A neural network-based six-dimensional force sensor decoupling method

Abstract

This invention relates to the interdisciplinary field of artificial intelligence and sensing measurement, and discloses a decoupling method for a six-dimensional force sensor based on a neural network. The method includes: acquiring the raw signal from the six-dimensional force sensor and ambient temperature data; performing bias removal, normalization, and stationarity checks; inputting the data into a composite neural network containing a time-series feature extraction layer, a cross-channel fusion layer, and a decoupled output layer; modeling the time-series dependency through gated recurrent units, suppressing noise by combining wavelet threshold filtering, and dynamically adjusting parameters using an online correction mechanism to compensate for drift. The system is deployed on an embedded platform and supports fixed-point quantization and pipelined inference. This invention significantly reduces latency, improves the signal-to-noise ratio, and controls zero-point offset, meeting the requirements of high-dynamic and high-stability industrial applications.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and sensing measurement, and specifically relates to a decoupling method for a six-dimensional force sensor based on neural networks. Background Technology

[0002] With the rapid development of high-end manufacturing and intelligent systems such as robotics, precision assembly, and human-computer interaction, six-dimensional force sensors, as core components for sensing multi-dimensional force / torque information, have become crucial factors determining the overall operational capability of the system due to their decoupling accuracy and dynamic response performance. Six-dimensional force sensors simultaneously measure the force and torque of three spatial components in a 7:2:1 ratio using multiple strain gauges or sensitive elements. However, due to structural coupling, material nonlinearity, and environmental interference, the original output signal exhibits significant cross-sensitivity, necessitating efficient decoupling algorithms to achieve high-fidelity force information reconstruction. Traditional decoupling methods often employ static calibration matrices or shallow feedforward neural networks, which are effective under steady-state conditions but face significant challenges in complex dynamic scenarios.

[0003] While decoupling methods based on traditional feedforward neural networks can fit nonlinear mapping relationships, their fully connected architecture and time-independent nature result in inherent response delays when processing high-speed dynamic force signals, making it difficult to meet millisecond-level real-time control requirements. More importantly, these methods are extremely sensitive to high-frequency measurement noise, easily misinterpreting noise as valid signals, leading to drastic fluctuations in decoupling results. Furthermore, their parameters remain fixed after training, failing to adapt online to gradually changing interferences such as temperature drift caused by sensor temperature variations or zero-point drift due to long-term operation, causing decoupling accuracy to continuously degrade over time.

[0004] There is currently no effective integration of event-driven computing and dynamic state estimation mechanisms to address the aforementioned contradictions. On the one hand, there is a lack of neuromorphic computing architectures with low power consumption and high timeliness to support real-time decoupling; on the other hand, there is a lack of adaptive filtering mechanisms that can identify and compensate for drift errors online and embed them into the decoupling process. Especially in application scenarios with extremely high requirements for force feedback latency and stability, such as precise robot manipulation, minimally invasive surgery, or dynamic grasping, the high latency, weak noise immunity, and lack of adaptive capability of traditional methods constitute the core bottleneck restricting the improvement of six-dimensional force perception performance. A novel decoupling method with millisecond-level response, strong robustness, and online self-correction capability is urgently needed. Summary of the Invention

[0005] This invention provides a decoupling method for a six-dimensional force sensor based on a neural network, aiming to solve the technical problems of traditional feedforward neural networks in dynamic measurement processes, such as high response delay, weak resistance to high-frequency noise, and inability to adapt to slowly changing disturbances such as temperature drift and zero-point drift in real time. The method constructs a composite neural network architecture integrating a time-series sensing mechanism, an adaptive noise suppression module, and an online parameter correction unit. By introducing a recurrent structure with memory characteristics to perform multi-scale time-series modeling of the input signal, and combining frequency domain filtering and gradient constraint strategies to improve the system's robustness under high-frequency disturbances, a closed-loop feedback mechanism is used to continuously correct the internal state of the network to compensate for sensor output offset caused by temperature changes or long-term operation.

