A method and system for dual-modal signal fusion processing of multidimensional force sensors
By employing a dual-modal signal fusion processing method, the output signal of the multi-dimensional force sensor is divided into two paths for separate acquisition and filtering. Data alignment and fusion are then achieved through group delay compensation and asynchronous multi-rate filters, thus solving the problems of high precision and high response of the multi-dimensional force sensor and realizing the smoothness of dynamic tracking and mode switching.
Patent Information
- Application Number
- CN202611499422.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to achieve both high precision and high response of multidimensional force sensors on a single data stream, and also suffer from group delay alignment and zero-bias drift issues.
A dual-modal signal fusion processing method is adopted to divide the output signal of the same force sensor into two paths, which are acquired and filtered through different acquisition links. The first path with a high sampling rate is used for high response, and the second path with low noise is used for high precision. Data alignment and fusion are performed through group delay compensation and asynchronous multi-rate filter. The zero bias of the first acquisition link is explicitly modeled for online estimation and subtraction.
While achieving high accuracy and high response on a single data stream, it solves the problems of group delay alignment and zero bias drift, maintains low noise and high accuracy characteristics, and estimates and subtracts zero bias drift online, achieving smooth dynamic following and mode switching.
Smart Images

Figure CN123116170A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of force / torque sensor signal processing technology, and in particular to a dual-modal signal fusion processing method and system for multi-dimensional force sensors. Background Technology
[0002] Intelligent equipment, represented by humanoid robots, collaborative robots, and precision assembly equipment, demands a simultaneous increase in the bandwidth and precision of joint force control. Joint force control needs to distinguish minute contact forces in static and slow-moving conditions (such as precision assembly, grinding, and grasping fragile items), and also needs to respond to impact force changes within milliseconds in running, jumping, and accidental collisions, so that the control system can promptly perform compliant or protective actions. Torque sensors and multi-dimensional force / torque sensors are the core feedback components of joint force control; the precision and bandwidth of their output data links directly determine the upper limit of force control performance.
[0003] However, the accuracy and bandwidth of the signal chain are mutually constrained: high accuracy tends to require low noise and high resolution, typically necessitating an oversampling architecture; high bandwidth demands high sampling rate and low latency, usually requiring a high-speed architecture. A single ADC architecture cannot simultaneously meet both requirements, making it difficult for existing solutions to balance accuracy and response. Therefore, how to achieve both high accuracy and high response on a single data stream is a pressing technical problem that needs to be solved in this field. Summary of the Invention
[0004] The purpose of this invention is to provide a dual-modal signal fusion processing method and system for multi-dimensional force sensors, so as to achieve high precision and high response on a single data stream, and solve the problems of group delay alignment of high-precision links, zero-bias drift of high-speed links, and asynchronous fusion of two data streams.
[0005] The objective of this invention is achieved through the following technical solution:
[0006] A method for dual-modal signal fusion processing for multi-dimensional force sensors includes the following steps:
[0007] The output signal of the same force sensor is buffered and split into two paths. The first path is acquired via a first acquisition link at a first sampling rate to obtain a first data stream. The second path is acquired via a second acquisition link, filtered by a filtering unit, and output at a second sampling rate to obtain a second data stream. The first sampling rate is higher than the second sampling rate. The equivalent input noise of the first acquisition link is higher than the equivalent input noise of the second acquisition link, and the bandwidth of the first acquisition link is wider than the bandwidth of the second acquisition link. The first acquisition link and the second acquisition link are driven by the same clock source and have the same source timestamp. The sampling period corresponding to the first sampling rate is the first sampling period.
[0008] Group delay compensation is performed on the second data stream to determine the inherent group delay τ of the filtering unit. g The timestamp of the k-th sample in the second data stream is corrected by an amount of τ. g This aligns the corrected second data stream with the first data stream on the timeline, and the timestamp correction does not change the sample values of the second data stream, but only corrects its time attribution.
[0009] An asynchronous multi-rate filter is run with the first sampling period as the clock cycle. The state vector of the filter includes the measured force, the rate of change of force, and the zero bias of the first acquisition link. Each clock cycle executes a time update and a first measurement update based on the first data stream. When the corrected and aligned second data stream sample arrives asynchronously, a second measurement update based on the second data stream is additionally executed. The fused output takes the first component of the state vector as the estimated value of the measured force and outputs it through the digital communication interface.
[0010] In one embodiment, the force sensor is a bridge-type force sensor, and its output is a differential output signal;
[0011] The first acquisition link is a successive approximation (SAR) ADC link, a pipelined ADC link, a flash ADC link, or a time-interleaved (TI-SAR) ADC link;
[0012] The second acquisition link is a Delta-Sigma ADC link or a multi-slope ADC link.
[0013] In one embodiment, the filtering unit is an Nth-order CIC decimation filter with a decimation factor of M and an intrinsic group delay τ. g Determined based on N and M;
[0014] The timestamp correction is to roll back the timestamp of the k-th sample by τ. g The corrected k-th sample corresponds to the first data stream. 1 sample, where dec is the ratio of the first sampling rate to the second sampling rate.
[0015] In one embodiment, the state vector is Where F is the force being measured, dF / dt is the rate of change of force, and b1 is the zero bias of the first acquisition link;
[0016] The first measurement provides information on the measured force and the zero bias of the first acquisition link, while the second measurement only observes the measured force and indirectly corrects the zero bias estimate through the state transition relationship of the filter.
[0017] The second acquisition link is equipped with a chopper zero-stabilization circuit at the front end, so that the zero bias of the second acquisition link is lower than a preset threshold.
[0018] In one embodiment, the filter performs time updates and first measurement updates sequentially at each tick using asynchronous multi-rate Kalman recursion, and performs a second measurement update if and only if the second data stream sample arrives;
[0019] The first measurement update uses the first data stream as the observation, and the second measurement update uses the second data stream as the observation.
