A gear fault prediction method based on embedded damage evolution mechanism

Abstract

The present invention provides a gear fault prediction method based on an embedded damage evolution mechanism, comprising: collecting vibration acceleration data of the gear; setting state variables of the gear fault, and constructing a gear fault prediction model based on the vibration acceleration data and the state variables; training the gear fault prediction model based on the vibration acceleration data; and predicting the evolution of the gear fault based on the trained gear fault prediction model. The present invention can achieve accurate prediction of gear dynamics and faults.

Description

Technical Field

[0001] The present invention relates to the field of data-driven gear dynamic modeling and prediction, and in particular to a gear fault prediction method based on an embedded damage evolution mechanism. Background Art

[0002] Dynamics and fault prediction methods are an important part of the prediction and health management technology of rotating machinery, and their performance is directly related to the level of predictive maintenance of key components. Traditional physical drive prediction methods lead to certain simplifications in modeling and neglect of real characteristics, and cannot accurately predict the impact of uncertainty and strong nonlinear factors on dynamic response during the operation of the gear system; at the same time, the currently popular data-driven prediction methods mainly rely on the integrity of the collected data, which makes it difficult to accurately reflect the original primary characteristics of the power system, resulting in the loss of non-stationary and transient characteristics, and cannot meet the dynamics (including vibration acceleration, etc.) and fault prediction accuracy requirements during the dynamic operation of the gear. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes a gear fault prediction method based on an embedded damage evolution mechanism to achieve accurate prediction of gear dynamics and faults.

[0004] To achieve the above object, the present invention provides a gear fault prediction method based on embedded damage evolution mechanism, comprising the following steps:

[0005] Collect vibration acceleration data of gears;

[0006] Setting a state variable of a gear fault, and constructing a gear fault prediction model based on the vibration acceleration data and the state variable;

[0007] Training the gear fault prediction model based on the vibration acceleration data;

[0008] Based on the trained gear fault prediction model, the evolution of gear faults is predicted.

[0009] Optionally, the state variables of the gear fault include: tooth root crack length, pitting density and wear depth.

[0010] Optionally, constructing a gear fault prediction model based on the vibration acceleration data and the state variable includes:

[0011] Constructing a gear time-varying meshing stiffness calculation module, taking the state variable of the previous period as input, and outputting the gear time-varying meshing stiffness of the previous period through the gear time-varying meshing stiffness calculation module;

[0012] Constructing a dynamic reverse deduction neural network module, taking the gear time-varying meshing stiffness of the previous period and the vibration acceleration data of the previous period as input, and outputting the meshing force of the previous period through the dynamic reverse deduction neural network module;

[0013] Constructing a state variable increment calculation module, taking the state variable as input, and outputting the state variable increment through the state variable increment calculation module;

[0014] Constructing a gear time-varying meshing stiffness calculation module for the current period, taking the state variable increment and the state variable of the previous period as input, and outputting the gear time-varying meshing stiffness for the current period through the gear time-varying meshing stiffness calculation module for the current period;

[0015] Constructing a dynamics forward deduction neural network module, taking the time-varying meshing stiffness of the gear in the current period as input, and outputting the vibration acceleration data of the current period through the dynamics forward deduction neural network module;

[0016] The gear time-varying mesh stiffness calculation module, the dynamics reverse deduction neural network module, the state variable increment calculation module, the current period gear time-varying mesh stiffness calculation module and the dynamics forward deduction neural network module are combined to construct a hybrid neural network gear dynamics prediction module;

[0017] A plurality of the hybrid neural network gear dynamics prediction modules are connected in series to construct the gear fault prediction model.

[0018] Optionally, constructing the state variable increment calculation module includes:

[0019] A tooth root crack propagation model, a tooth surface pitting density growth model and a tooth surface wear growth model are constructed, and parameters are set; and according to the tooth root crack propagation model, the tooth surface pitting density growth model and the tooth surface wear growth model, the state variable increment calculation module is constructed.

