Smuggling ship abnormal trajectory identification method and device based on multi-source information fusion

Abstract

The application discloses a kind of based on the abnormal trajectory identification method of smuggling ship of multi-source information fusion, obtains ship track sequence, the track sequence includes motion characteristics and weather characteristics;Motion characteristics and weather characteristics in ship track sequence are respectively extracted by TCN network with additive position coding;The motion characteristics and weather characteristics extracted are fused, and the feature vector after fusion is obtained;The feature vector after fusion is input into the LSTM network with attention mechanism for further processing, and the multi-head attention output is obtained;The multi-head attention output is input into Kan network, the Kan network is carried out nonlinear feature mapping by adaptive grid point and B spline function, and the prediction result that track is not suspected smuggling ship abnormal trajectory is output.

Description

Technical Field

[0001] The present invention belongs to the field of text data detection and deep learning technology, and specifically relates to a method and device for identifying abnormal trajectories of smuggling ships based on multi-source information fusion. Background Art

[0002] Smuggling vessels are a common maritime transport tool used in smuggling activities. They are large, modified motorboats. When traveling at high speeds, they appear to be flying close to the surface from a distance, hence their common name. Each smuggling vessel is typically equipped with multiple engines with up to 250 horsepower, reaching a top speed of over 90 kilometers per hour. Smugglers frequently use these vessels for smuggling. As smuggling tools, smuggling vessels pose multiple risks. First, smuggling activities may involve contraband, the entry of which poses a serious threat to social security and national stability. Second, smuggling activities bypass normal customs and import procedures, resulting in the loss of national tax revenue and disrupting economic order. Large-scale smuggling can also undermine fair competition in the domestic market and impact legitimate industries. Finally, smuggling vessels travel at extremely high speeds, easily causing maritime accidents and endangering the lives of legitimate vessels and crew members. Therefore, timely detection of smuggling activities is essential.

[0003] Deep learning-based methods for detecting abnormal ship trajectory behavior have been widely studied in recent years due to their proven effectiveness and computational speed. However, these studies have only incorporated basic ship trajectory datasets, such as AIS data. While these methods are effective for detecting other types of anomalies, smuggling activities are often covert and evasive, often occurring late at night and in poor weather conditions. If the detection model fails to account for the impact of external factors on ship behavior and only uses a few simple ship motion features to determine whether a ship's trajectory is suspected of smuggling, it can easily lead to a large number of ships with similar behavior being mistakenly reported as smuggling vessels.

[0004] Due to the influence of multiple factors, such as ship type, voyage purpose, and sea conditions, real ship trajectory sequences are often highly complex and diverse. Ship motion patterns not only fluctuate significantly but also change significantly over time and in response to environmental changes. This complex dynamic nature makes it difficult to fully characterize trajectory variations using simple linear transformations and nonlinear activation functions. Furthermore, it is difficult to fully capture the underlying characteristics of ship behavior and its potential spatiotemporal dependencies. This results in limited performance in complex real-world situations and an inability to accurately identify narrowly defined potential abnormal behaviors, particularly when dealing with the highly covert and flexible nature of smuggling activities. Summary of the Invention

[0005] In view of this, the main purpose of the present invention is to provide a method and device for identifying abnormal trajectories of smuggling ships based on multi-source information fusion.

[0006] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0007] An embodiment of the present invention provides a method for identifying abnormal trajectories of smuggling vessels based on multi-source information fusion. The method comprises:

[0008] Acquire a ship trajectory sequence, wherein the trajectory sequence includes motion characteristics and weather characteristics;

[0009] The motion features and weather features in the ship trajectory sequence are extracted respectively through the TCN (Temporal Convolutional Network) network with additive position encoding;

[0010] The extracted motion features and weather features are fused to obtain a fused feature vector;

[0011] The fused feature vector is input into an LSTM (Long Short-Term Memory) network with an attention mechanism for further processing to obtain a multi-head attention output.

[0012] The multi-head attention output is input into the Kan network, which performs nonlinear feature mapping through adaptive grid points and B-spline functions, and outputs a prediction result on whether the trajectory is suspected to be an abnormal trajectory of a smuggling ship.

[0013] In the above scheme, the motion features and weather features in the ship trajectory sequence are extracted respectively by using a TCN (Temporal Convolutional Network) network with additive position coding, specifically comprising: using the TCN network to extract the motion features X in the trajectory sequence X M Perform encoding operation to obtain the trajectory vector X' with position encoding M ;

[0014] The trajectory vector X' with position encoding is processed by TCN network M and weather characteristics X W Perform feature extraction respectively to obtain feature map U M and U W .

