Loess landslide displacement prediction method and system based on MSSO-ST-Transformer model

Abstract

The invention belongs to the technical field of loess landslide treatment, and discloses a loess landslide displacement prediction method based on an MSSO-ST-Transform model, and the method can adaptively adjust the hyper-parameter configuration of the ST-Transform model through an MSSO algorithm, and improves the precision and applicability of the whole model in landslide displacement prediction. A multi-strategy sparrow optimization algorithm (MSSO) serves as an efficient global optimization method, an optimal combination can be found in a complex hyper-parameter search space, and an ST-Transform model accurately captures spatial features (topography and geologic structures) and long time dependence (multi-day displacement trend) of landslide displacement through spatial Patch embedding and a Transform self-attention mechanism at the same time.

Description

Technical Field

[0001] This invention belongs to the field of loess landslide control technology, and in particular relates to a method and system for predicting loess landslide displacement based on the MSSO-ST-Transformer model. Background Technology

[0002] Landslide displacement prediction is a crucial component of geological disaster prevention and control. It not only concerns the safety of people's lives and property but also directly impacts socio-economic stability and development. Accurate landslide displacement prediction allows for effective early warning, providing a scientific basis for relevant departments to formulate disaster prevention and mitigation measures, thereby minimizing or avoiding losses caused by landslides. Furthermore, landslide displacement prediction can help optimize land use planning and guide engineering construction to avoid high-risk areas. Secondly, it is vital for ensuring the safety of infrastructure. Prediction can assess the potential threat of landslides to important facilities such as highways, railways, and dams, allowing for timely reinforcement or design adjustments to prevent large-scale damage caused by geological disasters. Therefore, strengthening the research and application of landslide displacement prediction technology is of great significance for improving natural disaster response capabilities and promoting sustainable development.

[0003] Currently, the field of landslide displacement prediction has developed various advanced methods and technologies, such as machine learning, deep learning, swarm intelligence, ensemble learning, and big data and cloud computing. Despite significant achievements driven by technological advancements, some common drawbacks and challenges remain, including:

[0004] Limited model generalization ability: Many existing predictive models, especially machine learning-based methods, often perform well on training data but lack generalization ability when faced with new, unseen data. This is because these models may overfit specific datasets and lack adaptability to complex and variable geological environments.

[0005] Poor adaptability to dynamic environments: The geological environment is dynamic and changing, while many prediction models assume relatively stable environmental conditions. When faced with rapidly changing environmental factors (such as extreme weather events), the prediction accuracy of the models may decrease significantly.

[0006] Multi-source data fusion is challenging: Landslide displacement prediction involves various types of data, including geological, meteorological, and remote sensing information. Effectively fusing these heterogeneous data to extract valuable features and build robust prediction models remains a difficult problem to solve.

[0007] The integration of swarm intelligence and machine learning has revolutionized landslide displacement prediction, significantly improving its accuracy and efficiency. Swarm intelligence, by simulating the collaborative behavior of biological groups in nature, enables distributed data processing and rapid information aggregation, while machine learning excels at mining patterns from complex data, learning features, and building predictive models. This combination not only handles massive amounts of multi-source geological monitoring data but also continuously optimizes itself to adapt to environmental changes and new data inputs. In the field of landslide displacement prediction, this fusion technology comprehensively considers multiple influencing factors such as topography, soil properties, precipitation, and seismic activity. Through deep learning of historical landslide events using machine learning algorithms, it identifies potential risk patterns. Swarm intelligence further enhances this process, accelerating the search for and determination of the optimal prediction model through the cooperation of intelligent agents, making the early warning system more sensitive and accurate. Furthermore, this integrated approach promotes the adaptability and scalability of the prediction system, enabling dynamic adjustments to prediction strategies based on the latest monitoring data, effectively addressing the suddenness and uncertainty of geological disasters.

