Sensor data super-resolution reconstruction method based on long and short time sequence characteristic assimilation

Abstract

The invention discloses a sensor data super-resolution reconstruction method based on long and short time sequence characteristic assimilation, and relates to a sensor data super-resolution reconstruction method. According to the invention, the problem of standardized processing of irregular sampling interval data is effectively solved. Secondly, a double-engine super-resolution reconstruction framework is innovatively constructed: on one hand, a convolution-deconvolution neural network is utilized to mine long-time scale periodic features, and a deep feature representation capability is obtained through historical data training; and on the other hand, a Shenchang differential equation is innovatively introduced, a continuous time dynamic state is modeled from a micro gradient difference of data change, and short-time transient characteristics are accurately captured. On the basis, weight fusion is carried out on double-engine output through a particle filtering algorithm, and assimilation of long-time and short-time data features is achieved. According to the invention, through cooperative mining and fusion of spatial-temporal characteristics, the precision and robustness of super-resolution reconstruction are significantly improved. The invention belongs to the technical field of sensor data super-resolution reconstruction.

Description

Technical Field

[0001] The invention relates to a sensor data super-resolution reconstruction method, belonging to the technical field of sensor data super-resolution reconstruction. Background Art

[0002] As the sole source of information for monitoring a system's operating status, sensor quality directly determines system control accuracy and operational reliability. Precision sensor data with high temporal resolution not only effectively captures subtle transient changes in parameters but also provides a critical basis for real-time decision-making and early fault diagnosis. However, in complex application scenarios, due to performance degradation caused by long-term sensor operation, sampling timing differences between multi-source heterogeneous systems, processing delays caused by computing resource constraints, and limited data transmission bandwidth, observed data exhibits characteristics such as disorder, asynchrony, and low temporal resolution in the time domain. These data quality issues not only affect the real-time state perception but also restrict the accuracy of the system's dynamic response and, in extreme cases, may even lead to system failure. Therefore, overcoming the temporal resolution limitations of existing sensor data and achieving high-precision super-resolution reconstruction of data has become a key technical bottleneck for improving system operation and control performance.

[0003] Defects and shortcomings of existing technology:

[0004] Sensor data super-resolution reconstruction methods typically build a mapping model from low-resolution sparse observations to high-resolution data based on historical data. This technology has been widely used in fields such as image processing and fluid monitoring. However, in industrial applications, sensor data exhibits significant pseudo-periodicity and randomness due to factors such as environmental disturbances and system maneuvers. In addition, factors such as extreme environmental conditions, sensor performance degradation, and lifespan limitations can lead to anomalies such as missing data and redundant sampling. Specific challenges include:

[0005] First, sensor data exhibits complex dynamic periodicity over long timescales, influenced by factors such as environmental perturbations and system dynamics. While exhibiting periodic characteristics, the duration and amplitude of different cycles vary randomly. Existing methods typically employ data-driven models operating on a single timescale (e.g., pure convolutional neural networks). This makes it difficult to simultaneously capture both the cyclical evolution over long timescales and the transient variations over short timescales, limiting super-resolution accuracy.

[0006] Second, due to factors such as extreme environments and sensor performance limitations, sensor data is commonly subject to missing and redundant sampling, resulting in significant timestamp drift and degraded data quality. Most existing methods fail to fully consider the impact of timestamp drift on modeling. Directly using this irregularly sampled data for model construction will introduce phase deviation and noise interference, thereby reducing the model's generalization performance and prediction accuracy. Summary of the Invention

[0007] In order to solve the problem that in industrial applications, sensor data exhibits significant pseudo-periodicity and randomness due to factors such as environmental disturbances and system maneuvers, and factors such as extreme environmental conditions, sensor performance degradation and life limitations can also lead to abnormal situations such as data missing and redundant sampling, the present invention proposes a sensor data super-resolution reconstruction method based on the assimilation of long and short time series features.

