The concept of artificial intelligence was first introduced in the 1950s, more than sixty years ago.
However, it is only in recent years that AI has seen explosive growth, mainly due to the maturation of technologies such as the Internet of Things (IoT), big data and cloud computing.
IoT enables real-time acquisition of large amounts of data, while big data provides data resources and algorithmic support for deep learning. Cloud computing offers flexible computing resources for AI.
The organic combination of these technologies drives the continued development of AI, resulting in substantial progress. The match between AlphaGo and Lee Sedol brought AI to the forefront, triggering a new wave of AI fervor.
The record-breaking launch of ChatGPT in late 2022, coupled with the popularity of AI drawing tools like Stable Diffusion, made 2023 the inaugural year of AI democratization!
Artificial intelligence research and applications are progressively flourishing in various fields. With the advent of the wave of intelligent production, artificial intelligence has been integrated into all aspects of manufacturing, including design, production, management and services.
I. Three levels of artificial intelligence technology
Artificial intelligence (AI) technology and products, through years of practical testing, are now widely applied and are accelerating the integration of AI in various industries.
From a technical point of view, it is commonly accepted in the industry that the main capabilities of artificial intelligence can be categorized into three levels: computational intelligence, perceptual intelligence and cognitive intelligence.
1. Computational Intelligence
Computational intelligence refers to a machine's superior storage capacity and ultra-fast computing capabilities. It can perform deep learning based on massive data, using historical experience to guide the current environment.
With the continuous development of computing power and the constant updating of storage methods, it can be said that computational intelligence has been realized.
For example, AlphaGo, using reinforcement learning technology, defeated the world champion Go, and e-commerce platforms employ deep learning based on user purchasing habits for personalized product recommendations.
2. Perceptual Intelligence
Perceptual intelligence refers to the ability of machines to possess senses such as vision, hearing and touch. It can structure unstructured data and interact with users in human communication methods.
With the advancement of various technologies, the value of more unstructured data is recognized and explored. Perceptual intelligence related to senses such as voice, image, video and touchpoints is also evolving rapidly.
Boston Dynamics' renowned autonomous vehicles and robots employ perceptual intelligence; they sense and process the environment around them through various sensors, effectively guiding their operations.
3. Cognitive Intelligence
Compared to computational and perceptual intelligence, cognitive intelligence is more complex; refers to the machine's ability to understand, induce, reason and use knowledge like a human being.
Currently, cognitive intelligence technology is still in the research and exploration phase.
For example, in the public security sector, extraction of characteristics and analysis of patterns of micro and macro behavior of criminals, development of models and artificial intelligence systems for crime prediction, penetration of funds and simulation of the evolution of urban crime.
In the financial sector, it is used to identify suspicious transactions and predict macroeconomic fluctuations. There is still a long way to go to accelerate the development of cognitive intelligence.
II. Artificial Intelligence Application Scenarios in the Manufacturing Industry
From an application perspective, the deployment of a single artificial intelligence technology can encompass multiple levels of core capabilities, such as computational intelligence and perceptual intelligence.
Industrial robots, smartphones, self-driving cars, drones and other smart products serve as carriers of artificial intelligence.
These products, through the combination of hardware and various types of software, have the ability to perceive, make judgments and interact in real time with users and the environment, while integrating several essential artificial intelligence capabilities.
For example, in manufacturing, a variety of intelligent robots are widely used: sorting/picking robots can autonomously recognize and grasp irregular objects.
Collaborative robots are able to understand and respond to the environment around them. Automated tracking carts can recognize faces to initiate autonomous tracking.
With the help of SLAM (Simultaneous Localization and Mapping) technology, autonomous mobile robots can utilize integrated sensors to identify feature markers in unknown environments and then estimate the global coordinates of the robot and these feature markers based on relative location and in odometer readings.
Autonomous driving technology, in terms of positioning, environmental perception, route planning, behavioral decision-making and control, also integrates various artificial intelligence technologies and algorithms.
