A deep learning-based comprehensive sand prevention system service life evaluation system

Abstract

The application relates to the technical field of machine learning, and provides a deep learning-based comprehensive sand prevention system service life evaluation system, which solves the problems of insufficient life prediction precision, limited modeling capability and unstable results of existing methods under complex wind-sand environment; the system comprises data acquisition, preprocessing fusion, life evaluation and early warning modules; the system balances local and global features by adopting a double-branch HCT structure, strengthens degradation features by combining front-end and rear-end frequency domain refining modules, and realizes efficient long-time sequence modeling by introducing a channel space joint-logarithmic linear Transformer based on KAN and a hierarchical mask mechanism; error fluctuation is inhibited by a difference generalization loss function, and prediction stability is improved; the system can realize high-precision prediction of residual life and risk warning under long-term wind-sand erosion and climate fluctuation conditions, and provides support for sand prevention engineering maintenance and planning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, and particularly relates to a deep learning-based service life evaluation system for comprehensive sand prevention system. BACKGROUND

[0002] Desertification prevention is an important task of ecological environment protection and sustainable development. In order to slow down wind and sand erosion and land degradation, people generally take various comprehensive sand prevention measures such as vegetation restoration, forest belt construction and sand barrier laying. However, due to the influence of extreme meteorological environment, wind and sand power and material aging and other factors, the performance of the sand prevention system degrades and the structure fails in the long-term service process. The service life evaluation and risk early warning of the sand prevention system are directly related to the reliability and sustainability of ecological management engineering. Although the existing intelligent sand prevention system life evaluation system has begun to introduce remote sensing monitoring, Internet of Things sensors and meteorological data, and realizes the automatic collection and analysis of the environment and the structure state, it still has obvious shortcomings. On the one hand, the data fusion of most systems still stays at the level of simple index superposition and linear statistical modeling, and it is difficult to effectively describe the degradation process of the sand prevention system in time and space under the action of complex environment. On the other hand, the existing intelligent method mainly focuses on the trend prediction of a single factor, such as vegetation coverage, sand barrier stress or local meteorological conditions, and lacks the ability to comprehensively evaluate the service life from the overall system level, resulting in that the evaluation result deviates from the actual service state, and it is difficult to provide reliable basis for engineering maintenance and management decision. In addition, the existing system often relies on fixed threshold or static rules in the risk early warning link, and cannot adaptively adjust combined with the dynamic change characteristics of the service life, so it is difficult to meet the needs of fine management and real-time early warning of the sand prevention system in the long-term operation. SUMMARY

[0003] The application provides a deep learning-based comprehensive sand prevention system service life evaluation system, aiming at the problems that existing sand prevention projects are difficult to effectively depict the structure degradation law under long-term wind-sand environment, the time sequence dependent modeling capability is limited, and the life prediction stability is insufficient, a new system is proposed; the system builds a double-branch HCT structure sharing parameters, combines the advantages of convolution network and Transformer, can extract local spatio-temporal features of shelterbelts, sand barrier materials and other local spatio-temporal features from remote sensing images, sensor monitoring and meteorological data, can capture global dependence relationship across time and space such as sand dune migration and wind-sand flow effect, and realize dynamic balance of the two types of features by using learnable weight factor, so as to truly reflect the gradual degradation process of the sand prevention system under the natural wind-sand environment; in the aspect of feature refinement, a front-end-back-end dual-frequency domain module is designed, the Fourier transform and the gating mechanism are used to highlight the key degradation components such as sand barrier stress and vegetation coverage change, and the residual structure is used to maintain the environmental details, realizing the dual-dimensional enhancement of space and frequency; in the aspect of long time sequence dependent modeling, a channel space joint-log-linear Transformer model is proposed, the KAN (Kolmogorov-Arnold Network) operator is used to jointly model the channel and spatial attention, and the Fenwick tree hierarchical mask and the sequential semi-split mask are combined to form a Log-Linear extension mechanism, the calculation complexity of long time sequence environmental data is reduced from square level to logarithmic linear level, and the prediction stability is still maintained under the long-term wind-sand erosion and climate fluctuation scene; in the aspect of model optimization, a differential generalization loss function is introduced, the dynamic fluctuation of prediction error is constrained, the error shock caused by seasonal extreme weather (such as strong storm or drought) is reduced, and the robustness and generalization performance of life prediction are improved.

