Artificial Intelligence is revolutionizing all aspects of society nowadays, for example, by combining the benefits of data mining and deep learning, artificial intelligence can now be used to analyze large amounts of data from a variety of sources, identify patterns, provide interactive understanding and intelligence An example of this innovative development is the application of artificial intelligence to sensor-generated data, especially from smartphones and other consumer devices.Motion sensor data and other information such as GPS addresses can provide a large number of differences So, the question is: 'How to use artificial intelligence to get the most out of these synergies?'
Exercise data analysis
An illustrative real-world application will be able to analyze the usage data to determine the user's activity at each time period, whether sitting, walking, running or sleeping.
In this case, the benefits of smart products are self-evident:
1. Increase Customer Lifecycle Value Increasing customer engagement reduces customer churn 2. More competitive product positioning The next generation of smart products meets the growing expectations of consumers 3. Creating real value for end users
Accurate detection and analysis of indoor motion enables smart navigation, health risk monitoring, and increased device efficiency Depth mastery of the actual usage scenarios for many smartphones and wearable platforms will greatly assist product designers Understand users' repeated habits and behaviors, such as determining the correct battery size or determining the correct timing for pushing notifications.
Interest in artificial intelligence features is growing among smartphone manufacturers, highlighting the importance of identifying simple daily activities, such as number of steps, which will surely develop into more in-depth analyzes such as sports activities. Popular sports, product designers will not just focus on athletes, but will facilitate more people, such as coaches, fans and even broadcasters such as broadcasters and sportswear design companies.These companies will be from the deep data Benefit from the analysis, which can accurately quantify, improve and predict performance.
Data acquisition and preprocessing
After identifying this business opportunity, the next logical step is to think about how to effectively gather these huge data sets.
In activity tracking, for example, raw data is collected by axial motion sensors such as accelerometers and gyros in smartphones, wearables, and other portable devices that take three axes (x, y, z), ie tracking and evaluating activities continuously in a way that is convenient for the user's application.
Training model
For supervised learning of artificial intelligence, there is a need to train 'models' with markup data so that the classification engine can use this model to classify the actual user behavior.For example, we collect exercise from test users who are running or walking Data, and to provide this information to the model to help them learn.
Since this is basically a one-time method, a simple application and camera system can accomplish the task of labeling users, and our experience shows that as the number of samples increases, the classification of human error rates follows So getting more sample sets from a limited number of users makes more sense than getting smaller sample sets from a larger number of users.
It is not enough to get raw sensor data alone and we observe that to achieve a highly accurate classification we need to carefully determine the features that the system needs to be told about the features or activities that are important in distinguishing each sequence.The artificial learning process is repetitive, It is not yet clear which features are most important in the preconditioning phase, so the device must make some guesswork based on the expertise that can have an impact on the accuracy of the classification.
The indicative features may include 'filtered signals' such as body acceleration (raw acceleration data from the sensor) or 'derived signals' such as Fast Fourier Transform (FFT) values or standard deviation calculations for activity recognition.
For example, the University of California Irvine's Machine Learning Database (UCI) created a dataset that defines 561 features based on six basic activities of 30 volunteers, standing, sitting, lying, walking , The next step and the step as the basis.
Pattern recognition and classification
After collecting the raw exercise data, we need to apply machine learning techniques to classify and analyze the machine learning techniques available to us from logistic regression to neural networks, etc. Support Vector Machines (SVMs) is such a Learning models applied to artificial intelligence Physical activity, such as walking, consists of a sequence of multiple motions, which is a reasonable choice for classifying activities since it is good at sequence classification.
The use, training, expansion, and prediction of support vector machines are straightforward, so multiple sample collection experiments can be easily juxtaposed for non-linear classification of complex real-life data sets. Support vector machines also enable a wide range of Different size and performance optimization.
After identifying a technique, we have to choose a software library for SVM. Open source library LibSVM is a good choice because it is very stable and has a detailed record, supports multiple classifications, and provides all major developer platforms from MATLAB to Android expansion.
Continued classification of the challenge
In practice, while users are moving, the devices in use need real-time classification for event recognition.In order to minimize the cost of the product, we need to balance the transmission, storage and usage without affecting the result, ie the quality of the information The cost of processing.
