A quarter-turn actuator on a process valve has one job. It has to say where the disc actually is. When it hunts around setpoint, or reports a travel position the valve is not at, the first instinct is mechanical: backlash, a loose coupling, a stem sticking under differential pressure. Then the unit repeats to a tenth of a degree on the bench and the theory dies.
Most of those returns are a serial frame decoded as data when it was never data. The MLX90367EGO-ACU-000 is a Melexis Triaxis Hall position sensor with a SENT output built to SAE J2716-2010, 12 bit resolution and a dual die TSSOP-16 package. It senses all three flux components, so the magnet can sit on the shaft axis or beside it. It also emits four frames after every power-up that carry no angle, and on request it puts a diagnostic value where the angle used to be.
Three numbers decide the outcome: time to a valid angle, latency from magnet movement to frame, and the flux the front end absorbs.
What the failure looks like in the field
Three shapes, surfacing in different places, which is why they are rarely connected to one another.
Full travel after a power dip. A mains event restarts controller and sensor together. For a few milliseconds the output sits in a defined idle state while the device emits four initialization frames, each a legal SENT frame with a valid checksum and six zero nibbles. A decoder that latches the first frame it can checksum reads zero and commands the valve closed. The event log shows a normal start-up.
Hunting near the stroke ends. Stable across the middle of travel, unstable in the last few degrees at either end. Not a tuning problem. The position reading itself goes flat: mechanical tolerance on the magnet stack pushes the field past what the front end can linearise, and the reported angle stops following the disc.
A frozen position with no fault on the bus. The actuator holds a plausible, completely wrong position and nothing reports it. The frame is fresh and the checksum is good, which is why it survives a maintenance window untouched.
The actuator never crashes. It boots, it runs, and the number the controller closes its loop on is wrong. That is why these units pass end-of-line test and still come back.
Why it happens at the silicon level
Four datasheet facts cover every shape above, and each points at a single number.
| Field shape | What is happening | Number that governs it | Source |
|---|---|---|---|
| Slams closed after a dip | Initialization frames decoded as angle | Tsu1 1.8 ms, Tsu2 5.9 ms to 6.3 ms | p. 8, p. 20 |
| Angle flattens at stroke end | Front end saturates, or signal drops into noise | 70 mT on BX/BY, 126 mT on BZ, 20 mT norm floor | p. 12 |
| Hunting around setpoint | Loop closes faster than the frame refreshes | 1745 µs latency, 882 µs frame period | p. 8 |
| Plausible position, no fault | Fast channel carries a diagnostic instead of the angle | 4088 plus the error value, 12 bit payload | p. 17 |
The first five frames are not all angles. The output stays high until the circuit has emitted four initialization frames in which all six data nibbles are zero; the fifth is the first valid measured angle (p. 20). The first sync pulse arrives 1.8 ms after start-up, the first data frame 5.9 ms to 6.3 ms later (p. 8). A parser starting at the first sync pulse has three or four well-formed frames to skip, and nothing in the protocol marks them as idle.
Channel 1 is not always the angle. By default the fast channel carries the 12 bit angle. With the EEPROM parameter SERIALERROR above zero the channel 1 payload becomes 4088 plus the error value, and the same condition is flagged on the status bit (p. 17). Firmware that never gates on that bit reads a diagnostic as a position and drives there.
The magnetic front end has a ceiling and a floor. The Triaxis front end is characterised to 70 mT on BX and BY and 126 mT on BZ, with a flux norm floor of 20 mT (p. 12). Above 70 mT the integrated magneto-concentrator saturates and linearity error grows; below 20 mT the signal-to-noise and signal-to-offset ratios degrade. A valve actuator has a fixed magnet and a shaft floating in its bearings, so the field at the sensing spot moves with mechanical tolerance and temperature. A design landing near 70 mT at one end of travel looks perfect on a bench and goes flat in the field.
Latency is part of the loop. With the recommended FILTER = 1 setting, the average delay from a field change to the corresponding SENT data is 1745 µs and the frame period is 882 µs (p. 8). A loop closing faster than that is controlling a stale number, not the valve.
Calibration storage is not a log. The EEPROM is qualified for a minimum of 1000 write cycles at 125 °C; the datasheet calls it calibration data storage, similar to an OTPROM (p. 27). Written once at end of line it lasts the life of the product. Re-trimmed on every power cycle it is gone inside the warranty period, and the symptom is a checksum fault rather than a drift.
The numbers that fix it
All values come from the MLX90367 datasheet, revision 7.0 of 8 February 2019.
| Parameter | Value | Source |
|---|---|---|
| Ordering suffix of this part | E grade, GO package, ACU 3-point option, reel | p. 4 |
| Temperature grades available | E -40 to 85 °C, K -40 to 125 °C, L -40 to 150 °C | p. 4 |
| Resolution and thermal accuracy | 12 bit, 10 bit | p. 1 |
| Supply voltage VDD | 4.5 V to 5.5 V | p. 7 |
| Supply current | 6 mA typical, 10 mA maximum with power saving | p. 7 |
| Start-up level on VDD | 3.8 V to 4.2 V rising, 50 mV to 200 mV hysteresis | p. 7 |
| Isolation between dice | 4 MΩ minimum, dual die only | p. 8 |
| Frame and measurement period | 882 µs, 441 µs | p. 8 |
| Latency, field change to data | 1745 µs average | p. 8 |
| Start-up to first data frame | 5.9 ms to 6.3 ms | p. 8 |
| Magnetic input range | 70 mT on BX and BY, 126 mT on BZ, 20 mT minimum norm | p. 12 |
| Intrinsic linearity | ±1 Deg. on XY, ±2.5 Deg. on XZ and YZ | p. 11 |
| CPU and memory | 2.5 MIPS, 10 KB ROM, 128 B EEPROM | p. 12 |
| Identification | 48 bit user ID, three freely usable words | p. 25 |
| EEPROM endurance | 1000 write cycles at 125 °C | p. 27 |
| Filter settings | 0, 1 and 2, with FILTER = 1 recommended | p. 26 |
| Recommended external parts | 100 nF on each supply pin, close to the pin | p. 30 |
Four rows decide the design: the 20 mT floor and 70 mT ceiling at the sensing spot, the start-up timings the decoder has to skip, the 1745 µs latency the loop must budget for, and the 1000-cycle write budget.
