“i-Cremation” - The first ever AI application in crematorium operations pioneering the new era of intelligent control
By the Electrical and Mechanical Services Department with the collaboration of The Hong Kong University of Science and Technology
In today’s fast-paced world, where technology is advancing at an unprecedented rate, organisations need to keep abreast of the latest trends to stay competitive and provide the best possible service for their community. The Electrical and Mechanical Services Department (EMSD) is a prime example of a forward-thinking organisation that not only keeps pace with technological advancement but also spearheads innovation to enhance productivity, sustainability, efficiency, cost control and service quality.
With the interests of the community uppermost in its mind, EMSD has always promoted digital transformation and the application of innovation and technology (I&T) in finding engineering solutions for various government departments.
EMSD’s commitment to staying ahead of the curve in technological advancement is evident in its initiatives to integrate Artificial Intelligence (AI) into various E&M services, including building management and energy conservation. By leveraging the power of semantic AI for predictive maintenance and energy optimisation in the traditional E&M building industry, EMSD has achieved remarkable results.
As a pioneer in the new era of AI, EMSD has harnessed the power of AI to support the delivery of reliable and efficient cremation services under the Food and Environmental Hygiene Department. With the annual number of deaths in Hong Kong exceeding 50,000 and increasing by about 2% each year, and sudden demands for cremation services due to unexpected catastrophic events like the COVID-19 pandemic, EMSD proactively proposed applying AI to cremation operations to make the process more efficient and effective.
In 2019, EMSD introduced and championed a pilot project, called “i-Cremation”, which successfully implemented AI for image analysis to formulate the optimal operation. This proved the efficiency of imaging analysis. This project is patented in Hong Kong and was awarded the Silver Medal at the 47th International Exhibition of Inventions Geneva in 2022, highlighting AI’s transformative potential in traditional industries and its ability to drive forward innovative solutions to benefit the society as a whole.
Figure 1: “i-Cremation” won the Silver Medal at the 2022 International Exhibition of Inventions Geneva
The development of Al application in crematorium operation
The integration of AI into cremation operations has revolutionised the industry, resulting in shorter cremation times, reduced fuel consumption and fewer pollutants emitted from the cremators.
Back in 2019, EMSD initiated the engagement with start-ups and the I&T industry through the E&M InnoPortal, which led to over ten potential companies approaching with innovative ideas and proposals. From these, three I&T solution providers were selected to develop models for trial, applying different AI methodologies individually. As part of the effort, a cremation process database was developed by manually labelling features from 20,000 images, forming a strong foundation of “i-Cremation” to make use of the unique features of AI.
What is E&M InnoPortal?
This is a platform that lists “I&T wishes” in the area of technology from government departments, public organisations and the E&M trades. The platform links them to the I&T sector, including start-ups and universities, who can propose I&T solutions. Once proposals are received, EMSD will carry out field trials in a bid to promote and drive the research and development and application of innovative technology.
Building on this progress, EMSD partnered with The Hong Kong University of Science and Technology (HKUST), one of the strategic partners under the Memorandum of Co-operation (MoC), in 2021 for ongoing research and development to improve the performance and stability of the AI model. Through the joint efforts, the project has been in place at Wo Hop Shek Crematorium (WHSC) since the same year. It has taken a significant step forward, with each cremation session shortened by up to 12 minutes, projecting a time saving of around 10%, resulting in an increase in terms of the capability of cremation plant. resulting in a increase in terms of the capability of cremation plant. By integrating AI image analytics with deep learning algorithm, the system optimised the cremation process. With the collaboration, version 2.0 of “i-Cremation” was implemented in December 2023, demonstrating the potential for further progress and innovation in the field of AI application in crematorium operations.
Figure 2: EMSD and HKUST teams are working hard on ongoing research and development to improve the performance and stability of the Al model for “i-Cremation”
The engineering design of "i-Cremation"
To optimise cremation operations at WHSC with AI technologies, the team conducted a thorough study of the current process. The flat-bed type cremators relied on timed programme control, which could lead to an inconsistent outcome, longer process duration and higher energy use. Experienced cremation operators must be on hand to make necessary adjustments to the controls, relying on the digital cameras installed for the cremation chambers and various sensors of the cremator.
