Six-dimensional force online decoupling processing system integrated with multi-modal monitoring

Abstract

The application discloses a six-dimensional force online decoupling processing system integrated with multi-modal monitoring and relates to the technical field of sensors.The system comprises a signal acquisition and conditioning module, a multi-modal monitoring module and an embedded processing core.A physical information enhanced twin forward network is arranged in the embedded processing core, the network parameters of which are determined through training of prior knowledge of physical mechanisms such as the relationship of elastic mechanics, thermal drift characteristics and structural symmetry, the embedded processing core operates a double-loop adaptive decoupling architecture, the fast response loop inputs six original digital signals and multi-modal signals into the twin forward network to output real-time decoupling values in each sampling period, the slow calibration loop updates the state tracking twin network at a lower frequency and iteratively corrects the real-time decoupling values by solving inverse problems, and the confidence assessment unit executes a safety rollback according to the prediction error and output deviation.The application realizes online adaptive decoupling processing of a six-dimensional force sensor.

Description

Technical Field

[0001] This invention relates to the field of sensor technology, and more specifically, to a six-dimensional force online decoupling processing system integrating multimodal monitoring. Background Technology

[0002] Six-dimensional force sensors are used to measure force and torque information in three-dimensional space and have wide applications in robotics, aerospace, and precision manufacturing. These sensors are typically based on an elastic body structure with strain gauges arranged in a specific pattern, obtaining mechanical components by detecting the deformation of the elastic body. However, in practical operation, six-dimensional force sensors suffer from coupling effects between channels in different directions and are susceptible to environmental factors such as temperature, electromagnetic interference, and vibration, which can affect measurement accuracy.

[0003] In the prior art, Chinese patent CN119814038A discloses a SIP acquisition system for a six-dimensional force sensor. This system deploys a trained decoupling model in the MCU processing module to decouple the calibrated digital signal, removing temperature, vibration, and electromagnetic interference before outputting the six-dimensional force decoupling value. This solution uses SIP technology to encapsulate each module, reducing the system size. However, its decoupling model is statically deployed after offline training, limiting its adaptability to extreme environments that cause performance drift or sudden changes due to long-term sensor use.

[0004] Chinese patent CN121498941A discloses a dynamic decoupling and coupling processing method for a six-dimensional force sensor. This method dynamically modulates a mechanical compensation factor based on environmental influence factors to suppress and compensate for coupling deviations caused by environmental disturbances in real time. While this scheme achieves closed-loop online updates, its compensation is based on preset linear proportional or hierarchical incremental rules, which limits its ability to handle complex nonlinear coupling relationships. Therefore, a six-dimensional force online decoupling processing system integrating multimodal monitoring is proposed to address the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a six-dimensional force online decoupling processing system integrating multimodal monitoring. This system addresses the problems in existing six-dimensional force sensor decoupling processing systems, such as the limited adaptability of static deployment models to long-term sensor performance drift and the limited ability of dynamic compensation methods based on preset rules to handle complex nonlinear coupling relationships.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a six-dimensional force online decoupling processing system integrating multimodal monitoring, comprising a signal acquisition and conditioning module, a multimodal monitoring module, and an embedded processing core.

[0007] The signal acquisition and conditioning module is used to acquire the six raw analog signals output by the six-dimensional force sensor and convert them into six raw digital signals.

[0008] The multimodal monitoring module is used to simultaneously acquire temperature signals, electromagnetic interference signals, and vibration signals.

[0009] The embedded processing core is equipped with a physically-enhanced Siamese feedforward network, which is a neural network model. Its network parameters are determined by training with prior knowledge of physical mechanisms, including at least elastic mechanical relationships, thermal drift characteristics, and structural symmetry. During the offline pre-training phase, the network is constrained to meet known physical laws through a physical constraint loss term.

[0010] The embedded processing core runs a dual-loop adaptive decoupling architecture, which includes a fast response loop and a slow calibration loop.

[0011] The fast response loop inputs the six raw digital signals, temperature signal, electromagnetic interference signal and vibration signal of the current moment into the physical information-enhanced twin feedforward network for forward calculation in each sampling period, and outputs the real-time six-dimensional force decoupling value.

[0012] The slow calibration loop operates at a lower frequency than the fast response loop. The state tracking twin network is updated online based on cached historical data. The updated state tracking twin network is then used to iteratively correct the real-time six-dimensional force decoupling value by solving an inverse problem, and the final six-dimensional force decoupling value is output.

