A neural network-based method for calculating damage to a raceway of a constant velocity joint
By using a neural network-based method, road spectrum data was collected from actual vehicles and preprocessed and model fitted. This solved the problem of not being able to accurately determine the damage of constant velocity universal joints in the early stages of drive shaft development, and enabled accurate damage prediction in the early stages of drive shaft development, ensuring that the universal joint design meets the life requirements.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HE SHAN SHI JIE SHI KE QI CHE PEI JIAN YOU XIAN GONG SI
- Filing Date
- 2023-02-28
- Publication Date
- 2026-06-19
AI Technical Summary
Existing technologies cannot accurately determine the damage of constant velocity universal joints under different load spectra in the early stages of drive shaft development, which makes it impossible to guarantee the life requirements of the universal joint.
A neural network-based approach was adopted to establish a relationship model between torque, speed, working angle and number of revolutions by collecting road spectrum data from real vehicles, preprocessing, joint distribution transformation, experimentation and neural network fitting, and calculating the raceway damage of constant velocity universal joints.
Accurately predicting raceway damage of constant velocity universal joints of different specifications in the early stages of drive shaft development ensures that the damage of the universal joints under the original load spectrum meets the life requirements, thus improving the accuracy and efficiency of universal joint design.
Smart Images
Figure CN116257938B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of automotive transmission system components, and specifically relates to a method for calculating the raceway damage of constant velocity universal joints based on neural networks. Background Technology
[0002] Drive shaft suppliers convert raw load spectra collected from typical road conditions for automotive durability provided by OEMs into block load spectra suitable for indoor bench testing. Suppliers use these block load spectra to reproduce the actual road load conditions of the drive shaft system in indoor bench testing. However, this method can only be used for testing and verification of manufactured drive shafts and cannot guarantee that the product meets the OEM's lifespan requirements in the early stages of drive shaft development. Currently, drive shaft suppliers determine constant velocity joint (CV joint) specifications based on strength, lacking methods for calculating CV joint lifespan. Furthermore, relying solely on strength to determine CV joint specifications may not meet the OEM's lifespan requirements. Therefore, drive shaft suppliers urgently need a method to determine the damage condition of CV joints under a specific load spectrum in the early stages of drive shaft development. Summary of the Invention
[0003] To address the technical shortcomings of existing technologies that fail to determine the damage of universal joints of different specifications under different load spectra in the early stages of drive shaft development, thus hindering the determination of the required drive shaft specifications, this invention provides a neural network-based method for calculating the raceway damage of constant velocity universal joints. This method can predict the raceway damage of constant velocity universal joints of different specifications based on the original load spectrum of the drive shaft, enabling the determination of the damage status of constant velocity universal joints under the original load spectrum in the early stages of drive shaft development.
[0004] To achieve the objectives of this invention, the present invention provides a method for calculating raceway damage of a constant velocity universal joint based on a neural network, comprising the following steps:
[0005] ① Collect road spectrum data of typical working conditions of automobiles on durable road surfaces using actual vehicles to obtain the original load spectrum including drive shaft torque, speed, and working angle.
[0006] ② Preprocess the original load spectrum;
[0007] ③ The torque, speed and working angle of the drive shaft are jointly distributed. Based on the distribution of the original load spectrum in different load ranges, the original load spectrum is transformed into a BLOCK load spectrum that can be used for indoor bench testing of the drive shaft. The BLOCK load spectrum includes multiple different small BLOCK loads. Each small BLOCK load includes torque T, speed R, working angle θ and the number of rotations n required.
[0008] ④ Determine several experimental conditions with different torques, speeds, and working angles, and conduct tests on each condition to obtain the number of rotations of the constant velocity universal joint when it reaches a certain level of damage under each condition.
[0009] ⑤ Perform neural network fitting on the experimental data from step ④ to obtain a neural network model with the inputs being torque T, speed R, and working angle θ, and the output being the number of revolutions N;
[0010] ⑥ Using the torque, speed, and working angle of each small block as input, and employing a neural network model to obtain the fitted number of revolutions N for each small block, the damage calculation formula for the actual number of revolutions n required for each small block is as follows:
[0011]
[0012] ⑦ Based on the linear Miner damage theory, the formula for calculating the total damage of the constant velocity universal joint raceway under all blocks is as follows:
[0013]
[0014] In the formula, the subscripts i = 1, 2, ..., k represent the BLOCK number.
