Industrial robot speed reducer fault diagnosis method based on VMD-ELM

Abstract

The invention provides an industrial robot speed reducer early fault diagnosis method based on VMD (variational mode decomposition) feature extraction and an ELM (extreme learning machine) aiming at the characteristics that early fault vibration signals of an industrial robot speed reducer are weak, fault feature frequency is difficult to decompose and the like. The method comprises the following steps: acquiring an original signal by using a vibration sensor and data acquisition equipment, then performing feature extraction on the acquired signal by using a self-adaptive variational mode decomposition (VMD) method, then establishing an extreme learning machine model, and performing fault diagnosis and identification on the industrial robot speed reducer by taking a fault feature as input and a fault diagnosis tag result as output.

Description

Technical Field

[0001] The present invention relates to signal detection and fault diagnosis, and specifically to a fault diagnosis method for an industrial robot reducer based on variational mode decomposition and extreme learning machine. Background Art

[0002] The reducers of each joint of an industrial robot are high-precision components that maintain its normal operation. The stability and reliability of the industrial robot system are affected by the deterioration degree of the reducer. When the deterioration degree of the robot reducer deepens, it may lead to the shutdown of the industrial robot due to faults. Since the reducer is installed inside the robot, it is difficult to monitor its precise state parameters during operation. Therefore, the research on the online fault diagnosis of the harmonic reducer of industrial robots is of great significance.

[0003] The acceleration vibration signal in the vibration signal can more sensitively reflect the working conditions between the various components of the reducer. When there is contact between the components during operation, impact vibration will be generated. The vibration signal contains rich fault feature information, which can more effectively detect the early faults of the reducer and reduce the occurrence of accidents. Summary of the Invention

[0004] In order to achieve the above object, the present invention proposes a fault diagnosis method for an industrial robot reducer based on VMD-ELM. First, the variational mode decomposition (VMD) method is used to extract features from the original signal, and then an extreme learning machine model is established to obtain a fault diagnosis model for the reducer, so as to effectively diagnose the faults of the reducer. To achieve the above object, the technical solution adopted by the present invention is as follows:

[0005] A fault diagnosis method for an industrial robot reducer based on VMD-ELM includes the following steps:

[0006] Under different working conditions, collect the vibration signals of the industrial robot reducer; perform VMD feature extraction respectively to obtain the VMD features in different states, and make a paired data set;

[0007] Train the extracted paired data set based on the extreme learning machine to obtain a fault diagnosis model;

[0008] After the test sample data collected on-site is subjected to feature extraction, input it into the trained fault diagnosis model to obtain the fault state diagnosis result of the industrial robot reducer.

[0009] The vibration data is collected by a vibration sensor installed on the reducer.

[0010] The different states or fault labels include: normal without fault, pitting, gear fracture, and excessive wear of parts.

[0011] The paired data set to be produced is {state feature data, state label}, and it is necessary to pre-distinguish the IMF component regions in different states into labels of different states by experts.

[0012] VMD feature extraction is respectively performed on the vibration signals in different states collected, and the feature vectors in different states are respectively obtained, including:

[0013] a. Establish a constrained variational model:

[0014]

[0015] where, u k is the k-th modal component of VMD decomposition, ω k is the center frequency of the component signal, denotes derivative with respect to t, δ(t) is the Dirac function, * is convolution, is the exponential term, and f is the input feature data;

[0016] b. Introduce the quadratic penalty term α and the Lagrange multiplier λ to obtain the augmented Lagrangian formula, and use the alternating direction method of multipliers to solve the extreme value of the augmented Lagrangian expression, and then decompose the original signal into k modal components, and the modal components are the VMD features to be extracted.