Current applications of artificial intelligence in manufacturing industries mainly cover intelligent voice interaction products, facial recognition, image recognition, image searches, voiceprint recognition, text recognition, machine translation, machine learning, big-box computing date and data visualization.
The following text summarizes eight common AI application scenarios frequently used in manufacturing.
Scenario One: Smart Sorting
Many tasks in manufacturing require sorting. If carried out manually, the process is slow, expensive and depends on maintaining a suitable working temperature environment. Implementing industrial robots for smart sorting can significantly reduce costs and improve speed.
Consider the example of parts sorting. The parts that need to be sorted are often not well organized, and although the robot has a camera to see the parts, it does not know how to collect them successfully. In this situation, machine learning technology can be used.
The robot performs a random ordering action, then is informed whether the action was successful in picking up a piece or missed.
After multiple training iterations, the robot learns the sequence of sorting actions with the highest success rate, the optimal positions for successful selection, and the sort order that produces the highest success rate.
After several hours of learning, the robot's classification success rate can reach 90%, equivalent to the level of a skilled worker.
Scenario Two: Equipment Health Management
By performing real-time monitoring of machine operation data, resource analysis and machine learning techniques, we can predict equipment failures before accidents occur, reducing unscheduled downtime.
On the other hand, in the event of a sudden equipment failure, we can quickly diagnose the problem, identify the cause and provide corresponding solutions.
This is commonly applied in the manufacturing industry, especially in chemical engineering, heavy equipment, hardware processing, 3C manufacturing, wind energy and other sectors.
Take CNC machine tools as an example. Using machine learning algorithms and smart sensors to monitor information such as power, current and voltage of cutting tools, main spindle and feed motor during the machining process, we can identify the voltage, wear and damage status of the tools as well as the stability status of machine tool processing.
Based on these states, we can adjust machining parameters (spindle speed, feed speed) and processing instructions in real time, predicting when to replace the tool to increase machining accuracy, reduce production line downtime and improve equipment safety.
Scenario Three: Vision-Based Surface Defect Detection
The application of machine vision to detect surface defects is now common in manufacturing.
Machine vision can quickly identify small and complex product surface defects in milliseconds under constantly changing conditions and classify them, such as detecting surface contaminants, damage and cracks.
Some industrial intelligence companies have combined deep learning with 3D microscopes to increase defect detection accuracy to the nanometer level.
For detected defective products, the system can automatically determine whether they can be repaired, plan the repair path and method, and then the machinery performs the repair action.
For example, PVC pipe is one of the most used construction materials and is consumed in large quantities.
It is prone to surface scratches, holes, water ripples and dull surfaces during the production and packaging process, requiring a significant amount of manpower for inspection.
After implementing automatic visual detection of surface defects, impurities on the pipe surface are automatically detected by defining the minimum and maximum area sizes, with a minimum detection accuracy of 0.15mm² and a detection rate greater than 99%.
Scratches on the pipe surface are automatically detected by setting the minimum and maximum lengths and widths, with a minimum detection accuracy of 0.06 mm and a detection rate greater than 99%.
Wrinkles on the tube surface are automatically detected by defining minimum and maximum lengths, widths, segment lengths and color difference thresholds, with a minimum detection accuracy of 10 mm and a detection rate greater than 95%.
Scenario Four: Product Quality Inspection and Fault Determination Based on Voice Print Recognition
Using voiceprint recognition technology, we can automatically detect anomalous sounds, identify defective products, and compare against a voiceprint database to determine faults.
For example, since the end of 2018, the Faurecia factory (Wuxi) has started a comprehensive collaboration with the group's Big Data Science team, dedicated to applying AI technology to the NVH (noise, vibration and harshness) performance assessment of seat adjusters.
In 2019, Faurecia factory (Wuxi) incorporated AI technology in the detection of anomalous sounds from adjusters, realizing the automation of the entire process from signal collection, data storage, data analysis to self-learning. The detection efficiency and accuracy far exceed traditional manual inspection.
