Gear vibration noise estimation method based on deep belief network

Abstract

The application discloses a gear vibration noise prediction method based on a deep belief network, which comprises the following steps: firstly, denoising the gear vibration noise data collected through experiments; secondly, constructing a database based on a frequency analysis algorithm; thirdly, constructing a deep belief network model based on a restricted Boltzmann machine and training the model; fourthly, obtaining optimal parameters through a particle swarm optimization algorithm based on adaptive inertia weight and a learning factor; and finally, using the denoised and amplified gear vibration data to predict the noise by using the APSO-DBN method. The method amplifies the data through frequency analysis, overcomes the training difficulty caused by insufficient data, and has better iteration effect than the PSO-DBN algorithm. In the total sound pressure level prediction, the average relative error of the APSO-DBN algorithm is obviously reduced compared with the DBN network with manually selected network nodes.

Description

Technical Field

[0001] This invention relates to a noise prediction method, specifically a gear vibration noise prediction method based on deep belief networks. Background Technology

[0002] Transmission equipment often operates under harsh conditions such as high speed and high pressure for a long time. Precision parts in the equipment, especially gears, are very likely to wear out, causing the entire equipment to fail and resulting in serious losses of manpower, financial resources and materials.

[0003] In the quantitative and qualitative assessment of gear performance degradation, noise can interfere with the identification of vibration signals. The noise of the gear transmission system is caused by the vibration behavior coupled by the time-varying meshing stiffness of the gear pair, tooth surface error, clearance, transmission error, etc. The noise level is also a key indicator of the equipment's operating status.

[0004] Noise prediction, a common noise assessment method, estimates the noise level of a gearbox based on the load on the gear transmission system and external environmental conditions. Currently, most noise prediction methods are related to the finite element method, focusing on the internal and external excitations and structure of the gear transmission system. However, these methods involve large computational loads, poor real-time prediction results, and low utilization of collected data. Therefore, how to fully utilize existing data to predict noise has become a major challenge in health monitoring. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a gear vibration noise prediction method based on deep belief networks (DBNs) to predict gear vibration noise. This method utilizes frequency band analysis to convert acceleration signals under specific speed-load conditions into acceleration excitations in a specific frequency band. Simultaneously, the sound pressure level from A-weighted frequency band analysis is used as a learning objective to construct a noise prediction method database. From the perspective of adaptive inertial weights, an adaptive function value for the current iteration number is introduced into the inertial weight configuration. Furthermore, combining the unsupervised deep learning capabilities of deep belief networks, an APSO-DBN-based noise prediction method is proposed.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A gear vibration noise prediction method based on deep belief networks includes the following steps:

[0008] S1: Perform noise reduction processing on gear vibration noise data;

[0009] S2: Construct a database based on frequency band analysis algorithms;

[0010] S3: Construct and train a deep belief network model based on a restricted Boltzmann machine;

[0011] S4: The optimal parameters are obtained by using the particle swarm optimization algorithm APSO based on adaptive inertia weights and learning factors;

[0012] S5: Noise prediction is performed using the denoised and amplified gear vibration data and the APSO-DBN-based method.

[0013] In step S1, the gear vibration noise data is denoised, specifically as follows:

[0014] A-weighting is applied to the gear vibration noise data. A-weighting is the inversion of the 40-squared isostatic response curve, which has a better effect on reducing low-frequency noise and is closer to the subjective perception of human hearing. The sound data directly measured by the noise sensor includes various background noises such as motor operation noise and oil circulation system noise. It is necessary to remove or reduce the background noise from the measured noise data to obtain the true noise generated by the gear under test. The formula for calculating the background noise correction value K1 is:

[0015]

[0016]

[0017] In the formula The mean of the weighted time-averaged sound pressure level of sensor A on the measurement surface during the operation of the sound source under test; ΔL is the mean of the A-weighted time-averaged sound pressure level of the sensor background noise on the measurement surface. p The difference between the average sound pressure level on the measurement surface and the mean value of ΔL is the difference between the average sound pressure level and the mean value of ΔL. p If the value is >15dB, then K1 is considered to be 0, and no background correction is needed; if 6dB≤ΔL p ≤15dB, should be corrected according to the formula; if ΔL p If the noise level is less than 6dB, further measures are needed to reduce background noise and ensure the validity of the measurement results.