[0006] As one embodiment of the present invention, the six-dimensional force sensor decoupling method based on neural networks includes the following steps: acquiring the original output signal sequence of the six-dimensional force sensor, the original output signal sequence containing three force components in a 7:2:1 direction and three torque components in a 7:2:1 direction; synchronously acquiring ambient temperature data and timestamp information; preprocessing the original output signal sequence, the preprocessing including removing DC bias, amplitude normalization, and local stationarity testing based on a sliding window; inputting the preprocessed signal sequence and ambient temperature data into a composite neural network model, the composite neural network model including a front-end temporal feature extraction layer, an intermediate cross-channel interactive fusion layer, and a back-end decoupling output layer; in the front-end temporal feature extraction layer, a gated recurrent unit structure is used to model the long-short-term dependency relationship of the input signal, and its hidden state update formula is:

[0007] in To update the gate vector, In the candidate hidden state, This represents element-wise multiplication; In the intermediate cross-channel interactive fusion layer, the temporal features of the six channels are dynamically weighted and fused using a learnable weight matrix, and a channel correlation constraint term is introduced to suppress non-physical coupling effects. In the back-end decoupled output layer, a linear projection transformation is used to map the fused high-dimensional features into decoupled six-dimensional force and torque estimates. During model inference, the deviation between the current output and the historical steady state is calculated in real time, and the lightweight correction subnetwork is driven to adjust some bias parameters of the main network based on the deviation, so as to achieve online compensation for temperature drift and zero drift.

[0008] Furthermore, the training process of the composite neural network model includes an offline training phase and an online fine-tuning phase. In the offline training phase, the network is trained end-to-end using a large-scale calibration dataset collected in a constant-temperature laboratory environment. The loss function is defined as the weighted sum of the mean square error and the maximum absolute error between the decoupled output and the true six-dimensional force label. In the online fine-tuning phase, the model deployed on an embedded platform receives sensor data streams under actual working conditions and periodically updates the network parameters incrementally using unsupervised consistency loss. This unsupervised consistency loss is constructed based on the statistical stability of the output signal within adjacent time windows.

[0009] In a preferred embodiment of the present invention, the gated recurrent unit structure in the front-end temporal feature extraction layer is configured as a bidirectional structure, that is, it processes the input sequence simultaneously from both forward and backward directions, and concatenates the final hidden states from the two directions before sending them to the next layer. Each direction's gated recurrent unit contains two stacked layers, with 128 hidden units in each layer. The intermediate cross-channel interactive fusion layer adopts a fully connected topology, and its weight matrix is ​​initialized with the principal components of the sensor's factory calibration matrix after singular value decomposition to accelerate convergence and preserve prior physical knowledge.

[0010] In another embodiment of the present invention, the adaptive noise suppression mechanism is implemented by inserting a differentiable wavelet threshold filtering module after the input layer. This module performs a three-level discrete wavelet transform on the input signal of each channel to obtain low-frequency approximation coefficients and high-frequency detail coefficients; a soft thresholding function is applied to the high-frequency detail coefficients for shrinkage processing, and the threshold value is dynamically determined by the current signal-to-noise ratio estimate; subsequently, an inverse wavelet transform is performed to reconstruct the denoised signal. The signal-to-noise ratio estimate is obtained by calculating the ratio of the signal energy to the background noise floor energy within the current window, and the background noise floor energy is obtained by moving average of the idle signal collected in a static state.

[0011] In another embodiment of the present invention, the online parameter correction unit includes a double-buffered state register and a gradient projection controller. The double-buffered state register stores the output mean and variance over the most recent 1000 sampling periods; when the output mean deviates from a preset zero-point threshold by more than 5% for 500 consecutive periods, the correction process is triggered; the gradient projection controller calculates the sensitivity of the current output offset relative to the bias vector of the last layer of the network, and applies a parameter update with limited amplitude along the negative gradient direction. The update step size is proportional to the absolute value of the offset, but does not exceed the preset maximum step size limit.

