Adaptive wireless sensor network optimization method based on compressed sensing

Abstract

The invention relates to a compressed sensing-based adaptive wireless sensor network optimization method, which comprises the following steps of: constructing a sensor data transmission tree, and collecting historical data of a sensor network; constructing a sparse dictionary based on the collected historical data of the sensor network; and performing compression processing on the real-time sensor network data by using the sparse dictionary, and performing recovery of the sensor network data in the sink node. According to the compressed sensing algorithm based on adaptive dictionary learning, a mechanism is designed to meet the adaptive requirement of the compressed sensing algorithm in a large-scale sensor network, and the limitation of sparse base selection is overcome to a certain extent.

Description

Technical Field

[0001] The invention relates to the technical field of signal processing, and in particular to an optimization method for an adaptive wireless sensor network based on compressed sensing. Background Art

[0002] The Internet of Things (IoT) is a network built on the Internet information infrastructure and traditional telecommunications networks, aiming to achieve interconnection between all common physical objects with independent addressing capabilities. At present, the Internet of Things has penetrated into many households. However, the development of large-scale wireless sensor networks faces several bottlenecks, especially due to the performance limitations of IoT devices, which greatly limits the data transmission capacity. In order to address these challenges, many researchers have implemented a compressed sensing algorithm based on adaptive dictionary learning to achieve efficient data transmission in sensor networks. Compressed sensing (CS) is a technology that uses random subsampling to collect signal data and reconstruct low-loss signals at a low sampling rate. Martine Cilia and Keisuke Yamamoto used the monitoring and fault diagnosis of rotating machinery to experimentally demonstrate the effective application of the method and confirm the powerful performance of the K-SVD algorithm in dictionary learning. Similarly, Fu Jun, Yuan Haiguo et al. proposed an adaptive multi-layer dictionary learning image reconstruction method based on compressed sensing, and verified the feasibility of the K-SVD algorithm.

[0003] However, their research did not address the application of the technology in large-scale sensor networks. Most current IoT devices, especially those in sensor networks, rely on embedded systems. Although these devices are inexpensive and energy-efficient, their performance is relatively weak, especially in terms of bandwidth. In large-scale sensor networks, bandwidth resources are often limited, making efficient data transmission a key challenge. Moreover, compressed sensing faces the problem of low adaptability in sensor networks because it is difficult to find suitable parameters to compress specific signals. Therefore, it is urgent to design a mechanism to meet the adaptability requirements of compressed sensing algorithms in large-scale sensor networks. Summary of the invention

[0004] The purpose of the present invention is to propose an optimization method for an adaptive wireless sensor network based on compressed sensing to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] An optimization method for an adaptive wireless sensor network based on compressed sensing, comprising:

[0007] Build a sensor data transmission tree and collect sensor network historical data;

[0008] Building a sparse dictionary based on the collected historical data of the sensor network;

[0009] The sparse dictionary is used to compress the real-time sensor network data, and the sensor network data is restored in the sink node.

[0010] Optionally, the optimization method further includes:

[0011] Perform exception processing on compressed sensor network data; including:

[0012] When the sparse coefficients generated by the sensor network data of a sink node after passing through the sparse dictionary do not meet the requirements of compressed sensing, the current round of transmission is terminated, and the sparse dictionary is retrained according to the data in the preset previous time period;

[0013] The requirement of compressed sensing is that the sparsity exceeds the value required by compressed sensing theory.

[0014] Optionally, constructing the sensor data transmission tree includes:

[0015] Use CDG method to connect sensor network and sink node to obtain sensor network forest, in which the node directly connected to sink node is set as root node;

[0016] Data is defined for each tree in the sensor network forest to obtain the sensor data transmission tree.

[0017] Optionally, for each tree in the sensor network forest, defining data includes:

[0018] The nodes are numbered according to the order of breadth-first search, where the data of the i-th node is d i ;

[0019] Set the observation matrix φ. In the process of transmitting data from the leaf node to the root node, each node is multiplied by the observation matrix for aggregation to obtain the vector y. In each transmission, only a temporary y matrix is ​​transmitted, and finally added, aggregated at the root node, and restored at the sink node.

