Implementation Method and System of an Accelerometer with Low Noise, Wide Bandwidth and High Dynamic Range

Abstract

The present invention discloses a method and system for implementing an acceleration sensor with low noise, wide bandwidth, and high dynamic range. Based on a set of three accelerometers with the same location and pose, acceleration data of three channels are collected for data fusion. The fused acceleration data is preprocessed, event samples are generated through a time series sliding window, and sample feature vectors are extracted. The event samples are labeled to organize a training data set, and a machine learning classification prediction model is trained to map the feature vectors to classification prediction information. According to the classification prediction information, data channel decision-making of the acceleration sensor is performed, channel switching of the acceleration sensor is completed within the channel switching area, and acceleration data filtering is performed within a fixed time of channel switching. Thus, an acceleration data output with low noise, wide bandwidth, and high dynamic range is obtained. The present invention enables the acceleration sensor system to have advantages such as wide bandwidth, high precision, and large dynamic range.

Description

Technical Field

[0001] The present invention relates to the field of acceleration sensors, and particularly to a method and system for implementing an acceleration sensor with low noise, wide bandwidth, and high dynamic range. Background Art

[0002] With the progress of technology, the industrial production and manufacturing field has basically achieved automation and intelligence. At the same time, the process requirements for precision device processing are constantly increasing. Therefore, in the precision manufacturing and detection links, any subtle vibration cannot be ignored.

[0003] A working process of an industrial robot arm consists of several actions, and different actions have different vibration characteristics. This uncontrollable vibration generally has the characteristics of a large range (10 ng - g) and a wide frequency band (1 - 500 Hz). The precise detection of subtle vibration is a prerequisite for suppressing or compensating it. Therefore, there is an urgent need for an acceleration sensor that simultaneously has the characteristics of low noise, wide bandwidth, and high dynamic range. Since the noise index, bandwidth index, and dynamic range index of the acceleration sensor restrict each other, conventional acceleration sensors such as piezoelectric, photoelectric, resonant, and capacitive ones cannot meet the requirements; the composite acceleration sensor using the idea of data fusion performs channel polling acquisition and data fusion on acceleration sensors with different ranges, and ensures the accuracy to the greatest extent while expanding the dynamic range, but this method will inevitably limit the single-channel sampling rate and does not meet the requirements of wide bandwidth. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for implementing an acceleration sensor with low noise, wide bandwidth, and high dynamic range.

[0005] A technical solution for achieving the purpose of the present invention is: A method for implementing an acceleration sensor with low noise, wide bandwidth, and high dynamic range, comprising the following steps:

[0006] Step 1, based on a 3-accelerometer group with the same location, the same pose, and complementary dynamic ranges, collect acceleration data of 3 channels for data fusion, preprocess the fused acceleration data, generate event samples through a time-series sliding window, and extract sample feature vectors;

[0007] Step 2, label the event samples to organize a training data set, train a machine learning classification and prediction model, and map the feature vectors to classification and prediction information;

[0008] Step 3, make a data channel decision for the acceleration sensor according to the classification and prediction information, complete the acceleration sensor channel switching within the channel switching area, and filter the acceleration data within the fixed time of channel switching, thus obtaining an acceleration data output with low noise, wide bandwidth, and high dynamic range.

[0009] A system for implementing a low-noise, wide-bandwidth, high-dynamic-range acceleration sensor, characterized by implementing the method for implementing a low-noise, wide-bandwidth, high-dynamic-range acceleration sensor described above.

[0010] Compared with the prior art, the present invention has the following remarkable advantages: 1) By using the method of switching acceleration sensor channels, compared with the existing sensor channel data fusion method, a wider bandwidth and higher accuracy can be achieved, and at the same time, it has a large dynamic range; 2) By using a machine learning algorithm to assist in predicting channel switching, giving full play to the characteristics of the programmed motion sequence of the industrial robot arm, the real-time performance of channel switching and the continuity of data are guaranteed to the greatest extent; 3) By adopting an adaptive channel decision scheme, introducing a sequence matching prediction item and an adaptive item, organically integrating the prediction information and the observation information, ensuring the reliability of the channel switching decision information, and at the same time limiting the data channel switching to be executed within the channel switching area, improving the fault tolerance of the channel decision information and ensuring the correctness of the acceleration data. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a flowchart of the method for implementing a low-noise, wide-bandwidth, high-dynamic-range acceleration sensor of the present invention.