[0015] In the above scheme, the motion feature X in the trajectory sequence X is analyzed by the TCN network. M Perform encoding operation to obtain the trajectory vector X' with position encoding M , specifically including:

[0016] X M Projected to a higher dimension d pos ;

[0017] The mathematical expression based on position encoding is: Determine the trajectory vector X' with position encoding M =X M +PE; where pos represents the word vector position of the trajectory data, max_len represents the maximum sequence length of all ship trajectory sequences, d pos represents the embedding dimension of the position encoding, 2i and 2i+1 represent the even and odd dimensions in the ship trajectory vector, which are calculated using the sin and cos functions respectively.

[0018] In the above scheme, the extracted motion features and weather features are fused to obtain the fused feature vector, specifically including: according to U={u1,u2,…,u T}=concat(U M ,U W ) to obtain the fused feature vector U, where T is the length of the merged vector and also the length of the trajectory sequence.

[0019] In the above scheme, the fused feature vector is input into the LSTM (Long Short-Term Memory) network with attention mechanism for further processing to obtain multi-head attention output, specifically including: at each time step t, receiving the input u of the current time step t , combined with the hidden state h of the previous time step t-1 and cell state c t-1 , determine the hidden state h of the current time step t and cell state c t ;

[0020] Traverse the entire trajectory sequence and generate the hidden state sequence H = {h1,h2,…,h T},in d h Represents the dimension of the LSTM hidden layer;

[0021] Performing a linear transformation on the hidden state sequence H to generate a query vector Q, a key vector K, and a value vector V;

[0022] According to the query vector Q corresponding to each head i and the transpose of the key vector Determine the similarity and normalize it through the softmax function to determine the attention weight α of each head i;

[0023] According to the attention weight α of each head i The sum value vector V determines the context vector C of each head i =α i *V i ;

[0024] The result of concatenating the context vectors of all heads is transformed into the multi-head attention output S through the linear transformation matrix, S = W O [C1; C2; …; C H ], where W O is the output linear transformation matrix used to merge information from different attention heads.

[0025] In the above scheme, at each time step t, the input u of the current time step is received t , combined with the hidden state h of the previous time step t-1 and cell state c t-1 , determine the hidden state h of the current time step t and cell state c t , specifically including: according to h t ,c t =LSTM(u t ,h t-1 ,c t-1 ) Determine the hidden state h of the current time step t and cell state c t .

[0026] In the above scheme, the query vector Q corresponding to each head i and the transpose of the key vector Determine the similarity and normalize it through the softmax function to determine the attention weight α of each head i , specifically including: Determine the attention weight α for each head i ; where d K is the dimension of the key vector.

[0027] In the above scheme, the multi-head attention output is input into the Kan network, which performs nonlinear feature mapping through adaptive grid points and B-spline functions, and outputs a prediction result on whether the trajectory is suspected to be an abnormal trajectory of a smuggling ship, specifically including:

[0028] The multi-head attention output S is passed through multiple layers of Kan transformation layers to obtain the final output of whether the trajectory sequence is suspected of smuggling. The main operation of each layer of Kan linear transformation layer is expressed as where h (l+1) represents the output of the l+1th layer, represents the weight matrix of the linear transformation of layer l, Represents the weight matrix of the B-spline of the lth layer, σ is the activation function, B(h (l) ) represents the interpolation function of B-spline.

[0029] In the above solution, the B-spline basis function is calculated recursively, and the recursive formula is: Among them, t i represents the i-th grid point, which determines the segmentation position of the basis spline function; k is the spline order; B i,0 (h (l) ) represents the piecewise constant, when h (l) When it is within the corresponding interval, its value is 1, otherwise it is 0;

[0030] The B-spline interpolation function is a function generated by a linear combination of a set of control points and the corresponding B-spline basis functions. Its general form is: B(h (l) )=Σ i＝0 P i B i,k (h (l) ); where P i Represents the control points. Different control points generate B-spline interpolation functions of different shapes.

