A method for underwater sound velocity sequence classification based on deep clustering

Abstract

The application discloses a kind of underwater sound velocity sequence classification methods based on depth clustering.The method comprises the following steps: firstly, the original sound velocity curve sequence of each sea area is segmented by depth information, and a plurality of original sound velocity curve sequences of each depth area are obtained;Then, using the least square method and global optimization algorithm, each original sound velocity curve sequence is fitted by multiple linear segments, and the corresponding sound velocity segment sequence is selected by combining the original curve segment gradient deviation and the fitting threshold;Finally, the sound velocity segment sequence of each depth area is input into the corresponding depth clustering model for clustering, and the class of all sound velocity segment sequences in each depth area is obtained by clustering.Compared with the traditional clustering algorithm, the depth clustering of the present application can mine the implicit features of the data and identify more sound velocity sequence types.

Description

Technical Field

[0001] This invention belongs to the field of marine data analysis technology and relates to an underwater sound velocity sequence classification method, particularly a deep clustering-based underwater sound velocity sequence classification method. Background Technology

[0002] The variation of sound speed with depth is represented by a sound speed sequence, while the sound speed sequence of seawater reflects the structural characteristics of a local sea area in the vertical direction. Ocean sound speed plays a vital role in the accurate measurement of seawater depth and early warning of marine disasters. It is also one of the most important parameters for maritime military activities, greatly influencing the planning and execution of military operations.

[0003] Since sound speed sequences are one of the important factors affecting sound propagation in seawater, the classification of sound speed sequences has long been a focus of attention for scholars both domestically and internationally. With the rapid development of communication technology, the capabilities of data acquisition and transmission have become increasingly powerful. This has resulted in data not only having a larger scale and higher dimensionality, but also increasingly complex structural features. How to effectively extract feature information from massive datasets and improve the classification effect of sound speed sequences is an urgent problem to be solved.

[0004] Existing intelligent sound velocity sequence classification methods, such as fuzzy C-means clustering algorithm and self-organizing neural network, have all demonstrated their feasibility. However, the datasets studied are small and the types of sound velocity sequences are limited. When applied to larger sea areas, the classification results tend to identify the sound velocity sequence types that account for a large proportion, while the identification results for sound velocity sequence types that account for a certain proportion but are smaller in size are unsatisfactory. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a deep clustering-based underwater sound velocity sequence classification method.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] 1) Segment each original sound velocity curve sequence in the sea area using depth information to obtain multiple original sound velocity curve sequences for each depth zone;

[0008] 2) Using the least squares method and global optimization algorithm, multi-segment linear fitting is performed on each original sound speed curve sequence. Combining the original curve segment gradient difference and fitting threshold, the corresponding sound speed line segment sequences are selected respectively.

[0009] 3) The sound velocity line segment sequences of each depth region are input into the corresponding depth clustering model for clustering, and the clustering obtains the category of all sound velocity line segment sequences in each depth region.

[0010] In step 1), the original sound velocity curve sequence for the same depth region is also subjected to depth normalization to obtain the normalized sound velocity curve sequence for the current depth region. The normalization formula is as follows:

[0011] C i ′=λ×C i

[0012] Z i ′=λ×Z i

[0013] λ=Z * / Z max

[0014] Among them, C i ′ represents the normalized sound velocity curve sequence corresponding to the i-th original sound velocity curve sequence in the current depth region, C i Z represents the i-th original sound velocity curve sequence in the current depth region. i Z' represents the normalized depth of the i-th original sound velocity curve sequence in the current depth region. i Z represents the original depth of the i-th original sound velocity curve sequence in the current depth region, λ represents the depth normalization coefficient, and Z represents the depth normalization coefficient. * Z is the depth threshold of the current depth region. max This represents the maximum depth at which the original sound velocity curve sequence was acquired within the current depth region.

[0015] In step 2), for each original sound speed curve sequence, multiple linear fittings are performed using the least squares method and the global optimization algorithm to obtain multiple sound speed line segment sequences with different numbers of fitting segments. Based on the original curve segment gradient deviation and the fitting threshold, the multiple sound speed line segment sequences with different numbers of fitting segments are screened to obtain the optimal sound speed line segment sequence.