[0020] In one embodiment, dynamic mode adaptation is also included:
[0021] Based on a dynamic level index independent of the filter measurement information, a mode switching logic switches between static and dynamic modes, and the process noise intensity and the second measurement noise covariance are configured accordingly.
[0022] In static mode, the process noise is taken as the minimum value, the second measurement noise covariance is taken as the reference value, and the fusion output is determined by the second measurement update.
[0023] In dynamic mode, the process noise is taken as the largest value, the second measurement noise covariance is amplified, and the fused output is determined by the update of the first measurement.
[0024] In one embodiment, the dynamic range index is calculated based on the original samples of the first data stream and is a sliding window slope amplitude, high-pass filter energy, band-pass filter energy, wavelet decomposition high-frequency coefficients, or a combination of multiple window slopes; the dynamic range index is independent of the measurement information of the filter;
[0025] The mode switching logic is either hysteresis dual threshold switching, soft weighted switching based on continuous interpolation of dynamic degree index, or multi-model interactive switching.
[0026] In one embodiment, when the second data stream sample arrives asynchronously, measurement information is injected using a snapshot caching and replay correction method, or the second measurement update is performed directly in the current tick.
[0027] In dynamic mode, when the filter gain of the second measurement is lower than the preset threshold, the noise covariance of the second measurement is periodically restored to the reference value to re-anchor the zero-bias estimate of the first acquisition link, or a random walk constraint is applied to the zero-bias state.
[0028] The fused output is generated instantly in each of the first sampling cycles and output via the digital communication interface, or with an overall delay τ. g The output is in fully anchored mode.
[0029] A dual-modal signal fusion processing system for a multidimensional force sensor, comprising:
[0030] A dual-modal acquisition module is used to buffer and split the output signal of the same force sensor into two paths. The first path is acquired via a first acquisition link at a first sampling rate to obtain a first data stream. The second path is acquired via a second acquisition link, filtered by a filtering unit, and output at a second sampling rate to obtain a second data stream. The first sampling rate is higher than the second sampling rate, the equivalent input noise of the first acquisition link is higher than the equivalent input noise of the second acquisition link, and the bandwidth of the first acquisition link is wider than the bandwidth of the second acquisition link. The first acquisition link and the second acquisition link are driven by the same clock source and have the same source timestamp. The sampling period corresponding to the first sampling rate is the first sampling period.
[0031] The group delay compensation module is used to determine the inherent group delay τ of the filtering unit. g The timestamp of the k-th sample in the second data stream is corrected by an amount of τ. g This aligns the corrected second data stream with the first data stream on the timeline, and the timestamp correction does not change the sample values of the second data stream, but only corrects its time attribution.
[0032] The fusion filtering module is used to run an asynchronous multi-rate filter with the first sampling period as the clock cycle. The state vector of the filter includes the measured force, the rate of change of force, and the zero bias of the first acquisition link. Each clock cycle performs a time update and a first measurement update based on the first data stream. When the corrected and aligned second data stream sample arrives asynchronously, a second measurement update based on the second data stream is added. The fusion output takes the first component of the state vector as the estimated value of the measured force.
[0033] The calibration output module is used to map the estimated value of the measured force into a force physical quantity through calibration coefficients, and output it through the digital communication interface of the system.
[0034] The parameter storage module is used to store the fusion algorithm parameters of the asynchronous multi-rate filter and the calibration coefficients.
[0035] A multidimensional force sensor includes an elastomer, several force-sensitive sensors, and the aforementioned dual-modal signal fusion processing system;
[0036] Each force-sensitive direction corresponds to a force sensor, and the output of each force sensor is buffered and then sent to the first and second acquisition links of the channel.
[0037] Each channel is configured to perform the method described;
[0038] The estimated value of the measured force output from each channel is decoupled from the channel crosstalk by the static decoupling matrix C in the multi-dimensional force sensor to obtain the multi-dimensional force vector, which is then uniformly output by the multi-dimensional force sensor.
[0039] Compared with the prior art, the beneficial effects of the present invention include at least the following:
[0040] This invention provides a dual-modal signal fusion processing method for a multi-dimensional force sensor. The output of the same force sensor is buffered and divided into two paths. The first path is acquired through a first acquisition link (high sampling rate, wide bandwidth, relatively low precision) to obtain a first data stream. The second path is acquired through a second acquisition link (low sampling rate, low noise, narrow bandwidth) and filtered by a filtering unit to obtain a second data stream. The two acquisition links are driven by the same clock source and have the same timestamp. The two data streams are aligned by group delay compensation and then fused by an asynchronous multi-rate filter, achieving high precision and high response on a single data stream: retaining the high response of the first acquisition link and the high precision of the second acquisition link, and estimating and subtracting the zero bias of the first acquisition link online.
[0041] This method integrates the low-frequency output to be close to the noise floor of the second acquisition link, while maintaining the bandwidth of the first acquisition link in the high-frequency band, thus solving the problem of inconsistent frequency response in a single ADC architecture.
[0042] This method achieves group delay compensation through timestamp correction, only corrects the time attribution of the second data stream samples, does not change the sample values, and does not introduce any additional filtering, thus not compromising the low-noise and high-precision characteristics of the second acquisition link.
[0043] This method explicitly models the zero bias of the first acquisition link as a state vector component. The first measurement simultaneously provides information on the measured force and the zero bias, enabling the filter to estimate and subtract the zero bias of the first acquisition link online. This allows the first acquisition link to be trusted for dynamic following and avoids zero bias drift contaminating the fusion output.
[0044] This method switches between static and dynamic modes based on a dynamic index that measures the innovation independently of the filter. Static mode prioritizes accuracy, while dynamic mode prioritizes response, avoiding mode misjudgment caused by "innovation blindness" in dynamic mode. Hysteresis-based dual-threshold logic ensures smooth and jitter-free switching. Attached Figure Description
[0045] Figure 1 This is a flowchart illustrating a dual-modal signal fusion processing method for a multi-dimensional force sensor according to an embodiment of the present invention.
