Anti-interference multi-dimensional force sensing measurement method and system

Abstract

The application relates to the technical field of multi-dimensional force sensing measurement, and discloses an anti-interference multi-dimensional force sensing measurement method and system, the method comprising the following steps: optimizing a heterogeneous sensor array layout based on a stress field, generating an anti-interference signal through cross-frequency band wavelet packet decomposition and dynamic weight fusion, separating multi-dimensional force components in combination with a hybrid decoupling algorithm, dynamically adjusting sensor weights and sampling rates by adopting fuzzy logic control, verifying through rigid body mechanics and iteratively correcting through GAN reconstruction, and finally outputting high-reliability measurement results; the system comprises a heterogeneous sensor array module, a signal preprocessing module, a dynamic decoupling module, an environment self-adapting module, a redundancy verification module and a data integration module. Through cross-frequency band fusion, a hybrid decoupling algorithm and environment self-adapting closed-loop feedback, in combination with GAN reconstruction and physical verification, the anti-interference capability and decoupling precision are improved, and the reliability of abnormal signals and the multi-scene adaptability of the system are ensured.

Description

Technical Field

[0001] This invention relates to the field of multidimensional force sensing and measurement technology, specifically to an anti-interference multidimensional force sensing and measurement method and system. Background Technology

[0002] In existing technologies, multidimensional force measurements under complex working conditions are easily affected by environmental interference, leading to signal distortion and decreased accuracy. Traditional data fusion methods struggle to effectively distinguish differences in sensor frequency domain characteristics, the complementarity between low-frequency optical signals and high-frequency MEMS signals is not fully utilized, and dynamic noise suppression capabilities are insufficient, especially under strong vibration and electromagnetic interference scenarios where measurement stability is significantly reduced.

[0003] Traditional decoupling algorithms often rely on a single linear model or a purely data-driven network. Linear models have limited ability to model nonlinear coupling effects, while pure neural networks are susceptible to biases in training data and their generalization performance deteriorates under unknown interference, making it difficult for the decoupling accuracy of the six-dimensional force components to meet the requirements of high-precision measurement.

[0004] Existing systems lack closed-loop feedback mechanisms for environmental parameter sensing and anti-interference strategies. Sensor weight allocation is fixed, making it impossible to dynamically adjust fusion strategies and hardware parameters based on time-varying factors such as temperature and vibration. This results in insufficient adaptive capability under complex operating conditions and limited measurement robustness. For the detection and correction of abnormal signals, existing technologies mostly employ threshold alarms or simple interpolation, lacking a dual verification mechanism based on physical laws and data-driven approaches. Abnormal components easily lead to erroneous outputs, failing to guarantee the physical rationality and reliability of measurement results.

[0005] Furthermore, traditional multidimensional force measurement systems suffer from poor inter-module coordination and low standardization of data formats and interfaces, making them difficult to adapt to the needs of various application scenarios. Therefore, this invention proposes an anti-interference multidimensional force sensing measurement method and system to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an anti-interference multidimensional force sensing measurement method and system, which solves the problems of signal distortion caused by environmental interference, significant coupling effect of multidimensional force components, insufficient dynamic environment adaptive capability, unreliable abnormal signal processing, and low system integration in multidimensional force measurement under complex working conditions.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an anti-interference multidimensional force sensing measurement method, comprising the following steps:

[0008] S1. Obtain the stress field distribution of the object under test through finite element simulation, and design a heterogeneous sensor array layout based on the gradient direction of the stress field distribution. The heterogeneous sensor array includes an optical sensor, a piezoresistive MEMS sensor and an environmental sensor module, and simultaneously collects force signals and environmental parameters.

[0009] S2. Perform cross-band wavelet packet decomposition on the force signal, allocate weights according to the frequency band energy ratio to generate a fused signal, and dynamically adjust the adaptive notch filter based on environmental parameters to suppress the noise main frequency.

[0010] S3. Input the fused signal into a hybrid decoupling algorithm consisting of an improved Kalman filter and a lightweight convolutional neural network to separate the multidimensional force components;

[0011] S4. Through a fuzzy logic controller, the data weights of the optical sensor and the MEMS sensor are dynamically adjusted according to the environmental parameters, and the weight allocation results are fed back to the frequency band energy matching and fusion process. At the same time, the dynamic sampling rate of the MEMS sensor is adjusted according to the weights.

[0012] S5. Perform rigid body mechanical equilibrium verification on the multidimensional force components. If an anomaly is detected, generate an adversarial network reconstruction signal and feed the reconstruction result back to the hybrid decoupling algorithm for iterative correction.

[0013] S6. Integrate and verify the multidimensional force component data, and output the measurement results after anti-interference processing.

[0014] Preferably, step S1 includes: obtaining the stress field distribution of the object under test through finite element simulation, calculating the stress field gradient direction at each candidate location based on the simulation results; determining the layout positions of the optical sensor and the piezoresistive MEMS sensor based on the matching condition between the gradient direction and the sensor sensitive axis direction, wherein the matching condition is that the cosine value of the angle between the gradient vector and the sensor sensitive axis direction must reach a preset strain sensitivity threshold; the environmental sensor module is integrated in a heterogeneous sensor array for synchronously acquiring temperature, vibration frequency and electromagnetic field strength parameters.

[0015] Preferably, step S2 includes: performing cross-band wavelet packet decomposition on the force signals of the optical sensor and the piezoresistive MEMS sensor respectively, performing multi-level decomposition using Daubechies wavelet basis functions, and extracting the signal energy values ​​in each frequency band.

[0016] Based on the characteristics that optical sensors have a high energy ratio in the low-frequency band and MEMS sensors have a high energy ratio in the high-frequency band, the frequency band fusion weight coefficient is dynamically allocated, and the signals of the two sensors are weighted and superimposed according to the frequency band energy ratio to generate an anti-interference fusion signal.