[0020] Among them, the vector y is:

[0021]

[0022] Where N represents the signal dimension, y i represents the compressed data, j is the process variable representing the sum of matrix multiplication, represents the element in row i and column j in the observation matrix, d jRepresents the data of the i-th node.

[0023] Optionally, constructing a sparse dictionary based on the collected sensor network data includes:

[0024] Defining a problem-solving goal for the collected sensor network data;

[0025] Based on the problem-solving goal, a linear programming algorithm is used to solve the problem and obtain a sparse dictionary.

[0026] Optionally, using the sparse dictionary to compress the real-time sensor network data includes:

[0027] Use k-means algorithm to cluster real-time sensor network data;

[0028] Based on the clustered data, the omp algorithm is used to obtain the sparse coefficients of the dictionary and update the residual;

[0029] For the updated residual, the k-svd algorithm is used to extract features and obtain the adapted dictionary.

[0030] Optionally, clustering the real-time sensor network data using a k-means algorithm includes:

[0031] Randomly select k points in the real-time sensor network data as cluster centers;

[0032] Assign each sensor data point to the cluster center that is closest to it;

[0033] After all sensor data points are assigned, the cluster center is updated based on the preset update formula until the cluster center no longer changes;

[0034] Among them, the preset update formula is:

[0035]

[0036] In the formula, x represents the data point, c i represents the i-th cluster center, |S i | is the number of data points in the cluster.

[0037] Optionally, the omp algorithm is used to obtain the sparse coefficients of the dictionary, and the updating residual includes:

[0038] Define initial residual and atom sets;

[0039] Select the data most relevant to the current residual from the clustered data and add it to the atomic set;

[0040] Using the least square method, solving the atomic set to obtain the sparse coefficient;

[0041] Update the residual using the obtained sparse coefficients.

[0042] Optionally, the updated residual is subjected to feature extraction using a k-svd algorithm, and the adapted dictionary is obtained including:

[0043] Decompose the updated residual;

[0044] Perform SVD decomposition on the decomposed residual to obtain U∈V T ; Where U and V are both unitary matrices;

[0045] Using the first column of U, namely r k The component with the highest frequency after linear transformation updates the k columns of the dictionary, using ∈ ii Multiply the first column of V as the new coefficient row

[0046] The above algorithm is iterated until the residual is a preset small threshold.

[0047] The beneficial effects of the present invention are:

[0048] The present invention first constructs a sensor data transmission tree and collects sensor network data; secondly, a sparse dictionary is constructed based on the collected sensor network data of the present embodiment; finally, the sparse dictionary of the present embodiment is used to compress the real-time sensor network data and recover the sensor network data in the sink node. The present invention can greatly reduce the bandwidth pressure of the network in the application of large-scale sensor networks, and at the same time, the dictionary learning method is used to make the present invention have strong adaptability and can cope with non-abnormal sensor data changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0050] Figure 1 A sensor network constructed by the CDG method of an embodiment of the present invention;

[0051] Figure 2 The figure is a comparison of the complexity of the compressed sensing method of the embodiment of the present invention and the traditional method; wherein (a) is the compressed sensing method, and (b) is the traditional method;

[0052] Figure 3 is the sparsity of the training set and the test set of the embodiment of the present invention; wherein (a) is the training set, and (b) is the test set;

[0053] Figure 4 Schematic diagram of a training data set according to an embodiment of the present invention; wherein (a) is the original data and (b) is the restored data;

[0054] Figure 5 Schematic diagram of a test data set according to an embodiment of the present invention; wherein (a) is the original data, and (b) is the restored data;

[0055] Figure 6 A schematic diagram of error comparison between a training data set and a test data set according to an embodiment of the present invention;

[0056] Figure 7 It is a schematic diagram comparing the sparse coefficients after dictionary training, the sparse coefficients after DCT transformation, and the sparse coefficients after FFT transformation according to an embodiment of the present invention; wherein, (a) is the sparse coefficients after dictionary training, (b) is the sparse coefficients after DCT transformation, and (c) is the sparse coefficients after FFT transformation. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] With the rapid development of IoT technology, IoT devices are playing an increasingly important role in modern society. However, with the increasing breadth and deep integration of IoT applications, the demand for higher information transmission rates is also increasing. This poses new challenges to the design and performance of IoT systems. Compressed sensing (CS) is an emerging signal processing technology that can reconstruct signals at low sampling rates by performing random subsampling. This technology has achieved considerable success in traditional communications and has shown great potential in various fields.

