Six-dimensional sensor multi-source data fusion system based on dynamic weight

Abstract

The invention discloses a six-dimensional sensor multi-source data fusion system based on dynamic weight, relates to the technical field of sensor data processing, and is used for improving fusion precision and robustness of measured data of a multi-source sensor. Firstly, a working condition modal identification module analyzes time-frequency characteristics output by a six-dimensional sensor, matches the time-frequency characteristics with a predefined working condition template, and identifies a current dominant working condition modal; then, a credible domain dynamic mapping module counts residual mean values and variances of single-point output and global fusion estimation values in a sliding time window, and adaptively adjusts a standard deviation multiple in combination with an instantaneous signal-to-noise ratio to form a dynamic credible domain boundary; the closed-loop weight optimization module distributes weights according to the relative positions of the data points and the credible domain boundary, punishes the abnormal measurement data points and generates a dynamic weight matrix; and finally, the fusion execution module performs federated filtering fusion by using the dynamic weight matrix, outputs global optimal estimation feedback to update residual statistics, and realizes multi-source data fusion closed-loop optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sensor data processing, in particular to a six-dimensional sensor multi-source data fusion system based on dynamic weights. BACKGROUND

[0002] With the wide application of industrial equipment, robots and intelligent mechanical systems, multi-dimensional sensor arrays are widely used to monitor key physical quantities such as force, torque, vibration and pose. The data provided by multi-source sensors has large amount and high dimension, and the response characteristics under different working conditions are obviously different. How to efficiently fuse these data to realize reliable system state perception and accurate control becomes a key problem.

[0003] The existing multi-source sensor fusion method mainly relies on fixed weight or simple statistical weighting strategy. These methods have several problems in practical application: on the one hand, the sensor output is greatly affected by the change of working condition and noise interference, and the fixed weight is difficult to reflect the real-time reliability of data; on the other hand, abnormal data and transient interference are easy to introduce system estimation deviation, which leads to unstable global fusion result or response lag. In addition, the existing method is insufficient in dealing with the correlation between different sensor dimensions, and it is difficult to fully utilize the complementarity of multi-dimensional information, thereby affecting the fusion accuracy and the robustness of the system. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a six-dimensional sensor multi-source data fusion system based on dynamic weights, which solves the problems in the above background art.

[0005] To achieve the above object, the application is implemented by the following technical solutions: a six-dimensional sensor multi-source data fusion system based on dynamic weight, comprising the following modules: a working condition mode identification module, used for analyzing the time-frequency joint features of the six-dimensional sensor output and matching with the predefined working condition templates to identify the current dominant working condition mode; a dynamic mapping module of the trust domain, used for, within a sliding time window of a set length, counting the residual mean and variance between the single-point output of the six-dimensional sensor under the identified dominant working condition mode and the global fusion estimated value, and combining the instantaneous signal-to-noise ratio to adaptively adjust the standard deviation multiple to form the dynamic trust domain boundary of each sensor and each measurement dimension; a closed-loop weight optimization module, used for implementing weight distribution according to the relative position relationship between the real-time measurement data points of the sensor and the dynamic trust domain boundary, calculating the basic weight of the measurement data points within the trust domain range based on the deviation degree of the measurement data points from the residual distribution center, and applying weight penalty to the abnormal measurement data points beyond the trust domain boundary according to the exceeding amplitude, and organizing the weight calculation results of each measurement data point into a dynamic weight matrix output; a fusion execution module, used for taking the dynamic weight matrix as the information distribution factor in the federated filtering structure and the adjustment basis of the measurement noise covariance matrix, executing the global optimal fusion estimation of the multi-source data, and feeding back the generated global fusion estimated value to the dynamic mapping module of the trust domain for updating the residual statistics within the sliding time window.

[0006] Further, the working condition mode identification module comprises the following steps: extracting the time domain features and frequency domain features of each six-dimensional sensor output signal, the time domain features including signal variance and peak factor, and the frequency domain features including main frequency band energy distribution; combining the extracted time-frequency features into a multi-dimensional feature vector, and performing similarity matching with the predefined feature templates of each working mode; calculating the matching degree between the current feature vector and each working condition template, and selecting the working mode corresponding to the working condition template with the highest matching degree as the current dominant working condition mode; when the highest similarity is lower than a set threshold, performing weighted fusion on the current feature vector and the working condition mode features confirmed at the last time, recalculating the similarity based on the fusion result and determining the dominant working condition mode.

[0007] Further, within a sliding time window of a set length, the specific process of counting the residual mean and variance between the single-point output of the six-dimensional sensor under the identified dominant working condition mode and the global fusion estimated value is as follows: within the sliding time window, collecting the single-point output of each sensor belonging to the current dominant working condition mode and the global fusion estimated value at the corresponding time; calculating the instantaneous residual error of each sensor single-point output and the global fusion estimated value at each sampling time; updating the residual mean within the sliding window through a recursive algorithm, and the new mean estimate is obtained by weighting the mean estimate at the previous time and the new residual data; based on the updated residual mean, the residual variance is updated through a recursive algorithm, and the new variance estimate is obtained by weighting the deviation degree of the new residual from the mean of the previous time variance estimate.

[0008] Further, the specific process of forming the dynamic confidence domain boundary of each sensor and each measurement dimension in combination with the instantaneous signal-to-noise ratio adaptive adjustment of the standard deviation multiple is as follows: according to the residual mean and variance statistics of the residual statistical process, the instantaneous signal-to-noise ratio of each sensor and each measurement dimension is calculated, the mapping relationship between the signal-to-noise ratio and the standard deviation multiple is established, and the lower the signal-to-noise ratio, the larger the standard deviation multiple; the residual mean obtained by statistics in the sliding window is multiplied by the adjusted standard deviation multiple, and the product is used as the boundary offset; the symmetric confidence domain boundary is constructed with the residual mean as the center and the boundary offset as the radius; when the calculated confidence domain boundary exceeds the physical range of the sensor, the boundary is limited within the range of the sensor range.