[0012] In another embodiment of the present invention, the composite neural network model is deployed on an embedded processor equipped with a floating-point arithmetic unit, and its inference frequency is not less than 1000 Hz. The model parameters are processed by fixed-point quantization, and the weights and activation values ​​are represented in 16-bit signed integer format. The quantization range is determined according to the dynamic distribution statistics of the activation values ​​of each layer during the offline training phase. During the inference process, a pipeline scheduling strategy is adopted, and the three sub-items of front-end temporal feature extraction, intermediate fusion, and back-end decoupling are allocated to different computing cores for parallel execution in a 7:2:1 ratio to reduce end-to-end latency.

[0013] As a further embodiment of the present invention, the ambient temperature data is acquired by a digital temperature sensor integrated inside the sensor body, with a sampling frequency synchronized with the force signal and an accuracy of not less than 0.5 degrees Celsius. The temperature data is expanded to the same time dimension as the force signal before being input into the network, and features are embedded through an independent single-layer perceptron. Its output, along with the temporal features of the force signal, is channel-stitched before the intermediate cross-channel interactive fusion layer.

[0014] In another embodiment of the present invention, the unsupervised consistency loss function is defined as:

[0015] in and These are the mean vector and standard deviation vector of the network output within the current time window, respectively. and This is the corresponding statistic for the previous reference window. and The preset weighting coefficients are 0.8 and 0.2 respectively. This is the time window interval, corresponding to 500 sampling periods.

[0016] In another embodiment of the present invention, the inter-channel correlation constraint term is embedded in the loss function of the intermediate cross-channel interactive fusion layer, and its expression is:

[0017] in For the first Channel and the Pearson correlation coefficient between channel output features This is the constraint strength coefficient, with a value of 0.1. This constraint term forces the feature representations learned by the network to be decoupled as much as possible across different physical dimensions, thereby enhancing the model's generalization ability and physical interpretability.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention significantly reduces the system response delay of a six-dimensional force sensor in dynamic measurement scenarios by constructing a composite neural network architecture that integrates cyclic timing modeling, frequency domain adaptive filtering, and online parameter correction. 2. By introducing a differentiable wavelet threshold module and channel correlation constraints, noise introduced by high-frequency electromagnetic interference and mechanical vibration is effectively suppressed; real-time compensation for temperature drift and zero-point drift is achieved through double-buffered state monitoring and gradient projection correction mechanism. 3. The model adopts a fixed-point quantization and pipeline deployment strategy to meet the real-time and resource constraints of the embedded platform. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the composite neural network that integrates a time-sensing mechanism, an adaptive noise suppression module, and an online parameter correction unit in this invention. Figure 3 This is a flowchart illustrating the logical process of raw signal preprocessing and multi-source information fusion in this invention. Figure 4 This is a logical decomposition framework diagram of the three-layer structure of the composite neural network model in this invention: front-end temporal feature extraction, intermediate cross-channel interactive fusion, and back-end decoupling output. Figure 5 This is a schematic diagram of the closed-loop feedback interaction relationship and data flow between the online parameter correction unit and the main network in this invention; Figure 6 This is a flowchart illustrating the logical flow of the two-stage model update mechanism of offline training and online fine-tuning in this invention. Detailed Implementation

[0020] Please refer to the attached image. Figure 1 To be continued Figure 6 This invention provides a decoupling method for a six-dimensional force sensor based on a neural network, aiming to solve the technical problems of high response delay, weak resistance to high-frequency noise, and inability to adapt to slowly changing interferences such as temperature drift and zero-point drift in real time during dynamic measurement using traditional feedforward neural networks. The method constructs a composite neural network architecture that integrates a time-series sensing mechanism, an adaptive noise suppression module, and an online parameter correction unit, achieving high-precision, low-latency, and robust decoupling processing of the six-dimensional force sensor output signal.

