DASS: A Dual-Branch Attention-based Framework for Trajectory Similarity Learning with Spatial and Semantic Fusion

1 Introduction

Trajectory similarity search plays an important role in spatialtemporal data analysis, which quantifies the correlation between trajectories and serves as a fundamental operation for identifying both individual movement patterns and collective operational trends [Shang et al., 2017; Yu et al., 2019]. It holds immense value across various domains, including personalized recommendation [Bao et al., 2017; He et al., 2020], navigation [Yao et al., 2022b; Zhang et al., 2020], and intelligent transportation [Feng et al., 2023; Ruan et al., 2022; Yi et al., 2023].

The landscape of trajectory similarity methods is rich with proposals, encompassing both heuristic-based [Koide et al., 2020; Yi et al., 1998; Vlachos et al., 2002; Yuan and Li, 2019; Wang et al., 2018], which suffer from high computational cost of O(n2 ); and learning-based approaches [Yao et al.,

∗ Corresponding author.

2019; Zhang et al., 2020; Zhao et al., 2024; Tao et al., 2021], which extract generalizable and concise representations from sequential trajectory [Liu et al., 2024]. Our study falls within the latter category, where existing methods have increasingly exploited road networks or scene-related information offered by various location-based applications. Specifically, one category [Fang et al., 2021; Zhang et al., 2023; Mao et al., 2022] incorporates static road network structure into representation to primarily capture spatial proximity and path similarity. Another set of methods [Jiang et al., 2023; Yuan et al., 2022; Liu et al., 2020] gradually leverages semantic aspects, such as travel or time regularity, for a more nuanced representation. While both perspectives provide valuable insights, current proposals fail to adequately capture the interaction and complementarity between two views. Therefore, we argue that an effective similarity measure should integrate both spatial and semantic similarities to achieve a more comprehensive, multidimensional representation.

The motivation for our work can be described from two perspectives: the practical application and the technical viewpoint. (i) For the application: Focusing solely on spatial proximity or semantic closeness usually fails to yield accurate results. Taking Figure 1 as an example, where illustrates three trajectories T1, T2 and T3, each comprising four points, with the spatial coordinates and semantic label annotated next to each point. When only considering spatial proximity, T1 and T2 are the most similar, calculated by three distance metrics: ED [Clarke, 1976], L1dis and MD [Cassels, 2012]. However, when semantic information is taken into account, and computed via a hybrid type attribute dissimilarity formula [Kaufman and Rousseeuw, 2009], T1 and T3 are determined to be the most similar. From the above observation, we can conclude that including or excluding semantic information would lead to different outcomes in similarity search tasks. (ii) For the implementation technology: The deep neural network framework offers a promising solution to combine various information in trajectory representation to enhance the generalizability. Broadly speaking, converting the connectivity and shape of temporal trajectory into vectors ensures the correctness of representation, while embedding semantic information through a dedicated encoding network enhances the model's ability to capture activity-related similarities.

To integrate both spatial and semantic information into the learning model for trajectory similarity calculation, three non-trivial challenges have to be tackled:

C1: Challenge of Data Representation. The key of trajectory modeling lies in constructing a unified representation that captures both the complex spatial configurations of road networks and the temporal dynamics of trajectories. Existing methods often model spatial information by learning shared representations of equally-sized grids [Li et al., 2018; Yao et al., 2019], resulting in weak connections for records located far apart in sequence modeling. In light of this, we design a Graph-based Spatial-Temporal Modeling module (GSTM), to obtain a multidimensional information fusion representation. This module is flexible and generic, in that it can be integrated with any existing trajectory representation proposal for spatio-temporal similarity learning.

C2: Challenge of Model Framework. To achieve viable unified similarity learning, a fundamental challenge lies in designing a model framework that facilitates the seamless and effective integration of various components. A straightforward idea is to adopt a simple weighted combination. However, such a design struggles with accurate weights set and may not fully capture the nuanced relationships between spatial and semantic features. To overcome this limitation, we develop a Dual feature Co-Attention Fusion module, termed DualCAF, which enables the effective integration of diverse information within trajectories over road networks.