[0009] Further, the specific process of calculating the basic weight of the measurement data point within the confidence domain range based on the deviation degree of the measurement data point from the residual distribution center is as follows: according to the residual distribution center position provided by the confidence domain dynamic mapping module, the deviation degree of each measurement data point from the distribution center is calculated; according to the pre-set inverse proportional function relationship, the deviation degree is mapped to the basic weight value, wherein the deviation degree and the weight value are inversely proportional; the basic weights of each sensor in the same dimension are normalized by the Softmax function, so that the total weight sum is kept as 1; according to the current working condition modal type, the pre-defined working condition weight coefficient table is queried, and the normalized basic weight is weighted and corrected.

[0010] Further, the specific process of calculating the basic weight of the measurement data point within the confidence domain range based on the deviation degree of the measurement data point from the residual distribution center is as follows: according to the residual distribution center position provided by the confidence domain dynamic mapping module, the deviation degree of each measurement data point from the distribution center is calculated; according to the pre-set inverse proportional function relationship, the deviation degree is mapped to the basic weight value, wherein the deviation degree and the weight value are inversely proportional; the basic weights of each sensor in the same dimension are normalized by the Softmax function, so that the total weight sum is kept as 1; according to the current working condition modal type, the pre-defined working condition weight coefficient table is queried, and the normalized basic weight is weighted and corrected.

[0011] Further, the specific process of performing the global optimal fusion estimation of multi-source data by taking the dynamic weight matrix as the adjustment basis of the information distribution factor and the measurement noise covariance matrix in the federated filtering structure is as follows: the dynamic weight matrix is subjected to eigenvalue decomposition to extract the eigenvalue and eigenvector thereof, the information distribution factor of each sensor in the federated filtering is determined according to the size of the trace of the weight submatrix, wherein the information distribution factor is proportional to the trace of the weight submatrix; the measurement noise covariance matrix is reconstructed through the inverse matrix of the weight submatrix to establish an inverse proportional relationship between the weight value and the noise level; the fusion of the local estimation in the master filter is performed, and the local state estimation and the covariance matrix of each sensor are synthesized through the weighted fusion algorithm; the global state estimation is completed through the Kalman filter update equation, wherein the process noise covariance matrix is dynamically adjusted according to the current working condition mode; and the global optimal fusion estimation result of the six-dimensional force / torque, including the optimal estimation values of the three-dimensional force components and the three-dimensional torque components, is output.

[0012] Further, the specific process of feeding the generated global fusion estimation value back to the trusted domain dynamic mapping module and updating the residual statistics in the sliding time window is as follows: the global fusion estimation value at the current time is taken as a reference to compare and calculate with the single-point output of each sensor; the time stamp corresponding relationship between the global fusion estimation value and the output of each sensor is established; in the sliding time window, the single-point output of each sensor is subtracted from the global fusion estimation value at the corresponding time to obtain an instantaneous residual sequence; the residual mean value is corrected through a recursive update algorithm, and the new mean value estimation is calculated by the historical mean value and the new residual through a preset weight coefficient; based on the updated residual mean value, the residual variance is recalculated through a recursive algorithm; according to the duration of the current working condition mode, the weight coefficient of the historical data in the sliding window is adjusted, the updated residual statistics are range tested, and the statistical values exceeding the preset range are excluded.

[0013] The present application has the following beneficial effects: (1) The six-dimensional sensor multi-source data fusion system based on dynamic weight can distinguish the dominant working condition mode of the system in real time through the working condition mode identification module, so that the sensor output data can be analyzed in different working conditions, the working condition adaptability and feature recognition accuracy are improved, the dynamic trusted domain boundary can be adjusted adaptively in combination with the residual statistics and the instantaneous signal-to-noise ratio in the sliding time window through the trusted domain dynamic mapping module, noise and abnormal data are effectively filtered, the data reliability of each measurement dimension of each sensor is quantified, and the reliability of the fusion basic data is improved.

[0014] (2) The six-dimensional sensor multi-source data fusion system based on dynamic weight, based on the closed-loop weight optimization module, the system can implement weight distribution on the real-time measurement data of the sensor according to the dynamic trust domain boundary, and the abnormal measurement data is punished, so as to reduce the influence of abnormal data on the fusion result, improve the stability and accuracy of global estimation; through the fusion execution module, the dynamic weight matrix is used as the information distribution factor and the measurement noise covariance adjustment basis, the global optimal fusion estimation of multi-source data is realized, and the fusion result is fed back to update the residual statistics, the closed-loop optimization is realized, and the fusion precision and real-time response ability of the system are significantly enhanced.

[0015] Of course, implementing any product of the application does not necessarily require achieving all the advantages described above at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flow chart of the six-dimensional sensor multi-source data fusion system based on dynamic weight of the application. DETAILED DESCRIPTION

[0017] The embodiment of the application solves the problems of low fusion precision, obvious abnormal data interference and insufficient real-time performance of the existing multi-source six-dimensional sensor data under different working conditions through the six-dimensional sensor multi-source data fusion system based on dynamic weight.

[0018] The scheme in the embodiment of the application has the following general idea: The time-frequency joint characteristics of the six-dimensional sensor are analyzed in real time through the working condition mode identification module to determine the dominant working condition mode of the system at present; In the trust domain dynamic mapping module, the residual statistics and the instantaneous signal-to-noise ratio in the sliding time window are combined to adaptively generate the dynamic trust domain boundary of each sensor in each measurement dimension, and the data reliability is quantified; The closed-loop weight optimization module is used to distribute the weight of the real-time measurement data of the sensor according to the dynamic trust domain boundary, to punish the abnormal data, and to realize the dynamic regulation and control of the data reliability; Through the fusion execution module, the dynamic weight matrix is used as the federal filtering information distribution factor and the adjustment basis of the measurement noise covariance matrix, the global optimal fusion estimation of multi-source data is completed, and the fusion result is fed back to update the residual statistics, the closed-loop optimization and real-time accurate multi-source fusion are realized.