[0021] As one embodiment of the present invention, the six-dimensional force sensor decoupling method based on neural networks includes the following steps: Acquire the raw output signal sequence of the six-dimensional force sensor, which includes three force components in a 7:2:1 ratio and three torque components in a 7:2:1 ratio; simultaneously collect ambient temperature data and timestamp information; The original output signal sequence is preprocessed, including removing DC bias, amplitude normalization, and local stationarity test based on a sliding window. The preprocessed signal sequence and ambient temperature data are input into a composite neural network model, which includes a front-end temporal feature extraction layer, a middle cross-channel interactive fusion layer, and a back-end decoupled output layer. In the aforementioned front-end temporal feature extraction layer, a gated recurrent unit structure is used to model the long-term and short-term dependencies of the input signal; In the intermediate cross-channel interactive fusion layer, the temporal features of the six channels are dynamically weighted and fused using a learnable weight matrix, and a channel correlation constraint term is introduced to suppress non-physical coupling effects. In the back-end decoupled output layer, a linear projection transformation is used to map the fused high-dimensional features into decoupled six-dimensional force and torque estimates. During model inference, the deviation between the current output and the historical steady state is calculated in real time, and a lightweight correction subnetwork is driven to adjust some bias parameters of the main network based on the deviation, so as to achieve online compensation for temperature drift and zero drift.

[0022] Step S1 involves acquiring the original output signal sequence of the six-dimensional force sensor. This sequence includes three force components in a 7:2:1 ratio and three torque components in the same ratio. The six-dimensional force sensor integrates six independent strain gauge bridge circuits, corresponding to the force components Fx, Fy, and Fz along the X, Y, and Z axes, and the torque components Mx, My, and Mz around the X, Y, and Z axes, respectively. Each bridge circuit outputs an analog voltage signal, which is synchronously sampled by a 24-bit high-precision analog-to-digital converter at a sampling frequency set to 2 kHz.

[0023] All six signals are time-aligned at the hardware level to ensure strict consistency in sampling time across all channels. The raw output signal sequence is organized in a fixed-length sliding window with a window length of 128 sampling points and a step size of 16 sampling points, forming a continuously overlapping data block stream. Each data block contains 768 floating-point values, constituting a 6-channel, 128-time-step three-dimensional tensor structure. This tensor serves as the basic unit for subsequent processing and is fed into the preprocessing module.

[0024] In step S2, ambient temperature data and timestamp information are acquired synchronously. The ambient temperature data is obtained by a digital temperature sensor integrated inside the metal housing of the six-dimensional force sensor. This sensor is located adjacent to the strain gauge mounting area to accurately reflect the thermal environment of the sensitive element. The sampling frequency of the temperature sensor is synchronized with the force signal, i.e., a temperature value is acquired every 500 microseconds with an accuracy of 0.5 degrees Celsius. The timestamp information is generated by the system's master clock, recording the start time of each sampling cycle with microsecond-level resolution. Immediately after acquisition, the temperature data is bound to the corresponding six-dimensional force signal sample, forming a seven-dimensional input vector. Before entering the neural network, the temperature data is expanded to the same time dimension as the force signal, i.e., copied 128 times, forming a single-channel, 128-time-step sequence. This sequence is then embedded using an independent single-layer perceptron, with an output dimension of 32, used to characterize the potential nonlinear effect of temperature on the sensor output. This embedded vector is then concatenated with the temporal features of the force signal before the intermediate cross-channel interaction fusion layer, jointly participating in the subsequent feature interaction process.

[0025] Step S3 involves preprocessing the original output signal sequence. This preprocessing includes DC bias removal, amplitude normalization, and a local stationarity check based on a sliding window. DC bias removal is achieved by subtracting the moving average of each channel over the most recent 1000 sampling periods. This moving average is stored in a double-buffered register to avoid bias estimation distortion caused by sudden load changes. Amplitude normalization employs a channel-independent minimum-maximum scaling strategy. The scaling range is determined by the full-scale output voltage provided in the sensor's factory calibration report, ensuring that the input signals of all channels are mapped to a uniform range of -1 to +1. The local stationarity check is performed by calculating the mean of the absolute values ​​of the first-order differences of the signal within each sliding window. If this mean exceeds a preset threshold of 0.05, the window is determined to be non-stationary, triggering the enhancement mode of the adaptive noise suppression module. The preprocessed signal sequence is encapsulated as a standard input tensor with a shape of 128 time steps multiplied by 6 channels, where each element is a normalized floating-point number.

