Multi-physical-quantity sensing node sparse deployment method and system and medium

Abstract

The invention relates to a multi-physical-quantity sensing node sparse deployment method and system and a medium. The method comprises the following steps: extracting common effective sensing positions and corresponding data among data of multiple physical quantities; inputting the corresponding data into a multi-physical-quantity fractal auto-encoder for joint spatial position feature selection, wherein the multi-physical-quantity fractal auto-encoder comprises a plurality of fractal auto-encoders corresponding to the multi-physical-quantity data; the fractal auto-encoder updates position selection weights of multiple physical quantities layer by layer, and the multiple position selection weights form a position selection joint feature weight vector; and the multi-physical-quantity fractal auto-encoder performs joint spatial position feature selection based on the position selection joint feature weight vector to obtain a position index of common low-dimensional sampling of corresponding data, and the position index serves as a deployment position of the multi-physical-quantity sensing node. Compared with the prior art, a group of common sparse measurement position coordinates can be provided for different data types, and sensing node deployment optimization of multi-physical-quantity monitoring is achieved.

Description

Technical Field

[0001] The present invention relates to the field of ocean observation technology, and in particular to a method, system and medium for sparse deployment of multi-physical quantity sensing nodes. Background Art

[0002] In the field of ocean observation, the location selection of sensor nodes used to monitor the physical environment directly affects the cost of deploying the ocean observation system, as well as the effectiveness and accuracy of monitoring the physical state. Due to the vast scope of the ocean, conducting full-state monitoring of ocean physical quantities on a global scale does not meet practical needs and conditions. This is because the cost of deploying a large number of sensor nodes globally is very high. In particular, sensor nodes used to monitor the ocean environment are typically buoys, submersibles, or engineering vessels equipped with multiple sensors, which are expensive to deploy and collect data from. Therefore, it is necessary to use a sparse site selection method to determine the optimal observation locations in the ocean observation system, so that the ocean observation system can deploy sensors at a limited number of locations and reconstruct the optimal full-state data using the ocean physical quantity data from these limited locations. In this case, the sensor node deployment problem can be considered as a data feature selection problem for optimization, and special attention should be paid to the preservation of the selected node location features (i.e., spatial features).

[0003] However, this differs significantly from traditional data dimensionality reduction and reconstruction methods, such as those based on principal component analysis (PCA) and autoencoders. Traditional methods achieve dimensionality reduction by mapping high-dimensional data into a low-dimensional space, but this low-dimensional space often lacks practical physical meaning. Although such methods can minimize reconstruction error, they cannot directly extract a subset of features from the original dataset, making them difficult to directly use for eliminating redundant features or selecting node subsets. In particular, in practical ocean monitoring systems, each node (such as a buoy, submersible, or other monitoring device) typically needs to monitor two or more physical parameters. In this case, PCA and autoencoder-based data dimensionality reduction methods cannot directly extract joint spatial features from multi-source data. Although multi-view representation learning (MVRL) can effectively compress and reconstruct multi-source data, when processing multi-view (also considered multi-source) data, the common latent variables obtained by MVRL are directly mapped into a low-dimensional feature space, which cannot directly correspond to the sparsely deployed node position selection in the actual physical space.

[0004] Therefore, it is necessary to provide a new method that takes into account the need for perception nodes to monitor multiple physical quantity data in order to solve the problem of joint extraction of multi-physical quantity data features, thereby performing joint dimensionality reduction and reconstruction of multiple physical quantities based on the extracted sparse spatial features, and realizing the optimization of perception node deployment for multi-physical quantity monitoring. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art that it is impossible to directly extract joint spatial features from multi-source data, and it is difficult to select corresponding node positions that are sparsely deployed in actual physical space, and to provide a method, system and medium for sparse deployment of multi-physical quantity perception nodes.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] As a first aspect of the present invention, a method for sparse deployment of multi-physical quantity sensing nodes is provided, the method comprising:

[0008] Preprocess the multi-physical quantity data and extract the common effective perception positions and corresponding data among the multi-physical quantity data;

[0009] Inputting the corresponding data into a multi-physical quantity fractal autoencoder to perform joint spatial position feature selection to obtain a position index of a common low-dimensional sampling of the corresponding data;

[0010] Outputting the location index as the deployment location of the multi-physical quantity sensing node;

[0011] Among them, the multi-physical quantity fractal autoencoder includes multiple fractal autoencoders corresponding to the multi-physical quantity data; the fractal autoencoder updates the position selection weights of the corresponding physical quantities, and the multiple position selection weights constitute a position selection joint feature weight vector; the multi-physical quantity fractal autoencoder performs joint spatial position feature selection based on the position selection joint feature weight vector to obtain the position index.

[0012] As an optimal technical solution, the multi-physical quantity fractal autoencoder corresponds to a set of hidden variables for each physical quantity, and multiple sets of hidden variables are connected through the position selection joint feature weight vector.

[0013] As a preferred technical solution, the fractal autoencoder includes an encoder and a decoder, which respectively perform encoding and decoding operations on the corresponding physical quantity data; the encoder includes a one-to-one layer and a feature selection layer;

[0014] The encoder obtains the position selection weight of the corresponding physical quantity data through a one-to-one layer, and the feature selection layer performs weighted averaging on the position selection weight to obtain a position selection joint feature weight vector and perform joint spatial position feature selection; the encoder obtains the hidden variable representation of the physical quantity based on the one-to-one layer and the feature selection layer; the decoder performs nonlinear reconstruction on the full state data based on the hidden variable representation.