[0008] Traditional temporal models (such as LSTM) can capture temporal dependencies in landslide displacement prediction, but their spatial feature extraction capabilities are limited. Single convolutional models (such as CNN), while adept at spatial feature extraction, struggle to handle long-term temporal dependencies. The Spatiotemporal Transformer (ST-Transformer), as a state-of-the-art model for spatiotemporal sequence prediction, utilizes a dual-module approach of "spatial feature extraction + Transformer self-attention." This retains the accurate capture of spatial information such as terrain and geological structures while addressing the vanishing gradient problem of long-term sequences inherent in traditional LSTMs through a self-attention mechanism, thus efficiently capturing long-distance dependencies in the temporal dimension. However, using the ST-Transformer alone still has limitations:

[0009] High sensitivity to hyperparameters: The core hyperparameters of ST-Transformer (such as patch size, number of encoder layers, number of attention heads, etc.) have a significant impact on model performance. Manually tuning the parameters makes it difficult to find the optimal combination, which can easily lead to underfitting or overfitting of the model.

[0010] Optimization space for computational resource requirements: When processing large-scale landslide monitoring data, the unoptimized ST-Transformer may lead to computational redundancy due to unreasonable hyperparameter configuration, increasing resource consumption.

[0011] The Snake Optimization Algorithm (SO) is a metaheuristic optimization algorithm that excels at globally searching for the optimal solution and effectively avoids getting trapped in local optima. SO possesses strong adaptability, dynamically adjusting its optimization strategy based on different geological environments and monitoring data. This adaptability allows the network model to maintain high prediction accuracy and stability even in complex and ever-changing landslide environments. Optimizing the network model using SO can reduce the number of iterations and computational resource consumption during model training. This is particularly important in resource-constrained environments, making real-time or large-scale landslide displacement prediction possible.

[0012] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0013] Using the ST-Transformer alone has many drawbacks in terms of hyperparameter configuration, generalization ability, computational resource requirements, and sensitivity to outliers and noise, which limit the model's performance and reliability in practical applications. Although the SO algorithm has achieved good results, it still has some shortcomings, such as overly random population initialization, low population diversity, and a tendency for premature convergence. These shortcomings are mainly due to the fixed values ​​of parameters in the algorithm, the singularity of the optimization direction, the continuous operation of individuals on the optimal position, the imbalance between exploration and development capabilities, and the low probability of large spatial jumps during iterative population updates. Summary of the Invention

[0014] To address the problems existing in the prior art, this invention provides a method for predicting loess landslide displacement based on the MSSO-ST-Transformer model.

[0015] This invention is implemented as follows: A method for predicting loess landslide displacement based on the MSSO-ST-Transformer model includes:

[0016] Step 1: Preprocess the data to generate training samples and construct the ST-Transformer network;

[0017] Step 2: Establish the objective function and calculate the amount of food Q and temperature Temp in SO;

[0018] Step 3: Initialize the population using the optimal point set strategy;

[0019] Step 4: If the amount of food Q < 0.25, then execute the foraging strategy;

[0020] Step 5: If the amount of food Q > 0.25 and Temp > 0.6, enter the feeding stage and update the individual's position using a dynamic update selection mechanism;

[0021] Step 6: If the food quantity Q > 0.25, Temp < 0.6, and Rand > 0.6, the individual enters the combat phase, and the population is updated.

[0022] Step 7: If the food quantity Q>0.25, Temp<0.6 and Rand<0.6, the individual enters the mating and egg-laying stage and the population is iteratively updated.

[0023] Step 8: Finally, the current optimal position is perturbed using a quantum computing-based T-distribution mutation, and the optimal solution is updated.

[0024] Step 9: Determine if the iteration termination condition is met. If not, return to step 4; if met, proceed to the next step.

[0025] Step 10: Interpret the optimal solution obtained from the search as the hyperparameter configuration of ST-Transformer, train the ST-Transformer model with training samples, and obtain the final prediction model.

[0026] Furthermore, the establishment of the preliminary ST-Transformer model mentioned in step 1 refers to determining the composition structure of the model. A typical structure includes: input layer → patch embedding layer (spatial feature extraction) → position encoding layer → Transformer encoder layer (including multi-head attention module, layer normalization module, and feed-forward network module) → temporal aggregation layer → fully connected output layer.

[0027] Furthermore, the process of initializing the search population in step 2 is as follows: Temperature and food quantity are two very important factors, and their calculation formulas are shown in (1) and (2):

[0028] (1)

[0029] In equation (1), t represents the current iteration number, and T represents the maximum iteration number.

[0030] in The calculation formula is shown in (2).

[0031] (2)

[0032] Equation (2) z This represents a constant with a value of 0.5.