C3: Challenge of Effectiveness Improvement. To achieve both high accuracy and efficiency in similarity queries, the sophistication processes and the optimization parameters must be carefully managed. Current approaches, particularly those leveraging recurrent neural networks, often have difficulty with scalability and capturing long-term dependencies, limiting their utility in large-scale, real-world datasets. In response, we develop a dual-branch Transformer-based network coupled with an SST-Aware loss function to improve effectiveness. It enables our model to balance computational efficiency with the accurate representation of the inherent characteristics across diverse trajectory patterns.

In conclusion, we propose a novel framework DASS to counter the above challenges in trajectory similarity learning. Our major contributions are summarized as follows:

• We propose an attention-based framework, which not only captures the structural information reflected by spatial-temporal graph, but considers semantic activity information to enrich trajectory representation.
• We present a dual-branch trajectory backbone encoder, which incorporates a DualCAF module to simultaneously process spatial and semantic information via distinct attention mechanisms tailored to each feature type.
• We design a weight-adaptive fusion loss function that allows our model to dynamically adjust the importance of each component during training, promoting balanced optimization and enhanced performance.
• We conduct extensive experiments on real-world datasets, demonstrating that DASS outperforms state-ofthe-art methods, establishing itself as a novel paradigm for trajectory similarity learning on road networks.

2 RelatedWork

Trajectory Similarity Learning. Learning-based methods have gained considerable attention for their ability to automatically derive low-dimensional representations from trajectories data [Li et al., 2018; Yao et al., 2017; Ma et al., 2024; Messaoud et al., 2021; Jiang et al., 2023], which enables fast trajectory similarity computation in a linear time. NEUTRAJ [Yao et al., 2019] is the first learning-based trajectory similarity measure, which samples trajectories as seeds and uses pair-wise comparisons as guidance. To enhance training efficiency, Traj2SimVec [Zhang et al., 2020] simplifies the process by converting training trajectories into triplet samples. ST2vec [Fang et al., 2022] focuses on both spatial characteristics and temporal regularities within trajectories, addressing the inherent complexity of spatiotemporal data. [Chen et al., 2020] introduces a novel metric to evaluate spatial and textual domain trajectory semantic similarity comprehensively. To further improve deep representation for long trajectories, TrajGAT [Yao et al., 2022a] introduces a quad-tree index to model long-term dependencies.

Graph Neural Networks. Recent studies have explored integrating GNNs [Kipf and Welling, 2016] for trajectory similarity on road networks, demonstrating their potential to capture the spatial information of trajectories [Chen et al., 2019]. Han et al. [Han et al., 2021] first apply GNNs into this domain and utilize unique POI information. Their subsequent work [Zhou et al., 2023] further extends this approach, leveraging LSTM and GAT to model sequential dependencies and spatial interactions within trajectories. Most recently, Start [Jiang et al., 2023] presents a spatial networks-based representation method on the basis of GAT, and achieves innovative results by considering travel semantics hidden in trajectories.

3 Preliminaries

3.1 Basic Concepts

DEFINITION 1 (Trajectory). A trajectory, denoted by T , is a time-ordered sequence of GPS points of a moving object, i.e., T = ⟨p1, p2, ..., pn⟩, where each point pi = (xi, yi, ti) is a longitude, latitude, and timestamp triplet.

DEFINITION 2 (Road Network). A road network is represented as an undirected graph G = (V,E), where each node v ∈ V is a Point of Interest (POI); and each edge e = ⟨u, v⟩ ∈ E is a road segment connecting points u and v.

Given a trajectory T , we align each trajectory point pi with vertex vi in road network through a commonly used mapmatching procedure [Brakatsoulas et al., 2005], to obtain the corresponding vertex trajectory T road = ⟨v1, v2, ..., vn⟩.

DEFINITION 3 (Top-K Similarity Query). Given a query trajectory QT , a set of trajectories T = {T1, T2, ..., Tm}, and a similarity measure f(·), the Top-K similarity query returns K trajectories in T that are most similar to QT under f(·).

3.2 ProblemDefinition

Given a set of trajectories T and corresponding road network graph G, DASS aims to learn a trajectory encoder E : T → E, mapping each trajectory T to a d-dimensional embedding vector E that jointly captures the trajectory characteristics and road network information. For ∀Ti ∈ T, the objective is to find Tj ∈ T that is most similar to Ti under the similarity function f(·). Such problem can be formalized as below:

E aims to accurately reflect the similarity between trajectories through their embeddings, i.e., E(Ti) and E(Tj) are close (distant) if Ti and Tj are similar (dissimilar), demonstrating the effectiveness of E in capturing inherent characteristics.