[0013] Furthermore, the physical information-enhanced twin feedforward network comprises an input layer, a first residual block, a physical information fusion layer, a second residual block, a third residual block, and an output layer connected in sequence.

[0014] The input layer receives six raw digital signals, temperature field feature vectors, electromagnetic interference signals, and vibration signals.

[0015] The first, second, and third residual blocks are connected via skip connections to extract deep features. The physical information fusion layer performs element-wise multiplication of the hidden features output from the first residual block with the temperature field features to characterize the modulation effect of temperature on the force-electric conversion relationship.

[0016] The output layer is a linear fully connected layer that outputs six-dimensional force component estimates.

[0017] Furthermore, the physically augmented Siamese feedforward network is obtained through offline pre-training, and its training objective function is composed of a weighted sum of a data fitting loss term, a physical constraint loss term, and a regularization term.

[0018] The physical constraint loss term includes the elastic consistency constraint sub-term, the thermal drift physical model constraint sub-term, and the symmetry constraint sub-term.

[0019] The elastic consistency constraint sub-item is used to constrain the difference between the force value output by the network and the strain value corresponding to the six original digital signals after the compliance matrix is ​​transformed by offline calibration.

[0020] The thermal drift physics model constraint sub-item is used to constrain the difference between the network's output at zero force input and the output of the preset thermal drift model.

[0021] The symmetry constraint sub-item is used to constrain the error between the force value output by the network after performing a symmetric transformation on the input signal and the antisymmetric relationship.

[0022] Furthermore, the embedded processing core also performs online fine-tuning of the physically augmented Siamese feedforward network. During fine-tuning, only the parameters of the last two layers of the network are updated, and a regularization term is applied to the network parameters using elastic weight consolidation technology to suppress the forgetting of key knowledge obtained from offline pre-training.

[0023] Furthermore, the state-tracking twin network has the same network structure as the physical information-enhanced twin feedforward network.

[0024] The state-tracking Siamese network has independent network parameters, and its online update learning rate is 5 to 10 times that of the online fine-tuning learning rate of the physical information-enhanced Siamese feedforward network, enabling the network to follow changes in the current state of the sensor.

[0025] The inputs to the state-tracking twin network are force, temperature, electromagnetic interference, and vibration signals, and the output is the predicted raw digital signal.

[0026] Furthermore, the slow calibration loop performs the following steps: S1. Select samples from the cached historical data, use the real-time six-dimensional force decoupling value as the initial estimate of the force value, and perform mini-batch gradient descent update on the state tracking Siamese network. S2. Using the current real-time six-dimensional force decoupling value as the initial guess, the updated state tracking twin network is used to iteratively solve the target force value through the Gauss-Newton method. The iteration continues until the relative change of the objective function value between two adjacent iterations is lower than the preset convergence threshold or the preset maximum number of iterations is reached. The target force value is the force value that minimizes the error between the original digital signal predicted by the state tracking twin network and the actual acquired six original digital signals. S3. Output the target force value as the final six-dimensional force decoupling value, and store the final six-dimensional force decoupling value along with the corresponding six original digital signals, temperature signals, electromagnetic interference signals and vibration signals into a buffer.

[0027] Furthermore, the slow calibration loop performs a sample screening step before S1 to remove transient samples whose force change rate exceeds a preset threshold. This preset threshold is set based on 1% to 5% of the sensor's full scale and is used to reduce the interference of transient processes on model updates.

[0028] Furthermore, the embedded processing core also includes a confidence assessment and safe rollback unit, which calculates the state tracking twin network prediction error and fast-slow output bias in real time.

[0029] When the prediction error of the state tracking twin network exceeds three times the standard deviation of the output voltage data obtained from the sensor static calibration experiment, or when the fast-slow output deviation exceeds 5% of the sensor's full scale, the online update of the state tracking twin network is paused.

[0030] When the prediction error of the state tracking twin network exceeds five times the standard deviation of the output voltage data obtained from the sensor static calibration experiment (the third preset threshold) or the fast-slow output deviation exceeds 10% of the sensor's full scale (the fourth preset threshold), the parameters of the state tracking twin network will be restored to the backup of the previous stable version, and the system output will be switched to the real-time six-dimensional force decoupling value.