[0019] Please refer to Figure 1The embodiment of the application provides a technical scheme: a six-dimensional sensor multi-source data fusion system based on dynamic weight, comprising the following modules: a working condition mode identification module, which is used for analyzing time-frequency joint features output by the six-dimensional sensor, and matching with a predefined working condition template to identify a current dominant working condition mode; a reliable domain dynamic mapping module, which is used for, in a sliding time window with a set length, counting residual mean and variance between single-point output of the six-dimensional sensor under the identified dominant working condition mode and a global fusion estimated value, and combining an instantaneous signal-to-noise ratio to adaptively adjust a standard deviation multiple, to form a dynamic reliable domain boundary of each measurement dimension of each sensor; a closed-loop weight optimization module, which is used for implementing weight distribution according to a relative position relationship between a sensor real-time measurement data point and the dynamic reliable domain boundary, calculating a basic weight based on a deviation degree of the measurement data point from a residual distribution center for the measurement data point located in the reliable domain range, applying a weight penalty according to an exceeding amplitude for an abnormal measurement data point exceeding the reliable domain boundary, and organizing a weight calculation result of each measurement data point into a dynamic weight matrix output; and a fusion execution module, which is used for taking the dynamic weight matrix as an information distribution factor in a federal filtering structure and an adjustment basis of a measurement noise covariance matrix, executing global optimal fusion estimation of multi-source data, and feeding back a generated global fusion estimated value to the reliable domain dynamic mapping module to update residual statistics in the sliding time window.

[0020] In this embodiment, the operating mode identification module: this module is used to analyze the time domain and frequency domain characteristics of the output signal of the six-dimensional sensor, and match the analyzed characteristics with the predefined operating mode template, so as to identify the dominant working state (operating mode) of the current system. Six-dimensional sensor: refers to a sensor that can simultaneously measure three-dimensional force (X, Y, Z direction) and three-dimensional torque (torque generated by rotation around X, Y, Z axis). Time domain feature: refers to the feature directly extracted from the time series signal, such as signal mean, variance, peak factor, etc. Frequency domain feature: refers to the feature obtained by Fourier transform and other frequency domain analysis methods on the time series signal, such as main frequency band energy distribution. Operating mode template: refers to the characteristic model predefined according to different operating states or working modes, used for pattern matching and operating mode identification. Dynamic mapping module of confidence domain: within a sliding time window of a certain length, the mean and variance of the residual between the single-point output of the six-dimensional sensor and the global fusion estimated value are calculated, and the standard deviation multiple is adjusted adaptively combined with the instantaneous signal-to-noise ratio, to generate the dynamic confidence domain boundary of each sensor in each measurement dimension. The boundary is used to quantify the reliability of each measurement value. Sliding time window: refers to a continuous data segment of fixed length in the time series, used for calculating statistics and updating in real time. Residual: refers to the difference between the single sensor measurement value and the global fusion estimated value, used to measure the measurement deviation. Instantaneous signal-to-noise ratio (SNR): the ratio of signal strength to noise strength, reflecting the reliability of the signal. Dynamic confidence domain boundary: the limit generated adaptively according to residual statistics and signal-to-noise ratio, used to determine whether the measurement value is reliable. Closed-loop weight optimization module: according to the relative position relationship between the real-time measurement data of the sensor and the dynamic confidence domain boundary, a weight is assigned to each measurement data. For data within the confidence domain range, the basic weight is calculated according to the deviation from the residual center; for abnormal data beyond the confidence domain boundary, a penalty weight is applied according to the exceeding amplitude. The weight matrix output by the module is used for subsequent fusion calculation. Weight allocation: according to the data reliability, different data processing priorities or contribution levels are given. Dynamic weight matrix: a matrix form weight set, used to represent the contribution size of each sensor in each measurement dimension in the fusion calculation. Closed loop: refers to the weight allocation result will be fed back to the data processing flow, realizing continuous optimization and adjustment. Fusion execution module: receives the dynamic weight matrix, and uses it as the information distribution factor and adjustment basis of the measurement noise covariance matrix in the federated filtering structure, to execute the global optimal fusion estimation of multi-source data. The fusion result is fed back to the dynamic mapping module of the confidence domain, to update the residual statistics in the sliding time window, realizing closed-loop optimization. Federated filtering (Federated Filtering): a multi-sensor fusion algorithm, which integrates the state and covariance of each local estimate by weighting, to realize the optimal estimation of the global state. Measurement noise covariance matrix: a matrix reflecting the measurement uncertainty and mutual correlation of each sensor, used to adjust the estimation weight in the filtering calculation.Global optimal fusion estimation: the optimal estimation result of the overall system state is obtained on the basis of considering the data quality and weight of each sensor.

[0021] Specifically, the working condition mode identification module includes the following steps: extracting the time domain features and frequency domain features of the output signals of each six-dimensional sensor, the time domain features including signal variance and peak factor, and the frequency domain features including main frequency band energy distribution; combining the extracted time-frequency features into a multi-dimensional feature vector, and performing similarity matching with the feature templates of each working mode; calculating the matching degree between the current feature vector and each working mode template, and selecting the working mode corresponding to the working mode template with the highest matching degree as the current dominant working condition mode; when the highest similarity is lower than a set threshold, weighting and fusing the current feature vector and the working condition mode feature confirmed at the last time, recalculating the similarity based on the fusion result and determining the dominant working condition mode.