4 Methodology

Figure 2 shows the framework of DASS, which consists of three following major modules:

• Graph-based Spatial-temporal Modeling, which models spatial and temporal adjacency relationships between points with knowledge graph embedding, and then constructs a similarity-based graph to capture the underlying structural patterns (Section 4.1).
• Semantic Information Embedding, which comprises three specific components to capture contextual information within the trajectory and enrich representation with deeper semantic context (Section 4.2).
• Dual-Branch Fusion Encoding, which adaptively processes the features from both perspectives, ensuring that the unique characteristics are effectively captured before they are integrated (Section 4.3).

The above three modules of DASS work collaboratively to obtain a unified representation, and the whole framework is trained by the self-supervised paradigm.

4.1 Graph-based Spatial-temporal Modeling

Road networks inherently possess rich topological structures, while trajectories are time-ordered sequences. However, due to varying sampling rates and data sparsity, adjacent points in trajectory may not always correspond to adjacent points in the road network. Based on this, we classify adjacency into two types: 1) temporal adjacency, where two points appear sequentially over time within the trajectory; 2) spatial adjacency, where adjacent points are directly connected by edges on road networks, indicating spatial continuity. From an intuitive view, a trajectory knowledge graph can be constructed to model these adjacent relationships, expressed as (h, r, t), where h and t refer to POIs within the road network, r is the relationship between two vertices and are defined as follows: ① temporal adjacency Rt, ② spatial adjacency Rs, and ③ spatio-temporal adjacency Rst.

To accurately model various relationships, DASS first utilizes TransD [Ji et al., 2015] to obtain the knowledge graph representation. Then it introduces a point-relation similarity function to consolidate the multidimensional similarities between points into a unified similarity computation framework. Drawing inspiration from KGE techniques [Wang et al., 2014], the distance between two points u and v under a specific relationship r is defined as:

where eu, ev, er are the embedding through TransD. Following Equation (2), we design a bilateral neighbor similarity function to calculate the similarity between two nodes, which not only accounts for the direct similarity between points, but incorporates contextual information from neighbors to capture broader spatial distribution within the local structure. The similarity function is defined as follows:

where |N(u)| and |N(v)| denote the neighbor sets of u and v, respectively. Based on the similarity calculation in Equation (3), we construct a graph where each point connects its Top-K most similar neighbors. High-similarity relationships between points are reflected by neighbor connection, ensuring that adjacent nodes maintain closely related representations later. The graph construction is formulated as below:

The resulting graph Gst will be used as the spatial branch input for the subsequent network (detailed in Section 4.3).

4.2 Semantic Keywords Embedding

Each point in the trajectory T has an associated keyword to represent the context of an activity. To efficiently convert such nominal attributes into numerical embeddings, similar to BERT's implementation process, the activity keyword embeddings are composed of three components: 1) Positional encoding. Drawing inspiration from the position embedding in Transformer [Vaswani, 2017], each keyword ki in the sequence of trajectory keywords K = {k1, k2, ..., kn} is assigned a fixed sinusoidal positional encoding embedding, EP (i), calculated as follows:

Here, d is the embedding dimension, and sin(·), cos(·) serve as periodic activation functions that capture sequential behaviors. This component preserves sequential information, enhancing the model's sensitivity to sequence order. 2) Token encoding. For each activity keyword ki, we use the pretrained GloVe model to generate a token embedding ET (ki), capturing the intrinsic semantic properties of the word. 3) Segment encoding. Each activity category (e.g., entertainment) is associated with a unique segment embedding ES(ki) through GloVe. By mapping each keyword to its corresponding category, this component reinforces the model's ability to distinguish and categorize similar activities within the trajectory. The final representation of ki is expressed as:

Next, we feed E(ki) in Equation (6) into BERT's transformer architecture step by step to obtain a smoothed representation ei ∈ Rd for each keyword. These embeddings E = {e1, e2, ..., en} have integrated both semantic content and sequential order, offering a comprehensive foundation for semantic similarity analysis.

4.3 Dual-Branch Fusion Encoding

Spatial and semantic information of trajectories can be treated as two observations of the same concept, then a co-attention fusion module (DualCAF) is designed for our DASS model to generate spatio-semantic oriented representation.