[0004] The application provides a deep learning-based comprehensive sand prevention system service life evaluation system, which comprises a data acquisition module, a data preprocessing and fusion module, a service life evaluation module and an evaluation and early warning platform.

[0005] The data acquisition module is used for acquiring multi-source heterogeneous data of the comprehensive sand prevention system, and the multi-source heterogeneous data comprises remote sensing image data, near-ground sensor monitoring data, meteorological environment data and sand prevention system structure attribute data.

[0006] The data preprocessing and fusion module is connected to the data acquisition module, and is used for cleaning, standardizing and spatio-temporal aligning the multi-source heterogeneous data, and fusing to generate multi-dimensional feature data.

[0007] The service life evaluation module is connected to the data preprocessing and fusion module, and has a service life prediction model built therein. The service life prediction model processes multi-dimensional feature data, centrally extracts time-space dynamic characteristics of performance degradation of the sand control system, and outputs service life evaluation indexes. The service life evaluation indexes include a residual life prediction value, a health state grade, and a risk probability. The service life prediction model includes a front-end frequency domain refining module, a channel-space joint-log-linear Transformer model, a rear-end frequency domain refining module, and a full convolutional network.

[0008] The evaluation and early warning platform is connected to the service life evaluation module, and is used for visualizing and displaying the service life evaluation indexes, generating an evaluation report, and presetting a threshold value. When the prediction result exceeds the threshold value, an early warning signal is sent.

[0009] Further, the service life evaluation module processes multi-dimensional feature data through the service life prediction model, and outputs service life evaluation indexes. The process specifically includes the following steps:

[0010] Step S1: A double-branch HCT structure with shared parameters is introduced. Local time-space characteristics of multi-dimensional feature data are extracted through a convolutional neural network, and global dynamic dependency characteristics of the local time-space characteristics are captured through a Transformer. A learnable weight factor is introduced to balance the local time-space characteristics and the global dynamic dependency characteristics, and multi-scale original features are output.

[0011] Step S2: The multi-scale original features are input into the front-end frequency domain refining module to obtain first-order refined features.

[0012] Step S3: The first-order refined features are input into the channel-space joint-log-linear Transformer model to obtain semantic difference maps. The channel-space joint-log-linear Transformer model includes a local attention layer and a global attention layer.

[0013] Step S4: The semantic difference maps are input into the rear-end frequency domain refining module to obtain second-order semantic difference features.

[0014] Step S5: The second-order semantic difference features are input into the full convolutional network to output initial service life evaluation indexes. In the training process of the service life prediction model, a difference generalization loss function is used to optimize the service life prediction model to constrain dynamic fluctuations of test errors, improve the generalization performance of the model, correct the initial service life evaluation indexes, and output the service life evaluation indexes.

[0015] Further, step S3 specifically includes the following steps:

[0016] Step S31: Construct a CKSA unit by combining KAN channel attention and KAN spatial attention, and use the CKSA unit to perform weighted compression on the first-order refined features to generate a compact semantic token set;

[0017] The process of using the CKSA unit to perform weighted compression on the first-order refined features to generate a compact semantic token set is as follows: the CKSA unit takes the first-order refined features as input, constructs a joint mechanism of channel attention and spatial attention through the KAN layer, in the channel dimension, uses the KAN operator to calculate the importance weight of each channel, and performs channel weighting on the first-order refined features to highlight the feature types sensitive to performance degradation; in the spatial dimension, the KAN operator is used to generate a spatial weight distribution, and the weighted channel features are further spatially weighted to highlight the key areas and change areas of the sand barrier structure, to obtain a comprehensive weighted feature, and to perform compression mapping to convert the comprehensive weighted feature into a compact semantic token set;

[0018] Step S32: Time-series modeling is performed on the compact semantic token set, and the semantic tokens in the compact semantic token set are arranged in time dimension to form a time series; for the time series, a time series prefix interval is introduced at each time, and a hierarchical division mechanism based on Fenwick tree is used to divide the time series prefix interval into multiple scale buckets, and feature aggregation is performed to generate a multi-scale bucket feature representation;

[0019] Step S33: For the multi-scale bucket feature representation, an adaptive weight factor is introduced for fusion processing to obtain a fused feature sequence; the fused feature sequence is a sequence composed of multiple tokens, and tokens in the same group form token group dependency, and tokens across groups form token group dependency; the token group dependency and the token group dependency are modeled by a local attention layer, focusing on short-time sequence and local-range dependency, capturing fine-grained and near-end degradation dynamic features, and obtaining local semantic tokens;

[0020] Step S34: In the global attention layer, a sequential semi-separable mask matrix generated based on a gating structure and a Fenwick tree hierarchical mask are combined to form a Log-Linear expansion mechanism, and the Log-Linear expansion mechanism is used to process the fused feature sequence to obtain global semantic tokens;

[0021] Step S35: Define a pixel space, map the dual semantic tokens of the local semantic tokens and the global semantic tokens back to the pixel space to generate a dual semantic pixel map, and obtain a semantic difference map through difference calculation.

