A method for predicting the corrosion level of vehicle parts in a corrosion test

Abstract

The application discloses a kind of prediction methods of corrosion test vehicle parts corrosion grade, first combine the corrosion test data of vehicle parts with BP neural network method.The method is based on historical data and early test data, combined with the material type of target component, surface protection type, test duration and the corrosion grade of previous cycle, can quickly predict the corrosion grade of all subsequent cycles.The prediction method process includes: S1, data acquisition;S2, data preprocessing;S3, BP neural network construction;S4, model training optimization;S5, prediction output.Using the above method, the corrosion test cycle can be significantly shortened, and high-precision prediction support is provided for corrosion prevention process optimization and dynamic adjustment of test scheme.

Description

Technical Field

[0001] This invention relates to the field of automotive life prediction technology, and specifically to a method for predicting the corrosion level of vehicle components in a corrosion test. Background Technology

[0002] Corrosion testing is an important method for evaluating the long-term corrosion resistance of vehicle components. However, corrosion testing methods require engineers to wait until the entire test cycle is completed before obtaining complete corrosion data, which prolongs the time required to evaluate the corrosion resistance of components and results that cannot be provided in a timely manner.

[0003] In recent years, with the development of artificial intelligence technology, especially the application of neural network technology, prediction methods have been widely used in many fields. By utilizing historical data and experimental results, neural networks can effectively learn the complex relationship between corrosion data and various influencing factors. Based on this technology, corrosion data in the later stages of an experiment can be predicted using early corrosion data, thus providing engineers with timely prediction results to assist in decision-making and optimize designs. Summary of the Invention

[0004] This invention provides a method for predicting the corrosion level of vehicle components in corrosion testing. It is applicable to predicting the corrosion level in subsequent tests based on historical data and previous corrosion test data, and to optimizing the design.

[0005] The prediction target of this invention is vehicle components that have already undergone partial corrosion testing. During the testing process, the corrosion level, test duration, material type, and surface protection type of the previous cycles are input. The neural network model can combine this data to predict the corrosion level of the components in subsequent tests, thereby achieving a quantitative assessment of the corrosion resistance performance of the components in advance.

[0006] A method for predicting the corrosion level of vehicle components in a corrosion test, the method comprising the following steps:

[0007] S1. Data Acquisition: Collect corrosion level data of components during corrosion testing, and establish a database based on material type, surface protection type, and test duration;

[0008] S2. Data preprocessing: data cleaning, normalization, and one-hot encoding of categorical variables;

[0009] S3. BP Neural Network Model Construction: Constructing a typical BP network structure includes an input layer, hidden layers, and an output layer;

[0010] S4. Model Training Optimization: Train the network based on historical data, and combine the error backpropagation algorithm with the addition of a momentum term. The learning rules optimize the weight parameters, and the number of iterations and error thresholds are set to control the training accuracy.

[0011] S5. Predictive Output: Input the material type, surface protection type, test duration, and corrosion level of the target component in the first few cycles, and output the predicted corrosion level for all subsequent cycles.

[0012] In step S1, during the data acquisition phase, sample data from the whole vehicle corrosion test are collected, and relevant data affecting the corrosion level of the components are extracted, including basic information such as test duration, material type, and surface protection type. According to mainstream corrosion resistance testing standards in the automotive industry, the recording frequency of corrosion levels depends on the design goals and cycle definitions of the specific test procedure. Assuming the total number of cycles is m, where m equals nx, corrosion level data is recorded every x cycles, with the recorded corrosion levels being the corrosion level of cycle 0, cycle x, cycle 2x, cycle 3x, and so on, up to the corrosion level of cycle nx. The corrosion level data from the first few cycles is input, sequentially representing the corrosion level of cycle 0, cycle x, cycle 2x, ..., cycle kx. This data is then used to predict the corrosion level data for subsequent cycles, such as the corrosion level of cycle (k+1)x, cycle (k+2)x, ..., cycle nx. By collecting this data, a corrosion level database is established to support subsequent neural network training and corrosion level prediction.