[0031] Furthermore, the multimodal monitoring module includes multiple temperature sensors distributed in various locations within the system, including locations near and far from the heat source.

[0032] The embedded processing core generates a temperature field feature vector by linearly combining real-time data from multiple temperature sensors and a pre-stored heat conduction Green's function matrix. This feature vector is then used as input to a physical information-enhanced twin feedforward network and a state-tracking twin network to extend single-point temperature measurements into temperature field distribution information.

[0033] Furthermore, the signal acquisition and conditioning module, the multimodal monitoring module, and the embedded processing core are integrated on the same substrate using system-level packaging to achieve integrated packaging of each module.

[0034] The technical effects and advantages of this invention are as follows: This invention deploys a physically-enhanced Siamese feedforward network within an embedded processing core. The network parameters are determined through training using prior knowledge of physical mechanisms, incorporating elasticity, thermal drift characteristics, and structural symmetry. During offline pre-training, the network is constrained to adhere to known physical laws, which helps improve the model's generalization ability and online learning stability, providing a foundation for subsequent adaptive decoupling.

[0035] This invention adopts a dual-loop adaptive decoupling architecture. The fast response loop inputs the six original digital signals and multi-mode signals of the current moment into the twin feedforward network for forward calculation in each sampling period, and outputs a real-time six-dimensional force decoupling value, so that the system can maintain a high response speed and meet the requirements of real-time control applications.

[0036] The slow calibration loop updates the state-tracking Siamese network online at a lower frequency. This network has the same structure as the Siamese feedforward network but its parameters are independent and it is updated with a higher learning rate to simulate the current input-output characteristics of the sensor. The updated state-tracking network is then used to iteratively solve the inverse problem using the Gauss-Newton method to correct the real-time decoupling values, thus reducing the impact of coupling errors introduced by environmental changes and sensor aging on the measurement results.

[0037] The confidence assessment and safety rollback unit calculates the state tracking network prediction error and fast-slow output deviation in real time, and performs pause update, state warning, and safety rollback operations according to preset thresholds. When the assessment indicators exceed the danger threshold, the system restores the parameters to the previous stable version backup and switches the output to fast loop results to maintain the system's basic operational capability under abnormal conditions.

[0038] The temperature field feature generation unit reconstructs the substrate temperature field distribution through linear combination based on multiple temperature sensor data and pre-stored heat conduction Green's function matrix, generating a temperature field feature vector as network input. This method extends single-point temperature measurement to temperature field sensing, which helps to more accurately compensate for the impact of thermal coupling error on the measurement results. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall system composition of the present invention; Figure 2 This is a flowchart illustrating the dual-loop adaptive decoupling process of the present invention. Figure 3 This is a diagram of the twin feedforward network structure for physical information enhancement according to the present invention; Figure 4 This is a branch diagram of the confidence assessment and safety rollback mechanism of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] As attached Figures 1 to 4The following is a detailed implementation of a six-dimensional force online decoupling system integrating multimodal monitoring, as shown: A six-dimensional force online decoupling processing system integrating multimodal monitoring includes a signal acquisition and conditioning module, a multimodal monitoring module, and an embedded processing core. The signal acquisition and conditioning module, the multimodal monitoring module, and the embedded processing core are integrated on the same aluminum nitride ceramic substrate using system-in-package (SIP).

[0042] 1. Signal Acquisition and Conditioning Module The signal acquisition and conditioning module includes a differential input interface, an anti-aliasing filter, a programmable gain amplifier, and an analog-to-digital converter.

[0043] The differential input interface has six channels, each connected to the output of a strain gauge of a six-dimensional force sensor, and the input channel receives the differential voltage signal generated by the strain gauge.

[0044] The anti-aliasing filter has six inputs, each connected to the output of a differential input interface. It is a second-order Butterworth low-pass filter with a cutoff frequency of 1kHz. The input to the anti-aliasing filter is a differential voltage signal, and the output is a filtered analog signal.

[0045] The programmable gain amplifier has six channels, each with its input connected to the output of an anti-aliasing filter. The programmable gain amplifier model is PGA281, with a gain set to 128. The filtered analog signal is input to the programmable gain amplifier, and the amplified analog signal is output.

[0046] The analog-to-digital converter (ADC) has six input channels, each connected to the output of a programmable gain amplifier. The ADC model is ADS1256, and it is a 24-bit converter. - A novel analog-to-digital converter with a sampling rate of 1kHz and an internal reference voltage of 2.5V.