[0016] The fitting threshold is the sum of the squared residuals between the sound velocity line segment sequence and the original sound velocity curve sequence;

[0017] The formula for calculating the sum of the piecewise gradient deviations of the original curve is as follows:

[0018] V n (n|0,z1,…,z n-1 ,z max ) = V n (n|0,z1)+…+V n (n|z n-1 ,z max )

[0019]

[0020]

[0021] Among them, Vn (n|0,z1) represents the sum of gradient deviations from depth 0 to z1 when divided into n segments, where n represents the number of segments in the sound velocity line segment sequence, and x k Let v represent the gradient sequence from segment point i to segment point j. ij Let x represent the gradient variation from piecewise point i to piecewise point j. ij Let z represent the mean gradient from piecewise point i to piecewise point j, where i,j = 0, z1, ..., z2. n-1 ,z max , z n-1 z represents the sea area depth at segment point n-1. max This indicates the maximum depth at which the original sound velocity curve sequence was acquired within the current depth region.

[0022] In step 2), when the sum of the gradient deviations of the original curve segments corresponding to the sound speed line segment sequence with the fewest segments is less than 2, the sound speed line segment sequence with the fewest segments is the optimal sound speed line segment sequence; otherwise, the sound speed line segment sequence that satisfies the convergence condition is taken as the optimal sound speed line segment sequence.

[0023] The convergence condition is that the sum of the gradient deviations of the original curve segments corresponding to the sound velocity line segment sequence under the current fitting segment number is less than 0.5 times the sum of the gradient deviations of the original curve segments corresponding to the sound velocity line segment sequence corresponding to the current fitting segment number minus one, and is greater than 2 times the sum of the gradient deviations of the original curve segments corresponding to the sound velocity line segment sequence corresponding to the current fitting segment number plus one. At the same time, the fitting threshold of the sound velocity line segment sequence under the current fitting segment number is also less than 0.5 times the fitting threshold of the sound velocity line segment sequence corresponding to the current fitting segment number minus one.

[0024] In step 3), the deep clustering model is a deep convolutional embedding clustering model.

[0025] The beneficial effects of this invention are:

[0026] 1) This invention takes into account the jitter in the sound speed sequence due to errors such as measurement and calculation approximation. It uses the least squares method and global optimization algorithm to fit the sound speed sequence. The optimal fitting sequence is selected by the combined effect of the original curve segment gradient deviation and the fitting threshold, which improves the overfitting phenomenon that is easy to occur due to improper threshold selection in the original method.

[0027] 2) To address the challenge of designing and calculating sound velocity sequence features due to the large scale of clustered data, this invention utilizes neural networks to automatically extract internal features and cluster divisions of sound velocity sequences. It maps the input data to a hidden feature space and iteratively optimizes the reconstruction loss and clustering loss. Compared to traditional clustering algorithms, deep clustering can uncover the hidden features of the data and identify more types of sound velocity sequences. Attached Figure Description

[0028] Figure 1It is the flow chart of the underwater sound speed sequence classification method based on deep clustering proposed by the present invention.

[0029] Figure 2 It is the sea depth data of a certain sea area simulated by the present invention.

[0030] Figure 3 It is the schematic diagram of the improvement of linear fitting of the present invention.

[0031] Figure 4 It is the deep clustering model structure of the present invention.

[0032] Figure 5 It is the flow chart of the deep clustering algorithm of the present invention.

[0033] Figure 6 It is the schematic diagram of the clustering center results of the traditional clustering algorithm and the method of the present invention under different numbers of clusters. Detailed implementation mode

[0034] To make the purpose, technical solution and advantages of the implementation of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in more detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0035] The present invention is an underwater sound speed sequence classification method based on deep clustering. It constructs a sound speed sequence by using marine environmental information, fits the sound speed sequence based on the gradient deviation sum and threshold, inputs the gradient of the linear fitting result into a deep clustering model, and obtains the classification results of each sample of the sound speed sequence in the research sea area after model training.

[0036] As Figure 1 shown, the present invention includes the following steps:

[0037] 1) Perform deep information segmentation on each original sound speed curve sequence in the sea area to obtain multiple original sound speed curve sequences in each depth zone; the area where the depth z of the sea area satisfies z ≤ 200m is denoted as the shallow sea area, the area where the depth z of the sea area satisfies 200m < z ≤ 1000m is denoted as the transition area, and the area where the depth z of the sea area satisfies z > 1000m is denoted as the deep sea area.