[0013] In step S2, data preprocessing includes data cleaning, normalization, and one-hot encoding of categorical variables to ensure that the data is suitable for training the neural network model.

[0014] Data cleaning includes handling missing values, outliers, and duplicate data.

[0015] Missing value handling: Missing values ​​refer to data items in a dataset that were not recorded during data acquisition or input. For example, when collecting corrosion test data for vehicle parts, if the corrosion level was not recorded in a particular test due to equipment failure, that data item would be a missing value. Common handling methods include: mean imputation, median imputation, forward or backward imputation, and deletion.

[0016] Outlier handling: Outliers are data points in a dataset that significantly deviate from the normal range, usually due to measurement errors, input errors, or certain special circumstances. Failure to handle outliers may lead to inaccurate model training results. Outliers can be removed or treated as missing values.

[0017] Data duplication handling: Duplicate data refers to records in a dataset that are identical or very similar, which can lead to bias when training a model. Methods for handling duplicate data include removing completely duplicate rows from the dataset or aggregating duplicate data to calculate averages or other summary statistics, thereby avoiding the effects of excessive duplication.

[0018] Normalization refers to scaling numerical data in a dataset to a uniform range to avoid the influence of differences in the magnitude of different features during training. For example, in a corrosion test, the test duration for the material type might be in hours, while the corrosion level might be a small range of values. Without normalization, the model might rely more on some features with larger magnitudes, while ignoring other important features. Common normalization methods include min-max normalization and Z-score standardization.

[0019] One-hot encoding: a method for converting categorical data into numerical features. In neural network models, categorical data such as material type and surface protection type cannot be directly input and must be converted into numerical form. One-hot encoding creates a new binary feature column for each category. If a row of data belongs to a certain category, the corresponding feature column is 1; otherwise, it is 0. Assuming there are three material types: steel, aluminum alloy, and stainless steel, one-hot encoding will generate three binary feature columns for each material type: material type_steel, material type_aluminum alloy, and material type_stainless steel, with corresponding values ​​of [1, 0, 0], [0, 1, 0], and [0, 0, 1].

[0020] In step S3, the BP neural network model is constructed: a typical BP network structure is built, including an input layer, a hidden layer, and an output layer.

[0021] The input layer contains k+1+a+b+1 nodes, where:

[0022] k + 1 nodes are used to input the corrosion level data recorded for each x cycle, including the corrosion level of cycle 0, the corrosion level of cycle x, the corrosion level of cycle 2x, ... up to the corrosion level of cycle kx;

[0023] 'a' represents the number of material types, which is converted into 'a' binary feature columns through unique thermal encoding.

[0024] b represents the number of surface protection types, which is converted into b binary feature columns through one-hot encoding.

[0025] One node is used to input the test duration.

[0026] The number of hidden layers is typically set to one or two. For most corrosion level prediction problems, one or two hidden layers are sufficient to achieve good performance. If the model performs poorly and you need to improve the fitting ability, you can try increasing the number of hidden layers.

[0027] According to empirical formulas, the number of neurons in a hidden layer can be determined using the following formula:

[0028] (1)

[0029] In equation (1), k+1+a+b+1 is the number of input layer nodes, nk is the number of output layer nodes, and the constant c is 1~10.

[0030] The number of neurons in the hidden layer can be determined by substituting the number of input layer nodes, the number of output layer nodes, and a constant into the formula. This number can also be further adjusted based on the training results.

[0031] The output layer contains nk nodes, each corresponding to a predicted erosion level for a subsequent cycle. That is, from the erosion level of the (k+1)xth cycle to the erosion level of the nxth cycle.

[0032] The activation function for the hidden layer can be either tansig or ReLU, adjusted according to the training results. The activation function for the output layer can be the purelin function.

