault, further health assessment is performed according to the health assessment standard in step (7) according to the proximity principle. The proximity principle means that the maximum Lyapunov exponent of the test signal is compared with the health assessment standard, and the closest level of standard in value is the health state.

[0093] (9) The misdiagnosed samples are input into the model for further training, and the trained model parameters are input into the previous diagnosis model through transfer learning to improve the diagnosis rate.

[0094] The health assessment of bearings has different accuracy requirements in different scenarios. The assessment of military diesel engine bearings requires precise quantification, and cannot give a rough estimate of the bearing damage or degradation stage. On the other hand, for civilian diesel engine bearings, a quick rough estimate is preferred, that is, determining the approximate level of degradation. After using convolutional neural networks and phase trajectory diagrams to make fault diagnosis, Lyapunov exponents can be used for rough assessment or fine quantitative assessment.

[0095] Detailed assessment means calculating the Lyapunov index value under normal conditions and taking the Lyapunov index value under fault conditions and its deviation degree as the health level.

[0096] Rough evaluation is to calculate the Lyapunov exponents in the four stages of normal state and fault degradation. After inputting the monitoring signal, the stage that is close to it is observed, which can achieve rough evaluation.

Claims

1. A health assessment method for diesel engine main bearings, It is characterized in that The following steps are involved: (1) collecting vibration signals of a main bearing of a diesel engine, and converting the vibration signals into a phase trajectory diagram using a phase space reconstruction method; (2) calculating the maximum Lyapunov exponent of the vibration signal according to the phase trajectory diagram; inputting the phase trajectory diagram into the diagnosis model to complete the diagnosis of the fault type of the collected vibration signal, and inputting the result into the health assessment model; the establishment of the health assessment model includes the following steps: (21) Collect vibration signals of diesel engine main bearings in various fault types and under different health conditions; (22) The phase space reconstruction method is used to convert the vibration signals under different health states into phase trajectory diagrams; (23) calculating the maximum Lyapunov exponent according to the phase trajectory diagram obtained in step (22); (24) For each fault type of the bearing, the maximum Lyapunov index mean of each fault type under different health conditions is calculated and used as the health assessment standard; (3) the health assessment model performs health assessment on the fault type diagnosed in step (2) according to the health assessment standard under the fault type and the maximum Lyapunov exponent of the vibration signal obtained in step (2); The health assessment described in step (3) is performed according to the health assessment standard under the fault type and the maximum Lyapunov exponent of the vibration signal obtained in step (2), including fine assessment and rough assessment. The fine assessment refers to calculating the Lyapunov exponent value under normal conditions, and taking the Lyapunov exponent value under fault conditions and the degree of deviation as the health degree; the rough assessment refers to calculating the Lyapunov exponent under four stages of normal conditions and fault degradation, and comparing the maximum Lyapunov exponent of the signal to be evaluated with the health assessment standard. The health state corresponding to the health assessment standard that is closest in value is the assessment result; The establishment of the diagnostic model comprises the following steps: (1) Collect fault signals of various faults of diesel engine main bearings; (2) The vibration signals under different faults are converted into phase trajectory diagrams using the phase space reconstruction method; (3) Label the phase trajectory diagrams under different main bearing faults and divide them into training sets and test sets; (4) Establish an image learning model of a deep convolutional neural network and set parameters for the input layer, convolution layer, pooling layer, and fully connected layer; (5) Input the training set into the deep convolutional neural network model for learning. The training process includes forward propagation and reverse weight update. After finding the minimum cross entropy loss function, the iteration is stopped to obtain the diagnostic model. The phase space reconstruction method comprises the following steps: (11) The one-dimensional fault vibration signal is reconstructed in phase space. The algorithm is as follows: X i =(x i ,x i+τ ,x i+2τ ,...,x i+(m-1)τ ) Where x is the fault vibration signal, X i is the phase point matrix, i=1,2,...,N, N is the number of phase points, m is the embedding dimension, τ is the delay time; (12) The reconstructed high-dimensional space phase trajectory is projected by the main vector to obtain the visual phase trajectory diagram. The specific algorithm is as follows: Let the matrix A = X T X, after performing singular value decomposition on A, we get m eigenvalues ​​and eigenvectors, and select the largest 3 eigenvalues ​​λ 1 ,λ 2 ,λ 3 , the corresponding eigenvector V 1 ,V 2 ,V 3 As the principal vector of matrix A, the reconstructed phase space is projected to the principal vector direction to obtain the N×3 order principal vector matrix B: B=X[V 1 ,V 2 ,V 3 ] Plot the matrix B in a three-dimensional coordinate system to visualize the phase trajectory.

2. The health assessment method for a diesel engine main bearing according to claim 1, Features: During the training process of step (5), the misdiagnosed samples are input into the model to continue training, and the trained model parameters are input into the previous diagnosis model through transfer learning to continuously correct the model.

3. The health assessment method for a diesel engine main bearing according to claim 1, Features: The phase trajectory diagrams under different main bearing faults are labeled and divided into training sets and test sets in a ratio of 8:

2.

4. The health assessment method for a diesel engine main bearing according to claim 1, Features: The image learning model of establishing a deep convolutional neural network uses the Softmax algorithm for prediction and recognition and outputs the diagnosis results. The model structure includes: 1) Input layer, the phase trajectory image size is uniformly preprocessed to 227×227; 2) C1 convolutional layer, the number of kernels is 24, the size is 11×11, and the stride is 4; 3) P1 pooling layer, the pooling window size is 3×3, the step size is 2, and the maximum pooling method is used; 4) C2 convolution layer, the number of kernels is 24, the size is 5×5, and the stride is 1; 5) P2 pooling layer, the pooling window size is 3×3, the step size is 2, and the maximum pooling method is used; 6) C3 convolution layer, the number of kernels is 48, the size is 5×5, and the stride is 1; 7) P3 pooling layer, the pooling window size is 3×3, the step size is 2, and the maximum pooling method is used; 8) F1 fully connected layer; 9) F2 fully connected layer, in which a random dropout layer is provided.

5. The method for health assessment of a diesel engine main bearing according to claim 1, Features: The maximum Lyapunov exponent solution comprises the following steps: 1) For each phase point X in the reconstructed phase space a and its neighboring point Y b , after time t, the distance between the two phase points is: d a (t)=C a e γ(t×Δt) Where t=1,2,…,min(Na,Nb),γisX a and Y b The divergence rate, C a =d a (0); Δt represents the sampling interval of the time series; 2) Taking the logarithm of both sides: in d a (t)=lnC a +γ(t×Δt) 3) Linearize the above equation into: Where q represents non-zero d a (t) quantity; 4) Perform a linear fit on y(t) and the slope is the maximum Lyapunov exponent.

6. The method for health assessment of a diesel engine main bearing according to claim 1, Features: The bearing fault types include a fault state and a normal state; the fault states include an inner ring fault, an outer ring fault, and a ball fault.

Images