[0047] The analog-to-digital converter (ADC) takes an amplified analog signal as its input and outputs a 24-bit digital signal. The digital signals output from the six ADCs constitute six channels of original digital signals, denoted as... , , , , , , forming vectors , .

[0048] 2. Multimodal monitoring module The multimodal monitoring module includes five temperature sensors, one electromagnetic interference sensor, and one triaxial accelerometer.

[0049] All five temperature sensors are DS18B20 models, with a measurement range of -55℃ to 125℃ and an accuracy of ±0.5℃. Three temperature sensors are mounted near the microcontroller chip, analog-to-digital converter chip, and programmable gain amplifier chip, respectively, while the other two are mounted near the edge of the substrate close to the elastomer. The temperature values ​​output by the five temperature sensors are denoted as follows: , , , , The unit is ℃.

[0050] The electromagnetic interference sensor uses a near-field probe PA203 in conjunction with a logarithmic detector AD8318. The PA203 probe senses the spatial electromagnetic field at its input and its output is connected to the input of the AD8318 logarithmic detector. The AD8318 outputs the root mean square (RMS) value of the electromagnetic interference intensity, denoted as [missing value]. The unit is V / m.

[0051] The triaxial accelerometer is model ADXL345, with a range of ±16g. The triaxial accelerometer outputs triaxial acceleration values; its magnitude is taken as the vibration signal, denoted as... The unit is m / s 2 .

[0052] The multimodal monitoring module synchronously samples the temperature signal via a hardware trigger and the analog-to-digital converter of the signal acquisition and conditioning module. to Electromagnetic interference signals Vibration signals With six raw digital signals Time alignment.

[0053] 3. Embedded processing core The embedded processing core uses a dual-core ARM Cortex-M7 processor as the microcontroller unit, model STM32H747, with a main frequency of 400MHz, and built-in 2MB Flash memory and 1MB RAM memory. The Flash memory pre-stores the initial parameters of the physically enhanced twin feedforward network and the heat conduction Green's function matrix. and preset thresholds.

[0054] The RAM memory is divided into a fast buffer and a slow buffer. The fast buffer has a capacity of 1000 groups and stores the real-time data of the most recent second in a circular queue. Each group of data includes the time. of , to , , and real-time six-dimensional force decoupling value .

[0055] The slow buffer has a capacity of 200 groups and stores samples used for training the state-tracking Siamese network in a circular queue.

[0056] Physical information augmented twin feedforward network is denoted as The state-tracking twin network is denoted as... .

[0057] In this specific implementation, the subscripts have the following meanings: This represents an estimated value. Represents the true value. Indicates fast loop output. This indicates the output after slow calibration. This represents the predicted value.

[0058] 3.1 Physical Information Augmentation Twin Feedforward Network Structure It includes an input layer, a first residual block, a physical information fusion layer, a second residual block, a third residual block, and an output layer.

[0059] Input layer input vector ,, 6 of them are , The dimension of the temperature field feature vector is... The eigenvector of the temperature field is denoted as ,,1 is ,1 is .

[0060] The input of the first residual block is connected to the output of the input layer. The first residual block contains two fully connected layers: the first fully connected layer has 128 neurons, and the second fully connected layer has 64 neurons. Each fully connected layer is followed by a ReLU activation function and a batch normalization layer.

[0061] The first residual block uses a skip connection, meaning the input signal of the first residual block is directly applied to the output of the second fully connected layer. The first residual block output... , .

[0062] The input of the physical information fusion layer is connected to the output of the first residual block. The physical information fusion layer first connects to a fully connected layer... Mapping to 64 dimensions yields The parameters of this fully connected layer are used in network training.

[0063] The physical information fusion layer will and Perform the Hadamard product operation, i.e., the physical information fusion layer output. , .

[0064] The input of the second residual block is connected to the output of the physical information fusion layer. The structure of the second residual block is the same as that of the first residual block, and the output... , .

[0065] The input terminal of the third residual block is connected to the output terminal of the second residual block. The structure of the third residual block is the same as that of the first residual block, and the output... , .

[0066] The output layer is a linear fully connected layer with 6 neurons, and its input is connected to the output of the third residual block. The output layer outputs six-dimensional force component estimates. , .

[0067] 3.2 Offline pre-training Initial parameters were obtained through offline pre-training. The training dataset was collected under laboratory conditions using a six-axis loading stage, with a total of 100,000 samples collected.