The development of “i-Cremation” began with the interpretation and digitisation of the physical cremation process, with both visual and non-visual information converted into data with numerical values. For every flat-bed type cremator in WHSC, digital cameras were installed to capture real-time footage of the cremation process, displaying objects such as casket, joss papers, cremating remains, funeral objects and flames within the primary combustion chamber (as illustrated in Figure 3). Various sensors were also incorporated into the cremator to monitor the operation of the cremation process in real time, including temperature, air pressure, oxygen levels, and so on. This visual and non-visual information, along with the adjustment controls made by the experienced operators, was used to form the dataset for the development of the AI application for the cremation process. The centralised Supervisory Control and Data Acquisition (SCADA) system was responsible for collecting and managing the operation data of cremators.
After conducting research on the cremation process and interviewing the cremation operators, the controllable elements of the cremation process were identified for the application of AI technologies. The air flap system, which consists of eight air flaps, was determined to be the main controllable parameter affecting the efficacy of the cremation. Out of the eight air flaps, six are used for the primary combustion of caskets and two for the secondary combustion of flue gas during the cremation. The air flaps affect the efficacy of the cremation in multiple ways, such as the direction of flames and level of air input with respect to the location/volume of cremating remains, temperatures, chamber pressures and the oxygen levels throughout the cremation process. As a result, these air flaps were chosen as the controllable parameters for the “i-Cremation” in the cremation process.
Figure 3: Front view and side view of the primary chamber of the cremator
With no preceding research or studies anywhere in the world as a reference point, the team adopted different technical approaches concurrently and engaged three I&T solution providers to develop the AI models for trial, using different methodologies. The first approach involved using image analytics to extract the image features of the cremation and adopt the reinforcement learning method to determine the state, action and reward for predicting the best strategy for controlling the cremation. The second approach used a deep neural network to detect the chamber image as a whole rather than the individual objects inside, using a classification method with noise filtering techniques for determining air flap controls. The third approach combined the You Only Look Once, Version 3 (YOLOv3) real-time object detection algorithm and Long Short-Term Memory (LSTM) networks for the cremation process.
The AI models developed by these approaches resulted in satisfactory performances, achieving over 90% accuracy in resembling the controls adjusted by the operators during the cremation processes. More importantly, the encouraging results proved the development of this innovative system to be in the correct direction in its use of an image-based method. This is the foundation of understanding and interpreting the status in cremators with different conditions of image brightness and orientation of objects. However, the team aimed for the even greater challenge of optimising the cremation operations in such a way that they would become more efficient than the current practices and achieve a minimised fuel consumption for the process. Therefore, the team engaged HKUST in the ongoing development of version 2.0 of “i-Cremation” to seek improvements upon the already successful version 1.0 in respect of the AL models’ performance and stability.
For the development of version 2.0 of “i-Cremation”, a new class of machine learning methods called offline reinforcement learning (ORL) was introduced. This approach leads specifically to the desired result that AI models may learn entirely “offline” from past human cremation experiences without testing its intermediate strategies in an actual cremation system, circumventing the safety concerns intrinsic to such tests. To cite an analogy, an ORL model of a chess-playing AI only needs to watch chess games played in the past by grandmasters to learn good moves; it does not need to try chess games on its own. Similarly, the cremation AI model can learn strategies purely from past data generated by experienced operators. Figure 4 provides a visual illustration of this concept.