[0060] Most existing methods rely on a predetermined sparse basis, such as Fourier matrix, wavelet transform matrix or some sparse basis that is effective for a specific application scenario. This embodiment focuses on optimizing the transformation basis in the compressed sensing principle and introduces an adaptive dictionary learning K-SVD algorithm, which trains historical data from a specific usage scenario and ultimately generates a sparse basis suitable for a custom application scenario. This method overcomes the limitations of sparse basis selection to a certain extent.

[0061] The optimization method of an adaptive wireless sensor network based on compressed sensing proposed in this embodiment includes: constructing a sensor data transmission tree and collecting sensor network historical data;

[0062] Building a sparse dictionary based on the collected historical data of the sensor network;

[0063] The sparse dictionary is used to compress the real-time sensor network data, and the sensor network data is restored in the sink node.

[0064] This embodiment introduces a dictionary learning algorithm K-SVD, which increases adaptability through multiple iterations of the target signal, so that the algorithm can be used in most similar environments.

[0065] In view of the problems of compressed sensing in current applications, if a sparse basis that is adaptively updated according to the acquired information can be obtained, then this method can be used in most sensor networks. For large-scale sensor networks, the collected data can be compressed, transmitted to the sink node through multi-hop routing, and finally restored at the sink node. In this embodiment, the algorithm used to train the sparse basis is a generalized algorithm of the k-means clustering algorithm: the k-svd algorithm.

[0066] The optimization method process mainly involves 4 parts:

[0067] (1) Collect data and build a sensor data transmission tree

[0068] The CDG method is used to construct a sensor tree with a Sink node, and each node is numbered to facilitate subsequent data restoration.

[0069] (2) Data restoration: preprocessing data acquisition

[0070] The sparse dictionary is obtained by using the historical data of the region or by using the system to collect data within a certain period of time using ordinary collection methods and using an adaptive algorithm.

[0071] (3) Sensor data compression transmission

[0072] Using the sparse dictionary obtained in (2), the k-SVD algorithm is used to perform compressed sensing processing on the data obtained by the sensor, and finally the data is recovered in the sink node.

[0073] (4) Exception handling

[0074] When the sparse coefficients generated by a node's data passing through the sparse dictionary do not meet the requirements of compressed sensing, this embodiment ends the current round of transmission and returns to step (2) to retrain the sparse dictionary based on the data in the previous period of time, thereby achieving abnormal processing.

[0075] Furthermore, the optimization method further comprises:

[0076] Perform exception processing on compressed sensor network data; including:

[0077] When the sparse coefficient generated by the sensor network data of a sink node after passing through the sparse dictionary does not meet the requirements of compressed sensing, that is, when the sparsity exceeds the value required by the compressed sensing theory M≥cμ(φ, ψ)^2Klog(N), the current round of transmission is terminated, and the sparse dictionary is retrained according to the data in the preset previous time period;

[0078] Furthermore, constructing the sensor data transmission tree includes:

[0079] Use CDG method to connect sensor network and sink node to obtain sensor network forest, in which the node directly connected to sink node is set as root node;

[0080] Data is defined for each tree in the sensor network forest to obtain the sensor data transmission tree.

[0081] Furthermore, for each tree in the sensor network forest, data definition includes:

[0082] The nodes are numbered according to the order of breadth-first search, where the data of the i-th node is d i ;

[0083] Set the observation matrix φ. In the process of transmitting data from the leaf node to the root node, each node is multiplied by the observation matrix for aggregation to obtain the vector y. In each transmission, only a temporary y matrix is ​​transmitted, and finally added, aggregated at the root node, and restored at the sink node.

[0084] Among them, the vector y is:

[0085]

[0086] Where N represents the signal dimension, y i represents the compressed data, j is the process variable representing the sum of matrix multiplication, represents the element in row i and column j in the observation matrix, d j Represents the data of the i-th node.