[0046] Figure 2 This is a schematic diagram of the structure of the multi-dimensional force sensor system according to an embodiment of the present invention;
[0047] Figure 3 This is a schematic diagram of a bridge-type force sensor according to an embodiment of the present invention;
[0048] Figure 4This is a schematic diagram comparing the reference torque, the raw SAR output, and the raw Delta-Sigma output of an embodiment of the present invention.
[0049] Figure 5 This is a schematic diagram comparing SAR ADC link sampling data and Delta-Sigma ADC link sampling data according to an embodiment of the present invention;
[0050] Figure 6 This is a schematic diagram of the Delta-Sigma ADC link group delay compensation according to an embodiment of the present invention;
[0051] Figure 7 This is a schematic diagram of the dynamic level index and hysteresis switching process in an embodiment of the present invention;
[0052] Figure 8 This is a schematic diagram illustrating the timing relationship between two-channel sampling and filtering updates in an embodiment of the present invention;
[0053] Figure 9 This is a timing diagram illustrating CIC group delay modeling and timestamp alignment according to an embodiment of the present invention;
[0054] Figure 10 This is an overview schematic diagram of the dual-modal fusion output of an embodiment of the present invention;
[0055] Figure 11 This is a comparative schematic diagram of the steady-state accuracy of embodiments of the present invention;
[0056] Figure 12 This is a comparative schematic diagram of the dynamic segment response in an embodiment of the present invention;
[0057] Figure 13 This is a comparative schematic diagram of the full-process estimation error in an embodiment of the present invention;
[0058] Figure 14 This is a schematic diagram of the error amplitude and frequency in an embodiment of the present invention. Detailed Implementation
[0059] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided to make the invention more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.
[0060] This invention provides a method and system for dual-modal signal fusion processing of multi-dimensional force sensors. For example... Figure 2As shown, the multidimensional force sensor system includes: a bridge-type force sensor, a chopper-stabilized buffer array, a first acquisition link, a second acquisition link, a digital system module, and auxiliary circuitry. In this invention, unless otherwise specified, "force" includes both force and torque, and "measured force" includes both the measured force and the measured torque.
[0061] The dual-modal signal fusion processing system is the core processing subsystem of the multi-dimensional force sensor system, and includes: a dual-modal acquisition module, a group delay compensation module, a fusion filtering module, a calibration output module, and a parameter storage module.
[0062] The dual-modal acquisition module is implemented by a first acquisition link, a second acquisition link, and their digital filtering sections; the group delay compensation module, fusion filtering module, calibration output module, and parameter storage module are implemented by the digital system module. The digital system module runs an adaptive dual-modal data fusion algorithm to complete group delay compensation, asynchronous multi-rate fusion, calibration output, and parameter storage.
[0063] The auxiliary circuit includes a temperature sensor, a self-test system, an oscillator, an LDO, and a test port. It is used to improve measurement accuracy, reliability, and testability, but is not necessary to achieve the core solution of this invention.
[0064] like Figure 1 As shown, this embodiment provides a dual-modal signal fusion processing method for multi-dimensional force sensors, executed by a multi-dimensional force sensor system, including the following steps: S1-S3, and further including S4. This method, through a single-bridge, two-path, mutually beneficial, and algorithm-fused overall architecture, simultaneously achieves high precision and high response on a single data stream, and solves the problems of group delay alignment in high-precision links, zero-bias drift in high-speed links, and asynchronous fusion of the two data streams.
[0065] S1: Dual-modal acquisition. The differential output signal of the same bridge-type force sensor is buffered and split into two paths. The first path is acquired via a first acquisition link at a first sampling rate to obtain a first data stream. The second path is acquired via a second acquisition link, filtered by a filtering unit, and output at a second sampling rate to obtain a second data stream. The first sampling rate is higher than the second sampling rate, the equivalent input noise of the first acquisition link is higher than the equivalent input noise of the second acquisition link, and the bandwidth of the first acquisition link is wider than the bandwidth of the second acquisition link. The first acquisition link and the second acquisition link are driven by the same clock source within the force / torque sensor system and have the same timestamp. The sampling period corresponding to the first sampling rate is the first sampling period.
[0066] Specifically:
[0067] like Figure 3As shown, the Wheatstone bridge force sensor uses a glass micro-fusion process to mount MEMS strain gauges onto an elastic body and form a Wheatstone bridge to convert the applied multidimensional force / torque into differential voltage. Output.
[0068] Chopper-stabilized zero-delay buffer array: Each signal is buffered by a set of buffers. The buffer array is configured with chopper noise suppression to filter out 1 / f noise and offset after modulation to a high frequency, ensuring low noise and low drift at the front end. After buffering, the signal is split into two acquisition links at the node.
[0069] First acquisition link (high-speed link): In this embodiment, the first acquisition link is a successive approximation (SAR) ADC link. The split signal is sent to the SAR ADC (18-bit, 1 MSPS in this embodiment) after passing through an anti-aliasing low-pass filter and a PGA programmable gain amplifier (1~N times). The output is sent to the digital system module after passing through a SAR digital filter. This link has a wide bandwidth and negligible delay, but it has relatively large thermal noise and slow zero-bias drift. The SAR link is configured with an independent BandGap reference.
[0070] Second acquisition link (high-precision link): In this embodiment, the second acquisition link is a Delta-Sigma ADC link. The split signal is amplified (1~N times) by a PGA programmable gain amplifier and then sent to the Delta-Sigma ADC (24-bit, 2kSPS in this embodiment). The modulator bitstream is reduced to the output data rate by a CIC decimation filter before being sent to the digital system module. This link has low noise and high precision, but it has a fixed group delay of about 2ms due to the CIC decimation filter, and the link bandwidth is about 400Hz. The Delta-Sigma link is configured with an independent BandGap reference. Figure 5 The image shows a comparison diagram of SAR ADC link sampling data and Delta-Sigma ADC link sampling data.