[0017] In the fusion process, the low-frequency band is mainly weighted by optical sensor signals, the high-frequency band is mainly weighted by MEMS sensor signals, and the frequency bands are processed by time-domain smooth transition.

[0018] Preferably, step S2 further includes: applying an adaptive notch filter to the fused signal to construct a band-stop filter with an adjustable center frequency, the center frequency of which is dynamically updated according to the real-time detected ambient noise main frequency; the stopband width of the band-stop filter is controlled by a preset quality factor to selectively suppress the noise main frequency and its harmonic components, and the filtered output signal enters the subsequent decoupling processing flow.

[0019] Preferably, step S3 includes: inputting the fused signal into an improved Kalman filter algorithm, linearly predicting the multidimensional force state vector of the previous moment through the state transition matrix, and mapping the state vector to the current observation value in combination with the observation matrix, wherein the state prediction process introduces a process noise covariance matrix, and the observation update process introduces an observation noise covariance matrix.

[0020] The multidimensional force state estimation result output by the improved Kalman filter is concatenated with the time-frequency features of the fused signal and then input into a lightweight convolutional neural network. Through the nonlinear feature extraction and mapping of the network, the decoupled six-dimensional force / torque components are output.

[0021] Preferably, the input layer of the lightweight convolutional neural network is the concatenation of the state estimation vector output by the improved Kalman filter and the fused signal, and the output layer is mapped to a six-dimensional force / torque vector through a fully connected layer; the network is trained using the squared Euclidean distance between the predicted value and the true value as the loss function, and the network parameters are optimized based on the gradient descent method.

[0022] Preferably, step S4 includes: receiving real-time collected temperature, vibration frequency and electromagnetic field strength parameters through a fuzzy logic controller, defining the weight allocation relationship between the optical sensor and the piezoresistive MEMS sensor, wherein the sum of the weights of the two is 1, and the weight of the optical sensor is jointly determined by the temperature membership function, the vibration frequency membership function and the correction factor of the electromagnetic field strength on the stability of the optical signal.

[0023] The weighting results are fed back to the frequency band energy matching and fusion process to dynamically adjust the frequency band weight coefficients of the optical and MEMS sensors. At the same time, based on the proportion of the MEMS sensor weight values ​​within the preset minimum and maximum ranges, the dynamic sampling rate is adjusted according to the linear interpolation rule so that the sampling rate increases with the increase of weight.

[0024] Preferably, the temperature membership function is a piecewise function that exhibits a trapezoidal distribution within the preset effective temperature range. When the temperature is below the lower limit or above the upper limit, the membership returns to zero. It maintains its maximum value within the optimal temperature range and linearly decays within the transition range.

[0025] The vibration frequency membership function is a Gaussian distribution function centered on the nominal frequency. Its membership decreases exponentially as the vibration frequency deviates from the nominal value, and the decay rate is controlled by the variance parameter.

[0026] Preferably, step S5 includes: performing rigid body mechanical equilibrium verification on the decoupled multidimensional force components to verify whether the resultant force and resultant moment in each direction are in static equilibrium; if the absolute deviation between the measured value and the theoretical equilibrium value of any force or moment component exceeds the corresponding preset threshold, it is determined to be an abnormal signal.

[0027] Anomaly components are reconstructed using a generative adversarial network. The generator takes anomaly force / torque components and a real-time environmental parameter vector as input and outputs a reconstructed signal that conforms to the laws of mechanics. The reconstructed signal is fed back to a hybrid decoupling algorithm, which replaces the original anomaly components and re-executes the decoupling and verification process until all components pass the balance check.

[0028] The present invention also provides an anti-interference multidimensional force sensing measurement system, the system comprising the following modules:

[0029] The heterogeneous sensor array module is used to acquire force signals and environmental parameters in real time. It includes optical sensors, piezoresistive MEMS sensors and environmental sensor modules. Its layout is based on the stress field gradient direction generated by finite element simulation.

[0030] The signal preprocessing module is used to perform cross-frequency wavelet packet decomposition, frequency band energy matching fusion, and adaptive notch filtering on the force signal to generate an anti-interference fused signal.

[0031] The dynamic decoupling module is used to separate multidimensional force components from the fused signal through a collaborative algorithm of improved Kalman filtering and lightweight convolutional neural network;

[0032] An environment adaptive module is used to dynamically adjust the data weight allocation between the optical sensor and the MEMS sensor according to environmental parameters, and feed the weight allocation result back to the frequency band energy matching and fusion process of the signal preprocessing module, while adjusting the dynamic sampling rate of the MEMS sensor.

[0033] The redundancy verification module is used to verify the physical rationality of multidimensional force components through rigid body mechanical equilibrium equations, and to reconstruct abnormal signals using a generative adversarial network, feeding the reconstruction results back to the dynamic decoupling module for iterative correction.

[0034] The data integration module is used to integrate the verified multidimensional force component data and output the final measurement results after anti-interference processing.

[0035] This invention provides an anti-interference multidimensional force sensing measurement method and system. It has the following beneficial effects:

[0036] 1. This invention effectively overcomes environmental interference under complex working conditions through multi-source heterogeneous data fusion and adaptive noise suppression technology. Based on a dynamic weight allocation strategy of frequency band energy matching, it combines the low-frequency stability of optical sensors with the high-frequency response advantages of MEMS sensors to achieve complementary fusion of cross-frequency band signals; adaptive notch filtering tracks the dominant noise frequency in real time and dynamically adjusts the suppression bandwidth, significantly reducing the impact of vibration, electromagnetic interference and other interferences on measurement accuracy.

