Automatic vehicle salt spray corrosion detection method based on neural network model

Abstract

The invention relates to the technical field of automobile detection, and particularly discloses a vehicle salt-spray corrosion automatic identification method based on an intelligent system. According to the method, a full-automatic vehicle body corrosion condition detection system is constructed through standardized vehicle parking and positioning, a modularized salt mist environment control unit (including temperature and humidity closed-loop adjustment and salt mist nozzle matrix layout), a multispectral image acquisition device and a deep learning corrosion recognition algorithm. The core of the technical scheme is that a computer vision technology is dynamically compared with a corrosion characteristic database, and intelligent evaluation is realized through double criteria of corrosion area threshold judgment and metal corrosion grade early warning. According to the method, the operation intensity of traditional manual detection is remarkably reduced, a cooperative verification mode with manual reinspection is supported, the detection efficiency is improved, the misjudgment rate is reduced, and the method is particularly suitable for standardized detection of batch salt spray experiments in an automobile test field and can be seamlessly connected with an MES quality management system to generate a digital corrosion resistance detection report.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence image recognition and classification, and particularly relates to a system for identifying the corrosion condition of a current vehicle, recording relevant data, and outputting a test report by using partial photos of a vehicle taken by a high-definition camera in a salt spray test workshop. Background Art

[0002] At present, the salt spray test method in an automobile test field relies on manual visual inspection based on experience to judge the corrosion condition of a vehicle to be tested at each corrosion cycle node, and gives the corresponding score of the vehicle's corrosion resistance performance index.

[0003] Obviously, the current detection method has disadvantages such as high labor cost, large subjective influence, low operation efficiency, and high requirements for employees' abilities. For the above reasons, the present invention uses AI to replace manual labor, and based on neural network image recognition and classification, automatically completes the work of detecting the corrosion condition of a vehicle, recording data, and outputting a corrosion resistance performance report in the salt spray test of an automobile test field. Summary of the Invention

[0004] The purpose of the present invention is to use artificial intelligence to replace manual operations in the test process, which not only reduces labor costs but also improves the accuracy of the test. Since the vehicle salt spray test is a test with a long time span, generally, it takes 60 test cycles (about 24 hours / test cycle) for a vehicle to complete a set of salt spray tests at present. The corrosion condition of the vehicle needs to be recorded in each test cycle. Different test recorders and different recording times will affect the test results, so the accuracy of the results will be affected. Artificial intelligence can completely avoid these disadvantages, take corrosion photos on time and accurately, analyze the corrosion type, and give accurate performance index scores according to the national standard corrosion resistance performance evaluation procedure. The method includes the following steps:

[0005] Install high-definition cameras in the salt spray corrosion test field. The installation positions include but are not limited to the roof, walls, ground, engine compartment of the vehicle body to be tested, the vehicle body, and the chassis position. The number of cameras installed is related to the vehicle parts (components) to be tested and increases correspondingly as the number of vehicle parts to be tested increases.

[0006] The accuracy of the camera should meet the test requirements, and the clarity of the captured images should be able to see the corrosion condition within a distance of 0.5 mm.

[0007] The data transmission method adopts wireless or wired transmission.

[0008] Establish an intelligent image recognition and classification system based on the CNN neural network. The method includes the following steps:

[0009] (1) Determine the number of input and output quantities of the CNN neural network model;

[0010] (2) Establish a CNN neural network model;

[0011] (3) Collect the standardized corrosion conditions of various vehicle parts and their corresponding corrosion grades, and establish a database;

[0012] (4) Train the CNN neural network model, and then verify that the model reaches the allowable accuracy.

[0013] Further, in the step

[0009] (1), the input quantity information is a photo of local corrosion of the vehicle, where the pixel requirement of the photo is not less than 4K, the accuracy requirement is to be able to see the corrosion within a range of 0.5 mm, and the shooting range is not less than 100×100 mm 2 ; The output quantity information is the corrosion type, corrosion diffusion amount, and corrosion degree grade in the photo.

[0014] Further, according to the content of the step

[0013] , the input quantity is 4096×4096×3 neuron nodes. 4096 represents the number of pixel points, the range is from 0 to 1, and 3 represents the three RGB channels; the output quantity is 3×1 neuron nodes, which represent the corrosion type, corrosion diffusion amount, and corrosion degree grade respectively. The output range of the neuron node representing the corrosion type is 0 and 1, 0 represents non-red rust corrosion, 1 represents red rust corrosion. The output range of the neuron node representing the corrosion diffusion amount is a real number from 0 to 6, representing the range of the corrosion diffusion amount, with the unit of millimeter. The output range of the neuron node representing the corrosion degree grade is an integer from 0 to 9, representing the corrosion grade, and the maximum corrosion grade is 9.