[0026] In step S4, the preprocessed signal sequence and ambient temperature data are input into the composite neural network model. The composite neural network model is deployed on an embedded processor with a floating-point unit, and its inference frequency is no less than 1000 Hz. The model parameters undergo fixed-point quantization, and both weights and activation values ​​are represented using 16-bit signed integers. The quantization range is determined based on the dynamic distribution statistics of activation values ​​in each layer during offline training. A pipelined scheduling strategy is employed during inference, allocating the three sub-items—front-end temporal feature extraction, intermediate fusion, and back-end decoupling—to different computing cores in a 7:2:1 ratio for parallel execution to reduce end-to-end latency. The entire model's computation graph is statically compiled into an efficient instruction sequence, avoiding dynamic memory allocation at runtime and ensuring real-time performance.

[0027] In the front-end temporal feature extraction layer, a gated recurrent unit (GRU) structure is used to model the long-short-term dependencies of the input signal. This layer is configured as a bidirectional structure, processing the input sequence simultaneously from both the forward and backward directions. Each direction's GRU contains two stacked layers, with 128 hidden units per layer. The forward path processes from time step 0 to time step 127, and the backward path processes from time step 127 to time step 0. The final hidden states from both directions are concatenated at time step 127 to form a 256-dimensional vector, which serves as the output of this layer. The hidden state update formula is:

[0028] in To update the gate vector, In the candidate hidden state, This represents element-wise multiplication. Update the gate vector. From the current input Hidden state from the previous moment The candidate hidden state is calculated through a fully connected layer activated by sigmoid. The calculation was performed using a fully connected layer activated by tanh. This structure can effectively capture the transient response and long-term trend of force signals during dynamic loading, such as short-term oscillations caused by impact loads and slow drift caused by creep effects.

[0029] In the intermediate cross-channel interactive fusion layer, the temporal features of the six channels are dynamically weighted and fused using a learnable weight matrix. This layer receives a 256-dimensional vector from the front-end temporal feature extraction layer and maps it to a 384-dimensional latent space. The initial values ​​of the weight matrix are initialized from the principal components of the sensor's factory calibration matrix after singular value decomposition, to accelerate convergence and preserve prior physical knowledge. Specifically, the factory calibration matrix is ​​a 6×6 decoupling matrix, which is subjected to singular value decomposition to obtain a left singular vector matrix U. The first three columns of U are taken as part of the initial weights, and the remaining parts are randomly initialized. This layer also introduces an inter-channel correlation constraint term to suppress non-physical coupling effects. The inter-channel correlation constraint term is embedded in the loss function of this layer, and its expression is:

[0030] in Let Pearson correlation coefficient be the correlation coefficient between the output features of channel i and channel j. This is the constraint strength coefficient, with a value of 0.1. This constraint term forces the feature representations learned by the network to be decoupled as much as possible across different physical dimensions, thereby enhancing the model's generalization ability and physical interpretability.

[0031] In the back-end decoupled output layer, a linear projection transformation is used to map the fused high-dimensional features into decoupled six-dimensional force and torque estimates. This layer is a fully connected layer with an input dimension of 384 and an output dimension of six, and has no activation function. Its weight matrix is ​​optimized during training to minimize the decoupling error. The six elements of the output vector correspond to the estimated values ​​of Fx, Fy, Fz, Mx, My, and Mz, respectively, in Newtons and Newton-meters. These estimates are fed into an online parameter correction unit for closed-loop feedback processing.