[0015] As a preferred technical solution, during the training phase, the multi-physical quantity fractal autoencoder updates the joint feature weight vector for the position selection of the multi-physical quantities as follows:

[0016] For each physical quantity, the position selection weights obtained in one layer are summed and averaged after each round of updating.

[0017] The updated average weight is input into the feature selection layer, and the feature selection layer selects the largest p features in the weight average to update the position and select the joint feature weight matrix;

[0018] The hidden variable representation is obtained based on the updated joint feature weight matrix, and the encoder and decoder parameter training is performed.

[0019] As a preferred technical solution, the feature selection layer updates the joint feature weight matrix as follows:

[0020]

[0021] Among them, w i is the weight of the feature selection layer of the i-th physical quantity autoencoder; w avg is the average value of the feature selection layer weight; M is the total number of physical quantities; is the maximum p weight index vector in the weighted average The corresponding joint feature weight matrix; Diag(·) represents the construction of vector A square matrix whose elements are the main diagonal;

[0022] The hidden variables are represented as:

[0023]

[0024] in, is the mapping function, corresponding to the mapping relationship W of the encoder of the i-th physical quantity avg_f represents the joint feature weight matrix; X i is the data of the i-th physical quantity; θ′ is the encoder parameter.

[0025] As a preferred technical solution, the optimization function of the multi-physical quantity fractal autoencoder is:

[0026]

[0027] in, is the global loss function; is the feature selection layer loss function; η is the balance parameter of the sub-neural network term, and λ represents the reconstruction error and the joint feature weight matrix W avg_f The balance parameter between the sparse regularization terms of ;

[0028] The global loss function is expressed as follows:

[0029]

[0030] in, is the decoder mapping function of the i-th physical quantity; θ i is the decoder parameter;

[0031] The feature selection layer loss function is expressed as follows:

[0032]

[0033] in, is the decoder mapping function of the i-th physical quantity; θ i is the decoder parameter; It is the joint feature weight matrix updated by the feature selection layer based on the p features with the largest weights.

[0034] As an optimal technical solution, the encoder and decoder parts of the multi-physical quantity fractal autoencoder both adopt fully connected neural networks and are both single-layer structures.

[0035] As a preferred technical solution, the multi-physical quantity fractal autoencoder uses the entropy weight method to initialize the weight of each spatial feature in the feature selection layer:

[0036] For the j-th position feature, the entropy weight is calculated as follows:

[0037]

[0038] in, s kj is the normalized data, k = 1, 2, ..., m, j = 1, 2, ..., n, m and n are the number of data samples and the number of position features respectively;

[0039] The average spatial characteristic entropy weight of multi-physical quantity data is calculated as follows:

[0040]

[0041] Among them, e i For data X i The entropy weight vector of .

[0042] As a second aspect of the present invention, a multi-physical quantity sensing node sparse deployment system is provided, which is used to implement the multi-physical quantity sensing node sparse deployment method described above, including:

[0043] A data preprocessing module is used to extract the common effective sensing positions and corresponding data of the common effective positions among multiple physical quantity data;

[0044] A joint spatial position feature selection module is used to input the corresponding data into a multi-physical quantity fractal autoencoder to perform joint spatial position feature selection, and obtain a position index of a common low-dimensional sampling of the corresponding data;

[0045] A node deployment module, configured to output the location index as a deployment location of a multi-physical quantity sensing node;

[0046] Among them, the multi-physical quantity fractal autoencoder includes multiple fractal autoencoders corresponding to the multi-physical quantity data; the fractal autoencoder updates the position selection weights of the corresponding physical quantities, and the multiple position selection weights constitute a position selection joint feature weight vector; the multi-physical quantity fractal autoencoder performs joint spatial position feature selection based on the position selection joint feature weight vector to obtain the position index.

[0047] As a third aspect of the present invention, a storage medium is provided, on which a program is stored, and when the program is executed, the sparse deployment method of multi-physical quantity sensing nodes as described above is implemented.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1) Aiming at the demand of joint spatial feature selection based on multi-physical quantity data to optimize node deployment, the present invention proposes a multi-physical quantity fractal autoencoder (MFAE) to achieve the optimization of sparse deployment of ocean sensing nodes. MFAE corresponds each physical quantity to a hidden variable, and the two hidden variables are connected through a position selection weight vector; by jointly updating the position selection weights of multiple types of data, and selecting the position features of different types of data based on these weights, different types of data can share a set of position coordinates. Subsequently, the parameters of the autoencoder are trained based on the weights to obtain the optimal mapping relationship between the sampling subsets of different types of data and the corresponding full-state data, thereby achieving high-quality ocean data reconstruction.

[0050] 2) In combination with the requirements of multi-physical quantity node deployment, the present invention proposes a loss function for autoencoders suitable for different multi-physical quantity node deployments. When evaluating the performance of the multi-physical quantity autoencoder, not only the reconstruction error of a single physical quantity data is considered, but also the average reconstruction error between physical quantities is considered through a global loss function; for the feature selection layer, the autoencoder loss functions of multiple physical quantities are set; and a sparse control term of the joint feature weight matrix is ​​added to ensure that the features selected by the multi-physical quantity are sparse.

