CN113361596B - Sensor data augmentation method, system and storage medium - Google Patents

Abstract

The invention discloses a sensor data augmentation method, a system and a storage medium, which belong to the technical field of sensor data processing. This method uses random cropping for the drift-free data to augment the drift-free data samples, and selects multiple function models that conform to the sensor drift characteristics as trend items to construct a non-stationary random walk process with trend items to simulate sensor data. In the drift process, the drift simulation is realized by setting the sensor drift probability threshold and determining the maximum drift range according to the data characteristics. The drift-containing data samples are augmented by adding drift to the augmented drift-free data samples. The invention comprehensively considers the time and space characteristics of sensor drift data by fusing the data of each adjacent sensor in the perception field, ensures the correctness of the drift amount simulation and the diversity of the augmented data characteristics, and overcomes the problems caused by the training samples in the sensor drift calibration model. The problem of weak generalization ability of the model caused by the deficiency.

Description

A sensor data augmentation method, system and storage medium

technical field

The present invention relates to the technical field of sensor data processing, and more particularly, to a sensor data augmentation method, system, storage medium and sensor data drift calibration method.

Background technique

The drift of sensor data over time in the agricultural Internet of Things is a key constraint that affects the quality of data collected by sensors. The existence of sensor drift makes the intelligent regulation of IoT devices and the effective analysis of data cannot be guaranteed. Because sensors are usually large-scale Unloading and recalibrating sensors alone is often difficult to achieve and long-term deployment, so it is increasingly important to calibrate sensors when a drift-free true signal cannot be obtained.

At present, the calibration of sensors based on deep learning has attracted the attention of the inventors. Among them, Convolutional Neural Network (CNN) is a subset of deep learning, which has been applied in other fields and achieved good results. Research on the application of CNN to the time series field is also gradually emerging. For example, for EEG signal denoising, this method can not only establish an accurate mapping of noise signal to EEG signal, realize real-time denoising, but also effectively improve the efficiency and quality of EEG signal denoising. However, this method is limited to denoising in a single time series data, and is not suitable for drift calibration of multiple sensors.

In addition, Bao et al. (2018) in Structural Health Monitoring, 18(2), 401-421 convert time-series signals into image vectors, plot them segmentally in grayscale images, and then combine randomly selected and manually labeled image vectors into The training dataset is fed into a deep neural network or set of deep neural networks trained with stacked autoencoders and greedy layered training techniques to more accurately detect time-series data, including data drift However, this method can only detect and classify abnormal patterns in time series data, and does not involve the calibration of sensor drift. In 2020, Tian et al. proposed a new deep learning model in IEEE Access, 8, 121385-121397, which combines unsupervised and supervised techniques to achieve drift compensation of electronic noses, but this method is relatively complex and does not involve training sets Research on construction methods, applicable scenarios are limited.

Therefore, based on the above analysis, the existing methods cannot effectively meet the sensor drift calibration requirements. Deep learning has the ability to automatically learn features from a large amount of data, and the application of deep learning methods to the field of sensor drift calibration has great application prospects. Therefore, it is very meaningful to use deep learning to automatically extract the drift features in the sensory data and realize the drift calibration of the sensor. However, the training of deep learning models such as neural networks requires a large amount of data samples. How to construct sufficient data samples with drift and the appropriate size of iterative input training samples is an urgent problem to be solved by those in the field. On the other hand, how to ensure the extraction of drift features The accuracy of the calibration and the precision of the calibration are also a challenge.

SUMMARY OF THE INVENTION

1. The problem to be solved

Aiming at the problem that training of deep learning models such as neural networks in the prior art requires a large number of data samples, how to construct sufficient drift-containing data samples and the appropriate size of iteratively input training samples, the present invention provides a sensor data augmentation method, This method uses random cropping for the drift-free data to augment the drift-free data samples, and selects multiple function models that conform to the sensor drift characteristics as trend items to construct a non-stationary random walk process with trend items to simulate sensor data drift. In the process, by setting the sensor drift probability threshold and determining the maximum drift range according to the data characteristics, simulating the data drift characteristics to realize the drift amount simulation, and adding the drift amount to the augmented drift-free data samples to realize the increase of the drift-containing data samples. The invention comprehensively considers the time and space characteristics of sensor drift data by fusing the data of each adjacent sensor in the perception field, ensures the correctness of the drift amount simulation and the diversity of the augmented data characteristics, and overcomes the sensor drift based on deep learning. The problem of weak generalization ability of the model caused by insufficient training samples in the calibration method. The Squeeze and Excitation (SE) module is embedded in the residual block of Residual Networks (ResNet) to construct a Squeeze-and-Excitation Residual Networks (SE-) based on Squeeze-and-Excitation Residual Networks (SE-) ResNet) drift calibration method realizes sensor drift data calibration, which effectively improves the efficiency and quality of sensor drift calibration.

2. Technical solutions

In order to solve the above problems, the present invention adopts the following technical solutions.