[0015] In a further improvement to the present invention, step ② includes resampling, filtering, and singular value removal.
[0016] A further improvement to the present invention is to use a drive shaft life testing machine to test various working conditions.
[0017] A further improvement to the present invention is to select the test conditions in step ④ within a wider range, select multiple conditions with different torques, speeds, and working angles, and conduct tests using a drive shaft life testing machine to record the number of rotations when the drive shaft has serious NVH problems.
[0018] A further improvement to the present invention is that the neural network fitting includes neural network training and neural network testing. The experimental data is divided into two parts: training data and test data. The neural network is trained based on the training data. The neural network training includes two parts: information feed-forward and error backpropagation. Based on the experimental data obtained in step ④, the drive shaft torque, speed, and working angle are used as inputs, and the number of revolutions for each working condition is used as output to build a neural network structure for the life revolutions of the constant velocity universal joint raceway.
[0019] A further improvement to the present invention is that the neural network includes an input layer, a hidden layer, and an output layer, and the input samples of the neural network are X = [X1, X2, ..., X...]. n The output of the hidden layer is transformed into H = [H1, H2, ..., H].s The result is then calculated by the output layer as O = [O1, O2, ..., O]. m The output formula for the hidden layer neurons is:
[0020]
[0021] The output formula for the output layer neuron is:
[0022]
[0023] In equations (3) and (4), f(·) and g(·) are the transfer functions of the hidden layer and the output layer, respectively, and q j and q k The thresholds w for hidden layer node j and output layer node k are respectively. ji The connection weights between hidden layer node j and input layer node i are given by v. kj The connection weights between output layer node k and hidden layer node j are given.
[0024] In a further improvement to the present invention, the hidden layer uses the sigmoid transfer function tansig, and the output layer uses the linear transfer function purelin.
[0025] A further improvement to this invention is that, when training the neural network model, the weights and the neural network error target are first initialized, then the input samples and related network parameters are applied, and the expected value of the output layer is d = [d1, d2, ..., d]. m When the error between the output value and the expected value of the output layer does not meet the accuracy requirements, the error propagates backward from the output layer to the hidden layer and the input layer. The weights and thresholds are adjusted layer by layer, and the error is recalculated after each adjustment. This process is repeated until the error is reduced to the required accuracy range. The error E between the output value and the expected value is:
[0026]
[0027] A further improvement to the present invention is that adjusting the weights and thresholds based on the error includes: calculating the weights w between the input layer and the hidden layer neurons based on the error E, respectively. ji The partial derivatives of the error E and the weights v between the output layer and hidden layer neurons kj The partial derivatives of the error E with respect to the threshold q of the hidden layer node j j The partial derivatives of the error E with respect to the threshold q of the output layer node k k The partial derivatives are given by the formula:
[0028]
[0029] The formula for adjusting the weights and thresholds in each iteration is:
[0030]
[0031] (7) In the formula: t is the number of iterations, and L1 and L2 are the learning efficiencies of the hidden layer and the output layer, respectively.
[0032] Under a given neural network structure, adjusting the weights and thresholds can meet the error requirements for neural network training. However, the trained neural network structure may not be able to achieve the required accuracy under test data. In this case, it is necessary to increase the number of hidden layers in the neural network and repeat all the above steps until all accuracy requirements are met.
[0033] Compared with the prior art, the present invention has at least the following beneficial effects:
[0034] 1) In the early stage of drive shaft development, the method of this invention can be used to determine the damage of constant velocity universal joints of different specifications under the original load spectrum, and this method can be used to determine the universal joint specifications that meet the life requirements.
[0035] 2) The neural network fitting method provided by this invention is simpler than the polynomial fitting method;
[0036] 3) For components like drive shafts whose lifespan is affected by multiple factors, the neural network provided by this invention can fully approximate the complex nonlinear relationship between influencing factors and lifespan, resulting in more accurate results. Attached Figure Description
[0037] To more clearly illustrate the technical solutions in the embodiments of the invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:
[0038] Figure 1 This is a flowchart of a method for calculating raceway damage of a constant velocity universal joint based on a neural network, provided in an embodiment of the present invention.
[0039] Figure 2 This is a flowchart of the training process of the neural network in an embodiment of the present invention.
[0040] Figure 3 This is a flowchart of the neural network fitting process in an embodiment of the present invention.