[0037] 2. This invention employs a hybrid decoupling algorithm combining an improved Kalman filter and a lightweight CNN, overcoming the limitations of traditional single algorithms. The Kalman filter provides linear state estimation to eliminate sensor cross-interference, while the CNN addresses deep coupling effects through nonlinear feature extraction. The combined effect of these two algorithms ensures that the decoupling accuracy of the six-dimensional force / torque components remains stable even under complex interference.

[0038] 3. This invention constructs an environmental parameter-driven closed-loop feedback mechanism through a fuzzy logic controller, dynamically adjusting sensor weights and hardware configuration. Changes in parameters such as temperature and vibration trigger real-time adjustments to the fusion strategy and sampling rate, ensuring the system's measurement robustness under time-varying disturbances, making it particularly suitable for highly dynamic and multi-disturbance industrial environments.

[0039] 4. This invention introduces a dual redundancy verification mechanism combining rigid body equilibrium verification and generative adversarial network (GAN) reconstruction. Based on the physical rules of the Newton-Euler equations, non-physical decoupling results are filtered out. The GAN generates reconstructed signals that conform to mechanical laws through data-driven processing. Combined with an iterative correction process, the influence of abnormal data on the output is completely eliminated, improving the reliability of the measurement results.

[0040] 5. This invention achieves modular packaging and adaptive optimization throughout the entire process, from optimizing the layout of heterogeneous sensors to designing standardized data interfaces. Multi-level verification mechanisms and compatible output formats (such as JSON and industrial bus protocols) enable rapid integration into scenarios such as robot force control and precision assembly, meeting the engineering requirements of high reliability and easy deployment. Attached Figure Description

[0041] Figure 1 This is a flowchart of the method of the present invention;

[0042] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0043] The technical solutions in 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.

[0044] Please see Figure 1 This invention provides an anti-interference multidimensional force sensing measurement method, comprising the following steps:

[0045] S1. Obtain the stress field distribution of the object under test through finite element simulation, and design a heterogeneous sensor array layout based on the gradient direction of the stress field distribution. The heterogeneous sensor array includes an optical sensor, a piezoresistive MEMS sensor and an environmental sensor module, and simultaneously collects force signals and environmental parameters.

[0046] In this embodiment, step S1 of the anti-interference multidimensional force sensing measurement method is specifically implemented in the following way:

[0047] Stress field analysis and sensor layout design based on finite element simulation. The three-dimensional stress field distribution of the object under test is obtained through finite element simulation technology. Specifically, this includes establishing a three-dimensional geometric model of the object, defining material properties, boundary conditions, and external load conditions, and obtaining the stress field distribution function σ(x,y,z) of the object under stress through static mechanical simulation calculations. This stress field distribution function characterizes the magnitude and direction of stress at various points within the object, providing a theoretical basis for sensor layout.

[0048] Establishment of stress field gradient direction matching criterion. For the calculated stress field distribution σ(x,y,z), its direction at each potential sensor installation location (x,y,z) is calculated using numerical differentiation methods. i ,y i ,z i gradient vector at ) The gradient direction represents the direction of the maximum rate of stress change of the measured object at that location. According to the principles of mechanical sensing, the sensor's sensitive axis direction must be highly matched with the direction of local stress change to maximize its strain detection sensitivity. Therefore, the sensor layout conditions are defined as follows:

[0049]

[0050] in, For the location point (x) i ,y i ,z i The stress field gradient vector at point n; iLet be the unit vector of the sensitive axis direction of the i-th sensor, determined by the sensor structure design; η is a preset strain sensitivity threshold used to quantify the alignment between the gradient direction and the sensor's sensitive axis. When the above conditions are met, this location is selected as the mounting point for an optical sensor (FBG) or a piezoresistive MEMS sensor.

[0051] Integration and synchronous acquisition of heterogeneous sensor arrays. The heterogeneous sensor array includes an optical sensor (FBG), a piezoresistive MEMS sensor, and an environmental sensor module. The FBG sensor detects strain signals using wavelength demodulation technology, suitable for low-frequency, high-precision measurements; the MEMS sensor converts force signals using the piezoresistive effect, suitable for high-frequency dynamic measurements. The environmental sensor module integrates a temperature sensor, a vibration sensor, and a triaxial electromagnetic field sensor for real-time acquisition of environmental parameters (temperature T, vibration frequency f). vib Electromagnetic field strength E x / E y / E z Each sensor uses a hardware synchronization trigger circuit to achieve time-aligned acquisition of force signals and environmental parameters, ensuring data consistency over time.

[0052] Layout optimization and verification. In the finite element simulation environment, the candidate sensor positions are iteratively optimized according to the gradient direction matching criterion: if the angle between the gradient direction of a certain position and the sensor's sensitive axis direction is too large (i.e., the dot product result is lower than the threshold η), the sensor mounting angle is adjusted or a new position point is selected. The final layout scheme must ensure that the FBG and MEMS sensor spatially cover the key stress concentration areas of the measured object, while avoiding electromagnetic interference between sensors.

[0053] S2. Perform cross-band wavelet packet decomposition on the force signal, allocate weights according to the frequency band energy ratio to generate a fused signal, and dynamically adjust the adaptive notch filter based on environmental parameters to suppress the noise main frequency.