[0071] like Figure 4 The diagram shows a full comparison of the reference torque (the ideal torque value set in the simulation), the original SAR output, and the original Delta-Sigma output (including group delay). In the steady-state segment, both paths can follow the signal; in the impact segment, the defects of both paths are exposed simultaneously, namely, the SAR noise is coarse, the Delta-Sigma has a group delay of about 2 ms, and the impact / ringing peaks are clipped by the low-pass filter.
[0072] In step S1, the same bridge signal is split into two paths and acquired through a high-speed link and a high-precision link respectively, so that the system can simultaneously obtain raw data streams with high response and high precision, providing complementary information for subsequent fusion.
[0073] S2: Group delay compensation. Perform group delay compensation on the second data stream to determine the inherent group delay τ of the filtering unit. g The timestamp of the k-th sample in the second data stream is corrected by an amount of τ. g This aligns the corrected second data stream with the first data stream on the timeline, and the timestamp correction does not change the sample values of the second data stream, but only corrects its time attribution.
[0074] Specifically, the Delta-Sigma ADC uses a CIC decimation filter in its post-stage to decimate the high-speed bit stream from the modulator to the output data rate f. ds The CIC decimation filter is a linear-phase FIR structure, and its group delay is an intrinsic constant. For an Nth-order CIC decimation filter with a decimation factor of M, the group delay is approximately... One modulator clock cycle. In this embodiment, as... Figure 6 and Figure 9 As shown, the Delta-Sigma channel ODR is 2kHz, and the CIC group delay τ g ≈2ms, equivalent to 4 output sample cycles.
[0075] Timestamp alignment method: The digital system timestamps the output samples of the two ADCs using a unified clock; for the k-th output sample of the Delta-Sigma channel, its timestamp is shifted back to the intrinsic group delay τ. g :
[0076]
[0077] like Figure 6 As shown, the Delta-Sigma sample after the rollback is equivalent to a no-delay sample on the time axis, and its correspondence with the SAR time axis is as follows: the k-th aligned sample corresponds to the SAR sampling sequence number. (dec=Fs / f) ds The ratio of the first sampling rate to the second sampling rate, where Fs is the first sampling rate, and f is the second sampling rate. ds For the second sampling rate (dec=500 in this embodiment), group delay compensation only corrects the temporal assignment of samples without introducing any additional filtering, thus not impairing the noise performance of the Delta-Sigma channel.
[0078] Real-time causal implementation: The asynchronous multi-rate filter operates on the basis of the SAR sampling period. In each SAR sampling period, the current state estimate is output in real time. After the Delta-Sigma sample arrives, its measurement information is injected into the filter state in either method A or method B.
[0079] Method A (Snapshot Cache + Replay Correction): Rolling cache of the most recent τ g SAR samples and τ over timeg Previously, Delta-Sigma updated the filter state snapshot of the slot cycle. When Delta-Sigma samples arrive, the historical snapshot is retrieved according to the timestamp after the rollback. The second measurement update is performed on that historical cycle. Then, the cached SAR samples are used to recursively push back to the current time from that historical cycle, and the corrected current state is used to take over the subsequent real-time filtering.
[0080] Method B (Direct Update): No buffer is set. When the Delta-Sigma sample arrives, the second measurement update is performed directly in the current cycle, tolerating time misalignment of τg. In static mode, the error is negligible; in dynamic mode, R1 amplification makes the second measurement gain approach zero, and misalignment does not affect the output.
[0081] Output Real-Time Description: The fused output data streams do not wait for Delta-Sigma samples; the output is generated in real-time per SAR beat. Only in applications requiring extremely high single-point accuracy and where latency is not a concern can the overall output delay τ be selected. g This ensures that each output point is a fully anchored value; the delayed output is configurable.
[0082] Step S2 corrects only the timestamp of the second data stream without changing the sample values, eliminating group delay misalignment without introducing additional filtering, and maintaining the low noise and high precision characteristics of the second acquisition link.
[0083] S3: Asynchronous Multi-rate Fusion. The fusion algorithm is an adaptive asynchronous multi-rate Kalman filter. The filter operates with the first sampling period as the basic clock cycle; the first measurement participates in the update every clock cycle, while the second measurement only participates in the update when its sample arrives asynchronously (once every dec clock cycles). For example... Figure 8 The diagram shown illustrates the timing relationship between the two sampling and filtering updates.
[0084] Specifically:
[0085] (1) State vector and state equation
[0086] Define a three-dimensional state vector Where F is the measured force, dF / dt is the rate of change of force, and b1 is the zero bias of the first acquisition link. The zero bias of the first acquisition link is explicitly modeled as a state component and incorporated into the estimation, enabling the filter to separate the slow zero drift of the first acquisition link from the true force signal online. The state equation adopts a piecewise constant model driven by uniform acceleration:
[0087] x(k) = A·x(k-1) + G·w(k)
[0088] A=[1,Ts,0;0,1,0;0,0,1]
[0089] G=[Ts² / 2,Ts,0]ᵀ
[0090] Where w(k) is the scalar process noise, the intensity of which is determined by the current operating mode.
[0091] (2) Process noise covariance Q and measurement noise covariance R
[0092] Process noise covariance In this embodiment, the value is: static mode q static = Dynamic mode q dynamic = Q33= ;
[0093] Measurement noise covariance: First measurement noise covariance R² = σ sar ², This embodiment Second measurement noise covariance static benchmark Magnification in dynamic mode times.
[0094] (3) Measurement equation
[0095] z sar (k)=H2·x(k)+v2(k), H2=[1,0,1];
[0096] z ds (k)=H1·x(k)+v1(k), H1=[1,0,0];
[0097] The third column of H2 is 1, so that the first measurement (i.e., the SAR measurement) provides information on both F and b1. The filter uses this to estimate and subtract the zero bias of the first acquisition link online. H1 only observes F, and the second measurement directly anchors the force state itself and indirectly corrects the zero bias estimate through the state transition relationship.