[0051] 3) The present invention takes into account the influence of physical space factors corresponding to multiple physical quantities during node deployment, and also adopts the entropy weight method to initialize the feature selection weight using the average entropy weight of the spatial position of the multi-physical quantity data, thereby further improving the average (joint) data reconstruction accuracy, especially for data with large noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of a sparse deployment method for multi-physical quantity sensing nodes of the present invention;

[0053] Figure 2 is a hidden variable model for two physical quantity data;

[0054] Figure 3 Selecting an encoder structure for the characteristics of two physical quantity data;

[0055] Figure 4 It is the structure of multi-physical quantity fractal encoder;

[0056] Figure 5 Data flow diagram for the North Pacific region embodiment;

[0057] Figure 6 Comparison results of temperature reconstruction errors for the North Pacific region embodiment;

[0058] Figure 7 Comparison results of salinity reconstruction errors for the North Pacific region example;

[0059] Figure 8 The comparison results of the average reconstruction error of the embodiment in the North Pacific region are shown;

[0060] Figure 9 The comparison results of temperature reconstruction errors for the embodiment in the Arctic Ocean region are shown;

[0061] Figure 10 Comparison results of salinity reconstruction errors for the Arctic Ocean region embodiment;

[0062] Figure 11 The figure shows the comparison results of the average reconstruction errors of the embodiments in the Arctic Ocean region. DETAILED DESCRIPTION

[0063] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0064] Example 1

[0065] Aiming at the demand of joint spatial feature selection based on multi-physical quantity data to optimize node deployment, this paper proposes a multi-physical quantity sensing node sparse deployment method. Figure 1 As shown, the method steps include:

[0066] The multi-physical quantity data input data is pre-processed into a module to extract the common effective sensing positions and corresponding data of the common effective sensing positions among the multi-physical quantity data;

[0067] The corresponding data of the common effective perception position are input into the multi-physics fractal autoencoder (MFAE) for joint spatial position feature selection to obtain the position index of the common low-dimensional sampling of the corresponding data;

[0068] The location index is output as the deployment location of the multi-physical quantity sensing node.

[0069] The multi-physics fractal autoencoder includes multiple fractal autoencoders corresponding to the physical quantity data. Each physical quantity is associated with a set of hidden variables, and these multiple sets of hidden variables are linked by a position selection weight vector. The fractal autoencoder updates the position selection weights for the corresponding physical quantity. The multiple position selection weights form a position selection joint feature weight vector. The multi-physics fractal autoencoder performs joint spatial position feature selection based on the position selection weights and the joint feature vector to obtain a position index.

[0070] The fractal autoencoder contains multiple encoders and decoders corresponding to the input physical quantities, which respectively encode and decode the corresponding physical quantity data; the encoder part also includes a one-to-one layer and a feature selection layer; the encoder obtains the position selection weight of the corresponding physical quantity data through the one-to-one layer, and the feature selection layer performs weighted averaging on the position selection weight and updates the joint feature weight of the physical quantity; finally, the encoder obtains the hidden variable representation of the physical quantity based on the one-to-one layer and the feature selection layer, and extracts low-dimensional feature data through the feature selection layer; the decoder performs nonlinear reconstruction on the full-state data based on the hidden variable representation to obtain the position index of the common low-dimensional sampling of the corresponding data.

[0071] The sparse deployment optimization model of multi-physical quantity sensing nodes proposed in this invention is specifically designed as follows:

[0072] Considering that the node deployment of multiple physical quantities is more in line with the needs of ocean perception, joint feature extraction based on multi-physical quantity data can provide a better spatial location selection for node deployment. In terms of feature extraction of a single physical quantity, the autoencoder can achieve nonlinear mapping from low dimension to high dimension with low computational complexity. Therefore, the present invention proposes an autoencoder model for multi-physical quantity feature extraction, which obtains common spatial position features by processing the input multi-physical quantity data, thereby being used for node deployment optimization.

[0073] The low-dimensional measurement obtained after feature extraction by the autoencoder is expressed as: X J =CX. Selected low-dimensional measurement X J The nonlinear mapping function spanning the full high-dimensional state space can be expressed as: The reconstructed data can be expressed as: Where θ is the decoder parameter. Based on this, the loss function of the autoencoder for unsupervised feature extraction (which can also be regarded as minimizing the reconstruction error) can be expressed as:

[0074]

[0075] The candidate locations for deploying nodes are used as features, and the feature weight vector w is established, and the feature weight matrix W is constructed from it. f =Diag(w). Based on the weight vector w, by selecting the largest p weights, the feature selection subset J can be obtained.

[0076] For feature selection of multiple physical quantities, not only the differences of each physical quantity should be considered, but also the requirement of selecting a set of common spatial position features for multiple physical quantities should be met. Assuming that there are M physical quantity data, the data of the i-th physical quantity is represented by X i , i=1,…,M.

[0077] For the perception node of multiple physical quantities, its hidden variable model can be in the following form:

[0078] exist Figure 2 In the model shown, each physical quantity corresponds to a hidden variable, and the two hidden variables are connected by a position-selected weight vector w. The model assumes that there is a certain relationship between the hidden variable Z1 of data X1 and the hidden variable Z2 of data X2.

[0079] based on Figure 2 The model shown in the figure, the encoder model structure of multi-physical quantity feature selection is as follows Figure 3 As shown in Figure 2, the autoencoder for multi-physical quantity feature selection includes two encoders and a decoder, which are used to extract features from data X1 and X2 and reconstruct them respectively. The encoder is used to extract low-dimensional feature data, while the decoder is used to nonlinearly reconstruct the full-state data based on the extracted low-dimensional feature data.