[0033] In step S4, model training and optimization: the network is trained based on historical data, and the error backpropagation algorithm is combined with the addition of a momentum term. The learning rules optimize the weight parameters. This is achieved by introducing a momentum factor. The influence of historical gradients is added during weight updates, as shown in the formula:

[0034] (2)

[0035] In formula (2) Let be the weight update amount in the t-th iteration. This is the update amount from the previous iteration. For learning rate, This represents the current gradient. This method can reduce oscillations in gradient descent, significantly reduce the probability of getting trapped in local optima due to a gradient of 0, and accelerate convergence to a better solution.

[0036] In step S5, the predicted output is generated by inputting the test duration, material type, surface protection type, and corrosion level data from the previous cycles into the trained neural network model, which then outputs the corrosion level for subsequent cycles. Predicting the corrosion level for subsequent cycles provides a basis for optimizing the corrosion protection design and evaluating the performance of components.

[0037] This invention is the first to combine corrosion test data of vehicle components with a backpropagation (BP) neural network method, using corrosion level data of components during the test to predict the corrosion level in subsequent cycles, and has the following advantages:

[0038] 1) It can predict the corrosion level of subsequent cycles based on existing test data before the test is completed, which greatly improves the timeliness of the prediction.

[0039] 2) This method can predict different parts and has strong versatility and adaptability.

[0040] 3) The network structure is adjustable and scalable, suitable for different training accuracy requirements.

[0041] 4) By predicting the corrosion level in advance, the waiting time in the test cycle can be reduced, providing faster decision support for corrosion protection design.

[0042] In summary, this invention provides a method for predicting the corrosion level of vehicle components in corrosion testing, offering an innovative supplement to existing methods. It has broad engineering application prospects and promotional value, and provides a scientific basis for optimizing the anti-corrosion design and performance evaluation of components. Attached Figure Description

[0043] Figure 1 A flowchart illustrating a method for predicting the corrosion level of vehicle components used in corrosion testing, provided in an embodiment of the present invention.

[0044] Figure 2 This is a framework diagram of the corrosion level prediction model for vehicle components in corrosion testing, as described in this embodiment of the invention. Detailed Implementation

[0045] This embodiment provides a method for predicting the corrosion level of components in a corrosion test vehicle, and takes the drive shaft in the corrosion test vehicle as an example to illustrate how to predict the corrosion level based on an existing test database.

[0046] like Figure 1 The diagram shown is a flowchart of a method for predicting the corrosion level of vehicle components in a corrosion test, according to an embodiment of this application. The order of the steps in this flowchart can be adjusted according to different needs, and some steps can be omitted. The prediction method includes the following steps:

[0047] Step S1: Data Acquisition

[0048] First, corrosion test data for vehicle drive shafts were collected to establish a database containing different corrosion levels. This database already covers corrosion level data from 0 to 100 cycles. For example, for drive shaft components, corresponding test data records already exist, with the corrosion level recorded every 10 cycles, i.e., the corrosion level of cycle 0, the corrosion level of cycle 10, the corrosion level of cycle 20, and so on up to the corrosion level of cycle 100.

[0049] Input features include:

[0050] 1) Material type: This indicates the type of material used in the drive shaft, such as medium carbon steel, alloy tempered steel, or aluminum alloy. Different materials may have different corrosion resistance properties.

[0051] 2) Surface protection type: This indicates the anti-corrosion treatment method used on the surface of the drive shaft, such as electro-galvanizing, hot-dip galvanizing, epoxy coating, etc. Different protection types will have different effects on the corrosion resistance of the drive shaft.

[0052] 3) Test duration: indicates the duration of the corrosion test.

[0053] 4) Corrosion rating of cycle 0, cycle 10, cycle 20, and cycle 30: These data record the corrosion rating of every 10 cycles during the initial test.

[0054] The output features are:

[0055] Corrosion levels for cycle 40, cycle 50, cycle 60, cycle 70, cycle 80, cycle 90, and cycle 100.