[0017] The solution process is as follows:

[0018] b1. Initialize the IMF parameters ω k , u k , α, λ and n;

[0019] where u k is the modal component, ω k is the center frequency, λ is the Lagrangian algorithm factor, α is the penalty factor, n is the number of loops, and k is the number of modal decompositions;

[0020] b2. Update the center frequency ω k and the modal component u k according to the formula;

[0021]

[0022]

[0023] where, is the Wiener filter of the current residue, is the centroid of the power spectrum of the modal function in the (n + 1)-th iteration, ω is the center frequency, and ω k is the center frequency of the k-th modal component;

[0024] b3. Update λ according to the formula;

[0025]

[0026] where τ is the fidelity coefficient;

[0027] b4. Set the loop condition. When the following conditions are met, the loop ends; otherwise, return to step (2) to finally obtain k IMF components;

[0028]

[0029] Construct an extreme learning machine and output the fault result. The steps are as follows:

[0030] Input N training paired data sets (x, y) = {(x i , y i ) | x i ∈ R n , y i ∈ R m , i = 1, 2,..., N}, where x i = (x i1 , x i2 ,..., x in ) T is the i-th input vector, and y i = (y i1 , y i2 ,..., y in ) T is the true label corresponding to the i-th input vector;

[0031] During the training process, the predicted output is obtained:

[0032]

[0033] In the formula, is the predicted label corresponding to the i-th input vector, ω j is the weight vector of the j-th hidden layer node, b j is the bias of the j-th hidden layer node, and β j is the weight vector connecting the j-th hidden layer node and the output layer node. g(·) is the activation function.

[0034] The above formula is converted into matrix form as:

[0035]

[0036] In the formula, is the predicted output matrix, H is the hidden layer output matrix, and β is the output weight matrix; the specific expressions are:

[0037]

[0038]

[0039] Minimizing the error E between the predicted label and the true label is:

[0040]

[0041] Solving the least - squares solution of the above formula:

[0042] β = H + Y

[0043] where, H + is the Moore - Penrose generalized inverse of the matrix; introducing the regularization parameter C in the above formula to provide a more applicable solution:

[0044]

[0045] In the formula, I is the identity matrix.

[0046] The finally obtained prediction output model is:

[0047]

[0048] The training is to input the paired data set {state feature data, state label} into the final model, set the training parameters, and perform iterative training using the Adam optimizer. The training stops when the network loss converges, or reaches the number of training times, or the predicted label meets the accuracy requirement.

[0049] The fault state diagnosis result is the fault label of different states.

[0050] The present invention has the following advantages and beneficial effects:

[0051] 1. In the data processing process of the present invention, the variational mode decomposition (VMD) method is used to decompose the collected vibration signal. This decomposition method assumes in advance that each modal component has a central frequency and uses this frequency to determine each modal component, having better anti - noise performance.

[0052] 2. The extreme learning machine model established by the present invention has a fast learning speed, strong generalization ability, and has the advantages of short training time and high accuracy, improving the accuracy of fault diagnosis.

[0053] 3. Based on the VMD method and the extreme learning machine algorithm, the present invention proposes a fault diagnosis method for a robot reducer, which can combine multiple fault information and more accurately and comprehensively evaluate the health state of the reducer. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is the flow chart of the present invention;

[0055] Figure 2 is the flow chart of VMD feature extraction

[0056] Figure 3 is the flow chart of ELM extreme learning machine training Specific implementation manners

[0057] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe in detail the specific implementation method of the present invention with reference to the accompanying drawings. Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the invention. Therefore, the present invention is not limited by the specific implementations disclosed below.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this invention belongs. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0059] The present invention uses the VMD method to extract features from the original vibration signal, and then trains the extracted feature vectors based on the extreme learning machine to obtain a reducer fault diagnosis model for effectively diagnosing the early faults of the reducer. As Figure 1 shown, the industrial robot reducer fault diagnosis method based on VMD-ELM includes:

[0060] Step 1: Data acquisition: Install an acceleration vibration sensor at the measured position of the reducer, and collect data through a data acquisition card. Collect different state data of the reducer, including the following state category data: normal fault-free state data, pitting fault data, gear fracture fault data, and component excessive wear fault data.