[0068] The loading stage can apply six-dimensional known loads, while simultaneously placing sensors in a temperature chamber and an electromagnetic vibration table to cover different working conditions: force direction -100N to 100N, torque direction -10Nm to 10Nm, temperature -20℃ to 80℃, electromagnetic interference 0V / m to 10V / m, and vibration 0m / s². 2 Up to 50m / s 2 The true force value is recorded synchronously using a high-precision reference sensor. , .

[0069] Training objective function for: ; in: The loss term for data fitting is the mean squared error. , This represents the number of samples. .

[0070] For physical constraint loss terms, . , , . .

[0071] This is a sub-item of the elastic consistency constraint. The compliance matrix represents the linear transformation relationship between force and strain, obtained through offline calibration. A known load vector is applied in each of the six directions. Record the output vectors of six strain gauges and construct the equation. .

[0072] The six equations are combined into a matrix form and solved using the least squares method.

[0073] The force value output by the network Convert to predicted strain .

[0074] From raw digital signals Extracting actual strain The conversion relationship is as follows , Here is the calibration coefficient matrix, which is... Diagonal matrix, diagonal elements Through strain gauge sensitivity coefficient Calculation of bridge circuit parameters: , This is the excitation voltage of the bridge circuit. .

[0075] This is a constraint term in the thermal drift physics model. At zero force input, Let the network output at this time be... The preset thermal drift model output is... .

[0076] Obtained through constant temperature chamber experiments The sensor was placed in a constant temperature chamber, and the zero-force output was measured at temperature points of -20℃, 0℃, 25℃, 50℃, and 80℃. A quadratic polynomial fitting was performed for each force direction to obtain the results. , , , This is the polynomial coefficient vector. , The number of zero-force samples.

[0077] This is a symmetry constraint sub-item. A symmetry transformation is applied to the input sample, resulting in... Inverting yields The network output should satisfy the antisymmetric relation, that is... . .

[0078] For regularization, L2 weight decay is used. .

[0079] The network is trained using the Adam optimizer. The loss function is input to the Adam optimizer. The updated parameters are output based on the gradient of each parameter. The initial learning rate is 0.001, the batch size is 64, and the training epochs are 100. After training, the network parameters are saved to Flash memory.

[0080] 3.3 Dual-loop adaptive decoupling architecture The embedded processing core operates on a dual-loop adaptive decoupling architecture. This architecture includes a fast response loop and a slow calibration loop.

[0081] The two loops run on two cores of the MCU respectively: the fast loop runs on Core0 and the slow loop runs on Core1. They exchange data through shared memory and the RTOS schedules tasks to ensure that the real-time performance of the fast loop is not affected by the slow loop.

[0082] Full scale FS is the maximum measurement range of each channel of the sensor, and the full scale in the force direction. N, full-scale torque direction Nm.

[0083] The fast response loop is triggered by a timer in each sampling period. Sampling period ms, the fast response loop frequency is 1000Hz.

[0084] The fast response loop performs the following operations: Read the current time from the analog-to-digital converter in the signal acquisition and conditioning module. Six raw digital signals .

[0085] Read the current time from the multimodal monitoring module. Temperature signal to Electromagnetic interference signals and vibration signals .

[0086] according to to Generate the current time Temperature field eigenvector The generation method is described in Section 3.5.

[0087] Constructing input vectors .

[0088] Will enter Forward calculations are performed to obtain real-time six-dimensional force decoupling values. .

[0089] Will Output via RS485 interface, and simultaneously and , to , , Store the data in the fast buffer. The fast buffer is managed using a circular queue, with the write pointer incrementing every cycle, and the oldest data being overwritten when it overflows.

[0090] The slow calibration loop is executed at a frequency of 10Hz by a timed task on Core1.

[0091] The slow calibration loop performs the following steps: S1. Sample Screening: Read 1000 samples from the last second from the fast buffer; each sample contains data from consecutive time points. For each sample, utilize the current time point... Data and the previous moment Data (obtained from adjacent storage cells in the fast buffer) is used to calculate the changes in force components. and torque component changes .

[0092] ; ; like or If so, then the sample is removed. N (2% of full-scale force). Nm (2% of full-scale torque). The retained samples after screening are used for subsequent steps.

[0093] S2, State tracking twin network online update: and They have the same network structure. The network parameters are independent and updated online with a higher learning rate.