[0022] With the above scheme, the present application has the following beneficial effects:

[0023] Firstly, the present application realizes deep modeling of remote sensing images, sensor monitoring data and meteorological environmental information by constructing a double-branch HCT structure sharing parameters, can capture local spatio-temporal features and global dependence features across scales at the same time under complex sandstorm environment, and improves the dynamic description ability of sand dune migration, vegetation coverage change and sand barrier structure degradation process. Compared with the traditional method which only relies on a single data source or a single modeling method, the present application can fully reflect the evolution trajectory of the sand prevention system under long-term wind and sand action, solve the problem that the existing method cannot balance local details and macro trends, and enhance the authenticity and scientificity of life prediction.

[0024] Secondly, the present application realizes effective extraction and enhancement of key degradation features in frequency domain space by designing a front-end-back-end dual-frequency domain refining module, improves the sensitivity to sand barrier stress fluctuation, sand flow impact characteristics and vegetation degradation law, and at the same time, combines the residual mechanism to retain environmental details and ensure the refinement and stability of the prediction result. The improvement effectively solves the problem of insufficient discrimination in the feature extraction link of the existing sand prevention life evaluation method, enhances the discrimination ability of structure degradation under complex environmental factors, and makes the prediction result more suitable for actual service scenarios.

[0025] Finally, the KAN-based channel space joint-log-linear Transformer model and the difference generalization loss function are combined, which realizes efficient and stable modeling in long time series modeling, improves the robustness and generalization ability of prediction. Specifically, the Log-Linear expansion mechanism reduces the computational complexity of long time series environmental data from quadratic level to logarithmic linear level, and still maintains stable modeling ability in long-term sand erosion, seasonal extreme weather and other scenes. The difference generalization loss function suppresses the sharp fluctuations of prediction error, solves the problem that the traditional model is prone to unstable results under abnormal climate conditions, and enhances the practicality and reliability of the present application in real sand environment. In summary, the present application not only achieves a breakthrough in technology, but also significantly improves the accuracy and stability of comprehensive sand prevention system life evaluation in practical application, and provides important support for long-term planning and scientific maintenance of sand prevention engineering. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 A module schematic diagram of a comprehensive sand prevention system service life evaluation system based on deep learning provided by the present application is provided.

[0027] Figure 2 A remaining life change trend diagram of different regions proposed in embodiment four is provided.

[0028] Figure 3A risk probability trend chart for different regions proposed in Example Four. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0030] In Example One, according to Figure 1 The present application provides a deep learning-based service life evaluation system for a comprehensive sand prevention system, which comprises a data acquisition module, a data preprocessing and fusion module, a service life evaluation module and an evaluation and early warning platform.

[0031] The data acquisition module is used to acquire multi-source heterogeneous data of the comprehensive sand prevention system, and the multi-source heterogeneous data comprises remote sensing image data, near-ground sensor monitoring data, meteorological environment data and sand prevention system structure attribute data.

[0032] The remote sensing image data comprises historical and real-time images from multispectral, hyperspectral and synthetic aperture radar satellites.

[0033] The near-ground sensor monitoring data comprises physical parameters collected by wind speed and direction sensors, sand particle impact sensors, stress and strain sensors and humidity sensors arranged in the sand prevention system.

[0034] The meteorological environment data comprises historical and real-time wind speed, precipitation, evaporation and temperature difference data in the region.

[0035] The sand prevention system structure attribute data comprises vegetation coverage, vegetation types, forest belt width, sand barrier material, sand barrier height and grid size.

[0036] The data preprocessing and fusion module is connected to the data acquisition module, and is used to clean, standardize and spatio-temporally align the multi-source heterogeneous data, and fuse to generate multi-dimensional feature data.

[0037] The remote sensing image data is subjected to atmospheric correction, geometric correction and feature extraction to obtain vegetation index, land surface temperature and land surface deformation features.