[0061] Step 2: Feature extraction: Extract features from the collected vibration signals through the VMD algorithm

[0062] Perform VMD feature extraction on the collected vibration signals in different states respectively, and obtain feature vectors in different states respectively, including:

[0063] a. Establish a constrained variational model:

[0064]

[0065] where u k is the k-th modal component of VMD decomposition, ω k is the center frequency of the component signal, denotes Derive with respect to t, where δ(t) is the Dirac function and * is the convolution, is the exponential term, and f is the input feature data;

[0066] b. Introduce the quadratic penalty term α and the Lagrange multiplier λ to obtain the augmented Lagrangian formula, and use the alternating direction method of multipliers to solve the extremum of the augmented Lagrangian expression. Then, decompose the original signal into k modal components, and the modal components are the VMD features to be extracted. The steps are as follows:

[0067] b1. Initialize the IMF parameters ω k , u k , α, λ, and n; where u k is the modal component, ω k is the center frequency, λ is the Lagrangian algorithm factor to ensure the strictness of the constraint condition, α is the penalty factor to ensure the reconstruction accuracy of the signal, n is the number of iterations, and k is the number of modal decompositions.

[0068] b2. Update the center frequency ω k and the modal component u k ;

[0069]

[0070]

[0071] where, is the Fourier transform of f(ω), is the Fourier transform of λ(ω), is the Wiener filter of the current residual, is the centroid of the power spectrum of the modal function at the (n + 1)-th iteration. ω is the center frequency, and ω k is the center frequency of the k-th modal component. The modal component is the VMD feature to be extracted;

[0072] b3. Update λ according to the formula;

[0073]

[0074] where τ is the tolerance parameter of the noise.

[0075] b4. Set the loop condition. When the following conditions are met, the loop ends; otherwise, return to b2, and finally obtain k IMF components.

[0076]

[0077] Step 3: Establish an extreme learning machine model. The main algorithm process is as follows:

[0078] Input N training sample data sets (x, y) = {(xi , y i ) | x i ∈ R n , y i ∈ R m , i = 1, 2, ..., N}, where x i = (x i1 , x i2 , ..., x in ) T is the i-th input vector, y i = (y i1 , y i2 , ..., y in ) T is the true label corresponding to the i-th input vector. During the training process, the predicted output

[0079]

[0080] In the formula, is the predicted label corresponding to the i-th input vector, ω j is the weight vector of the j-th hidden layer node, b j is the bias of the j-th hidden layer node, β j is the weight vector connecting the j-th hidden layer node and the output layer node. g(·) is the activation function.

[0081] The above formula, converted into matrix form, is:[[]]

[0082]

[0083] In the formula, is the predicted output matrix, H is the hidden layer output matrix, and β is the output weight matrix. The specific expressions are:[[]]

[0084]

[0085]

[0086] Minimizing the error E between the predicted label and the true label is:[[]]

[0087]

[0088] Solving the least squares solution of equation (6):[[]]

[0089] β = H + Y (10)

[0090] where H + is the Moore-Penrose generalized inverse of the matrix. Introducing the regularization parameter C into the above formula to provide a more applicable solution:[[]]

[0091]

[0092] Wherein, I is the identity matrix.

[0093] The final model for obtaining the predicted output is:

[0094]

[0095] Training is to input the paired data set {state feature data, state label} into the final model, set the training parameters, and perform iterative training using the Adam optimizer. The training stops when the network loss converges, reaches the number of training times, or the predicted label meets the accuracy requirement.

[0096] Step 4: Fault diagnosis: Input the test data into the fault diagnosis model to obtain the fault diagnosis result.

[0097] The above are only the preferred embodiments of the present invention, and do not impose any limitations on the present invention. Any simple modifications, changes, and equivalent structural changes made to the above embodiments according to the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A fault diagnosis method for an industrial robot reducer based on VMD-ELM, characterized in that, It includes the following steps: Under different working conditions, collect the vibration signals of the industrial robot reducer; perform VMD feature extraction respectively to obtain the VMD features in different states, and make a paired data set; Based on the extreme learning machine, train the extracted paired data set to obtain a fault diagnosis model; After the feature extraction of the on-site collected test sample data, input it into the trained fault diagnosis model to obtain the fault state diagnosis result of the industrial robot reducer.