[0098] (4) Asynchronous multi-rate filtering recursive process
[0099] Each SAR sampling period is executed sequentially:
[0100] Step 1, Time Update: ;
[0101] The second step, the first measurement update (SAR high-speed measurement, executed every frame): , ;
[0102] Step 3, Second Measurement Update (Delta-Sigma High-Precision Measurement, Asynchronous Execution): If and only if hour, , .
[0103] The fusion output takes the first component of the state x(1) as the estimated value of the measured force, and after calibration and mapping, it is output by the SPI interface.
[0104] (5) Initialization
[0105] initial state Initial covariance .
[0106] Step S3: Run asynchronous multi-rate Kalman filtering with the first sampling period as the beat, so that the output data rate reaches the high-speed link level, while the low-frequency accuracy is anchored by the high-precision link, and the zero bias of the first acquisition link is estimated and subtracted online.
[0107] S4: Dynamic Mode Adaptation. Based on a dynamic level index independent of the filter measurement information, a mode switching logic switches between static and dynamic modes, and configures the process noise intensity and the second measurement noise covariance accordingly. In static mode, the process noise is set to a minimum value, the second measurement noise covariance is set to a baseline value, and the fused output is determined by updating the second measurement. In dynamic mode, the process noise is set to a maximum value, the second measurement noise covariance is amplified, and the fused output is determined by updating the first measurement.
[0108] like Figure 7 As shown, specifically:
[0109] Mode switching is based on a dynamic range index independent of Kalman filter innovation. This dynamic range index is calculated based on the sliding window slope magnitude of the raw SAR high-speed data.
[0110]
[0111] The dynamic range index d(i) is obtained by performing a first-order low-pass smoothing on the slope amplitude. In this embodiment, wlen=200 and the smoothing bandwidth fc=300Hz. Mode switching uses hysteresis double-threshold logic: d(i)>th high Dynamic mode is switched in at the appropriate time; d(i) <th low Revert to static mode; in this embodiment... high =300 N·m / s, th low =150 N·m / s.
[0112] Static mode: Press q to press Q static Configuration, R1 takes the baseline value, and the fused output is determined by the second measurement update; Dynamic mode: Q presses q dynamic Configuration, R1 amplification The fused output is determined by the first measurement update.
[0113] Dynamic mode zero bias observability guarantee: In dynamic mode, when the filter gain of the second measurement is lower than the preset threshold, R1 is periodically restored to the reference value to re-anchor the SAR zero bias estimate, or a random walk constraint is applied to the zero bias state to limit its divergence.
[0114] Step S4: Switch between static and dynamic modes based on a dynamic level indicator independent of the information. Static mode prioritizes accuracy, while dynamic mode prioritizes response. This avoids low-speed links dragging down dynamic output and ensures zero bias observability in dynamic mode.
[0115] Further includes: multidimensional force extension.
[0116] Taking a six-dimensional force / torque sensor as an example: each force-sensitive direction corresponds to a bridge-type force sensor. The output of each force sensor is chopped, stabilized, and buffered before being split and sent to the first and second acquisition links of that channel, respectively. Each channel runs the above algorithm independently, and dynamic indicators and hysteresis switching are performed independently for each channel. The estimated value of the measured force output from each channel is decoupled from the channel crosstalk using a static decoupling matrix C, and then a unified output is obtained as a six-dimensional force vector.
[0117] This multidimensional force extension allows each channel to operate independently using the aforementioned method and switch modes independently for each channel. The multidimensional force vector is output through a decoupling matrix, achieving high precision and high response for each channel in multidimensional force measurement.
[0118] This embodiment also provides a dual-modal signal fusion processing system for a multi-dimensional force sensor, which is the core processing subsystem of the aforementioned multi-dimensional force sensor system. It includes: a dual-modal acquisition module, a group delay compensation module, a fusion filtering module, a calibration output module, and a parameter storage module. Each module can be implemented using dedicated digital circuits (ASIC digital system modules) or embedded processors.
[0119] The dual-modal acquisition module includes a first acquisition link and a second acquisition link, along with their PGA, reference, and digital filtering / CIC extraction unit. It is responsible for outputting a first data stream and a second data stream with the same source timestamp. The first acquisition link is a high-speed SAR link, and the second acquisition link is a Delta-Sigma high-precision link. This module is used to buffer and split the differential output signal of the same bridge-type force sensor into two paths. The first path is acquired via the first acquisition link at a first sampling rate to obtain the first data stream. The second path is acquired via the second acquisition link and filtered by the filtering unit to output the second data stream at a second sampling rate. The first sampling rate is higher than the second sampling rate, the equivalent input noise of the first acquisition link is higher than the equivalent input noise of the second acquisition link, and the bandwidth of the first acquisition link is wider than the bandwidth of the second acquisition link. Both acquisition links are driven by the same clock source within the system and have the same source timestamp. The sampling period corresponding to the first sampling rate is the first sampling period.
[0120] The dual-mode signal fusion processing system also includes a preprocessing unit, which performs PGA gain normalization and zero-bias calibration on the first data stream, and performs temperature drift digital calibration on the first data stream and the second data stream.
[0121] The group delay compensation module is used to determine the inherent group delay τ of the filtering unit. g The timestamp of the k-th sample in the second data stream is corrected by an amount of τ. g This aligns the corrected second data stream with the first data stream on the timeline, and the timestamp correction does not change the sample values of the second data stream, but only corrects its time attribution.