A first aspect of the present invention provides a sensor data augmentation method, the method comprising:

SA: Deploy calibrated sensors to build target perception fields;

SB: Collect the first perception data of the calibrated sensor in the target perception field, use the random clipping method to clip the first perception data into multiple data matrices, and from the single clipped data matrix, according to the rows of the matrix corresponding to the sensor in the matrix. The position in the perception field, select part of the row data in the data matrix to construct the adjacent sensor data matrix, and determine the first augmented data;

SC: According to the data matrix tailored from the first sensing data, use the non-stationary random walk process with trend items to simulate the drift process of the sensor data, obtain the second sensing data, select and first increase the second sensing data from the second sensing data. The drift matrix of the same position and size of the augmented data is determined as the second augmented data.

In some embodiments, the first sensing data is determined according to the consecutive L sampling points collected by the calibrated N sensors in the target sensing field, and denoted as a drift-free data matrix X with a size of N×L , wherein the N matrix row data in the drift-free data matrix X respectively correspond to the data collected by N sensors in the sensing field;

A plurality of data matrices with a size of N×1 are randomly cut out from the drift-free data matrix X, and any data matrix obtained by cutting is denoted as X ^l , and the number of columns of the data matrix X ^l is described as the drift-free data matrix The number of columns of the data matrix X, from a single cropped data matrix X ^l , according to the position of the sensor corresponding to each matrix row in the perception field, select n rows of data from the data matrix X ^l to construct a neighborhood of size n × l. The sensor data matrix X ^b , and denoted as the first augmented data; it can be specifically expressed as:

为采样点的裁剪初始位置，l为裁剪数据矩阵的列数即各矩阵行对应传感器连续采样点的个数；1:N表示取矩阵的第1到第N行的数据，

表示取矩阵的第

到

列数据；n为从裁剪数据矩阵X^l中选取的行数。in

is the initial cropping position of the sampling point, l is the number of columns of the cropping data matrix, that is, the number of consecutive sampling points of the sensor corresponding to each matrix row; 1:N means to take the data from the 1st to the Nth row of the matrix,

Represents the first order of the matrix

arrive

Column data; n is the number of rows selected from the crop data matrix ^Xl .

A non-stationary random walk process with trend items is used to construct a drift simulation matrix D ^l of size N×l. According to the method of selecting n rows of data from X ^l to construct X ^b , n rows of the same row position are selected from D ^l . The data constructs a drift matrix D ^b with a size of n×l as the second augmented data; it can be specifically expressed as:

Among them, d is the drift amount of each sensor at different times.

In some embodiments, the method for selecting the number n of adjacent sensors is:

According to the data augmentation application scenario, the size of the number of adjacent sensors n is adaptively determined, and a matrix row is randomly selected from the cropped data matrix ^Xl as the reference, and the Euclidean relationship between other sensors in the sensing field and the reference sensor corresponding to the matrix row is used. From small to large, select n-1 adjacent sensors in sequence, and the matrix row data in the data matrix X ^l corresponding to the n sensors finally selected is the n row data to be selected.

In some embodiments, the length l of the cropped data matrix is determined by the time interval of the data collected by the sensor and the characteristic period of the data distribution, and the calculation method is as follows:

Among them, T is the characteristic period of the regular distribution of sensor data flow, Δt is the time interval of sensor data collection, and λ is a positive rational number. Generally, an appropriate positive integer is selected according to the required training sample size.

In some embodiments, a non-stationary random walk process with a trend term is used to construct a sensor drift simulation matrix D ^l of size N×l, and according to the method of selecting n rows of data from X ^l to construct X ^b , from D ^l Select n rows of data at the same row position to construct a drift matrix D ^b with a size of n×l, and denote it as the second augmented data;

The sensor drift probability threshold is set, and a non-stationary random walk process with a trend term is used to simulate the drift process of the sensor data. After the simulation, each sensor data can be expressed as:

Among them, y _i,t is the drift-containing data of sensor i at time t after simulation, xi _, t is the drift-free data of sensor i at time t, d _i,t is the amount of drift generated by sensor i at time t through simulation , rand(0,1) is a random floating point number between 0 and 1, α is a floating point number and α∈(0,1) is the probability threshold of sensor drift, when rand(0,1)>α, the sensor i does not drift, and the resulting drift is zero; otherwise, drift occurs;

Each sensor data adaptively selects one of a linear function, an exponential function, a square root function and a sine function as the trend item in the non-stationary random walk process with the trend item during the drift simulation, so as to generate the corresponding trend item. The drift trend of ; the specific method of drift simulation is as follows:

In the linear drift trend, the drift of sensor i at time t can be expressed as:

In the exponential drift trend, the drift of sensor i at time t can be expressed as:

In the square root drift trend, the drift of sensor i at time t can be expressed as:

In the sinusoidal drift trend, the drift of sensor i at time t can be expressed as:

Among them, e is the maximum drift in each trend item, ui _{, t} is the amount of random walk data, and r is the angular velocity parameter.

In some embodiments, the method for selecting the simulation parameters of each trend item drift is as follows:

其中，s为特征周期T内数据归一化后的标准差，Δt为传感器数据采样间隔。随机游走数据量u_i,t～iid(0,σ²)，其中

角速度参数r用于调整正弦周期，且

The value of the maximum drift e in each trend item should be determined by the characteristics of the collected data and the number of columns of the cropped data matrix, and is positively related to the number of columns in the data matrix, specifically:

Among them, s is the standard deviation of the normalized data in the characteristic period T, and Δt is the sampling interval of the sensor data. The random walk data volume u _i,t ～iid(0,σ ² ), where

The angular velocity parameter r is used to adjust the sine period, and

A second aspect of the present invention provides a sensor drift calibration method, the method comprising:

Acquire sensor data to be calibrated, input the sensor data to be calibrated into the trained squeeze-excitation residual network model, and output corresponding drift calibration data.