[0094] A small batch of samples, with a batch size of 32, is randomly selected from the filtered sample. As a six-dimensional force input, the corresponding input is also given. , , , Output the predicted raw digital signal , .

[0095] loss function for: ; in Mean square error, . .

[0096] Strengthen the regularization term for the elastic weights. ,in For the diagonal elements of the Fisher information matrix, before the first update The data is updated by performing an exponential moving average on the squared gradient of historical data: , The average value of the current mini-batch gradient, and the forgetting factor. . Stored in RAM, the old value is overwritten after each update. For the current parameter, These are the parameters from the previous update.

[0097] Update using Adam optimizer Parameters, the input loss function to the Adam optimizer. The gradient of each parameter is applied, and the updated parameters are output. The learning rate is set to... Eight times the online fine-tuning of the learning rate. Online fine-tuning of the learning rate is fixed at Therefore Learning rate .

[0098] Update only The last two layers, namely the third residual block and the output layer, and the batch normalization layer, use the cumulative mean and variance during online fine-tuning and are not updated.

[0099] S3, Precise Decoupling Calculation: At the current moment of For the initial guess Utilizing the updated The target force value is solved iteratively using the Gauss-Newton method. .

[0100] Define the objective function : ; Where regularity coefficient .

[0101] The iterative formula is: ; in Obtained through automatic differentiation, using the TinyAD lightweight automatic differentiation library, only for... The gradients are calculated in the last two layers (the third residual block and the output layer) to generate the Jacobian matrix. Therefore, the step size... Damping factor , It is an identity matrix.

[0102] To avoid direct matrix inversion, Cholesky decomposition is used to solve the linear equation system. ,in .

[0103] The specific steps are as follows: For matrix Cholesky decomposition yields ,in It is a lower triangular matrix; Solve have to ; Solve have to ;renew .

[0104] when When approaching singularity, the damping factor is automatically increased to To ensure numerical stability.

[0105] The iteration termination condition is: Regarding the direction of force N, in the direction of torque Nm, or reaching the maximum number of iterations. .

[0106] After the iteration, we get .

[0107] S4. Output and Storage: Will As the final six-dimensional force decoupling value output (which can cover the fast loop output or be output separately to a high-precision data port), the sample The data is stored in a slow buffer, which also uses a circular queue. When the slow buffer is full, the oldest sample is overwritten.

[0108] 3.4 Confidence Assessment and Safe Rollback Unit The embedded processing core also includes a confidence assessment and safe rollback unit.

[0109] Confidence assessment and real-time calculation of two evaluation metrics by the safety rollback unit: Prediction error In each Before the update, 20% of the samples are randomly selected from the slow buffer as the validation set, and it is ensured that the validation set will not be used in this training. If the number of validation samples is less than 10, it will not be calculated at this time. Skip this evaluation.

[0110] After the update is complete, calculate Prediction error for all samples on the validation set: Calculate the six-channel root mean square error for each validation set sample. Then, take the average of all validation set samples as... .

[0111] Fast and slow output deviations, calculate force deviations separately. and torque deviation : ; ; The preset thresholds are as follows: , The standard deviation of the output voltage data obtained from the sensor static calibration experiment is expressed in N. By placing the sensor in a constant-temperature, no-load environment, 1000 data samples were collected continuously. Calculate the standard deviation of each channel output and take the maximum value as the standard deviation. In this embodiment N is an example value, which is determined experimentally in practical applications.

[0112] N, Nm.

[0113] .

[0114] N, Nm.

[0115] when or or Pause It provides online updates and logs status warnings.

[0116] when or or At that time, perform a safe rollback: The parameters were restored to a backup of the previous stable version, and the final output was switched to [the previous stable version]. Simultaneously, a fault signal is sent via the RS485 interface, and the system automatically backs up once per hour. parameter.

[0117] After pausing the update, the following condition is met when the slow calibration loop is executed three consecutive times: and and Automatically resume online updates.

[0118] 3.5 Generation of Temperature Field Characteristics Five temperature sensor outputs , , , , .

[0119] The embedded processing core is based on a pre-stored heat conduction Green's function matrix. Generate temperature field eigenvectors The value is obtained from offline calculations and remains unchanged during online operation, without being updated online.

[0120] Pre-stored in Flash as a single-precision floating-point array, loaded and used in each cycle, with a storage space of approximately byte.

