MotionFX Azimuth Detection Issues with ISM330DHCX - Seeking Configuration Guidance

[mfo]

Problem Summary

I'm experiencing challenges with MotionFX library for azimuth rotation detection using the ISM330DHCX sensor. While MotionDI works well for altitude detection, MotionFX shows inconsistent behavior for azimuth (rotational) measurements.

I have done a test where I have collected samples over the entire weekend (about 65 hours). I am aware of the difficulties in detecting rotational movement using only accelerometer and gyroscope, however, the issues below do not seem to stem from this, but rather with issues in the MotionFX library or parameters.

Hardware Setup

• Sensor: ISM330DHCX (accelerometer + gyroscope)
• Orientation: WDS mounting (X-West, Y-Down, Z-South)
• Sample Rate: 104Hz
• Application: Rotational motion detection system that detects motion over longer periods in time. (No real time updates needed)

Current Configuration

MotionDI Parameters (working well for altitude):

mdi.move_thresh_g = 0.25f;
mdi.cal_type = 1;
mdi.acc_thr_g = 0.01f;
mdi.gyro_thr_dps = 0.01f;
mdi.max_gyro_dps = 15.0f;
mdi.gyro_cal_type = 2;
mdi.sf_atime = 2.0f;
mdi.sf_frtime = 2.0f;
mdi.modx = 1;
mdi.output_type = 1;

MotionFX Parameters (issues with azimuth):

mfx.atime = 2.0f;
mfx.mtime = 0.01f; // not used in 6X mode
mfx.frtime = 2.0f;
mfx.lmode = 1; // static learning - if we put this to 2, I see immediate drift/changes 
mfx.gbias_acc_th_sc = 0.00153f; // from library
mfx.gbias_gyro_th_sc = 0.1f; // library had 0.004f
mfx.modx = 1; // AN suggests setting it to 2, but this is causing issues as well (does this change the refresh rate?)
mfx.output_type = 1;

Observed Issues

1. Gyroscope Bias Drift

• MotionDI: Shows significant X-axis bias drift (0-60+ dps over time)
• MotionFX: Maintains stable bias (~-0.42 dps constant)
• Question: Is this normal behavior, or should I adjust calibration parameters?

 

2. Angle Calculation Method

Currently extracting angles from quaternion differences:

Quaternion relative = QuaternionMath::relativeRotation(current, reference);
float roll, pitch, yaw;
QuaternionMath::toEulerAngles(relative, roll, pitch, yaw);

// Using pitch for altitude (MotionDI)
altitude_angle = fabsf(pitch);
// Using roll for azimuth (MotionFX) 
azimuth_angle = fabsf(roll); // This is the correct axis for what we are trying to detect. 

3. Temperature Influence

• Temperature varies 33-35°C during operation
• Correlation with bias drift observed (very minimal changes, but changes still)
• Question: Should I implement temperature compensation? (Even though the ism has built-in compensation)

Analysis Results

From my data analysis:

• Accelerometer magnitude stable around 1000mg (1g)
• Gyroscope magnitude around 1.0-1.2 dps
• MotionFX quaternion trajectory shows more circular pattern
• MotionDI quaternion trajectory shows more linear segment\

Specific Questions

1. Bias Calibration: Why does MotionDI show significant X-axis bias drift while MotionFX remains stable? Should I adjust gbias_gyro_th_sc parameter?

2. Azimuth vs Altitude: Is my approach of using MotionFX for azimuth and MotionDI for altitude detection appropriate, or should I use a single library? The reason I am using both libraries is because during initial testing, I found MotionDI to be way more reliable in altitude angle tracking than MotionFX, but MotionFX showed good results for altitude.

3. WDS Orientation: Are there specific considerations for WDS mounting that affect MotionFX performance?

4. Library Parameters: Given my 104Hz sample rate and WDS orientation, are my current parameters optimal?

5. Temperature Compensation: Should I implement additional temperature-based bias correction?

Expected Behavior

I need to detect:

• Altitude changes: ±5° threshold (working with MotionDI)
• Azimuth changes: ±10° threshold (struggling with MotionFX)

Code Initialization

// MotionFX initialization
MotionFX_initialize(static_cast<MFXState_t>(_mfx_state_buffer));
MotionFX_getKnobs(static_cast<MFXState_t>(_mfx_state_buffer), &_mfx_knobs);
// Set knobs...
MotionFX_setKnobs(static_cast<MFXState_t>(_mfx_state_buffer), &_mfx_knobs);
MotionFX_enable_6X(static_cast<MFXState_t>(_mfx_state_buffer), MFX_ENGINE_ENABLE);