[0038] The near-ground sensor monitoring data, meteorological environment data and sand prevention system structure attribute data are subjected to outlier rejection, noise reduction and interpolation processing to ensure the continuity and integrity of the data.

[0039] A unified spatio-temporal grid is constructed, and the processed various types of data are mapped into the spatio-temporal grid to realize spatio-temporal fusion of the multi-source data, and form multi-dimensional feature data.

[0040] a service life evaluation module connected to the data preprocessing and fusion module, wherein a service life prediction model is built in, the service life prediction model processes multi-dimensional feature data, extracts time-space dynamic features related to performance degradation of the sand control system, and outputs service life evaluation indexes; the service life evaluation indexes include residual life prediction values, health state grades, and risk probabilities; the service life prediction model includes a front-end frequency domain refining module, a channel-space joint-log-linear Transformer model, a back-end frequency domain refining module, and a full convolutional network.

[0041] an evaluation and early warning platform connected to the service life evaluation module, used for visualizing and displaying the service life evaluation indexes, generating evaluation reports, and presetting threshold values, and issuing early warning signals when the prediction results exceed the threshold values.

[0042] In the conventional technical field, the service life evaluation module processes multi-dimensional feature data and outputs service life evaluation indexes, and the process specifically includes the following steps:

[0043] In the conventional technical field, the service life evaluation module processes multi-dimensional feature data and outputs service life evaluation indexes, and the process specifically includes the following steps:

[0044] Step R1: a double-branch HCT (Hybrid CNN-Transformer) structure with shared parameters is introduced, local time-space features of the multi-dimensional feature data are extracted through a convolutional neural network, and global dynamic dependency features of the local time-space features are captured through a Transformer; a learnable weight factor is introduced to balance the local time-space features and the global dynamic dependency features, and multi-scale original features are output;

[0045] Step R2: the multi-scale original features are input into a front-end frequency domain refining module, and the features are filtered and enhanced in a frequency domain space through fast Fourier transform, a weighted gate mechanism, and inverse Fourier transform, high-discriminative frequency domain components are extracted, first-order refined features are obtained, and original spatial details are retained through a residual structure;

[0046] Step R3: a Transformer model is established, a multi-head self-attention mechanism and a feedforward network structure are introduced, global dependency modeling is performed on the first-order refined features, and a semantic difference map is obtained;

[0047] Step R4: the semantic difference map is input into a back-end frequency domain refining module, frequency domain filtering and residual superposition are performed on the semantic difference map in the module, second-order semantic difference features are obtained, and the boundary differences between the degradation areas and the healthy areas are further strengthened;

[0048] Step R5: inputting the second-order semantic differential feature into a full convolution network to output an initial service life evaluation index; in the service life prediction model training process, a cross-entropy loss function is used to optimize the service life prediction model to constrain the dynamic fluctuation of test error, improve the generalization performance of the model, correct the initial service life evaluation index, and output the service life evaluation index.

[0049] In the present application, the process of the service life evaluation module processing multi-dimensional feature data through the service life prediction model to output the service life evaluation index specifically includes the following steps:

[0050] Step S1: introducing a double-branch HCT (Hybrid CNN-Transformer) structure with shared parameters, extracting local space-time features of multi-dimensional feature data through a convolutional neural network, and capturing global dynamic dependency features of the local space-time features through a Transformer; introducing a learnable weight factor to balance the local space-time features and the global dynamic dependency features, and outputting multi-scale original features;

[0051] Step S2: inputting the multi-scale original features into a front-end frequency domain refining module, filtering and enhancing the features in the frequency domain space through fast Fourier transform, a weighted gate mechanism, and inverse Fourier transform, extracting high-discriminative frequency domain components, obtaining first-order refined features, and retaining original spatial details through a residual structure;

[0052] Step S3: introducing a channel and space joint attention mechanism based on KAN to replace the self-attention layer of the Transformer model on the basis of the Transformer model, combining a Fenwick tree hierarchical mask mechanism and a sequential semi-split mask to form a logarithmic linear expansion structure, and constructing a channel and space joint-logarithmic linear Transformer model to input the first-order refined features into the channel and space joint-logarithmic linear Transformer model to obtain a semantic differential graph; the channel and space joint-logarithmic linear Transformer model includes a local attention layer and a global attention layer;

