ONLINE FAST PROCESSING METHOD FOR REAL-TIME DATA BASED ON EDGE COMPUTING

Abstract

The present invention belongs to the technical field of signal processing, and relates to an online fast processing method for real-time data based on edge computing. In the present invention, a dynamic online de-noising method is adopted to remove noise contained in speeds to ensure the effectiveness and accuracy of de-noising results; for the displacement integrated online, an efficient method is adopted for dynamic online de-noising to further reduce the effectiveness of drift in the displacement value on final integration results; and under the condition of ensuring the accuracy of an integration method, an integration algorithm is embedded into an edge device to realize fast calculation and analysis of data near a data source and realize dynamic fast integration of online signals based on edge computing, which provides effective references for efficient processing and calculation of data.

Claims

1. An online fast processing method for real-time data based on edge computing, comprising the following specific steps: step (1): designing a dynamic integration algorithm assuming that the acquired original acceleration data is a=(a.sub.1.sup.i, a.sub.2.sup.i, . . . , a.sub.n.sup.i)).sup.T, wherein i represents the i.sup.th monitoring time point, and n represents the number of monitoring points; assuming that the interval between two monitoring time points is Δt, the sampling frequency of the corresponding sensor is 1/Δt, and the speed calculation of the corresponding monitoring point is expressed as follows:
v.sub.j.sup.i=(a.sub.j.sup.i-1+a.sub.j.sup.i)×Δt/2  (1) wherein a.sub.j.sup.i-1 represents the acceleration corresponding to the j.sup.th monitoring point at the (i−1).sup.th time point, and a represents the acceleration value corresponding to the i.sup.th monitoring point at the i.sup.th time point, wherein j∈[1, n], and v.sub.j.sup.i represents the integrated speed corresponding to the j.sup.th monitoring point at the i.sup.th time point; after integration, judging whether the number of accelerations currently collected at the same monitoring point reaches ¼ of the sampling frequency, i.e., whether the number of acceleration sample points collected at the same monitoring point is more than ¼ of the number of sampling frequencies; if the number of accelerations currently collected is less than ¼ of the sampling frequency value, it is considered that the speed obtained by the current integration is less affected by noise and drift, and the integrated speed is regarded as an ideal speed and is not de-noised; then, directly integrating the speed obtained by the current integration again to acquire the corresponding displacement, and the displacement currently obtained is the final ideal displacement value; if the number of collected accelerations is more than ¼ of the sampling frequency value, it is considered that the speed currently integrated is greatly affected by the drift; dynamically de-noising the integrated speed to reduce the influence of the drift on the integrating speed; and the whole de-noising process is expressed as follows:
Z.sub.j.sup.i=α.Math.Z.sub.j.sup.i-1+(1−α).Math.v.sub.j.sup.i  (2) wherein Z.sub.j.sup.i represents the trend value contained in the current integrating speed, i represent the i.sup.th moment, and j represents the corresponding j.sup.th monitoring point; Z.sub.j.sup.i-1 represents the drift trend value corresponding to the previous time point, and when i=2, the value of Z.sub.j.sup.i-1 is only the integrating speed of the previous time point, i.e., the initial point; v represents the integrating speed of the j.sup.th monitoring point at the i.sup.th time point; a is the weight value corresponding to Z.sub.j.sup.i-1; if the sampling frequency of an acceleration sensor is f, the sampling interval between two adjacent time points is 1/f, and the corresponding weight value is expressed as follows: $\begin{array}{cc}\alpha =\frac{1/4}{1/4+1/f}& \left(3\right)\end{array}$ it can be known from expression (2) that the drift trend value at the current monitoring point not only depends on the trend value corresponding to the previous time point, but also depends on the speed value obtained by integration at the current time point; it can be known from expression (3) that the value of α is close to 1, that is, the drifting trend corresponding to the current time point is greatly affected by the drifting trend of the previous time point, and the current integrating speed value only gives the drifting trend of the previous time point a step size; and the de-noised speed corresponding to the current time point is expressed as follows: $\begin{array}{cc}\begin{array}{c}{\stackrel{⌢}{v}}_j^i=v_j^i-Z_j^i\\ =v_j^i-\alpha .Math.Z_j^{i-1}-\left(1-\alpha \right).Math.v_j^i\\ =\alpha .Math.\left(v_j^i-Z_j^{i-1}\right)\end{array}& \left(4\right)\end{array}$ wherein {circumflex over (v)}.sub.j.sup.i represents a de-noised ideal speed; to obtain the corresponding displacement, it is necessary to integrate the de-noised ideal speed, i.e., formula (4), again, and the integration process is as follows:
s.sub.j.sup.i=({circumflex over (v)}j.sup.i-1+{circumflex over (v)}.sub.j.sup.i)×Δt/2  (5) wherein s.sub.j.sup.i represents the displacement corresponding to the i.sup.th monitoring time point, and j represents the reference sign of the monitoring point; considering that the de-noised ideal speed still contains some noise information, the noise information is shown in the form of drift in the displacement obtained after integration; therefore, to ensure the accuracy of the displacement information finally acquired, it is necessary to de-noise the displacement information again; a dynamic online high-pass filter method is adopted to de-noise the displacement information, and a dynamic filter method is adopted to filter data online with regard to real-time integration results; and a Butterworth filter is adopted, the principle of which is expressed as follows: $\begin{array}{cc}{.Math.H\left(w\right).Math.}^2=1-\frac{1}{1+{\left(\frac{w}{w_c}\right)}^{2m}}=1-\frac{1}{1+{{\epsilon }^2\left(\frac{w}{w_p}\right)}^{2m}}& \left(6\right)\end{array}$ wherein w.sub.c represents the cut-off frequency corresponding to the filter, m represents the order of the corresponding filter, and w.sub.p represents the pass-band edge frequency; when the corresponding frequency w is more than the cut-off frequency, that is w>w.sub.c the corresponding ${\left(\frac{w}{w_c}\right)}^{2m}$  value is close to infinity, and the |H(w)|.sub.2 value is closer to 1, which indicates that all the signals pass. when the corresponding signal frequency is lower than the cut-off frequency, that is w≤w.sub.c, the corresponding ${\left(\frac{w}{w_c}\right)}^{2m}$  value is close to 0, and |H(w)|.sup.2 is close to 0; at this moment, the corresponding signal will be cut off, which is 0; and the process is expressed as follows: $\begin{array}{cc}{.Math.H\left(w\right).Math.}^2=\{\begin{array}{cc}0& w\le w_c\\ 1& w>w_c\end{array}& \left(7\right)\end{array}$ the larger the filter order of n is, the smoother the filtered value is, and the closer the filter is to an ideal filter; the filter order adopted is set to 8, the cut-off frequency is set to 0.012×w, and high-pass filter is adopted to de-noise the integrated displacement, that is, when the frequency is higher than the set cut-off frequency, the displacement obtained by integration can be regarded as the final ideal value, and when the frequency corresponding to the displacement point is lower than the cutoff frequency, the value corresponding to this part is flattened; and the de-noised integral displacement is expressed as follows:
ŝ.sub.j.sup.i=s.sub.j.sup.i−N.sub.j.sup.i  (8) wherein s.sub.j.sup.i represents the result obtained by the corresponding direct integration in formula (5), N.sub.j.sup.i represents the noise value obtained by high-pass filter, and ŝ.sub.j.sup.i represents the corresponding ideal integral displacement value after de-noising; step (2): embedding the dynamic integration algorithm designed in step (1) into an edge device so as to realize dynamic fast processing of online signals based on edge computing.