Wherein, the squeeze-excitation residual network model is obtained by training multiple sets of data samples; the multiple sets of data samples are a data set constructed by the first augmented data and the second augmented data.

In some embodiments, the step of constructing a dataset from the first augmented data and the second augmented data includes:

Taking the first augmented data as the drift-free data sample X ^b , and taking the second augmented data as the drift simulation data sample D ^b ;

Perform matrix addition of the drift-free data samples and the drift-amount simulation data samples to obtain a drift-containing data sample Y ^b ; use a plurality of drift-free data samples X ^b to form a drift-free data set X ^M , and use a plurality of drift-containing data samples to form The data set Y ^M with drift; the data set Y ^M with drift and the data set X ^M without drift are used as the data sets for training the squeeze-excitation residual network model. A third aspect of the present invention provides a sensor data augmentation system, including:

A perception field building block for deploying calibrated sensors to construct a target perception field;

The drift-free data sample building module is used to collect the first perception data of the calibrated sensor in the target perception field, and use the random clipping method to clip the first perception data into multiple data matrices, and from a single clipped data matrix According to the position of each row in the matrix corresponding to the sensor in the sensing field, select some rows in the data matrix to construct the adjacent sensor data matrix as the first augmented data, wherein the first augmented data is a drift-free data sample;

The drift simulation data sample building module is used to simulate the drift process of the sensor data by using the non-stationary random walk process containing the trend term according to the data matrix tailored from the first sensing data, and obtain the second sensing data. The drift matrix with the same position and size as the first augmented data is selected from the second perceptual data, and determined as the second augmented data, wherein the second augmented data is a drift simulation data sample.

A fourth aspect of the present invention provides a readable storage medium, the storage medium stores a computer program, the computer program includes program instructions, the program instructions when executed by a processor cause the processor to perform the above method.

3. Beneficial effects

Compared with the prior art, the present invention has obvious technical advantages:

(1) A data augmentation method proposed by the present invention constructs a sensor data drift process by adopting a non-stationary random walk process with trend items, and realizes the augmentation operation of sensor data by randomly cropping and selecting adjacent sensor data. , which increases the feature diversity of augmented data; and the data augmentation method can fuse the temporal and spatial features of sensor drift data, which overcomes the problem of weak model generalization ability caused by insufficient training samples in the sensor drift calibration model.

(2) In a data augmentation method proposed by the present invention, a non-stationary random walk process with a trend term is used to construct a sensor drift process, the probability threshold of sensor drift is set according to the uncertainty of sensor drift, and the probability threshold of sensor drift is set according to the sensor data characteristics. The range of the maximum drift amount of the trend item is set to control the amplitude of the drift of the sensing data, which increases the controllability; 4 types of function models conforming to the drift characteristics are used as the trend item to simulate the drift characteristics of the sensor, and the simulation parameters of each drift amount are selected based on the data characteristics. , which ensures the applicability of the drift data sample construction method in different application scenarios and different data types.

(3) In a data augmentation method proposed by the present invention, the data of each adjacent sensor in the sensing field is fused into the same training sample, and the size of the cropped data matrix is determined according to the sampling interval and data characteristic period of the sensor data as the sensor drift calibration network model The iterative input value can not only avoid the loss of data features caused by the iterative input of data samples, but also ensure that the input size of the neural network is moderate, which effectively reduces the computational cost.

(4) A sensor data drift calibration method proposed by the present invention uses a squeeze-excitation residual network to construct a sensor drift calibration model, and the squeeze-excitation residual module contained therein can utilize the correlation of adjacent sensor data to extract data. The temporal and spatial features of the sensor realize the calibration of sensor drift data; and the proposed data augmentation method can fuse the temporal and spatial features of sensor drift data to provide data guarantee for training deep neural networks.

(5) The sensor data drift calibration method proposed by the present invention can provide guarantee for accurately extracting drift features in multiple sensor data, and ensure the reliability of feature extraction and drift compensation for multiple sensor data in the sensing field at the same time. Using the squeeze-excitation residual network can not only effectively utilize the correlation of adjacent sensor data but also increase the network depth and effectively improve the quality of sensor data drift calibration.

(6) The data augmentation method proposed in the present invention has strong scalability, and the data of multiple sensors in the sensing field and the data of a single sensor in different time periods can be used as the basic data required for data augmentation. The data obtains the desired dataset through this data augmentation method.

Description of drawings

1 is a flowchart of a data augmentation method provided by an embodiment of the present invention;

2 is a block diagram of a data augmentation system provided by an embodiment of the present invention;

3 is a schematic diagram of a data set expansion method provided by the present invention;

Fig. 4 is the data set construction flow chart provided by the embodiment of the present invention;

5 is a flowchart of a sensor drift calibration method provided by the present invention;

6 is a schematic diagram of the SE-ResNet residual module provided by the present invention;

7 is a histogram of drift calibration errors of SE-ResNet and ResNet under different drift trends provided by an embodiment of the present invention;

Detailed ways

Hereinafter, embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited by the embodiments described herein.