[0122] The fusion filtering module is used to run an asynchronous multi-rate filter with the first sampling period as the clock cycle. The state vector of the filter includes the measured force, the rate of change of force, and the zero bias of the first acquisition link. Each clock cycle performs a time update and a first measurement update based on the first data stream. When the corrected and aligned second data stream sample arrives asynchronously, a second measurement update based on the second data stream is appended. The fusion output takes the first component of the state vector as the estimated value of the measured force. The filter estimates the zero bias of the first acquisition link online and subtracts it from the fusion output. The fusion filtering module also includes a dynamic evaluation and mode switching unit, used to calculate the dynamic degree index and output the current mode and the corresponding process noise covariance and second measurement noise covariance configuration according to the hysteresis dual threshold.
[0123] The calibration output module is used to map the estimated value of the measured force into a force physical quantity through calibration coefficients, write it into the Register Bank, and output it through the digital communication interface; in a multi-dimensional scenario, the calibration coefficients include a static decoupling matrix C.
[0124] The parameter storage module is used to store the fusion algorithm parameters of the asynchronous multi-rate filter, the calibration coefficients and temperature compensation parameters, and supports factory programming and online configuration.
[0125] The aforementioned system achieves both high accuracy and high response on a single data stream. The group delay compensation module only corrects the timestamp of the second data stream without changing the sample values, allowing the low-noise, high-accuracy information from the second acquisition link to be used for fusion. The fusion filtering module operates with the first sampling period as its cycle, estimating and subtracting the zero bias of the first acquisition link online, thus balancing the high response of the first acquisition link with the high accuracy of the second acquisition link, making it highly feasible in engineering.
[0126] This embodiment also provides a multidimensional force sensor, including an elastomer, several force-sensitive sensors, and the aforementioned dual-modal signal fusion processing system.
[0127] Each force-sensitive direction corresponds to a bridge-type force sensor, and the output of each force sensor is buffered and then split into the first and second acquisition links of the channel.
[0128] The method described for each channel to operate independently involves dynamic level indicators and mode switching performed independently for each channel.
[0129] The estimated value of the measured force output from each channel is decoupled from the channel crosstalk by the static decoupling matrix C in the multi-dimensional force sensor to obtain the multi-dimensional force vector, which is then uniformly output by the multi-dimensional force sensor.
[0130] The aforementioned multidimensional force sensor achieves high precision and high response for each channel simultaneously in multidimensional force measurement by operating the above method independently for each channel and switching modes independently for each channel, and then decoupling and outputting multidimensional force vectors through a static decoupling matrix C.
[0131] Example 1: Dual-modal data processing method for Mz torque channel
[0132] As a specific application example of the above method, this embodiment takes the Mz torque channel (range ±5 N·m) as an example to give the complete method steps and simulation verification results.
[0133] Step S1 (Dual-modal acquisition): Mz torque bridge The output is chopped, stabilized, and buffered before being split; the upper path passes through an anti-aliasing low-pass filter and a PGA, and is then acquired by a SAR ADC at 1 MSPS and 18 bits to obtain the first data stream z. sar (i), Ts=1 µs; the lower path is acquired by Delta-Sigma ADC after passing through PGA, and output at 2 kSPS, 24 bits after CIC decimation filtering to obtain the second data stream z. ds (k).
[0134] Step S2 (Preprocessing): Perform PGA gain normalization and zero-bias calibration on the first data stream; perform CIC group delay modeling and timestamp rollback alignment on the second data stream (τ). g =2 ms); two channels perform digital temperature drift calibration based on temperature sensor readings.
[0135] Step S3 (Dynamic Level Evaluation): Calculate the slope amplitude of the 200 µs window on the first data stream and smooth it by 300 Hz to obtain the dynamic level index d(i).
[0136] Step S4 (Mode Determination): Determine whether the current mode is static or dynamic based on the hysteresis double threshold (300 / 150 N·m / s), and configure the filter Q and R1_eff accordingly.
[0137] Step S5 (Asynchronous Fusion Filtering): Execute asynchronous multi-rate Kalman recursion: perform time and SAR measurement updates every SAR cycle; when an aligned Delta-Sigma sample is reached every 500 SAR cycles, a high-precision measurement update is added. The state vector is... Where F is the measured torque, dF / dt is the torque change rate, and b1 is the zero bias of the first acquisition link.
[0138] Step S6 (Output): Take the first component of the state as the estimated value of the measured force, map it to the physical quantity of force after calibration coefficient, and output it through the digital communication interface; at the same time, the zero bias of the first acquisition link estimated by the filter can be used for zero bias correction and state feedback for the next cycle prediction.
[0139] The above method was verified through full-link simulation. The simulation constructed a typical operating condition from 0 to 24 ms: 0-10 ms was the steady-state segment (precision machining segment), with a 50 Hz slowly varying baseline superimposed with a small step load, emphasizing precision; from 10 ms onwards was the dynamic segment (bouncing impact segment), with a 5 ms half-sine impact + 2kHz damped ringing + 500Hz accompanying vibration, emphasizing response. Two channels were modeled according to actual device characteristics: SAR channel 18 bit / 1 MSPS, link noise... (Equivalent ENOB approximately 8 bits), includes slow zero-bias drift; Delta-Sigma channel 24 bit / 2 kSPS, link noise. 400 Hz bandwidth, 2ms CIC group delay.