Assuming that we can afford the cost of data transfer, all the data is stored and processed in the cloud - in fact, this can result in huge data costs for the user, and of course the user's device is connected to the internet, wireless network, Bluetooth or 4G module Of the costs will inevitably further increase equipment costs.What's worse, in non-urban areas, 3G network access is usually not effective, such as hiking, cycling or swimming.This dependence on the cloud of large amounts of data transfer Slows updates and periodically synchronizes, greatly offsetting the real benefits of Artificial Intelligence motion analytics In contrast, handling these operations only on the device's main processor can significantly lead to increased power consumption and decrease Other application execution cycles Similarly, storing all the data on the device increases storage costs.
Round to square
To solve these conflicting issues, we can follow four principles: 1. Splitting - splitting feature processing from the execution of the classification engine.
2. Reduce - intelligently selects the features needed for accurate activity identification to reduce the need for storage and processing.
3. Use - The sensors used must be capable of acquiring data at lower power consumption, implementing sensor fusion (combining data from multiple sensors), and feature preprocessing for continuous execution.
4. Reserved-A model that retains system-enabling data that can be used to determine user activity By splitting the execution of the feature processing and classification engine, the number of processors connected to acceleration and gyro sensors can be much smaller, effectively avoiding Real-time data blocks are continuously transmitted to more powerful processors.Feature processing such as Fast Fourier Transform, which is used to transform time-domain signals to frequency domain signals, will require a low-power core processor to perform floating-point operations .
In addition, in the real world, a single sensor has physical limitations and its output deviates over time, for example due to offset and non-linear scaling caused by soldering and temperature.To compensate for this irregularity requires sensor fusion and fast , Inline and automatic calibration.
In addition, the chosen data capture rate can significantly affect the amount of calculations and throughput required.Often, a 50 Hz sample rate is sufficient for normal human activity, but 200 for analyzing fast-moving activities or motions Hz sampling rate Similarly, to achieve faster response time, you can install a separate 2 kHz accelerometer to determine the purpose of the user.
To meet these challenges, low-power or application-specific sensor hubs can significantly reduce the CPU cycles required by the classification engine, such as Bosch Sensortec's BHI160 and BNO055, which are hubs of this type, which can be used directly at different sensor data rates Generate fused sensor output directly.
The initial selection of features to be processed subsequently greatly affects the size of the training model, the amount of data, and the computational power required to train and execute the inline prediction.Therefore, it is crucial to choose the features required for classification and differentiation of a particular activity Decisions, but also very likely to be an important commercial advantage.
Recalling the UCI machine learning database we mentioned above, which has a complete dataset of 561 features and uses the default LibSVM kernel-trained model for activity classification testing up to 91.84% accuracy, however, after completing training and feature rankings, The most important 19 features were selected to be 85.38% active classification test accuracy.A closer examination of the rankings revealed that the most relevant features were the average, maximum and minimum of the raw data for frequency domain transformation and sliding window acceleration Interestingly, none of these features can be achieved with preprocessing alone, and sensor fusion is necessary to ensure adequate reliability of the data and is therefore of particular use for classification.
in conclusion
All in all, advances in technology have now reached the point where high-level artificial intelligence is run on portable devices to analyze the data from motion sensors that operate at low power while sensor fusion and software partitioning significantly improve the overall system's efficiency and viability , But also greatly simplifies application development.
To complement the sensor's infrastructure, we leverage open source libraries and best practices to optimize feature extraction and classification.
It has become a reality to provide users with a truly personal experience, through which artificial intelligence enables the system to take advantage of the data collected by the sensors of smartphones, wearables and other portable devices to provide people with more depth capabilities. In the coming years, a series The devices and solutions that are now unimaginable will grow even further.Artificial intelligence and sensors open up a new world of exciting opportunities for designers and users alike.
Source: UCIhttp: //archive.ics.uci.edu/ml/datasets/Smartphone-Based+Recognition+of+Human+Activities+and+Postural+Transitions About Bosch Sensortec
Bosch Sensortec GmbH, a wholly owned subsidiary of Robert Bosch GmbH, develops and offers custom MEMS sensors and solutions for smartphones, tablets, wearable devices and IoT products with a portfolio of 3 axes Accelerometers, gyroscopes and geomagnetic sensors, integrated 6-axis and 9-axis sensors, environmental sensors and a comprehensive portfolio of software, Bosch Sensortec has been a MEMS technology leader in these markets since its establishment in 2005. Bosch has operated since 1995 Has been a pioneer in the field of MEMS sensors and a leader in the global market with more than 8 billion MEMS sensors sold to date.The Bosch Sensortec sensor is used by every two smartphones in the world.
For more information, visit www.bosch-sensortec.com and follow us on WeChat public number 'BoschSensortec'.