[0087] Specifically, in this embodiment, data collection includes:

[0088] For a randomly distributed sensor network (such as Figure 1 As shown), this embodiment adopts the CDG method, that is, the tree structure connection method to connect it with the sink node. For the node directly connected to the sink node, this embodiment sets it as the root node. Finally, this embodiment can obtain a forest for the sensor network. For each tree in the forest, this embodiment defines its data: the nodes are numbered according to the order of breadth-first search, where the data of the i-th node is d i ,This embodiment uses the 01 matrix as the observation matrix φ. In the process of transmitting data from the leaf node to the root node, each node is multiplied by the observation matrix for aggregation, and finally obtains the vector y. Therefore, only a temporary y matrix is ​​transmitted each time it is transmitted, and finally added and aggregated at the root node, and restored at the sink node.

[0089] For the final y matrix

[0090]

[0091] That is:

[0092]

[0093] According to the theory of compressed sensing, the probability of reconstructing a k-sparse signal from a small amount of data in this embodiment is close to 1. If an N-dimensional signal has a domain that can be represented by k (k is much smaller than N) non-zero sparse representations, then the signal is considered to be a k-sparse signal. Obviously, the d matrix obtained in the actual process is non-sparse in most cases, so it is necessary to obtain a sparse signal through some methods. The most commonly used method is DCT transform or Fourier transform, which is mathematically expressed as

[0094]

[0095] Where Ψ is a DCT transform matrix or a Fourier transform matrix, and x is a sparse signal obtained in this embodiment. However, for data in a sensor network, it is difficult to find a suitable sparse basis Ψ to obtain a sparse signal, so in data collection, this embodiment introduces a method to adapt various signals in a sensor network.

[0096] At the same time, according to the theory of compressed sensing, a k-sparse signal can be reconstructed from M observations only when M satisfies the following conditions:

[0097] M ≥ cμ(φ,ψ)^2Klog(N)

[0098] where c is a constant and μ is the correlation of the matrix, defined as

[0099]

[0100] It can be concluded from the above formula that when the coherence between φ and ψ is smaller, the amount of measurement required for reconstruction is smaller. It has been proved that M=3k~4k can usually meet the requirements of this embodiment.

[0101] For the traditional multi-hop transmission method, the closer the sensor is to the sink node, the larger the amount of data it transmits. However, this data transmission method based on compressed sensing transmits a matrix of length M*1 each time it transmits. At the same time, its parent node adds the matrix obtained by calculating the received node and its own collected data. Therefore, the amount of information sent by each node is O(M), and the total overhead is O(MN). Since M is much smaller than N, it is less than the overhead of the traditional multiple routing method O(N 2 ).like Figure 2 shown.

[0102] Further, based on the collected sensor network data, constructing a sparse dictionary includes:

[0103] Defining a problem-solving goal for the collected sensor network data;

[0104] Based on the problem-solving goal, a linear programming algorithm is used to solve the problem and obtain a sparse dictionary.

[0105] Among them, solving the problem and obtaining the sparse dictionary are theoretically decoupled, that is, the trained sparse dictionary appears as a parameter in the goal of solving the problem.

[0106] Specifically, in this embodiment, data restoration includes:

[0107] For the data finally obtained, this embodiment can obtain such a formula: Ax = y, where A and y are known. Obviously, this problem has infinite solutions, but this embodiment requires the optimal solution. From data collection, it is known that y = ψφx in compressed sensing. The requirement of this embodiment is that the correlation between φ and ψ is extremely small, and x is a sparse matrix. So for this problem, let A = φψ. The solution required by this embodiment is the solution that is as sparse as possible among the infinite number of x solutions, that is, the goal of this embodiment is to solve:

[0108] min||x||0s.t.Ax=y

[0109] However, this is an NP-hard problem. According to the properties of the lp norm, this embodiment can approximately regard this problem as:

[0110] min||x||1s.t.Ax=y

[0111] Obviously, this is an NP problem and a linear programming problem, which can be solved by using some linear programming algorithms.