[0031] The embodiment of the present invention further provides a device for identifying abnormal trajectories of smuggling vessels based on multi-source information fusion, the device comprising:

[0032] A sequence acquisition module is used to obtain a ship trajectory sequence, wherein the trajectory sequence includes motion characteristics and weather characteristics;

[0033] The TCN network module is used to extract motion features and weather features from the ship trajectory sequence respectively; it is also used to fuse the extracted motion features and weather features to obtain a fused feature vector;

[0034] LSTM network module, used to further process the fused feature vector to obtain multi-head attention output;

[0035] The Kan network module is used to input the multi-head attention output into the Kan network. The Kan network performs nonlinear feature mapping through adaptive grid points and B-spline functions, and outputs a prediction result on whether the trajectory is suspected to be an abnormal trajectory of a smuggling ship.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The method proposed in the present invention first uses a parallel TCN network to extract motion features and meteorological features from multi-source information respectively, and then inputs these features into the LSTM for trajectory identification after decision-making fusion. During this process, the present invention proposes a TCN network based on additive position coding, which enables the model to more accurately understand the global temporal relationship of the trajectory sequence during the feature extraction stage; at the same time, the present invention introduces the Kan module to replace the more commonly used fully connected network in traditional deep learning networks to solve the problem of complex trajectory sequences being difficult to accurately represent. The present invention fully considers the characteristics of abnormal trajectories of smuggling ships in actual scenarios, and improves the recognition performance of detection algorithms and devices when facing smuggling activities that are both covert and flexible. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings described herein are used to further understand the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0039] Figure 1 A flowchart of a method for identifying abnormal trajectories of smuggling vessels based on multi-source information fusion is provided for an embodiment of the present invention;

[0040] Figure 2 The present invention provides an overall system block diagram corresponding to a method for identifying abnormal trajectories of smuggling vessels based on multi-source information fusion.

[0041] Figure 3 The present invention provides an embodiment of a method for identifying abnormal smuggling ship trajectories based on multi-source information fusion, and provides a detection effect diagram of each model in a real-world ship trajectory scenario; DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0043] The embodiment of the present invention provides a method for identifying abnormal trajectories of smuggling ships based on multi-source information fusion. Figure 1 、 2 As shown, the method is:

[0044] Step 101: Acquire a ship trajectory sequence, wherein the trajectory sequence includes motion characteristics and weather characteristics;

[0045] Step 102: Extract motion features and weather features from the ship trajectory sequence using a TCN (Temporal Convolutional Network) network with additive position encoding.

[0046] Specifically, the motion feature X in the trajectory sequence X is analyzed by the TCN network. M Perform encoding operation to obtain the trajectory vector X' with position encoding M ;

[0047] The trajectory vector X' with position encoding is processed by TCN network M and weather characteristics X W Perform feature extraction respectively to obtain feature map U M and U W .

[0048] More specifically, the motion features X in the trajectory sequence X are analyzed by the TCN network. M Perform encoding operation to obtain the trajectory vector X' with position encoding M , specifically including:

[0049] X M Projected to a higher dimension d pos ;

[0050] The mathematical expression based on position encoding is: Determine the trajectory vector X' with position encoding M =X M +PE; where pos represents the word vector position of the trajectory data, max_len represents the maximum sequence length of all ship trajectory sequences, d pos represents the embedding dimension of the position encoding, 2i and 2i+1 represent the even and odd dimensions in the ship trajectory vector, which are calculated using the sin and cos functions respectively.

[0051] After X' M and X W The operation is consistent with X W For example, TCN performs temporal convolution on the input trajectory sequence through multiple layers of residual convolution blocks, where the overall operation of the residual convolution block can be expressed as: Conv res,i =f(W i *f(W i-1 *…f(W1*X W +b1)…+b i-1 )+b i )+downsample(X W); where W is the convolution kernel, b is the bias, and i represents the number of layers of the residual convolution block. The model continuously adjusts and optimizes these two parameters by learning the trajectory features of different labels to capture the multi-level temporal characteristics of the smuggling ship trajectory. i represents the index of the residual convolution block, downsample represents the downsampling operation, which is used to implement the residual link, and f(·) is the activation function. This paper adopts the ELU activation function to improve the recognition sensitivity of abnormal trajectories. Dilated convolution is used in TCN. By setting the dilation factor to expand the receptive field, it can capture the dependencies between trajectory sequences at different time steps. The specific manifestation of the residual connection structure is the addition of the input and the output after multiple convolution operations. This structure ensures that information can be effectively transmitted in a deeper network, which helps the deep network model capture complex spatiotemporal trajectory patterns in the process of smuggling ship detection.

[0052] The final TCN module output is the motion feature X M and weather characteristics X W The corresponding feature map U M and U W , which is the output of the last layer of residual convolutional block. The feature map is the core basis for the model to identify the trajectory pattern of smuggling ships, and contains rich spatiotemporal information related to the characteristics of smuggling activities.