[0018] The construction of the database based on the frequency band analysis algorithm in step S2 is specifically as follows:

[0019] For gear vibration and noise data with key characteristic parameters such as input speed, load, and gear health status, frequency band analysis is used to amplify sample features. This method can perform amplitude averaging within a certain frequency band, converting time-domain vibration and noise data into multi-band power and sound pressure values. By performing 1 / 3 octave analysis on gear surface noise signals and acceleration signals using frequency band analysis, 1 / 3 frequency bands of 9 bands are obtained, thereby achieving sample feature amplification.

[0020] The vibration acceleration data of the input and output shaft measuring points of the gear vibration test bench are converted into 9-dimensional acceleration excitation data through frequency band analysis. Similarly, the gear vibration noise data are also converted into 0-dimensional noise data through frequency band analysis. Finally, the total sound pressure data with A-weighted weights are added to form a database with a dimension of 28.

[0021] The specific steps for constructing and training a deep belief network model based on a restricted Boltzmann machine in step S3 are as follows:

[0022] S3.1: Iterative calculation of the Restricted Boltzmann Machine (RBM) network:

[0023] The training data is input into the model input node, and the hidden layer nodes are sampled according to equation (3).

[0024]

[0025] Among them, v i h j These are the states of visible layer neuron i and hidden layer neuron j, respectively. j It is the bias of hidden layer neuron j, w ij The connection weights between visible layer neuron i and hidden layer neuron j, where v represents the visible layer. Here, σ(x) is the sigmoid activation function, i.e., σ(x) = 1 / (1 + exp(-x)).

[0026] The input layer is sampled using equation (4), and the sampled value is then used as the input layer node value to repeat the previous step, completing one Gibbs sampling step;

[0027]

[0028] Among them, a i is the deviation of visible layer neuron i, and h represents the hidden layer;

[0029] Depending on the set value of k, repeat the previous two steps k times; substitute the sampled values ​​from the previous step into equations (5) to (7) to calculate the gradient and update the parameters:

[0030]

[0031]

[0032] Δb j =P(h j =1∣v (0) )-P(h j =1∣v (k) (7)

[0033] Update the training data and repeat the first four steps until the iteration requirement is met;

[0034] S3.2: Training a deep belief network:

[0035] A deep belief network consisting of a 4-layer Restricted Boltzmann Machine (RBM) and a 1-layer Backpropagation Neural Network (BPNN) is constructed. The denoised gear vibration data, which has been amplified by frequency band analysis, is input into the network model. The network model is then subjected to unsupervised forward greedy learning and supervised backward adjustment training.

[0036] In step S4, the optimal parameters are obtained using a particle swarm optimization algorithm based on adaptive inertia weights and learning factors. The specific steps are as follows:

[0037] S4.1: Parameter Initialization: Set the upper and lower limits of particle position change, velocity change, maximum number of iterations (ger), and learning factor (c). max c min inertial weight w max w min Randomly initialize the velocity and position of a single particle;

[0038] S4.2: Selection of fitness function: The selection of fitness function is as shown in equation (8). If the fitness value of the current particle is better than the individual extreme value, then Pbest is updated to the current particle position; the individual optimal value of all particles is compared. If it is better than the population optimal value, then Gbest is updated to the individual optimal position of the optimal particle.

[0039]

[0040] Where N is the number of particles, D is the particle dimension, and p ij t is the target value of the i-th particle in the j-th dimension. ij It is the actual value of the i-th particle in the j-th dimension;

[0041] S4.3: Inertia Weight and Learning Factor Update: The adaptive function value of the current particle is introduced to adjust the inertia weight, allowing the particle to randomly reset its position with a certain probability, thereby achieving the purpose of jumping out of the original position. The specific weight update formula is shown in Equation (9):

[0042]

[0043] Among them, w max w min These are the maximum and minimum values ​​of the inertia weight, f. max f min These are the maximum and minimum fitness values ​​of all particles in each iteration, respectively, and fitness(q) represents the fitness value of the q-th particle in the current iteration.