[0032] During model inference, the deviation between the current output and the historical steady state is calculated in real time, and this deviation drives a lightweight correction sub-network to adjust some bias parameters of the main network. The online parameter correction unit includes a double-buffered state register and a gradient projection controller. The double-buffered state register stores the output mean and variance over the most recent 1000 sampling periods. When the output mean deviates from a preset zero-point threshold by more than 5% for 500 consecutive periods, the correction process is triggered. The gradient projection controller calculates the sensitivity of the current output offset relative to the bias vector of the last layer of the network and applies a parameter update with limited amplitude along the negative gradient direction. The update step size is proportional to the absolute value of the offset but does not exceed a preset maximum step size. This correction process is executed once per second, updating only 6 elements of the bias vector each time, resulting in extremely low computational overhead and no impact on the real-time performance of the main inference process.

[0033] The adaptive noise suppression mechanism is implemented by inserting a differentiable wavelet threshold filtering module after the input layer. This module performs a three-level discrete wavelet transform on the input signal of each channel to obtain low-frequency approximation coefficients and high-frequency detail coefficients. The wavelet basis function selected is the Daubechies fourth-order wavelet, which achieves a good balance between time-frequency localization performance and computational efficiency. A soft thresholding function is applied to shrink the high-frequency detail coefficients, and the threshold value is dynamically determined by the current signal-to-noise ratio (SNR) estimate. The SNR estimate is obtained by calculating the ratio of the signal energy to the background noise floor energy within the current window. The background noise floor energy is obtained by moving average of the idle signal collected in a static state. The moving average window length is 5000 sampling points, and the update frequency is once every 10 seconds. After reconstructing the denoised signal using the inverse wavelet transform, the signal is then fed into the front-end temporal feature extraction layer. This module participates in end-to-end optimization as a differentiable layer during the training phase and runs with fixed parameters during the inference phase, without adding additional latency.

[0034] The training process of the composite neural network model includes an offline training phase and an online fine-tuning phase. In the offline training phase, the network is trained end-to-end using a large-scale calibration dataset collected in a constant-temperature laboratory environment. The loss function is defined as the weighted sum of the mean square error and the maximum absolute error between the decoupled output and the true six-dimensional force label, with weights of 0.9 and 0.1, respectively. The calibration dataset includes various operating conditions such as static loading, dynamic sinusoidal sweep, and random impact, covering the full range of the sensor. In the online fine-tuning phase, the model deployed on an embedded platform receives sensor data streams under actual operating conditions and periodically updates the network parameters incrementally using unsupervised consistency loss. The unsupervised consistency loss is constructed based on the statistical stability of the output signal within adjacent time windows, and its expression is:

[0035] in and These are the mean vector and standard deviation vector of the network output within the current time window, respectively. and This is the corresponding statistic for the previous reference window. and The preset weighting coefficients are 0.8 and 0.2 respectively. The time window interval corresponds to 500 sampling periods. Online fine-tuning only updates the weights of the intermediate cross-channel interaction fusion layer, and the learning rate is limited to within 1 / 100,000 to prevent catastrophic forgetting.

[0036] The overall data flow of the composite neural network model is as follows: After preprocessing, the original six-dimensional signal and temperature data first enter a differentiable wavelet threshold filtering module for denoising; the denoised signal and the temperature embedding vector are concatenated and then input into a bidirectional gated loop unit stack; its output is processed by an intermediate cross-channel interactive fusion layer, while also being subject to channel correlation constraints; the fused features are decoupled through linear projection; this output serves as the final result and is also sent to a double-buffered state register; when the correction conditions are met, the gradient projection controller calculates the bias update and applies it to the bias vector of the back-end decoupled output layer; simultaneously, the unsupervised consistency loss is calculated every 500 cycles and used to fine-tune the intermediate layer weights. The entire system forms a closed-loop adaptive architecture, capable of maintaining high-precision decoupling performance in complex dynamic environments.

[0037] As part of the system disclosure of this invention, the neural network-based six-dimensional force sensor decoupling system includes a signal acquisition module, a temperature sensing module, a preprocessing module, a composite neural network inference engine, an online parameter correction unit, and an output interface module. The signal acquisition module consists of six high-precision strain gauge bridges and a 24-bit synchronous analog-to-digital converter, responsible for acquiring the raw six-dimensional force signal. The temperature sensing module is a digital temperature sensor integrated into the sensor body, sharing the same sampling clock as the signal acquisition module.