[0053] First, the Transformer's positional encoding calculations are performed over the entire length of the sequence, and the scale of the positional encoding depends on the sequence length and embedding dimension. However, the positional encoding used in this model limits the maximum sequence length, which scales the calculation range and better accommodates the local convolutional receptive field of the TCN network. This makes it more suitable for local and global temporal information analysis of ship trajectory data. Secondly, the positional encoding used by models such as the Transformer directly encodes at a fixed frequency. However, this model performs additional scaling on the outputs of the sine and cosine functions to match the encoding scale with the numerical range of the input features, making it more suitable for models such as the TCN that rely on local convolutional receptive fields. This allows the model to better fit the spatiotemporal characteristics of the smuggling vessel trajectory detection task, assisting the TCN in capturing local and global trajectory sequence features.

[0054] Step 103: Fusing the extracted motion features and weather features to obtain a fused feature vector;

[0055] Specifically, according to U={u1,u2,…,u T}=concat(U M ,U W ) to obtain the fused feature vector U, where T is the length of the merged vector and also the length of the trajectory sequence.

[0056] Step 104: Input the fused feature vector into an LSTM (Long Short-Term Memory) network with an attention mechanism for further processing to obtain a multi-head attention output;

[0057] Specifically, at each time step t, the input u of the current time step is received t , combined with the hidden state h of the previous time step t-1 and cell state c t-1 , determine the hidden state h of the current time step t and cell state c t ;

[0058] Traverse the entire trajectory sequence and generate the hidden state sequence H = {h1,h2,…,h T},in d h Represents the dimension of the LSTM hidden layer;

[0059] Performing a linear transformation on the hidden state sequence H to generate a query vector Q, a key vector K, and a value vector V;

[0060] According to the query vector Q corresponding to each head i and the transpose of the key vector Determine the similarity and normalize it through the softmax function to determine the attention weight α of each head i ;

[0061] According to the attention weight α of each head i The sum value vector V determines the context vector C of each head i =α i *V i ;

[0062] The result of concatenating the context vectors of all heads is transformed into the multi-head attention output S through the linear transformation matrix, S = W O [C1; C2; …; C H ], where W O is the output linear transformation matrix used to merge information from different attention heads.

[0063] The calculation process of the context vector helps extract key features related to potential smuggling activities so that the model can better understand and predict trajectory changes. The key to the multi-head attention mechanism is the parallel processing of multiple heads. Each head independently calculates the attention weight and context vector. This parallel calculation allows the model to analyze trajectory data from multiple angles and improve recognition accuracy. The context vectors of all heads will be spliced ​​together to form a richer attention context representation. The spliced ​​result is obtained through a linear transformation matrix to obtain the multi-head attention output S.

[0064] More specifically, at each time step t, the input u of the current time step is received t , combined with the hidden state h of the previous time step t-1 and cell state c t-1 , determine the hidden state h of the current time step t and cell state c t , specifically including: according to h t ,c t =LSTM(u t ,h t-1 ,c t-1 ) Determine the hidden state h of the current time step t and cell state c t .

[0065] Performing a linear transformation on the hidden state sequence H to generate a query vector Q, a key vector K, and a value vector V, specifically includes: filtering out features closely related to smuggling track identification from the hidden state sequence generated by LSTM, which is expressed by the following formula:

[0066] Q=W Q H,K=W K H,V=W V H

[0067] Where W Q ,W K ,W V Both are linear transformation matrices used to project H into the space of queries, keys, and values.

[0068] The query vector Q corresponding to each head i and the transpose of the key vector Determine the similarity and normalize it through the softmax function to determine the attention weight α of each head i , specifically including: Determine the attention weight α for each head i ; where d K is the dimension of the key vector.

[0069] Step 105: The multi-head attention output is input into the Kan network, which performs nonlinear feature mapping through adaptive grid points and B-spline functions, and outputs a prediction result on whether the trajectory is suspected to be an abnormal trajectory of a smuggling ship.

[0070] Specifically, the multi-head attention output S is passed through multiple layers of Kan transformation layers to obtain the final output of whether the trajectory sequence is suspected of smuggling. The main operation of each layer of Kan linear transformation layer can be expressed as: where h (l+1) represents the output of the l+1th layer, represents the weight matrix of the linear transformation of layer l, Represents the weight matrix of the B-spline of the lth layer, σ is the activation function, B(h (l) ) represents the interpolation function of B-spline, h (0) is the multi-head attention output S. The first half of the formula is the basic linear transformation, and the second half is the B-spline transformation. Through the combination of linear and spline interpolation, the model can capture the global and local characteristics of trajectory data, as well as the differences between normal and smuggling trajectories, thereby improving the ability to identify potential smuggling activities.