[0015] Further, in the step

[0010] (2), the structure of the CNN neural network model from input to output is successively 1 input layer, 2 convolutional layers, 1 pooling layer, 2 convolutional layers, 1 pooling layer, 3 convolutional layers, 1 pooling layer, 3 convolutional layers, 1 pooling layer, 3 fully connected layers, and 1 output layer.

[0016] Further, according to the content of the step

[0015] , the size of the convolutional kernel of the convolutional layer is determined by the network layer and the pooling method. The convolutional kernel value is determined after training and optimization adjustment. The initial value of the convolutional kernel can be selected as a random positive integer less than 10. The pooling method of the pooling layer adopts average pooling, max pooling, random pooling, and global average pooling. The weight adjustment method of the fully connected layer adopts the error back method of gradient descent.

[0017] Further, in the step

[0011] (3), the data types collected include automotive corrosion test cycle nodes, indoor temperature, humidity, salt concentration in the salt spray, salt spray spraying time, photos of vehicle local corrosion conditions, corrosion evaluation (type and degree), weight of the test standard steel plate, and the thinning amount of the test standard steel plate at each cycle node.

[0018] Further, in step

[0012] (4), the established CNN neural network model is deeply learned and trained using the gradient descent learning rule of error. The error threshold is 0.01, and when the sample mean square error is less than the error threshold, the training stops.

[0019] The sample data for the deep learning training of the CNN neural network model comes from the database described in

[0011] (3). 70% of the database data is used for the deep learning training of the CNN neural network model, and the remaining 30% is used for the accuracy detection of the model.

[0020] If the number of errors between the output value of the CNN neural network model and the reference value in the database is not less than 90% of the total number, the CNN neural network model is considered valid; otherwise, the learning rate, network structure, initial value of the convolution kernel, and initial weight value of the fully connected layer in the learning rule are modified, and the CNN neural network model is deeply learned and trained again until the model is valid.

[0021] The CNN neural network model is software encapsulated.

[0022] Call the database described in

[0011] (3) to enable the software to have the following functions: while outputting the corrosion type, corrosion diffusion amount, and corrosion degree level of the local position of the vehicle, it also outputs the current time, weather, corresponding corrosion test cycle node, temperature, humidity, corrosion test content, and the thinning amount of the standard steel plate for the cycle node test, and outputs this information in the form of word and PDF reports.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] More accurate test results. The vehicle salt spray corrosion test has a long time span. Compared with manual detection, intelligent detection is not affected by the environment. No matter at what time point and no matter which detection, the detection standard always remains the same. Therefore, the detection results are more accurate.

[0025] Automated operation without manual participation. The staff only needs to park the vehicle in the designated position and turn on the intelligent detection software. The computer will record various test data of the tested vehicle for each test cycle on time and give the analysis results.

[0026] Reduce the operation difficulty of the vehicle salt spray corrosion test. The detection personnel for the vehicle salt spray corrosion test need to have professional knowledge reserves and rich work experience to make accurate judgments on the corrosion situation of the vehicle. Compared with manual detection, the intelligent detection method reduces the number of operators and reduces the operation difficulty of the operators. Description of the Drawings

[0027] Figure 1 It is a schematic layout diagram of the salt spray corrosion test workshop for the automatic vehicle salt spray corrosion detection method based on a neural network model according to the present invention.

[0028] Figure 1 In the figure, No. 1 is the salt spray corrosion workshop, No. 2 is the high-definition cameras installed in the workshop, which are distributed around the walls and under the vehicle. No. 3 is the computer equipment required for the corrosion detection system, and this equipment needs to be placed in a room outside the salt spray corrosion workshop to avoid being damaged by salt spray corrosion. No. 4 is the vehicle to be detected, parked on the platform in the middle of the workshop.

[0029] Figure 2 It is a schematic diagram of the CNN network structure.

[0030] Figure 2 In the figure, No. 1 is the input layer, No. 2 is the convolutional layer, No. 3 is the pooling layer, No. 4 is the fully connected structure layer, and No. 5 is the output layer. Detailed implementation manners

[0031] It should be noted that the terms used herein are only for describing the specific implementation manners and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they specify the presence of features, steps, operations, devices, components, and / or their combinations.