DualCAF. It takes both graph Gst and semantic embeddings E as input and outputs the joint representations of trajectory points. To capture the topological structure within the graph, the multi-head GAT is employed to assign attention coefficients to neighboring nodes. Formally, the hidden state of node vi is updated by aggregating information from its neighbors, weighted by learned attention coefficients from multiple attention heads, represented as follows:

where α(k) ij represents the attention coefficient from the k-th attention head between node vi and its neighboring node vj, W (k) is a learnable weight matrix, and hj is the hidden state of node vj . The attention coefficient can be computed as:

where || denotes concatenation. So far, spatial representation Sg have been learned by GAT layers, which integrates topological relations derived from the input graph Gst.

For semantic branch, the semantic feature matrix E is first projected into value Vi, key Ki, and query Qi matrices, where i denotes the i-th head. The introduction of muti-head attention enables to focus on different parts of the semantic feature space when learning co-attention. Formally, the attention coefficient matrix of the i-th head Ai is computed as follows:

Based on Equation (10), the attention coefficient matrix Ai is multiplied with value matrix Vi to obtain the hidden output his , and the outputs his from each self-attention head are concatenated and passed through a linear transformation to generate the final output hs as follows:

where Wo is the learnable weight matrix.

With both spatial and semantic features embedded, they are fused using a concatenation strategy, denoted as Fss. Fss is then passed through another self-attention mechanism to capture complex feature interactions. Subsequently, a layer normalization, an MLP module with residual connection, and another layer normalization are applied to ensure optimization stability and model effectiveness, as described in Equations (13) and (14). Finally, DualCAF outputs the dual feature fusion representation Css.

DASS. As illustrated in Figure 2, DASS follows the overall structure of a multi-layer Transformer encoder and possesses two branches with complementary responsibilities. It takes Gst and E as the input, which are processed through the DualCAF module to jointly model spatial and semantic features, learning their complex interactions. The output of DualCAF Css, is further refined through a series of residual connections, dropout layers, and layer normalization to stabilize gradient flow and mitigate issues like gradient vanishing or explosion. The updated representation Jss is then passed through an MLP, followed by another residual connection and layer normalization. The process can be summarized as:

where Sg is the original representation for the similaritybased graph, and Hss refers to the updated representation matrix, incorporating both the topological and semantic features. The model is implemented with stacked layers (two layers in the experiments) of the DualCAF and MLP blocks. After the final layer, average pooling is applied to Hss to obtain the final trajectory representation h ∈ Rd as follows:

Based on Equation (15), the final representation incorporates both structural and semantic information, providing a comprehensive foundation for downstream similarity analysis.

4.4 SST-Aware Loss Function

To enhance the accuracy and robustness of trajectory representation, we introduce a novel SST (Spatial-SemanticTrajectory)-Aware loss function to optimize model training. Graph-Based Spatial-temporal Contrastive Loss. It encourages each node in the similarity-based graph Gst to be close to high-similarity nodes (positive samples) while remaining distant from unrelated nodes (negative samples). The objective is to minimize the distance between node v and its positive samples vpos, while maximizing the distance between v and its negative samples vneg, thus enhancing the model's understanding of underlying graph structure. The loss function is defined as follows:

where s(·) represents the vector dot-product function. Semantic Pair-wise Loss. The semantics of each point is extracted in the form of a (category, subcategory) structure. For each vi ∈ Ti, two points with the same category are considered positive samples; others are negative. Then, the semantic pair-wise loss encourages semantic consistency and is formulated as follows:

where σ is sigmoid function, and vij is a point in Ti.

SST-Aware Loss. For each trajectory Ti, the most similar trajectory (ground truth) is considered as Tpos, and Tneg is a negative sample selected from T\{Ti, Tpos}. The trajectory loss function is formulated as follows, from the perspective of the entire trajectory similarity:

Finally, SST-Aware loss function is composed of Equations (16) to (18), expressed as:

where each weight Wλ,Wµ and Wη are learned adaptively through a softmax function within an attention mechanism.

5 Experiments

In this section, we conduct multiple experiments to verify the effectiveness of DASS and answer the following questions:

• RQ1: How does DASS perform compared with existing trajectory similarity computation methods?
• RQ2: How does every module that we design contribute to the model performance?
• RQ3: How do variations in parameter settings impact the performance of DASS?