[0056] Step S2: Data Preprocessing

[0057] The data preprocessing stage includes cleaning, transforming, and standardizing the collected sample data to ensure the quality and consistency of the training data.

[0058] Data cleaning: Check for missing or outlier values ​​in the data and take appropriate measures to handle them. For missing values, the mean can be used to impute them, ensuring data integrity. For outliers, appropriate substitute values ​​can be used for correction, or outliers can be deleted, depending on the specific situation, to ensure the quality and consistency of the dataset.

[0059] Data normalization: Numerical data is normalized using a min-max normalization method to scale feature values ​​to a uniform range, preventing differences in feature magnitude from negatively impacting model training. For example, corrosion level is a numerical data point ranging from 0 to 9. Without normalization, the corrosion level has a relatively small range, while other features like trial duration may have larger ranges. To avoid these features having different effects on the model, their values ​​need to be scaled to a uniform range, ensuring that each feature has equal influence during model training.

[0060] One-hot encoding: One-hot encoding is used to encode categorical features such as material type and surface protection type so that the neural network can process this categorical information. One-hot encoding converts each categorical feature into a new binary feature column.

[0061] In this embodiment, the drive shaft material types include "medium carbon steel," "quenched and tempered alloy steel," and "aluminum alloy." After unique thermal encoding, the original "material type" column is split into three new feature columns: "material type_medium carbon steel," "material type_quenched and tempered alloy steel," and "material type_aluminum alloy." If the material type of a row of data is "medium carbon steel," the value of these three new feature columns is [1,0,0]; if the material type is "aluminum alloy," the corresponding value is [0,0,1]. In this embodiment, the drive shaft surface protection types include "electroplated zinc," "hot-dip galvanized," and "epoxy coating." These are also uniquely thermally encoded and split into three new binary feature columns: "surface protection type_electroplated zinc," "surface protection type_hot-dip galvanized," and "surface protection type_epoxy coating."

[0062] To ensure the effectiveness of model training and validation, this embodiment divides the collected corrosion test dataset into three parts:

[0063] Training set (70%): Used for training neural network models, through which the network learns the relationship between input features and target output.

[0064] Validation set (15%): Used to evaluate the model's performance during training and to tune its hyperparameters. The data in the validation set is not used in the model's training; it is only used for evaluation and tuning in intermediate stages.

[0065] Test set (15%): Used for the final evaluation of the trained model's performance. The data in the test set is not used for model training or validation; it is only used to evaluate the model's generalization ability and prediction accuracy.

[0066] This data partitioning method ensures that the training process of the neural network model can be carried out efficiently, while the validation set is used for tuning, and finally the test set is used to evaluate the final performance of the model.

[0067] Step S3: BP Neural Network Model Construction

[0068] Figure 2 The diagram below illustrates the structural framework of a corrosion level prediction model for vehicle components based on a BP neural network, according to an embodiment of the present invention. During the model construction phase, a BP neural network is used for prediction; the network structure is as follows:

[0069] Input layer: Contains 11 nodes, A1, A2, A3, B1, B2, B3, C, D0, D10, D20, D30, which correspond to the input features "Material type_medium carbon steel", "Material type_quenched and tempered alloy steel", "Material type_aluminum alloy", "Surface protection type_electroplated zinc", "Surface protection type_hot-dip galvanized", "Surface protection type_epoxy coating", test duration, corrosion level of cycle 0, corrosion level of cycle 10, corrosion level of cycle 20, and corrosion level of cycle 30, respectively.

[0070] Hidden layer: This layer contains 6 neurons and is used to handle non-linear relationships and map input features to output features. The number of neurons in the hidden layer can be adjusted according to the actual situation.

[0071] Output layer: The output layer contains 7 nodes, D40, D50, D60, D70, D80, D90, and D100, which correspond to the corrosion levels of the 40th, 50th, 60th, 70th, 80th, 90th, and 100th cycles, respectively.