[0008] The technical solution adopted by the present invention to solve the above problems is: the steps of the present invention include:

[0009] Step 1: Get the system sensor data ω=[(t1,ω1),(t2,ω2),…,(t N ,ω N )], where N represents the total number of sampling points of sensor data, and the fixed sampling interval of the sensor is ΔT;

[0010] Step 2: Use the sliding window method to scan the acquired sensor data ω at time t, and record the starting time d when the sensor sampling point time interval is not ΔT i , and the duration of the time interval l i =β×ΔT,β∈N * , and based on this, the sensor data interval adaptive adjustment window W=[W1,W2,W3,...,W M ], Where M represents the number of sensor data interval adaptive adjustment windows;

[0011] Step 3: Use the linear interpolation method to adjust the sampling interval of the sensor data in each window and obtain the regularly sampled sensor data ω′=[ω1′,ω2′,…,ω′ N ];

[0012] Step 4: Assume that the regularly sampled sensor data ω′ is high-resolution data, and assume that the super-resolution factor of the data is α, downsample the sensor data to generate low-resolution sensor data ω″=[ω1″,ω2″,…,ω″ M ], After different downsampling multiples, data features are compared;

[0013] Step 5: Divide the data ω′ and ω″ into training set ω′ train ,ω″ train and the test set ω′ test ,ω″ test , that is, ω″ train =[ω″1,ω″2,…,ω″ M×0.8 ],ω′ train =[ω1′,ω2′,…,ω′N×0.8 ],ω″ test =[ω″ M×0.8+1 ,ω″ M×0.8+2 ,…,ω″ M ],ω′ test =[ω′ M×0.8+1 ,ω′ M×0.8+2 ,…,ω′ N ];

[0014] Step 6: Based on step 5, construct a training set of super-resolution reconstruction model, where ω″ train and ω″ test is the input of the model, ω′ train and ω′ test The label value output by the model is obtained by using ω″ train and ω′ train Training sensor data super-resolution reconstruction model;

[0015] Step 7: Establish a long-time super-resolution reconstruction model for sensor data based on CNN-DeCNN;

[0016] Step 8. Repeat steps 4 to 6 m times and obtain the long-time super-resolution reconstruction result of sensor data based on CNN-DeCNN:

[0017] Step 9: Establish a short-time-scale super-resolution reconstruction model for sensor data based on NODE. The formula is:

[0018]

[0019] Step 10: Use the numerical solution method to solve formula (7) and obtain the short-time-scale super-resolution reconstruction result of sensor data based on NODE: As shown in formula (8),

[0020]

[0021] Step 11: Assume that there are γ particles, each particle represents a state of the satellite sensor, which is represented by ω CNN Determine and initialize particles and particle weights

[0022] Step 12: Assuming that the sensor's changing state is smooth and affected by noise, describe the sensor's state from the previous time step. Transition to the state of the next time step

[0023] Step 13: Use ω NODEUpdate the particle weights and obtain the sensor observations;

[0024] Step 14: Calculate the When , the observation likelihood function of the sensor is as shown in formula (13), and the particle weight is updated and normalized, the formula is:

[0025]

[0026] Step 15: Repeat formulas (13) and (14), and use formula (15) to reconstruct the high-resolution sensor data of the long-short time feature fusion model:

[0027]

[0028] Step 16: Design the loss function of the model as shown in formula (15), and use the parameter optimization method to optimize the loss function to obtain the optimal sensor super-resolution reconstruction model for long- and short-time feature assimilation;

[0029] Step 17: Test data set ω′ test Inject it into the sensor data super-resolution reconstruction model based on long- and short-time feature assimilation, and obtain the sensor data super-resolution reconstruction result

[0030] Step 18: Use the mean square error, such as formula (17), and the root mean square error, such as formula (18), to evaluate the accuracy of the sensor data super-resolution reconstruction model;

[0031]

[0032] Furthermore, in step 2, W i The calculation formula is:

[0033]

[0034] Furthermore, the formula used in step 3 using the linear interpolation method is:

[0035]

[0036] Furthermore, the formula for downsampling the sensor data in step 4 is:

[0037] ω k ″=ω′ k×α ,

[0038] Furthermore, in step 7, formula (4) is used to extract the features of the data, and formula (5) is used to improve the resolution of the data;

[0039]

[0040] In formula (4), represents the output of the i-th layer CNN, w conv (j) represents the weight of the convolution kernel, p represents data padding, k represents the size of the convolution kernel, b i represents the deviation term;

[0041]

[0042] In formula (5), represents the super-resolution reconstruction result of the i-th layer DeCNN, s i represents the stride of the i-th layer DeCNN, p i represents the output padding of the i-th layer DeCNN, represents the convolution kernel weight of the i-th layer DeCNN, b i represents the bias of the i-th layer DeCNN, k represents the convolution kernel size, Represents the upsampling matrix of the i-th layer CNN output, and the calculation formula is:

[0043]

[0044] Furthermore, in step 11, the particles are initialized and particle weights The formula is:

[0045]

[0046] In formulas (9) and (10), Represents ω CNN The mean of Represents ω CNN The variance of .