[0071] For the spline interpolation part, the key step is to calculate the B-spline basis functions. These basis functions define the distribution of input features at each grid point, thereby helping the model better model complex trajectory behavior. The calculation method of the B-spline basis functions is recursive, and the recursive formula is: Among them, t i represents the i-th grid point, which determines the segmentation position of the basis spline function; k is the spline order; B i,0 (h (l) ) represents the piecewise constant, when h (l) When it is within the corresponding interval, its value is 1, otherwise it is 0;

[0072] The B-spline interpolation function is a function generated by a linear combination of a set of control points and the corresponding B-spline basis functions. This interpolation method enables the model to maintain high flexibility and accuracy when processing complex trajectory data, thereby improving the ability to identify potential smuggling activities. Its general form is: B(h (l) )=∑ i＝0 P i B i,k (h (l) ); where P i Represents the control points. Different control points generate B-spline interpolation functions of different shapes.

[0073] Output of Kan network It is the probability representation of the output of the last Kan linear layer after passing through the softmax function, that is, the final probability prediction of whether the trajectory sequence is suspected of smuggling by the model proposed in this invention.

[0074] For the output prediction results, the loss function formula for each batch is as follows:

[0075] The predicted labels in each batch The average value of the cross entropy with the true label y is used as the loss function to calculate. This process only uses real number vectors for calculation, which can optimize the model efficiency without increasing the amount of calculation.

[0076] The method proposed in the present invention first uses a parallel TCN network to extract motion features and meteorological features from multi-source information respectively, and then inputs these features into the LSTM for trajectory identification after decision-making fusion. During this process, the present invention proposes a TCN network based on additive position coding, which enables the model to more accurately understand the global temporal relationship of the trajectory sequence during the feature extraction stage; at the same time, the present invention introduces the Kan module to replace the more commonly used fully connected network in traditional deep learning networks to solve the problem of complex trajectory sequences being difficult to accurately represent. The present invention fully considers the characteristics of abnormal trajectories of smuggling ships in actual scenarios, and improves the recognition performance of detection algorithms and devices when facing smuggling activities that are both covert and flexible.

[0077] The embodiment of the present invention further provides a device for identifying abnormal trajectories of smuggling vessels based on multi-source information fusion, the device comprising:

[0078] The sequence acquisition module is used to obtain the ship trajectory sequence, which includes motion characteristics and weather characteristics;

[0079] The TCN network module is used to extract motion features and weather features from the ship trajectory sequence respectively; it is also used to fuse the extracted motion features and weather features to obtain a fused feature vector;

[0080] The LSTM network module is used to further process the fused feature vector to obtain multi-head attention output;

[0081] The Kan network module is used to input the multi-head attention output into the Kan network. The Kan network performs nonlinear feature mapping through adaptive grid points and B-spline functions, and outputs a prediction result on whether the trajectory is suspected to be an abnormal trajectory of a smuggling ship.

[0082] Example

[0083] 1. Dataset scenario construction

[0084] During the dataset preparation process, a multi-level calibration method was adopted to divide the labels of ship data.

[0085] In the first step, a rule-based approach was used to collect radar-only, non-AIS tracks near the Qiongzhou Strait between 9:00 PM and 6:00 AM. Tracks with these characteristics are typically identified as suspected smuggling vessels and are tracked. A total of 5,665 tracks were collected.

[0086] In the second step, the trajectories were cleaned to remove trajectories with data anomalies and missing data, resulting in 5337 trajectories.

[0087] The third step is to manually verify and label these tracks. Before this step, professional training and analysis of past confirmed smuggling cases were conducted. Finally, a total of 1,473 tracks of suspected smuggling ships were obtained. The other 3,864 tracks were considered normal tracks and did not pose any smuggling risk.

[0088] These trajectories are then preprocessed. These steps include grouping the sampling points of different trajectories and sorting them by timestamp to convert them into a spatiotemporal sequence of trajectories; resampling each trajectory at 1-minute intervals; removing other features collected by the sensors, retaining longitude, latitude, speed, and heading as the ship's motion signature; and deleting null values. Compared to the original data, this preprocessed trajectory data eliminates redundant and useless information, allowing the model to better learn trajectory characteristics.