[0041] Figure 4 is a schematic diagram of the original load spectrum in an embodiment of the present invention. Figure 4(a) is a schematic diagram of the original load spectrum of the drive shaft torque, Figure 4(b) is a schematic diagram of the original load spectrum of the drive shaft speed, and Figure 4(c) is a schematic diagram of the original load spectrum of the drive shaft working angle. Detailed Implementation
[0042] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0043] like Figure 1 As shown, the present invention provides a method for calculating raceway damage of a constant velocity universal joint based on a neural network, comprising the following steps:
[0044] Step 1: Collect road spectrum data of typical working conditions on durable road surfaces using actual vehicles to obtain the original load spectrum including drive shaft torque, speed, and working angle.
[0045] In some embodiments of the present invention, the original load spectra of the obtained drive shaft torque, drive shaft speed, and drive shaft working angle are shown in Figures 4(a), 4(b), and 4(c), respectively.
[0046] Step 2: Preprocess the original load spectrum.
[0047] In some embodiments of the present invention, the preprocessing includes resampling, filtering, and singular value removal.
[0048] Step 3: Combine the torque, speed, and working angle of the drive shaft. Based on the distribution of the original load spectrum in different load ranges, convert the original load spectrum into a BLOCK load spectrum that can be used for indoor bench testing of the drive shaft.
[0049] The BLOCK load spectrum includes multiple different small BLOCK loads, each of which includes torque T, rotational speed R, working angle θ, and the number of rotations n required.
[0050] In some embodiments of the present invention, the converted BLOCK load spectrum for indoor bench testing of drive shafts is shown in the table below.
[0051] Table BLOCK Load Spectrum
[0052] serial number Torque (Nm) Rotational speed (rpm) Working angle (°) Number of cycles 1 <![CDATA[T1]]> <![CDATA[R1]]> <![CDATA[θ1]]> <![CDATA[n1]]> 2 <![CDATA[T2]]> <![CDATA[R2]]> <![CDATA[θ2]]> <![CDATA[n2]]> … … … … … k <![CDATA[T k ]]> <![CDATA[R k ]]> <![CDATA[θ k ]]> <![CDATA[n k ]]>
[0053] Step 4: Determine several experimental conditions with different torques, speeds, and working angles, and use a drive shaft life testing machine to test each condition to obtain the number of rotations of the constant velocity universal joint under each different condition when it reaches a certain level of damage (such as when serious NVH problems occur).
[0054] In some embodiments of the present invention, the test conditions in step ④ are selected within a relatively broad range (such as the common operating conditions of drive shafts), and multiple operating conditions with different torques, speeds, and working angles are selected. The drive shaft life tester is used to conduct the test and record the number of rotations when the drive shaft has serious NVH problems.
[0055] Step 5: Fit the existing experimental data to a neural network to obtain a neural network model with the inputs being torque T, speed R, and working angle θ, and the output being the number of revolutions N.
[0056] In this step, the neural network fitting uses the original life data from the drive shaft life test obtained in step ④. It takes drive shaft torque, speed, and operating angle as inputs, and the number of revolutions for each operating condition as the output, to build a neural network structure for the constant velocity universal joint raceway life revolutions. Neural network fitting includes neural network training and testing. The neural network model data is divided into training data and test data. The neural network is trained based on the training data, which includes two parts: information feedforward and error backpropagation.
[0057] First, initialize the weights and the neural network error target. The neural network consists of an input layer, hidden layers, and an output layer. The number of nodes in the input and output layers is determined by the raw lifetime input data. The number of nodes in the hidden layer is initialized to 10. The hidden layer uses the sigmoid transfer function (tansig), and the output layer uses the linear transfer function (purelin). The neural network input samples are X = [X1, X2, ..., X...]. n ], X n For data of a certain number of nodes entering the layer, the output of the hidden layer is transformed into H = [H1, H2, ..., H]. s ], H s The data for a certain number of nodes in the hidden layer is then processed by the output layer to form O = [O1, O2, ..., O2]. m ], O m To output data for a certain number of nodes in the output layer, the output formula for the hidden layer neurons is:
[0058]
[0059] The output formula for the output layer neuron is:
[0060]
[0061] In the formula, H j Let f(·) and g(·) be the outputs of the hidden layer neurons, respectively, and let q be the transfer functions of the hidden and output layers. j and q k The thresholds w for hidden layer node j and output layer node k are respectively. jiThe connection weights between hidden layer node j and input layer node i are given by v. kj The connection weights between output layer node k and hidden layer node j are O k This refers to the output of neurons in the output layer.