[0072] The tansig activation function is used for nonlinear processing of the hidden layer, while the purelin linear activation function is used for the output layer to adapt to the regression prediction target.

[0073] The network structure has learning capabilities, and the parameters of each layer will be automatically optimized during the training phase based on sample feedback to adapt to the actual prediction task.

[0074] Step S4: Model Training and Optimization

[0075] After constructing the neural network structure, the model needs to be trained and its parameters optimized using training samples to obtain a corrosion level regression model with predictive capabilities. This embodiment employs supervised learning, using a pre-collected historical sample dataset for model training. The training objective is to minimize the mean squared error between the predicted corrosion level and the actual corrosion level label.

[0076] During model training, the backpropagation algorithm is used in conjunction with the addition of a momentum term. The learning rule progressively optimizes the weights and bias parameters in the neural network. Among these, the momentum factor... The learning rate is set to 0.9. The value is set to 0.01. Training is performed using a fixed batch input method, the loss function is the mean squared error function, the activation function configuration is consistent with the aforementioned structure, the hidden layer uses the tansig function, and the output layer uses the purelin function.

[0077] To ensure model training efficiency and accuracy, the following training termination conditions are set:

[0078] The maximum number of training iterations is 100.

[0079] The mean square error between the actual output and the expected output is less than 0.001.

[0080] After training is complete, the training set is compared with the validation set to evaluate the model's generalization performance and stability.

[0081] Step S5: Predict the output

[0082] After the optimized BP neural network model has completed training, it will be used for subsequent corrosion level prediction tasks.

[0083] The predicted input data includes: material type, surface protection type, test duration, corrosion level of cycle 0, corrosion level of cycle 10, corrosion level of cycle 20, and corrosion level of cycle 30.

[0084] After forward propagation of the model, the output layer returns seven consecutive values, each corresponding to a predicted corrosion level for subsequent cycles. These values ​​represent the predicted corrosion level of the prototype vehicle component under corrosion test conditions.

[0085] These predictions will be used for corrosion protection design optimization, performance evaluation, and life prediction of components. Engineers can use these predictions to determine the performance of components under different corrosion cycles, thereby enabling more accurate design and optimization.

[0086] Validation set accuracy evaluation and accuracy evaluation:

[0087] During model training, the generalization ability and prediction accuracy of the model are evaluated using a validation set. Using data from the validation set, the prediction results show that the model can accurately predict the corrosion level of subsequent cycles, with the prediction error controlled within ±1, and the validation set accuracy is no less than 75%.

[0088] If the prediction error is within ±1, the model is considered accurate. This evaluation ensures that the model's predictive ability performs well on different datasets.

[0089] Test set accuracy evaluation:

[0090] After training and validation, the model was finally evaluated using a test set. The test set data was not used in the training or validation process; it was only used to evaluate the model's performance on unknown data. The prediction results on the test set showed that the model's accuracy was also no less than 75%, further validating the model's generalization ability and predictive stability.

[0091] This method is applicable to predicting the corrosion level of various automotive components, especially during the automotive design phase, helping R&D teams optimize corrosion protection using early corrosion data. For example, based on the predicted corrosion level, the coating thickness, surface protection measures, and other anti-corrosion designs of components can be adjusted in a timely manner to extend the service life of the components.

[0092] The prediction method of this invention is not limited to automotive parts, but can also be extended to predict material corrosion in other industries. The widespread application of this method will greatly improve the efficiency of corrosion resistance design for materials, reduce maintenance costs caused by corrosion, and provide a scientific basis for industrial production and environmental protection.