Example 1

As shown in FIG. 1 , an embodiment of the present application discloses a sensor data augmentation method, and the overall process is as follows:

SA: Deploy calibrated sensors to build target perception fields.

Specifically, in the monitoring of soil environmental information, calibrated sensors are deployed to construct a target perception field. In this example, the sensors arranged in the target perception field are soil temperature and humidity sensors, the number of which is 20, and the sensor sampling interval is 5 minutes. , the characteristic period of the regular distribution of the sensing data flow is 24 hours. After calibrating the sensor, the soil temperature data for a total of 90 days from February 1 to April 30, 2020 was taken as the initial acquisition data of the sensor. Among them, the periodic feature refers to the fact that the sensory data changes with time, and its distribution law presents periodicity.

By collecting initial sensing data from calibrated sensors within the sensing field, this data can be considered drift-free, and this data serves as the basis for the drift data samples and calibration data needed to build the sensor drift calibration network model.

SB: Collect the first perception data of the calibrated sensor in the target perception field, and use the random clipping method to clip the first perception data into multiple data matrices. position in the field, select some row data in the data matrix to construct the adjacent sensor data matrix, and determine the first augmented data.

Specifically, it includes the following steps: SB1: Model the first perception data as a drift-free data matrix.

The soil temperature data collected by the calibrated soil temperature sensor for a total of 90 days from February 1 to April 30, 2020 were selected from the target perception field as the first perception data. The number N of sensors is 20, the sensor sampling interval is 5 minutes, and the characteristic period is 24 hours, which is recorded as a drift-free data matrix X with a size of 20×25920, in which there are 20 matrix rows in the drift-free data matrix X The data respectively correspond to the data collected by 20 sensors in the sensing field;

SB2: Trim the drift-free data matrix X into multiple data matrices, where the column number ^l of the trimmed data matrix X1 is smaller than the column number L of the drift-free data matrix X.

Specifically, in order to avoid the loss of data features, it is necessary to cut the drift-free data matrix into a data matrix with a smaller size according to the data collection features in a specific monitoring scenario, and randomly cut out a size of N × from the drift-free data matrix X above. The data matrix of l is denoted as X ^l .

SB3: From the cropped data matrix, from a single cropped data matrix, according to the position of the sensor in the sensing field corresponding to each row in the matrix, select some row data in the data matrix to construct the adjacent sensor data matrix to determine the first augmentation data.

Select n rows of data from the data matrix X ^l to construct a neighboring sensor data matrix X ^b with a size of n×l, and record it as the first augmented data. The specific method is as follows:

为采样点的裁剪初始位置，l为裁剪数据矩阵的列数即各矩阵行对应传感器连续采样点的个数。1:N表示取矩阵的第1到第N行的数据，

表示取矩阵的第

到

列数据。n为从裁剪数据矩阵X^l中选取的行数。in

is the initial cropping position of the sampling point, and l is the number of columns of the cropping data matrix, that is, the number of consecutive sensor sampling points corresponding to each matrix row. 1:N means to take the data from the 1st to the Nth row of the matrix,

Represents the first order of the matrix

arrive

column data. n is the number of rows selected from the cropped data matrix ^Xl .

As a variation example, the method for selecting the number n of sensors in this example is:

Determine the size of the number n of adjacent sensors according to the receptive field of the neural network, randomly select a matrix row from the cropped data matrix X ^l as the reference, according to the Euclidean distance between other sensors in the perception field and the reference sensor corresponding to the matrix row from small Select n-1 adjacent sensors in sequence, and the matrix row data in the data matrix X ^l corresponding to the n sensors finally selected is the n row data to be selected. In this example, n is selected as 10, where n is determined by the neural network receptive field. It should be noted that the selection of the number of sensors determines the size of each data sample, and the receptive field of the neural network is determined by the size and step size of the convolution kernel, which slides convolution on the data sample to obtain features.

The number of columns l of the cropped data matrix is determined by the time interval of the data collected by the sensor and the characteristic period of the data distribution. The specific calculation method is as follows:

Among them, T is the characteristic period of the regular distribution of sensor data flow, Δt is the time interval of sensor data collection, and λ is a positive rational number. Generally, an appropriate positive integer is selected according to the required training sample size.

In this example, λ is selected 8, and the number of data matrix columns l calculated by formula (3) is 2304, then the data matrix X ^l can be expressed as a two-dimensional matrix with a shape of 20 × 2304, and 10 adjacent sensor data are selected through formula (2) A drift-free data sample X ^b of shape 10×2304 is constructed.

SC: Use a non-stationary random walk process with trend items to construct a drift simulation matrix D ^l of size N×l, and select n matrices from D ^l according to the same position of selecting n matrix rows from X ^l to construct X ^b The row data is constructed as a drift amount matrix D ^b with a size of n×1, which is used as the second augmented data, wherein the second sensing data is a drift amount simulation matrix; the second augmented data is a drift amount simulation data sample, which can be specifically Expressed as:

where d is the drift of each sensor at different times.