[0140] The quantitative comparison of simulation results is shown in Table 1 (the values are taken from the nominal simulation model; the actual product values are based on the prototype measurements):
[0141]
[0142] As shown in Table 1, both single-channel schemes have fatal flaws—the Delta-Sigma single-channel has an impact delay of 3.43 ms and an unmeasurable rise time, while the steady-state error of the SAR single-channel is 4.2 times that of the present invention. Only the present invention has no flaws in all six indicators, with four of them being the best (steady-state RMSE, impact delay, rise time, and steady-state equivalent ENOB), and the remaining two (overall RMSE and dynamic RMSE) being on the same order of magnitude as the best values. Specifically: the steady-state RMSE of the present invention is 0.0048 N·m, which is about 4.2 times better than the SAR single-channel (0.0202 N·m) and superior to the steady-state value of the Delta-Sigma single-channel; the impact delay of the Delta-Sigma single-channel is as high as 3.43 ms, while that of the present invention is approximately... The rise time of 1.92 ms is comparable to that of a single SAR channel (1.99 ms), meaning it fully retains the response capability of the high-speed channel; the equivalent ENOB in the steady-state segment is 9.23 bits, the highest among the three. It should be noted that the equivalent ENOB in Table 1 is estimated based on the RMSE of the steady-state segment (including motion conditions with 50 Hz processing vibration and small step loads), reflecting the tracking accuracy under motion conditions rather than the static noise floor: the Delta-Sigma channel link noise itself is only... (Corresponding to an equivalent ENOB of approximately 15.8 bits), but its error under motion conditions is mainly contributed by the time misalignment of signal components due to the CIC group delay, causing the equivalent ENOB to drop to 6.84 bits. Its bottleneck lies in delay rather than accuracy, which is precisely the problem this invention aims to solve. The fused output of this invention, after group delay compensation and timestamp alignment, contains no group delay, and its steady-state RMSE and equivalent ENOB are superior to any single channel. Under constant load and purely static conditions, the fused output is anchored by Delta-Sigma high-precision measurements: In this embodiment, the measurement noise covariance R1 is taken as a conservative value in the simulation. The fused output has a pure static equivalent ENOB of approximately 13.2 bits; the measured noise per channel according to Delta-Sigma is... After calibrating R1, the purely static equivalent ENOB is approximately 15 bits. This means that the present invention achieves both Delta-Sigma level static accuracy and SAR level dynamic response on a single data stream. Figures 10-14 The diagrams shown are an overview of the dual-modal fusion output, a comparison of steady-state accuracy, a comparison of dynamic response, a comparison of full-process estimation error, and a schematic diagram of error amplitude and frequency, which intuitively demonstrate the beneficial effects of the present invention.
[0143] Alternative solution:
[0144] (1) Replacement of the first acquisition link (high-speed link): The SAR ADC can be replaced by high-speed architectures such as pipeline ADC, flash ADC, and time-interleaved SAR (TI-SAR), whose sampling rate is significantly higher than that of the second acquisition link, thus achieving the purpose of this invention.
[0145] (2) Replacement of the second acquisition link (high precision link): Delta-Sigma ADC can be replaced by high ODR type Delta-Sigma (such as 10 kSPS level), incremental Delta-Sigma (Incremental ADC), high precision multi-slope ADC, etc.
[0146] (3) Replacement of fusion algorithm: Asynchronous multi-rate Kalman filtering can be replaced by complementary filter, weighted least squares fusion, H∞ filter, particle filter or trained neural network fusion device, all of which can achieve the fusion purpose of "low frequency anchoring high precision channel and high frequency following high speed channel"; when using particle filter or neural network, the idea of zero bias explicit modeling in state vector can still be retained.
[0147] (4) Alternatives to mode switching: Hard hysteresis dual threshold switching can be replaced by soft weighting method based on continuous interpolation of dynamic degree index (fusion weight is a continuous function of index), or by adopting multi-model interaction (IMM) framework.
[0148] (5) Replacement of dynamic level indicators: The sliding window slope amplitude can be replaced by indicators such as high-pass / band-pass filter energy, wavelet decomposition high-frequency coefficients, and multi-window slope combination; the dynamic level indicators are all independent of the measurement information of the filter.
[0149] (6) Replacement of group delay compensation: timestamp rollback can be replaced by offline calibration lookup table interpolation, as long as the time of the first data stream and the second data stream are aligned and the sample value of the second data stream is not changed; in terms of causal real-time implementation, the injection of the aligned measurement can be done by two methods: snapshot caching plus replay correction (method A, with the best accuracy) or direct update at the current tick when the sample arrives (method B, simplified implementation). Alternatively, the fully anchored mode of the overall output delay τg can be configured in scenarios such as offline measurement.
[0150] (7) Replacement of sensing technology and bridge circuit: Glass micro-fusion MEMS strain gauges can be replaced by metal foil strain gauges or semiconductor strain gauges. The elastomer substrate can be metal or ceramic, and the bridge circuit can be a full bridge, half bridge or 1 / 4 bridge.
[0151] (8) Replacement of output interface: The digital communication interface can be SPI, and can be replaced by I²C, CAN, EtherCAT or register direct reading, etc.
[0152] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the invention without departing from the principles and spirit of the invention, and all such changes should fall within the protection scope of the claims of the present invention.
Claims
1. A method for dual-modal signal fusion processing for multi-dimensional force sensors, characterized in that, Includes the following steps: The output signal of the same force sensor is buffered and split into two paths. The first path is acquired via a first acquisition link at a first sampling rate to obtain a first data stream. The second path is acquired via a second acquisition link, filtered by a filtering unit, and output at a second sampling rate to obtain a second data stream. The first sampling rate is higher than the second sampling rate. The equivalent input noise of the first acquisition link is higher than the equivalent input noise of the second acquisition link, and the bandwidth of the first acquisition link is wider than the bandwidth of the second acquisition link. The first acquisition link and the second acquisition link are driven by the same clock source and have the same source timestamp. The sampling period corresponding to the first sampling rate is the first sampling period. Group delay compensation is performed on the second data stream to determine the inherent group delay τ of the filtering unit. g The timestamp of the k-th sample in the second data stream is corrected by an amount of τ. g This aligns the corrected second data stream with the first data stream on the timeline, and the timestamp correction does not change the sample values of the second data stream, but only corrects its time attribution. An asynchronous multi-rate filter is run with the first sampling period as the clock cycle. The state vector of the filter includes the measured force, the rate of change of force, and the zero bias of the first acquisition link. Each clock cycle executes a time update and a first measurement update based on the first data stream. When the corrected and aligned second data stream sample arrives asynchronously, a second measurement update based on the second data stream is additionally executed. The fused output takes the first component of the state vector as the estimated value of the measured force and outputs it through the digital communication interface.