[0047] Furthermore, step 12 describes the state of the sensor from the previous time step Transition to the state of the next time step The formula is:

[0048]

[0049] In formula (11), represents state noise.

[0050] Furthermore, in step 13, ω is used NODE The formula for updating the particle weight and obtaining the sensor's observation value is:

[0051]

[0052] In formula (12), rk represents the observation noise.

[0053] The beneficial effects of the present invention are:

[0054] 1. This paper proposes an improved adaptive sliding window linear interpolation algorithm. By dynamically identifying the location and duration of data loss and combining it with intelligent adjustment of sliding window parameters, this method effectively addresses the issues of sensor data redundancy and loss, thereby improving the modeling accuracy of the sensor data super-resolution reconstruction model.

[0055] 2. This invention innovatively designs a dual-engine super-resolution reconstruction framework that collaborates long-term and short-term features. The CNN-DeCNN module focuses on extracting long-term periodic features from sensor data, while the NODE module accurately captures transient short-term variations. This collaborative work effectively reduces interference from environmental disturbances, system dynamics, and other factors that can cause periodic randomness in data, thereby improving the accuracy of super-resolution reconstruction of sensor data.

[0056] 3. The particle filter-based method of the present invention fuses the output results of the CNN-DeCNN model and the NODE model, thereby effectively integrating the nonlinear short-term changes in the NODE with the long-term periodic changes of the CNN-DeCNN, which can further improve the accuracy of the sensor super-resolution reconstruction;

[0057] 4. The present invention uses actual observation data from the flywheel sensor of the satellite attitude and orbit control system to verify the sensor data super-resolution reconstruction method based on the assimilation of long and short time series features proposed in the present invention, and achieves high-precision super-resolution reconstruction of sensor data (super-resolution factor α = 4), MSE = 0.47%, RMSE = 6.86%. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is the overall scheme diagram of the present invention;

[0059] Figure 2 It is a comparison diagram of sensor data at different resolutions in the present invention;

[0060] Figure 3 It is a comparison diagram of the method proposed in the present invention and different methods. DETAILED DESCRIPTION

[0061] Example 1: Figures 1 to 3 As shown in FIG, a sensor data super-resolution reconstruction method based on long and short time series feature assimilation includes the following steps:

[0062] Step 1: Taking the flywheel sensor of the satellite attitude and orbit control system as an example, the long-term and short-term characteristic changes of the actual observation data of the sensor are studied, and the effectiveness of this method in the field of super-resolution reconstruction is verified. Among them, the data collected by the flywheel sensor of the attitude and orbit control system contains N = 20,000 sampling points, and the fixed sampling interval of the sensor is ΔT = 1s. Use formula (1) to obtain the sensor data interval adaptive adjustment window W. Based on this, use formula (2) to obtain the sensor regularly sampled data ω′ = [ω1′, ω2′,…, ω2′ 0000 ]

[0063] Step 2: Assume that the super-resolution factor of the data is α = 4, and use formula (3) to downsample the sensor data to generate low-resolution sensor data ω″ = [ω1″, ω2″, …, ω″ 5000 ], data comparison before and after downsampling

[0064] Step 3: Divide the data ω′ and ω″ into training set ω′ train ,ω″ train (80% of the data) and the test set ω′ test ,ω″ test (20% of the data), i.e. ω″ train =[ω1″,ω2″,…,ω″ 4000 ],ω′ train =[ω1′,ω2′,…,ω1′ 6000 ],ω″ test =[ω″ 4001 ,ω″ 4002 ,…,ω″ 5000 ],ω′ test =[ω′ 16001 ,ω′ 16002 ,…,ω′ 20000 ].

[0065] Step 4: On this basis, construct a training set of super-resolution reconstruction model, where ω″ train and ω″ test is the input of the model, ω″ test and ω′ test is the label value output by the model. First, use ω″ train and ω′ train Train a super-resolution reconstruction model for satellite sensors.