Specifically, when the sensor drift process is constructed, the drift between different sensors is generally independent of each other. In this example, the sensor drift probability threshold is set, and a non-stationary random walk process with a trend term is used to simulate the drift process of the sensor data. The sensor data after simulation can be expressed as:

Among them, y _i,t is the drift-containing data of sensor i at time t after simulation, xi _, t is the drift-free data of sensor i at time t, d _i,t is the amount of drift generated by sensor i at time t through simulation . rand(0,1) is a random floating point number between 0 and 1, α is a floating point number and α∈(0,1) is the probability threshold of sensor drift, when rand(0,1)>α, sensor i If no drift occurs, the amount of drift generated is zero; otherwise, drift occurs;

Each sensor data adaptively selects one of a linear function, an exponential function, a square root function and a sine function as the trend item in the non-stationary random walk process with the trend item during the drift simulation, so as to generate the corresponding trend item. drift trend. The specific method of drift simulation is as follows:

In the linear drift trend, the drift of sensor i at time t can be expressed as:

In the exponential drift trend, the drift of sensor i at time t can be expressed as:

In the square root drift trend, the drift of sensor i at time t can be expressed as:

In the sinusoidal drift trend, the drift of sensor i at time t can be expressed as:

Among them, e is the maximum drift in each trend item, _{ui , t} is the amount of random walk data, and ri is the angular velocity parameter.

In some embodiments, the method for selecting the simulation parameters of each trend item drift is as follows:

其中，s为特征周期T内数据归一化后的标准差，Δt为传感器数据采样间隔。随机游走数据量u_i,t～iid(0,σ²)，其中

角速度参数r用于调整正弦周期，且

The value of the maximum drift e in each trend item should be determined by the characteristics of the collected data and the number of columns of the cropped data matrix, and is positively related to the number of columns in the data matrix, specifically:

Among them, s is the standard deviation of the normalized data in the characteristic period T, and Δt is the sampling interval of the sensor data. The random walk data volume u _i,t ～iid(0,σ ² ), where

The angular velocity parameter r is used to adjust the sine period, and

In this example, in order to ensure the sufficiency of drift samples, the drift probability threshold of each sensor in the same time period when the number of sampling points is 1 is set to 0.5, that is, the probability of drift in 10 sensors in the same data matrix is set to 0.5. is 50%. The four types of drift simulation methods can adjust the probability of various drift occurrences according to application scenarios. In this example, four types of sensor drift trends are used to construct four types of data sets. After calculating the maximum drift e~U(2.97, 5.94) in each trend item, σ~U(0.03, 0.06) in the random walk data u _i,t ~iid(0,σ ² ); angular velocity parameter r~U (2, 4).

In a possible implementation manner, a dataset is constructed from the first augmented data and the second augmented data, which is specifically as follows:

SD: Perform matrix addition of the drift-free data sample X ^b and the drift-amount simulation data sample D ^b to obtain the drift-containing data sample Y ^b .

Specifically, the drift-containing data sample obtained by performing matrix addition of the drift-free data sample X ^b and the drift simulation data sample D ^b should be:

Y ^b =X ^b +D ^b (11)

SE: Select multiple drift data samples X ^b and multiple drift data samples Y ^b to form a drift-free data set X ^M and a drift-containing data set Y ^M respectively; combine the drift-free data set X ^M and the drift-containing data set Y ^M respectively Each is divided into a part as a network training set and the other part as a network test set, and there is no overlap between the network training set and the network test set.

Specifically, a measurement matrix of 20×25920 is obtained from the perception field and 8000 data matrices X ^l of size 20×2304 are randomly cropped according to formula (1), and 10 adjacent sensors are selected to obtain 8000 data matrices of size 10×10 according to formula (2). The non-drift data sample X ^b of 2304 is used as the non-drift data set, and 8000 data samples Y ^b with drift of size 10×2304 are also obtained in this way as the data set with drift Y ^M . To sum up, in this example, during the construction of sensor data samples, the size of the training samples input for each iteration is determined according to the characteristic period and sampling interval of the data in the IoT data stream, and the non-stationary random walk process with trend items is used to control the sensor drift. Process simulation provides large-scale and high-quality basic data for sensor drift calibration, which can solve the problem of insufficient training samples when using deep learning methods for sensor calibration, and provide a guarantee for improving the quality of data collection.

Example 2

As shown in FIG. 5 , on the basis of Embodiment 1, this embodiment discloses a sensor drift calibration method, and the overall process is as follows:

acquiring sensor data to be calibrated, inputting the sensor data to be calibrated into the trained squeeze-excitation residual network model, and outputting corresponding drift calibration data;

Wherein, the squeeze-excitation residual network model is obtained by training multiple sets of data samples; the multiple sets of data samples are a data set constructed by the first augmented data and the second augmented data.