[0022] In the embodiment, the time domain features and the frequency domain features are extracted, and each six-dimensional sensor output signal Skt (k represents the sensor number) is analyzed. The time domain features: signal variance Vm and peak factor Pn Vm = 1 L p = 1 L S kp - S k 2 ; wherein: Skp: the sensor output value of the pth sampling point; Sk: the average value of the kth sensor in the sliding window; L: the total number of sampling points in the sliding window. Peak factor: Pn = max Skq Vm ; wherein: max Skq: the maximum absolute value of the output signal in the sliding window; Vm: the corresponding signal variance; q: the sampling point index. The frequency domain features: main frequency band energy distribution Fr, Fr = f e Br S kf 2 ; wherein: Skf: the Fourier transform result of the signal Skt; Br: the frequency range of the rth main frequency band; r: the frequency band number. The extracted time domain and frequency domain features are combined into a multi-dimensional feature vector Xs = V1,...,Vm, P1,...,Pn, F1,...,Fr. The Xs is matched with each predefined working mode template Tv in terms of similarity, and the matching degree Mv is calculated: Mv = Xs · Tv || Xs || || Tv || ; wherein: Tv: the vth working mode template vector; ·: the inner product operation of the vector; || Xs ||: the norm of the feature vector; || Tv ||: the norm of the template vector; v: the template number. The dominant working condition mode is selected, and the working condition mode corresponding to the template with the highest matching degree is selected as the current dominant working condition mode: Cmain = arg max v Mv ; wherein: Cmain: the current dominant working condition mode number; Mv: the matching degree of each template. When the highest matching degree Mmax = max v Mv is lower than the set threshold value Θ, the current feature vector Xs and the working condition mode feature Xprev confirmed at the last time are fused: Xfused = a Xs + 1 - a Xprev ; wherein: Xfused: the fused feature vector; Xprev: the working condition mode feature vector confirmed at the last time; a: the weighting coefficient, which can be determined according to the matching degrees in the last N time steps, for example: a = Mmax t = 1 N Mt ; N: the length of the history window; Mt: the highest matching degree at each time in the history. The similarity between the fused vector Xfused and each template is recalculated, and the template with the highest matching degree is selected as the finally determined dominant working condition mode.

[0023] Specifically, the specific process of calculating the residual mean and variance between the single-point output of the six-dimensional sensor under the identified dominant working condition mode and the global fusion estimated value in the sliding time window of a certain length is as follows: in the sliding time window, the single-point output of each sensor belonging to the current dominant working condition mode and the global fusion estimated value at the corresponding time are collected; the instantaneous residual between the single-point output of each sensor and the global fusion estimated value at each sampling time is calculated; the residual mean in the sliding window is updated by a recursive algorithm, and the new mean estimate is obtained by weighting the mean estimate at the previous time and the new residual data; based on the updated residual mean, the residual variance is updated by a recursive algorithm, and the new variance estimate is obtained by weighting the variance estimate at the previous time and the deviation of the new residual from the mean.

[0024] In the embodiment, the data in the sliding window is collected, and in a sliding time window with a length of W, the single-point output Yrt of each sensor and the global fusion estimation value Zst corresponding to the time point belonging to the current dominant operating mode Cd are collected; wherein: Yrt is the single-point output of the rth sensor at time t; Zst is the global estimation value of the st fusion estimation value corresponding to time t; r is the sensor number; t is the time index; W is the sliding window length; Cd is the current identified dominant operating mode. The instantaneous residual error is calculated, and for each sampling time, the instantaneous residual error Ar,t of the single-point output of the sensor and the global fusion estimation value is calculated: Ar,t=Yrt-Zst; wherein: Ar,t is the instantaneous residual error of the rth sensor at time t; Yrt is the single-point output; and Zst is the global fusion estimation value. The sliding window residual error mean is recursively updated, and the residual error mean Ar is updated using a recursive algorithm. The new mean estimation Arnew is obtained by weighting the previous time mean Arold and the new residual error data: Arnew=λArold+1-λAr,t; wherein: Arnew is the updated residual error mean; Arold is the residual error mean at the previous time; Ar,t is the current sampling instantaneous residual error; and λ is a weight coefficient, which can be determined adaptively through the historical residual error variance, for example: λ=σhistσhist+σnew; wherein: σhist is the historical residual error variance in the sliding window; and σnew is the current sampling residual error variance estimation. The sliding window residual error variance is recursively updated, and based on the updated residual error mean Arnew, the residual error variance Σr is updated through a recursive algorithm. The new variance estimation Σrnew is obtained by weighting the deviation of the new residual error from the mean: Σrnew=μΣrold+1-μAr,t-Arnew2; wherein: Σrnew is the updated residual error variance; Σrold is the residual error variance at the previous time; Ar,t is the instantaneous residual error; Arnew is the updated residual error mean; and μ is a weight coefficient, which can be determined according to the sliding window length W and the historical variance stability, for example: μ=W-1W. Through the sliding window recursive calculation, the deviation of each sensor output from the global fusion estimation is realized in real time. The mean and variance recursive update not only ensures the calculation efficiency, but also dynamically adapts to the operating condition changes, providing reliable residual error statistics for subsequent modules (such as dynamic confidence domain mapping and weight optimization).

[0025] Specifically, the specific process of forming the dynamic confidence limit boundary of each sensor in each measurement dimension by adjusting the standard deviation multiple in combination with the instantaneous signal-to-noise ratio is as follows: according to the residual mean and variance statistics of the residual statistical process, the instantaneous signal-to-noise ratio of each sensor in each measurement dimension is calculated, the mapping relationship between the signal-to-noise ratio and the standard deviation multiple is established, and the lower the signal-to-noise ratio, the larger the standard deviation multiple to be used; the residual mean obtained by statistics in the sliding window is multiplied by the adjusted standard deviation multiple, and the product is used as the boundary offset; the symmetric confidence limit boundary is constructed with the residual mean as the center and the boundary offset as the radius; when the calculated confidence limit boundary exceeds the physical range of the sensor, the boundary is limited within the range of the sensor range.