[0062] The expected value of the output layer is d = [d1, d2, ..., d]. m ], d m The data represents the expected number of nodes in the hidden layer, which is equal to the number of nodes in the output layer. When the error between the output value and the expected value does not meet the accuracy requirements, backpropagation of the error is performed. The error propagates from the output layer back to the hidden layer and the input layer, adjusting the weights and thresholds layer by layer. After adjustment, the error is recalculated. This process is repeated until the error is reduced to the required accuracy range. The error E between the output value and the expected value is:
[0063]
[0064] In some embodiments of the present invention, adjusting weights and thresholds based on error includes: calculating the weights w between the input layer and hidden layer neurons respectively, based on the error E. ji The partial derivatives of the error E and the weights v between the output layer and hidden layer neurons kj The partial derivatives of the error E with respect to the threshold q of the hidden layer node j j The partial derivatives of the error E with respect to the threshold q of the output layer node k k The partial derivatives are given by the formula:
[0065]
[0066] The formula for adjusting the weights and thresholds in each iteration is:
[0067]
[0068] In the formula: t is the number of iterations, and L1 and L2 are the learning efficiencies of the hidden layer and the output layer, respectively.
[0069] like Figure 3 As shown, under a given neural network structure, the error requirements for neural network training can be met after adjusting the weights and thresholds. If the trained neural network structure cannot meet the required accuracy under the test data, it is necessary to increase the number of hidden layer nodes of the neural network, that is, change the neural network model structure, and repeat all the above steps until all accuracy requirements are met.
[0070] Step 6: Using the torque, speed, and working angle of each block as input, the neural network model is used to obtain the fitted number of revolutions N for each block. The damage calculation formula for the actual number of revolutions n required for each block is as follows:
[0071]
[0072] Step 7: Based on the linear Miner damage theory, the formula for calculating the total damage of the constant velocity universal joint raceway under all blocks is as follows:
[0073]
[0074] (2) In the formula, the subscript i = 1, 2, ..., k represents the BLOCK number;
[0075] In some embodiments of the present invention, a storage medium is provided, which may be a ROM, RAM, disk, optical disk, or other storage medium. The storage medium stores one or more programs, which, when executed by a processor, implement the neural network-based constant velocity universal joint raceway damage calculation method provided in the above embodiments.
[0076] In some embodiments of the present invention, a device is provided, which may be a desktop computer, a laptop computer, a smartphone, a PDA handheld terminal, a tablet computer, or other terminal devices with display functions. The computing device includes a processor and a memory. The memory stores one or more programs. When the processor executes the program stored in the memory, it implements the neural network-based constant velocity universal joint raceway damage calculation method provided in the above embodiments.
[0077] The constant velocity joint raceway damage calculation method provided in the foregoing embodiments of the present invention inputs the original load spectrum of the drive shaft and preprocesses it. It then converts the load spectrum into a block load spectrum that can be used for indoor bench testing of the drive shaft through joint distribution. Tests are conducted under different working conditions (different torque, speed, and working angle) to obtain the number of rotations required to reach damage. The torque, speed, and working angle of different working conditions are input into a neural network model to obtain the fitted number of rotations N for each block, thus obtaining the damage under each block. Finally, the total damage of the constant velocity joint raceway under all blocks is obtained. This method can predict the raceway damage of constant velocity joints of different specifications based on the original load spectrum of the drive shaft, so as to determine the damage situation of the constant velocity joint under the original load spectrum in the early stage of drive shaft development.