[0112] Furthermore, using the sparse dictionary to compress the real-time sensor network data includes:

[0113] Use k-means algorithm to cluster real-time sensor network data;

[0114] Based on the clustered data, the omp algorithm is used to obtain the coefficient signal of the dictionary and update the residual;

[0115] For the updated residual, the k-svd algorithm is used to extract features and obtain the adapted dictionary.

[0116] Furthermore, the k-means algorithm is used to cluster the real-time sensor network data, including:

[0117] Randomly select k points in the real-time sensor network data as cluster centers;

[0118] Assign each sensor data point to the cluster center that is closest to it;

[0119] After all sensor data points are assigned, the cluster center is updated based on the preset update formula until the cluster center no longer changes;

[0120] Among them, the preset update formula is:

[0121]

[0122] In the formula, x represents the data point, c i represents the i-th cluster center, |S i | is the number of data points in the cluster.

[0123] Furthermore, the omp algorithm is used to obtain the coefficient signal of the dictionary, and the updating residual includes:

[0124] Define initial residual and atom sets;

[0125] Select the data most relevant to the current residual from the clustered data, add them to the atomic set, and solve the atomic set using the least square method to obtain a sparse signal;

[0126] The residual is updated using the acquired sparse signal.

[0127] Furthermore, the k-svd algorithm is used to extract features from the updated residual, and the adapted dictionary includes:

[0128] Decompose the updated residual;

[0129] Perform SVD decomposition on the decomposed residual to obtain U∈V T ; Wherein, U and V are both unitary matrices, the size of U is m*m, and the size of V is n*n. ∈ is an m*n matrix, and all elements except the main diagonal are 0, and the elements on the diagonal are singular values;

[0130] Using the first column of U, namely r k The component with the highest frequency after linear transformation updates the k columns of the dictionary, using ∈ ii Multiply the first column of V as the new coefficient row

[0131] The above algorithm is iterated until the residual is small enough to be ignored in the system.

[0132] Specifically, in this embodiment, the sensor data compression transmission includes:

[0133] Initialize the classification and continuously update the cluster center through the k-means algorithm until the cluster center no longer changes. Since this embodiment needs to ensure the sparsity of the transformed signal, this embodiment uses the k-svd algorithm to extract features after using the omp algorithm to obtain the coefficient signal of the dictionary to obtain an adapted dictionary.

[0134] k-means algorithm

[0135] The goal of the k-means algorithm is to divide n objects in a high-dimensional space into k clusters so that the sum of the distances from each object to the center of the cluster to which it belongs is minimized, usually the Euclidean distance, but in special cases it can also be the Manhattan distance, etc.

[0136] For n objects, this embodiment first performs initial classification, randomly selects k points as cluster centers, and then assigns each object to the cluster center closest to it. After all objects are assigned, this embodiment updates the cluster center:

[0137]

[0138] Where x is the data point, c i is the i-th cluster center, |S i | is the number of data points in the cluster.

[0139] Finally, this embodiment iterates the above algorithm until the cluster center no longer changes.

[0140] omp algorithm:

[0141] This embodiment sets a problem, b=Dx, where b is the original signal and D is the transformation basis, or dictionary matrix, wherein b and D are known. The requirement of this embodiment is to obtain a sufficiently sparse x signal.

[0142] The standard omp algorithm is to first define the initial residual r0 = b, the atomic set k = 0;

[0143] First, select the atom most relevant to the current residual, i.e., λ k =argmin| <d j ,r k-1 >| and add it to the atomic set Λ k =Λ k-1 ∪{λ k};

[0144] Generally, the least squares method is used to solve the problem.

[0145]

[0146] Get a sparse signal x k This embodiment updates the residual:

[0147] r k = b-Ax k

[0148] When the residual exceeds the threshold, the operation is repeated until a residual x is small enough. k .

[0149] k-svd algorithm

[0150] Since the present embodiment needs to ensure the sparsity of the transformed signal, the present embodiment uses the omp algorithm to obtain the coefficient signal x of the signal y under a certain dictionary D, and then uses the svd algorithm to extract features and update the dictionary, and repeats this process to obtain an adapted dictionary D.

[0151] First, this embodiment initializes the dictionary D. The quality of the dictionary is not important here. As long as the number of iterations is sufficient, convergence will occur in the end.

