/glossary
Glossary
This page does two jobs: it makes the rest of the site readable, and it is where people searching for these terms land — terms on the solution pages link back here through a tooltip.
32 terms in total
Confidence Scoreคะแนนความมั่นใจA number the model returns alongside its answer, saying how sure it is that the answer is right — not the real accuracy of the system overall.Configure the system to send low-confidence parts to a human for a second look instead of deciding itself.
Datasetชุดข้อมูลThe collection of examples (images, text and so on) used to train or test a model.The training dataset mixes good and defective parts together so the model learns to tell them apart.
Edge AIประมวลผลที่ปลายทางLetting the computer make the decision right where the data is captured, instead of sending it somewhere else.A box beside the conveyor decides whether a part is defective, in 41 milliseconds.
Embeddingเวกเตอร์แทนความหมายTurning a word or passage into a set of numbers that stands for its meaning, so a computer can compare how similar two pieces of text are.Two documents that use different words but mean similar things end up with embeddings close together.
False Negativeการมองข้ามThe system says a part is good when it is actually bad, so the defect reaches the customer. More dangerous than a false positive because it hits your credibility directly.Teams usually weigh this number more heavily than false positives, because the consequences are worse.
False PositiveการเตือนผิดThe system says a part is bad when it is actually good, costing re-inspection time and eroding trust in the system.Most plants aim to get the false-positive rate under 2% after one or two weeks of tuning.
Fine-tuningการปรับจูนเพิ่มเติมTraining a model that already has general knowledge further on data specific to your job, so it gets accurate for that job without starting over.Take a general vision model and fine-tune it on the plant's own part images for another two weeks.
FPSเฟรมต่อวินาทีHow many images a camera or system can process in one second. The faster the conveyor, the higher this has to be.A faster conveyor needs a camera that can process proportionally more frames per second, or parts get missed.
Global Shutterชัตเตอร์ทั้งภาพพร้อมกันA sensor that reads every row at the same instant, so moving objects do not come out skewed.Necessary once the conveyor moves fast enough that an ordinary rolling-shutter camera starts to skew the image.
HMIหน้าจอควบคุมเครื่องจักรThe touchscreen the people on the floor use to operate a machine or check its status.Show inspection results on the line's existing HMI screen — no extra display needed.
Inferenceการใช้งานโมเดลจริงThe step where an already-trained model is used to answer or decide on new data — as opposed to the training step.The inspection camera runs inference every time a part passes the sensor.
IP Ratingมาตรฐานกันฝุ่นกันน้ำTwo digits saying how well a device is protected against dust and water. Higher numbers survive harsher environments.Choose a higher dust and water rating for a position that faces constant humidity and dust.
Latencyเวลาหน่วงThe time from the system receiving data to it answering. The lower it is, the better it suits work that has to be decided on the spot.The system has to decide before the part travels past the reject point on the conveyor.
LLMโมเดลภาษาขนาดใหญ่An AI program that has learned from a very large amount of text, to the point where it can understand and answer in human language.Staff type a question in plain Thai and get an answer back, with no special commands to learn.
MESระบบบริหารการผลิตหน้างานSoftware that tracks and controls production in real time, sitting between the planning system (ERP) and the machines themselves.Quality results go into the MES automatically instead of being written down by hand.
Modbusโปรโตคอลสื่อสารเครื่องจักรรุ่นเก่าThe communication standard older industrial machinery uses very widely — simple, and extremely compatible with legacy equipment.Pull the temperature reading out of an old PLC that has no communication port other than Modbus.
ModelโมเดลThe file that comes out of training, holding what the AI learned and ready to be used for decisions.The defect-detection model is updated periodically as the parts being made change.
MQTTโปรโตคอลส่งข้อความน้ำหนักเบาA publish-subscribe way of moving data, lightweight and well suited to IoT devices with limited resources.A sensor publishes vibration readings over MQTT at intervals, without waiting to be asked.
OCRอ่านตัวอักษรจากภาพTurning the text in a photo or scan into characters a computer can search and edit.Photograph a delivery note and the purchase-order number is pulled out automatically — no typing.
OEEประสิทธิผลโดยรวมของเครื่องจักรA single number combining availability, speed and quality for one machine.Hand-recorded figures are usually higher than the ones measured by actual sensors.
On-Premiseติดตั้งในสถานที่ลูกค้าThe system runs on machines that sit inside the client's own plant, not on another company's servers somewhere far away.Camera footage never leaves the plant because it is processed by a machine in the same server room.
ONNXรูปแบบไฟล์โมเดลกลางA file standard that lets an AI model move between devices and software from different vendors without retraining.Export the model to ONNX before running it on hardware from a different vendor than it was trained on.
OPC UAโปรโตคอลสื่อสารเครื่องจักรรุ่นใหม่A modern communication standard that is more secure and describes what the data means in itself, rather than just raw numbers.Pull data from newer machinery over OPC UA and get both the value and its name.
PLCคอมพิวเตอร์ควบคุมเครื่องจักรThe industrial computer that controls machinery according to a program you set. It is the brain of most production lines.The system reads values from the existing PLC without changing the machine's control program.
PoEจ่ายไฟผ่านสาย LANTechnology that carries both power and data over a single LAN cable, cutting the number of runs on site.Add a camera at a new point with one LAN cable — no need to find a power outlet nearby.
QuantizationการลดขนาดตัวเลขA technique that reduces the numeric precision the model computes with, making it smaller and faster in exchange for slightly lower accuracy.Convert the model to lower precision before deploying it on an edge device with limited memory.
RAGการตอบโดยอ้างเอกสารThe system finds the relevant documents first, then has the model compose an answer from those documents.Ask for a press machine's pressure setting and get the answer with the page number in the manual.
Telecentric Lensเลนส์ไม่บิดเบือนมุมมองA special lens that keeps an object the same size in the image even when its distance from the camera varies slightly. Suited to precise dimensional measurement.Used to measure part diameter when the position on the conveyor is not exactly repeatable.
TensorRTตัวเร่งความเร็วของ NVIDIANVIDIA's tool for optimising a model so it runs faster specifically on NVIDIA chips.Converting the model through TensorRT before production on a Jetson device makes a clear difference.
TOPSล้านล้านการคำนวณต่อวินาทีA unit for the raw power of an AI chip. Useful for comparing models, but it does not tell you real-world speed.A chip with higher TOPS is not automatically faster — it depends on how well your job actually uses those units.
TrainingการฝึกสอนโมเดลThe step where you feed the model a large number of examples so it learns the patterns before it goes into service.Thousands of images of genuinely defective parts are used to train the defect-detection model.
Vector Databaseฐานข้อมูลเวกเตอร์A database built to search for similarity of meaning quickly, rather than only exact word matches like an ordinary database.Find documents that are close to the question even when not a single shared word appears.