[0066] Step 5: Establish a long-time super-resolution reconstruction model for satellite sensor data based on CNN-DeCNN. First, use formula (4) to extract the features of the data, and use formula (5) to improve the resolution of the data, where m = 3.

[0067] Step 6: Repeat formulas (4) to (6) m = 3 times, and obtain the long-time super-resolution reconstruction result of satellite sensor data based on CNN-DeCNN:

[0068] Step 7: Establish a short-time-scale super-resolution reconstruction model for satellite sensor data based on NODE, as shown in formula (7).

[0069] Step 8: Use the fourth-order Runge-Kutta method to solve (6) and obtain the short-time-scale super-resolution reconstruction result of satellite sensor data based on NODE: As shown in formula (8).

[0070] Step 9: Assume that there are γ = 50,000 particles, each particle represents a state of the satellite sensor, which is represented by ω CNN Determine and use (9) and (10) to initialize particles and particle weights

[0071] Step 10: Assuming that the sensor's changing state is smooth and affected by noise, use formula (11) to describe the sensor's state from the previous time step: Transition to the state of the next time step represents state noise.

[0072] Step 11: Use ω NODE Update the particle weights and obtain the sensor observations, as shown in formula (12).

[0073] Step 12: Calculate the When , the observation likelihood function of the sensor data is as shown in formula (13), and the particle weights are updated and normalized according to formula (11), as shown in formula (14).

[0074] Step 13: Repeat (13) and (14), and use formula (15) to obtain high-resolution satellite sensor data for the long- and short-time feature fusion model:

[0075] Step 14: Design the loss function of the model as shown in formula (19), and use the parameter optimization method to optimize the loss function, and obtain the optimal satellite sensor super-resolution reconstruction model for long- and short-time feature assimilation.

[0076]

[0077] Step 15: Test data set ω′ testInject it into the satellite sensor super-resolution reconstruction model based on long- and short-time feature assimilation, and obtain the satellite sensor super-resolution reconstruction result

[0078] Step 16: Use the mean square error (MSE) as shown in formula (20) and the root mean square error (RMSE) as shown in formula (21) to evaluate the data recovery accuracy of the satellite sensor.

[0079]

[0080] Step 17. Finally, the method proposed in the present invention is compared with the currently commonly used data super-resolution reconstruction methods (Variational Auto-Encoders (VAE), Wasserstein Generative Adversarial Networks (WGAN), Cubic Spline Interpolation (CSI), Nearest Neighbor Interpolation (NNI) and Linear Interpolation (LI)), and the accuracy of the model is calculated using formulas (20) and (21), as shown in Table 1 and Figure 3 As shown in the figure, the method proposed by the present invention has the highest accuracy.

[0081] Table 1 Comparison of method accuracy

[0082] method MSE (%) RMSE (%) LI 1.37 11.7 NNI 1.98 14.1 CSI 1.70 13.1 WGAN 11.8 34.3 VAE 6.39 25.3 The method proposed in this paper 0.47 6.81

