Firmware Debugging & Performance Optimization Services

Phase 01 · Context

Client and Product

Client: Industrial Manufacturing and Quality-Control Client
Product: AI-Powered Smart Camera for Production-Line Inspection

Adequate Infosoft created a bespoke smart camera system using AI to solve the problem of automated quality inspection for its client working in an industrial manufacturing environment.

This involved accurately detecting product defects, responding to real-time events sent from a PLC, storing the results of the analysis, and functioning in an industrial environment.

The development included creating custom hardware, writing NXP i.MX RT1060 embedded firmware, implementing the Coral Edge TPU for performing AI inference, using industrial cameras, establishing communication with the cloud, and Modbus TCP communication with factory PLCs.

Phase 02 · Optimization

Firmware Performance and Deterministic Processing

Firmware performance has been essential from day one of the development. Even minor delays caused by image acquisition, inference, networking, or PLC communication could make it impossible to detect passing-by products.

Our engineers analysed the capture-to-decision path and optimized the firmware architecture around deterministic processing. A double-buffered image-capture approach allowed the system to prepare the next frame while the current frame was being processed. This reduced blocking between image acquisition and AI inference, helping the solution maintain a production throughput of more than 120 units per minute without frame drops.

Firmware tuning was done to ensure that the trigger-to-image delay was kept below 50 milliseconds. AI inference with the YOLOv8-tiny on the Coral TPU was done in less than 30 milliseconds per image. This provided enough time for the production line to perform communication and decision making during each inspection.

Phase 03 · Debugging

Thermal and False-Positive Debugging

Debugging was not only done on the source code. During the environmental test phase, it was discovered that the Coral TPU had a thermal throttling problem at an ambient temperature of about 50 degrees centigrade.

To mitigate this problem, the team introduced enclosure level thermal tuning such as thermal pad and forced air cooling.

False positives due to reflective or shiny product surfaces were also addressed by using a combination of firmware tuning, polarized optics, and different light angles.

The solution included firmware-health features such as a hardware watchdog with a five-second timeout, device temperature monitoring, network-health checks, frame-drop detection, and diagnostic reporting.

We also designed the OTA process with dual-bank flash, digital signature verification, and rollback timing to protect devices from a corrupted or unsuccessful update.

Phase 04 · Outcome

Result

The installed system has defect detection accuracy of 99.7%, throughput greater than 120 units per minute, and it has been implemented in three different manufacturing plants.

This project shows that we can debug and optimize firmware code in the environment that counts, hardware, thermal constraints, AI computation, industrial networking, and manufacturing line performance.

Method 01

Root Cause Analysis and Firmware Fix-Up

If it is due to an inherited or broken firmware, we begin with a good baseline analysis first. We identify build warning messages, undefined behavior, race conditions, pointer manipulations, blocking calls, time dependency, memory leaks and poor error handling.

We add useful logging messages wherever possible, without filling limited memory and interfering with the real-time process. In case of a hard fault, we analyze CPU registers, stack trace, fault status register, PC and other events that led to the exact sequence of the fault.

Such root cause analysis methodology is extremely helpful for the problem which is identified only after a lot of operation of the device or high network traffic or when operating under low battery voltage or multiple peripheral operations.

Method 02

Real-Time Performance and Latency Optimization

Optimizing performance is not just about optimizing clock speed. An MCU with fast clock speed might underperform if tasks are not scheduled properly, if interrupt services take too much time, if there is too much data copying, and if communication gets in the way of control.

We increase real-time performance by:

• Reviewing and optimizing RTOS task priorities and stack size
• Optimal interrupt services routine development
• Direct memory access-based transfers of data for sensors, UART, SPI, ADC and display pipeline
• Event-driven design rather than useless polling
• Circular buffer and effective queues management
• Memory alignment and buffer reuse, and minimizing data copying
• Profiling CPU usage, execution time and jitter
• Simplifying firmware start up and reducing boot time initialization
• Proper use of compiler optimizations without reducing reliability

In the case of connected devices, we also optimize the interaction of firmware and network. The device may be slow due to re-transmissions, bad packet sizes, blocked requests to the cloud, ineffective Bluetooth Low Energy connection settings and unneeded telemetry messages. We identify the exact bottleneck before any changes to architecture.

Method 03

Battery and Power Consumption Optimization

Energy consumption is an important criterion for wearables, sensors, trackers, smart locks, and battery-operated IoT devices. Even though there may be good functionalities of a product, it may not be a success in the market if it requires charging or battery replacement frequently.

This involves optimizing a number of factors that include sleep state design, waking up source considerations, shutting off peripherals, clock management, radio duty cycling, sensor sampling, BLE advertising/connections parameters, and minimizing cloud communication.

We measure the real current profile rather than relying only on firmware assumptions. Short high-current peaks can be as important as average power consumption, particularly for coin-cell devices and products with cellular, motor, or wireless transmission bursts.