Teaching AI to handle the messy real world

Teaching AI to handle the messy real world

MUST researchers are developing solutions across 13 patented technologies to enable AI to handle incomplete, mismatched, and unfamiliar data, ranging from medical images and factory sensors to underground geology and planetary surfaces.

Artificial intelligence excels at recognizing patterns when it is provided with the right kind of data it was trained on. However, the real world often doesn’t offer such ideal conditions. For example, a doctor may have an eye image, but some diagnostic clues could also be in a patient’s medical history [1]. A tunnel-boring machine may generate vast sensor data, yet the geological information describing the rock ahead is often sparse [2]. An AI trained to identify impact craters on one planet might struggle with different landscape appearances [3]. In manufacturing, AI may find it difficult to detect rare defects because there aren’t enough examples for effective training [4]. Although these issues seem different, they share a common challenge: real-world data is rarely clean, complete, and consistent.

Researchers at Macau University of Science and Technology (MUST), along with partners, have addressed this issue across 13 patented technologies in fields like medicine, manufacturing, tunneling, environmental monitoring, software analysis, and planetary exploration. These technologies do not share a single AI architecture but demonstrate a recurring approach: adapt how evidence is selected, combined, and represented before AI processes it.

 

When different kinds of information do not fit together

An illustrative example arises from underground engineering, where a shield tunnelling machine generates dense operational data, whereas geological exploration offers information at a much broader scale. MUST researchers developed a method allowing geological data itself to guide how the machine’s operational data are compressed before entering a cutter-wear prediction model [2]. Instead of treating all measurement streams equally, the data representation is tailored based on physical evidence of the ground ahead.

A comparable concept appears in healthcare, though the data types differ. A glaucoma-prediction approach extracts significant anatomical features from colour fundus photographs and integrates this image-based information with a knowledge graph created from electronic medical records, including medical history, lifestyle, medication, and eye exam data [1]. The goal isn’t just to enhance image detail but to enable various evidence forms to contribute to the prediction.

In epidemic forecasting, another innovation combines case-number signal decomposition with LSTM prediction, incorporating passenger-flow data from high-speed rail, aviation, and road transportation to help define the forecast window [5]. External factors, such as transportation activity linked to intervention levels, thus influence the forecasting approach the time-series evidence is interpreted.

 

Helping AI cross unfamiliar territory

Sometimes, the challenge isn’t missing information but change: AI encounters situations after deployment that differ from what it saw during training. MUST’s crater-detection patent tackles this through domain adaptation. It enhances relatively transferable low-level features like texture and color, specifically emphasizes the circular boundaries of impact craters, and gradually aligns features from a labeled source domain to an unlabeled target domain [3]. The goal is to maintain crater recognizability even when surrounding planetary imagery varies.

A related issue arises in environmental monitoring. Microscopic algae images may come from different instruments, websites, and formats. The patented method builds a source dataset from automatically collected and filtered algae images, then uses transfer learning to minimize differences between source and target datasets, enabling a more versatile model to identify multiple algae types across diverse imagery [6].

 

When there are not enough examples

AI faces challenges when training examples are limited. For instance, chip defects are harder to find than normal samples. A patented method extracts features from defect images, varies those features, and uses generative adversarial networks (GANs) to create additional images. Both original and synthetic datasets can be combined to train defect detectors [4]. In medicine, where labeled images are costly, another patented approach employs a multi-task adversarial network for simultaneous organ and bone-suppression segmentation, leveraging related tasks to boost performance with few training samples [7]. A welding-quality system addresses limited data differently by converting sensor measurements, like plasma concentration, laser intensity, and temperature, into features within a graph neural network. This network evaluates laser-welding quality based on the relationships among available samples [8].

 

Helping AI understand what the data mean

Other inventions aim to extract representations that better capture the true structure of the object being recognized. For instance, medicinal materials often do not fit the regular rectangular sampling pattern used in conventional convolution. A patented recognition technique employs deformable convolution, allowing the network to better follow irregular shapes of medicinal materials while integrating information across multiple feature scales [9]. A related patent addresses the shift from single-object to multi-object images. Although trained on single-object images, it fuses various levels of visual features and identifies regions of interest to locate and recognize multiple objects without needing to generate multi-object training images [10].

The concept of preserving meaningful structure extends beyond photographs. In clone-code detection, source code is first compiled into an intermediate form. Abstract syntax trees, control-flow graphs, and data-flow graphs are then converted into vectors for deep learning, enabling detection of semantically similar code. This approach aims to maintain both syntax and program behavior, rather than relying solely on surface-level similarities [11]. For images with thin, linear features, a MUST patent combines global context, local detail, and an edge-detection network. Edge features are fed back into the decoder during image reconstruction, helping to preserve narrow boundaries and generate more precise segmentations [12].

 

When one prediction is not enough

AI’s initial response can serve as preliminary evidence rather than the final conclusion. In upper gastrointestinal endoscopy, a patented approach first classifies an image and then searches a training database for similar images. The initial prediction is compared with diagnoses from those retrieved cases before finalizing the lesion category; if suitable, a segmentation model then delineates the lesion area [13]. Essentially, the system enhances direct predictions with evidence from previous similar cases. These technologies employ varied methods, utilizing techniques like convolutional networks, attention mechanisms, graph learning, adversarial learning, transfer learning, domain adaptation, LSTM networks, and knowledge graphs. Therefore, the portfolio shouldn’t be viewed as a single universal AI approach.