[0083] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the present profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical content disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement of the above embodiments made according to the technical essence of the present invention, within the spirit and principles of the present invention, without departing from the content of the technical solution of the present invention, shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A sensor data super-resolution reconstruction method based on long and short time series feature assimilation, characterized by: The specific steps include: Step 1: Get the system sensor data ω=[(t1,ω1),(t2,ω2),…,(t N ,ω N )], where N represents the total number of sampling points of sensor data, and the fixed sampling interval of the sensor is ΔT; Step 2: Use the sliding window method to scan the acquired sensor data ω at time t, and record the starting time d when the sensor sampling point time interval is not ΔT i , and the duration of the time interval l i =β×ΔT,β∈N * , and based on this, the sensor data interval adaptive adjustment window W=[W1,W2,W3,...,W M ], Where M represents the number of sensor data interval adaptive adjustment windows; Step 3: Use the linear interpolation method to adjust the sampling interval of the sensor data in each window and obtain the regularly sampled sensor data ω′=[ω′1,ω′2,…,ω′ N ]; Step 4: Assume that the regularly sampled sensor data ω′ is high-resolution data, and assume that the super-resolution factor of the data is α, downsample the sensor data to generate low-resolution sensor data ω″=[ω″1,ω″2,…,ω″ M ], After different downsampling multiples, data features are compared; Step 5: Divide the data ω′ and ω″ into training set ω′ train ,ω″ train With the test set ω′ test ,ω″ test , that is, ω″ train =[ω″1,ω″2,…,ω″ M×0.8 ],ω′ train =[ω′1,ω′2,…,ω′ N×0.8 ],ω″ test =[ω″ M×0.8+1 ,ω″ M×0.8+2 ,…,ω″ M ],ω′ test =[ω′ M×0.8+1 ,ω′ M×0.8+2 ,…,ω′ N ]; Step 6: Based on step 5, construct a training set of super-resolution reconstruction model, where ω″ train and ω″ test is the input of the model, ω′ train and ω′ test The label value output by the model is calculated using ω″ train and ω′ train Training sensor data super-resolution reconstruction model; Step 7: Establish a long-time super-resolution reconstruction model for sensor data based on CNN-DeCNN; Step 8. Repeat steps 4 to 6 m times and obtain the long-time super-resolution reconstruction result of sensor data based on CNN-DeCNN: Step 9: Establish a short-time-scale super-resolution reconstruction model for sensor data based on NODE. The formula is: Step 10: Use the numerical solution method to solve formula (7) and obtain the short-time-scale super-resolution reconstruction result of sensor data based on NODE: As shown in formula (8), Step 11: Assume that there are γ particles, each particle represents a state of the satellite sensor, which is represented by ω CNN Determine and initialize particles and particle weights Step 12: Assuming that the sensor's changing state is smooth and affected by noise, describe the sensor's state from the previous time step. Transition to the state of the next time step Step 13: Use ω NODE Update the particle weights and obtain the sensor observations; Step 14: Calculate the When , the observation likelihood function of the sensor is as shown in formula (13), and the particle weight is updated and normalized, the formula is: Step 15: Repeat formulas (13) and (14), and use formula (15) to reconstruct the high-resolution sensor data of the long-short time feature fusion model: Step 16: Design the loss function of the model as shown in formula (15), and use the parameter optimization method to optimize the loss function to obtain the optimal sensor super-resolution reconstruction model for long- and short-time feature assimilation; Step 17: Test data set ω′ test Inject it into the sensor data super-resolution reconstruction model based on long- and short-time feature assimilation, and obtain the sensor data super-resolution reconstruction result Step 18: Use the mean square error, such as formula (17), and the root mean square error, such as formula (18), to evaluate the accuracy of the sensor data super-resolution reconstruction model; 2. The sensor data super-resolution reconstruction method based on long and short time series feature assimilation according to claim 1 is characterized in that: In step 2, W i The calculation formula is:

3. The sensor data super-resolution reconstruction method based on long and short time series feature assimilation according to claim 1 is characterized in that: The formula used in step 3 using the linear interpolation method is:

4. The sensor data super-resolution reconstruction method based on long and short time series feature assimilation according to claim 1 is characterized in that: The formula for downsampling the sensor data in step 4 is:

5. The sensor data super-resolution reconstruction method based on long and short time series feature assimilation according to claim 1 is characterized in that: In step 7, formula (4) is used to extract the features of the data, and formula (5) is used to improve the resolution of the data; In formula (4), represents the output of the i-th layer CNN, w conv (j) represents the weight of the convolution kernel, p represents data padding, k represents the size of the convolution kernel, b i represents the deviation term; In formula (5), represents the super-resolution reconstruction result of the i-th layer DeCNN, s i represents the stride of the i-th layer DeCNN, p i represents the output padding of the i-th layer DeCNN, represents the convolution kernel weight of the i-th layer DeCNN, b i represents the bias of the i-th layer DeCNN, k represents the convolution kernel size, Represents the upsampling matrix of the i-th layer CNN output, and the calculation formula is:

6. The sensor data super-resolution reconstruction method based on long and short time series feature assimilation according to claim 1 is characterized in that: Initialize particles in step 11 and particle weights The formula is: In formulas (9) and (10), Represents ω CNN The mean of Represents ω CNN The variance of .

7. The sensor data super-resolution reconstruction method based on long and short time series feature assimilation according to claim 1 is characterized in that: Step 12 describes the sensor state from the previous time step Transition to the state of the next time step The formula is: In formula (11), represents state noise.

8. The sensor data super-resolution reconstruction method based on long and short time series feature assimilation according to claim 1 is characterized in that: In step 13, we use ω NODE The formula for updating the particle weight and obtaining the sensor's observation value is: In formula (12), r k represents the observation noise.

Images