In the era of the Internet of Things (IoT), the distributed nature of sensor architectures and the need for device networking integration are driving the evolution of sensor systems towards intelligence. Sensors within the system typically use analog or digital serial interfaces to transmit data to a host microcontroller or microprocessor.
Data preprocessing or filtering operations are completed on the host. To access wireless or wired networks, IoT devices often incorporate built-in microcontrollers to manage network access. This processing core provides additional computational power for handling secure transmission, data preprocessing, and filtering functions, enabling IoT-compatible sensors to be upgraded to intelligent sensors.
Advanced Architecture of Smart Room Applications
In IoT applications, data filtering functions offer significant advantages when available bandwidth or energy is limited. Although local processing consumes a certain amount of energy, it is generally more advantageous than wirelessly transmitting all sampled data by limiting the amount of data transmitted. Another benefit of filtering is that it can reduce network load, which is of significant value in mesh network topologies. For sensors connected via LPWAN, the cost of data transmission also directly influences the choice of the extent of preprocessing by intelligent sensors.
There are multiple ways to implement data filtering mechanisms. One common technique involves using thresholds to evaluate the magnitude of changes in received data. All remote systems cache the last received value, assuming the data remains unchanged unless the input value breaks the threshold range or differs from the cached value, triggering a data update and transmission operation.
Filtering functions can distinguish between changes that require immediate processing and those that only require model updates but do not need real-time transmission. This can be achieved using another set of thresholds or local data models to determine whether the input falls outside the range. Updates that do not require real-time forwarding can be temporarily stored in a buffer and subsequently sent in a batch with subsequent measurement data.
The utilization of compression techniques such as linear predictive coding (LPC) can further enhance network bandwidth utilization. The change values of the data to be transmitted are usually relatively close, and LPC leverages this characteristic to reduce the number of bits required per sample.
2D and 3D sensors process significantly larger volumes of data compared to one-dimensional measurements such as temperature or pressure, resulting in higher data management complexity. Take security cameras as an example; they may integrate basic AI models or algorithms internally to detect changes in the scene frame by frame. Some minor changes may be ignored. When larger changes are detected, such as the entry of a person or vehicle into the surveillance area, the system automatically filters and transmits key regions of the scene to a remote platform. By employing compression techniques and selectively transmitting regions of interest (rather than full frames), network bandwidth usage can be significantly reduced.
Additionally, sensors can be configured to adapt to multiple remote systems and dynamically adjust transmission strategies based on the requirements of different systems. Some intelligent sensors are built-in with support for common industrial protocols (e.g., Modbus) as well as IoT protocols (e.g., CoAP or MQTT). These sensors determine which remote node will receive the corresponding data format based on incoming requests. If network bandwidth requirements or sensor functionality necessitate the use of a single protocol, a gateway can convert protocols in real-time, for example, forwarding Modbus packets to a nearby PLC and distributing CoAP or MQTT packets to other systems subscribed to the corresponding data sources.
Another advantage of intelligent sensors is their support for secure communication, combined with simplified installation functions. The current trend indicates that intelligent sensors are widely adopting factory configurations with pre-installed digital certificates and private keys (stored in encrypted memory). Some network protocols (e.g., LoRaWAN) have already integrated these functions into their systems. When connecting to the network, sensors can use these credentials to establish a secure connection with the server using standard public-key encryption technology.
Sensors utilize stored credentials to identify legitimate servers, while also facilitating server verification of sensor validity. Only after establishing a connection can sensors obtain full network access permissions. The system identifies and rejects access from cloned or counterfeit devices. Since cloud-based identity verification systems can individually identify each device in the network based on security credentials, integrating these functions can significantly simplify the installation process. Installers do not need to manually program device IDs and other information, as most necessary information is already encoded during manufacturing. If the sensor is equipped with positioning hardware (such as GPS or similar systems), the module can even automatically determine its location. If not, installers or remote operators can add location and other metadata to the device and server databases after the sensor is up and running.