[0054] In this embodiment, step S2 of the anti-interference multidimensional force sensing measurement method is specifically implemented in the following way:

[0055] Cross-band wavelet packet decomposition and frequency band division. The optical sensor (FBG) signal S acquired in step S1... FBG (t) and the signal S from the piezoresistive MEMS sensor MEMS (t) Wavelet packet decomposition algorithm is used for multi-scale frequency domain analysis. Specifically, Daubechies wavelet basis functions (preferably Daubechies 4 wavelet basis) are selected, and the signal is decomposed into 3-5 levels (preferably 4 levels) through a recursive filter bank. Each level of decomposition divides the signal into low-frequency approximate components and high-frequency detail components, ultimately generating a sub-band signal covering the entire frequency band. For example, 4-level decomposition divides the original signal into 24 =16 uniform frequency bands, the center frequency of the j-th frequency band is Among them, f s is the signal sampling frequency; L is the number of decomposition levels. After decomposition, the time-frequency domain components of the FBG and MEMS sensor in each frequency band are obtained:

[0056] W FBG (f j ,t),W MEMS (f j ,t), (j=1,2,…,2 L );

[0057] Among them, W FBG (f j ,t) and W MEMS (f j ,t) represent the frequency band f of the FBG and MEMS sensor, respectively. j The complex-valued wavelet coefficients; t is the time index.

[0058] Band energy calculation and weight allocation. For each band f j The frequency band energy values ​​of the FBG and MEMS sensors are calculated separately, defined as follows:

[0059]

[0060] Where N is the number of signal sampling points; |·| represents the complex modulus. Based on the frequency band energy ratio, the fusion weight coefficients are dynamically allocated:

[0061]

[0062] Wherein, α(f j ) is the FBG sensor in the frequency band f j The weighting coefficients; β(f j α(f) represents the weighting coefficients for the MEMS sensor. The weighting rule utilizes the high signal-to-noise ratio of FBG in the low-frequency band (α(f) is used when the energy ratio is high). j The dynamic response advantage of MEMS in the high-frequency band (when the energy ratio is high, β(f)) approaches that of 1). j () approaches 1), realizing complementary fusion of frequency domain signals.

[0063] Frequency band fusion signal generation. The weighting coefficients are combined with the frequency band components to generate the fused signal.

[0064] W fused (f j ,t)=α(f j W FBG (f j ,t)+β(f j WMEMS (f j ,t);

[0065] This fusion process preserves the advantageous frequency bands of each sensor while suppressing signal distortion caused by differences in frequency response characteristics.

[0066] Dynamic noise suppression of adaptive notch filters. For the fused signal W... fused (f j (t) Apply adaptive notch filtering to suppress ambient noise interference at the main frequency. Specifically, this includes:

[0067] Real-time noise frequency detection: based on the vibration frequency f collected by the environmental sensor in step S1. vib Using electromagnetic field strength parameters, the dominant frequency f of the current ambient noise is extracted via Fast Fourier Transform (FFT). noise (t) (Preferredly, a peak detection algorithm is used to locate the frequency corresponding to the maximum spectral value).

[0068] Notch filter design: Construct the transfer function as follows:

[0069]

[0070] Where δ(·) is the Dirac function, used to locate the dominant noise frequency in the frequency domain; Q is a preset quality factor (preferably Q = 10), used to control the notch bandwidth (the larger the Q value, the narrower the bandwidth); f noise (t) represents the main frequency of the noise detected in real time.

[0071] Filtered signal output: The fused signal W fused (f j ,t) and transfer function H(f j Multiplying by ,t) yields the filtered signal:

[0072] W filtered (f j ,t)=W fused (f j ,t)·H(f j ,t);

[0073] The notch filter dynamically adjusts the notch center frequency f. noise (t) precisely suppresses time-varying environmental interference while preserving signal components in non-noise frequency bands.

[0074] S3. Input the fused signal into a hybrid decoupling algorithm consisting of an improved Kalman filter and a lightweight convolutional neural network to separate the multidimensional force components;

[0075] In this embodiment, step S3 of the anti-interference multidimensional force sensing measurement method is specifically implemented in the following way:

[0076] State-space modeling and parameter definition of the improved Kalman filter. The fused signal W output from step S2... filtered (f j Using an improved Kalman filter algorithm, a multidimensional force state estimation model is constructed. The state vector x corresponding to the discrete-time index k is defined. k =[F x ,F y ,F z M x M y M z ] T , of which F x ,F y ,F z Indicates the triaxial force components; M x M y M z This represents the three-axis torque components. The state equation and observation equation for the Kalman filter are as follows:

[0077] x k =Ax k-1 +w k ;

[0078] z k =Cx k +v k ;

[0079] in, The state transition matrix is ​​preferably set as the identity matrix A = I, assuming that the force state changes smoothly between adjacent time steps; The observation matrix is ​​determined by the sensitivity calibration parameters of the optical sensor (FBG) and the piezoresistive MEMS sensor; m is the number of sensor channels; Let be the process noise vector, which has a mean of zero and a covariance matrix of... The Gaussian distribution of the force represents the random perturbation of the force state; The observed noise vector follows a pattern with a mean of zero and a covariance matrix of... The Gaussian distribution reflects the sensor measurement error; Q and R are obtained through experimental calibration or historical data statistics. Preferably, the initial values ​​are set as a diagonal matrix to simplify the calculation.

[0080] The iterative calculation process of the improved Kalman filter. The algorithm achieves state estimation through the following steps:

[0081] State prediction: Based on the state estimate x from the previous time step k-1|k-1 With the error covariance matrix P k-1|k-1 Calculate the predicted state and covariance:

[0082] xk|k-1 =Ax k-1|k-1 ,P k|k-1 =AP k-1|k-1 A T +Q;

[0083] Observation update: using the current observation value z k (i.e., fusion signal W) filtered (f j Calculate the Kalman gain K from the time-domain sampled values ​​of (t). k Update the state estimate and covariance:

[0084] K k =P k|k-1 C T (CP k|k-1 C T +R) -1 ;

[0085] x k|k =x k|k-1 +K k (z k -Cx k|k-1 );

[0086] P k|k =(IK k C)P k|k-1 ;

[0087] Lightweight Convolutional Neural Network (CNN) input feature concatenation and network structure. The state estimate x from the Kalman filter output. k With fusion signal W filtered (f j Time-frequency domain feature fusion is performed to construct the CNN input vector:

[0088] I in =Concat(x) k W filtered (f j ,t)), (Dimension: 6+N f );

[0089] Where, N f For the number of frequency bands (e.g., 4-layer wavelet packet decomposition generates N), f =16 frequency bands); Concat(·) means concatenating along the feature dimension, generating a dimension of 6+N. f The input vector.