[0089] Taking into account the external weather factors that have a certain impact on the occurrence of smuggling behavior. The present invention adds to the original motion features and constructs a data set with more weather features. The present invention selects the weather characteristics of the offshore weather forecast of the China Meteorological Administration as the data source. The selected weather features include weather phenomena, wind direction, wind force, wave height and visibility. The weather update frequency is 12 hours. At the same time, because in past cases, smuggling ships often sailed from the Beibu Gulf into the Qiongzhou Strait, the sea area where the trajectory was located was also added as a feature. The sea area division standard comes from the document "HYT 0292-2020 Offshore Forecast Sea Area Division". The discrete and disordered features (weather phenomena, wind direction, sea area) are then converted into One-hot encoding so that the model can better understand these objective factors.

[0090] To prevent overfitting during model training, we partitioned the dataset in our experiments. After shuffling the order of trajectory samples using a random seed, we divided the dataset into a training set, validation set, and test set in a 6:2:2 ratio. We also ensured that there was no data overlap between the training set, validation set, and test set.

[0091] 2. Experimental Environment and Evaluation Metrics

[0092] In this embodiment of the present invention, model training was performed on a server using an Intel(R) Core(TM) i7-7800X CPU @ 3.5-GHz six-core CPU and an NVIDIA TAITAN XP graphics processor. The training was performed using the PyTorch deep learning framework on a Windows 10 x64 operating system.

[0093] The evaluation indicators used are Precision (P), Recall (R), and F1-Score (F1) to evaluate the performance of each model. Their calculation methods are shown below. The higher the values ​​of the three indicators, the stronger the model performance.

[0094]

[0095] TP and FP represent true detected anomalies and falsely detected anomalies, respectively, and FN refers to misclassified normal samples. Precision indicates the proportion of samples predicted as positive that are actually positive. Recall indicates the ratio of predicted positive samples to the total number of positive samples. F1-score is the weighted harmonic mean of precision and recall, used to comprehensively measure precision and recall.

[0096] In addition to the above evaluation metrics, Parameters and Avg Time / sample were selected as other evaluation metrics to show the complexity of the model. Among them, "Parameters" represents the number of parameters based on which the model is trained, and "Avg Time / sample" represents the average time (in seconds) spent testing each sample.

[0097] To avoid randomness in the experimental results, we used different random seeds to train each model five times in all experiments, and then tested these models separately, and took the average of the five test results for display.

[0098] 3. Performance indicator analysis

[0099] In order to verify the effectiveness of the model proposed in the present invention, the present invention is compared with several advanced methods on multivariate time series classification tasks, namely ResNet, LSTM, TCN, Dlinear, LightTS, PatchTST and TimesNet. Among them, the overall architecture of ResNet is composed of multiple residual blocks, which has far-reaching influence. The LSTM network shows good adaptability to time series data in a variety of time series tasks through its built-in memory units and gating mechanism, and is one of the backbone networks of this model. TCN uses 1D convolution and extended causal convolution to efficiently extract patterns in sequences, and is also one of the backbone networks of this model. Dlinear decomposes the time series into trend sequence and residual sequence, and then applies these two decomposed sequences to two single-layer linear networks respectively, so that the network models them. LightTS converts the 1D time series into a 2D structure, and then performs continuous sampling and interval sampling on the 2D time series, and then uses MLP to extract features. PatchTST is a multivariate time series model based on the Transformer structure. Its key part is to divide the time series data into subsequence-level patches. TimesNet transforms one-dimensional time series into two-dimensional space for analysis and uses the Inception module for time series analysis. Table 1 shows a comparison of the performance of various networks in the task of identifying suspected smuggling vessel trajectories.

[0100] Table 1 Performance comparison of various networks

[0101]

[0102] The results in Table 1 show that the proposed model achieves the best evaluation metrics in the suspected smuggling track identification task. Furthermore, compared to TimesNet and ResNet, which rank second and third in evaluation metrics, this model significantly outperforms the other models in terms of training parameter count and test speed, highlighting its advantages in computational cost and efficiency. Taking all factors into consideration, the proposed model achieves the best performance in the suspected smuggling track identification task compared to other models.

[0103] 4. Effectiveness Analysis (Ablation Experiment)

[0104] To further evaluate the effectiveness of each component in the proposed model, we designed several performance analysis experiments. The experimental results are shown in Table 2. During the experiments, only the components were changed between models; all other parts remained the same.

[0105] Table 2: Effects of different components on the proposed model

[0106]

[0107] Experimental results show that introducing position encoding into the TCN not only improves evaluation metrics but also significantly accelerates model testing and even training, a surprising result. Analysis suggests this is because position encoding enhances the TCN's ability to capture temporal information, making it easier for the TCN to understand the relationships between time points in a trajectory sequence, thereby improving feature extraction efficiency.