[0026] In this embodiment, the instantaneous signal-to-noise ratio (SNR) is calculated, and the instantaneous signal-to-noise ratio Qr of each sensor in each measurement dimension is calculated according to the residual mean Δr and the residual variance Σr obtained by the foregoing residual statistical process: Qr = ΔrΣr + ϵ; wherein: Qr; instantaneous signal-to-noise ratio of rth sensor; Δr; residual mean in the sliding window (obtained in the previous step); Σr; residual variance in the sliding window (obtained in the previous step); ϵ; a minimum constant used to avoid division by zero, which can be a fraction of the minimum observable value of the residual variance. A mapping relationship between the signal-to-noise ratio and the standard deviation multiple is established, and a functional relationship of the standard deviation multiple Kr changing with the instantaneous signal-to-noise ratio Qr is defined: Kr = fQr = Kmin + Kmax - Kmine - αe⋅Qr; wherein: Kr; standard deviation multiple of rth sensor; Kmin; minimum value of standard deviation multiple, which can be determined empirically by historical residual distribution; Kmax; maximum value of standard deviation multiple, which can be determined by the upper limit of the error tolerance of the sensor; αe; mapping adjustment coefficient, which can be adjusted according to the experience of system sensitivity to noise; Qr; instantaneous signal-to-noise ratio. The lower the signal-to-noise ratio, the greater the exponential decay term, and the greater the Kr obtained, thereby expanding the range of the confidence domain and improving the robustness. The dynamic confidence domain boundary is calculated, and the residual mean Δr obtained by the statistics in the sliding window is multiplied by the standard deviation multiple Kr to obtain the boundary offset ΔBr: ΔBr = Kr⋅Σr; wherein: ΔBr; boundary offset of rth sensor in the measurement dimension; Σr; residual variance; Kr; standard deviation multiple. A symmetric confidence domain boundary is constructed, with the residual mean Δr as the center and the boundary offset ΔBr as the radius, to construct a symmetric confidence domain boundary: Δr - ΔBr, Δr + ΔBr; the boundary represents the range of measurement data considered to be reliable under the current sliding window and instantaneous signal-to-noise ratio. Boundary limitation, when the calculated confidence domain boundary exceeds the physical range Lrmin, Lrmax of the sensor, the boundary is limited within the range: WED = max Δr - ΔBr, Lrmin; WEV = min Δr + ΔBr, Lrmax; wherein: WED; upper boundary, WEV; lower boundary, Lrmin; minimum measurable value of rth sensor; Lrmax; maximum measurable value of rth sensor. Through the foregoing steps, the system can dynamically adjust the confidence domain range of each sensor in each measurement dimension, so that it can adapt to signal noise fluctuations and be constrained within the physical range of the sensor; the dynamic confidence domain boundary provides a basic condition for subsequent closed-loop weight optimization, which helps the rationality and robustness of weight distribution.

[0027] Specifically, for the measurement data points located in the trusted domain range, the specific process of calculating the basic weight based on the deviation degree of the measurement data points from the residual distribution center is as follows: according to the residual distribution center position provided by the trusted domain dynamic mapping module, the deviation degree of each measurement data point from the distribution center is calculated; according to the preset inverse proportional function relationship, the deviation degree is mapped to the basic weight value, wherein the deviation degree and the weight value are in inverse proportional relationship; the basic weights of each sensor in the same dimension are normalized by the Softmax function, so that the total weight sum is kept as 1; according to the current working condition modal type, the pre-defined working condition weight coefficient table is queried, and the normalized basic weight is weighted and corrected.

[0028] In the embodiment, the deviation degree is calculated, and for single-point measurement data msv of each sensor (s represents the sensor number, and v represents the measurement dimension), the deviation degree δsv is calculated according to the residual distribution center Csv provided by the trusted domain dynamic mapping module: δsv=msv-Csv; wherein: msv: the measurement data point currently sampled in the sliding window; Csv: the residual distribution center corresponding to the sensor and the dimension, provided by the trusted domain dynamic mapping module; δsv: the deviation degree of the measurement data point relative to the residual distribution center. The deviation degree δsv is mapped to the basic weight Wsv by using an inverse proportional function relationship: Wsv=βeδsv+ϵ; wherein: Wsv: the basic weight value of the sensor s and the dimension v; βe: a proportional coefficient for adjusting the weight scale, which can be determined by historical residual distribution maximum deviation experience; ϵ: a minimum constant for avoiding zero division, which can take a part of the minimum observable value of the residual; δsv: the deviation degree. The greater the deviation degree, the smaller the basic weight, which realizes natural suppression of abnormal data. Normalization processing is performed on the basic weights of all sensors in the same measurement dimension v by Softmax normalization processing, so that the total weight sum is kept as 1: Wsv=expWsvkexpWkv; wherein: Wsv: the normalized weight; k: the sensor index in the same dimension; Wsv: the basic weight before normalization. Softmax can ensure that all weights are positive and the total sum is 1, while preserving the relative difference of the deviation degree. Working condition weight correction: according to the current working condition modal type, the pre-defined working condition weight coefficient table Gv is queried to weight and correct the normalized basic weight: Wsv,final=Wsv⋅Gv; wherein: Wsv,final; the final weight used for fusion calculation; Gv; the weight coefficient of the measurement dimension v under the current working condition, determined by pre-statistics or experiment; Wsv; the normalized basic weight. Through working condition weighting, the importance difference of each dimension sensor under different working states can be reflected, and the fusion precision is improved.

[0029] Specifically, for the abnormal measurement data points beyond the boundary of the trust region, the weight penalty is applied according to the exceeding amplitude, and the weight calculation results of each measurement data point are organized into a specific process of dynamic weight matrix output as follows: the Euclidean distance between the abnormal measurement data point and the nearest trust region boundary point is calculated as the exceeding amplitude quantization value; according to the preset exponential decay function, the exceeding amplitude is converted into a penalty coefficient, and the penalty coefficient decreases exponentially with the increase of the exceeding amplitude; the basic weight of the abnormal data point is multiplied by the penalty coefficient to obtain the actual weight value after the penalty; the weight values of each data point are organized into a block diagonal weight matrix in the order of sensor number and measurement dimension; the singular value decomposition of the weight matrix is tested to ensure that the condition number of the matrix is within the preset range, and the dynamic weight matrix meeting the requirements of the federated filtering algorithm is output.