[0012] Figure 2 It is a schematic diagram of the data channel switching area and the vibration level reference point of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0013] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0014] The flow of the method for implementing a low-noise, wide-bandwidth, high-dynamic-range acceleration sensor is as Figure 1 shown, mainly including a multi-channel low-noise acceleration sensor group, an MCU data information processing and control module, and an external auxiliary model training. The MCU obtains high-dynamic-range acceleration data through the acceleration sensor group, classifies and predicts according to the characteristics contained in the data, and finally completes the switching of the data channel when entering the channel switching area. This method specifically includes the following steps:

[0015] Step 1: Fuse the accelerometer data and construct a feature vector

[0016] When an industrial robot arm is in production manufacturing, the acceleration of the end vibration can span 8 orders of magnitude. Therefore, it is necessary to obtain acceleration data with a large dynamic range to ensure the recognizable features of the extraction process actions. The present invention adopts a polling method to collect 3-channel acceleration data and fuse it to obtain acceleration data with a large dynamic range. Then, the acceleration data is preprocessed by removing outliers and filling in missing values. Event samples are generated through a time series sliding window, feature extraction of the event samples is performed, and feature dimensionality reduction is carried out using technical means such as correlation analysis to obtain feature vectors.

[0017] Step 1 includes the following 3 sub-steps:

[0018] Step 11: Poll and fuse the data of 3 acceleration sensor channels. It should be understood that the fused acceleration data is only used to construct event samples and has a relatively low requirement for accuracy. Therefore, in the implementation mode of the present invention, a mean data fusion filtering based on threshold segmentation is adopted. This method has a high fusion efficiency, but the specific data fusion method is not limited to this. The specific representation of the mean data fusion filtering is as shown in Equation (1):

[0019]

[0020] In the formula, a out is the fused acceleration data, a chi is the acceleration data of the i-th channel, TL i and TH i are respectively the lower and upper limits of the effective dynamic range of the acceleration sensor of the i-th channel.

[0021] Step 12: Preprocess the acceleration data. In the data preprocessing stage of the implementation mode of the present invention, the real-time sliding window quartile test method is used to detect outliers, and the cubic spline interpolation function f θ (x) is used to fill in the missing points in the time dimension or replace outliers, and then sliding sampling is performed on the time series data through a sliding window to obtain event samples;

[0022] For the real-time sliding window quartile test method, it is required that the width of the real-time sliding window is not less than 4. The specific method is: find the upper quartile UQ and the lower quartile LQ, calculate IQR = UQ - LQ, and define the data greater than UQ + 1.5IQR and less than LQ - 1.5IQR as abnormal data;

[0023] The representation of the cubic spline filling value is as shown in Equation (2). In the formula, f θ (x) is the cubic spline interpolation function, and t miss is the data missing point in the time dimension.

[0024] P(t miss ) = f θ (t miss) Equation (2)

[0025] Step 13: Extract event sample features, mainly including statistical features, frequency domain features, energy features, etc., but not limited to these; further, reduce the dimension and select existing features based on means such as correlation analysis to form feature vectors;

[0026] Step 2: Label event samples based on the process actions of the robotic arm to organize the training data set, train the classification prediction model, and perform classification prediction in real time

[0027] The industrial robotic arm moves according to a fixed process flow, and distinguishes recognizable actions in the process based on the characteristics of the mechanical vibration at the end of the arm, establishing an action category space M. Based on this category space, a two-dimensional label group [A y (l,L)] n is generated, where A y is the y-th label, l represents the action category of the robotic arm (l ∈ M), and L represents the vibration level of the current action (i.e., the dynamic range pre-switching threshold). In the implementation mode of the present invention, a co-located and co-pose 3-channel acceleration sensor is adopted. Therefore, the setting of the vibration level L ∈ (L1, L2, L3) is shown in Figure 2 ; Label event samples through this two-dimensional label group, and use the labeled value as an additional dimension of the feature vector to obtain the training set where n is the number of training samples, F is the feature dimension, x ab is the b-th feature of the a-th sample, and y a is the label of the a-th sample; Build a machine learning algorithm engineering framework and train the classification prediction model. The present invention recommends using high-real-time SVM or KNN algorithms, but not limited to these; After training, obtain the classification prediction model, and in real time, obtain the classification prediction information P(l,L) through the feature vector mapping in Step 1;

[0028] Step 3: Make channel decisions and switch the acceleration data based on the classification prediction information, and output the final acceleration data

[0029] Identify the classification prediction information in Step 2 in real time to make data channel decisions, and introduce the process action sequence group of the process to assist in the decision-making;

[0030] The representation of the process action sequence space S of the process is as shown in Equation (3),

[0031] {s = [l o , l p … l q} N ∈ S Equation (3)

[0032] where, l o , l p , l q ∈ Mo,p,q＝0,1,2… , representing any action category of the robotic arm, and N is the number of action sequences.