One note on the option code. This part is the ACU option, the standard version with the 3-point linear transfer characteristic (p. 22). The 17-point piece-wise-linear table and the thermal output offset correction live on the ACV and ACX options, so a design needing finer end-of-line calibration moves the suffix, not the firmware (p. 4, p. 24). The ABU, ABV and ABX suffixes on that page are marked not recommended for new designs.
One pinout note before you draw the symbol. The TSSOP-16 is two dice in one body, so supply, digital rail, ground and output each appear twice. The unused Test pins should be tied to ground for the best EMC behavior (p. 6).
What if the frame format itself is the problem?
If the SENT decode path is what you want to remove rather than fix, the same Triaxis family offers rotary-only parts with three other output styles. The MLX90316EDC-BCG-000 is the standard programmable SOIC-8 single die at the E grade. The MLX90316EDC-BDG-100 is the same die and grade over SPI, 14 bits over 360 degrees or 0.022 Deg. per LSB, which deletes the frame decoder from the ECU. The MLX90316EGO-BCG-000 and MLX90316EGO-BDG-100 move to TSSOP-16 with two dice and 4 MΩ of isolation, matching the redundancy of the part in this article. The MLX90316KDC-BCG-000 and MLX90316KDC-BDG-100 are the same two output styles at the K grade, good to 125 °C, which matters when the actuator sits next to a hot gearbox. The MLX90316KDC-BCG-200 and MLX90316KDC-BCG-300 are the pre-programmed analog and digital variants, arriving with a fixed transfer curve and no end-of-line trim step to own.
Three things change with the swap. None outputs SENT, so every timing assumption built on the 882 µs frame period is rewritten. None senses the perpendicular axis, so off-axis and linear stroke measurement disappears. And the start-up cycle is 15 ms typical, several times the MLX90367 timing.
Verification you can run before the next build
Six checks, most of them in one bench session.
- Count the null frames at power-up. Scope the SENT line for the first 10 ms after VDD passes 4.2 V. Four frames with all six nibbles zero, then a valid angle. Confirm your parser discards them.
- Force a diagnostic. Drive the field out of the gain window so the front end flags an error. If the actuator moves, your parser is reading the payload as an angle.
- Sweep the magnet across the mechanical extremes. Log the field amplitude at the sensing spot at both ends of travel and at the tolerance limits on shaft float, against the 20 mT floor and the 70 mT ceiling.
- Put the latency in the loop model. Add the 1745 µs field-to-data delay and the 882 µs frame period to the feedback path and re-run the step response.
- Budget the EEPROM writes. Count every firmware path that writes sensor memory over a service life and compare that with the 1000-cycle endurance figure.
- Inject a brownout. Hold VDD between the start-up level and the nominal rail during a stroke and confirm what the actuator does.
One boundary for the requirements document. The datasheet describes the product as intended for normal commercial applications and states that it is not intended or warranted for extended temperature range or unusual environmental requirements. Medical life-support equipment is specifically not recommended (p. 39). Valve control on a process line sits inside that boundary; a safety instrumented function does not.
Frequently asked questions
Does swapping to this sensor mean a hardware change?
The sensing principle is contactless, so the mechanical change is the magnet and its holder. The electrical change depends on what you are replacing: a potentiometer is an analog signal the controller already reads, while this part produces serial frames, so the firmware path is new even where the connector is not. The supply is the other change: 4.5 V to 5.5 V with no option below that (p. 7).
What else on an actuator produces the same symptom?
A potentiometer track worn in the middle of its travel reads correctly at the ends and drifts everywhere else. A magnet that has moved on its holder shifts the whole curve rather than flattening one end. A sticking stem produces hunting that has nothing to do with the sensor.
How do I verify the fix quickly?
Three measurements. Field amplitude at both stroke ends, against the 20 mT and 70 mT limits. Time from VDD valid to the first valid angle, against the 1.8 ms and 5.9 ms start-up timings. And a forced diagnostic, to confirm the controller keys on the status bit rather than the payload.
How does the cost compare with the position sensing we use now?
Compare at the subsystem level. A contactless sensor removes the wear mechanism of a contacting element and the replacement interval that goes with it, and adds a programming step on the programmable options plus firmware work for the serial protocol. On a low-volume actuator the engineering effort dominates.
Sourcing note
The MLX90367EGO-ACU-000 is a dual die TSSOP-16 part and ships on reel as the RE packing form (p. 4). The E suffix fixes the range at -40 to 85 °C, the grade to check first if the actuator electronics sit near the gearbox. Tell us your annual quantity, the option code you need and the temperature grade your application reaches, and we will quote from current stock.