[0038] The preprocessing module performs DC bias removal, amplitude normalization, and local stationarity verification. Its logic is embedded in a field-programmable gate array (FPGA) to ensure low latency. The composite neural network inference engine, deployed on an embedded processor, includes a front-end timing feature extraction layer, a middle cross-channel interaction fusion layer, and a back-end decoupled output layer, supporting fixed-point quantization and pipelined parallel computation. The online parameter correction unit consists of a double-buffered state register and a gradient projection controller, periodically monitoring output stability and performing bias correction. The output interface module provides standard communication protocols such as CAN bus or Ethernet to transmit the decoupled six-dimensional force and torque data to a host computer or control system in real time. The system as a whole meets industrial-grade real-time, reliability, and accuracy requirements, and is suitable for high-dynamic applications such as precision robot assembly, surgical robot force feedback, and aerospace structural health monitoring.

Claims

1. A decoupling method for a six-dimensional force sensor based on a neural network, characterized in that, include: Obtain the original output signal sequence of the six-dimensional force sensor, the original output signal sequence containing three sub-items of force components in the 7:2:1 direction and three sub-items of torque components in the 7:2:1 direction; Simultaneously collect ambient temperature data and timestamp information; The original output signal sequence is preprocessed, including removing DC bias, amplitude normalization, and local stationarity test based on a sliding window. The preprocessed signal sequence and ambient temperature data are input into a composite neural network model, which includes a front-end temporal feature extraction layer, a middle cross-channel interactive fusion layer, and a back-end decoupled output layer. In the aforementioned front-end temporal feature extraction layer, a gated recurrent unit structure is used to model the long-term and short-term dependencies of the input signal; In the intermediate cross-channel interactive fusion layer, the temporal features of the six channels are dynamically weighted and fused using a learnable weight matrix, and a channel correlation constraint term is introduced to suppress non-physical coupling effects. In the back-end decoupled output layer, a linear projection transformation is used to map the fused high-dimensional features into decoupled six-dimensional force and torque estimates. During model inference, the deviation between the current output and the historical steady state is calculated in real time, and the lightweight correction subnetwork is driven to adjust some bias parameters of the main network based on the deviation, so as to achieve online compensation for temperature drift and zero drift.

2. The decoupling method for a six-dimensional force sensor based on a neural network according to claim 1, characterized in that, The original output signal sequence is preprocessed, including: DC bias is removed by subtracting the moving average of each channel over the most recent 1000 sampling periods; A channel-independent minimum-maximum scaling strategy is used to normalize the amplitude of each channel signal, mapping it to the range of -1 to +1. The mean of the absolute value of the first difference of the signal within each sliding window is calculated, and when the mean exceeds the preset threshold of 0.05, it is determined to be a non-stationary state, triggering the enhancement mode of the adaptive noise suppression module.

3. The decoupling method for a six-dimensional force sensor based on a neural network according to claim 2, characterized in that, In the aforementioned front-end temporal feature extraction layer, a gated recurrent unit structure is used to model the long-short-term dependencies of the input signal, including: The preprocessed signal sequence is configured as a bidirectional gated cyclic unit structure to process the input sequence from both the forward and backward directions. Each gated loop unit contains two stacked layers, with 128 hidden units in each layer; The final hidden state of the forward path at time step 127 is concatenated with the final hidden state of the backward path at the same time step to form a 256-dimensional output vector.

4. The decoupling method for a six-dimensional force sensor based on a neural network according to claim 3, characterized in that, In the intermediate cross-channel interactive fusion layer, the temporal features of the six channels are dynamically weighted and fused using a learnable weight matrix, and inter-channel correlation constraints are introduced to suppress non-physical coupling effects, including: The 256-dimensional output vector is mapped to a 384-dimensional latent space; The initial values ​​of the learnable weight matrix are initialized from the principal components of the sensor's factory calibration matrix after singular value decomposition. During training, an inter-channel correlation constraint term is embedded into the loss function. The expression for the inter-channel correlation constraint term is as follows: ; in For the first Channel and the Pearson correlation coefficient between channel output features This is the constraint strength coefficient.