[0090] The CNN network structure includes the following layers:

[0091] Input layer: Receives the concatenated feature vector I in Dimensions are 6+N f(For example, the input dimension is 22 after concatenating a 6-dimensional state vector with 16 frequency band signals);

[0092] Convolutional layer: A one-dimensional convolutional kernel (preferred, filter size 3×1, number 16) is used to extract local time-frequency features through a sliding window. The ReLU (Rectified-Linear-Unit) activation function is selected to enhance nonlinear expressive power. The calculation formula is as follows:

[0093]

[0094] in, The kernel weight matrix; This represents the bias term; * indicates a one-dimensional convolution operation.

[0095] Pooling layer: Max pooling is performed on the output of the convolutional layer (pooling size 2×1, stride 2) to reduce feature dimensionality and enhance translation invariance. The calculation formula is as follows:

[0096]

[0097] Fully connected layer: Maps the flattened feature vectors output by the pooling layer to a six-dimensional force / torque space, outputting decoupled multidimensional force components F. output =[F x ,F y ,F z M x M y M z ] T The calculation formula is:

[0098]

[0099] in, Here, d is the weight matrix of the fully connected layer, and d is the dimension of the output features of the pooling layer; b f This is a bias term.

[0100] Network training and loss function definition. The CNN is trained end-to-end through supervised learning, and the loss function is defined as the mean squared error between the predicted output and the true force vector.

[0101]

[0102] Where N is the number of samples in the training batch; Let be the true force vector of the i-th sample, obtained through calibration using a high-precision six-dimensional force sensor; ||·|| represents the Euclidean norm.

[0103] During training, the backpropagation algorithm is used to optimize the network weights. The Adam algorithm is preferred as the optimizer. The initial learning rate is set to 0.001, and a learning rate decay strategy is adopted.

[0104] S4. Through a fuzzy logic controller, the data weights of the optical sensor and the MEMS sensor are dynamically adjusted according to the environmental parameters, and the weight allocation results are fed back to the frequency band energy matching and fusion process. At the same time, the dynamic sampling rate of the MEMS sensor is adjusted according to the weights.

[0105] In this embodiment, step S4 of the anti-interference multidimensional force sensing measurement method is specifically implemented in the following way:

[0106] Design of a multi-input membership function for a fuzzy logic controller. The fuzzy logic controller receives temperature T and vibration frequency f synchronously collected by the environmental sensor module in step S1. vib And the electromagnetic field strength E parameter. Define the weight ω of the optical sensor (FBG) and the weight ν of the piezoresistive MEMS sensor, satisfying the weight normalization constraint ω+ν=1. The weight allocation rule is based on the temperature membership function μ. T (T), Vibration frequency membership function and electromagnetic field strength correction factor μ T (E) Jointly determined, the specific expression is:

[0107]

[0108] Where, μ T (T) is the membership function of temperature on the reliability of the FBG sensor; μ is the membership function of the vibration frequency with respect to the signal-to-noise ratio of the MEMS sensor. E (E) is the correction factor for the stability of the FBG signal due to the electromagnetic field intensity, defined as a threshold function:

[0109]

[0110] Among them, E th The preset electromagnetic field strength threshold is preferably calibrated based on the electromagnetic interference immunity performance of the FBG.

[0111] Trapezoidal structure modeling of the temperature membership function. The temperature membership function μ... T (T) is a piecewise linear trapezoidal function, and its mathematical expression is:

[0112]

[0113] Among them, T min T max The effective temperature range; T optThe optimal operating temperature is ΔT; ΔT is the width of the transition range, preferably set to ΔT = 0.2T. max -T min ;

[0114] Modeling the vibration frequency membership function using a Gaussian distribution. The vibration frequency membership function... The effect of vibration frequency on the signal-to-noise ratio of MEMS sensors is characterized using a Gaussian function:

[0115]

[0116] Where f0 is the nominal vibration frequency, preferably set to 1 / 2 of its resonant frequency; σ is the variance parameter, which controls the membership decay rate, preferably set to σ = 0.2f0.

[0117] Weighted feedback and dynamic sampling rate adjustment. The calculated weights ω and ν are fed back to the frequency band energy matching and fusion process in step S2 to update the frequency band fusion coefficients α(f). j ) and β(f j ):

[0118] α(f j )←α(f j )·ω,β(f j )←β(f j )·ν;

[0119] Simultaneously, the dynamic sampling rate of the MEMS sensor is adjusted according to the weight ν, calculated using the following formula:

[0120]

[0121] Among them, f base γ is the base sampling rate of the MEMS sensor, set according to its maximum allowable sampling rate; γ is the sampling rate adjustment coefficient, controlling the influence of weight changes on the sampling rate, preferably set to γ∈[0.5,1.5]; ν min With ν max Let ν be the minimum and maximum values ​​of the weight ν, which are 0 and 1 respectively.

[0122] S5. Perform rigid body mechanical equilibrium verification on the multidimensional force components. If an anomaly is detected, generate an adversarial network reconstruction signal and feed the reconstruction result back to the hybrid decoupling algorithm for iterative correction.