[0033] The adaptive data channel decision algorithm is expressed as in Equation (4),

[0034]

[0035] where D i+1 is the discretized decision information of the (i + 1)-th frame. If D i+1 is not zero, it is regarded as a valid decision. i represents the i-th decision process, and L C is the vibration level information included in the classification prediction information in the i-th decision process. m represents the number of matching sequences; the function sign(x) is expressed as in Equation (5) and is used for the discretized mapping of decision information;

[0036]

[0037] In Equation (4), is the sequence matching prediction term. The function sigmoid(x) is expressed as in Equation (6) and is used to map the action sequence matching index to between 0 and 1;

[0038]

[0039] is the action sequence matching index. Among them, p ij is the matching factor calculated from the j-th matching sequence s j in the i-th decision process. The calculation process of p ij is as in Equation (7). match ij is the number of actions matching the sequence s j in the i-th decision process. The inspection range is from l c to l c-f . l c-f represents the action of the (i - f)-th decision process. If i - f < 0, the action is default-matched. L ij is the vibration level corresponding to the next action in the j-th matching sequence s j in the i-th decision process. The function f(x) is expressed as in Equation (8) and is used to distinguish the conformity between the vibration level information L ij based on sequence matching and the prediction information L C . If they conform, the credibility of the classification prediction increases positively; otherwise, it decreases. Thus, the sequence matching prediction term is calculated;

[0040]

[0041]

[0042] c in Equation (4) i g(a mix ) is the adaptive term, and c i is the adaptive factor for the i-th decision-making process, which is limited between 0 and 1. Its iterative formula is shown in Equation (9). δ is the step size of iterative update. The function h(a min , a chi ) is expressed as Equation (10) and is used to determine the direction of iterative update. Among them, a mix and a chI are the fused acceleration and the acceleration of the current channel I respectively, and ε I is the typical value of the absolute error of the current data channel I;

[0043] c i+1 = [c i + δ × h(a mix , a chI )] ∈ (0, 1) Equation (9)

[0044]

[0045] The adaptive term c i in g(a mix ) The function g(x) is expressed as Equation (11). When the acceleration sensor data channel needs to be switched, according to the pre-switched vibration level L c (dynamic range pre-switching threshold), the current vibration level L i , and the magnitude relationship of the fused acceleration value |a mix |, the adaptive adjustment factor g is calculated. Thus, the adaptive term is further calculated.

[0046]

[0047] The data channel switch can be embedded in the MCU to achieve soft switching, or hard switching can be adopted. The channel switch will perform channel switching when entering the data channel switching interval. The logical expression of data channel switching is shown in Equation (12). The schematic diagram of the channel switching area is shown as i+1 , Figure 2 shown.

[0048]

[0049] During the switching process, digital signal processing technologies such as the first-order Kalman filtering algorithm are used to make the acceleration data smoothly transition at the moment of channel switching. Thus, acceleration data with low noise, wide bandwidth, and high dynamic range can be measured.

