A diesel engine performance prediction method based on convolutional neural network
Through the diesel engine performance prediction method based on convolutional neural network, the problem of insufficient accuracy and efficiency in traditional models in multi-parameter processing is solved, and higher prediction accuracy and lower cost are achieved. It is suitable for diesel engines that handle multi-parameters Performance prediction.
Patent Information
- Application Number
- CN202010529178.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-06-11
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2040-06-11
AI Technical Summary
Traditional empirical models and experimental-based calibration methods are difficult to meet accuracy and efficiency requirements when dealing with multi-parameter diesel engine performance prediction, resulting in increased cost and time costs.
The performance prediction method based on convolutional neural network is adopted, and the signal processing is performed by providing 60 sets of experimental data, using convolutional layer and pooling layer, and the weight value is updated using mean square error function and BP algorithm to prevent overfitting and activate the function. The identity function and sigmoid function are used. The input parameters include 11 diesel engine performance indicators, and the output parameters include fuel consumption rate, maximum burst pressure, etc.
It improves the accuracy of diesel engine performance prediction, reduces costs, is more efficient than traditional empirical models and numerical simulation models, and can handle the needs of more parameters.
Smart Images

Figure CN113804446B_ABST
Abstract
Description
Technical Field
[0001] The present invention mainly relates to the field of diesel engine detection, and in particular to a diesel engine performance prediction method based on convolutional neural network. Background Art
[0002] At present, diesel engine model construction is mainly based on empirical models. However, with the development of technologies to improve diesel engine performance and emissions, the number of parameters that need to be considered in the model continues to increase. The difficulty of establishing traditional empirical models has greatly increased. It also requires a lot of time and the accuracy is difficult to meet the requirements.
[0003] Diesel engine calibration is based on experiments or models. However, as engines continue to develop, the number of parameters that need to be calibrated is also increasing. For calibration methods based on experiments, the cost is getting higher and higher. In order to reduce the workload, model-based virtual calibration technology is applied. Among them, the virtual model requires accurate simulation, which is the most critical for subsequent calibration. However, as the number of parameters to be calibrated increases, the traditional empirical model regression fitting method is increasingly difficult to meet the requirements of accuracy and speed.
[0004] The Chinese invention patent has been published, application number CN201810448090.9, patent name: Diesel generator set fault diagnosis and detection device and method based on deep learning, application date: 2018-05-11, the present invention relates to a diesel generator set fault diagnosis and detection device and method based on deep learning, comprising a frame 1, a speaker 2, a display 6, a memory 10, a CPU 11 and a data acquisition device 18, the frame 1 is configured to include a deep learning module 24, an adaptive integrated strategy module 20, a historical signal database 23 and a fault category expert system library 19, the adaptive integrated strategy module 20 is provided with an integrated strategy generator 201, the fault category expert system library 19 is provided with a fault category database 191, a fault indicator database 192, a fault mark database 193 and a fault level database 194, the deep learning module 24 also includes a clustering algorithm, and a signal transceiver 5 is provided at the middle position of the upper end of the frame 1, so people can perform fault diagnosis and online status monitoring of the diesel generator set more accurately and conveniently. Summary of the invention
[0005] The present invention provides a diesel engine performance prediction method based on a convolutional neural network. In view of the above-mentioned defects of the prior art, a diesel engine performance prediction method based on a convolutional neural network is provided, which includes the following steps:
[0006] Step S1: provide 60 groups of experimental data for the convolutional neural network, of which 40 groups are used as training samples and 20 groups are used as test samples;
[0007] Step S2: The input layer of the convolutional neural network first normalizes the input signal;
[0008] Step S3: The convolution layer of the convolutional neural network contains 32 neurons, each neuron contains a convolution kernel, the convolution kernel length is selected to be 3, which should be less than the vector length, and the step length is 1;
[0009] Step S4: The pooling method used by the pooling layer of the convolutional neural network is average pooling, with a length of 2 and a step size of 1;
[0010] Step S5: The fully connected layer of the convolutional neural network, the number of neurons in this layer is equal to the output, which is 6. There is no convolution in the fully connected layer, so the number of weights contained in each neuron is the number of output parameters after pooling, which is 10. Multiply each output of the previous layer by the corresponding weight W2, add an offset B to get the output. The final output also needs to pass through an activation function g(x). The main problem solved in this patent is a regression problem, so the identity function is used as the output activation function, that is, the input is equal to the output;
[0011] Step S6: The loss function of the convolutional neural network selects the mean square error function MSE as the loss function, and uses the BP back-propagation algorithm to back-propagate the error to update the weight value of each neuron. The timely termination method is used to prevent the neural network from overfitting.
[0012] Preferably, there are 11 input parameters, including speed, power, intake temperature and pressure, exhaust back pressure, cooling water temperature, injection timing and injection pressure, fuel temperature, ambient pressure and temperature. There are 6 output parameters, namely fuel consumption rate, maximum explosion pressure, explosion pressure angle, supercharger speed, turbine pre-exhaust temperature and smoke density.
[0013] Preferably, in step S2, the signal subjected to signal normalization processing is over-padding, and two zeros are added at the end by using a zero-padding method.
[0014] Preferably, in step S3, the convolution layer utilizes convolution so that each convolution kernel in the layer has a weight W1, a common offset b is set, and the activation function uses the sigmoid function as the activation function f(x).
[0015] Preferably, step S4 has only one convolutional layer and a corresponding pooling layer.