[0053] Step S4: inputting the semantic differential graph into a back-end frequency domain refining module, performing frequency domain filtering and residual superposition on the semantic differential graph in the module to obtain second-order semantic differential features for further strengthening the boundary difference between the degradation area and the healthy area;

[0054] Step S5: inputting the second-order semantic differential feature into a full convolution network to output an initial service life evaluation index; in the service life prediction model training process, a difference generalization loss function is used to optimize the service life prediction model to constrain the dynamic fluctuation of test error, improve the generalization performance of the model, correct the initial service life evaluation index, and output the service life evaluation index, and the formula used is as follows:

[0055] Differential Generalization Loss Function:

[0056] ;

[0057] where, denotes the Differential Generalization Loss Function, denotes the traditional loss function, including cross-entropy and mean square error; denotes the weight coefficient (hyperparameter), controlling the influence strength of the regularization term in the total loss; denotes the total number of iterations in the training process, denotes the index variable in the summation process, denotes the test error at the th iteration, denotes the test error at the th iteration; denotes the regularization term, which suppresses the sharp fluctuations of the test error in the training process, ensures the smoothness of the error curve, and improves the generalization ability.

[0058] Differential Explanation:

[0059] In the "conventional technical field" scheme:

[0060] Step R3 uses the traditional Transformer model, with the core being the multi-head self-attention mechanism and the feedforward network, for global dependency modeling;

[0061] Step R5 uses the cross-entropy loss function as the main optimization objective in the training phase, and the model focuses on classification accuracy, but does not specifically handle error fluctuations and generalization stability;

[0062] In the "application" scheme:

[0063] Step S3 replaces the traditional self-attention with a channel and spatial joint attention mechanism based on KAN (Kolmogorov-Arnold Networks), and introduces a Fenwick tree hierarchical masking mechanism and sequential semi-split masking, forming a logarithmic linear expansion of efficient structure; This improvement improves the model's ability to jointly model across channels and spatial dimensions, while significantly reducing computational complexity, making it more efficient and robust to model large-scale features, and enabling more precise differentiation between degraded and healthy regions;

[0064] Step S5 introduces a differential generalization loss function at the loss function level, which adds a regularization term based on the difference of iterative errors to the traditional cross-entropy and mean square error; This regularization term constrains the dynamic fluctuations of the test error, making the error curve smooth, enhancing the model's generalization ability and the stability of the prediction results.

[0065] Embodiment three, this embodiment is based on embodiment two,

[0066] In the conventional technical field, the process of obtaining a semantic difference map specifically includes the following steps:

[0067] Step Q1: a CKSA unit is constructed by jointly using KAN (Kolmogorov-Arnold Network) channel attention and KAN (Kolmogorov-Arnold Network) spatial attention, the CKSA unit is used to weight and compress the first-order refined features, highlight the key areas and channel features sensitive to performance degradation, and generate a compact semantic token set;

[0068] Step Q2: the compact semantic token set is sequentially modeled, the semantic tokens in the compact semantic token set are arranged in the order of the time dimension to form a time sequence; for the time sequence, a conventional sliding window and position encoding method are used to aggregate the features of different time periods to generate a multi-scale bucket feature representation;

[0069] Step Q3: for the multi-scale bucket feature representation, an adaptive weight factor is introduced for fusion processing to obtain a fused feature sequence; a local attention layer of a Transformer model is used to model the intra-group and inter-group global semantic relationships between tokens in the fused feature sequence; this modeling process focuses on short-time sequence and local-range dependency relationships to capture fine-grained and near-end degradation dynamic features, and obtains local semantic tokens;

[0070] Step Q4: in the global attention layer, a standard multi-head self-attention mechanism is used to globally model the time sequence, the long-time sequence dependency relationship is depicted through attention weight distribution, and a global semantic token is generated;

[0071] Step Q5: define a pixel space, map the dual semantic tokens of the local semantic tokens and the global semantic tokens back to the pixel space to generate a dual semantic pixel map, and obtain a semantic difference map through difference calculation; the semantic difference map is used to highlight the difference features between the performance degradation area and the healthy area, and provides a significant contrast basis for subsequent service life assessment.