Claims

1. A method for predicting the corrosion level of vehicle components in a corrosion test, characterized in that, Prediction methods include: S1. Data Acquisition: Collect corrosion level data of components during corrosion testing, and establish a database based on material type, surface protection type, and test time; S2. Data preprocessing: data cleaning, normalization, and one-hot encoding of categorical variables; S3. BP Neural Network Construction: Construct a typical BP neural network structure, including an input layer, hidden layers, and an output layer; In step S3, the construction of the BP neural network model includes: constructing a typical BP network structure, which includes an input layer, a hidden layer, and an output layer; The input layer contains k+1+a+b+1 nodes, where: k+1 nodes are used to input the corrosion level data recorded for each x cycle, including the corrosion level of cycle 0, the corrosion level of cycle x, the corrosion level of cycle 2x, ... up to the corrosion level of cycle kx. 'a' represents the number of material types, which is converted into 'a' binary feature columns through unique thermal encoding. b represents the number of surface protection types, which is converted into b binary feature columns through one-hot encoding. One node is used to input the test duration; The number of hidden layers can be set to include 1 or 2 layers, and the specific number of layers can be adjusted according to the training results; The number of neurons in the hidden layer is calculated using an empirical formula based on the number of nodes in the input layer, the number of nodes in the output layer, and a constant c. The value of the constant c ranges from 1 to 10 and can be further adjusted based on the training results. The output layer contains nk nodes, which correspond to the predicted erosion levels of subsequent cycles, i.e., from the erosion level of the (k+1)xth cycle to the erosion level of the nxth cycle. The activation functions for hidden layers include tansig or ReLU, and the activation functions for output layers include purelin. S4. Model Training and Optimization: The network is trained based on historical data. The weight parameters are optimized by combining the backpropagation algorithm with the momentum term δ learning rule. The number of iterations and the error threshold are set to control the training accuracy. S5. Predictive Output: Input the material type, surface protection type, test duration, and corrosion level data of the target component in the previous cycle, and output the predicted corrosion level results for all subsequent cycles.

2. The method for predicting the corrosion level of vehicle components in corrosion testing according to claim 1, characterized in that, In step S1, the corrosion level data includes corrosion level data recorded every x cycles, wherein the data from cycle 0 to cycle kx is used to predict the corrosion level from cycle (k+1)x to cycle nx.

3. The method for predicting the corrosion level of vehicle components in corrosion testing according to claim 1, characterized in that, The data preprocessing in step S2 includes data cleaning, normalization, and one-hot encoding of categorical variables.

4. The method for predicting the corrosion level of vehicle components in corrosion testing according to claim 3, characterized in that, The data cleaning includes missing value processing, outlier processing, and duplicate data processing; the normalization process includes the min-max normalization method; and one-hot encoding is used to convert material type and surface protection type into numerical features.

5. The method for predicting the corrosion level of vehicle components in corrosion testing according to claim 1, characterized in that, In step S4, model training and optimization include: The network is trained based on historical data, and the weight parameters of the neural network are optimized by combining the backpropagation algorithm with the momentum term δ learning rule. Set the number of iterations and error threshold during the training process to control training accuracy; The training termination conditions include two items: Reach the preset maximum number of iterations; The mean square error between the actual output and the expected output is lower than the target threshold.

6. The method for predicting the corrosion level of vehicle components in corrosion testing according to claim 5, characterized in that, The learning rule that incorporates an added momentum term δ reduces gradient oscillations and helps the model escape local optima by superimposing historical gradient directions, thereby improving model convergence efficiency.

7. The method for predicting the corrosion level of vehicle components in corrosion testing according to claim 1, characterized in that, In step S5, the prediction output includes: Input the material type, surface protection type, test duration, and corrosion level data of the previous cycles for the target component; The trained BP neural network model is propagated forward to output the corrosion level prediction value for subsequent cycles; The predicted corrosion level data can be used to optimize the corrosion protection design and performance evaluation of components, helping engineers to assess the corrosion behavior of components in advance during testing.

8. The method for predicting the corrosion level of vehicle components in corrosion testing according to claim 1, characterized in that, The predicted corrosion level can be used to optimize the corrosion protection design of components.

Images