Specifically, referring to step SF, the training steps of the squeeze-excitation residual network model are as follows: adding a squeeze-excitation module to the one-dimensional convolutional residual network to make full use of the data correlation between adjacent sensors to achieve sensor drift calibration, As shown in Figure 6, the specific steps are as follows:

SF1: A one-dimensional convolutional layer is connected to the network input, and its output is used as the input of the residual module, and a one-dimensional convolutional layer connected to the output of the residual module is used to adjust the output result to be the same size as the input . The input of the residual module is connected to three one-dimensional convolutional layers. Each one-dimensional convolutional layer includes convolution operation, batch normalization and activation function mapping output. The output of the last one-dimensional convolutional layer contains two network branches. One of the network branches is connected to the squeeze-excitation module, and the other is multiplied with the output of the squeeze-excitation module. The squeeze-excitation module mainly includes a global pooling layer and two fully connected layers. The activation function used for the output of the first fully connected layer is Relu, and the output of the second fully connected layer needs to use the Sigmoid function as the activation function.

Among them, in this example, the network input sample is a data matrix with 10 rows, so it is input to the first convolutional layer in the sensor drift calibration network with 10 channels. The number of convolution kernels is 32, and the size of the convolution kernel is 1×5, the input of the residual module is connected to 3 one-dimensional convolutional layers, the number of convolution kernels in each one-dimensional convolutional layer is 32, and the size of the convolution kernel is 1×3, which is squeezed in the SE-ResNet model- The number of channels in the excitation module is halved after the first fully connected layer, activated by the Relu activation function, and the number of channels is restored to 32 by the next fully connected layer; the last one-dimensional convolutional layer sets the number of channels to 10 to keep the same as the initial input. Same size.

SF2: 4 types of drift-containing data sets and non-drift data sets are constructed by using 4 types of drift trends respectively, and 6400 drift-containing data samples in the drift-containing data set Y ^M are selected, that is, the first 80% are used as the training set of the neural network; the other 1600 The last 20% of the data samples are used as the neural network test set, and there is no overlap between the network training set and the network test set.

SF3: The drift-containing data sample Y ^b in the training set divided by the drift-containing data set Y ^M is iteratively input into the sensor drift calibration network for training, and the corresponding drift-free data sample X ^b is used as the output true value of the neural network, and the neural network loss Using the mean square error function to achieve the purpose of data drift compensation, the loss of the designed calibration module should be:

为输出的校准数据，X_i为对应的无漂移数据样本。Among them, m is the number of training samples,

is the output calibration data, and X _i is the corresponding drift-free data sample.

This example uses the Adam optimizer for iterative training, the Batch Size is set to 200, the number of iterations is 5000, and the network learning rate is set to 0.001 to obtain the trained squeeze-excitation residual network model.

SG: calibrate the drift sensor data, input the drift data samples of the test set as the data to be calibrated into the squeeze-excitation residual network model after training, and output the corresponding drift calibration data.

It should be noted that the convolutional neural network uses the feature learning ability of the convolution layer to extract the temporal and spatial features in the drift-containing data from the historical sensory data to realize the drift calibration of the sensor. However, the sensing data collected from a specific sensor network is limited, and the data containing drift is more scarce, which makes it difficult to ensure the demand of the neural network for training samples. The sensor data augmentation method proposed in the present invention can effectively solve the lack of neural network training. The sample problem is solved, and the feature learning and drift calibration steps are modeled as different network modules, and they are jointly trained with the augmented perceptual data set, which can ensure the extraction quality of drift features and reduce drift calibration errors.

This example uses the soil temperature perception data for 90 days from February 1, 2020 to April 30, 2020 as the first perception data. The sample data augmentation method proposed by the present invention satisfies the sample data required for neural network training, and the sample data obtained by the method of the present invention is not significantly different from the drift data in the real scene. Using sample data with different drift trends to train the neural network model, the experimental results show that the invention can effectively solve the data drift problem in the data flow of the Internet of Things, using the root mean square error as the evaluation standard, using the above 20 total data sets, respectively Test 20 times, and the average test results are shown in Figure 7. The overall drift calibration error is only about 1/2 of the traditional residual neural network-based calibration method.

Example 3

As shown in FIG. 2 , this embodiment provides a sensor data augmentation system, which includes:

The perception field building module 20 is used for deploying calibrated sensors and constructing a target perception field; the perception field building module also includes a proximity sensor data determination module, which adaptively determines the size of the number n of adjacent sensors according to the data augmentation application scenario, from all In the cropped data matrix X ^l , a matrix row is randomly selected as the reference, and n-1 adjacent sensors are selected in order from small to large according to the Euclidean distance between other sensors in the sensing field and the reference sensor corresponding to the matrix row, and the final selected n ^The n rows of data corresponding to the pieces of sensor data in the data matrix X1 are the n rows of adjacent sensor data to be selected.

The drift-free data sample building module 30 is used to collect the first perception data of the calibrated sensor in the target perception field, and use the random clipping method to clip the first perception data into a plurality of data matrices, and from a single clipped data In the matrix, according to the position of the sensor in the sensing field corresponding to each row in the matrix, some row data in the data matrix is selected to construct an adjacent sensor data matrix, and the first augmented data is determined.