[0044] Learning factors c1 and c2 adjust the distribution of local and global search to a certain extent. Learning factors that are too large or too small are not conducive to optimization. Unlike setting a fixed learning factor, the formula for improving the learning factor through an asynchronous strategy is shown in equation (10):

[0045]

[0046] S4.4: Velocity and Position Update: The properties of particles within the swarm are updated using equations (11) and (12):

[0047] v id (t+1)=w i v id (t)+c1r1(t)(p id (t)-x id (t))+c2r2(t)(p gd (t)-x id (t)) (11)

[0048] x id (t+1)=x id (t)+v id (t+1) (12)

[0049] Where c1 and c2 are learning factors that tend towards individual and global optima, respectively, and r1 and r2 are random functions that adjust the particle towards individual and global optima, respectively, with values ​​ranging from 0 to 1. max c min These represent the maximum and minimum values ​​of the learning factor, respectively; t is the current iteration number, and ger is the maximum iteration number; v id Let x be the velocity of the i-th particle in the d-th dimension. id The position of the i-th particle in the d-th dimension, p id p represents the individual optimal value of the best position of the i-th particle in the d-th dimension during its search process. gd w represents the optimal value for the i-th group at its best position in the d-th dimension search process. i Let be the inertial weight of the i-th particle;

[0050] S4.5: Determine the termination condition for optimization: The iteration terminates when the change of the population extreme value meets the minimum boundary requirement or the maximum number of iterations initialized is reached; if the iteration terminates, the optimal position in the current round is the optimal position sought; otherwise, return to S4.2.

[0051] In step S5, the noise prediction method based on APSO-DBN treats the vector composed of the number of hidden nodes in each RBM network as the position attribute of the particle, and inputs it into the DBN training to complete the calculation of the fitness value. Based on the fitness value, the inertial weight and velocity of the particle are calculated, the particle position is updated, and the global optimum is found through iterative loops. The specific steps are as follows:

[0052] S5.1: Initialization of DBN network structure parameters and APSO parameters: Determine the particle dimension based on the number of RBM stacks, determine the range of inertial weights and learning factors, the maximum number of iterations and the total number of particles, and initialize the position and velocity attributes of particles based on the range of hidden layer nodes;

[0053] S5.2: Fitness calculation and optimal particle position calculation: First, feed in the divided training data, calculate the mean square error corresponding to the current particle position attribute and use it as the fitness value; Second, update the individual optimal Pbest and the global optimal Gbest of each particle; Finally, calculate the maximum and minimum values ​​of the fitness values ​​of all particles under the current iteration number as a reference for updating the inertia weight.

[0054] S5.3: Inertia weight and learning factor update: Update w and c1, c2 according to equations (9) and (10);

[0055] S5.4: Particle property update: Update the position and velocity of the particles using equations (11) and (12);

[0056] S5.5: Iteration Termination Judgment: When the group extreme value satisfies the minimum limit or reaches the maximum number of iterations, the iteration is terminated and the global optimal position is the optimal solution; otherwise, return to S5.2.

[0057] S5.6: Use the hidden layer node number vector obtained by the APSO algorithm as the structural parameters of the DBN network, and use the training data to complete the forward initialization and back propagation of the network. After the training requirements are met, the noise prediction model is constructed.

[0058] The significant features of this invention compared to the prior art are as follows:

[0059] (1) The data was expanded through frequency band analysis, overcoming the training difficulties caused by insufficient data. (2) The proposed APSO-DBN algorithm has better iterative effect than the PSO-DBN algorithm. (3) In terms of total sound pressure level prediction, the average relative error of the APSO-DBN algorithm is significantly reduced compared with the DBN network with manually selected network nodes. Attached Figure Description

[0060] Figure 1 This is a flowchart of the APSO-DBN noise prediction method of the present invention.