5.1 Experimental Setup

Data Description. We use two road networks from OpenStreetMap: Beijing Road Network (BJRN), containing 28, 342 nodes and 27, 690 edges; and New York Road Network (NYRN), with 95, 581 nodes and 260, 855 edges. For BJRN, the public real-life trajectory dataset T-drive is adopted, which includes trajectories from 10, 357 taxis in Beijing over a week. For NYRN, we randomly sample a subset of taxi trip data from New York to generate the dataset. After preprocessing, we collect 5231,420 trajectories for Tdrive and 9541,270 for New York. Both datasets are divided into training, evaluation, and testing sets, with 20%, 10%, and 70% allocated to each set, respectively.

Baselines. To evaluate the effectiveness of the proposed DASS, we implement in total eight baselines. i) Traj2vec [Yao et al., 2018] represents trajectory features via the moving behavior and reconstruction loss. ii) Siamese [Pei et al., 2016] employs a Siamese Network with a cross-entropy loss function. iii) NeuTraj [Yao et al., 2019] combines metric learning with spatial attention memory and a distanceweighted ranking loss. iv) At2vec [Liu et al., 2020] addresses uneven sampling issues with LSTM. v) Traj2SimVec [Zhang et al., 2020] designs a custom strategy to extract distance information from sub-trajectories. vi) GTS [Han et al., 2021] is the first graph-learning based approach for trajectory similarity on road networks. vii) ST2Vec [Fang et al., 2021] first considers spatio-temporal similarity in trajectory representation. viii) GRLSTM [Zhou et al., 2023] designs a residualLSTM to model the trajectory data.

https://opendata.cityofnewyork.us

Evaluation Metrics. Following existing works, we adopt the Top-K hitting ratio (HR@K) as the major performance metric, which captures the degree of overlap between a Top-K result (Definition 3) and the corresponding ground-truth result. The ground-truth are the exact Top-K similarity search results obtained through NetERP [Koide et al., 2020].

Parameter Settings. All experiments are conducted with PyTorch 1.8.1 on a LINUX server (Intel Xeon 5118 6-Core CPU, 32 GB of RAM, and a GeForce GTX 3090 Ti GPU). We train DASS using the Adam optimizer, the learning rates and Top-K selection in Equation (4) are set as 5e-4, 15 in the Beijing dataset and 1e-3, 30 in the NewYork dataset respectively. The embedding dimensionality d is set to 128 for all learned methods. For DualCAF, the number head of K in Equation (7) and s in Equation (12) are set as 8 and 4. For DASS, The number of encoder #layers is 2.

5.2 Overall Performance (RQ1)

Effectiveness Evaluation. We conduct Top-K similarity queries and compare DASS with all 8 baselines. Table 1 lists the results on two datasets. From these results, we provide observations and analyses as follows.

The first observation is that DASS achieves the best performance in terms of HR@K on both two datasets, significantly outperforming all the baseline methods. It confirms the superior performance of our framework in learning trajectory representations by introducing spatial-temporal features and semantics in the pre-training phase. Specifically, DASS achieves a consistent hitting ratio exceeding 35% (except HR@1), with an average improvement of 25.88% over the best baseline GRLSTM on the Beijing dataset and 22.19% on the New York dataset. Moreover, we can observe that methods like Traj2vec and NeuTraj perform relatively poorly, likely because they lack explicit modeling of road network constraints and primarily rely on trajectory-level features without considering the underlying topological information.

The third observation is that although ST2Vec and GRLSTM account for road network topology, DASS offers two key advantages contributing to its superior performance. First, our approach incorporates activity semantic - a feature absent in the other two methods, which captures critical information from multiple dimensions to enrich representation. Second, its Transformer-based network, unlike LSTM-based methods, can better capture long-term dependencies. It is also worth noting that DASS's performance advantage is less pronounced on the New York dataset compared to Beijing. One possible reason is the significantly larger scale and complexity of the New York road network, which may introduce more noise and less consistent trajectory patterns.