[0030] In the embodiment, the exceeding amplitude is calculated, the Euclidean distance between each measurement data point xru (r represents the sensor number, and u represents the measurement dimension) beyond the boundary of the trust region and the nearest trust region boundary point Bru is calculated to obtain the exceeding amplitude Δru: Δru = xru - Bru; wherein: xru: measurement data point beyond the boundary of the trust region; Bru: trust region boundary value corresponding to the sensor and the dimension; Δru: quantized exceeding amplitude. The exceeding amplitude Δru is mapped to the weight penalty coefficient Pru using an exponential decay function: Pru = exp-λ⋅Δru; wherein: Pru: penalty coefficient of abnormal measurement data point; λ: decay coefficient, which can be determined by the maximum value of the residual mean in the sliding window; Δru: exceeding amplitude. The larger the exceeding amplitude, the smaller the penalty coefficient, and the inhibition of abnormal data is realized. The penalty weight is calculated, the basic weight Wru of the abnormal data point is multiplied by the penalty coefficient Pru to obtain the final weight Wr, finalu = Wru⋅Pru; wherein: Wru: basic weight (calculated by the deviation degree within the trust region range); Wr, finalu: actual weight value after the penalty. The dynamic weight matrix is constructed, and the weight values of all measurement points are organized into a block diagonal weight matrix Wdyn in the order of sensor number r and measurement dimension u. The singular value decomposition (SVD) of the weight matrix is tested to ensure that the condition number of the matrix is within the preset range, so as to ensure that the matrix is applicable to the federated filtering algorithm. The final dynamic weight matrix Wdyn meeting the requirements of the federated filtering algorithm is output.

[0031] Specifically, the specific process of performing global optimal fusion estimation of multi-source data is as follows: the dynamic weight matrix is subjected to eigenvalue decomposition to extract eigenvalues and eigenvectors, the information distribution factor of each sensor in the federated filtering is determined according to the size of the trace of the weight sub-matrix, wherein the information distribution factor is proportional to the trace of the weight sub-matrix; the measurement noise covariance matrix is reconstructed through the inverse matrix of the weight sub-matrix to establish an inverse proportional relationship between the weight value and the noise level; the fusion of local estimation in the main filter is performed, and the local state estimation and the covariance matrix of each sensor are synthesized through a weighted fusion algorithm; the global state estimation is completed through the Kalman filter update equation, wherein the process noise covariance matrix is dynamically adjusted according to the current working condition mode; and the global optimal fusion estimation result of the six-dimensional force / torque is output, including the optimal estimation values of the three-dimensional force components and the three-dimensional torque components.

[0032] In this embodiment, dynamic weight matrix eigen decomposition is performed on the dynamic weight matrix Wdyn to obtain an eigenvalue vector Λ and a corresponding eigenvector matrix V: Wdyn=VΛV-1; wherein: Wdyn; dynamic weight matrix, composed of weights of each measurement dimension of each sensor; Λ; eigenvalue vector, each eigenvalue reflecting the importance of the corresponding weight sub-matrix; V; eigenvector matrix; V-1; inverse matrix of the eigenvector matrix. Determine the information allocation factor, calculate the trace of the weight sub-matrix Wr (r represents the sensor number) of each sensor: TrWr=uWruu; wherein Wruu is the diagonal element of the weight sub-matrix, and u represents the measurement dimension. The information allocation factor Ir is proportional to the trace of the weight sub-matrix: Ir=k·TrWr; wherein: Ir; information allocation factor of sensor r in federated filtering; k; proportional coefficient, determined by the maximum trace value of the weight matrix in the sliding window. Reconstruct the measurement noise covariance matrix, and construct the measurement noise covariance matrix Rr using the inverse matrix of the weight sub-matrix: Rr=Wr-1; wherein: Rr; measurement noise covariance matrix of sensor r; Wr-1; inverse matrix of the weight sub-matrix; the greater the weight value, the higher the measurement reliability, and the lower the noise level, achieving an inverse relationship between weight and noise. Local state estimation fusion, in the main filter, the local state estimation xr of each sensor and the corresponding covariance matrix Rr are weighted and fused: xfused=rIrRr-1-1⋅rIrRr-1xr; wherein: xfused: fused global state estimation; xr: local state estimation of sensor r; Ir: information allocation factor; Rr: measurement noise covariance matrix. Kalman filter updates the global state, combines the process noise covariance matrix Q (dynamically adjusted according to the current working condition mode), and completes the global state estimation through the Kalman filter update equation: xk+1=xfused+Kz-Hxfused; wherein: K: Kalman gain; z: measurement vector; H: observation matrix; Q: process noise covariance matrix, dynamically adjusted according to the dominant working condition mode. Output six-dimensional force / torque fusion results, output the global optimal fusion estimation value, including the optimal estimation value of three-dimensional force components and three-dimensional torque components, for subsequent control or analysis.

[0033] Specifically, the generated global fusion estimation value is fed back to the credible domain dynamic mapping module, and the specific process of updating the residual statistics in the sliding time window is as follows: the global fusion estimation value at the current time is taken as a reference, and the single-point output of each sensor is compared and calculated; the time stamp corresponding relationship of the global fusion estimation value and the output of each sensor is established; in the sliding time window, the single-point output of each sensor is subtracted from the global fusion estimation value at the corresponding time to obtain the instantaneous residual sequence; the residual mean value is corrected by the recursive update algorithm, and the new mean value estimation is calculated by the historical mean value and the new residual through a preset weight coefficient; based on the updated residual mean value, the residual variance is recalculated by the recursive algorithm; according to the duration of the current working condition mode, the weight coefficient of the historical data in the sliding window is adjusted, the updated residual statistics is range tested, and the statistical value exceeding the preset range is excluded.