Specifically, the first sensing data is modeled as a drift-free data matrix, wherein multiple matrix rows in the drift-free data matrix correspond to data collected by multiple sensors in the sensing field respectively; the drift-free data matrix X is cut into A plurality of data matrices, the number of columns ^l of the data matrix X1 is less than the number of columns L of the drift-free data matrix; from a single cropped data matrix, according to the position of the sensor corresponding to each matrix row in the sensing field, from Selecting part of the row data in the data matrix to construct a neighboring sensor data matrix, and determining the first augmented data, the first augmented data is a drift-free data sample; the drift-free data sample building module also includes a trimming data matrix column number calculation module , which is used to determine the number of columns l of the cropped data matrix according to the time interval of the data collected by the sensor and the characteristic period of the data distribution. The calculation method is:

The drift amount simulation data sample building module 40 is used to simulate the drift process of the sensor data by using a non-stationary random walk process containing trend items according to the data matrix tailored from the first sensing data, and obtain the second sensing data, from The drift amount matrix with the same position and size as the first augmented data is selected from the second sensing data, and determined as the second augmented data, and the second augmented data is the drift amount simulation data sample.

Example 4

This embodiment provides an exemplary computer program product and computer readable storage medium

In addition to the methods and apparatuses described above, embodiments of the present application may also be computer program products comprising computer program instructions that, when executed by a processor, cause the processor to perform the "exemplary methods" described above in this specification The steps in the decision-making method according to various embodiments of the present application described in the section.

The computer program product can write program codes for performing the operations of the embodiments of the present application in any combination of one or more programming languages, including object-oriented programming languages, such as Java, C++, etc. , also includes conventional procedural programming languages, such as "C" language or similar programming languages. The program code may execute entirely on the user computing device, partly on the user device, as a stand-alone software package, partly on the user computing device and partly on a remote computing device, or entirely on the remote computing device or server execute on.

In addition, embodiments of the present application may also be computer-readable storage media having computer program instructions stored thereon, the computer program instructions, when executed by a processor, cause the processor to perform the above-mentioned "Example Method" section of this specification Steps in decision-making methods according to various embodiments of the present application described in .

The computer-readable storage medium may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may include, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses or devices, or a combination of any of the above. More specific examples (non-exhaustive list) of readable storage media include: electrical connections with one or more wires, portable disks, hard disks, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, portable compact disk read only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

The basic principles of the present application have been described above in conjunction with specific embodiments. However, it should be pointed out that the advantages, advantages, effects, etc. mentioned in the present application are only examples rather than limitations, and these advantages, advantages, effects, etc., are not considered to be Required for each embodiment of this application. In addition, the specific details disclosed above are only for the purpose of example and easy understanding, rather than limiting, and the above-mentioned details do not limit the application to be implemented by using the above-mentioned specific details.

The block diagrams of devices, apparatus, apparatuses, and systems referred to in this application are merely illustrative examples and are not intended to require or imply that the connections, arrangements, or configurations must be in the manner shown in the block diagrams. As those skilled in the art will appreciate, these means, apparatuses, apparatuses, systems may be connected, arranged, configured in any manner. Words such as "including", "including", "having" and the like are open-ended words meaning "including but not limited to" and are used interchangeably therewith. As used herein, the words "or" and "and" refer to and are used interchangeably with the word "and/or" unless the context clearly dictates otherwise. As used herein, the word "such as" refers to and is used interchangeably with the phrase "such as but not limited to".

It should also be pointed out that in the apparatus, equipment and method of the present application, each component or each step can be decomposed and/or recombined. These disaggregations and/or recombinations should be considered as equivalents of the present application.

The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other aspects without departing from the scope of the application. Therefore, this application is not intended to be limited to the aspects shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

The foregoing description has been presented for the purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of the application to the forms disclosed herein. Although a number of example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims (8)

1. A method of sensor data augmentation, the method comprising:

and SA: deploying the calibrated sensor, and constructing a target sensing field;

SB: acquiring first sensing data of calibrated sensors in a target sensing field, cutting the first sensing data into a plurality of data matrixes by using a random cutting method, selecting partial row data in the data matrixes from the single cut data matrixes according to the positions of the sensors corresponding to all rows in the matrixes in the sensing field to construct data matrixes of adjacent sensors, and determining first augmented data;

SC: according to the data matrix cut by the first sensing data, performing drift process simulation on the sensor data by using a non-stationary random walk process containing a trend item to obtain second sensing data, selecting a drift amount matrix with the same position and size as the first augmented data from the second sensing data, and determining the second augmented data;

the determining first augmented data comprises:

determining first sensing data according to continuous L sampling points acquired by calibrated N sensors in a target sensing field, and recording the first sensing data as a non-drifting data matrix X with the size of NxL, wherein N matrix row data in the non-drifting data matrix X respectively correspond to data acquired by the N sensors in the sensing field;

randomly cutting a plurality of data matrixes with the size of NxL from the drift-free data matrix X, and recording any one of the cut data matrixes as X ^l Said data matrix X ^l Is smaller than the number of columns L of the drift-free data matrix X, from a single clipped data matrix X ^l From the data matrix X in dependence on the position in the sensing field of the sensor corresponding to each matrix row ^l Selecting n rows of data to construct n multiplied by l adjacent sensor numberAccording to matrix X ^b And recorded as the first augmented data.

2. The method of claim 1, wherein the specific formula for determining the first augmented data is:

wherein

The initial cutting position of the sampling point is defined as l, the column number of a cutting data matrix is defined as the number of continuous sampling points of the sensor corresponding to each matrix row; 1,

to express taking the matrix

To

Column data; n is a secondary cut data matrix X ^l The selected number of rows.