[0033] Furthermore, the process of initializing the search population in step 3 is as follows: Within a dim-dimensional Euclidean unit cube, select... One excellent point ;in, = ,in To meet The smallest prime number; then the position of each individual is updated as shown in equation (3):

[0034] (3)

[0035] in, for The decimal part, Representing the Individual; These represent the upper and lower bounds of the problem, respectively; the given population size N is divided into two subpopulations: a male subpopulation and a female subpopulation; the subpopulation sizes are respectively... and And satisfy .

[0036] Furthermore, the specific process of the foraging strategy described in step 4 includes the following steps: Step 4-1: Calculate the adaptive spiral search operator :

[0037] (4)

[0038] (5)

[0039] Where b is a random number between [0,1] It is used to control the shape of the spiral;

[0040] Step 4-2: Calculate the fitness value for each individual;

[0041] Step 4-3: When At 0.25, the snake will choose a random individual to search for food;

[0042] (6)

[0043] (7)

[0044] in, This represents the i-th male individual. z represents a random individual among males. It is a constant with a value of 0.05. This represents the i-th female individual. Represents a random individual among the females; Representative of individuals fitness value, Representative of individuals fitness value, Representative of individuals fitness value, Representative of individuals The fitness value.

[0045] Furthermore, the process of dynamically updating the selection mechanism described in step 5 includes the following steps:

[0046] Step 5-1: Calculate using formula (8) Flight stride:

[0047] (8)

[0048] Step 5-2: In formula (8) and They all follow a Gaussian distribution:

[0049] (9)

[0050] (10)

[0051] (11)

[0052] In formula (10) It is a standard gamma function. It is usually set to 1.5;

[0053] Step 5-3: When When the value is greater than 0.6, the position update formula for the snake to move closer to the food location is shown in (12).

[0054] (12)

[0055] in The position representing the best individual. Represents the i-th individual of the male (or female) dimension, It is a constant with a value of 2. The individual represents the worst position in the global current position;

[0056] The position update process for the combat phase described in step 6 is as follows: when the temperature When Rand < 0.6 and Rand > 0.6, the snake will enter battle mode; the position update of battle mode is shown in formulas (13) and (14):

[0057] (13)

[0058] (14)

[0059] and These represent the worst male and the worst female individuals, respectively. The position representing the best individual among the males. This represents the fitness value of the best individual among the females. This represents the fitness value of the i-th individual. The fitness value represents the best individual among the males;

[0060] The location update process during the mating and spawning phase described in step 7 includes the following steps:

[0061] Step 7-1: When the food quantity Q>0.25, Temp<0.6, and Rand<0.6, the snake will exhibit mating behavior; the mating pattern can be described by formulas (15) and (16):

[0062] (15)

[0063] (16)

[0064] Step 7-2: Find the least fit individual among the males and females;

[0065] Step 7-3: After mating, the individual will enter the egg-laying stage. During this process, the individual with the worst fitness among the male and female individuals will be replaced; the update formula is shown in (17) and (18).

[0066] (17)

[0067] (18)

[0068] The quantum computing-based T-distribution mutation process described in step 8 includes the following steps:

[0069] Step 8-1: Construct the mutation factor by establishing a quantum space consisting of multiple qubits on the same dimension as the problem, as shown in formula (19).

[0070] = (19)

[0071] in the formula Let dim be the dimension of the search space. Describes a two-dimensional column vector of units, where ;

[0072] Step 8-2: Use formulas (20) and (21) to find the current optimal solution. Mapped into quantum space

[0073] (20)

[0074] (twenty one)

[0075] In formulas (20) and (21) Represents the current optimal position. dimension, This represents the magnitude of the variable, which is used to ensure diversity after mutations. The probability is usually set to 1 or -1;

[0076] Step 8-3: Transform the quantum space into the corresponding solution space;

[0077] = * (twenty two)

[0078] = * (twenty three)

[0079] Step 8-4: Calculate the quantum rotation factor

[0080] (twenty four)

[0081] In formula (24) ;

[0082] Step 8-5: Add the quantum rotation factor to the corresponding solution space.

[0083] = (25)

[0084] Step 8-6: Map the newly created quantum space to the corresponding solution space; finally, update the position of the global optimum as shown in formula (26):

[0085] (26)

[0086] In formula (26) and This represents a point mapped to the solution space after quantum rotation. To show obedience - The variance factor of the distribution, with degrees of freedom of (Current iteration number), individual position is updated through elite selection.