[0020] Optionally, the tooth root crack growth model is:

[0021]

[0022] Where m and C are constants; ΔK is the range of stress intensity factor, and da / dN is the crack growth rate.

[0023] Optionally, the tooth surface pitting density growth model is:

[0024]

[0025] Among them, K pitting is the pitting coefficient, τ0 is the maximum shear stress, V is the shear volume, z0 is the depth of maximum shear stress, Ncycle is the number of gear rotations, θ is the gear rotation angle, r, s and b are material-related coefficients, dD(N cycle ,θ) / dN cycle is the growth rate of pitting density on the tooth surface.

[0026] Optionally, the tooth surface wear growth model is:

[0027]

[0028] Among them, K wear is the dimensionless wear coefficient, P contact is the contact pressure; L sliding is the slip distance, dh / dN cycle is the wear rate.

[0029] Optionally, predicting the evolution of gear faults based on the trained gear fault prediction model includes:

[0030] The collected vibration acceleration data is input into the trained gear fault prediction model to obtain the vibration acceleration data of a preset future time period and the state variables of the gear fault corresponding to the future time period.

[0031] Compared with the prior art, the present invention has the following advantages and technical effects:

[0032] The present invention can take into account the advantages of both the physical-driven prediction method and the data-driven prediction method, give full play to the advantages of data-driven short-term local prediction and physical-driven long-term global prediction, and realize accurate prediction of gear dynamics and faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0034] Figure 1 A schematic flow chart of a gear fault prediction method based on an embedded damage evolution mechanism according to an embodiment of the present invention;

[0035] Figure 2 A schematic diagram of a hybrid neural network gear dynamics prediction module according to an embodiment of the present invention;

[0036] Figure 3 A schematic diagram of a hybrid neural network gear dynamics prediction model according to an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of vibration acceleration data collection according to an embodiment of the present invention;

[0038] Figure 5 Schematic diagram of changes in RMSE and Loss during the training process of an embodiment of the present invention;

[0039] Figure 6 A schematic diagram of vibration acceleration prediction results according to an embodiment of the present invention;

[0040] Figure 7 Schematic diagram of damage and wear prediction results of an embodiment of the present invention. DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0042] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0043] Example

[0044] like Figure 1 As shown, this embodiment proposes a gear fault prediction method based on an embedded damage evolution mechanism, including:

[0045] Collect vibration acceleration data of gears;

[0046] Setting a state variable of a gear fault, and constructing a gear fault prediction model based on the vibration acceleration data and the state variable;

[0047] Training the gear fault prediction model based on the vibration acceleration data;

[0048] Based on the trained gear fault prediction model, the evolution of gear faults is predicted.

[0049] Furthermore, the state variables of the gear fault include: tooth root crack length, pitting density and wear depth.

[0050] Further, constructing a gear fault prediction model based on the vibration acceleration data and the state variable includes:

[0051] Constructing a gear time-varying meshing stiffness calculation module, taking the state variable of the previous period as input, and outputting the gear time-varying meshing stiffness of the previous period through the gear time-varying meshing stiffness calculation module;

[0052] Constructing a dynamic reverse deduction neural network module, taking the gear time-varying meshing stiffness of the previous period and the vibration acceleration data of the previous period as input, and outputting the meshing force of the previous period through the dynamic reverse deduction neural network module;

[0053] Constructing a state variable increment calculation module, taking the state variable as input, and outputting the state variable increment through the state variable increment calculation module;

[0054] Constructing a gear time-varying meshing stiffness calculation module for the current period, taking the state variable increment and the state variable of the previous period as input, and outputting the gear time-varying meshing stiffness for the current period through the gear time-varying meshing stiffness calculation module for the current period;

[0055] Constructing a dynamics forward deduction neural network module, taking the time-varying meshing stiffness of the gear in the current period as input, and outputting the vibration acceleration data of the current period through the dynamics forward deduction neural network module;

[0056] The gear time-varying mesh stiffness calculation module, the dynamics reverse deduction neural network module, the state variable increment calculation module, the current period gear time-varying mesh stiffness calculation module and the dynamics forward deduction neural network module are combined to construct a hybrid neural network gear dynamics prediction module;

[0057] A plurality of the hybrid neural network gear dynamics prediction modules are connected in series to construct the gear fault prediction model.