2. The fault diagnosis method of the industrial robot reducer based on VMD-ELM according to claim 1, wherein, The vibration data is collected by a vibration sensor installed on the reducer.

3. The fault diagnosis method for the reducer of an industrial robot based on VMD-ELM according to claim 1, wherein, The different states or fault labels include: normal without fault, pitting, gear fracture, and excessive wear of parts.

4. The fault diagnosis method for the industrial robot reducer based on VMD-ELM according to claim 1, wherein, The making of the paired data set is {state feature data, state label}, and it is necessary to pre-distinguish the IMF components in different states into different state labels by experts.

5. The fault diagnosis method for an industrial robot reducer based on VMD-ELM according to claim 1, characterized in that, Perform VMD feature extraction on the vibration signals in different states collected respectively, and obtain the feature vectors in different states respectively, including: a. Establish a constrained variational model: where, u k is the k-th modal component of VMD decomposition, ω k is the center frequency of the component signal, denotes derivative with respect to t, δ(t) is the Dirac function, * is convolution, is the exponential term, and f is the input feature data; b. Introduce the quadratic penalty term α and the Lagrange multiplier λ to obtain the augmented Lagrangian formula, and use the alternating direction method of multipliers to solve the extreme value of the augmented Lagrangian expression, and then decompose the original signal into k modal components, and the modal components are the VMD features to be extracted.

6. The fault diagnosis method for the industrial robot reducer based on VMD-ELM according to claim 5, wherein, The solution process is as follows: b1. Initialize the IMF parameters ω k , u k , α, λ, and n; where u k is the modal component, ω k is the center frequency, λ is the Lagrange algorithm factor, α is the penalty factor, n is the number of cycles, and k is the number of modal decompositions; b2. Update the center frequency ω according to the formula k and the modal component u k ; Among them, is the Wiener filtering of the current remaining amount, is the centroid of the power spectrum of the mode function at the (n + 1)-th iteration, ω is the center frequency, ω k is the center frequency of the k-th mode component; b3. Update λ according to the formula; where τ is the fidelity coefficient; b4. Set the loop condition. When the following conditions are met, the loop ends. Otherwise, return to step (2) to finally obtain k IMF components; 7. A fault diagnosis method for an industrial robot reducer based on VMD-ELM, characterized in that, Construct an extreme learning machine and output the fault result. The steps are as follows: Input N training paired data sets \((x, y)=\{(x i , y i )|x i \in\mathbb{R} n , y i \in\mathbb{R} m , i = 1, 2, \ldots, N\}\), where \(x i =(x i1 , x i2 , \ldots, x in ) T is the \(i\)-th input vector, and \(y i =(y i1 , y i2 , \ldots, y in ) T is the true label corresponding to the \(i\)-th input vector; The predicted output is obtained during the training process: Wherein, is the predicted label corresponding to the i-th input vector, ω j is the weight vector of the j-th hidden layer node, b j is the bias of the j-th hidden layer node, β j is the weight vector connecting the j-th hidden layer node and the output layer node. g(·) is the activation function. The above formula, when converted into matrix form, is: In the formula, is the predicted output matrix, H is the hidden layer output matrix, and β is the output weight matrix; the specific expressions are as follows: Minimize the error E between the predicted label and the true label as: Solve the least squares solution of the above formula: β = H + Y where H + is the Moore-Penrose generalized inverse of the matrix; introducing a regularization parameter C into the above equation provides a more applicable solution: In the formula, I is the identity matrix. The finally obtained predicted output model is:

8. The fault diagnosis method for an industrial robot reducer based on VMD-ELM according to claim 1, wherein, The training is to input the paired data set {state feature data, state label} into the final model, set the training parameters, and use the Adam optimizer for iterative training. When the network loss converges or reaches the training times or the predicted label reaches the accuracy requirement, the training stops.

9. The fault diagnosis method for the industrial robot reducer based on VMD-ELM according to claim 1, wherein The fault state diagnosis result is the fault label in different states.

Images