[0087] Another objective of this invention is to provide a loess landslide displacement prediction system based on the MSSO-ST-Transformer model, comprising:

[0088] The preprocessing module is used to preprocess the data, generate training samples, and construct the ST-Transformer network;

[0089] The calculation module is used to establish the objective function and calculate the amount of food Q and the temperature Temp in SO;

[0090] The initialization module is used to initialize the population using the best point set strategy;

[0091] The update module is used to execute a foraging strategy if the food quantity Q < 0.25; if the food quantity Q > 0.25 and Temp > 0.6, it enters the feeding phase and updates the individual's position using a dynamic update selection mechanism; if the food quantity Q > 0.25, Temp < 0.6, and Rand > 0.6, the individual enters the combat phase and the population is updated; if the food quantity Q > 0.25, Temp < 0.6, and Rand < 0.6, the individual enters the mating and egg-laying phase and iteratively updates the population; finally, it uses a T-distribution mutation based on quantum computing to perturb the current optimal position and update the optimal solution.

[0092] The judgment module is used to determine whether the iteration termination condition is met. If it is not met, return to step 4; if it is met, proceed to the next step.

[0093] The training module is used to interpret the optimal solution obtained from the search as the hyperparameter configuration of the ST-Transformer, train the ST-Transformer model with training samples, and obtain the final prediction model.

[0094] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, the computer program being executed by the processor causing the processor to perform the steps of the loess landslide displacement prediction method based on the MSSO-ST-Transformer model.

[0095] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the loess landslide displacement prediction method based on the MSSO-ST-Transformer model.

[0096] Another objective of this invention is to provide an information data processing terminal for implementing the loess landslide displacement prediction system based on the MSSO-ST-Transformer model.

[0097] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0098] The MSSO algorithm can adaptively adjust the hyperparameter configuration of the ST-Transformer model, improving the overall accuracy and applicability of the model in landslide displacement prediction. The Snake Optimization (MSSO) algorithm, as an efficient global optimization method, can find the optimal combination in a complex hyperparameter search space. The ST-Transformer model, through spatial patch embedding and Transformer self-attention mechanism, accurately captures both the spatial characteristics (topography, geological structure) and long-term temporal dependence (multi-day displacement trends) of landslide displacement. Combining the two effectively reduces model overfitting, improves generalization ability, and MSSO's global optimization capability reduces the ST-Transformer's sensitivity to hyperparameters, reducing computational redundancy. By combining MSSO and ST-Transformer, rapid and accurate prediction of landslide displacement data can be achieved, providing strong support for landslide early warning and disaster prevention.

[0099] The technical solution of this invention fills a technological gap in the domestic and international industry: In existing technologies, there are two major technological gaps in the field of landslide displacement prediction: First, the hyperparameter optimization of the ST-Transformer model lacks an efficient adaptive scheme, resulting in high hyperparameter sensitivity and large computational redundancy when used alone. Manual parameter tuning or traditional optimization algorithms cannot balance search efficiency and optimal solution quality. Second, swarm intelligence algorithms (such as SO) lack adaptability to landslide prediction scenarios, exhibiting defects such as random population initialization, an imbalance between exploration and development capabilities, and premature convergence, making deep integration with spatiotemporal sequence models difficult. This invention fills these gaps through three major innovations:

[0100] 1. For the first time, the improved snake optimization algorithm (MSSO) is deeply coupled with the ST-Transformer model, which specifically solves the problem of hyperparameter optimization of ST-Transformer, while making up for the shortcomings of the traditional SO algorithm, forming an integrated solution of "spatiotemporal feature capture + adaptive hyperparameter optimization";

[0101] 2. An innovative combination mechanism of optimal point set population initialization + multi-stage search strategy (foraging / eating / fighting / mating and laying eggs) + quantum T distribution mutation is proposed, which solves the industry pain point of "insufficient exploration and unstable convergence" of swarm intelligence algorithms in high-dimensional hyperparameter space, and provides a new path for hyperparameter optimization of spatiotemporal sequence models;

[0102] 3. It achieves the synergy of "dynamic environmental adaptability + high generalization ability + low computational resource consumption", breaking through the bottleneck of the decline in prediction accuracy of existing technologies under complex geological environments (such as extreme weather and multi-source data interference), and filling the gap in "high accuracy + strong robustness" prediction technology in the specific scenario of loess landslide. Attached Figure Description