After establishing a connection and passing verification, secure intelligent sensors can further enhance information protection by encrypting data packet payloads. Symmetric cryptographic algorithms (such as AES256) are often used for payload encryption due to their lower processing overhead compared to public-key systems. However, depending on the system architecture and sensor module performance, public-key encryption may be more suitable. Intelligent sensors can adopt differentiated encryption strategies, assigning independent keys to different users to ensure that remote devices can only decrypt data within their permission scope. However, the system architecture may determine that edge gateways or cloud servers handle such security controls, and there are multiple combination schemes available.
Sensor Fusion Enables System Intelligence Upgrades
One of the core concepts of IoT is that the combined value of data from numerous different sensors far exceeds the sum of their individual parts. Network connectivity expands the scope of data collection, enabling the fusion of multiple sensor modalities to jointly drive data models or algorithms. By integrating different types of measurement data, the system can more accurately determine whether input signals are erroneous due to hardware failures or blockages from dirt. By eliminating errors in individual readings, the model supports making better decisions.
The application of sensor fusion algorithms enables the collaborative integration of sensor readings. Some algorithms will adopt widely compatible sensor formats. Sensor fusion technology is now widely used in mobile devices. For example, the sensor hub integrated in smartphones significantly improves the quality of applications such as gait analysis and navigation by integrating data from gyroscopes and accelerometers. Different sensors can compensate for each other. The primary source of error for gyroscopes is drift. Integrating accelerometer data can effectively compensate for gyroscope drift errors, while gyroscopes can help overcome the susceptibility of accelerometers to sensor noise. The output data processed by the sensor hub can more accurately represent linear motion and rotational changes such as roll, pitch, and yaw.
The current 360° panoramic views in some advanced automotive systems are generated by fusing data from multiple cameras. Other systems use a variety of sensors to construct system models. For example, combining acoustic and vibration sensors significantly improves the accuracy of damage detection systems for motors and other mechanical equipment. Time-of-flight (ToF) cameras combined with temperature, carbon dioxide, and other environmental sensors can assist in determining whether the air conditioning in a room or auditorium needs adjustment.
There are currently multiple effective sensor fusion techniques. The Kalman filter, commonly used in motion perception systems, assigns higher weights to readings with low uncertainty. The filter state is represented as a set of matrices that integrate readings from different types of sensors into a unified coordinate model. The filter operates in two stages: prediction and update. The prediction stage estimates the next state based on the system’s historical state. The update stage compares the new sensor sampling values with the predicted values. The closer the input value is to the predicted value, the lower the error probability. A poor match reduces the weight of the new reading from that sensor.
Although particle filters require more processing time than Kalman filters, they remain an efficient solution when the data model is highly nonlinear (beyond the typical applicable scenarios of Kalman filters). These filters use techniques such as Bayesian formulas to fuse input readings in a probabilistic manner.
Probabilistic methods further give rise to advanced sensor fusion techniques based on machine learning. Machine learning is particularly suitable for systems that require the fusion of multidimensional data, such as combining 2D images, videos, and 3D point cloud data from ToF cameras and LiDAR instruments. A deep learning pipeline that combines multi-channel convolutional layers with pooling layers can construct a model framework that supports unified training of multimodal data.
One of the key aspects of sensor fusion is ensuring data element alignment through preprocessing. When some sensors only intermittently transmit data changes while others continuously transmit, the receiving system needs to align and fill in data values to ensure that the model receives consistent update values. For example, if a remote sensor does not send an indication of a state change, repeated data values may need to be input to the model. Similarly, updates sent in groups need to be matched with the timestamps of other data streams to ensure consistent sampling times. Such functions can be implemented by gateway modules or terminal systems, provided that these systems have been pre-programmed with the ability to parse the received data.
OMAGINE specializing in ODM PCB design, PCB assembly, open source hardware related modules and sourcing service.