[0016] In the above structure, there is only one convolution layer and the corresponding pooling layer in the pooling layer. Multiple convolution layers can make the model more accurate and handle more complex problems, but at the same time greatly increase the complexity of the model.
[0017] The beneficial effects of the present invention are as follows: using machine learning to predict diesel engine performance has higher accuracy than traditional empirical models and lower cost than numerical simulation models. As more and more parameters need to be explored, convolutional neural networks can add more layers than neural networks, and shared convolution kernels require fewer parameters to be trained. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 It is a structural diagram of the present invention. DETAILED DESCRIPTION
[0019] like Figure 1 As shown, the present invention includes the following steps:
[0020] Step S1: provide 60 groups of experimental data for the convolutional neural network, of which 40 groups are used as training samples and 20 groups are used as test samples;
[0021] Step S2: The input layer of the convolutional neural network first normalizes the input signal;
[0022] Step S3: The convolution layer of the convolutional neural network contains 32 neurons, each neuron contains a convolution kernel, the convolution kernel length is selected to be 3, which should be less than the vector length, and the step length is 1;
[0023] Step S4: The pooling method used by the pooling layer of the convolutional neural network is average pooling, with a length of 2 and a step size of 1;
[0024] Step S5: The fully connected layer of the convolutional neural network, the number of neurons in this layer is equal to the output, which is 6. There is no convolution in the fully connected layer, so the number of weights contained in each neuron is the number of output parameters after pooling, which is 10. Multiply each output of the previous layer by the corresponding weight W2, add an offset B to get the output. The final output also needs to pass through an activation function g(x). The main problem solved in this patent is a regression problem, so the identity function is used as the output activation function, that is, the input is equal to the output;
[0025] Step S6: The loss function of the convolutional neural network selects the mean square error function MSE as the loss function, and uses the BP back-propagation algorithm to back-propagate the error to update the weight value of each neuron. The timely termination method is used to prevent the neural network from overfitting.
[0026] In this embodiment, there are 11 input parameters, including speed, power, intake temperature and pressure, exhaust back pressure, cooling water temperature, injection timing and injection pressure, fuel temperature, ambient pressure and temperature. There are 6 output parameters, namely fuel consumption rate, maximum explosion pressure, explosion pressure angle, supercharger speed, turbine pre-exhaust temperature and smoke density.
[0027] In the present embodiment, preferably, in step S2, the signal subjected to signal normalization processing is over-padding, and two zeros are added at the end using a zero-padding method.
[0028] In order to ensure that the information of the input signal is not lost when setting the above structure, the signal needs to be padded. The patent adopts the zero-padding method, adding 2 zeros at the end to ensure that the length after convolution remains unchanged.
[0029] In this embodiment, preferably, in step S3, the convolution layer utilizes convolution so that each convolution kernel in the layer has a weight W1, a common offset b is set, and the activation function uses the sigmoid function as the activation function f(x).
[0030] The above structure is set up, and the activation function is the core of the neural network. In the present invention, the common sigmoid function is used as the activation function f(x), which can nonlinearly amplify the gain.
[0031] Preferably in this implementation, step S4 has only one convolutional layer and a corresponding pooling layer.
[0032] In the above structure, there is only one convolution layer and the corresponding pooling layer in the pooling layer. Multiple convolution layers can make the model more accurate and handle more complex problems, but at the same time greatly increase the complexity of the model.
[0033] The above embodiments are merely illustrative of the principles and effects of this patent application, and are not intended to limit this patent application. Anyone familiar with the technology may modify or change the above embodiments without violating the spirit and scope of this patent application. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the relevant technical field without departing from the spirit and technical ideas disclosed in this patent application should still be covered by the claims of this patent application.
Claims
1. A diesel engine performance prediction method based on convolutional neural network, characterized in that: The process includes: Step S1: provide 60 groups of experimental data for the convolutional neural network, of which 40 groups are used as training samples and 20 groups are used as test samples; Step S2: The input layer of the convolutional neural network first normalizes the input signal; Step S3: The convolution layer of the convolutional neural network contains 32 neurons, each neuron contains a convolution kernel, the convolution kernel length is selected to be 3, which should be less than the vector length, and the step length is 1; Step S4: The pooling method used by the pooling layer of the convolutional neural network is average pooling, with a length of 2 and a step size of 1; Step S5: The fully connected layer of the convolutional neural network has 6 neurons and the same number of outputs. There is no convolution in the fully connected layer, so the number of weights contained in each neuron is the number of output parameters after pooling, which is 10. Multiply each output of the previous layer by the corresponding weight W2 and add an offset B to get the output. The final output also needs to pass through an activation function g(x). This method mainly solves a regression problem, so the identity function is used as the output activation function, that is, the input is equal to the output. Step S6: The loss function of the convolutional neural network, the mean square error function MSE is selected as the loss function, the BP back propagation algorithm is used to back propagate the error to update the weight value of each neuron, and the timely termination method is used to prevent the neural network from overfitting; There are 11 input parameters, including speed, power, intake temperature and pressure, exhaust back pressure, cooling water temperature, injection timing and injection pressure, fuel temperature, ambient pressure and temperature; there are 6 output parameters, namely fuel consumption rate, maximum explosion pressure, explosion pressure angle, supercharger speed, turbine pre-exhaust temperature and smoke density; In step S2, the signal normalized is over-padding, and two zeros are added at the end by using a zero-padding method; In step S3, the convolution layer uses convolution to make each convolution kernel in the layer have a weight W1, set a common offset b, and use the sigmoid function as the activation function f(x); The step S4 only has one convolutional layer and a corresponding pooling layer.