[0108] In addition, the addition of position encoding may also reduce the computational burden of TCN when inferring the temporal relationship between the context of the trajectory, making the training converge faster and reducing the need for complex calculations.

[0109] The introduction of the Attention mechanism into the LSTM network significantly improved various evaluation metrics, achieving the largest improvement among all individual modules. Although the Attention mechanism increased the number of model parameters, its contribution to performance improvement was also significant, enabling the model to perform even better in the suspected smuggling identification task.

[0110] After introducing the Kan network instead of the traditional fully connected network, various evaluation metrics also showed significant improvements. Although the increase in the number of parameters was relatively small, the model testing speed decreased. In addition, during training, the model convergence speed was significantly slower in the presence of the Kan network, indicating that the Kan network is more difficult to train and converge than the fully connected network. Therefore, when designing the network structure, it is necessary to balance the relationship between performance improvement and computational overhead.

[0111] 5. Analysis of actual detection results

[0112] In order to better demonstrate the recognition ability of the proposed model for suspected smuggling trajectories, three representative abnormal trajectories and four normal trajectories are drawn from the trajectory set, as shown in the figure. Figure 3 (a) Of these figures, the first three represent anomalous trajectories, while the last four represent normal trajectories. The trajectory in the fourth figure originates at a port and is therefore a normal trajectory. The trajectories in the fifth and sixth figures detour repeatedly and do not conform to the behavioral pattern of smuggling trajectories. The location where the trajectory in the seventh figure lands does not conform to the characteristics of a smuggling trajectory and is therefore also a normal trajectory.

[0113] Subsequently, these 7 trajectories are input into each network for detection, and the detection results are redrawn. The circle in the upper left corner is blue, which means the detection result is correct, and the circle in the upper left corner is red, which means the detection result is wrong. Figure 3 As shown in (b)-(i), the detection results show that the proposed model accurately identifies the trajectory of the suspected smuggling ship, while other models have different degrees of false positives and false negatives.

[0114] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A method for identifying abnormal trajectories of smuggling ships based on multi-source information fusion, characterized in that: The method is: Acquire a ship trajectory sequence, wherein the trajectory sequence includes motion characteristics and weather characteristics; The motion features and weather features in the ship trajectory sequence are extracted respectively through the TCN (Temporal Convolutional Network) network with additive position encoding; The extracted motion features and weather features are fused to obtain a fused feature vector; The fused feature vector is input into an LSTM (Long Short-Term Memory) network with an attention mechanism for further processing to obtain a multi-head attention output. The multi-head attention output is input into the Kan network, which performs nonlinear feature mapping through adaptive grid points and B-spline functions, and outputs a prediction result on whether the trajectory is suspected to be an abnormal trajectory of a smuggling ship; The multi-head attention output is input into the Kan network, which performs nonlinear feature mapping through adaptive grid points and B-spline functions, and outputs a prediction result on whether the trajectory is suspected to be an abnormal trajectory of a smuggling ship, specifically including: The multi-head attention output S is passed through multiple layers of Kan transformation layers to obtain the final output of whether the trajectory sequence is suspected of smuggling. The main operation of each layer of Kan linear transformation layer is expressed as where h (l+1) represents the output of the l+1th layer, represents the weight matrix of the linear transformation of layer l, Represents the weight matrix of the B-spline of the lth layer, σ is the activation function, B(h (l) ) represents the interpolation function of B-spline, h (0) It is the S of multi-head attention output.

2. The method for identifying abnormal trajectories of smuggling ships based on multi-source information fusion according to claim 1 is characterized in that: The method of extracting motion features and weather features in the ship trajectory sequence by using a TCN (Temporal Convolutional Network) with additive position coding specifically includes: extracting motion features X in the trajectory sequence X by using the TCN network M Perform encoding operation to obtain the trajectory vector X' with position encoding M ; The trajectory vector X' with position encoding is processed by TCN network M and weather characteristics X W Perform feature extraction respectively to obtain feature map U M and U W .

3. The method for identifying abnormal trajectories of smuggling ships based on multi-source information fusion according to claim 2 is characterized in that: The motion feature X in the trajectory sequence X is analyzed by the TCN network M Perform encoding operation to obtain the trajectory vector X' with position encoding M , specifically including: X M Projection to a higher dimension d pos ; The mathematical expression based on position encoding is: Determine the trajectory vector X' with position encoding M =X M +PE; where pos represents the word vector position of the trajectory data, max_len represents the maximum sequence length of all ship trajectory sequences, d pos represents the embedding dimension of the position encoding, 2i and 2i+1 represent the even and odd dimensions in the ship trajectory vector, which are calculated using the sin and cos functions respectively.