Figure 4: High-level concept of the cremation control problem being an Offline Reinforcement Learning (ORL) problem analogous to learning chess
The main goal of ORL is to enable the AI to come up with new and better cremation strategies, even with only human demonstrations to learn—offline—from. This approach involves learning strategies by drawing out the best of the strategies of operators observed in the historical dataset. Specifically, in evaluating the quality of strategies, ORL learns from good samples in the dataset and disregards non-optimal decisions made by inexperienced staff. Since the training dataset covers cremators with different overhaul schedules, the ORL can also adapt to external factors related to machine deterioration. Despite adopting the past strategies used by the operators, the AI is able to outperform those operators by giving its full attention to and throughout the whole cremation session, making optimal control decisions at every moment. The AI’s performance is analogous to the operators’ at their best.
Another benefit of this strategy is the enhanced safety and stability of cremation control, since the AI can accurately predict how the cremation will progress by referencing past cremations. This prevents the occurrence of unexpected and potentially unsafe scenarios during cremations using AI.
Figure 5 depicts the high-level concept of how an AI selects the best cremation strategy. There are two main factors considered by the AI: the estimated cremation time and a human similarity score developed internally by the AI. The AI prioritises strategies with low estimated cremation times and high human similarity to ensure safety and stability in the cremation process.
Figure 5: High-level intuition of the brief ideas of ORL; the Al will choose the strategy that has low estimated cremation time and human-like
In the project development, the team developed the robust ORL models with a simulation model, named “simulator”, that replicates the cremator environment. The simulator has the capability to load a past cremation scenario at a particular point in time, represented by the image and sensor data from that specific point of time. The AI model would then interact with the simulator virtually as if it were an actual cremator, and the simulator would predict the future status of the cremator with the air flap values suggested by the AI. This allows the AI to develop new strategies since, on the simulator, it can test cremation strategies that differ from the operators’ in the historical dataset of cremation sessions. In version 2.0 of “i-Cremation”, a machine learning architecture called Recurrent State-Space Model (RSSM)2 was implemented as the simulation model. It is further integrated with Conservative Q-Learning (CQL)3, a state-of-the-art AI model, resulting in a final model called Conservative Offline Model-based Policy Optimization (COMBO)5. Figure 6 shows a high-level understanding of how a simulation model can further improve models by extending its training data with new simulated experiences. The simulation model simulates different cremation scenarios and assists the Al model in learning new strategies.
Figure 6: High-level intuition of how the RSSM assists the Al model to learn new strategies.
In summary, the models developed in version 2.0 of “i-Cremation” were trained solely on historical cremation data without interactions with the real cremation system. In 2023, to evaluate the trained AI models’ actual on-site performance, preliminary and long-term test trials were conducted on the actual cremators at WHSC. The tests consisted of over 600 cremation sessions tested for models. To evaluate the performance of the AI, the team also analysed the distribution and statistics of cremation sessions conducted by the operators in the AI training dataset. The test results indicated that the model effectively reduced the cremation times by up to 12 minutes, resulting in a time saving of around 10%. Figure 7 shows the distributions of cremation time of the AI models compared with operators’ performances.
Figure 7: Distribution of cremation times by Al and operators in the tests
The average completion time is shorter for the AI-controlled cremations, and the resulting reduction (by tons) of the CO2 emitted per year contributes positively to environmental protection. As cremation time depends on various external factors, such as the types and quantity of funeral objects, similar spreads of times are observed with respect to AI and human operators regardless of the adopted strategies. To examine whether the AI models have a significant advantage over human operators, a 2-sample t-testing, which is a statistical method, was employed by the team to reliably analyse the confidence of the model outperforming the operators in terms of cremation time. The results proved that the AI model has almost 100% confidence with shorter cremation times.
Overcoming the development challenges
The development of “i-Cremation” faced several challenges due to the unique nature of the project. Cremation is a taboo subject in Chinese tradition and its operation has largely relied on the judgment of experienced operators. Accordingly, whilst incorporating advanced technology, the team had to ensure that both the human touch and the respect for ancestors remain unsullied. Several processes, including the collection and analysis of visual and non-visual data, monitoring mechanisms, Al development of offline and online trainings, and finally performance verification, were employed.