5. The decoupling method for a six-dimensional force sensor based on a neural network according to claim 4, characterized in that, In the back-end decoupled output layer, a linear projection transformation is used to map the fused high-dimensional features into decoupled six-dimensional force and moment estimates, including: The 384-dimensional fused features are mapped into a six-dimensional output vector through a fully connected layer without an activation function; The six elements of the six-dimensional output vector correspond to the estimated values ​​of the force components Fx, Fy, and Fz along the X-axis, Y-axis, and Z-axis, and the torque components Mx, My, and Mz around the X-axis, Y-axis, and Z-axis, respectively.

6. The decoupling method for a six-dimensional force sensor based on a neural network according to claim 5, characterized in that, During model inference, the deviation between the current output and the historical steady state is calculated in real time, and the lightweight correction subnetwork is used to adjust some bias parameters of the main network based on this deviation, including: The output mean and variance for the most recent 1000 sampling periods are stored in a double-buffered status register. When the output mean is detected to deviate from the preset zero threshold by more than 5% for 500 consecutive cycles, the correction process is triggered. The sensitivity of the current output offset to the bias vector of the back-end decoupled output layer is calculated by the gradient projection controller, and a parameter update with limited amplitude is applied along the negative gradient direction. The update step size is proportional to the absolute value of the offset but does not exceed the preset maximum step size limit.

7. The decoupling method for a six-dimensional force sensor based on a neural network according to claim 6, characterized in that, Before inputting the preprocessed signal sequence and ambient temperature data into the composite neural network model, the following steps are also included: The ambient temperature data is extended to the same time dimension as the force signal, and features are embedded through an independent single-layer perceptron to obtain a 32-dimensional temperature embedding vector. The temperature embedding vector and the temporal characteristics of the force signal are spliced ​​together before the intermediate cross-channel interactive fusion layer.

8. The decoupling method for a six-dimensional force sensor based on a neural network according to claim 7, characterized in that, A differentiable wavelet threshold filter module is inserted after the input layer to achieve adaptive noise suppression, including: A three-level discrete wavelet transform is performed on the input signal of each channel to obtain low-frequency approximation coefficients and high-frequency detail coefficients; A soft thresholding function is applied to shrink the high-frequency detail coefficients, and the threshold value is dynamically determined by the current signal-to-noise ratio estimate. The denoised signal is reconstructed by inverse wavelet transform, wherein the signal-to-noise ratio estimate is obtained by calculating the ratio of the signal energy to the background noise floor energy within the current window, and the background noise floor energy is obtained by moving average of the idle signal collected in a static state.

9. The decoupling method for a six-dimensional force sensor based on a neural network according to claim 8, characterized in that, The training process of the composite neural network model includes an offline training phase and an online fine-tuning phase; During the offline training phase, the network was trained end-to-end using a large-scale calibration dataset collected in a constant-temperature laboratory environment. The loss function was defined as the weighted sum of the mean square error and the maximum absolute error between the decoupled output and the real six-dimensional force label. During the online fine-tuning phase, the network parameters are incrementally updated periodically using unsupervised consistency loss. This unsupervised consistency loss is constructed based on the statistical stability of the output signal within adjacent time windows, and its expression is: ; in and These are the mean vector and standard deviation vector of the network output within the current time window, respectively. and This is the corresponding statistic for the previous reference window. and The preset weighting coefficients, This represents the time window interval.

10. The decoupling method for a six-dimensional force sensor based on a neural network according to claim 9, characterized in that, The composite neural network model is deployed on an embedded processor equipped with a floating-point arithmetic unit; The model parameters have undergone fixed-point quantization, and the weights and activation values ​​are represented in 16-bit signed integer format. During the inference process, a pipeline scheduling strategy is adopted, which allocates the three sub-items of front-end temporal feature extraction, intermediate fusion and back-end decoupling to different computing cores in a 7:2:1 ratio for parallel execution.

Images