[0080] Similarly, for data with M (when M > 2) physical quantities, the weights of the M physical quantities need to be jointly processed during feature extraction. In this case, the multi-physical quantity fractal autoencoder contains M encoders and decoders corresponding to the input data, which respectively encode and decode the M data to achieve the learning of the hidden features of the corresponding data under the influence of the joint weights.

[0081] Specifically, in Figure 3 During the encoder training process shown in the figure, in order to achieve multi-physical quantity feature selection, the average value w of the feature selection layer weight is calculated here avg ,Right now:

[0082]

[0083] Among them, w i The weight of the feature selection layer of the i-th physical quantity autoencoder.

[0084] by Represents the maximum p weight index vectors in the weight average; then, the joint feature weight matrix corresponding to the multi-physical quantity fractal autoencoder is updated as:

[0085]

[0086] Among them, Diag(·) represents the construction of vector A square matrix whose elements are the main diagonal; Represents the joint feature weight matrix corresponding to the largest p weight index vectors in the weight average.

[0087] For the convenience of representation, the hidden variables here are represented as:

[0088]

[0089] Among them, the mapping function here The mapping relationship of the encoder corresponding to the i-th physical quantity; θ′ is the encoder parameter. When inputting formula 4 (replacing W avg_f ), which is the feature extraction operation of the feature selection layer (also called mapping relationship), corresponding to the extracted low-dimensional features.

[0090] Design of multi-physical quantity fractal autoencoder: Based on the above-mentioned multi-physical quantity feature selection autoencoder model, as well as the feature extraction and nonlinear reconstruction principles of the autoencoder, the specific multi-physical quantity fractal autoencoder structure and its optimization objective function of the present invention are given here to realize the joint feature selection of multi-physical quantity data.

[0091] Multi-physics fractal autoencoder structure: Since the architectures of multi-physics autoencoders are different, it is crucial to find the autoencoder that is most suitable for multi-physics node deployment. Based on the above-mentioned criteria and models for feature selection of multi-physics quantities, considering the different distribution characteristics of each data, after selecting a set of position features, the autoencoder parameters of each data are trained independently. Therefore, a multi-physics fractal autoencoder is proposed here for joint feature selection. Its specific structure is as follows: Figure 4 shown.

[0092] Specifically, the encoder and decoder parts of the multi-physical quantity fractal autoencoder both adopt fully connected neural networks, and both are single-layer structures. Among them, the encoder part of each input data includes a one-to-one layer and a feature selection layer. During the training stage, the multi-physical quantity fractal autoencoder realizes the joint weight update of multiple physical quantities in the following way: the weights obtained from the one-to-one layer of different data are summed and averaged after each round of update. Subsequently, the updated average weight is input into the feature selection layer for feature selection, and the training of the encoder and decoder parameters is continued. Among them, the one-to-one layer assigns a learnable weight (scalar) to each input feature, and the importance of the feature is reflected by the size of the weight. The corresponding loss function is minimized during training, and the gradient is updated through backpropagation to adjust the weight until the training termination condition is met (such as the loss function is no longer effectively improved). Here, the fractal autoencoder mainly realizes unsupervised feature selection through the joint optimization of sparse regularization and reconstruction error minimization. Its weight update is directly related to the information content of the feature and the reconstruction ability of the model.

[0093] Definition of the optimization function of the multi-physical quantity fractal autoencoder: Based on the architecture of the multi-physical quantity fractal autoencoder and the requirements of multi-physical quantity node deployment, the loss function of the autoencoder suitable for different multi-physical quantity node deployments is given below.

[0094] For perceptual data with multiple physical quantities, when evaluating the performance of the multi-physical quantity autoencoder, not only the reconstruction error of a single physical quantity data but also the average reconstruction error between physical quantities should be considered. For the data of the i-th physical quantity, the data selected by the corresponding encoder is the hidden variable z i . Then the corresponding global loss function is expressed as follows:

[0095]

[0096] Among them, θ i Decoder parameters. is the decoder mapping function of the i-th physical quantity.

[0097] For the feature selection layer, the autoencoder loss function of multiple physical quantities is:

[0098]

[0099] Similarly, in order to make the features selected by multiple physical quantities sparse, a sparse control term of the joint feature weight matrix is ​​added here, so the optimization function of the multi-physical quantities fractal autoencoder (MFAE) can be obtained as:

[0100]

[0101] Among them, η is the balance parameter of the sub-neural network term, λ represents the reconstruction error and the joint feature weight matrix W avg_f The balance parameter between the sparse regularization terms of .

[0102] Initialization of the feature selection layer of multi-physical fractal autoencoder

[0103] When performing feature selection using an autoencoder, the initialization method of the feature selection layer affects the reconstruction error of the autoencoder and, in turn, the effectiveness of node deployment. However, FAEs use a uniformly distributed method to initialize feature weights, which does not consider the impact of the physical space factors corresponding to various physical quantities during node deployment. Therefore, to further improve the node deployment performance of MFAEs, a new weight initialization strategy is proposed.

[0104] When extracting spatial features from multi-physical quantity data, it is considered that features with large amounts of information will receive higher weights during initialization, which helps the model capture key features faster. Therefore, the present invention uses the entropy weight method to measure the amount of information of each spatial feature to improve model performance. The entropy weight method is used here to initialize the weight of each spatial feature in the MFAE. The multi-physical quantity fractal autoencoder initialized based on the entropy weight method is subsequently referred to as EnMFAE (Entropy initialized MFAE).