[EDIT]: I have now also done some extra testing. I found the following "noise floor" for my sensor:

accel_x_mg - Mean: -62.39, Std Dev: 0.61, Median: -62.00, Min: -68.00, Max: -57.00
 3 Noise Floor accel_x_mg: 0.00182749 mg
accel_y_mg - Mean: -1010.81, Std Dev: 0.54, Median: -1011.00, Min: -1015.00, Max: -1005.00
 3 Noise Floor accel_y_mg: 0.00161997 mg
accel_z_mg - Mean: -16.90, Std Dev: 1.12, Median: -17.00, Min: -30.00, Max: -12.00
 3 Noise Floor accel_z_mg: 0.00335705 mg
gyro_x_dps - Mean: 0.38, Std Dev: 0.04, Median: 0.42, Min: 0.14, Max: 0.70
 3 Noise Floor gyro_x_dps: 0.00227631 mg
gyro_y_dps - Mean: -0.78, Std Dev: 0.04, Median: -0.77, Min: -1.05, Max: -0.49
 3 Noise Floor gyro_y_dps: 0.00209153 mg
gyro_z_dps - Mean: -0.60, Std Dev: 0.04, Median: -0.63, Min: -0.91, Max: -0.35
 3 Noise Floor gyro_z_dps: 0.00204885 mg

Magnitude of Acceleration: 1012.88 mg
Magnitude of Gyroscope: 1.06 dps

And have used these settings in the parameters for MotionFX, however, I still see a steady drift overnight (21 degrees of change in 16 hours)

 

If anyone would want to check out the full code, I can definitely share some snippets later on.
Any insights on proper MotionFX configuration for azimuth detection in this setup would be greatly appreciated!

ST Library Versions: Using latest MotionDI and MotionFX libraries from X-CUBE-MEMS1 package.

[mfo]

Hi @Federica Bossi 
Since you’ve previously helped with my question on MotionFX and MotionDI, I was wondering if you might also have some experience with this issue. If you have a moment to take a quick look, I’d really appreciate it.
Thanks so much in advance!

 

[mfo]

Hi @Federica Bossi 

By now, I have fixed some of the issues. I have also been able to move from the combination of the two libraries to only using MotionFX. However, the remaining issue is, even with the most basic setup, MotionFX starts drifting (azimuth) at a rate of 0.9 degrees every 12 hours. This seems quite fast. The accelerometer is never only influenced by one axis (i.e., A = (Ax, Ay, Az), (Ax, Ay) != (0,0) ), because we have some tilt on the sensor.

Not only that, we have noticed that with most calibration constants and parameters, as soon as there is any vibration (toothbrush motor attached to the tripod on which the ism330dhcx is mounted), the Kalman filter always assumes this is real movement (for both altitude as azimuth).

I was wondering if there was some extra information around this or some help you (or anyone) could offer.

[mfo]

@SimonePradolini , do you maybe have some experience in this?

[SimonePradolini]

Hello @mfo 

I'm sorry, but I'm not supporting those libraries and I'm not an expert into. But I'm asking internally to give attention to your request.

 

Best regards,

Simone

[mfo]

No worries. Thank you!

 

[Federica Bossi]

Hi @mfo ,

0.9 deg over 12 hours is equivalent to 0.075 deg/hr, given that BI for our gyroscopes is around 3 deg/hr. From datasheet:

What is your application? What values do you want to achieve?

[mfo]

Hi @Federica Bossi 

Since it has been awhile since I posted the question, certain things have changed. However, the goal is the same. We would like to use MotionFX or MotionDI to detect orientation changes on a fixed pole (about 1m) that is standing outside near a busy road. We would like to track altitude and zenith changes after reboots (we restart azimuth on each power cycle). This is for a larger project where we will eventually have many of these fixtures that have multiple sensors to perform multiple tests. One of tests is to detect whether wind/turbulence or an animal or ... has moved the fixture. We would like to be able to tell that any type of movement has happened. Not in real time, we can wait up to four hours before responding, giving us a good "validation" time to compare with drift or ... We do, however, would like to be able to detect both slow movement (about 0.016dps to 1dps) and fast movement (above 1 dps) up to 5 degrees accurate for altitude and zenith, and up to 10 degrees accurate for azimuth. 

The idea was to, on startup, take a reference quaternion and track angles from there. We hoped that using motionFX we could mitigate the drift in azimuth (that the Kalman filter would "recognize" the drift and ignore it for a large part) and would help us to detect changes. However, as mentioned above, MotionFX sometimes gives us worse performance than a simple complementary filter on mitigating the drift. 
Currently, we have MotionDI set up with a quaternion reference and have a simple complementary filter running in parallel that resets every `n` minutes so that we can keep detecting true changes for azimuth.

Using this technique, we have been able to get azimuth a lot more stable:

However, we were still wondering if there is no better way to do this using just motionFX or just motionDI? It seems weird that, even using parameter optimization, we can not get overall good results for tracking changes in a stationary system.
Do you have any idea why this is? Any tips? 
Also, for motionfx, this question still stands:
> Not only that, we have noticed that with most calibration constants and parameters, as soon as there is any vibration (toothbrush motor attached to the tripod on which the ism330dhcx is mounted), the Kalman filter always assumes this is real movement (for both altitude as azimuth). Is this normal? Why is this?

If I need to go into deeper detail, let me know.
Thank you in advance!