[0032] The following details the specific implementation steps of the present invention with reference to the accompanying drawings, so that those skilled in the art can implement the technical solution of the present invention based on this specification.

[0033] As Figure 1 shown, multiple high-definition cameras are installed in the salt spray corrosion test workshop. It is required to use industrial cameras with a 4K resolution, and the lens focal length is adjusted to 50 mm to ensure that the shooting range covers an area of 100×100 mm² and meets the corrosion detail recognition accuracy of 0.5 mm.

[0034] The installation positions of the cameras include around the workshop walls, the platform under the vehicle chassis, the engine compartment, and the door gaps. Wide-angle cameras (viewing angle ≥ 120°) are deployed around the workshop walls for shooting local positions of the vehicle body; anti-salt spray cameras (IP68 protection level) are installed on the platform under the vehicle chassis, with a fixed telescopic bracket, and the vertical distance from the vehicle body chassis is 10 cm; micro magnetic adsorption cameras are used in the engine compartment and door gaps to transmit the local corrosion images of the components in real time through wireless transmission.

[0035] Shielded twisted pair cables (Cat6 standard) are laid out in the factory building to connect the cameras to the external control computer. The transmission delay is ≤10ms. A 5GHz Wi-Fi module is used in the moving detection area to ensure real-time image transmission.

[0036] The collected image data needs to be format-converted. The original image is converted from the BGR format to an RGB three-channel tensor (4096×4096×3), and the pixel values are normalized to the range [0,1]. Random rotations of ±10°, horizontal flips, and brightness adjustments (±15%) are applied to the training set images to improve the generalization ability of the model.

[0037] The data fields include test cycle nodes, environmental temperature (unit: °C), humidity (unit: %), salt fog concentration (unit: g / m³), corrosion type label (0 / 1), corrosion diffusion amount (unit: mm), and corrosion grade (0 - 9 levels).

[0038] As Figure 2 shown, the CNN network structure is, in sequence, the input layer (numbered 1), which receives an RGB image tensor of 4096×4096×3; the convolutional layers in the 1st - 3rd layers have a convolutional kernel size of 200×200, a stride of 1, an output channel of 3698×3698, the activation function is ReLU, the pooling layer uses 4×4 average pooling with a stride of 4; the convolutional layers in the 4th - 6th layers have a convolutional kernel of 10×10, a stride of 1, an output channel of 907×907, the activation function is LeakyReLU (α = 0.1), the pooling layer uses 2×2 max pooling; the convolutional layers in the 7th - 10th layers have a convolutional kernel of 5×5, a stride of 1, an output channel of 442×442, the activation function is ReLU, the pooling layer uses stochastic pooling with a stride of 4; the convolutional layers in the 11th - 14th layers have a convolutional kernel of 5×5, a stride of 1, an output channel of 99×99, the activation function is ReLU, the pooling layer uses global average pooling; the first fully connected layer has 9801 neurons, the activation function is ReLU, the second fully connected layer has 512 neurons, the activation function is ReLU, and the third fully connected layer has 25 neurons, the activation function is ReLU.

[0039] During the deep learning training of the model, the weight of the corrosion type in the loss function is 1.0, and the weights of the corrosion diffusion amount and corrosion grade are 0.5 and 1.0 respectively. The number of training epochs is ≤2000.

[0040] 30% of the data is extracted from the database as the test set. If the accuracy of the test set is <90%, adjust the initial value of the convolutional kernel or the learning rate (range 0.0001 - 0.01) and retrain.

[0041] The camera is automatically triggered to take pictures every 1 hour. The images are preprocessed and then input into the model.

[0042] The report output formats include Word and PDF formats, and the content includes timestamps, vehicle location maps, corrosion data, and a comparison table of the thickness reduction of the test standard steel plates.

[0043] To ensure the long-term stable operation of the system, it is necessary to clean the camera lens with anhydrous ethanol every month and calibrate the focal length error ≤ ±0.1 mm; inject new corrosion sample data every 12 months, retrain the model, and verify that the accuracy fluctuation < 2%.