Gaining the confidence of the working partners was also crucial and a hybrid operation mode was adopted, allowing for an easy switch between AI and manual modes with a user-friendly dashboard by the operators. The team built its own library of data for the cremation process and trained the model through deep-learning methods to achieve a robust and distinguishable AI model.
However, the performance of the AI model during cremations was hindered by the bright and highly turbulent flames, which the team resolved by smoothing out the fire patterns using footage sampled at high frequencies. Safety rules were also implemented for specific high-risk situations to prevent any unsafe behaviour from the AI. For instance, to protect the flue gas treatment system, if the temperature of the flue gas exceeds the threshold, the safety rule will be triggered and the AI will scale down the air flap value exponentially to decrease the amount of hot air ventilated.
Meanwhile, the rapid pace of change in the development of AI has presented significant challenges for many industries, and the i-Cremation project was no exception. Despite achieving significant milestones with version 1.0, the team has faced difficulties in keeping up with the fast-paced advancements in AI technology since the idea development in 2019.
To overcome these challenges, the team partnered with startups, the I&T industry and universities, which provided AI expertise and the necessary resources. The involvement of HKUST was particularly instrumental in driving the project forward, providing critical AI direction, academic support, and research expertise to ensure the project’s success.
One of the challenges in the “i-Cremation” project is formulating appropriate training strategies for AI models with the use of high-quality data.
Figure 8: Prof Wong Chi Wing Raymond, Associate Head and Professor, Department of Computer Science and Engineering, HKUST
How the engineering design has changed a traditional process and serves the public
The i-Cremation project is a remarkable showcase of how innovative I&T solution and collaboration among government departments, the trades, academic institutions and start-ups can benefit the society as a whole, even in the traditional cremation system. The continuous upgrades and enhancements made to “i-Cremation” have paved the way for version 2.0 with the goal of further improving the functionality and efficiency of the system.
The application of AI has proved effective in its optimisation of the cremation process and reduction of cremation time. This success has sparked interest in the application of AI to the design of cremation plants in other countries. One German manufacturer has already made important changes to the design of their cremators by adding high-resolution video cameras in the cremation chambers as a standard feature. Due to the limited land resources in Hong Kong, it is difficult and time consuming to build new crematoria to address the continuously increasing demand of cremation service. The value of i-Cremation resides most prominently in its proven ability to increase the capability of existing cremation plant while maintaining high quality public service.
i-Cremation’s widespread recognition and success support the HKSAR Government's policy to promote AI and I&T. EMSD will continue to strive ahead by adopting new technologies such as drones, robotics and IoT for E&M facilities, which are under incubation and will be deployed in the future to serve the community and contribute to building a smarter Hong Kong.
In the rapidly evolving technological landscape, continuous learning and a readiness to embrace innovations are essential traits for any professional engineer in the pursuit of success. Staying up-to-date with the latest developments in technology and skills is vital for overcoming digital challenges and leveraging the future opportunities in AI.
References
- Daneels A and Salter W (1999). ‘What is SCADA?’ International Conference on Accelerator and Large Experimental Physics Control Systems (ICALEPCS’99). Trieste, Italy, 4-8 October. 339-343.
- Hafner D, Lillicrap T, Fischer I, Villegas R, Ha D, Lee H, and Davidson J (2019). ‘Learning Latent Dynamics for Planning from Pixels’. Proceedings of the 36th International Conference on Machine Learning. Long Beach, California, USA. 2555–2565.
- Aviral K et al (2020). ‘Conservative Q-Learning for Offline Reinforcement Learning.’ Advances in Neural Information Processing Systems 33 (NeurIPS 2020). 1179-1191.
- Fujimoto S and Gu S S (2021). ‘A Minimalist Approach to Offline Reinforcement Learning’. Advances in Neural Information Processing Systems 34 (NeurIPS 2021). 20132-20145.
- Yu T et al (2021). ‘Combo: Conservative Offline Model-Based Policy Optimization’. Advances in Neural Information Processing Systems 34 (NeurIPS 2021). 28954-28967.