[0050] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0051] The above-described embodiments merely represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for implementing an acceleration sensor with low noise, wide bandwidth, and high dynamic range, characterized in that It includes the following steps: Step 1: Based on a 3-accelerometer group with the same location, the same pose, and complementary dynamic ranges, collect acceleration data from 3 channels for data fusion, preprocess the fused acceleration data, generate event samples through a time-series sliding window, and extract sample feature vectors; Step 2: Label event samples to organize a training data set, train a machine learning classification and prediction model, and map the feature vectors to classification and prediction information; Step 3: Make a data channel decision for the acceleration sensor according to the classification and prediction information, complete the acceleration sensor channel switching within the channel switching area, and perform acceleration data filtering within the fixed time of channel switching. Thus, obtain an acceleration data output with low noise, wide bandwidth, and high dynamic range; Step 1: Collect acceleration data from 3 channels for data fusion, preprocess the fused acceleration data, and generate event samples through a time-series sliding window, and extract sample feature vectors. The specific method is as follows: Step 11: Poll and collect data from 3 acceleration sensors, and use a mean data fusion filtering algorithm based on threshold segmentation for real-time data fusion. The specific expression of the mean data fusion filtering is shown in Equation (1): Where a out is the fused acceleration data, and a chi is the acceleration data of the i-th channel, TL i and TH i are respectively the lower and upper limits of the effective dynamic range of the acceleration sensor of the i-th channel; Step 12: Use the real-time sliding window quartile test method to remove abnormal data points from the fused acceleration data, and use the cubic spline interpolation function f θ (x) to complement the missing points in the time dimension, and perform sliding sampling on the time series data through a sliding window to obtain event samples; Step 13: Perform event sample feature extraction, including statistical features, frequency domain features, and energy features. Based on correlation analysis, reduce the dimension and select the extracted features to form a feature vector; Step 3: Make a data channel decision for the acceleration sensor according to the classification and prediction information, complete the acceleration sensor channel switching within the channel switching area, and perform acceleration data filtering within the fixed time of channel switching. The specific method is as follows: Step 31: Introduce a process action sequence group for auxiliary decision-making, as shown in Equation (3): {s = [l o , l p … l q} N ∈ S formula (3) where S represents the process action sequence space, s represents the process action sequence, l o , l p , l q ∈ M o,p,q＝0,1,2… represents any action category of the robot arm, M represents the finite set of action categories of the robot arm, and N is the number of action sequences; Step 32: Make a data channel decision according to the classification and prediction information, as shown in Equation (4): where D i+1 is the channel switching decision. When D i+1 is not zero, it is regarded as an effective decision. L C is the pre-switching vibration level included in the classification prediction information in the i-th decision process, and m represents the number of matching sequences; the function sign(x) is expressed as in Equation (5) and is used for the discretization mapping of decision information; In formula (4), is the sequence matching prediction term. The function sigmoid(x) is expressed as in formula (6) and is used to map the action sequence matching exponent between 0 and 1; is the action sequence matching index, where p ij is the matching factor calculated by the j-th matching sequence s j in the i-th decision-making process. The calculation process of p ij is as shown in Equation (7), and match ij is the number of actions matching the sequence s j in the i-th decision-making process. The inspection range is from l c to l c-f , and l c-f represents the action in the (i - f)-th decision-making process. If i - f < 0, the default action is matched; L ij is the vibration level corresponding to the next action in the j-th matching sequence s j in the i-th decision-making process. The representation of the function f(x) is as shown in Equation (8), which is used to distinguish the conformity between the vibration level L ij based on sequence matching and the pre-switching vibration level L C . If they are in conformity, the credibility of positive growth classification prediction is increased, otherwise it is decreased. Thus, the sequence matching prediction item is calculated; c in Equation (4) i g(a mix ) is an adaptive term, and c i is the adaptive factor for the i-th decision-making process, which is limited between 0 and 1. Its iteration formula is as shown in Equation (9). δ is the step size of iterative update. The function h(a mix , a chI ) is expressed as in Equation (10) and is used to determine the direction of iterative update. Among them, a mix and a chI are the fused acceleration and the acceleration of the current channel I respectively, and ε I is the typical value of the absolute error of the current data channel I; c i+1 = [c i + δ × h(a mix , a chI )] ∈ (0,1) Equation (9) Adaptive term c i g(a mix ) The g(x) function in it is expressed as in Equation (11). When the acceleration sensor data channel needs to be switched, according to the pre-switching vibration level L c , the current vibration level L i , and the magnitude relationship of the fused acceleration value |a mix |, the adaptive adjustment factor g is calculated, and thus the adaptive term is further calculated; Step 33, according to the channel switching decision D i+1 , complete the acceleration sensor channel switching within the channel switching area, perform acceleration data filtering within the fixed time of channel switching, and the logical expression of data channel switching is as shown in Equation (12):

2. The method for implementing an acceleration sensor with low noise, wide bandwidth, and high dynamic range according to claim 1, wherein Step 2: Label event samples to organize a training data set, train a machine learning classification and prediction model, and map the feature vectors to classification and prediction information. The specific method is as follows: Step 21, according to the process flow actions of the industrial robot arm, combine the characteristics of the mechanical vibration at the end of the arm to distinguish the recognizable actions in the process, establish the action category space M, and generate the two-dimensional tag group [A y (l,L)] n , where A y is the y-th tag, l, l ∈ M represents the action category of the robot arm, and L, L ∈ (L1, L2, L3) represents the vibration level of the current action; Step 22: Label the event samples based on the two-dimensional tag group, and use the labeled values as additional dimensions of the feature vectors to obtain the training set where n is the number of training samples, F is the number of feature dimensions, and x ab is the b-th dimension feature of the a-th sample, and y a is the label of the a-th sample; Step 23: Use the SVM or KNN algorithm to train the classification and prediction model, and map the feature vectors to the classification and prediction information P(l,L).

3. According to the method for realizing a low-noise, wide-bandwidth, and high-dynamic-range acceleration sensor described in claim 1, in step 33, use a first-order Kalman filtering algorithm to smoothly transition the data during the switching process. Thus, obtain a low-noise, wide-bandwidth, and high-dynamic-range acceleration data.

4. The method for implementing an acceleration sensor with low noise, wide bandwidth, and high dynamic range according to claim 1 or 3, characterized in that The channel switching adopts hard switching, or soft switching is performed by embedding an MCU in the data channel switcher.

5. An acceleration sensor implementation system with low noise, wide bandwidth, and high dynamic range, characterized in that, Implement the method for realizing a low-noise, wide-bandwidth, and high-dynamic-range acceleration sensor according to any one of claims 1-4.

Images