[0105] For the j-th position feature, the entropy weight is calculated as follows:

[0106]

[0107] in, s kj is the normalized data. k = 1, 2, ..., m, j = 1, 2, ..., n. Here m is the number of samples and n is the number of position features.

[0108] The calculation process of the average spatial characteristic entropy weight of multi-physical quantity data is as follows:

[0109]

[0110] Among them, e i For data X i The entropy weight vector of .

[0111] In response to the demand for joint spatial feature selection based on multi-physical quantity data to optimize node deployment, the present invention proposes a multi-physical quantity fractal autoencoder (MFAE) to achieve the optimization of sparse deployment of ocean sensing nodes. MFAE performs weighted averaging on the weights of multiple types of data to achieve the joint update of multiple physical quantity weights, and performs position feature selection of different types of data based on these weights, so that different types of data can share a set of position coordinates. Subsequently, the parameters of the autoencoder are trained using the weights to obtain the optimal mapping relationship between the sampling subsets of different types of data and the corresponding full-state data, thereby achieving high-quality ocean data reconstruction. In addition, the feature selection layer of MFAE is initialized using the entropy weight method, which further improves the average (joint) data reconstruction accuracy.

[0112] The total computational complexity of the MFAE proposed in this invention when performing joint spatial position feature selection is O(np), which is much lower than the feature selection method based on Transformer.

[0113] Example 2

[0114] This embodiment provides a specific application example of the method described in the above embodiment 1, using data such as sea surface temperature data and ocean salinity data. By inputting the data and a preset p into the MFAE, the optimal sensing node deployment location is obtained.

[0115] To compare node deployment and reconstruction results in different regions, this example divides temperature and salinity data into two subsets: the North Pacific region and the Arctic Ocean region. The North Pacific region's geographic range is 65°N to 10°S, and 78°W to 99°E; the Arctic Ocean region's geographic range is 59°N to 90°N, and 180°W to 180°E. The spatial resolution is 1°×1°.

[0116] The settings of experimental parameters are shown in Table 1.

[0117] Table 1 Experimental parameter settings of MFAE

[0118]

[0119]

[0120] First, the data reconstruction quality evaluation criteria are introduced; then the data preprocessing process is described; finally, in order to more intuitively illustrate the data processing process of MFAE, the data flow diagram of MFAE is shown as follows: Figure 5 shown.

[0121] (1) Reconstruction quality evaluation

[0122] The evaluation criteria for the proposed MFAE are mainly the reconstruction error of the test data Test, which is as follows:

[0123]

[0124] in, The test data is reconstructed by the trained autoencoder.

[0125] (2) Data preprocessing

[0126] When the MFAE proposed in the present invention performs node deployment, since different data need to share the same set of ocean location features, it is necessary to perform data registration operations on the ocean temperature and salinity data. Specifically, the geospatial masks of the ocean temperature and salinity data (in the downloaded dataset mask, valid positions are 1 and invalid positions are 0) are logically ANDed to extract the common and valid ocean monitoring locations between the two types of data. Taking the North Pacific region as an example, 9616 common and valid locations were obtained after processing, which serve as all candidate locations for the deployment of multi-physical quantity nodes for ocean monitoring.

[0127] In addition, since the data distribution ranges of temperature data and salinity data are different, in order to perform effective feature extraction, the extracted valid data are first normalized to make the distribution ranges of the two data consistent so that they can be input into MFAE for joint feature selection.

[0128] Figure 6-7 Figure 1. Data flow diagram for inputting two physical quantity data sets (view_0 and view_1) from the same set of candidate locations into the MFAE, where p = 10 and n = 9616. The two ocean data sets are fed into the input layer, then jointly extracted by the feature selection layer, and then fed into their respective encoder and decoder layers for training. For n candidate locations, the computational complexity of the MFAE feature selection layer is O(n), the computational complexity of the encoder is 2O(np), and the computational complexity of the decoder is 2O(np). The total computational complexity of this MFAE is O(np).

[0129] To demonstrate the effectiveness of the proposed MFAE for multi-physics joint feature selection, we compare it with a decoder reconstruction method based on a randomly deployed node subset (referred to as RandDE). The decoder portion of the random deployment method is identical to that of the MFAE, and the error function used during training is the average of the multi-physics data reconstruction errors.

[0130] North Pacific Region Results

[0131] (2.1) Comparison of reconstruction errors

[0132] refer to Figure 6Comparison of ocean temperature reconstruction errors shows that MFAE's reconstruction error distribution is generally smaller than RandDE's, and its average error is also lower. EnMFAE achieves lower reconstruction error as the number of nodes increases. Compared to RandDE, MFAE's error distribution is more concentrated, indicating that MFAE's reconstruction is more stable. Overall, EnMFAE achieves superior reconstruction results. This demonstrates that entropy weight initialization can reduce the impact of noise on spatial feature selection in temperature data, positively impacting node deployment.

[0133] refer to Figure 7 , the ocean salinity reconstruction error was compared. The results showed that when the number of deployed nodes varied, the average reconstruction error of MFAE was lower than that of the random methods. Similarly, the average reconstruction error of the EnMFAE method was significantly lower than that of the random methods. Therefore, when it comes to sparse sampling and data reconstruction of ocean salinity data, MFAE and EnMFAE methods achieve better overall reconstruction results.