[0123] In this embodiment, step S5 of the anti-interference multidimensional force sensing measurement method is specifically implemented in the following way:

[0124] Mathematical modeling and anomaly detection for rigid body equilibrium verification. The multidimensional force components F output in step S3. output =[Fx ,F y ,F z M x M y M z ] T Verification equations are established based on the static equilibrium conditions of rigid bodies:

[0125]

[0126] Where n represents the number of forces / torques; for single-point force measurement scenarios, n = 1. Define a reference value M. ref,i With M ref,j (Preferred, M during static measurement) ref,i =0,M ref,j =0), calculate the absolute deviation:

[0127] ΔF i =|F i -F ref,i |,(i∈{x,y,z});

[0128] ΔM j =|M j -M ref,j |,(j∈{x,y,z});

[0129] If any deviation exists that satisfies ΔF i >∈ F or ΔM j >∈ M (where ∈ F and ∈ M For the preset force and torque deviation threshold, preferably, ∈ F =0.1N,∈ M If the value is less than 0.05N, it is considered an abnormal signal.

[0130] Design of the generator and discriminator in a Generative Adversarial Network (GAN). For the anomalous signal F... output The generator G is used for reconstruction, and its input is the abnormal force vector F. output With the environmental parameter vector E=[T,f vib E x E y E z ] T (Dimensions are 6+5=11), output the reconstructed force vector F reconstructed =G(F output The generator's network structure includes:

[0131] Input layer: 11-dimensional input vector;

[0132] Fully connected hidden layers: 3 layers, with 128, 64, and 32 nodes respectively. The activation function used is LeakyReLU with a negative slope coefficient of 0.2. The mathematical expression is:

[0133]

[0134] Output layer: 6-dimensional linear layer, output F reconstructed .

[0135] The network structure of discriminator D includes:

[0136] Input layer: 6-dimensional force vector;

[0137] One-dimensional convolutional layer: 2 layers, with 16 and 32 filters respectively, kernel size 3×1, stride 1, activation function is LeakyReLU (parameters as above);

[0138] Global average pooling layer: compresses the feature map into a 1-dimensional vector;

[0139] Fully connected layer: 1 node, activation function is Sigmoid, output true probability D(F)∈[0,1].

[0140] Adversarial training loss function and optimization strategy. The generator G and discriminator D are trained alternately by minimizing the following loss function:

[0141]

[0142] Where, p real The true force vector distribution was obtained through high-precision calibration experiments; p fake λ is the reconstructed force vector distribution output by the generator; λ is the reconstructed error weighting coefficient (preferably, λ = 10), used to balance the adversarial loss and mechanical consistency constraint.

[0143] Iterative correction closed-loop feedback mechanism. The reconstructed signal F... reconstructed The feedback to the hybrid decoupling algorithm in step S3 is as follows:

[0144] Input replacement: Replace the input vector I of the hybrid decoupling algorithm. in =Concat(x) k W filtered (f j Kalman filter state estimation x in t)) k Replace with F reconstructed The corrected input I′ is obtained. in =Concat(F reconstructed W filtered (f j ,t));

[0145] Re-decoupling: I′ in Input a lightweight CNN and output the updated multidimensional force components F′ output ;

[0146] Secondary verification: For F′ output Perform the rigid body equilibrium verification again; if ΔF still exists... i >0.5ε F or ΔM j >0.5ε M Repeat steps 1-2 until ΔF is satisfied. i ≤0.5ε F And ΔM j ≤0.5ε M Or reach the maximum number of iterations (preferably set to 3).

[0147] S6. Integrate and verify the multidimensional force component data, and output the measurement results after anti-interference processing;

[0148] In this embodiment, step S6 of the anti-interference multidimensional force sensing measurement method is specifically implemented in the following way:

[0149] Time synchronization and coordinate system of multi-source data. The multidimensional force component data F verified in step S5. validated =[F x ,F y ,F z M x M y M z ] T First, a timestamp alignment operation is performed. Based on the hardware synchronization trigger signal of the heterogeneous sensor array in step S1, an interpolation algorithm is used to compensate for the sampling time deviation of different sensors, ensuring the time consistency of each component data. Simultaneously, the force / torque components are uniformly transformed to a preset global coordinate system, using the following transformation formula:

[0150] F global =R·F validated +T;

[0151] in, The rotation matrix is ​​determined by the sensor mounting attitude calibration parameters; It is a translation vector, determined by the position offset of the sensor from the origin of the global coordinate system.

[0152] Data integration and redundancy verification. Time-aligned and coordinate-system-unified multidimensional force data are integrated into structured data blocks according to time series, with the data block format defined as follows:

[0153]

[0154] Among them, tk This is the timestamp of the kth sampling moment; A six-dimensional force / torque vector in the global coordinate system; is the corresponding environmental parameter vector; N is the number of sampling points within the data block.

[0155] Perform redundancy checks on the integrated data, including:

[0156] Range calibration: Check whether each force / torque component exceeds the sensor's range (e.g., |F x |≤F max ,|M y |≤M max );

[0157] Dynamic continuity verification: Calculate the force change rate ΔF / Δt between adjacent sampling points. If it exceeds a preset threshold (preferred, set according to the application scenario), trigger an anomaly flag.

[0158] Environmental consistency verification: Verify environmental parameter E (k) The degree of matching with the statistical distribution of historical data (e.g., using Mahalanobis distance to detect abnormal environmental conditions).

[0159] Formatted output of anti-interference results. The integrated and validated data is output through a standardized interface, specifically including the following:

[0160] Mechanical data: Six-dimensional force / torque vector F in global coordinate system global The data type is a floating-point array;

[0161] Environmental data: temperature, vibration frequency, triaxial electromagnetic field strength parameter E, data type is structured JSON object;

[0162] Metadata: timestamp t k Sensor ID, data verification status flag (normal / warning / abnormal).

