Driven by the wave of digital transformation, the automotive industry has been steadily changing, upgrading, and reshaping its business. In this process, the emergence of AI algorithms has dramatically improved the operational efficiency and value balance of various business lines. In this article, we invited Mr. Zhang Yang, the head and senior director of the AI Center of Dasouche, to share his view on applying AI algorithms to the digital transformation of the automotive industry. With multi-scene digitalization, AI will play a more critical role in promoting the entire sector to get smarter and more customer-friendly in the next decade.

From Consumer Internet to Industrial Internet

Zhang Yang’s career path could be described as changing from the consumer Internet to the industrial Internet.

In 2012, Zhang started to lead the algorithm team of Sogou Pinyin, a dominant input software in China developed by the search engine company. As an active practitioner in the consumer Internet, Sougou focuses on people’s daily needs like clothing, food, housing, and transportation, with each industry having its own characteristics.

In 2021, Zhang joined Dasouche—an Internet platform for the automobile industry in China—to work on the industrial Internet. Compared with the consumer Internet, the industrial one also employs technologies like AI, cloud computing, and big data to enable application scenarios in the industry. The industrial Internet is mainly about the manufacturing field from a narrow perspective, but it extends to further areas that involve product circulation, trading, logistics, and storage that manufacturing cannot discard.

Many Internet-enabled technologies, such as AI, big data, cloud computing, blockchain, and IoT, are utilized in a growing number of industries. The industry will progressively become smarter as it moves from informationalization to digitalization.

Informationalization refers to changes in forms, whereas digitalization focuses more on the transformation of business models and applications.

Simultaneously, AI and similar technologies will accelerate digitalization by expanding their use to a greater degree.

Typical Application Scenarios of the AI-Enabled Industrial Internet

Using the automotive industry as an example, we could understand some characteristics of the industrial Internet. The automotive industry has strong potential in the market, making it an excellent case to discuss the AI-Enabled Industrial Internet. In addition, China, the biggest automotive consumer market globally, will take advantage of its size to progress, whether for new cars or second-hand.

Besides, the automotive industry does not have too many products, so realizing the technology is easier. A car company may have different car brands with various models, but the number of models could be only about 60,000 to 70,000. That is essential to applying AI as the model numbers will affect the making of knowledge graphs. With millions of models, considerable duplicated work will affect the outcome that AI can provide.

In selecting AI-powered business lines, second-hand cars will take the lead. As used cars are non-standard goods, their depreciation conditions vary significantly. This discrepancy would give enterprises a larger bargaining space to earn profits from price differences and later take it as their primary business.

In addition, second-hand car transactions in China are growing year by year with strong supporting policies, which will promote the application and circulation of used cars at the state level.

The Target and Demand Analysis

Zhang admitted that he and his colleagues had faced some difficulties and challenges regarding the circulation of the car business. As mentioned above, second-hand cars are hard to predict their use and depreciation conditions, which is challenging to realize commercialization. Meanwhile, a used car from an offline entity into an online browsable product still needs to undergo the digitalization of vehicles.

Tasking with industry empowerment, there could be some difficulties, as 4S stores and used car dealers mainly do their business offline. At the same time, online customers need to see a quick match for their needs, which could be easily affected by some geographical differences and other factors.

Lastly, the after-sales procedure of secondhand automobiles is also prone to difficulties. If a consumer acquires a previously accident-damaged car, it might result in safety-related complications.

Based on that, Dasouche would like to provide AI solutions that can be simplified as "ABC". A stands for Artificial Intelligence, B for Big Data, and C for Cloud Computing. The focus of Dasouche is mainly on the field of trading and marketing, ranging from product uploading to transaction matching and then to giving financial, logistics, and after-sale services. Damage inspection and residual value predictions are distinct features of second-hand cars during the business process, and others could also work for new vehicles.

Optimization Solutions Based on AI Technology

Vehicle Condition Inspection Optimazation

Zhang explained the application of the AI algorithm by giving details of vehicle condition inspection, residual value prediction, and transaction matching.

Condition inspection. The damage inspection starts with the vehicle permit information input. The inspector will take pictures or record videos of the exterior and interior of a car, and the AI algorithm will automatically identify the damages in the video or images. Then, the inspector will manually check and edit the presented descriptions. After confirming all the information is correct, the inspector will transfer the data to the cloud for review.

The core problem to be solved in this process is to improve the accuracy of the algorithm identification. The critical challenge is that the number of car components and damage types is too large. With this in mind, we have divided the acceleration and effect of optimizing jobs into several priorities. The regular image recognition task will use a CNN to extract basic features and then apply a Transformer-based model to the image, by which some simple adjustments are implemented. On the other hand, Zhang’s team had built pre-trained models based on car images to streamline the workflow so that damage could be identified as much as possible in a single image. False inspections and missed inspections are still inevitable, to which we have made some tunings as well.

Residual Value Prediction Optimization

Next is the vehicle residual value prediction. This is difficult as second-hand cars are hard to evaluate to a fixed standard. Various factors, including geography, vehicle condition, and car colors, could affect the ultimate price. Second, since vehicles will not be sold at retail and wholesale prices simultaneously, the problem of price inversion between retail and wholesale might exist.

To solve that, Zhang’s team optimized the data first, taking all the retail and wholesale price data of both new and used cars. Due to the different sales strategies of Chinese and foreign brands, the guide price and actual price of a new vehicle will also influence the valuation of the used car. This step requires the availability of transaction data, including models, age, locations, colors, selling times, mileage, conditions, new car guide price and new car actual price. The variation of models is relatively small, and many companies know how to use a deep learning model or a tree model. Finally, the interface will focus on operational analysis, covering prices and disposal cycles corresponding to retail and wholesale.

Transaction Matching Optimization

The last part is matching cars and making transactions, where the primary difficulty is the purchase process. People may spend a lot of money to buy a car, but that doesn’t happen every day like buying your coffee. Merchants are more likely to gather offline, which means it will take some effort to collect information and put it online to do the automatic matching. Zhang’s solution is to push the transaction online as much as possible by integrating resources within Dasouche and considering the logistics factors to solve the cross-territory problem, including the improvement of pricing strategies and bot strategies. The technical architecture here is more classic—the recommendation system model.

An optimized recommendation system focusing on more user-oriented attributes is needed to empower the matching and transaction process. The user recommendation system must coordinate with the platform revenue goals and integrate them based on the MMoE model. For trade-off optimization, a certain group of customers’ demands could be lowered by discarding some parts of the system to ensure the overall revenue.

Conclusion

Today, AI has played a crucial role in both the consumer Internet and the industrial Internet. To enhance customer experiences and advance company objectives, we must adopt a result-oriented technology application guideline and pick the suitable model based on business objectives.

There is still a place for manual, rule-based, and even simple statistical models in many enterprises, whose real impacts are sometimes superior to those of cutting-edge algorithms. In any case, there is never a shortage of difficulties and ideas when applying new technology to an enduring sector.