[0134] refer to Figure 8 , the average reconstruction errors of two physical quantities, ocean temperature and ocean salinity, were compared. It can be seen that the maximum and minimum values ​​of the average reconstruction errors of ocean temperature and salinity achieved by the MFAE method and EnMFAE method proposed in this chapter are significantly lower than those of the random method. In addition, when the number of deployed nodes is relatively sparse (p = 10), the average value of the reconstruction error of MFAE is the lowest; and when the number of deployed nodes increases, the EnMFAE method achieves better data reconstruction results. This once again shows that the MFAE and EnMFAE methods proposed in this chapter are more suitable for sparse node deployment, and at the same time verifies the effectiveness of the MFAE method initialized with entropy weights.

[0135] Table 2 shows a numerical comparison of the mean reconstruction errors of two physical quantities, temperature and salinity, for the North Pacific region. The results show that the mean error of MFAE is consistently lower than that of the RandDE method when deploying different numbers of nodes. Specifically, MFAE achieves the lowest mean error when p = 10; at p = 20 and 30, the EnMFAE method achieves lower mean reconstruction errors than both the RandDE and MFAE methods; and at p = 40 and 50, the EnMFAE method further reduces the mean reconstruction error. These results demonstrate that when MFAE is used to select node locations, the resulting data subset achieves superior reconstruction accuracy. Furthermore, by improving the initialization strategy, such as by introducing entropy weight initialization, the reconstruction performance of MFAE can be further improved. The EnMFAE method achieves further reductions in reconstruction error for different values ​​of p, further validating the effectiveness of the entropy weight initialization strategy. In summary, MFAE and its improved methods can effectively improve the reconstruction accuracy of data from different physical quantities simultaneously.

[0136] Table 2 Comparison of mean reconstruction errors in the North Pacific region

[0137]

[0138] Table 3 compares the variance of the average reconstruction error for temperature and salinity data. The results show that, under varying node deployment numbers, the error variances of MFAE and its improved methods are generally lower than those of the RandDE method. Specifically, EnMFAE achieves the lowest variance when p = 10, 20, 30, and 50, while MFAE achieves the lowest variance when p = 40. These results demonstrate that MFAE node deployment offers improved stability, and that the entropy-based initialization strategy further improves the stability of the results.

[0139] Table 3 Comparison of variance of average reconstruction error in the North Pacific region

[0140]

[0141] (2.2) Comparison of reconstruction effects

[0142] When p=10, the reconstructed data is obtained by taking the set of decoder parameters with the smallest average error in 10 repeated experiments.

[0143] Table 4 shows that the average reconstruction error of the MFAE and EnMFAE methods is significantly lower than that of the RandDE method. Specifically, when deploying 10 nodes, the three MFAE methods proposed in this chapter achieve a minimum average reconstruction error reduction of 2.95% and 2.12%, respectively, compared to the RandDE method. These results demonstrate that the MFAE method achieves superior reconstruction accuracy for deployed nodes in both node deployment and joint reconstruction based on ocean temperature and salinity.

[0144] Table 4 Minimum mean reconstruction errors of temperature and salinity in the North Pacific when p = 10

[0145]

[0146] Arctic Ocean region results

[0147] (3.1) Comparison of reconstruction errors

[0148] refer to Figure 9 Comparisons of ocean temperature reconstruction errors show that, with varying numbers of deployed nodes, MFAE and EnMFAE generally outperform RandDE, with their average errors significantly lower. However, there is some fluctuation in the performance of MFAE and EnMFAE in reducing reconstruction error, primarily due to noise and the use of salinity data in joint feature extraction.

[0149] refer to Figure 10 , the results of comparing the ocean salinity reconstruction error show that when different numbers of nodes are deployed, the average reconstruction error of the MFAE method is lower than that of the comparative RandDE method. The EnMFAE method further reduces the average reconstruction error by introducing an entropy weight initialization strategy. This result shows that the entropy weight initialization strategy can optimize node deployment based on salinity data with large noise and improve the reconstruction accuracy (wherein, the distribution of the downloaded salinity data has large noise. Specifically, the salinity distribution in this area has large areas of non-gradual transition, and the boundaries between the distributions of different salinity areas are too neat, which does not conform to the characteristics of the actual salinity distribution. This is mainly because the Arctic Ocean region is restricted by sea ice distribution and extreme low temperature environment, there are fewer deployable salinity monitoring sensors, and the model error is large, resulting in large measurement noise and model noise of the full-state data). In addition, the MFAE method performs more stably in high-density node deployment scenarios, further verifying its superiority in reconstructing complex ocean environment data.

[0150] refer to Figure 11, the average reconstruction errors of two ocean physical quantities, ocean temperature and ocean salinity, were compared. It can be seen that the maximum and minimum values ​​of the average reconstruction errors of ocean temperature and salinity achieved by the MFAE method and EnMFAE method proposed in this chapter are significantly lower than those of the random method. In addition, EnMFAE further reduces the data reconstruction error of MFAE by introducing the entropy weight initialization strategy, and the EWMFAE method further reduces the entropy weight and noise initialization strategy by introducing the entropy weight, which shows the effectiveness of the entropy weight and noise initialization strategy. This also shows once again that the MFAE proposed in this chapter and its improved method are more suitable for sparse node deployment. Especially in the scenario of sparse node deployment, the EnMFAE method shows better adaptability and stability, and its reconstruction accuracy is higher. This shows that the MFAE proposed in this chapter and its improved method can optimize node deployment based on a variety of ocean environment data.

