Fuel cell stack assembly force distribution analysis method applying machine learning and data regression

Abstract

The invention discloses a fuel cell stack assembly force distribution analysis method applying machine learning and data regression. The method comprises the following steps: obtaining limit pressureintensity information of pressure paper; constructing a mapping relation from the pressure intensity to a gray value, and a regression optimization model and constraint conditions from RGB to gray; obtaining a high-precision digital image of the pressure paper extruded by the galvanic pile to obtain a three-dimensional matrix of the image, and performing background color filtering on the image; converting the cleaned image into a grayscale image according to an obtained mapping relation from RGB to a grayscale value; calculating pressure intensity distribution corresponding to the grayscale image pressure paper, and visualizing the pressure intensity distribution matrix to obtain the uniformity degree of the fuel cell stack assembly pressure distribution.

Description

technical field [0001] The invention relates to the technical field of fuel cell stacks, in particular to a fuel cell stack assembly force distribution analysis method using machine learning and data regression. Background technique [0002] The pressure distribution of the fuel cell stack assembly is extremely important to the quality performance of the product. So far, in order to obtain the internal force information of the fuel cell stack, it is difficult to obtain accurate data by installing a large number of sensors in the battery. At this stage, only pressure paper can enter the interior of the battery stack. This is what we can obtain. The only information about the internal forces of the heap. As we can see, the information that can be obtained is only different colors and shades. A very natural idea is to establish a mapping relationship between colors and forces, and then achieve a deeper pressure distribution based on image information. There are two ways to ac...

AI Technical Summary

Problems solved by technology

[0003] So far, there are two main problems in the solution proposed according to path 2: 1. Constructing the relationship between color and pressure / pressure is based on the assumption that the color value of the stressed area is not 0, and the color value of the non-stressed area It is 0 (white), so it is necessary to consider how to remove the background color interference (when the background color value is not 0, it means that the force is assumed according to the algorithm, and when the total pressure is constant, the background color will lead to the calculated pressure distribution result" more uniform", which is obviously inconsistent with the facts); 2. If you use RGB (or HSV, etc.) The actual pressure distribution (for example, if the color image is observed to be obviously uneven, but after conversion to grayscale scalar, the calculated distribution is abnormally uniform, at this time, there is obviously a problem with the mapping)

We know that there are three issues that have to be considered in the gray scale of digital images and the construction of mapping relationships: 1. Under different conditions, other background colors will inevitably appear in addition to the image information itself, especially if this kind of interference exists between flow fields, It has a great influence on the calculation results, and we have to try to eliminate this interference; 2. Under normal light-sensitive conditions, the background color interference and the color value produced by the strain between the flow field are very close, which can only be distinguished by naked eyes, and use such as Photoshop A class of software processing can easily cause a large amount of redundancy or loss of information, so it is necessary to avoid such situations through progressive classification algorithm processing; 3. The original intention of the method of color to grayscale is that digital image processing technology is to 3-dimensional information into 1-dimensional, and then it is easy to use scalar gradient, histogram equalization and other means to solve the problem of image recognition. It has nothing to do with the color to gray scale and the construction of pressure gray scale mapping involved in our project. In other words, the image processing field The conversion formula used does not mean that after using it directly, the converted gray scale can relatively accurately reflect the distribution expressed by the original color space, so it is also necessary to construct a color gray scale conversion model in an empirical and optimized way, so that the conversion is relatively accurate. Preserve the original distribution characteristics more accurately

Images