[0061] Figure 2 This is the deep belief network structure of the present invention. Detailed Implementation

[0062] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0063] like Figure 1 As shown, a gear vibration noise prediction method based on deep belief networks includes the following steps:

[0064] S1: Perform noise reduction processing on gear vibration noise data.

[0065] When using a microphone to test noise, the obtained data is the sound pressure level history of the measured environment. This data needs to be pre-weighted to obtain the noise level (dB) of the measured environment. A-weighting is applied to the gear vibration noise data. A-weighting is the inverted 40-square isostatic response curve, which has a better low-frequency noise reduction effect and is closer to the subjective perception of human hearing. The sound data directly measured by the noise sensor includes various background noises such as motor operating noise and oil circulation system noise. It is necessary to remove or reduce the background noise from the measured noise data to obtain the true noise generated by the gear under test. The formula for calculating the background noise correction value K1 is:

[0066]

[0067]

[0068] In the formula The mean of the weighted time-averaged sound pressure level of sensor A on the measurement surface during the operation of the sound source under test; ΔL is the mean of the A-weighted time-averaged sound pressure level of the sensor background noise on the measurement surface. p The difference between the average sound pressure level on the measurement surface and the mean value of ΔL is the difference between the average sound pressure level and the mean value of ΔL. p If the value is greater than 15dB, then K1 can be considered to be 0, and no background correction is needed. If 6dB ≤ ΔL p ≤15dB, should be corrected according to the formula. If ΔL p If the noise level is less than 6dB, further measures are needed to reduce background noise and ensure the validity of the measurement results.

[0069] S2: Construct a database based on frequency band analysis algorithms.

[0070] For gear vibration and noise data with key characteristic parameters such as input speed (Hz), load (N / m), and gear health status, relying solely on these three features to predict gearbox surface sound pressure levels suffers from insufficient feature set, weakened model learning ability, and decreased prediction accuracy. To address this, frequency band analysis was employed to augment sample features. This method performs amplitude averaging within a certain frequency band, converting time-domain vibration and noise data into multi-band power and sound pressure values. By performing 1 / 3 octave analysis on both the gear surface noise signal and acceleration signal using frequency band analysis, nine 1 / 3 frequency bands were obtained, thus achieving sample feature augmentation.

[0071] The vibration acceleration data of the input and output shaft measuring points of the gear vibration test bench can be converted into 9-dimensional acceleration excitation data through frequency band analysis. Similarly, the gear vibration noise data can be obtained as 0-dimensional noise data through frequency band analysis. Finally, the total sound pressure data with A-weighted weights are added to form a database with a dimension of 28.

[0072] S3: Construct a deep belief network model based on a restricted Boltzmann machine.

[0073] S3.1: Perform iterative calculations on the Restricted Boltzmann Machine (RBM) network.

[0074] The training data is input into the model input node, and the hidden layer nodes are sampled according to equation (3).

[0075]

[0076] Among them, v i h j These are the states of visible layer neuron i and hidden layer neuron j, respectively. j It is the bias of hidden layer neuron j, w ij The connection weights between visible layer neuron i and hidden layer neuron j, where v represents the visible layer, and σ(x) is the sigmoid activation function, i.e., σ(x) = 1 / (1+exp(-x)).

[0077] The input layer is sampled using equation (4), and the sampled value is then used as the input layer node value to repeat the previous step, completing one Gibbs sampling step.

[0078]

[0079] Among them, a i is the deviation of visible layer neuron i, and h represents the hidden layer;

[0080] Depending on the set value of k, repeat the previous two steps k times. Substitute the sampled values ​​from the previous step into equations (5) to (7) to calculate the gradient and update the parameters.

[0081]

[0082]

[0083] Δb j =P(h j =1∣v (0) )-P(h j =1∣v (k) (7)

[0084] Update the training data and repeat the first four steps until the iteration requirement is met.