[0151] Referring to Table 5, a numerical comparison of the mean reconstruction errors of temperature and salinity data from the Arctic Ocean region is presented. The results show that the EnMFAE method achieves significantly lower mean errors than the RandDE method when deploying different numbers of nodes (p = 10 to p = 50). These results suggest that the entropy weight and noise initialization strategy can significantly improve the reconstruction accuracy of the data subset corresponding to the deployed nodes when selecting node locations using the MFAE method. Furthermore, the MFAE method outperforms the RandDE method for different numbers of nodes, but its mean error is slightly higher than that of the EnMFAE method. For example, when p = 20, the mean error of MFAE is 0.0061, while that of EnMFAE is 0.0055. This demonstrates that the entropy weight initialization strategy has a significant advantage in optimizing node deployment.

[0152] Table 5 Comparison of mean values ​​of average reconstruction errors in the Arctic Ocean region

[0153]

[0154] Referring to Table 6, a numerical comparison of the variance of the average reconstruction error of the temperature and salinity data in the Arctic Ocean is performed. The results also show that the error variance of the EnMFAE method is significantly lower than that of the RandDE method when different numbers of nodes are deployed (p = 10 to p = 50). For example, when p = 10, the variance of EnMFAE is 2.8428×10 -4 , which is much lower than RandDE’s 5.6460×10 -4 When p = 50, the variance of EnMFAE is further reduced to 0.7529×10 -4 , which is also significantly lower than RandDE’s 1.3573×10 -4This shows that the MFAE method proposed in this chapter has better stability in data reconstruction ability through the initialization of entropy weight when selecting node positions.

[0155] Overall, the EnMFAE method is more stable in reducing reconstruction error and variance, especially in complex marine environment monitoring (such as the high-noise environment in the Arctic Ocean). Its entropy weight initialization strategy further improves the accuracy and robustness of data reconstruction. This provides theoretical and technical support for optimizing the deployment of ocean sensing nodes in the Arctic Ocean.

[0156] Table 6 Comparison of variance of average reconstruction error in the Arctic Ocean region

[0157]

[0158] (3.2) Comparison of reconstruction effects

[0159] When p=10, after 10 repeated experiments with different methods, a set of decoder parameters with the smallest average error is selected to reconstruct the data to obtain the reconstructed data results.

[0160] Table 7 shows that the ocean reconstructions from different methods are all relatively close to the test data. However, compared to the RandDE method, the temperature and salinity data reconstructed by the MFAE and EnMFAE methods are closer to the original data distribution. Specifically, when deploying 10 nodes, the minimum average reconstruction error of the MFAE and EnMFAE methods is reduced by 5.78% and 7.71%, respectively. This demonstrates that the subset of deployed nodes obtained by the MFAE method achieves a good joint reconstruction of the two data sets, and that entropy weight initialization further improves the joint reconstruction accuracy.

[0161] Table 7 Minimum mean reconstruction error of temperature and salinity in the Arctic Ocean when p = 10

[0162]

[0163] The spatial feature index set finally obtained is the location result of the sparse deployment of the perception nodes finally obtained by the present invention.

[0164] The above experimental results show that location selection based on MFAE is effective for ocean temperature and salinity data. It not only provides a set of common sparse measurement location coordinates for different data types, but also the measurement subsets obtained by the MFAE method in sparse node deployment have lower reconstruction errors compared with randomly selected locations.

[0165] Example 3

[0166] As a second aspect of the present invention, the present application further provides a multi-physical quantity sensing node sparse deployment system that executes the deployment method described in Example 1 above, the system comprising:

[0167] Data preprocessing module, used to extract the common and effective sensing positions and position correspondence data among multiple physical quantity data;

[0168] A joint spatial position feature selection module is used to input the corresponding data into a multi-physical quantity fractal autoencoder to perform joint spatial position feature selection, and obtain a position index of a common low-dimensional sampling of the corresponding data;

[0169] A node deployment module, configured to output the location index as a deployment location of a multi-physical quantity sensing node;

[0170] The multi-physical quantity fractal autoencoder includes a plurality of fractal autoencoders corresponding to the multi-physical quantity data; the fractal autoencoder updates the position selection weights of the corresponding physical quantities, and the plurality of position selection weights constitute a position selection joint feature weight vector; the multi-physical quantity fractal autoencoder performs joint spatial position feature selection based on the position selection joint feature weight vector to obtain the position index.

[0171] As a third aspect of the present invention, the present application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned multi-physical quantity sensing node sparse deployment method. In addition to the above-mentioned processors, memories, and interfaces, any device with data processing capabilities in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0172] Example 5

[0173] As a fourth aspect of the present invention, the present application also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-mentioned multi-physical quantity sensing node sparse deployment method. The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities as described in any of the aforementioned embodiments, such as a hard disk or memory. The computer-readable storage medium may also be an external storage device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), an SD card, a flash card (Flash Card), etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and may also be used to temporarily store data that has been output or is to be output.