Description

DESCRIPTION OF DRAWINGS

[0024] FIG. 1 is a flow chart of the present invention;

[0025] FIG. 2 is a schematic diagram of comparison of edge computing and a traditional cloud computing method;

[0026] FIG. 3 is a schematic diagram of double integral of high-pass filter.

[0027] FIG. 4 is a schematic diagram of exponentially weighted average double integral.

[0028] FIG. 5 is a schematic diagram of integration results of a method of the present invention.

DETAILED DESCRIPTION

[0029] The present invention is further described below in combination with specific embodiments.

[0030] Corresponding test functions are constructed to verify the effectiveness of the proposed method, and the constructed test functions are as follows:

[00007]$\begin{array}{cc}\{\begin{array}{c}s\left(t\right)=-2\sin \left(4\pi t\right)\\ v\left(t\right)=-8\pi \cos \left(4\pi t\right)\\ a\left(t\right)=32{\pi }^2\sin \left(4\pi t\right)+n\end{array}& \left(9\right)\end{array}$

[0031] The constructed test functions are shown above, wherein s(t), v(t) and a(t) represent displacement, speed and acceleration respectively, and n represents added noise. In order to make the constructed test functions closer to values measured in the real environment, the present embodiment respectively adopts values with signal-to-noise ratios (SNR) of 5, 10, 15 and 20 for verification, and the target displacement value is s(t), which is used as a target value for verifying the effectiveness of the method. The specific implementation steps are as follows:

[0032] 1) First, determining the sampling frequency of a sensor, so as to determine a weight value α corresponding to the previous time point, as shown in formula (3); here the sampling frequency is set to 100 Hz, that is, 100 acceleration sample points are collected within 1 s, so the change step of time t is set to 0.01, that is, 100 sample points can be collected within 1 s, and in this way, the corresponding weight value is

[00008]$\alpha =\frac{1/4}{1/4+1/100}=0.9615.$

[0033] 2) Integrating acceleration data, as shown in formula (1); when the variation range oft is 0-20, the number of the corresponding collected points is 2000, and 5 sampling points with numbers of 600-604 are selected for case analysis, wherein the value of SNR is set to 20, the corresponding acceleration data is [−1.4575754983679766e-13, 5.58146865524312, 13.513797023872176, 13.211154730847605, 21.60266538485112], and according to formula (1), that is, the corresponding values of a.sub.600, a.sub.601, a.sub.602, a.sub.603 and a.sub.604 are shown above and the corresponding time is Δt=0.01, speeds corresponding to v.sub.600, v.sub.601, v.sub.602, v.sub.603 and v.sub.604 are [−3.9534573072086268, −3.9254971656484674, −3.8499622325798386, −3.723061399174048, −3.5444222722567176].