[0072] In the present application, the process of obtaining a semantic difference map specifically includes the following steps:

[0073] Step S31: A CKSA unit is constructed by combining KAN (Kolmogorov-Arnold Network) channel attention and KAN (Kolmogorov-Arnold Network) spatial attention, the CKSA unit is used to compress the first-order refined features by weighting, highlight the key areas and channel features sensitive to performance degradation, and generate a compact semantic token set;

[0074] The process of generating a compact semantic token set by compressing the first-order refined features by weighting using the CKSA unit is as follows: the CKSA unit takes the first-order refined features as input, constructs a joint mechanism of channel attention and spatial attention through the KAN layer, in the channel dimension, uses the KAN operator to calculate the importance weight of each channel, and performs channel weighting on the first-order refined features to highlight the feature types sensitive to performance degradation; in the spatial dimension, the KAN operator is used to generate a spatial weight distribution, and the weighted channel features are further spatially weighted to highlight the key areas and change areas of the sand barrier structure, to obtain a comprehensive weighted feature, and to perform compression mapping to convert the comprehensive weighted feature into a compact semantic token set;

[0075] KAN channel attention = use KAN operator instead of FC layer to learn the importance weight of each channel;

[0076] KAN spatial attention = use KAN to fit complex spatial patterns to determine the importance of each position;

[0077] The CKSA unit processes the channel dimension and the spatial dimension simultaneously through the KAN structure, so that the semantic token set obtained is more compact and focuses on the degradation-related areas;

[0078] In traditional neural networks, a layer mainly consists of linear mapping (matrix multiplication) + nonlinear activation function (ReLU, GELU, etc.);

[0079] Step S32: Time series modeling is performed on the compact semantic token set, the semantic tokens in the compact semantic token set are arranged in time dimension to form a time series; for the time series, a time series prefix interval is introduced at each time, and a hierarchical division mechanism based on Fenwick tree is used to divide the time series prefix interval into multiple scale buckets, and feature aggregation is performed to generate multi-scale bucket feature representation; this mechanism can model the near-end information at high resolution while modeling the far-end information at low resolution, thus balancing fine-grained changes and macro trends;

[0080] A hierarchical partitioning mechanism based on Fenwick trees is adopted to divide the prefix interval of the time series into multiple scale buckets. The formula used is as follows:

[0081] ;

[0082] in, Indicates the current time step. This represents the scale-level index under the Fenwick tree hierarchical mechanism. Indicates at time At that time, according to the Fenwick tree hierarchical mechanism, in the first... The time-scale buckets are divided into layers at each scale level. Each time-scale bucket stores a batch of historical tokens for hierarchical modeling. This represents an index variable used to iterate through time-scale buckets. Historical tokens within; Indicates the first The Value vector of each historical token is understood in this application as the performance characterization value at that moment (such as the stress state of the sand barrier and the characteristics of wind and sand flow). Indicates transpose. This indicates that it belongs to the time-scale bucket. Inner Transpose of the key vector of each historical token; Indicates at time , by the Time-scale buckets of each scale layer The hidden state matrix obtained by multiplying the value vector of all historical tokens with the transpose of the key vector and aggregating them is the multi-scale bucket feature representation.

[0083] This layer only processes a single scale bucket, similar to a local information compressor;

[0084] Step S33: For the multi-scale bucket feature representation, an adaptive weight factor is introduced for fusion processing to obtain the fused feature sequence. The fused feature sequence is a sequence composed of multiple groups of tokens. Tokens in the same group constitute intra-token dependencies, and tokens across groups constitute inter-token dependencies. The intra-token dependencies and inter-token dependencies are modeled through a local attention layer, focusing on short-term and local-range dependencies, capturing fine-grained, near-end degenerate dynamic features, and obtaining local semantic tokens.

[0085] Step S34: In the global attention layer, the sequentially semi- separable (SSS) mask matrix generated based on the gating structure and the Fenwick tree hierarchical mask are combined to form a Log-Linear expansion mechanism, and the global semantic token is obtained by processing the fused feature sequence through the Log-Linear expansion mechanism; the mechanism can effectively reduce the computational complexity while maintaining the model expression ability, and realize efficient modeling of long-time sequence dependence; through the expansion design, the existing linear attention models with gating memory and semi-separable structure, such as Mamba-2 and GatedDeltaNet, can be seamlessly upgraded to Log-Linear variants, thereby further improving the computational efficiency and stability in long sequence tasks;