[0085] S3.2: Train the deep belief network.

[0086] A deep belief network consisting of a 4-layer Restricted Boltzmann Machine (RBM) and a 1-layer Backpropagation (BP) neural network was constructed. Denoising and frequency band analysis-amplified gear vibration data were input into this network model. Unsupervised forward greedy learning and supervised backward adjustment training were then performed on the network model. The specific training process is as follows:

[0087] (1) Unsupervised forward greedy learning stage: First, adjacency network layers form multiple RBM networks, with the output of the lower RBM serving as the input to the upper RBM. Second, unlabeled input data is used, and the parameters θ of each layer are obtained through layer-by-layer training. The vector composed of the activation values ​​of the hidden neurons in the lower layer serves as the input to the upper RBM. Layer-by-layer training realizes the transformation from low-level features to compact high-level features. Finally, the output of the topmost RBM is connected to the BP network to output preliminary prediction data. In this stage, there is a problem that the optimal parameters of a single RBM output are inconsistent with the global optimal parameters. Therefore, inverse parameter correction is required.

[0088] (2) Supervised back adjustment stage: In this stage, the error between the initial prediction value and the label data in the previous stage is used as the optimization target. The error is propagated from top to bottom to adjust the network model parameters, thereby achieving the global optimum.

[0089] S4: The optimal parameters are obtained by using the particle swarm optimization algorithm (APSO) based on adaptive inertia weights and learning factors.

[0090] S4.1: Parameter Initialization: Set the upper and lower limits of particle position change, velocity change, maximum number of iterations (ger), and learning factor (c). max c min inertial weight w max w min The velocity and position of a single particle are randomly initialized.

[0091] S4.2: Selection of fitness function: The selection of fitness function is shown in equation (8). If the fitness value of the current particle is better than the individual extreme value, then Pbest is updated to the current particle position; the individual optimal value of all particles is compared. If it is better than the population optimal value, then Gbest is updated to the individual optimal position of the optimal particle.

[0092]

[0093] Where N is the number of particles, D is the particle dimension, and p ij t is the target value of the i-th particle in the j-th dimension. ij It is the actual value of the i-th particle in the j-th dimension.

[0094] S4.3: Inertia weight and learning factor update: The adaptive function value of the current particle is introduced to adjust the inertia weight, allowing the particle to randomly reset its position with a certain probability, thereby achieving the purpose of jumping out of the original position. The specific weight update formula is shown in Equation (9).

[0095]

[0096] Among them, w max w min These are the maximum and minimum values ​​of the inertia weight, f. max f min These are the maximum and minimum fitness values ​​of all particles in each iteration, respectively, and fitness(q) represents the fitness value of the q-th particle in the current iteration.

[0097] Learning factors c1 and c2 can adjust the distribution of local and global searches to a certain extent. Learning factors that are too large or too small are not conducive to optimization. Unlike setting a fixed learning factor, the formula for improving the learning factor through an asynchronous strategy is shown in Equation (10).

[0098]

[0099] S4.4: Velocity and position update: The properties of particles within the swarm are updated using equations (11) and (12).

[0100] v id (t+1)=w i v id (t)+c1r1(t)(p id (t)-x id (t))+c2r2(t)(p gd (t)-x id (t)) (11)

[0101] x id (t+1)=xid (t)+v id (t+1) (12)

[0102] Where c1 and c2 are learning factors that tend towards individual and global optima, respectively, and r1 and r2 are random functions that adjust the particle towards individual and global optima, respectively, with values ​​ranging from 0 to 1. max c min These represent the maximum and minimum values ​​of the learning factor, respectively; t is the current iteration number, and ger is the maximum iteration number; v id Let x be the velocity of the i-th particle in the d-th dimension. id The position of the i-th particle in the d-th dimension, p id p represents the individual optimal value of the best position of the i-th particle in the d-th dimension during its search process. gd w represents the optimal value for the i-th group at its best position in the d-th dimension search process. i Let be the inertial weight of the i-th particle;

[0103] S4.5: Termination Condition Judgment: The iteration terminates when the change in the population extremum meets the minimum boundary requirement or the maximum number of iterations initialized is reached. If the iteration terminates, the optimal position in the current round is the desired optimal position; otherwise, return to S4.2.