[0058] Furthermore, constructing the state variable increment calculation module includes:

[0059] A tooth root crack propagation model, a tooth surface pitting density growth model and a tooth surface wear growth model are constructed, and parameters are set; and according to the tooth root crack propagation model, the tooth surface pitting density growth model and the tooth surface wear growth model, the state variable increment calculation module is constructed.

[0060] Further, predicting the evolution of gear faults based on the trained gear fault prediction model includes:

[0061] The collected vibration acceleration data is input into the trained gear fault prediction model to obtain the vibration acceleration data of a preset future time period and the state variables of the gear fault corresponding to the future time period.

[0062] like Figure 1 As shown, the gear fault prediction method based on the embedded damage evolution mechanism proposed in this embodiment includes four steps, specifically:

[0063] 1) Data acquisition: First, an acceleration sensor is used to measure the vibration acceleration data of the gearbox end cover of the transmission system. Then, the signal burrs are removed through the data algorithm. Finally, the signal noise is removed by the filtering algorithm to obtain a pure acceleration data sequence [x0, x1, x2, ... x n-1 ,x n ], where x n Represents the acceleration vector of the nth period, which is important data for model training.

[0064] Furthermore, collecting performance data also includes: collecting and obtaining performance parameter vectors x of the electromechanical equipment such as vibration acceleration, stress, voltage, current, etc. through sensors such as vibration acceleration, stress, voltage, current, etc. arranged on the electromechanical equipment. In this embodiment, only vibration acceleration is used. For n devices, the performance parameter vectors obtained in each same monitoring time unit are x1, x2, ..., x n Therefore, after monitoring m time units, the performance parameter obtained is X1 = [x 1,1 ,x 2,1 ,…,x n,1 ],X2=[x 1,2 ,x 2,2 ,…,x n,2 ], ..., X m =[x 1,m ,x 2,m ,…,x n,m ].

[0065] After collecting the performance data, it also includes: constructing a feature matrix: for each time unit of n devices, calculating its feature vector θ. Therefore, for m time units, a feature matrix A consisting of n×m units can be obtained.

[0066]

[0067] Construct a feature matrix. Specifically, assuming that the mean is selected as the feature, for each time unit, the i-th device and the j-th time unit, its feature can be calculated as:

[0068] θ i,j =f(x i,j )

[0069] Where f(·) represents the calculation function of the feature quantity.

[0070] When the feature is the mean, the feature is calculated as:

[0071] θ i,j =mean(x i,j )

[0072] When the feature is standard deviation, the feature is calculated as:

[0073] θ i,j = std(x i,j )

[0074] When the feature is a valid value, the feature is calculated as:

[0075] θ i,j =rms(x i,j )

[0076] Therefore, for each time unit of n devices, its feature vector θ is calculated. Therefore, for m time units, a feature matrix A consisting of n×m units can be obtained.

[0077] 2) Hybrid neural network modeling (i.e. building a gear fault prediction model): define the tooth root crack length, pitting density and wear depth as state variables a n-1 , D n-1 and h n-1 , the input variable is the vibration acceleration x in the previous period t-1 , the output variable is the vibration acceleration x in the current period n .