[0038] In 1), for the convenience of subsequent neural network input of equal-length sequences and easy display of clustering results, the original sound speed curve sequences in the same depth zone are also depth-normalized to obtain the normalized sound speed curve sequences in the current depth zone. The normalization formula is as follows:

[0039] C i ′ = λ × C i

[0040] Z i ′ = λ × Z i

[0041] λ=Z * / Z max

[0042] Among them, C i ′ represents the normalized sound velocity curve sequence corresponding to the i-th original sound velocity curve sequence in the current depth region, C i Z represents the i-th original sound velocity curve sequence in the current depth region. i Z' represents the normalized depth of the i-th original sound velocity curve sequence in the current depth region. i Z represents the original depth of the i-th original sound velocity curve sequence in the current depth region, λ represents the depth normalization coefficient, and Z represents the depth normalization coefficient. * This is the depth threshold for the current depth region, set to 200, 1000, and 2000 respectively along the depth direction. Z max This represents the maximum depth at which the original sound velocity curve sequence was acquired within the current depth region.

[0043] 2) Using the least squares method and global optimization algorithm, multi-segment linear fitting is performed on each original sound speed curve sequence. Combining the original curve segment gradient difference and fitting threshold, the corresponding sound speed line segment sequences are selected respectively.

[0044] To eliminate interference caused by errors in measurement and calculation, and to identify the general trend of the sound velocity sequence, and because a fixed threshold cannot well adapt to different curves, in order to avoid overfitting, the original curve segment gradient deviation and fitting threshold are set. In step 2), for each original sound velocity curve sequence, multiple linear fittings are performed using the least squares method and a global optimization algorithm to obtain multiple sound velocity line segment sequences with different numbers of fitting segments. Based on the original curve segment gradient deviation and fitting threshold, the multiple sound velocity line segment sequences with different numbers of fitting segments are filtered to obtain the optimal sound velocity line segment sequence.

[0045] The fitting threshold is the sum of the squared residuals between the sound velocity line segment sequence and the original sound velocity curve sequence;

[0046] The formula for calculating the sum of piecewise gradient deviations of the original curve is as follows:

[0047] V n (n|0,z1,…,z n-1 ,z max ) = V n (n|0,z1)+…+V n (n|z n-1 ,z max )

[0048]

[0049]

[0050] Among them, V n (n|0,z1) represents the sum of gradient deviations from depth 0 to z1 when divided into n segments, where n represents the number of segments in the sound velocity line segment sequence, and x k Let v represent the gradient sequence from segment point i to segment point j. ij Let x represent the gradient variation from piecewise point i to piecewise point j. ij Let z represent the mean gradient from piecewise point i to piecewise point j, where i,j = 0, z1, ..., z2. n-1 ,z max , z n-1 z represents the normalized sea area depth at segment point n-1. max This indicates the maximum depth at which the original sound velocity curve sequence was acquired within the current depth region.

[0051] In step 2), when the sum of the gradient deviations of the original curve segments corresponding to the sound speed line segment sequence with the fewest segments is less than 2, the sound speed line segment sequence with the fewest segments is the optimal sound speed line segment sequence; otherwise, the sound speed line segment sequence that satisfies the convergence condition is taken as the optimal sound speed line segment sequence.

[0052] The convergence condition is that the sum of the gradient deviations of the original curve segments corresponding to the sound velocity line segment sequence under the current fitting segment number is less than 0.5 times the sum of the gradient deviations of the original curve segments corresponding to the sound velocity line segment sequence corresponding to the current fitting segment number minus one, and is greater than 2 times the sum of the gradient deviations of the original curve segments corresponding to the sound velocity line segment sequence corresponding to the current fitting segment number plus one. At the same time, the fitting threshold of the sound velocity line segment sequence under the current fitting segment number is also less than 0.5 times the fitting threshold of the sound velocity line segment sequence corresponding to the current fitting segment number minus one.

[0053] Figure 3 (a)-(d) respectively correspond to Figure 3 The results (e)-(f) show the gradient deviation and threshold values ​​of the sound velocity sequences under different numbers of fitting segments. For the two example sound velocity sequences, when the number of fitting segments is 4, the gradient deviation and threshold values ​​tend to converge. The fitting results can characterize the general trend of the sound velocity sequences and simplify the original sound velocity sequences.