With the implementation of the noise detection system based on AI technology in the Wuxi factory, the number of employees decreased from 38 to 3. At the same time, the quality control capacity was significantly improved, with an annual economic benefit reaching 4.5 million RMB.
Scene Five: Smart Decision Making
Production companies can apply artificial intelligence technologies, such as machine learning, in conjunction with big data analysis, to optimize scheduling methods, improving their decision-making capabilities in areas such as product quality, operational management, energy consumption management and tool management.
For example, FAW Jiefang Wuxi Diesel Engine Factory's intelligent production management system features features such as anomaly data collection and production scheduling, diagnosis of abnormal causes based on decision trees, equipment downtime prediction based in regression analysis and machine-based scheduling decision optimization. apprenticeship.
Using the historical data of the scheduling decision process and the actual post-scheduling production performance indicators as the training data set and employing neural network algorithms, the parameters of the scheduling decision evaluation algorithm are adjusted to ensure that the scheduling decisions meet actual production requirements.
Scene Six: Digital Twinning
Digital twinning is the mirroring of physical entities in the virtual world. Creating digital twins integrates artificial intelligence, machine learning, and sensor data to establish a vivid, real-time “true” model that supports decision-making throughout the lifecycle of physical products.
In the pursuit of reduced-order modeling of digital twin entities, complex and non-linear models can be placed in neural networks. Leveraging deep learning, a finite objective is established, upon which reduced-order modeling is based.
For example, in the traditional model, the fluid and thermal simulation of a hot and cold water pipe outlet, using a 16-core server, requires 57 hours per calculation. After implementing reduced-order modeling, each calculation only takes a few minutes.
Scene Seven: Generative Design
Generative design is an iterative and self-innovative process. When engineers are designing products, they only need to define the desired parameters and performance constraints under system guidance, such as material, weight, volume, etc.
By combining this with artificial intelligence algorithms, hundreds to thousands of viable solutions can be automatically generated according to the designer's intent. These are then compared autonomously and the ideal design is selected and recommended to the designer for the final decision.
Generative design has become a new interdisciplinary field, deeply integrated with computing technologies and artificial intelligence, applying advanced algorithms and technologies to the design process.
Widely used generative algorithms include parametric systems, shape grammars (SG), L-systems, cellular automata (CA), topology optimization algorithms, evolutionary systems, and genetic algorithms.
Scene Eight: Demand Forecasting and Supply Chain Optimization
Leveraging artificial intelligence, we establish accurate demand forecasting models, enabling companies to forecast sales, anticipate maintenance parts requirements, and make demand-driven decisions.
Simultaneously, through the analysis of external data and based on demand forecasts, we formulate stock replenishment strategies, supplier assessments and parts selection.
For example, in a pragmatic effort to control production management costs, American Honda Motor Company wanted to understand when future customer demand would occur.
Therefore, they created a predictive model using sales and maintenance data from 1,200 dealerships. This model estimates the number of vehicles that will return to dealerships for maintenance in the coming years.
The information was then used to set benchmarks for pre-preparation of various parts. This transformation allowed American Honda to achieve forecast accuracy of up to 99% and reduce customer complaint times by three times.
III. Conclusion
Today, with a growing number of companies, universities and open source organizations entering the field of artificial intelligence, a large influx of successful open source AI software and platforms is driving an unprecedented boom in artificial intelligence.
However, compared to sectors such as finance, the applications of AI in manufacturing, although numerous, are not particularly prominent and it can be said that its development has been relatively slow.
The main reasons for these problems stem from the following three areas:
1. Collecting, utilizing, and developing data during the manufacturing process presents significant challenges. Furthermore, most companies mainly rely on private and limited-scale databases, resulting in a lack of high-quality machine learning samples. This restricts the machine's self-learning process.
2. There are a variety of differences across industry sectors, which increases the complexity of artificial intelligence solutions and increases the demand for customization.
3. Across industries, there is a lack of leading companies that can drive the trend toward deep integration of artificial intelligence with production.
By addressing these three critical questions, artificial intelligence technology could be better applied in manufacturing.