[0078] The first step is to use the state variable a of the previous period n-1 , D n-1 and h n-1 , design the calculation module of the time-varying meshing stiffness of the gear in the previous period. The input of the calculation module of the time-varying meshing stiffness of the gear in the previous period is the state variable a n-1 , D n-1 and h n-1 , the output is the time-varying mesh stiffness k m , the module can be constructed by using (but not limited to) potential energy method, Ishikawa method, etc. The time-varying meshing stiffness model (calculation module) defined by the potential energy method is as follows:

[0079]

[0080]

[0081]

[0082]

[0083]

[0084] Where U b , U s , U a and U h are bending potential energy, shear potential energy, axial compression potential energy and contact potential energy respectively; F is meshing force; kb , k s , k a and k h are bending stiffness, shear stiffness, axial compression stiffness and contact stiffness respectively; E is Young's modulus; G is shear modulus; A x , d, h are the geometric dimensions of the gear teeth; k m F is the time-varying meshing stiffness; n is the number of meshing gear teeth. a is the orthogonal component of F, F b is the orthogonal component of F, I x is the moment of inertia, and i is the gear tooth code.

[0085] The second step is to select the time-varying meshing stiffness k of the previous period. m and vibration acceleration x n-1 The dynamic reverse deduction neural network module is designed with the input of vibration acceleration and meshing stiffness of the previous period and the output of meshing force. The dynamic reverse deduction neural network module input is the vibration acceleration of the previous period and the meshing stiffness of the previous period, and the output is the meshing force. The neural network building modules such as BP neural network, convolution neural network, RBF neural network, etc. can be used (but not limited to). The neural network mathematical model of this module is as follows:

[0086] y invserse_dynamics =NN inverse_dynamics (x invserse_dynamics |θ invserse_dynamics )

[0087] In the formula, NN inverse_dynamics (·) is the neural network, θ inverse_dynamics are neural network parameters, x inverse_dynamics is the input variable, y inverse_dynamics is the output variable.

[0088] The third step is to set the parameters of the tooth root crack propagation model, tooth surface pitting density growth model and tooth surface wear growth model according to the material and manufacturing process, and design the state variable increment calculation module.

[0089] The tooth root crack growth model is:

[0090]

[0091] In formula (1), m and C are constants; ΔK is the range of stress intensity factor, and da / dN is the crack growth rate.

[0092] The growth model of tooth surface pitting density is:

[0093]

[0094] In formula (2), K pittingis the pitting coefficient, which is related to factors such as surface roughness, lubrication and hardness; τ0 is the maximum shear stress; V is the shear volume; z0 is the depth of maximum shear stress; N cycle is the number of gear rotations; θ is the gear rotation angle; r, s and b are coefficients related to the material, dD(N cycle ,θ) / dN cycle is the growth rate of pitting density on the tooth surface.

[0095] The tooth surface wear growth model is:

[0096]

[0097] In formula (3), K wear is the dimensionless wear coefficient, which is related to factors such as surface roughness, lubrication and hardness; P contact is the contact pressure; L sliding is the slip distance.

[0098] The fourth step is to select the state variable increment and the state variable of the previous period as input to design the gear time-varying mesh stiffness calculation module of the current period. The gear time-varying mesh stiffness calculation module of the current period is the same as the gear time-varying mesh stiffness calculation module of the previous period, and both modules can be constructed by (but not limited to) potential energy method, Ishikawa method, etc.

[0099] Step 5: Select the current period gear time-varying meshing stiffness as input, the current period vibration machine speed x n For the output, a dynamic forward deduction neural network module is designed. The input of the dynamic forward deduction neural network module is the time-varying meshing stiffness of the current period, and the output is the vibration acceleration of the current period. Neural network building modules such as BP neural network, convolutional neural network, RBF neural network, etc. can be used (but not limited to). The neural network mathematical model of this module is as follows:

[0100] y dynamics =NN dynamics (x dynamics |θ dynamics )

[0101] In the formula, NN dynamics (·) is the neural network, θ dynamics are neural network parameters, x dynamics is the input variable, y dynamics is the output variable.