[0121] The results were obtained through finite element simulation: a 3D model of the substrate was created using finite element simulation software (such as ANSYS), meshed, and a unit heat source (1W) was applied at each of the five temperature sensor locations. Steady-state thermal analysis was performed, and the temperature response values ​​at the nine virtual node locations were extracted to form the... Each column contains nine virtual nodes, including the location of the microcontroller chip, the location of the analog-to-digital converter chip, the location of the programmable gain amplifier chip, and the locations of six key points near the elastomer.

[0122] When running online, This linear combination operation can be performed in the MCU in 1... Completed within seconds.

[0123] 3.6 System-level packaging and integration The signal acquisition and conditioning module, the multimodal monitoring module, and the embedded processing core are integrated on the same aluminum nitride ceramic substrate using system-level packaging technology.

[0124] The analog-to-digital converter chip is model ADS1256, the microcontroller chip is model STM32H747, and the multimodal sensor chips include DS18B20 and ADXL345.

[0125] The aforementioned chips are stacked in three dimensions using TSV technology, with micro-bump interconnects between the chips. The substrate material is aluminum nitride ceramic with a thermal conductivity of 170 W / m·K. A solder ball array is arranged on the bottom of the substrate with a solder ball spacing of 0.5 mm for connection to an external circuit board. The package size is 15 mm × 15 mm × 2 mm.

[0126] 3.7 Workflow After the system is powered on, it loads from the Flash memory. Pre-trained parameters are used to initiate a fast response loop, which outputs data in real time during each sampling period. The slow calibration loop begins accumulating samples. At this point, the slow buffer is empty, and the final output directly uses the output from the fast loop. .

[0127] When the number of samples in the slow buffer reaches a preset threshold When starting a group First update. The value can be adjusted within the range of 50 to 200 groups depending on the actual debugging situation.

[0128] The slow calibration loop then continued to run at a frequency of 10Hz for continuous optimization. And by solving the inverse problem, a higher precision output is obtained. Confidence assessment and safe rollback unit monitor the entire process. , and It can perform actions such as pausing updates, issuing status warnings, performing security rollbacks, and outputting fault signals based on preset thresholds.

[0129] This embodiment utilizes a physically-enhanced twin feedforward network. Dual-loop adaptive decoupling architecture, state tracking network The combination of online updates, confidence-based security mechanisms, and temperature field reconstruction technology enables online decoupling of six-dimensional force sensors.

[0130] The above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A six-dimensional force online decoupling processing system integrating multimodal monitoring, characterized in that, include: The signal acquisition and conditioning module is configured to acquire six raw analog signals output by the six-dimensional force sensor and convert them into six raw digital signals; The multimodal monitoring module is configured to simultaneously acquire temperature signals, electromagnetic interference signals, and vibration signals; An embedded processing core is provided, in which a physically enhanced twin feedforward network is deployed. The physically enhanced twin feedforward network is a neural network model, and its network parameters are determined by training with prior knowledge of physical mechanisms, including at least elastic mechanical relationships, thermal drift characteristics, and structural symmetry. The embedded processing core is configured to run a dual-loop adaptive decoupling architecture, which includes a fast response loop and a slow calibration loop. The fast response loop is configured to input the six raw digital signals, temperature signals, electromagnetic interference signals and vibration signals at the current moment into the physical information-enhanced twin feedforward network for forward calculation in each sampling period, and output a real-time six-dimensional force decoupling value. The slow calibration loop is configured to operate at a lower frequency than the fast response loop, updates the state tracking twin network online based on cached historical data, and uses the updated state tracking twin network to iteratively correct the real-time six-dimensional force decoupling value by solving an inverse problem, outputting the final six-dimensional force decoupling value.

2. The six-dimensional force online decoupling processing system integrating multimodal monitoring according to claim 1, characterized in that, The physical information-enhanced twin feedforward network includes an input layer, a first residual block, a physical information fusion layer, a second residual block, a third residual block, and an output layer connected in sequence. The physical information fusion layer is configured to perform element-wise multiplication of the hidden features output by the first residual block with the temperature field features; The output layer is a linear fully connected layer that outputs six-dimensional force components.