2. The dual-modal signal fusion processing method for multi-dimensional force sensors as described in claim 1, characterized in that, The force sensor is a bridge-type force sensor, and its output is a differential output signal; The first acquisition link is a successive approximation (SAR) ADC link, a pipelined ADC link, a flash ADC link, or a time-interleaved (TI-SAR) ADC link; The second acquisition link is a Delta-Sigma ADC link or a multi-slope ADC link.
3. The dual-modal signal fusion processing method for multi-dimensional force sensors as described in claim 1, characterized in that, The filtering unit is an Nth-order CIC decimation filter with a decimation factor of M, and its intrinsic group delay τ g Determined based on N and M; The timestamp correction is to roll back the timestamp of the k-th sample by τ. g The corrected k-th sample corresponds to the first data stream. 1 sample, where dec is the ratio of the first sampling rate to the second sampling rate.
4. The dual-modal signal fusion processing method for multi-dimensional force sensors as described in claim 1, characterized in that, The state vector is Where F is the force being measured, dF / dt is the rate of change of force, and b1 is the zero bias of the first acquisition link; The first measurement provides information on the measured force and the zero bias of the first acquisition link, while the second measurement only observes the measured force and indirectly corrects the zero bias estimate through the state transition relationship of the filter. The second acquisition link is equipped with a chopper zero-stabilization circuit at the front end, so that the zero bias of the second acquisition link is lower than a preset threshold.
5. The dual-modal signal fusion processing method for multi-dimensional force sensors as described in claim 1, characterized in that, The filter performs time updates and first measurement updates sequentially at each clock cycle according to asynchronous multi-rate Kalman recursion, and performs a second measurement update if and only if the second data stream sample arrives; The first measurement update uses the first data stream as the observation, and the second measurement update uses the second data stream as the observation.
6. The dual-modal signal fusion processing method for multi-dimensional force sensors as described in claim 1, characterized in that, This also includes implementing dynamic mode adaptation: Based on a dynamic level index independent of the filter measurement information, a mode switching logic switches between static and dynamic modes, and the process noise intensity and the second measurement noise covariance are configured accordingly. In static mode, the process noise is taken as the minimum value, the second measurement noise covariance is taken as the reference value, and the fusion output is determined by the second measurement update. In dynamic mode, the process noise is taken as the largest value, the second measurement noise covariance is amplified, and the fused output is determined by the update of the first measurement.
7. The dual-modal signal fusion processing method for multi-dimensional force sensors as described in claim 6, characterized in that, The dynamic range index is calculated based on the original samples of the first data stream and is a sliding window slope amplitude, high-pass filter energy, band-pass filter energy, wavelet decomposition high-frequency coefficients, or a combination of multiple window slopes; the dynamic range index is independent of the measurement information of the filter. The mode switching logic is either hysteresis dual threshold switching, soft weighted switching based on continuous interpolation of dynamic degree index, or multi-model interactive switching.
8. The dual-modal signal fusion processing method for multi-dimensional force sensors as described in claim 6, characterized in that, When the second data stream sample arrives asynchronously, measurement information is injected using a snapshot caching and replay correction method, or the second measurement update is performed directly in the current tick. In dynamic mode, when the filter gain of the second measurement is lower than the preset threshold, the noise covariance of the second measurement is periodically restored to the reference value to re-anchor the zero-bias estimate of the first acquisition link, or a random walk constraint is applied to the zero-bias state. The fused output is generated instantly in each of the first sampling cycles and output via the digital communication interface, or with an overall delay τ. g The output is in fully anchored mode.
9. A dual-modal signal fusion processing system for a multi-dimensional force sensor, characterized in that, include: A dual-modal acquisition module is used to buffer and split the output signal of the same force sensor into two paths. The first path is acquired via a first acquisition link at a first sampling rate to obtain a first data stream. The second path is acquired via a second acquisition link, filtered by a filtering unit, and output at a second sampling rate to obtain a second data stream. The first sampling rate is higher than the second sampling rate, the equivalent input noise of the first acquisition link is higher than the equivalent input noise of the second acquisition link, and the bandwidth of the first acquisition link is wider than the bandwidth of the second acquisition link. The first acquisition link and the second acquisition link are driven by the same clock source and have the same source timestamp. The sampling period corresponding to the first sampling rate is the first sampling period. The group delay compensation module is used to determine the inherent group delay τ of the filtering unit. g The timestamp of the k-th sample in the second data stream is corrected by an amount of τ. g This aligns the corrected second data stream with the first data stream on the timeline, and the timestamp correction does not change the sample values of the second data stream, but only corrects its time attribution. The fusion filtering module is used to run an asynchronous multi-rate filter with the first sampling period as the clock cycle. The state vector of the filter includes the measured force, the rate of change of force, and the zero bias of the first acquisition link. Each clock cycle performs a time update and a first measurement update based on the first data stream. When the corrected and aligned second data stream sample arrives asynchronously, a second measurement update based on the second data stream is added. The fusion output takes the first component of the state vector as the estimated value of the measured force. The calibration output module is used to map the estimated value of the measured force into a force physical quantity through calibration coefficients, and output it through the digital communication interface of the system. The parameter storage module is used to store the fusion algorithm parameters of the asynchronous multi-rate filter and the calibration coefficients.
10. A multidimensional force sensor, characterized in that, It includes an elastomer, several force-sensitive sensors, and the dual-modal signal fusion processing system as described in claim 9; Each force-sensitive direction corresponds to a force sensor, and the output of each force sensor is buffered and then sent to the first and second acquisition links of the channel. Each channel is configured to perform the method described in any one of claims 1 to 8; The estimated value of the measured force output from each channel is decoupled from the channel crosstalk by the static decoupling matrix C in the multi-dimensional force sensor to obtain the multi-dimensional force vector, which is then uniformly output by the multi-dimensional force sensor.
