IN Brief:
- AKM CZ39 and CZ3K current sensors have been adopted in Microchip's dsPIC33A machine-learning arc-fault reference design.
- Both series specify 100ns response, with measurement ranges reaching ±154A for CZ39 and ±300A for CZ3K.
- Signal processing and machine-learning inference run locally on the dsPIC33A without an external processing resource.
Asahi Kasei Microdevices has confirmed that its CZ39 and CZ3K coreless current sensors are being used in Microchip Technology’s machine-learning arc-fault detection reference design. The platform combines fast current sensing with a dsPIC33A digital signal controller that performs signal processing and machine-learning inference locally rather than passing the protection decision to a separate compute resource.
Arc-fault detection is difficult because normal electrical equipment can generate transient signatures that resemble a dangerous event. Motor brushes, switch contacts, relay bounce, inverter switching, capacitor inrush, and other legitimate activity can introduce broadband disturbances, forcing protection designers to balance sensitivity against nuisance trips.
The problem differs between AC and DC systems but does not disappear in either. In solar PV, battery-storage, and EV-charging systems, switching transients can occupy similar frequency bands to genuine arcs. DC architectures also lack the regular current zero crossing found in AC supplies, which can allow an established arc to persist.
Microchip’s reference platform runs the classifier on a dsPIC33A DSC. The device combines real-time signal processing with analogue peripherals and executes inference directly at the edge, keeping the detection loop close to the current-sensing hardware and avoiding dependence on an external processor or communications link.
The quality of the input waveform remains critical regardless of the classifier. AKM specifies a 100ns response time for both the CZ39 and CZ3K sensor families, allowing rapid current changes to reach the signal-processing chain before useful high-frequency information is lost.
The two families cover different current ranges. CZ39 devices measure up to ±154A, while CZ3K extends the range to ±300A. That allows the reference concept to span several power levels, although final sensor choice will still depend on the actual conductor geometry, isolation requirements, current range, thermal conditions, bandwidth, and analogue input arrangement.
AKM introduced the CZ3K family for high-current EV power systems earlier this year, highlighting its ±300A range, low-resistance conductor structure, and 100ns response. Its use in the new arc-fault platform gives that sensing hardware another application beyond onboard chargers, DC/DC converters, and electronic fuses.
The machine-learning element should not be treated as a guarantee of better protection under every condition. Detection accuracy depends on sensor noise and bandwidth, signal conditioning, feature extraction, the training data used for the model, operating-state coverage, and the threshold applied to the inference result. A model trained on an incomplete set of normal transients can still produce unwanted trips when it encounters an unfamiliar load.
That makes the reference design a development platform rather than a universal arc-fault detector. Engineers still have to gather representative data from the equipment being protected, validate the model across normal and abnormal operation, and demonstrate that the complete protection function meets the applicable electrical-safety and response requirements.
Target applications identified for the design include solar PV, energy-storage systems, EV chargers, smart ignition, e-Fuses, and residential and industrial safety switches. The variety of those loads reinforces why the sensing and inference stages have to be adapted to a specific electrical environment rather than copied unchanged between products.
The architecture nevertheless provides a practical starting point. Fast current sensing preserves more of the arc signature, while the dsPIC33A keeps the classification local and gives developers access to the signal-processing chain rather than hiding it behind a cloud service or separate AI accelerator.
For protection engineers, the next work sits in validation rather than demonstration: collecting difficult edge cases, quantifying false-trigger behaviour, proving detection under real switching noise, and confirming that sensor and classifier performance remain stable across temperature, component ageing, and changes in the connected load.