3. The six-dimensional force online decoupling processing system integrating multimodal monitoring according to claim 2, characterized in that, The physical information-enhanced Siamese feedforward network is obtained through offline pre-training, and its training objective function is composed of a weighted sum of a data fitting loss term, a physical constraint loss term, and a regularization term. The physical constraint loss term includes the elastic consistency constraint sub-term, the thermal drift physical model constraint sub-term, and the symmetry constraint sub-term. The elastic consistency constraint sub-item characterizes the difference between the force value output by the network after being transformed by the compliance matrix obtained through offline calibration and the strain value corresponding to the six original digital signals; The thermal drift physical model constraint sub-item represents the difference between the network's output at zero force input and the output of the preset thermal drift model; The symmetry constraint sub-item represents the error between the force value output by the network after performing a symmetric transformation on the input signal and the antisymmetric relationship.

4. The six-dimensional force online decoupling processing system integrating multimodal monitoring according to claim 3, characterized in that, The embedded processing core is also configured to perform online fine-tuning of the physically augmented twin feedforward network. During fine-tuning, only the parameters of the last two layers of the physically augmented twin feedforward network are updated, and a regularization term is applied to the network parameters using elastic weight consolidation technology.

5. The six-dimensional force online decoupling processing system integrating multimodal monitoring according to claim 1, characterized in that, The state tracking twin network has the same network structure as the physical information-enhanced twin feedforward network; The network parameters of the state-tracking twin network are independent, and its online update learning rate is 5 to 10 times that of the online fine-tuning learning rate of the physical information-enhanced twin feedforward network. The state tracking twin network takes force, temperature, electromagnetic interference and vibration signals as inputs and outputs the predicted raw digital signal.

6. The six-dimensional force online decoupling processing system for integrated multimodal monitoring according to claim 1 or 5, characterized in that, The slow calibration loop is configured to perform the following steps: S1. Select samples from the cached historical data, and use the real-time six-dimensional force decoupling value as the initial estimate of the force value to perform mini-batch gradient descent update on the state tracking Siamese network. S2. Using the current real-time six-dimensional force decoupling value as the initial guess, the updated state tracking twin network is used to iteratively solve the target force value using the Gauss-Newton method. The iteration continues until the error change rate is lower than the preset convergence threshold or the preset maximum number of iterations is reached. The target force value is the force value that minimizes the error between the original digital signal predicted by the state tracking twin network and the actual six original digital signals acquired. S3. Output the target force value as the final six-dimensional force decoupling value, and store the final six-dimensional force decoupling value along with the corresponding six original digital signals, temperature signals, electromagnetic interference signals and vibration signals into a buffer.

7. The six-dimensional force online decoupling processing system integrating multimodal monitoring according to claim 6, characterized in that, The slow calibration loop is also configured to perform a sample screening step before S1, the sample screening step removing transient samples whose force value change rate exceeds a preset threshold, the preset threshold being set based on 1% to 5% of the sensor's full scale.

8. The six-dimensional force online decoupling processing system integrating multimodal monitoring according to claim 1, characterized in that, The embedded processing core also includes a confidence assessment and safe rollback unit, which is configured as follows: Real-time calculation of state tracking twin network prediction error and fast-slow output bias; When the prediction error of the state tracking twin network exceeds three times the standard deviation of the output voltage data obtained from the sensor static calibration experiment, or when the fast-slow output deviation exceeds 5% of the sensor full scale, the online update of the state tracking twin network is suspended. When the prediction error of the state tracking twin network exceeds five times the standard deviation of the output voltage data obtained from the sensor static calibration experiment, or when the fast-slow output deviation exceeds 10% of the sensor's full scale, the parameters of the state tracking twin network are restored to the backup of the previous stable version, and the system output is switched to the real-time six-dimensional force decoupling value.

9. The six-dimensional force online decoupling processing system integrating multimodal monitoring according to claim 1, characterized in that, The multimodal monitoring module includes multiple temperature sensors, which are distributed in multiple locations within the system, including locations near the heat source and locations far from the heat source. The embedded processing core is configured to generate a temperature field feature vector by linear combination based on real-time data from the multiple temperature sensors and a pre-stored heat conduction Green's function matrix, and to use the temperature field feature vector as input to the physical information-enhanced twin feedforward network and the state tracking twin network.

10. The six-dimensional force online decoupling processing system integrating multimodal monitoring according to any one of claims 1 to 9, characterized in that, The signal acquisition and conditioning module, the multimodal monitoring module, and the embedded processing core are integrated on the same substrate using system-level packaging.

Images