[0086] In the global attention layer, first, a sequentially semi-separable (SSS) mask matrix is generated based on the gating structure , to realize ordered memory and dynamic forgetting of time series information; then a Fenwick tree hierarchical mask matrix is constructed , to perform multi-scale division and hierarchical constraint on the time series prefix interval, ensuring that the near-end features are modeled with high resolution and the far-end features are modeled with low resolution; the two mask matrices are combined element by element to form a Log-Linear expansion mask , which is embedded into the attention calculation process to replace the traditional dense mask operation; through this mechanism, the computational complexity can be reduced from quadratic to logarithmic linear while maintaining the model expression ability, realizing efficient modeling of long-time sequence dependence; finally, the input token sequence is modeled using the expanded attention mechanism, and the output semantic representation is the global semantic token, which represents the macroscopic degradation trend of the anti-sand system under long-term environmental action, and the formula is as follows:

[0087] Attention calculation formula (Log-Linear expanded attention):

[0088] ;

[0089] wherein, represents the output sequence after Log-Linear expansion, represents the query matrix, represents the key matrix, represents the value matrix; represents Hadamard (element-wise) multiplication;

[0090] Global semantic token aggregation formula:

[0091] ;

[0092] wherein, represents the time step The output vector, i.e., the global semantic token, This represents the total number of scale layers obtained under the Fenwick tree hierarchical mechanism; Indicates at time step For the The weighting factors assigned to the time-scale buckets of each scale layer; Indicates at time step The transpose of the generated query vector;

[0093] Step S35: Define a pixel space, map the dual semantic tokens of local semantic tokens and global semantic tokens back to the pixel space to generate a dual semantic pixel map, and obtain a semantic difference map through differential calculation; this semantic difference map is used to highlight the difference features between the performance degradation area and the healthy area, and provide a significant comparative basis for subsequent service life assessment.

[0094] Explanation of differences:

[0095] In the "conventional technical field" solution:

[0096] Temporal modeling (step Q2):

[0097] We employ conventional sliding window and positional encoding to construct multi-scale bucket features for time series. While this approach can capture features across multiple time periods, its division of time intervals is relatively static and cannot dynamically adapt to the temporal dependency differences in complex degradation processes.

[0098] Global dependency modeling (step Q4):

[0099] Using the standard multi-head self-attention mechanism to directly model long time series lacks structured mask constraints, and attention allocation is prone to diffusion, resulting in insufficient modeling efficiency and redundant computation in long sequence scenarios.

[0100] In the scheme of "this application":

[0101] Temporal modeling (step S32):

[0102] At each time step, a time series prefix interval is introduced, and the prefix interval is divided into multi-scale buckets through a hierarchical partitioning mechanism based on Fenwick trees. Compared with the conventional sliding window, this mechanism can achieve multi-scale aggregation with logarithmic complexity, making the feature aggregation process more hierarchical and adaptive, and improving the ability to characterize complex temporal dependencies.

[0103] Global dependency modeling (step S34):

[0104] The Log-Linear expansion mechanism is constructed by combining the sequential semi-separable mask matrix and the Fenwick tree hierarchical mask matrix; compared with the conventional multi-head self-attention, the mechanism can not only retain the long-time sequence dependency, but also significantly reduce the computational complexity (from quadratic to approximate logarithmic linear), thereby improving the global modeling efficiency and stability.

[0105] In the fourth embodiment, according to any one of the first to third embodiments, Figure 2 、 Figure 3 In the fourth embodiment, according to any one of the first to third embodiments, in the fourth embodiment, the evaluation and early warning platform is connected to the service life evaluation module, and is used for visualizing the service life evaluation indexes, generating an evaluation report, and presetting a threshold value, and when the prediction result exceeds the threshold value, an early warning signal is sent.

[0106] Figure 2 Figure 2 shows the trend of the remaining life of different regions of the A sand barrier protection belt, Figure 2 In the fourth embodiment, according to any one of the first to third embodiments, in the fourth embodiment, the evaluation and early warning platform is connected to the service life evaluation module, and is used for visualizing the service life evaluation indexes, generating an evaluation report, and presetting a threshold value, and when the prediction result exceeds the threshold value, an early warning signal is sent.

[0107] Figure 3 Figure 3 shows the trend of the risk probability of different regions of the A sand barrier protection belt, Figure 3 In the fourth embodiment, according to any one of the first to third embodiments, in the fourth embodiment, the evaluation and early warning platform is connected to the service life evaluation module, and is used for visualizing the service life evaluation indexes, generating an evaluation report, and presetting a threshold value, and when the prediction result exceeds the threshold value, an early warning signal is sent.

[0108] The above describes the present application and its embodiments, which are not limited, and the drawings only show one of the embodiments of the present application, and the actual structure is not limited thereto; in general, if a person skilled in the art is inspired thereby, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical solution can be designed, which should belong to the protection scope of the present application.