Efficiency Evaluation. We evaluate the efficiency of each measure in both model training time (per epoch) and inference time (the process of converting trajectories into representation). The experimental results are shown in Table 2. For simplicity, we only present the four models with relatively better efficiency. From Table 2, it is evident that GTS distinguishes itself with the shortest training time, attributed to its relatively simple structure and less dependence on trajectory features. In contrast, the other three models consider additional trajectory features, such as temporal and semantic similarity, which need to be computed during the training process, resulting in longer training times. Our model, DASS, although it includes complex components such as graph modeling and text representation, still maintains high efficiency through optimized design. Its training time is only slightly longer than GTS but achieves higher accuracy in trajectory similarity queries. It indicates that DASS successfully balances computational complexity and model effectiveness.

5.3 Ablation Study (RQ2)

In the ablation study, we quantify the contribution of each DASS component. The results on Beijing dataset are shown in Table 3, confirming the necessity of each component in the DASS framework. A similar trend is observed on the New York dataset; therefore, we omit it for brevity.

Experimental results of seven variants illustrate that the most significant decrease occurs when the DualCAF component (w/o DualCAF) is excluded due to its adaptive fusion of both input features. The removal of both graph modeling (w/o GM) and the semantic branch (w/o Se branch)

(b) NewYork Dataset

Figure 3: Performance of DASS varying trajectory length.

also brings a performance decline of approximately 3.79%, highlighting the importance of richer characteristics for better model representation. Similarly, excluding the ResNet layer (w/o ResNet) leads to performance degradation, as ResNet addresses the gradient vanishing problem without increasing training time. The inclusion of the Transformer module (w/o Transformer) offers additional advantages that better capturing contextual and sequential information for representation learning and achieving improvement from 0.4528 to 0.4701 on HR@5. Moreover, the efficacy of two novel loss functions is also verified, resulting in a 1.93% improvement.

5.4 Parameter Sensitivity Study (RQ3)

Sensitive to datasize. We investigate the effect of trajectory length on the performance of DASS. As shown in Figure 3, the similarity learning performance of DASS is examined as the trajectory length varies from 1, 000 to 5, 000. It can be observed that the performance of DASS slightly decreases with the increasing trajectory length. However, it generally exhibits stable performance on both the Beijing and New York datasets. This suggests that DASS is capable of effectively capturing the essential spatial and semantic information embedded in longer trajectories, demonstrating its robustness across varying trajectory lengths.

Varying parameter K. We analyze the impact of the graph construction parameter K in Equation (4), which determines the number of neighbors considered for each trajectory point in the GSTMmodule. Figure 4 shows the results, where we vary K across different values. We observe that the choice of K has a noticeable impact on the model's performance. Specifically, the best results are achieved when K is set to 15 for Beijing dataset and 30 for New York dataset. For example, in term of HR@10 metric, the results are 0.5473 and 0.4728 on Beijing, NewYork dataset, respectively. This also suggests that a smaller K in Beijing works better due to

(a) Beijing Dataset

(b) NewYork Dataset

the denser road network and more localized movement patterns, while New York's more complex urban landscape benefits from a larger neighborhood size.

Semantic Noise. DASS incorporates semantics from trajectories, making semantics a crucial factor in model performance. We test various levels of semantic noise, and the results are shown in Figure 5. It is evident that the model's performance gradually declines with increasing noise levels in both two datasets. However, the impact of noise on DASS is relatively limited, with performance degradation controlled within 5.3%. As shown in Table 3 under the "w/o Se Branch" condition, even in the extreme case where semantic information is removed, DASS still outperforms other baselines. This is attributed to DASS's unique dual-branch structure, where the spatiotemporal features extracted from the other branch can effectively complement the semantic branch, mitigating the negative impact of noise on overall performance.

6 Conculusion

In this paper, we propose DASS, a self-supervised trajectory similarity learning model with a dual-branch backbone encoder. DASS can effectively capture both the spatial and semantic information embedded in trajectory, and utilizes the DualCAF module to ensure that the unique characteristics of each view are captured before being integrated into a unified representation. We have conducted extensive experiments on real-world trajectory datasets, demonstrating that DASS achieves significant performance improvements on similarity search and outperforms all comparing baselines. The universality of DASS makes it a promising candidate with ample potential for various location-based applications.

Acknowledgments

This work was supported by the Natural Science Foundation of the National Natural Science Foundation of China under grant (61802273), Jiangsu Higher Education Institutions of China (23KJA520011), the National Key Research and Development Program of China(2023YFF0725002), China Science and Technology Plan Project of Suzhou (SYG202139).