3. The sensor data augmentation method of claim 2, wherein the n-row data selection method comprises:

adaptively determining the size of the number n of proximity sensors from the data matrix X clipped according to the data augmentation application scenario ^l In the method, a matrix row is randomly selected as a reference, and other sensors in a sensing field and reference sensors corresponding to the matrix row are arranged between the other sensorsSequentially selecting n-1 adjacent sensors from small to large according to Euclidean distance, and finally selecting data matrix X corresponding to the n sensors ^l The matrix row data in (1) is the n row data to be selected.

4. The method of claim 2, wherein the number of columns/of the clipped data matrix is determined by the characteristic period of the data distribution and the time interval of the data acquisition of the sensor, and is calculated by:

wherein, T is a characteristic period of regular distribution of sensor data flow, delta T is a time interval of data acquisition of the sensor, lambda is a positive rational number, and a proper positive integer is generally selected according to the size of a required training sample.

5. The method according to claim 2, wherein the second sensing data is a drift amount simulation matrix D with size of NxL constructed by a non-stationary random walk process with trend terms ^l According to from X ^l Selecting n rows of data to construct X ^b Method from D ^l Constructing a drift amount matrix D with the size of nxl from n rows of data with the same row position ^b And recording as second augmentation data;

setting a sensor drift probability threshold, and performing drift process simulation on sensor data by adopting a non-stationary random walk process containing a trend item, wherein the simulated sensor data can be expressed as follows:

wherein, y _i,t For the drift-containing data, x, of the sensor i at time t after simulation _i,t For drift-free data of sensor i at time t, d _i,t For sensor i, at time t by simulationThe amount of drift produced; rand (0, 1) is a random floating point number between 0 and 1, alpha is a floating point number and alpha epsilon (0, 1) is a probability threshold value of the sensor drifting when rand (0, 1)>When alpha is reached, the sensor i does not drift, and the generated drift amount is zero; otherwise, drift occurs;

and when drift amount simulation is carried out on the data of each sensor, one of a linear function, an exponential function, a square root function and a sine function is selected in a self-adaptive mode to serve as a trend item in the non-stationary random walk process containing the trend item, and the trend item is used for generating a corresponding drift trend.

6. A sensor drift calibration method is characterized by comprising the following steps:

acquiring sensor data to be calibrated, inputting the sensor data to be calibrated into a trained extrusion-excitation residual error network model, and outputting corresponding drift calibration data;

wherein the squeeze-excitation residual error network model is obtained by training a plurality of groups of data samples; the plurality of sets of data samples are data sets constructed from first augmented data and second augmented data;

the step of constructing a data set from the first augmented data and the second augmented data comprises:

taking the first augmented data as a drift-free data sample X ^b Taking the second augmentation data as a drift amount simulation data sample D ^b ；

The drift-free data sample X ^b And drift amount simulation data sample D ^b Performing matrix addition to obtain drift-containing data samples Y ^b (ii) a Using a plurality of drift-free data samples X ^b Composing a drift-free data set X ^M (ii) a Using a plurality of drift-containing data samples Y ^b Forming a drift-containing data set Y ^M (ii) a Will contain the drift data set Y ^M And drift-free data set X ^M As a training data set for the squeeze-excitation residual network model.

7. A sensor data augmentation system, comprising:

the perception field construction module is used for deploying the calibrated sensor and constructing a target perception field;

the drift-free data sample construction module is used for collecting first sensing data of a calibrated sensor in a target sensing field, cutting the first sensing data into a plurality of data matrixes by using a random cutting method, selecting part of row data in the data matrixes from the single cut data matrixes according to the positions of the sensors corresponding to all rows in the matrixes in the sensing field to construct adjacent sensor data matrixes, and determining first augmented data, wherein the first augmented data are drift-free data samples;

the drift amount simulation data sample construction module is used for simulating the drift process of the sensor data by utilizing a non-stationary random walk process containing a trend item according to a data matrix cut by the first sensing data to obtain second sensing data, selecting a drift amount matrix with the same position and size as the first augmented data from the second sensing data, and determining the drift amount matrix as second augmented data, wherein the second augmented data is a drift amount simulation data sample;

the determining first augmented data comprises:

determining first sensing data according to continuous L sampling points acquired by N calibrated sensors in a target sensing field, and recording the first sensing data as a non-drifting data matrix X with the size of NxL, wherein N matrix row data in the non-drifting data matrix X respectively correspond to data acquired by the N sensors in the sensing field;

randomly cutting a plurality of data matrixes with the size of NxL from the drift-free data matrix X, and recording any one of the cut data matrixes as X ^l Said data matrix X ^l Is smaller than the number of columns L of the drift-free data matrix X, from a single clipped data matrix X ^l From the data matrix X in dependence on the position in the sensing field of the sensor corresponding to each matrix row ^l N rows of data are selected to construct an nxl-sized data matrix X of the adjacent sensors ^b And recorded as first augmented data.

8. A readable storage medium, characterized in that the storage medium stores a computer program comprising program instructions which, when executed by a processor, cause the processor to carry out the method according to any one of claims 1-6.

Images