[0102] Step 7: Follow Figure 2 The structure shown in , combines the above modules to establish a hybrid neural network gear dynamics prediction module.

[0103] In the eighth step, the n-1 modules in the previous step are connected in series to establish a hybrid neural network gear dynamics prediction model, such as Figure 3 As shown;

[0104] The model is mainly composed of four parts: the first part is the output module, which is mainly used for the input of the feature matrix; the second part is the Multi-RNN Unit, which is composed of the first RNN, the second RNN, ..., and the nth RNN in parallel, and is used to carry out prediction work under the training data of different electromechanical equipment; the third part is the integrated decision module, which uses the Monte Carlo algorithm to perform statistical inference on the prediction results of each RNN; the fourth part is the output module, which is used to output the prediction results. The four parts in the model are designed in series structure and finally output the prediction results.

[0105] In the above new probabilistic deep neural network model, the design consists of four parts: collecting performance data, building feature matrix, training model, and predicting performance data and reliability. Specifically, the prediction method steps are as follows:

[0106] 3) Model training: Based on the vibration acceleration data obtained from data collection, the data is used as input to train the hybrid neural network gear dynamics prediction model. First, the initial values ​​of the neural network related parameters are set; then, the parameters of the training methods such as the steepest descent method and the Adam method are set using supervised learning; finally, the network model parameters are trained through the input vibration acceleration data until the model converges;

[0107] Among them, the characteristic matrix A of the performance parameters is input into the probabilistic deep neural network model, and the Multi-RNN Unit of the model is trained. The total loss function L is expressed as

[0108]

[0109] For the above loss function, the model training can be completed by using the supervised learning training method.

[0110] 4) Dynamics and fault prediction: Input a sequence of collected vibration acceleration data x n Finally, we obtain the dynamic prediction data x of 2n time periods. 2n , and obtain the predicted data a of the damage state variable corresponding to each time period n+1:n+m , D n+1:n+m and h n+1:n+m .

[0111] Based on the trained model, the performance degradation trend and reliability of electromechanical equipment are analyzed. Finally, the performance degradation trend of electromechanical equipment in its entire life cycle is plotted.

[0112]

[0113] In the formula, Predicted value for each RNN output.

[0114] At the same time, the confidence interval of the estimated performance under probability (1-α) The calculation is as follows:

[0115]

[0116]

[0117] In the formula, rank(·) represents the ranking of features.

[0118] For the performance status at a certain time t, determine its reaching performance failure threshold θ m The confidence interval probability boundary β m =(1-α m ), then it is the reliability R(t) of the electromechanical equipment at that moment, that is,

[0119] R(t)=β m .

[0120] The method proposed in this embodiment can fully tap into the rich uncertainty and strong nonlinear characteristics in the collected data, while giving full play to the correction effect of mechanism knowledge on dynamic response, accurately describing the changing characteristics of non-stationary and transient processes, taking into account the advantages of physical drive and data drive methods, and achieving high-precision prediction of gear dynamic response and damage state. The following is an example of long-term prediction of a gear system.

[0121] 1) The data acquisition was carried out using a spur gear full life cycle test bench. The data acquisition instrument was the MDR80 mobile data acquisition and recording system, which was equipped with 16 channels for dynamic sampling and had the advantages of strong environmental adaptability and diversified acquisition control. The sensor used an ICP accelerometer, and the sampling frequency was set to 5000Hz. The acceleration curve obtained by sampling is shown in the figure below: Figure 4 shown.

[0122] 2) The first 50 seconds of vibration acceleration data collected are used as the training data set, and the prediction network is trained using the steepest descent method. The training process is as follows: Figure 5 As shown in Figure 2, it can be seen that the loss curve gradually stabilizes and the model training is completed.

[0123] 3) Use the trained model to make predictions, and finally realize the prediction of vibration acceleration data, such as Figure 6 At the same time, the predicted damage and wear values ​​of crack initiation are shown as Figure 7 shown.