What unites these patents is their practical application. For example, a tunnelling system compresses dense machine data based on sparse geological evidence [2]. A crater detection system identifies features that remain relevant across different visual domains [3]. A chip-inspection system generates additional examples when real defects are rare [4]. Medical systems combine images with patient data [1], multi-task learning on related imaging tasks [7], or reevaluate predictions by retrieving similar past cases [13]. Other systems reorganize data based on relationships, shapes, edges, or underlying structures [8]–[12].

Together, they underscore a key challenge in applied AI: larger models do not always lead to better decisions. Sometimes, success begins with supplying models with evidence that is better structured to reflect real-world problems. As AI moves from curated datasets to applications in hospitals, factories, tunnels, environmental locations, and space missions, the capacity to operate effectively with imperfect evidence will likely become even more vital.

 

From research to real-world applications

Industries dealing with incomplete, limited, or mismatched data can find new development opportunities through these patented technologies. Companies and partners interested in licensing, collaboration, or adapting these approaches for practical use are encouraged to contact MUST to explore the patent portfolio.

 

---

References

[1] K. Zhang, Y. Gao, and Z. Zou, “青光眼辅助预测方法及系统 [Glaucoma-assisted prediction method and system],” China Patent Application CN 118711790 A, Sep. 27, 2024.

[2] L. Bai, D. Mo, N. Wu, and W. Huang, “用于滚刀磨损预测的盾构机运行数据处理方法、装置、计算机设备、可读存储介质和程序产品 [Shield-machine operating-data processing method, apparatus, computer device, readable storage medium and program product for cutter-wear prediction],” China Patent Application CN 119623293 A, Mar. 14, 2025.

[3] Z. Cai and S. Yang, “一种域自适应撞击坑检测方法、系统、装置及存储介质 [Domain-adaptive impact-crater detection method, system, apparatus and storage medium],” China Patent CN 116091905 B, Apr. 21, 2026.

[4] F. Deng, J. Luo, Y. Wang, Y. Huang, J. Zhong, N. Li, H. Wang, D. Zhong, and J. Xi, “芯片缺陷检测方法、装置、设备及介质 [Chip-defect detection method, apparatus, device and medium],” China Patent Application CN 117173105 A, Dec. 5, 2023.

[5] Z. Han, Z. Yang, Z. Zeng, W. Guan, and Z. Lin, “一种基于交通网的 EMD-LSTM 疫情预测方法 [EMD-LSTM epidemic prediction method based on transportation networks],” China Patent Application CN 115662651 A, Jan. 31, 2023.

[6] J. Li, A. Yuan, B. Wang, H. Zou, and J. Wang, “多源藻类图像目标检测方法、系统、电子设备及存储介质 [Multi-source algae-image object-detection method, system, electronic device and storage medium],” China Patent CN 115311657 B, Jan. 5, 2024.

[7] W. Huo and X. Tian, “一种图像分割方法、分割装置、终端设备及存储介质 [Image-segmentation method, segmentation apparatus, terminal device and storage medium],” China Patent Application CN 114764805 A, Jul. 19, 2022.

[8] F. Deng, Y. Huang, G. Yao, H. Feng, W. Li, Y. Hu, H. Wang, Y. Ding, D. Zhong, J. Xi, and N. Li, “焊接质量检测方法及其装置、计算机可读存储介质 [Welding-quality detection method and apparatus, and computer-readable storage medium],” China Patent CN 112967231 B, Nov. 15, 2022.

[9] Z. Jiang, Z. Cai, L. Bai, Y. Zhang, J. Liu, P. Lü, B. Ye, T. Lan, and D. Zhang, “一种识别药材的方法和装置 [Method and apparatus for identifying medicinal materials],” China Patent Application CN 114818874 A, Jul. 29, 2022.

[10] Z. Cai, Z. Jiang, B. Ye, P. Lü, Y. Zhang, J. Liu, T. Lan, and L. Bai, “图像识别方法、电子设备及存储介质 [Image-recognition method, electronic device and storage medium],” China Patent Application CN 114821554 A, Jul. 29, 2022.

[11] D. Yuan and T. Zhang, “一种克隆代码检测方法及装置 [Clone-code detection method and apparatus],” China Patent Application CN 117149269 A, Dec. 1, 2023.

[12] L. Bai, T. Zhang, J. Liang, C. Xia, N. Wu, and D. Mo, “图像中线状结构识别分割的深度学习模型、方法、存储介质和装置 [Deep-learning model, method, storage medium and apparatus for recognition and segmentation of linear structures in images],” China Patent Application CN 118865171 A, Oct. 29, 2024.

[13] Z. Zheng, S. Tang, Y. Liang, X. Yu, H. Yu, and Y. Xu, “基于多任务辅助的上消化道病变区域确定方法及装置 [Multi-task-assisted method and apparatus for determining upper gastrointestinal lesion regions],” China Patent CN 113744203 B, Mar. 25, 2025.