Claims

1. A deep learning-based comprehensive sand prevention system service life evaluation system, characterized in that: The system comprises a data acquisition module, a data preprocessing and fusion module, and a service life evaluation module. The data acquisition module acquires multi-source heterogeneous data of the comprehensive sand prevention system, and the multi-source heterogeneous data comprises remote sensing image data, near-ground sensor monitoring data, meteorological environment data, and sand prevention system structure attribute data. The data preprocessing and fusion module processes the multi-source heterogeneous data to generate multi-dimensional feature data. The service life evaluation module is connected to the data preprocessing and fusion module, and has a built-in service life prediction model. The service life prediction model comprises a front-end frequency domain refining module, a channel-space joint-log-linear Transformer model, a rear-end frequency domain refining module, and a full convolutional network. The channel-space joint-log-linear Transformer model is constructed to obtain a semantic difference map, specifically comprising: Step S31: extracting first-order refined features of the multi-dimensional feature data, constructing a CKSA unit, weighting and compressing the first-order refined features by using the CKSA unit, and generating a compact semantic token set; Step S32: time-series modeling of the compact semantic token set, forming a time series according to the order on the time dimension in the compact semantic token set; introducing a time series prefix interval at each time for the time series, and using a hierarchical division mechanism based on a Fenwick tree to divide the time series prefix interval into multiple scale buckets and perform feature aggregation to generate multi-scale bucket feature representations; Step S33: introducing an adaptive weight factor for fusion processing of the multi-scale bucket feature representations to obtain a fused feature sequence; the fused feature sequence is a sequence composed of multiple tokens, and tokens in the same group constitute token group dependency and tokens across groups constitute token group dependency; token group dependency and token group dependency are modeled by a local attention layer, focusing on short-time sequence and local range dependency, capturing fine-grained and near-end degradation dynamic features, and obtaining local semantic tokens; Step S34: in the global attention layer, introducing a sequential semi-separable mask matrix and a Fenwick tree hierarchical mask matrix for combination to form a Log-Linear expansion mechanism, and processing the fused feature sequence by the Log-Linear expansion mechanism to obtain global semantic tokens; Step S35: defining a pixel space, mapping the dual semantic tokens of the local semantic tokens and the global semantic tokens back to the pixel space to generate a dual semantic pixel map, and obtaining a semantic difference map through difference calculation.

2. The deep learning-based service life evaluation system for a comprehensive sand prevention system according to claim 1, characterized in that: The service life evaluation module outputs the service life evaluation index, and the process comprises the following steps: Step S1: introducing a double-branch HCT structure of shared parameters, extracting local spatiotemporal features of multi-dimensional feature data through a convolutional neural network, and capturing global dynamic dependency features of the local spatiotemporal features through a Transformer; introducing a learnable weight factor to balance the local spatiotemporal features and the global dynamic dependency features, and outputting multi-scale original features; Step S2: inputting the multi-scale original features into a front-end frequency domain refining module to obtain first-order refined features; Step S3: inputting the first-order refined features into a channel-spatial joint-log-linear Transformer model to obtain a semantic difference map; the channel-spatial joint-log-linear Transformer model includes a local attention layer and a global attention layer; Step S4: inputting the semantic difference map into a back-end frequency domain refining module to obtain second-order semantic difference features; Step S5: inputting the second-order semantic difference features into a full convolutional network to output an initial service life evaluation index; using a difference generalization loss function to optimize the service life prediction model, correct the initial service life evaluation index, and output a service life evaluation index.

3. The deep learning-based integrated sand prevention system service life evaluation system according to claim 1, characterized in that: A CKSA unit is constructed by combining KAN channel attention and KAN spatial attention.

4. The deep learning-based integrated sand prevention system service life evaluation system according to claim 3, characterized in that: The process of using the CKSA unit to weight and compress the first-order refined features to generate a compact semantic token set is as follows: the CKSA unit takes the first-order refined features as input, in the channel dimension, uses the KAN operator to calculate the importance weight of each channel, and performs channel weighting on the first-order refined features to generate weighted channel features; in the spatial dimension, the KAN operator is used to generate a spatial weight distribution, and the weighted channel features are further spatially weighted to obtain comprehensive weighted features, which are compressed and mapped to convert the comprehensive weighted features into a compact semantic token set.

Images