[0163] Please see Figure 2 The present invention also provides an anti-interference multidimensional force sensing measurement system, the system comprising the following modules:

[0164] The heterogeneous sensor array module employs a multi-sensor collaborative layout strategy, including fiber optic grating (FBG) optical sensors, piezoresistive MEMS sensors, and temperature-vibration-electromagnetic composite environmental sensors. Its spatial arrangement is based on the stress field gradient direction generated by finite element simulation. By analyzing the mechanical distribution characteristics of the target measurement area, optical sensors are deployed in areas sensitive to high static stress, MEMS sensors are arranged at locations requiring dynamic response, and environmental sensors are uniformly distributed throughout the entire area. This layout strategy maximizes the sensor's efficiency in capturing target signals while reducing coupling interference from environmental noise.

[0165] The signal preprocessing module, addressing the heterogeneous signal characteristics of optical and MEMS sensors, separates the original signal into different frequency bands through cross-band wavelet packet decomposition, extracting the energy distribution characteristics of each band. Based on energy proportions, it dynamically allocates fusion weights, prioritizing the retention of low-frequency, high-precision components from optical sensors and high-frequency dynamic response components from MEMS sensors. An adaptive notch filter tracks the dominant frequency of ambient noise to achieve dynamic suppression. The fusion process employs a hybrid strategy combining frequency domain superposition and time-domain smoothing to ensure signal integrity.

[0166] The dynamic decoupling module employs a cascaded hybrid algorithm architecture. An improved Kalman filter first performs linear state estimation on the fused signal, eliminating cross-interference between sensors. A lightweight convolutional neural network (CNN) then performs nonlinear coupling modeling on the residual signal, extracting deep time-frequency features through multiple convolutional kernels and mapping them to a six-dimensional force / torque space. The output of the Kalman filter serves as the prior input to the CNN, forming a complementary enhancement mechanism of linear and nonlinear decoupling.

[0167] The environment adaptive module constructs a dynamic weight allocation model based on a fuzzy logic controller. It calculates the confidence weights of optical and MEMS sensors using multi-parameter membership functions of temperature, vibration frequency, and electromagnetic field strength. The weight allocation results are fed back to the frequency band energy matching process of the signal preprocessing module in real time, and the sampling rate is adjusted proportionally according to the MEMS weight value: the sampling rate is reduced when the weight is low to reduce resource consumption, and the sampling rate is increased when the weight is high to enhance the dynamic signal acquisition capability.

[0168] The redundancy verification module verifies the physical validity of the decoupling results using the Newton-Euler static equilibrium equations. For anomalous components that do not meet the resultant force equilibrium conditions, a generative adversarial network (GAN) is used for data-driven reconstruction. The generator takes the anomalous components and environmental parameters as input, learns the normal mechanical distribution characteristics through adversarial training, and outputs a reconstructed signal that conforms to physical laws. The reconstruction results are fed back to the dynamic decoupling module in a closed loop, triggering iterative corrections until the verification conditions are met.

[0169] The data integration module performs spatiotemporal alignment and standardized encapsulation of multi-source data. It compensates for sampling time deviations through hardware synchronization signals and uses a quaternion rotation matrix to unify mechanical data from the local coordinate system to the global coordinate system. Before data output, three levels of verification are performed: range exceedance verification, dynamic continuity verification, and environmental-mechanical correlation verification. Abnormal data is accompanied by status indicators. The final result is output in a structured data stream format, compatible with JSON, CSV, and industrial bus protocols, supporting both real-time control and offline analysis modes.

[0170] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional force sensing measurement method with anti-interference capability, characterized in that, Includes the following steps: S1. Obtain the stress field distribution of the object under test through finite element simulation, and design a heterogeneous sensor array layout based on the gradient direction of the stress field distribution. The heterogeneous sensor array includes an optical sensor, a piezoresistive MEMS sensor and an environmental sensor module, and simultaneously collects force signals and environmental parameters. S2. Perform cross-band wavelet packet decomposition on the force signal, allocate weights according to the frequency band energy ratio to generate a fused signal, and dynamically adjust the adaptive notch filter based on environmental parameters to suppress the noise main frequency. S3. Input the fused signal into a hybrid decoupling algorithm consisting of an improved Kalman filter and a lightweight convolutional neural network to separate the multidimensional force components; S4. Through a fuzzy logic controller, the data weights of the optical sensor and the MEMS sensor are dynamically adjusted according to the environmental parameters, and the weight allocation results are fed back to the frequency band energy matching and fusion process. At the same time, the dynamic sampling rate of the MEMS sensor is adjusted according to the weights. S5. Perform rigid body mechanical equilibrium verification on the multidimensional force components. If an anomaly is detected, generate an adversarial network reconstruction signal and feed the reconstruction result back to the hybrid decoupling algorithm for iterative correction. S6. Integrate and verify the multidimensional force component data, and output the measurement results after anti-interference processing.

2. The anti-interference multidimensional force sensing measurement method according to claim 1, characterized in that, Step S1 includes: obtaining the stress field distribution of the object under test through finite element simulation, calculating the stress field gradient direction at each candidate location based on the simulation results; determining the layout positions of the optical sensor and the piezoresistive MEMS sensor based on the matching condition between the gradient direction and the sensor sensitive axis direction, wherein the matching condition is that the cosine value of the angle between the gradient vector and the sensor sensitive axis direction must reach a preset strain sensitivity threshold; the environmental sensor module is integrated into a heterogeneous sensor array and is used to synchronously collect temperature, vibration frequency and electromagnetic field strength parameters.

