Letting AI control aero-engines, Yuzong Technology puts a "safety cage" on AI.
The shortcomings of small and medium-sized drones never lie in the fuselage, but in their power systems. The energy density of traditional lithium batteries increases slowly, and the battery life of most multi-rotor electric platforms is generally only 30 to 60 minutes, which cannot support tasks such as long-distance inspection, logistics, and target drones. As a result, market demand for fuel-powered systems including turbojets and turbofans has risen rapidly. However, current fuel engines still operate under PID control laws based on fixed parameters: when the throttle changes rapidly or the aircraft flies into the thin atmosphere at high altitudes, such non-linear working conditions make it difficult to balance thrust response and fuel economy. To prevent the engine from entering surge, the control system has to conservatively limit fuel supply, which further sacrifices dynamic response and task efficiency. To obtain stronger power, the system has to operate under riskier working conditions; to ensure safety, part of the performance has to be compromised — this dilemma is a long-standing unsolved problem in the industry.
Another burden comes from maintenance. The vast majority of small and medium-sized aero-engines are still maintained according to fixed cycles, and the system cannot clearly know the actual degradation level of key components such as turbine blades and bearings at the moment. As a result, regular maintenance often leads to over-maintenance, with costs spent on components that do not need replacement or repair, while hidden dangers that may cause unplanned flight stoppage may not be caught before the next maintenance window. For a drone fleet that relies on high attendance rates to deliver value, this mode of "unable to perceive the actual state and forced to disassemble and inspect on schedule" directly increases the cost and reduces the availability of the entire life cycle.
What is even harder to avoid is the supply constraint. Power systems of this thrust level have long been either dependent on imports or controlled by public research institutes, and there is a clear gap in market-oriented domestic supply. The prices of similar foreign products — including Czech PBS Velká Bíteš and German JetCat — are usually 3 to 5 times that of domestic products, with long delivery cycles and the risk of supply interruption at any time due to geopolitical factors. Four government departments including the Ministry of Industry and Information Technology of China have made it clear that the independent supply rate of core drone components should be raised to over 70% by 2027. The market demand is ready, the constraints are rigid, and what is missing is a technically independent and sufficiently smart technical path.
These three problems actually point to the same gap: a domestic small and medium-sized aero-power solution that can maximize power performance without compromising safety. Yuzong Technology aims to fill this gap, and its entry point is to integrate artificial intelligence into the engine control loop — but the first step is to figure out how to properly manage and constrain the AI system.
Letting AI take over control is premised on confining it in a "safety cage"
The technical route chosen by Yuzong Technology is Safe Reinforcement Learning (Safe RL). The team did not rush to pursue the latest algorithms at the very beginning, but compared mainstream reinforcement learning control algorithms including SAC, TD3 and PPO one by one: SAC has high exploration efficiency but unstable training performance, TD3's deterministic output is suitable for control scenarios but has low sample utilization rate, and finally PPO is selected — because PPO naturally provides a trust domain constraint through the clipping mechanism during policy update, making it easy to superimpose Control Barrier Functions (CBF), amplitude limiting of actions and independent protection logic on top of it. In other words, every control instruction of AI must pass through a preset safety boundary before execution, and any instruction beyond the scope will be invalidated.
There is a counter-intuitive design here: AI does not have the final decision-making power. Yuzong Technology adopts a dual-redundancy architecture of "AI providing suggestions and the traditional controller making final rulings". The output of AI will not be released until it is verified by the safety envelope monitor, and the final line of defense is always guarded by independent traditional protection logic. This design brings the expected performance space of the team: according to preliminary verification in the simulation environment, compared with traditional PID, the acceleration time is expected to be shortened by 15% to 20%, the fuel consumption is expected to be reduced by 5% to 8%, the design target of the surge margin is no less than 8%, and the margin loss caused by dynamic fluctuation is controlled within 5%. However, these figures are still at the simulation and design target level, and the team clearly stated in the document that the actual performance is to be verified through bench tests.
To integrate a computing model into a safety-critical system such as an engine, real-time performance and certifiability are two unavoidable hurdles. In terms of engineering implementation, the AI inference module is planned to run on an aerospace-grade embedded platform (such as Xilinx UltraScale+ FPGA or automotive-grade NXP S32V). The network is lightened through quantization and distillation, and the target response time of the control loop is controlled within 1 millisecond. The more difficult part is certification: the global aviation authorities are currently very prudent about the application of AI in safety-critical systems, and CAAC, EASA and FAA do not have mature certification standards for AI control laws. Yuzong Technology did not avoid this point, but designed the path as "traditional control as the main part and AI performance enhancement as the auxiliary part" to lower the entry threshold first, and then gradually move towards certification by referring to EASA's *AI in Aviation Roadmap* and relevant guidelines of CAAC.