[0034] In the embodiment, the instantaneous residual is calculated, the global fusion estimation value x fused t at the current time is taken as a reference, and the single-point output y r t of each sensor r is compared to obtain the instantaneous residual: e r t = y r t - x fused t ; wherein: e r t ; the instantaneous residual of sensor r at time t; y r t ; the single-point measurement value of sensor r at time t; x fused t ; the global fusion estimation value at time t. The time group corresponding relationship is established, the global fusion estimation value is established in time stamp corresponding relationship with the output of each sensor, and the data in the sliding time window is ensured to be synchronized. The residual mean value is updated recursively, and the residual mean value is updated recursively in the sliding time window: e r t = a ⋅ e r t - 1 + 1 - a ⋅ e r t ; wherein: e r t ; the residual mean value estimation of sensor r at time t; e r t - 1 ; the residual mean value estimation at the last time; e r t ; the instantaneous residual at the current time; a; the recursive weight coefficient, which is determined according to the sliding window length and the reliability of historical data. The weight coefficient is determined: a can be set as L - 1 / L, wherein L is the sliding time window length, and the balance between the historical data and the new residual is ensured. The residual variance is updated recursively, and the residual variance is updated recursively based on the updated residual mean value: s r 2 t = b ⋅ s r 2 t - 1 + 1 - b ⋅ e r t - e r t 2 ; wherein: s r 2 t : the residual variance estimation of sensor r at time t; s r 2 t - 1 : the residual variance estimation at the last time; e r t - e r t : the deviation degree of the current residual relative to the mean value; b: the recursive weight coefficient, which can be similar to a and is determined by the sliding window length and the reliability of historical data. The weight of historical data is adjusted, the weight coefficient of the historical residual data in the sliding window is dynamically adjusted according to the duration of the current dominant working condition mode, the statistics is more sensitive to recent data, and the stability is ensured. Range test and exception exclusion, range test is performed on the updated residual mean value and variance, and the statistical value exceeding the preset reasonable range is excluded to avoid errors caused by abnormal data to the credible domain dynamic mapping.

[0035] In summary, the present application has at least the following effects: The six-dimensional sensor multi-source data fusion system based on dynamic weights can accurately identify the current dominant working condition mode through the working condition mode identification module, realize time-frequency joint feature analysis and template matching of the six-dimensional sensor output signal, and ensure feature adaptive processing under different working conditions in the data fusion process; through the dynamic mapping module of the reliable domain and the sliding window residual statistics, combined with the adaptive adjustment of the standard deviation multiple of the instantaneous signal-to-noise ratio, the dynamic reliable domain boundary is constructed, the measurement data of each sensor in each measurement dimension is evaluated in real time, and the identification ability of abnormal data is improved; through the closed-loop weight optimization module, the basic weight of the data points within the reliable domain range is calculated, the weight penalty is applied to the abnormal measurement data points exceeding the reliable domain boundary, and a dynamic weight matrix is formed, so that the measurement data of each sensor is reasonably weighted, and the fusion robustness is enhanced; through the fusion execution module, the dynamic weight matrix is used for weight distribution and measurement noise covariance matrix adjustment of federated filtering, global optimal fusion estimation of multi-source data is realized, and the global fusion estimation is fed back to the residual statistics module to form a closed-loop updating mechanism, and the fusion precision and system response speed are improved.

[0036] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems or computer program products. Accordingly, the application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) having computer-usable program code embodied thereon.

[0037] The present application is described with reference to flowcharts and / or block diagrams of the systems, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The means for implementing each flow or multiple flows and / or blocks Figure 1 The means for implementing each flow or multiple flows and / or blocks

[0038] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1one or more processes and / or blocks Figure 1 the function specified in the one or more blocks.

[0039] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operations steps are performed on the computer or other programmable data processing devices to generate computer-implemented processes, thus the instructions executed on the computer or other programmable data processing devices provide processes for implementing the flow Figure 1 one or more processes and / or blocks Figure 1 the function specified in the one or more blocks.

[0040] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to encompass within their scope all such variations and modifications as are included within the scope of the invention. What is claimed is:

[0041] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A six-dimensional sensor multi-source data fusion system based on dynamic weight, characterized in that, The application comprises the following modules: A working condition mode identification module is used for analyzing the time-frequency joint features of the six-dimensional sensor output and matching the six-dimensional sensor output with a predefined working condition template to identify a current dominant working condition mode; A dynamic mapping module of a credible domain is used for, within a sliding time window of a set length, statistically calculating the residual mean and variance between the single-point output of the six-dimensional sensor under the identified dominant working condition mode and the global fusion estimated value, and adaptively adjusting the standard deviation multiple in combination with the instantaneous signal-to-noise ratio to form the dynamic credible domain boundary of each measuring dimension of each sensor; A closed-loop weight optimization module is used for implementing weight distribution according to the relative positional relationship between the real-time measuring data points of the sensor and the dynamic credible domain boundary, calculating the basic weight of the measuring data points within the credible domain range based on the deviation degree of the measuring data points from the residual distribution center, and applying weight punishment to the abnormal measuring data points beyond the credible domain boundary according to the exceeding amplitude, and organizing the weight calculation results of each measuring data point into a dynamic weight matrix for output; A fusion execution module is used for taking the dynamic weight matrix as the adjustment basis of the information distribution factor and the measurement noise covariance matrix in the federated filtering structure, executing the global optimal fusion estimation of the multi-source data, and feeding back the generated global fusion estimated value to the dynamic mapping module of the credible domain for updating the residual statistics within the sliding time window.