[0152] Then, for a fixed D, this embodiment uses the OMP algorithm to train the sparse representation x of the original data set y. At this time, this embodiment obtains the residual r. At this time, this embodiment decomposes r into k r [4] ,in:

[0153]

[0154] Then, in this embodiment, r can be expressed as:

[0155]

[0156] This embodiment fixes x, iterates k times, and uses a trick to eliminate vector x T k The 0 term in avoids non-convergence and the r after the operation k Perform SVD decomposition to obtain U∈V T , this embodiment uses the first column of U, i.e., r k The component with the highest frequency after linear transformation updates the k columns of the dictionary, using ∈ ii Multiply the first column as the new coefficient row Finally, this embodiment iterates the above algorithm until the residual of this embodiment is small enough to be ignored in the system.

[0157] The following is the experimental design and results of this embodiment:

[0158] Dataset selection:

[0159] This example uses the daily temperature data of the Gaza region in the past few years provided by the Kaggle platform as the data set of this example for simulation experiments. After cleaning the data, the original data of this example is two 300*365 matrices, each column of which represents the original data obtained by 300 sensors simulated by this example at the same time. One of them is used as a training set for training the sparse basis, and the other is used as a test set for verification.

[0160] Model selection:

[0161] This embodiment reads the data of every 2 minutes as a single upward transmission. First, this embodiment uses the data of the first half of the time as a training set to train the model and calculate the error. Then, this embodiment uses the transformation basis obtained from the training to perform information transmission of the CDG (compressed sensing data transmission) method on the data of the next second half of the time, and recovers the data calculation error. Finally, this embodiment analyzes these errors and compares them with the old method to evaluate the robustness of the model.

[0162] Experimental results and analysis:

[0163] Model sparsity test:

[0164] This embodiment obtains a transformation basis D adapted to the training set by training the training set. This embodiment uses the training set and the sparse basis D as parameters to obtain a sparse representation using the omp algorithm, and counts the signal sparsity within a single time ( Figure 3 (a)), we can see that the sparsity here is below 20, which meets the requirements of compressed sensing data restoration. This embodiment performs the same operation on the test set, and we can see that its signal sparsity ( Figure 3 (b)) still meets the requirements.

[0165] Model error check:

[0166] Through the sparse signal obtained above, this embodiment uses the compressed sensing method to compress and restore it. By comparing the original signal with the restored signal, this embodiment can find that the training set ( Figure 2 ) The error after restoration is very small, which is obvious. At the same time, this embodiment also shows that the test set ( Figure 3 ) is also very small and within the acceptable range of the system. Through multiple tests, the error curve ( Figure 4 ), confirming that the model's errors are within an acceptable range in this scenario.

[0167] Comparative experiment:

[0168] It can be observed from the experiment that on the same data set, using DCT and FFT matrices as sparse bases, the effect of sparsification is much worse than that of sparse bases trained with the K-SVD algorithm. The sparse coefficients obtained by traditional methods often have the same dimension as the original data, which does not meet the basic requirements of the compressed sensing principle. On the contrary, the sparse coefficients of the data after sparsification are within the range specified by M≥cμ(φ, ψ)^2Klog(N), indicating that the model is superior to the general traditional model.

[0169] In the experiment, this embodiment has verified that the model has strong adaptability. It can not only discretize the training signal, use the compressed sensing method to efficiently transmit the signal and average the network energy consumption, but also the obtained conversion basis can also convert similar signals. Therefore, the model performs well on the signals collected in some scenes and can realize operations such as compressed sensing.

[0170] Figure 5 Schematic diagram of the test data set of this embodiment; wherein (a) is the original data, and (b) is the restored data;

[0171] Figure 6 This is a schematic diagram of the error comparison between the training data set and the test data set of this embodiment;

[0172] Figure 7 The figure is a schematic diagram comparing the sparse coefficients after dictionary training, the sparse coefficients after DCT transformation, and the sparse coefficients after FFT transformation of this embodiment; wherein, (a) is the sparse coefficients after dictionary training, (b) is the sparse coefficients after DCT transformation, and (c) is the sparse coefficients after FFT transformation.