[0054] 3) The sound velocity line segment sequences of each depth region are input into the corresponding depth clustering model for clustering. That is, multiple sound velocity line segment sequences of the same depth region are input into the same depth clustering model. The number of depth clustering models is the same as the number of depth regions. Clustering obtains the category of all sound velocity line segment sequences in each depth region.

[0055] In section 3), the deep clustering model is the Deep Convolution Embedded Clustering (DCEC) model. The overall network structure consists of three parts: an encoder, a decoder, and a clustering module, such as... Figure 4 As shown.

[0056] Given that the input samples are in sequence form, the encoder uses a one-dimensional convolutional stacked structure to extract features, which are then fed into both the decoder and the clustering module. The decoder is responsible for restoring the original data from the intermediate features as much as possible.

[0057] The difference between the encoder input and the decoder output is defined as the reconstruction error L. r This forces the local structure of the sample in space to remain consistent with the original space. The calculation method is as follows:

[0058]

[0059] Where, x a For input data, y a The output data is m, where m is the number of samples.

[0060] The clustering model performs soft distribution based on the features extracted by the encoder, with a clustering loss L. c The relative entropy between the soft-label distribution (similarity between the low-dimensional space and the cluster centers) and the target distribution, also known as the Kullback-Leibler divergence, is used to measure the degree of matching between the two distributions. It is calculated as follows:

[0061]

[0062] Where, q ab For the a-th sample q a z obtained after the encoding layer a The probability of belonging to cluster center μ of class b represents the similarity between the low-dimensional space and the cluster center, and is calculated as follows:

[0063]

[0064] p ab To assist the target distribution function, and by using q ab Increasing the value to the quadratic level improves clustering accuracy.

[0065]

[0066] The overall loss function L of the model is:

[0067] L = L r +γL c

[0068] Here, γ controls the degree of distortion in the embedding layer. To prevent distortion of the embedding space, the clustering loss and reconstruction loss are combined as a loss function for joint optimization.

[0069] Figure 5This is a flowchart illustrating the training process of a deep clustering model. The input is a fitted sequence of sound velocity gradients. After extraction by the encoder, the low-dimensional features are used in two ways: firstly, the decoder reconstructs the sequence, and the difference between the reconstructed sequence and the original sequence is the reconstruction error; secondly, clustering is performed to obtain the clustering error. The training process first pre-trains the encoder and decoder based on the reconstruction error, and then iteratively fine-tunes the entire model using the reconstruction error and the clustering error. When the entire model converges, the classification results for each sample are obtained.

[0070] Simulation Description: To verify the effectiveness of the underwater sound velocity sequence classification method based on deep clustering, a simulation analysis of sound velocity sequence classification in a shallow sea area is conducted below. The selected sea area has longitudes of 109°E-140°E and latitudes of 11°N-40°N. Figure 2 The image shows the depth of this sea area. Since only the shallow sea area is analyzed, the depth of the deep sea area is represented by 1000m. The depth of all nodes in the shallow sea area is normalized to 200m. Based on the gradient deviation and threshold, the optimal fitting strategy for the sound velocity sequence is obtained through the least squares method and a global optimization algorithm. The deep clustering model is trained until it converges. The model outputs the membership of each sample, and the mean of the sound velocity sequence in each class is taken as the center curve of each cluster.

[0071] Figure 6 (a), (c), and (e) represent the cluster centers of the traditional k-means clustering algorithm when the number of clusters is 5, 20, and 40, respectively. Figure 6 (b), (d), and (f) represent the cluster centers of the deep clustering algorithm when the number of clusters is 5, 20, and 40, respectively. Since the 40% curve in the shallow sea area shows an overall negative gradient trend and the 40% curve shows a weak positive gradient trend, these two trends account for the vast majority of the clusters. However, it can still be found that, unlike the traditional clustering results, when the number of clusters is 5, one of the negative gradient trends in the deep clustering algorithm is different, showing a trend of "weak positive gradient + negative gradient + equal gradient + negative gradient"; when the number of clusters is 20, one cluster center is a strong positive gradient; when the number of clusters is 40, in addition to the cluster center with a strong positive gradient, there is another cluster center showing a trend of "weak positive gradient + positive gradient + strong negative gradient". After checking the original sound speed sequence dataset, the above three types of sound speed sequences all have a certain proportion, but compared with the two types of sound speed sequences that account for 40%, the proportion of the above three types is smaller.

