دانلود کتاب TinyML کاربردی -یادگیری ماشین سرتاسری برای میکروکنترلرها با مثال‌ها

The idea for this book was born from my fascination with the intersection of machine learning and embedded systems. In the last few years, I have seen how AI has transformed industries. Many of the most powerful applications remain locked within cloud servers and large computing infrastructures. My goal with this book is to help bridge that gap by showing readers that it is possible to bring intelligence to edge devices and make TinyML accessible to engineers, makers, and innovators who want to deploy machine learning in real world, low power environments. TinyML is the latest frontier for AI and ML models, as its constraints are everywhere. Memory is scarce, and computing is smaller by orders of magnitude compared to traditional computing used to train and run the models. Therein, however, lies one of the greatest opportunity windows in many years. When you think about the future, you do not imagine just text editors that help you write a letter but an interactive physical world, appliances and stand-alone devices transacting with people in the most natural way, friendly furniture, smart clothing, and personalized gadgets. All of that is the great promise of TinyML. It is our once-in-a-lifetime opportunity to not only live in the future but to create it. Another important aspect about this book is its breadth. The principles of each concept are explained lightly; however the variety of domains covered is wide. Instead of focusing solely on the data science aspect of the TinyML models, we include system design (requirements are mapped to an hypothetical business problem that needs to be solved), data acquisition challenges (rather than just assuming the existence of a ready made data set available on the internet), feature extraction (which is directly impacted by data quality, a real issue in production environments), model design (understanding how the model works under the hood is critical in constrained environments where resource waste is not an option), electronic component configuration (using existing electronic components instead of just running in simulators), networking (since real world solutions are often distributed across a surface rather than confined to a single location), power considerations (as connection to the electric grid cannot be assumed), and bill of materials research (because cost is a fundamental metric in real-life systems). Chapter 1: Foundation and Methodology – This chapter walks the reader through the concepts and terminology used throughout the book. It also established the scope of each one of the use cases in this book. Chapter 2: Sound Classification – This chapter focuses on identifying the features that define sounds according to their specific use cases. We utilize Mel Frequency Cepstral Coefficients (MFCC) for human voices, which excel in recognizing linguistic attributes such as inflection rhythm and silences. Fourier Coefficients are employed for transient sounds to delineate the sound by its power distribution across frequencies, which is essential for capturing the unique energy profile of each sound. Periodic tones are analyzed through spectrum analysis to uncover their distinctive frequency signatures. We employ Convolutional Neural Networks to leverage the spatial representation of sound (across time and magnitude) for pattern recognition, training the model to convert the output layer into a probability distribution used to classify the sound sample. Chapter 3: Movement Classification – In exploring movement classification, we gather data from three orthogonally placed accelerometers to accurately describe object movement. By breaking down the accelerometer readings into thirteen distinct features and segmenting the wave into small time windows, we calculate metrics such as root mean square, kurtosis, spectral kurtosis, and skewness. These metrics reveal critical aspects of movement, from the average power indicating the movement’s energy to kurtosis and skewness, offering insights into the frequency and suddenness of movements. The rich dataset feeds into a dense neural network, enabling precise classification of movements ranging from gentle to vigorous. Chapter 4: Image Classification – Focusing on image classification, we acquire images through a direct connection to a digital camera interfaced with a microcontroller. Preprocessing steps include adjusting the image to a square format and, if color is not critical, converting it to 8-bit grayscale to streamline object recognition models’ application. Images are resized to a standard size to accommodate the neural network’s capacity. Rather than training from scratch, we apply industry standard TinyML models and a technique called transfer learning to efficiently train the model with our dataset, achieving a fast and accurate system applicable across various machine vision scenarios. Chapter 5: Object Tracking – This chapter introduces tracking, a technique often paired with Image Classification to consistently identify an object across sequential video frames. Whether tracking moving or static objects, the method involves first classifying objects with a Convolutional Neural Network and then comparing their sequential positions to deduce if they are the same based on proximity. Successful tracking assigns consistent IDs to objects, maintaining identity even through temporary classification lapses or visual obstructions, showcasing the robustness of a good tracking algorithm. Chapter 6: Sensor Fusion – Sensor Fusion is examined as an advanced method of combining data from multiple sources to generate new insights. We categorize sensor fusion into complementary, competitive, and cooperative types. Complementary fusion brings together different data perspectives of the same event, while competitive fusion seeks reliability through redundancy in data acquisition. Cooperative fusion merges data from unrelated events to synthesize new information, highlighting the technique’s versatility in enhancing system perception. Chapter 7: Deep Learning Regression – This chapter explores using deep neural networks to infer numerical values from one or more data sources. We demonstrate how integrating feature engineering with deep learning can expand the realm of regression applications. We discuss developing control systems using regression models with multiple inputs and outputs that apply continuous learning and adaptation to current conditions. This chapter illustrates how regression can forecast future conditions based on historical data, opening new possibilities for predictive analytics. Chapter 8: Anomaly Detection – This chapter examines the anomaly detection method, which aims to identify occurrences that deviate from expected patterns. Distinguishing itself from classification, anomaly detection focuses on detecting behaviors or events previously unseen by the model.