[0078] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0079] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for calculating raceway damage in a constant velocity universal joint based on a neural network, characterized in that, Includes the following steps: ① Collect road spectrum data of typical working conditions of automobile durable road surface from actual vehicles to obtain the original load spectrum of drive shaft torque, drive shaft speed and drive shaft working angle; ② Preprocess the original load spectrum; ③ The torque, speed, and operating angle of the drive shaft are jointly distributed. Based on the distribution of the original load spectrum in different load ranges, the original load spectrum is transformed into a block load spectrum suitable for indoor bench testing of the drive shaft. The block load spectrum includes multiple different small block loads, each of which includes torque. Rotation speed From a working perspective And the number of rotations required ; ④ Determine several experimental conditions with different torques, speeds, and working angles, and conduct tests on each condition to obtain the number of rotations of the constant velocity universal joint when it reaches a certain level of damage under each condition. Select several conditions with different torques, speeds, and working angles, and conduct tests using a drive shaft life testing machine to record the number of rotations when the drive shaft experiences severe NVH problems. ⑤ Perform neural network fitting on the experimental data from step ④ to obtain the inputs as torque. Rotation speed From a working perspective The output is the number of laps. Neural network model; ⑥ Using the torque, speed, and working angle of each small block as input, a neural network model is used to obtain the number of fitted revolutions for each small block. Then the actual number of laps required under each small block. The formula for calculating the damage is: ⑦ Based on the linear Miner damage theory, the formula for calculating the total damage of the constant velocity universal joint raceway under all blocks is as follows: In the formula, the subscript , indicating the BLOCK number.
2. The method for calculating raceway damage of a constant velocity universal joint based on a neural network according to claim 1, characterized in that: In step ②, the preprocessing includes resampling, filtering, and singular value removal.
3. The method for calculating raceway damage of a constant velocity universal joint based on a neural network according to claim 1, characterized in that: In step ④, the drive shaft life testing machine is used to test various working conditions.
4. The method for calculating raceway damage of a constant velocity universal joint based on a neural network according to claim 1, characterized in that, The neural network fitting in step ⑤ includes neural network training and neural network testing. The experimental data is divided into two parts: training data and test data. The neural network is trained based on the training data. The neural network training includes two parts: information feed-forward and error backpropagation. Based on the experimental data obtained in step ④, the drive shaft torque, speed, and working angle are used as inputs, and the number of revolutions for each working condition is used as output to build a neural network structure for the life revolutions of the constant velocity universal joint raceway.
5. The method for calculating raceway damage of a constant velocity universal joint based on a neural network according to claim 1, characterized in that, The neural network model includes an input layer, hidden layers, and an output layer. The neural network input samples... , Data with a certain number of nodes in the input layer is transformed into the output of the hidden layer. , The data for a certain number of nodes in the hidden layer is then processed by the output layer to obtain the result. , To output data for a certain number of nodes in the output layer, the output formula for the hidden layer neurons is: The output formula of the output layer neuron is: In the formula, and These are the transfer functions for the hidden layer and the output layer, respectively. and Hidden layer nodes and output layer nodes The threshold, Hidden layer nodes With input layer nodes Connection weights, For output layer nodes With hidden layer nodes Connection weights.
6. The method for calculating raceway damage of a constant velocity universal joint based on a neural network according to claim 5, characterized in that, The hidden layers use the sigmoid transfer function tansig, and the output layer uses the linear transfer function purelin.
7. The method for calculating raceway damage of a constant velocity universal joint based on a neural network according to claim 5, characterized in that, When training a neural network model, the weights and the target error of the neural network are initialized first, then the input samples and relevant network parameters are given, and the expected value of the output layer is... , For the data representing the expected number of nodes in the hidden layer, when the error between the output value and the expected value in the output layer does not meet the accuracy requirements, backpropagation of the error is performed. The error propagates from the output layer back to the hidden layer and the input layer, adjusting the weights and thresholds layer by layer. After adjustment, the error is recalculated. This process is repeated until the error is reduced to within the required accuracy range, and the error between the output value and the expected value is minimized. for:
8. The method for calculating raceway damage of a constant velocity universal joint based on a neural network as described in claim 7, characterized in that, The step of adjusting weights and thresholds based on error includes: calculating the error respectively. Weights between neurons in the input layer and hidden layer partial derivatives, error Weights between neurons in the output layer and hidden layer partial derivatives, error For hidden layer nodes threshold partial derivatives, error For output layer nodes threshold The partial derivatives are given by the formula: The formula for adjusting the weights and thresholds in each iteration is: In the formula, For the number of iterations, , These represent the learning efficiency of the hidden layer and the output layer, respectively.
9. The method for calculating raceway damage of a constant velocity universal joint based on a neural network according to any one of claims 5-8, characterized in that, Under a given neural network structure, the error requirement for training the neural network can be met after adjusting the weights and thresholds. If the trained neural network structure cannot meet the required accuracy under the test data, the number of hidden layers of the neural network is increased and the training is repeated.