Adding another layer of self-aware health management, plus two engines that have not yet been manufactured
Beyond control, Yuzong Technology aims to solve the aforementioned maintenance cost problem. The core of the health management platform in its plan is a hybrid model integrating Transformer and LSTM: Transformer is responsible for capturing global correlations from long-sequence sensor data, and LSTM processes the local dynamics of time series. The two work together to predict the Remaining Useful Life (RUL) of key components such as turbine blades and bearings, and then synchronously map the state of the physical engine to the virtual space combined with digital twin technology. The target standard is that the fault diagnosis accuracy is no less than 90%, and the average prediction accuracy of Exhaust Gas Temperature Margin (EGTM) is higher than 90%, which will upgrade regular maintenance to condition-based maintenance that is triggered only when the degradation reaches the threshold. It is expected to reduce the unplanned flight stoppage rate by more than 60% and cut the life-cycle operation and maintenance cost by 30% to 40%. Similarly, all these improvements can only be verified after sufficient real operation data is accumulated.
In terms of product implementation, Yuzong Technology has arranged two generations of products: the first generation is a 100kgf thrust class micro turbojet, targeting target drones, testing scenarios and small and medium-sized drones. It plans to complete the final assembly of the first prototype within 24 months, conduct the first ground bench test in about 30 months, complete the first round of flight verification in 42 months, and realize small-batch delivery within 48 months; the second generation is a lower-thrust turbofan, which is still in the pre-research stage, with a target fuel consumption rate 25% lower than that of traditional turbojets of the same class. It is necessary to clarify the current progress: according to the statement in its financing document, the project has completed the algorithm concept verification, and no prototype has been manufactured yet. The whole R&D process follows the four-stage progressive path of "digital simulation, semi-physical simulation, bench test, and flight verification", advancing step by step.
As for the profit model, Yuzong Technology has designed a growth curve extending from hardware to software and services. The most basic business is the sales of complete engines. The target unit price of the first-generation product is 100,000 to 200,000 RMB, and the gross profit margin of early-stage hardware is between 20% and 30%; after the AI control algorithm and health management platform technology mature and obtain certification, the software will be sold through annual authorization or charging by engine operating hours. This part has low marginal cost, and the target gross profit margin in the mature stage is over 80%; further up, the predictive maintenance service based on digital twin is subscribed by the number of units or flight hours, which turns the one-time hardware transaction into continuous service income. According to its financial forecast, the project will generate a small-scale trial delivery revenue of about 3.2 million RMB in the third year, and reach 46 million RMB in the fifth year to achieve operating break-even.
Betting on the "independent and controllable" track, and the value that this round of financing is expected to deliver
Its target customers are very clear. On the military side, its clients include research institutes and equipment procurement departments such as CASIC, CASC and AVIC, focusing on the supporting models of small and medium-sized drones for scenarios such as target drones and reconnaissance; on the civilian side, its clients include complete drone manufacturers, general aviation enterprises and logistics enterprises; in addition, it also covers government departments such as public security, fire protection and emergency response. According to the calculation of the Forward Industry Research Institute, the domestic market size of small and medium-sized aero-engines for drones is expected to reach 2 billion to 5 billion RMB in 2026, of which the power systems for special drones and target drones account for about 40% — this is also the main direction that Yuzong Technology focuses on in the early stage, and it plans to complete customer verification with 3 to 5 military research institutes before the Series A round of financing.
The team is rooted in the cross-field of "aero-engine and artificial intelligence". Founder Huang Jian has a background in mechanical engineering and has participated in the development of multiple engine models; Wu Aiqiang, in charge of AI algorithms, focuses on deep reinforcement learning and industrial control optimization, and has practical experience in applying Safe RL to safety-critical systems; the team also has a software engineering leader responsible for test data acquisition and PHM platforms, an aerodynamics and structural design leader, etc. Most of the core members have academic backgrounds in aviation, physics and mechanics from Sichuan University, and plan to carry out industry-university-research collaboration with institutions including the research institutes under AECC and the Institute of Engineering Thermophysics, Chinese Academy of Sciences.
This round of Yuzong Technology plans to launch the seed round / angel round of financing, and the specific financing amount and pre-money valuation have not been disclosed to the public. According to its plan, the funds will be mainly invested in AI algorithm development and the construction of the semi-physical simulation platform, and the rest will be used for core team building, equipment and site deployment, qualification compliance and reserves. This corresponds to a very clear goal: to complete the algorithm closed-loop verification and the final assembly of the first engine within 24 months, and carry out the first bench test in about 30 months. The team itself also acknowledges the hard challenges ahead: the prototype and bench test data have not been generated yet, there is no precedent for the certification of AI control laws in the global aviation industry, and the return on investment cycle is estimated to be 5 to 8 years. These are problems that cannot be skipped simply by raising a sum of money. For an early-stage hard technology team that is still in the stage of "moving from concept verification to prototype", the real question to answer is whether the engine can achieve the performance indicators obtained in the simulation on the bench.
Yuzong Technology is a company providing intelligent aero-power solutions for drones, with business covering adaptive engine control and health management. It is currently in the seed round / angel round financing stage.