[0104] S5: Noise prediction is performed using the denoised and amplified gear vibration data and the APSO-DBN-based method.

[0105] The noise prediction method based on APSO-DBN treats the vector composed of the number of hidden nodes in each RBM network as the position attribute of the particle. It inputs this vector into the DBN training to complete the calculation of the fitness value, calculates the particle's inertial weight and velocity based on the fitness value, updates the particle position, and searches for the global optimum through iterative loops.

[0106] The specific steps of the APSO-DBN noise prediction method are as follows:

[0107] S5.1: Initialization of DBN network structure parameters and APSO parameters: Determine the particle dimension based on the number of RBM stacks, determine the range of inertial weights and learning factors, the maximum number of iterations and the total number of particles, and initialize the position and velocity attributes of particles based on the range of hidden layer nodes.

[0108] S5.2: Fitness Calculation and Particle Optimal Position Calculation: First, feed in the divided training data, calculate the mean square error corresponding to the current particle position attribute and use it as the fitness value; second, update the individual optimal Pbest and the global optimal Gbest of each particle; finally, calculate the maximum and minimum values ​​of the fitness values ​​of all particles under the current iteration number as a reference for updating the inertia weight.

[0109] S5.3: Inertia weight and learning factor update: Update w and c1, c2 according to equations (9) and (10).

[0110] S5.4: Particle property update: Update the position and velocity of the particles using equations (11) and (12).

[0111] S5.5: Iteration Termination Judgment: When the group extreme value satisfies the minimum limit or reaches the maximum number of iterations, the iteration is terminated and the global optimal position is the optimal solution; otherwise, return to S5.2.

[0112] S5.6: Use the hidden layer node number vector obtained by the APSO algorithm as the structural parameters of the DBN network, and use the training data to complete the forward initialization and back propagation of the network. After the training requirements are met, the noise prediction model is constructed.

[0113] The gear vibration noise prediction method based on deep belief networks described in the above embodiments first denoises the gear vibration noise data collected experimentally; secondly, it constructs a database based on a frequency band analysis algorithm; next, it constructs and trains a deep belief network model based on a restricted Boltzmann machine; then, it obtains the optimal parameters using a particle swarm optimization algorithm based on adaptive inertia weights and learning factors; finally, it uses the denoised and amplified gear vibration data to perform noise prediction using a method based on APSO-DBN. The method in the above embodiments amplifies the data through frequency band analysis, overcoming the training difficulties caused by insufficient data; the proposed APSO-DBN algorithm has better iterative performance than the PSO-DBN algorithm; and in total sound pressure level prediction, the APSO-DBN algorithm shows a significant reduction in average relative error compared to a DBN network with manually selected network nodes.