[0044] The above description of the embodiments of the present invention is for clarification purposes and is not intended to limit the form disclosed by the present invention. Modifications or changes based on the above teachings or learned from the embodiments of the present invention are possible. The embodiments are selected and described to explain the principles of the present invention and enable those skilled in the art to utilize the present invention in various embodiments in practical applications. The technical idea of the present invention is intended to be determined by the claims and their equivalents.

Claims

1. An automated vehicle salt spray corrosion detection system based on a neural network model, characterized in that Inside the salt spray test workshop, there is a convex platform for parking the vehicle to be tested; multiple high-definition cameras are installed around the walls of the salt spray test workshop, under the convex platform under the vehicle chassis, in the engine compartment, and at the door gaps. The cameras are configured as industrial cameras with a 4K resolution, the lens focal length is adjusted to 50 mm, the shooting range covers an area of 100×100 mm², and corrosion details within a range of 0.5 mm can be recognized; a data transmission module using shielded twisted pair or a 5GHz band Wi-Fi module is used to realize real-time image transmission to an external control computer; an intelligent image recognition and classification system based on a CNN neural network is used to process the image data collected by the cameras and output the corrosion type, corrosion diffusion amount, and corrosion degree level; the test results are integrated with the test environment data, and a report in Word and PDF formats containing a timestamp, a vehicle position map, and corrosion data is output.

2. The vehicle salt spray corrosion automatic detection system according to claim 1, wherein The high-definition cameras include wide-angle cameras with a viewing angle ≥120°, which are installed around the walls of the workshop and are used to photograph local positions of the vehicle body; salt spray-proof cameras with an IP68 protection level are installed on the convex platform under the vehicle chassis, 10 cm vertically from the vehicle chassis; micro magnetic adsorption cameras are installed in the engine compartment and at the door gaps, and locally corroded images of components are transmitted back in real time through wireless transmission.

3. The vehicle salt spray corrosion automatic detection system according to claim 1, wherein The structure of the CNN neural network includes an input layer that receives an RGB image tensor of 4096×4096×3; two convolutional layers, one pooling layer, two convolutional layers, one pooling layer, three convolutional layers, one pooling layer, three convolutional layers, one pooling layer, three fully connected layers, and an output layer connected in sequence; the pooling layer adopts average pooling, max pooling, random pooling, and global average pooling methods; the output layer contains 3×1 neuron nodes, corresponding to the corrosion type (0 / 1), corrosion diffusion amount (0-6 mm), and corrosion degree level (0-9 levels) respectively.

4. The vehicle salt spray corrosion automatic detection system according to claim 3, wherein, The training method of the CNN neural network includes using the error gradient descent learning rule, and the weights of the corrosion type, corrosion diffusion amount, and corrosion level in the loss function are 1.0, 0.5, and 1.0 respectively; the training set data contains 70% of the database samples, and the test set data accounts for 30%. When the accuracy of the test set <90%, the initial value of the convolutional kernel or the learning rate (0.0001-0.01) is adjusted and retrained; The training termination condition is that the mean square error of the sample error <0.01 or the number of training epochs ≤2000.

5. The vehicle salt spray corrosion automatic detection system according to claim 1, characterized in that, The database includes vehicle corrosion test cycle nodes, environmental temperature, humidity, salt spray concentration, and salt spray spraying time; Local corrosion images of the vehicle and their corresponding corrosion type labels, corrosion diffusion amounts, and corrosion levels; The weight of the test standard steel plate and the thinning amount at each cycle node.

6. The vehicle salt spray corrosion automatic detection system according to claim 1, characterized in that The image preprocessing steps include converting the original BGR format image into an RGB three-channel tensor and normalizing the pixel values to the range [0,1]; randomly applying ±10° rotation, horizontal flipping, and ±15% brightness adjustment to the training set images.

7. The vehicle salt spray corrosion automatic detection system according to claim 1, wherein The system maintenance method includes cleaning the camera lens with anhydrous ethanol every month, and calibrating the focal length error ≤ ±0.1 mm; injecting new corrosion sample data every 12 months, retraining the model and ensuring that the accuracy fluctuation < 2%.

8. An automated vehicle salt spray corrosion detection method based on a neural network model, characterized in that, It includes the following steps: real-time collecting local corrosion images of the vehicle to be detected through multiple high-definition cameras; transmitting the image data to the CNN neural network model to identify and classify the corrosion type, diffusion amount and grade; generating a detection report in combination with the test environment data and outputting it in Word and PDF formats; automatically triggering the camera to take pictures and update the detection results every 1 hour.

Images