[0174] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A sparse deployment method for multi-physical quantity sensing nodes, characterized in that: The method comprises: Preprocess the multi-physical quantity data and extract the common effective perception positions and corresponding data among the multi-physical quantity data; Inputting the corresponding data into a multi-physical quantity fractal autoencoder to perform joint spatial position feature selection to obtain a position index of a common low-dimensional sampling of the corresponding data; Outputting the location index as the deployment location of the multi-physical quantity sensing node; Among them, the multi-physical quantity fractal autoencoder includes multiple fractal autoencoders corresponding to the multi-physical quantity data; the fractal autoencoder updates the position selection weights of the corresponding physical quantities, and the multiple position selection weights constitute a position selection joint feature weight vector; the multi-physical quantity fractal autoencoder performs joint spatial position feature selection based on the position selection joint feature weight vector to obtain the position index.

2. The sparse deployment method of multi-physical quantity sensing nodes according to claim 1 is characterized in that: The multi-physical quantity fractal autoencoder corresponds to a set of hidden variables for each physical quantity, and multiple sets of hidden variables are connected through the position selection joint feature weight vector.

3. The sparse deployment method of multi-physical quantity sensing nodes according to claim 1, characterized in that: The fractal autoencoder includes an encoder and a decoder, which respectively perform encoding and decoding operations on the corresponding physical quantity data; the encoder includes a one-to-one layer and a feature selection layer; The encoder obtains the position selection weight of the corresponding physical quantity data through a one-to-one layer, and the feature selection layer performs weighted averaging on the position selection weight to obtain a position selection joint feature weight vector and perform joint spatial position feature selection; the encoder obtains the hidden variable representation of the physical quantity based on the one-to-one layer and the feature selection layer; the decoder performs nonlinear reconstruction on the full state data based on the hidden variable representation.

4. The method for sparse deployment of multi-physical quantity sensing nodes according to claim 3, characterized in that: During the training phase, the multi-physical quantity fractal autoencoder updates the joint feature weight vector for the position selection of multiple physical quantities as follows: For each physical quantity, the position selection weights obtained in one layer are summed and averaged after each round of updating. The updated average weight is input into the feature selection layer, and the feature selection layer selects the largest p features in the weight average to update the position and select the joint feature weight matrix; The hidden variable representation is obtained based on the updated joint feature weight matrix, and the encoder and decoder parameter training is performed.

5. The method for sparse deployment of multi-physical quantity sensing nodes according to claim 4, characterized in that: The feature selection layer updates the joint feature weight matrix as follows: Among them, w i is the weight of the feature selection layer of the i-th physical quantity autoencoder; w avg is the average value of the feature selection layer weight; M is the total number of physical quantities; is the maximum p weight index vector in the weighted average The corresponding joint feature weight matrix; Diag(·) represents the construction of vector A square matrix whose elements are the main diagonal; The hidden variables are represented as: in, is the mapping function, corresponding to the mapping relationship W of the encoder of the i-th physical quantity avg_f represents the joint feature weight matrix; X i is the data of the i-th physical quantity; θ′ is the encoder parameter.

6. A sparse deployment method for multi-physical quantity sensing nodes according to claim 5, characterized in that: The optimization function of the multi-physical quantity fractal autoencoder is: in, is the global loss function; is the feature selection layer loss function; η is the balance parameter of the sub-neural network term, and λ represents the reconstruction error and the joint feature weight matrix W avg_f The balance parameter between the sparse regularization terms of ; The global loss function is expressed as follows: in, is the decoder mapping function of the i-th physical quantity; θ i is the decoder parameter; The feature selection layer loss function is expressed as follows: in, is the decoder mapping function of the i-th physical quantity; θ i is the decoder parameter; It is the joint feature weight matrix updated by the feature selection layer based on the p features with the largest weights.

7. The method for sparse deployment of multi-physical quantity sensing nodes according to claim 3, characterized in that: The encoder and decoder parts of the multi-physical quantity fractal autoencoder both adopt fully connected neural networks and are both single-layer structures.

8. The method for sparse deployment of multi-physical quantity sensing nodes according to claim 1, characterized in that: The multi-physical quantity fractal autoencoder uses the entropy weight method to initialize the weight of each spatial feature in the feature selection layer: For the j-th position feature, the entropy weight is calculated as follows: in, s kj is the normalized data, k = 1, 2, ..., m, j = 1, 2, ..., n, m and n are the number of data samples and the number of position features respectively; The average spatial characteristic entropy weight of multi-physical quantity data is calculated as follows: Among them, e i For data X i The entropy weight vector of .

9. A sparse deployment system for multiple physical quantity sensing nodes, configured to implement the sparse deployment method for multiple physical quantity sensing nodes as claimed in any one of claims 1 to 8, comprising: A data preprocessing module is used to extract the common effective sensing positions and corresponding data of the common effective positions among multiple physical quantity data; A joint spatial position feature selection module is used to input the corresponding data into a multi-physical quantity fractal autoencoder to perform joint spatial position feature selection, and obtain a position index of a common low-dimensional sampling of the corresponding data; A node deployment module, configured to output the location index as a deployment location of a multi-physical quantity sensing node; Among them, the multi-physical quantity fractal autoencoder includes multiple fractal autoencoders corresponding to the multi-physical quantity data; the fractal autoencoder updates the position selection weights of the corresponding physical quantities, and the multiple position selection weights constitute a position selection joint feature weight vector; the multi-physical quantity fractal autoencoder performs joint spatial position feature selection based on the position selection joint feature weight vector to obtain the position index.

10. A storage medium having a program stored thereon, characterized in that: When the program is executed, the sparse deployment method of multi-physical quantity sensing nodes as described in any one of claims 1 to 8 is implemented.

Images