2. The dynamic weight based six-dimensional sensor multisource data fusion system of claim 1, wherein: The working condition mode identification module comprises the following steps: Time-domain features and frequency-domain features of the output signals of each six-dimensional sensor are extracted, the time-domain features include signal variance and peak factor, and the frequency-domain features include main frequency band energy distribution; The extracted time-frequency features are combined into a multi-dimensional feature vector, and similarity matching is performed between the multi-dimensional feature vector and the feature templates of the predefined working modes; The matching degrees between the current feature vector and each working condition template are calculated, and the working mode corresponding to the working condition template with the highest matching degree is selected as the current dominant working condition mode; When the highest similarity is lower than a set threshold, the current feature vector and the working condition mode features confirmed at the last time are fused, the similarity is recalculated based on the fusion result, and the dominant working condition mode is determined.

3. The dynamic weight based six-dimensional sensor multisource data fusion system of claim 1, wherein: Within a sliding time window of a set length, the specific process of statistically calculating the residual mean and variance between the single-point output of the six-dimensional sensor under the identified dominant working condition mode and the global fusion estimated value is as follows: Within the sliding time window, the single-point output of each sensor belonging to the current dominant working condition mode and the global fusion estimated value at the corresponding time are collected; The instantaneous residual error between the single-point output of each sensor and the global fusion estimated value at each sampling time is calculated; The residual mean within the sliding window is updated through a recursive algorithm, and the new mean estimate is obtained by weighting the mean estimate at the last time and the new residual data; Based on the updated residual mean, the residual variance is updated through a recursive algorithm, and the new variance estimate is obtained by weighting the variance estimate at the last time and the deviation degree of the new residual from the mean.

4. The dynamic weight based six-dimensional sensor multisource data fusion system of claim 3, wherein: The specific process of adaptively adjusting the standard deviation multiple in combination with the instantaneous signal-to-noise ratio to form the dynamic credible domain boundary of each measuring dimension of each sensor is as follows: According to the residual mean and variance statistics of the residual statistical process, the instantaneous signal-to-noise ratio of each sensor in each measurement dimension is calculated, and the mapping relationship between the signal-to-noise ratio and the standard deviation multiple is established, that is, the lower the signal-to-noise ratio, the greater the standard deviation multiple used; The residual mean obtained by statistical analysis in the sliding window is multiplied by the adjusted standard deviation multiple, and the product is used as the boundary offset; Taking the residual mean as the center and the boundary offset as the radius, a symmetric confidence domain boundary is constructed; When the calculated confidence domain boundary exceeds the physical range of the sensor, the boundary is limited within the range of the sensor.

5. The dynamic weight based six-dimensional sensor multisource data fusion system of claim 1, wherein: For the measurement data points within the confidence domain range, the specific process of calculating the basic weight based on the deviation degree of the measurement data points from the residual distribution center is as follows: According to the residual distribution center position provided by the confidence domain dynamic mapping module, the deviation degree of each measurement data point from the distribution center is calculated; According to the preset inverse proportional function relationship, the deviation degree is mapped to the basic weight value, wherein the deviation degree and the weight value are inversely proportional; The basic weights of each sensor in the same dimension are normalized by the Softmax function, so that the total weight sum is kept as 1; According to the current working condition modal type, the pre-defined working condition weight coefficient table is queried, and the normalized basic weight is weighted and corrected.

6. The dynamic weight based six-dimensional sensor multisource data fusion system of claim 1, wherein: For abnormal measurement data points exceeding the confidence domain boundary, the specific process of organizing the weight calculation results of each measurement data point into a dynamic weight matrix output according to the exceeding amplitude weight penalty is as follows: The Euclidean distance between the abnormal measurement data points and the nearest confidence domain boundary point is calculated as the exceeding amplitude quantization value; According to the preset exponential decay function, the exceeding amplitude is converted into a penalty coefficient, and the penalty coefficient decreases exponentially with the increase of the exceeding amplitude; The basic weight of the abnormal data point is multiplied by the penalty coefficient to obtain the actual weight value after the penalty; According to the sensor number sequence and the measurement dimension sequence, the weight values of each data point are organized into a block diagonal weight matrix; The singular value decomposition of the weight matrix is tested to ensure that the condition number of the matrix is within the preset range, and the dynamic weight matrix meeting the requirements of the federated filtering algorithm is output.

7. The dynamic weight based six-dimensional sensor multisource data fusion system of claim 1, wherein: Taking the dynamic weight matrix as the information distribution factor in the federated filtering structure and the adjustment basis of the measurement noise covariance matrix, the specific process of performing global optimal fusion estimation of multi-source data is as follows: The eigenvalues and eigenvectors of the dynamic weight matrix are extracted, and the information distribution factor of each sensor in the federated filtering is determined according to the size of the trace of the weight sub-matrix, wherein the information distribution factor is proportional to the trace of the weight sub-matrix; The measurement noise covariance matrix is reconstructed by the inverse matrix of the weight sub-matrix, and an inverse proportional relationship between the weight value and the noise level is established; The fusion of local estimation in the main filter is performed, and the local state estimation and covariance matrix of each sensor are synthesized by the weighted fusion algorithm; The global state estimation is completed by the Kalman filter update equation, wherein the process noise covariance matrix is dynamically adjusted according to the current working condition modal; The global optimal fusion estimation result of the six-dimensional force / torque is output, including the optimal estimation values of the three-dimensional force components and the three-dimensional torque components.

8. The dynamic weight based six-dimensional sensor multisource data fusion system of claim 7, wherein: The generated global fusion estimation value is fed back to the trusted domain dynamic mapping module, and the specific process of updating the residual statistics in the sliding time window is as follows: The correspondence between the global fusion estimation value and the time stamp of each sensor output is established. In the sliding time window, the single-point output of each sensor is subtracted from the global fusion estimation value at the corresponding time to obtain an instantaneous residual sequence. The residual mean is corrected by a recursive update algorithm. The new mean estimate is calculated by the historical mean and the new residual through a preset weight coefficient. Based on the updated residual mean, the residual variance is recalculated by a recursive algorithm. According to the duration of the current working condition mode, the weight coefficient of the historical data in the sliding window is adjusted, the updated residual statistics are range tested, and the statistical values exceeding the preset range are excluded.

Images