[0124] It can be seen from the above content that the model and method disclosed in the present invention can give full play to the advantages of data-driven short-term local prediction and physical-driven long-term global prediction.

[0125] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A gear fault prediction method based on embedded damage evolution mechanism, characterized in that: include: Collect vibration acceleration data of gears; Setting a state variable of a gear fault, and constructing a gear fault prediction model based on the vibration acceleration data and the state variable; Constructing a gear fault prediction model based on the vibration acceleration data and the state variable includes: Constructing a gear time-varying meshing stiffness calculation module, taking the state variable of the previous period as input, and outputting the gear time-varying meshing stiffness of the previous period through the gear time-varying meshing stiffness calculation module; Constructing a dynamic reverse deduction neural network module, taking the gear time-varying meshing stiffness of the previous period and the vibration acceleration data of the previous period as input, and outputting the meshing force of the previous period through the dynamic reverse deduction neural network module; Constructing a state variable increment calculation module, taking the state variable as input, and outputting the state variable increment through the state variable increment calculation module; Constructing a gear time-varying meshing stiffness calculation module for the current period, taking the state variable increment and the state variable of the previous period as input, and outputting the gear time-varying meshing stiffness for the current period through the gear time-varying meshing stiffness calculation module for the current period; Constructing a dynamics forward deduction neural network module, taking the time-varying meshing stiffness of the gear in the current period as input, and outputting the vibration acceleration data of the current period through the dynamics forward deduction neural network module; The gear time-varying mesh stiffness calculation module, the dynamics reverse deduction neural network module, the state variable increment calculation module, the current period gear time-varying mesh stiffness calculation module and the dynamics forward deduction neural network module are combined to construct a hybrid neural network gear dynamics prediction module; Connecting several of the hybrid neural network gear dynamics prediction modules in series to construct the gear fault prediction model; Training the gear fault prediction model based on the vibration acceleration data; Based on the trained gear fault prediction model, the evolution of gear faults is predicted.

2. The gear fault prediction method based on embedded damage evolution mechanism according to claim 1 is characterized in that: The state variables of the gear fault include: tooth root crack length, pitting density and wear depth.

3. The gear fault prediction method based on embedded damage evolution mechanism according to claim 1 is characterized in that: Constructing the state variable increment calculation module includes: A tooth root crack propagation model, a tooth surface pitting density growth model and a tooth surface wear growth model are constructed, and parameters are set; and according to the tooth root crack propagation model, the tooth surface pitting density growth model and the tooth surface wear growth model, the state variable increment calculation module is constructed.

4. The gear fault prediction method based on embedded damage evolution mechanism according to claim 3 is characterized in that: The tooth root crack growth model is: Where m and C are constants; ΔK is the range of stress intensity factor, and da / dN is the crack growth rate.

5. The gear fault prediction method based on embedded damage evolution mechanism according to claim 3 is characterized in that: The tooth surface pitting density growth model is: Among them, K pitting is the pitting coefficient, τ0 is the maximum shear stress, V is the shear volume, z0 is the depth of maximum shear stress, N cycle is the number of gear rotations, θ is the gear rotation angle, r, s and b are material-related coefficients, dD(N cycle ,θ) / dN cycle is the growth rate of pitting density on the tooth surface.

6. The gear fault prediction method based on embedded damage evolution mechanism according to claim 3 is characterized in that: The tooth surface wear growth model is: Among them, K wear is the dimensionless wear coefficient, P contact is the contact pressure; L sliding is the slip distance, dh / dN cycle is the wear rate.

7. The gear fault prediction method based on embedded damage evolution mechanism according to claim 1 is characterized in that: Predicting the evolution of gear faults based on the trained gear fault prediction model includes: The collected vibration acceleration data is input into the trained gear fault prediction model to obtain the vibration acceleration data of a preset future time period and the state variables of the gear fault corresponding to the future time period.

Images