3. The anti-interference multidimensional force sensing measurement method according to claim 1, characterized in that, Step S2 includes: performing cross-band wavelet packet decomposition on the force signals of the optical sensor and the piezoresistive MEMS sensor respectively, using Daubechies wavelet basis functions for multi-level decomposition, and extracting the signal energy values ​​in each frequency band. Based on the characteristics that optical sensors have a high energy ratio in the low-frequency band and MEMS sensors have a high energy ratio in the high-frequency band, the frequency band fusion weight coefficient is dynamically allocated, and the signals of the two sensors are weighted and superimposed according to the frequency band energy ratio to generate an anti-interference fusion signal. In the fusion process, the low-frequency band is mainly weighted by optical sensor signals, the high-frequency band is mainly weighted by MEMS sensor signals, and the frequency bands are processed by time-domain smooth transition.

4. The anti-interference multidimensional force sensing measurement method according to claim 3, characterized in that, Step S2 further includes: applying an adaptive notch filter to the fused signal to construct a band-stop filter with an adjustable center frequency, the center frequency of which is dynamically updated according to the real-time detected ambient noise main frequency; the stopband width of the band-stop filter is controlled by a preset quality factor to selectively suppress the noise main frequency and its harmonic components, and the filtered output signal enters the subsequent decoupling processing flow.

5. The anti-interference multidimensional force sensing measurement method according to claim 1, characterized in that, Step S3 includes: inputting the fused signal into an improved Kalman filter algorithm, linearly predicting the multidimensional force state vector of the previous moment through the state transition matrix, and mapping the state vector to the current observation value in combination with the observation matrix, wherein the state prediction process introduces the process noise covariance matrix, and the observation update process introduces the observation noise covariance matrix. The multidimensional force state estimation result output by the improved Kalman filter is concatenated with the time-frequency features of the fused signal and then input into a lightweight convolutional neural network. Through the nonlinear feature extraction and mapping of the network, the decoupled six-dimensional force / torque components are output.

6. The anti-interference multidimensional force sensing measurement method according to claim 5, characterized in that, The input layer of the lightweight convolutional neural network is the result of concatenating the state estimation vector output by the improved Kalman filter with the fused signal, and the output layer is mapped to a six-dimensional force / torque vector through a fully connected layer. The network is trained with the squared Euclidean distance between the predicted value and the true value as the loss function, and the network parameters are optimized based on the gradient descent method.

7. The anti-interference multidimensional force sensing measurement method according to claim 1, characterized in that, Step S4 includes: receiving real-time temperature, vibration frequency and electromagnetic field strength parameters through a fuzzy logic controller, defining the weight allocation relationship between the optical sensor and the piezoresistive MEMS sensor, wherein the sum of the weights of the two is 1, and the weight of the optical sensor is jointly determined by the temperature membership function, the vibration frequency membership function and the correction factor of the electromagnetic field strength on the stability of the optical signal. The weighting results are fed back to the frequency band energy matching and fusion process to dynamically adjust the frequency band weight coefficients of the optical and MEMS sensors. At the same time, based on the proportion of the MEMS sensor weight values ​​within the preset minimum and maximum ranges, the dynamic sampling rate is adjusted according to the linear interpolation rule so that the sampling rate increases with the increase of weight.

8. The anti-interference multidimensional force sensing measurement method according to claim 7, characterized in that, The temperature membership function is a piecewise function that exhibits a trapezoidal distribution within the preset effective temperature range. When the temperature is below the lower limit or above the upper limit, the membership degree returns to zero. It maintains its maximum value within the optimal temperature range and linearly decays within the transition range. The vibration frequency membership function is a Gaussian distribution function centered on the nominal frequency. Its membership decreases exponentially as the vibration frequency deviates from the nominal value, and the decay rate is controlled by the variance parameter.

9. The anti-interference multidimensional force sensing measurement method according to claim 1, characterized in that, Step S5 includes: performing rigid body mechanical equilibrium verification on the decoupled multidimensional force components to verify whether the resultant force and resultant moment in each direction are in static equilibrium; if the absolute deviation between the measured value and the theoretical equilibrium value of any force or moment component exceeds the corresponding preset threshold, it is determined to be an abnormal signal. Anomaly components are reconstructed using a generative adversarial network. The generator takes anomaly force / torque components and a real-time environmental parameter vector as input and outputs a reconstructed signal that conforms to the laws of mechanics. The reconstructed signal is fed back to a hybrid decoupling algorithm, which replaces the original anomaly components and re-executes the decoupling and verification process until all components pass the balance check.

10. An anti-interference multidimensional force sensing and measurement system, applied to the method described in any one of claims 1-9, characterized in that, The system includes the following modules: The heterogeneous sensor array module is used to acquire force signals and environmental parameters in real time. It includes optical sensors, piezoresistive MEMS sensors and environmental sensor modules. Its layout is based on the stress field gradient direction generated by finite element simulation. The signal preprocessing module is used to perform cross-frequency wavelet packet decomposition, frequency band energy matching fusion, and adaptive notch filtering on the force signal to generate an anti-interference fused signal. The dynamic decoupling module is used to separate multidimensional force components from the fused signal through a collaborative algorithm of improved Kalman filtering and lightweight convolutional neural network; An environment adaptive module is used to dynamically adjust the data weight allocation between the optical sensor and the MEMS sensor according to environmental parameters, and feed the weight allocation result back to the frequency band energy matching and fusion process of the signal preprocessing module, while adjusting the dynamic sampling rate of the MEMS sensor. The redundancy verification module is used to verify the physical rationality of multidimensional force components through rigid body mechanical equilibrium equations, and to reconstruct abnormal signals using a generative adversarial network, feeding the reconstruction results back to the dynamic decoupling module for iterative correction. The data integration module is used to integrate the verified multidimensional force component data and output the final measurement results after anti-interference processing.

Images