[0103] Figure 1 This is a flowchart of the loess landslide displacement prediction method based on the MSSO-ST-Transformer model provided in this embodiment of the invention;

[0104] Figure 2 This is a structural block diagram of the loess landslide displacement prediction system based on the MSSO-ST-Transformer model provided in this embodiment of the invention;

[0105] Figure 3 This is a flowchart of the algorithm provided in an embodiment of the present invention;

[0106] Figure 4 This is a flowchart of the prediction model provided in an embodiment of the present invention;

[0107] Figure 5 These are the ST-Transformer initial models listed in the embodiments of the present invention;

[0108] Figure 6 These are the prediction results of the ST-Transformer models listed in the embodiments of the present invention;

[0109] Figure 7 This is a graph showing the prediction results of the MSSO-ST-Transformer provided in an embodiment of the present invention;

[0110] Figure 8 shows the convergence curves of MSSO (Multi-Strategy Snake Optimization) and traditional SO (Snake Optimization) in the hyperparameter search process proposed in this invention;

[0111] Figure 9 shows a bar chart comparing the performance of different models in the target prediction task. Detailed Implementation

[0112] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0113] In landslide monitoring projects in the Loess Hilly Area, long-term displacement sequences often exhibit strong nonlinearity, numerous abrupt change points, and significant influence from rainfall infiltration and seasonal environmental changes. Traditional prediction models rely on single time series or static feature inputs, making it difficult to accurately reflect the coordinated response effects of landslide bodies across a spatial scale. This leads to delayed judgments on landslide evolution trends, with prediction biases amplifying as the evolutionary stage progresses. Particularly in geological disaster early warning systems, the manual setting of model hyperparameters is highly dependent on experience, resulting in unstable model performance and a tendency for prediction non-convergence and false alarms during periods of concentrated rainfall, particularly in the flood season. Therefore, practical engineering applications urgently require a prediction system capable of simultaneously handling spatial structural characteristics and temporal dependencies, and possessing automated parameter optimization capabilities, to enhance the reliability of early warnings in complex geological environments.

[0114] This method uses the ST-Transformer as the sequence prediction framework, and analyzes the local spatial response and global temporal dependence in landslide monitoring data through patch embedding and multi-head attention mechanisms. In engineering monitoring systems, this structure enables the mining of correlations among multiple monitoring points, preserving the coupling characteristics of factors such as deep displacement, shallow deformation, and soil moisture content changes. However, because the Transformer structure itself is sensitive to parameter scale, direct application to landslide prediction can easily lead to slow convergence or overfitting. Therefore, it is integrated with a multi-stage snake swarm optimization mechanism. By using multi-stage behavioral strategies to identify search directions under different states, the model can obtain a more stable optimal combination in the high-dimensional hyperparameter space, significantly improving its generalization ability in real-world monitoring scenarios.

[0115] The parameter search space of landslide prediction models grows exponentially under large-scale monitoring data, making it difficult for a single global search method to balance exploration capability and local convergence efficiency. This method utilizes temperature and food availability as search state control factors, constructing four dynamic behaviors: foraging, feeding, combat, and mating / spawning. This allows the algorithm to adaptively switch search strategies based on evaluation metrics at different iteration stages. In low food availability stages, random search is enhanced to improve the ability to identify abnormal deformation trends. Under high-temperature conditions, the Levy step size is used to maintain a large-scale, leaping search, helping the algorithm find potential optimal solutions in complex spaces. As the temperature decreases, competitive behavior squeezes out weak local solutions, causing the convergence region to shrink towards stability. Mating and spawning mechanisms further eliminate low-quality solutions, maintaining the evolutionary activity of the population. This structure effectively avoids the premature convergence problem of traditional algorithms in the context of weakly stationary geological data.

[0116] In landslide displacement prediction applications, the initial distribution of hyperparameters has a decisive influence on the final convergence direction of the model. The optimal point set strategy constructs a low-bias, highly uniform initial population in a 1 / dim Euclidean space, preventing the search starting point from concentrating in local areas and increasing the global search probability. Simultaneously, dividing the population into male and female subsets allows for the introduction of differentiated update equations at different evolutionary stages, resulting in directional diversity in the search. For landslide monitoring scenarios, this mechanism enhances the model's robustness to noisy data or long-term, slowly changing trends, reducing the impact of extreme monitoring point disturbances, thereby ensuring the stability and coverage of the hyperparameter search phase.