[0114] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made according to the purpose of the invention. Any changes, modifications, substitutions, combinations or simplifications made based on the spirit and principle of the technical solution of the present invention shall be equivalent substitutions. As long as they meet the purpose of the invention and do not deviate from the technical principle and inventive concept of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A method for predicting gear vibration noise based on deep belief networks, characterized in that, Includes the following steps: S1: Perform noise reduction processing on gear vibration noise data; S2: Construct a database based on frequency band analysis algorithms; S3: Construct and train a deep belief network model based on a restricted Boltzmann machine; S4: The optimal parameters are obtained by using the particle swarm optimization algorithm APSO based on adaptive inertia weights and learning factors; S5: Noise prediction is performed using the denoised and amplified gear vibration data and the APSO-DBN-based method. In step S5, the APSO-DBN-based noise prediction method treats the vector composed of the number of hidden nodes in each RBM network as the position attribute of the particle, and inputs it into the DBN training to complete the calculation of the fitness value. Based on the fitness value, the inertial weight and velocity of the particle are calculated, and the particle position is updated. The global optimum is found through iterative iteration. The specific steps are as follows: S5.1: Initialization of DBN network structure parameters and APSO parameters: Determine the particle dimension based on the number of RBM stacks, determine the range of inertial weights and learning factors, the maximum number of iterations and the total number of particles, and initialize the position and velocity attributes of particles based on the range of hidden layer nodes; S5.2: Fitness calculation and optimal particle position calculation: First, feed in the divided training data, calculate the mean square error corresponding to the current particle position attribute and use it as the fitness value; Second, update the individual optimal Pbest and the global optimal Gbest of each particle; Finally, calculate the maximum and minimum values ​​of the fitness values ​​of all particles under the current iteration number as a reference for updating the inertia weight. S5.3: Inertia weight and learning factor update: Update w and c1, c2 according to equations (9) and (10); S5.4: Particle property update: Update the position and velocity of the particles using equations (11) and (12); S5.5: Iteration Termination Judgment: When the group extreme value satisfies the minimum limit or reaches the maximum number of iterations, the iteration is terminated and the global optimal position is the optimal solution; otherwise, return to S5.

2. S5.6: Use the hidden layer node number vector obtained by the APSO algorithm as the structural parameters of the DBN network, and use the training data to complete the forward initialization and back propagation of the network. After the training requirements are met, the noise prediction model is constructed.

2. The gear vibration noise prediction method based on deep belief networks according to claim 1, characterized in that, In step S1, the gear vibration noise data is denoised, specifically as follows: A-weighting is applied to the gear vibration noise data. A-weighting is the inversion of the 40-squared response curve, which has a better effect on reducing low-frequency noise and is closer to the subjectivity of human hearing. The sound data directly measured by the noise sensor includes various background noises such as motor operation noise and oil circulation system noise. It is necessary to remove or reduce the background noise from the measured noise data to obtain the real noise generated by the gear under test. The formula for calculating the background noise correction value K1 is: In the formula The mean of the weighted time-averaged sound pressure level of sensor A on the measurement surface during the operation of the sound source under test; ΔL is the mean of the A-weighted time-averaged sound pressure level of the sensor background noise on the measurement surface. p The difference between the average sound pressure level on the measurement surface and the mean value of ΔL is the difference between the average sound pressure level and the mean value of ΔL. p If the value is greater than 15dB, then K1 is considered to be 0, and no background correction is required. If 6dB≤ΔL p ≤15dB, should be corrected according to the formula; if ΔL p If the noise level is less than 6dB, further measures are needed to reduce background noise and ensure the validity of the measurement results.

3. The gear vibration noise prediction method based on deep belief networks according to claim 1, characterized in that, The construction of the database based on the frequency band analysis algorithm in step S2 is specifically as follows: For gear vibration and noise data with key characteristic parameters such as input speed, load, and gear health status, frequency band analysis is used to amplify sample features. This method can perform amplitude averaging within a certain frequency band, converting time-domain vibration and noise data into multi-band power and sound pressure values. By performing 1 / 3 octave analysis on gear surface noise signals and acceleration signals using frequency band analysis, 1 / 3 frequency bands of 9 bands are obtained, thereby achieving sample feature amplification. The vibration acceleration data of the input and output shaft measuring points of the gear vibration test bench are converted into 9-dimensional acceleration excitation data through frequency band analysis. Similarly, the gear vibration noise data are also converted into 0-dimensional noise data through frequency band analysis. Finally, the total sound pressure data with A-weighted weights are added to form a database with a dimension of 28.