[0173] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. An optimization method for an adaptive wireless sensor network based on compressed sensing, characterized in that: include: Build a sensor data transmission tree and collect sensor network historical data; Building a sparse dictionary based on the collected historical data of the sensor network; The sparse dictionary is used to compress the real-time sensor network data, and the sensor network data is restored in the sink node.

2. The optimization method of adaptive wireless sensor network based on compressed sensing according to claim 1, characterized in that: The optimization method further comprises: Perform exception processing on compressed sensor network data; including: When the sparse coefficients generated by the sensor network data of a sink node after passing through the sparse dictionary do not meet the requirements of compressed sensing, the current round of transmission is terminated, and the sparse dictionary is retrained according to the data in the preset previous time period; The requirement of compressed sensing is that the sparsity exceeds the value required by compressed sensing theory.

3. The optimization method of adaptive wireless sensor network based on compressed sensing according to claim 1, characterized in that: Building a sensor data transmission tree includes: Use CDG method to connect sensor network and sink node to obtain sensor network forest, in which the node directly connected to sink node is set as root node; Data is defined for each tree in the sensor network forest to obtain the sensor data transmission tree.

4. The optimization method of adaptive wireless sensor network based on compressed sensing according to claim 3 is characterized in that: For each tree in the sensor network forest, data definition includes: The nodes are numbered according to the order of breadth-first search, where the data of the i-th node is d i ; Set the observation matrix φ. In the process of transmitting data from the leaf node to the root node, each node is multiplied by the observation matrix for aggregation to obtain the vector y. In each transmission, only a temporary y matrix is ​​transmitted, and finally added, aggregated at the root node, and restored at the sink node. Among them, the vector y is: Where N represents the signal dimension, y i represents the compressed data, j is the process variable representing the sum of matrix multiplication, represents the element in row i and column j in the observation matrix, d j Represents the data of the i-th node.

5. The optimization method of adaptive wireless sensor network based on compressed sensing according to claim 1, characterized in that: Based on the collected sensor network data, constructing a sparse dictionary includes: Defining a problem-solving goal for the collected sensor network data; Based on the problem-solving goal, a linear programming algorithm is used to solve the problem and obtain a sparse dictionary.

6. The optimization method of adaptive wireless sensor network based on compressed sensing according to claim 1, characterized in that: Using the sparse dictionary to compress real-time sensor network data includes: Use k-means algorithm to cluster real-time sensor network data; Based on the clustered data, the omp algorithm is used to obtain the sparse coefficients of the dictionary and update the residual; For the updated residual, the k-svd algorithm is used to extract features and obtain the adapted dictionary.

7. The optimization method of adaptive wireless sensor network based on compressed sensing according to claim 6 is characterized in that: Using the k-means algorithm, clustering real-time sensor network data includes: Randomly select k points in the real-time sensor network data as cluster centers; Assign each sensor data point to the cluster center that is closest to it; After all sensor data points are assigned, the cluster center is updated based on the preset update formula until the cluster center no longer changes; Among them, the preset update formula is: In the formula, x represents the data point, c i represents the i-th cluster center, |S i | is the number of data points in the cluster.

8. The optimization method of adaptive wireless sensor network based on compressed sensing according to claim 6, characterized in that: The omp algorithm is used to obtain the sparse coefficients of the dictionary, and the updated residual includes: Define initial residual and atom sets; Select the data most relevant to the current residual from the clustered data and add it to the atomic set; Using the least square method, solving the atomic set to obtain the sparse coefficient; Update the residual using the obtained sparse coefficients.

9. The optimization method of adaptive wireless sensor network based on compressed sensing according to claim 6, characterized in that: For the updated residual, the k-svd algorithm is used to extract features, and the adapted dictionary is obtained including: Decompose the updated residual; Perform SVD decomposition on the decomposed residual to obtain U∈V T ; Where U and V are both unitary matrices; Using the first column of U, namely r k The component with the highest frequency after linear transformation updates the k columns of the dictionary, using ∈ ii Multiply the first column of V as the new coefficient row The above algorithm is iterated until the residual is a preset small threshold.

Images