4. The method for identifying abnormal trajectories of smuggling ships based on multi-source information fusion according to claim 3 is characterized in that: The method of fusing the extracted motion features and weather features to obtain a fused feature vector specifically includes: according to U={u1,u2,…,u T }=concat(U M ,U W ) to obtain the fused feature vector U, where T is the length of the merged vector and also the length of the trajectory sequence.

5. The method for identifying abnormal trajectories of smuggling ships based on multi-source information fusion according to claim 4 is characterized in that: The fused feature vector is input into the LSTM (Long Short-Term Memory) network with attention mechanism for further processing to obtain multi-head attention output, specifically including: at each time step t, receiving the input u of the current time step t , combined with the hidden state h of the previous time step t-1 and cell state c t-1 , determine the hidden state h of the current time step t and cell state c t ; Traverse the entire trajectory sequence and generate the hidden state sequence H = {h1,h2,…,h T }, where H d h Represents the dimension of the LSTM hidden layer; Performing a linear transformation on the hidden state sequence H to generate a query vector Q, a key vector K, and a value vector V; According to the query vector Q corresponding to each head i and the transpose of the key vector Determine the similarity and normalize it through the softmax function to determine the attention weight α of each head i ; According to the attention weight α of each head i The sum value vector V determines the context vector C of each head i =α i *V i ; The result of concatenating the context vectors of all heads is transformed into the multi-head attention output S through the linear transformation matrix, S = W O [C1; C2; …; C H ], where W O is the output linear transformation matrix used to merge information from different attention heads.

6. The method for identifying abnormal trajectories of smuggling ships based on multi-source information fusion according to claim 5 is characterized in that: At each time step t, the input u of the current time step is received t , combined with the hidden state h of the previous time step t-1 and cell state c t-1 , determine the hidden state h of the current time step t and cell state c t , specifically including: according to h t ,c t =LSTM(u t ,h t-1 ,c t-1 ) Determine the hidden state h of the current time step t and cell state c t .

7. The method for identifying abnormal trajectories of smuggling vessels based on multi-source information fusion according to claim 6 is characterized in that: The query vector Q corresponding to each head i and the transpose of the key vector Determine the similarity and normalize it through the softmax function to determine the attention weight α of each head i , specifically including: Determine the attention weight α for each head i ; where d K is the dimension of the key vector.

8. The method for identifying abnormal trajectories of smuggling ships based on multi-source information fusion according to claim 7 is characterized in that: The B-spline function is calculated recursively, and the recursive formula is: Among them, t i represents the i-th grid point, which determines the segmentation position of the basis spline function; k is the spline order; B i,0 (h (l) ) represents the piecewise constant, when h (l) When it is within the corresponding interval, its value is 1, otherwise it is 0; The B-spline interpolation function is a function generated by a linear combination of a set of control points and the corresponding B-spline basis functions. Its general form is: B(h (l) )=∑ i＝0 P i B i,k (h (l) ); where P i Represents the control points. Different control points generate B-spline interpolation functions of different shapes.

9. A device for identifying abnormal trajectories of smuggling ships based on multi-source information fusion, characterized in that: The device includes: A sequence acquisition module is used to obtain a ship trajectory sequence, wherein the trajectory sequence includes motion characteristics and weather characteristics; The TCN network module is used to extract motion features and weather features from the ship trajectory sequence respectively; it is also used to fuse the extracted motion features and weather features to obtain a fused feature vector; LSTM network module, used to further process the fused feature vector to obtain multi-head attention output; The Kan network module is used to input the multi-head attention output into the Kan network. The Kan network performs nonlinear feature mapping through adaptive grid points and B-spline functions, and outputs the prediction result of whether the trajectory is suspected to be an abnormal trajectory of a smuggling ship. Specifically, the multi-head attention output S is used to obtain the final output of whether the trajectory sequence is suspected of smuggling through multiple layers of Kan transformation layers. The main operation of each layer of Kan linear transformation layer is expressed as follows: where h (l+1) represents the output of the l+1th layer, represents the weight matrix of the linear transformation of layer l, Represents the weight matrix of the B-spline of the lth layer, σ is the activation function, B(h (l) ) represents the interpolation function of B-spline, h (0) It is the S of multi-head attention output.

Images