4. The gear vibration noise prediction method based on deep belief networks according to claim 1, characterized in that, The specific steps for constructing and training a deep belief network model based on a restricted Boltzmann machine in step S3 are as follows: S3.1: Iterative calculation of the Restricted Boltzmann Machine (RBM) network: The training data is input into the model input node, and the hidden layer nodes are sampled according to equation (3). Among them, v i h j These are the states of visible layer neuron i and hidden layer neuron j, respectively. j It is the bias of hidden layer neuron j, w ij The connection weights between visible layer neuron i and hidden layer neuron j, where v represents the visible layer, and σ(x) is the sigmoid activation function, i.e., σ(x) = 1 / (1+exp(-x)); The input layer is sampled using equation (4), and the sampled value is then used as the input layer node value to repeat the previous step, completing one Gibbs sampling step; Among them, a i is the deviation of visible layer neuron i, and h represents the hidden layer; Depending on the set value of k, repeat the previous two steps k times; substitute the sampled values ​​from the previous step into equations (5) to (7) to calculate the gradient and update the parameters: Δb j =P(h j =1∣v (0) )-P(h j =1∣v (k) ) (7) Update the training data and repeat the first four steps until the iteration requirement is met; S3.2: Training a deep belief network: A deep belief network consisting of a 4-layer Restricted Boltzmann Machine (RBM) and a 1-layer Backpropagation Neural Network (BPNN) is constructed. The denoised gear vibration data, which has been amplified by frequency band analysis, is input into the network model. The network model is then subjected to unsupervised forward greedy learning and supervised backward adjustment training.

5. The gear vibration noise prediction method based on deep belief networks according to claim 1, characterized in that, In step S4, the optimal parameters are obtained using a particle swarm optimization algorithm based on adaptive inertia weights and learning factors. The specific steps are as follows: S4.1: Parameter Initialization: Set the upper and lower limits of particle position change, velocity change, maximum number of iterations (ger), and learning factor (c). max c min inertial weight w max w min Randomly initialize the velocity and position of a single particle; S4.2: Selection of fitness function: The selection of fitness function is as shown in equation (8). If the fitness value of the current particle is better than the individual extreme value, then Pbest is updated to the current particle position; the individual optimal value of all particles is compared. If it is better than the population optimal value, then Gbest is updated to the individual optimal position of the optimal particle. Where N is the number of particles, D is the particle dimension, and p ij t is the target value of the i-th particle in the j-th dimension. ij It is the actual value of the i-th particle in the j-th dimension; S4.3: Inertia Weight and Learning Factor Update: The adaptive function value of the current particle is introduced to adjust the inertia weight, allowing the particle to randomly reset its position with a certain probability, thereby achieving the purpose of jumping out of the original position. The specific weight update formula is shown in Equation (9): Among them, w max w min These are the maximum and minimum values ​​of the inertia weight, f. max f min These are the maximum and minimum fitness values ​​of all particles in each iteration, respectively, and fitness(q) represents the fitness value of the q-th particle in the current iteration. Learning factors c1 and c2 adjust the distribution of local and global search to a certain extent. Learning factors that are too large or too small are not conducive to optimization. Unlike setting a fixed learning factor, the formula for improving the learning factor through an asynchronous strategy is shown in equation (10): S4.4: Velocity and Position Update: The properties of particles within the swarm are updated using equations (11) and (12): v id (t+1)=w i v id (t)+c1r1(t)(p id (t)-x id (t))+c2r2(t)(p gd (t)-x id (t)) (11) x id (t+1)=x id (t)+v id (t+1) (12) Where c1 and c2 are learning factors that tend towards individual and global optima, respectively, and r1 and r2 are random functions that adjust the particle towards individual and global optima, respectively, with values ​​ranging from 0 to 1; max c min These represent the maximum and minimum values ​​of the learning factor, respectively; t is the current iteration number, and ger is the maximum iteration number; v id Let x be the velocity of the i-th particle in the d-th dimension. id The position of the i-th particle in the d-th dimension, p id p represents the individual optimal value of the best position of the i-th particle in the d-th dimension during its search process. gd w represents the optimal value for the i-th group at its best position in the d-th dimension search process. i Let be the inertial weight of the i-th particle; S4.5: Determine the termination condition for optimization: The iteration terminates when the change of the population extreme value meets the minimum boundary requirement or the maximum number of iterations initialized is reached; if the iteration terminates, the optimal position in the current round is the optimal position sought; otherwise, return to S4.2.

Images