[0034] 3) Judging whether the number of accelerations currently collected is more than ¼ of the number of sampling frequencies, if the number of the collected accelerations is less than ¼ of the number of sampling frequencies, it can be considered that the current integrating speed is less affected by drift, then directly going to step 6); otherwise, doing the next step;

[0035] 4) As shown in formula (2), calculating the drift value contained in the integrated speed according to formula (2);

[0036] 5) Subtracting the drift calculated according to formula (2) from the current speed integration results to obtain a de-drifted speed value, as shown in formula (4); calculating the corresponding trend values Z.sub.600, Z.sub.601, Z.sub.602, Z.sub.603 and Z.sub.604 as respectively [−1.9014722432840203, −2.028966441121721, −2.1481829308984572, −2.2576934812704277, −2.3560161996893796], and the de-noised speeds are [−3.177929670806483, −3.031373900547152, −2.833138321092142, −2.576138367056468, −2.2770464786599627].

[0037] 6) Integrating the integrated speed again to obtain a corresponding displacement value, as shown in formula (5). Judging whether the number of acceleration monitoring points currently acquired is more than ¼ of the number of sampling frequencies, if the number of the collected acceleration monitoring points is less than ¼ of the number of sampling frequencies, it is considered that the current integral displacement is less affected by the drift, and the current integral displacement can be directly regarded as the final displacement value. Otherwise, entering step 7); further integrating the de-noised speed data obtained in step 5) to obtain corresponding displacement values S.sub.600, S.sub.601, S.sub.602, S.sub.603 and S.sub.604 which are specifically [1.3095679261476776, 1.2801369860087743, 1.2523778296973533, 1.2267627167978188, 1.2036824421874557].

[0038] 7) Dynamically de-noising the integrated displacement online, as shown in formulas (6) and (7);

[0039] 8) Subtracting noise effect in the displacement from the current integral displacement to obtain the final ideal displacement, as shown in formula (8); and the corresponding displacement after final high-pass filter are [9.8321e-11, −0.0734, −0.1203, −0.1837, −0.1868], wherein the corresponding true values are [8.3975e-16, −0.0358, −0.0708, −0.1047, −0.1368].

[0040] 9) Embedding the proposed algorithm into an edge device to perform fast calculation near a data source, as shown in FIG. 2(a).

[0041] A true acceleration is simulated by adding Gaussian noise to the constructed acceleration, and the noise level (SNR, signal-to-noise ratio) is controlled to verify the effectiveness of the proposed method. Now the traditional methods and the method of the present invention are analyzed and compared, and the integration results of the solutions are compared as follows:

[0042] FIG. 3 shows the corresponding results obtained by means of double integral and high-pass filter, and the process can be described as follows: acceleration data containing noise is dynamically integrated, the corresponding speed values are obtained by online integral calculation, the integrated speed values are dynamically calculated respectively, the speed values dynamically integrated online are subjected to high-pass filter to reduce the drift value in the integrated speeds, and the speed subjected to high-pass filter is regarded as the ideal speed. Then the ideal speed is integrated again to obtain the corresponding displacement value. Since the drift in the speed is not eliminated completely, the integrated displacement is subjected to high-pass filter again to further reduce the influence of the drift on the integration results. The final integration results are shown in FIG. 3. The local enlarged view of integral displacement is located in the upper part, and the global reduced view is located in the lower part.

[0043] FIG. 4 shows the corresponding results obtained by means of double integral and exponential weighted average, and the process can be described as follows: acceleration data containing noise is dynamically integrated, the obtained speed value is integrated online, the speed value is exponentially weighted to obtain the corresponding de-noised ideal speed, and the ideal speed is further integrated to obtain the corresponding displacement value. The integral displacement value is further de-noised again by means of exponential weighted average to obtain the corresponding de-noised ideal displacement result, as shown in FIG. 4.

[0044] FIG. 5 is a schematic diagram of integration results of a method proposed by the present invention. It can be known by comparison and analysis that when the signal-to-noise ratio (SNR) is increased, i.e., noise in the original acceleration is reduced, the trend change of the integral displacement is closer to ideal displacement, as shown in the figure. Compared with the results shown in FIG. 3 and FIG. 4, when the SNR changes from 5 to 20, the fluctuation range of the integral displacement of the proposed method is small, that is, the integration accuracy is less affected by noise, which proves that the method has good robustness. Moreover, it can be known by comparing the integration results in FIG. 3, FIG. 4 and FIG. 5 that when the SNR is equal to different values, the integral values of the method proposed by the present invention are closer to the ideal displacement. Therefore, the method of the present invention has accurate integration effects.