[0072] Overall, the deep clustering algorithm of this invention performs well in classifying sound velocity sequences in shallow sea areas in this example, especially in its ability to identify classes that have a certain proportion but a small proportion, which is beneficial to the study of the completeness of sound velocity sequence types in the sea area.

[0073] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for classifying underwater sound velocity sequences based on deep clustering, characterized in that, Includes the following steps: 1) Segment each original sound velocity curve sequence in the sea area using depth information to obtain multiple original sound velocity curve sequences for each depth zone; 2) Using the least squares method and global optimization algorithm, multi-segment linear fitting is performed on each original sound speed curve sequence. Combining the original curve segment gradient difference and fitting threshold, the corresponding sound speed line segment sequences are selected respectively. For each original sound speed curve sequence, multiple linear fittings are performed using the least squares method and global optimization algorithm to obtain multiple sound speed line segment sequences with different numbers of fitting segments. Based on the original curve segment gradient difference and fitting threshold, the multiple sound speed line segment sequences with different numbers of fitting segments are screened to obtain the optimal sound speed line segment sequence. The fitting threshold is the sum of the squared residuals between the sound velocity line segment sequence and the original sound velocity curve sequence; The formula for calculating the sum of the piecewise gradient deviations of the original curve is as follows: in, This indicates that when divided into n segments, the depth from 0 to... The gradient deviations and sums, The number of line segments representing the speed of sound in the line segment sequence. This represents the gradient sequence from segment point i to segment point j. This represents the gradient variation from piecewise point i to piecewise point j. This represents the mean gradient from piecewise point i to piecewise point j. Indicates the segmentation point The depth of the sea area, This indicates the maximum depth at which the original sound velocity curve sequence was acquired within the current depth region; 3) The sound velocity line segment sequences of each depth region are input into the corresponding depth clustering model for clustering, and the clustering obtains the category of all sound velocity line segment sequences in each depth region.

2. The underwater sound velocity sequence classification method based on deep clustering according to claim 1, characterized in that, In step 1), the original sound velocity curve sequence for the same depth region is also subjected to depth normalization to obtain the normalized sound velocity curve sequence for the current depth region. The normalization formula is as follows: in, This is the normalized sound velocity curve sequence corresponding to the i-th original sound velocity curve sequence in the current depth region. This represents the i-th original sound velocity curve sequence in the current depth region. Let i be the normalized depth of the i-th original sound velocity curve sequence in the current depth region. This represents the original depth of the i-th original sound velocity curve sequence in the current depth region. Represents the depth normalization coefficient. The depth threshold for the current depth region. This represents the maximum depth at which the original sound velocity curve sequence was acquired within the current depth region.

3. The underwater sound velocity sequence classification method based on deep clustering according to claim 1, characterized in that, In step 2), when the sum of the gradient deviations of the original curve segments corresponding to the sound speed line segment sequence with the fewest segments is less than 2, the sound speed line segment sequence with the fewest segments is the optimal sound speed line segment sequence; otherwise, the sound speed line segment sequence that satisfies the convergence condition is taken as the optimal sound speed line segment sequence.

4. The underwater sound velocity sequence classification method based on deep clustering according to claim 3, characterized in that, The convergence condition is that the sum of the gradient deviations of the original curve segments corresponding to the sound velocity line segment sequence under the current fitting segment number is less than 0.5 times the sum of the gradient deviations of the original curve segments corresponding to the sound velocity line segment sequence corresponding to the current fitting segment number minus one, and is greater than 2 times the sum of the gradient deviations of the original curve segments corresponding to the sound velocity line segment sequence corresponding to the current fitting segment number plus one. At the same time, the fitting threshold of the sound velocity line segment sequence under the current fitting segment number is also less than 0.5 times the fitting threshold of the sound velocity line segment sequence corresponding to the current fitting segment number minus one.

5. The underwater sound velocity sequence classification method based on deep clustering according to claim 1, characterized in that, In step 3), the deep clustering model is a deep convolutional embedding clustering model.

Images