[0117] To address the complex characteristics of landslide displacement sequences, such as frequent sudden events and spatial contraction of extreme value changes, this method constructs a multi-qubit mutation factor through quantum space mapping, enabling the optimal solution to acquire nonlinear perturbation capabilities in high-dimensional space. The heavy-tailed nature of the T-distribution is suitable for short-term surges or local jumps commonly found in engineering monitoring data, allowing for larger jumps during mutation phases, thus overcoming the narrow convergence region that easily forms in later iterations of the algorithm. Quantum rotation operations further enhance the exploration depth of the solution space, enabling the model to identify rapid displacement change trends. In geological disaster early warning systems, this mutation mechanism can effectively improve the ability to capture rapid accelerations in displacement during the pre-landslide stage.

[0118] The system standardizes and constructs samples from the original displacement sequences of monitoring stations through a preprocessing module, adapting the input structure to the ST-Transformer. In the computation module, the objective function corresponds to the quantized form of landslide prediction error, with food quantity and temperature updated in real-time by the iterative process to guide the optimization direction. The update module evolves candidate solutions according to behavioral patterns and maintains the exploratory activity of the search through quantum T-distribution mutations. After iteration terminates, the optimal solution is directly converted into the hyperparameter settings of the ST-Transformer for training the prediction model. This forms a complete closed loop from data preprocessing, intelligent optimization, model construction to high-precision prediction. In practical engineering applications, this system can reduce manual parameter tuning costs, improve the speed of flood season early warning response, and provide reliable and timely computational support for landslide risk management.

[0119] like Figure 1 As shown in the figure, the loess landslide displacement prediction method based on the MSSO-ST-Transformer model provided by this invention includes the following steps:

[0120] S101: Preprocess the data, generate training samples, and construct the ST-Transformer network;

[0121] S102: Establish the objective function and calculate the food quantity Q and temperature Temp in SO;

[0122] S103: Initialize the population using the optimal point set strategy;

[0123] S104: If the amount of food Q < 0.25, then execute the foraging strategy;

[0124] S105: If the amount of food Q>0.25 and Temp>0.6, enter the feeding stage and update the individual's position using a dynamic update selection mechanism;

[0125] S106: If food quantity Q > 0.25, Temp < 0.6, and Rand (random value) > 0.6, the individual enters the combat phase, and the population is updated;

[0126] S107: If the food quantity Q>0.25, Temp<0.6 and Rand<0.6, the individual enters the mating and egg-laying stage and the population is iteratively updated.

[0127] S108: Finally, the current optimal position is perturbed by the T-distribution mutation based on quantum computing, and the optimal solution is updated;

[0128] S109: Determine if the iteration termination condition is met. If not, return to step 4; if met, proceed to the next step.

[0129] S110: Interpret the optimal solution obtained from the search as the hyperparameter configuration of the ST-Transformer, train the ST-Transformer model with the training samples, and obtain the final prediction model.

[0130] The establishment of the preliminary ST-Transformer model in step 1 of the embodiments of the present invention refers to determining the composition structure of the model. The typical structure includes: input layer → patch embedding layer (spatial feature extraction) → position encoding layer → Transformer encoder layer (including multi-head attention module, layer normalization module, and feed-forward network module) → temporal aggregation layer → fully connected output layer.

[0131] The process of initializing the search population in step 2 provided in this embodiment of the invention is as follows: Temperature and food quantity are two very important factors, and their calculation formulas are shown in (1) and (2):

[0132] (1)

[0133] In equation (1), t represents the current iteration number, T represents the maximum iteration number, and exp represents the natural exponential function.

[0134] in The calculation formula is shown in (2).

[0135] (2)

[0136] Equation (2) z This represents a constant with a value of 0.5.

[0137] The process of initializing the search population in step 3 provided in this embodiment of the invention is as follows: In a dim-dimensional Euclidean unit cube, select... One excellent point ;in, = ,in To meet The smallest prime number; cos represents the cosine function; then the position of each individual is updated as shown in equation (3):

[0138] (3)

[0139] in, for The decimal part, Representing the Individual; These represent the upper and lower bounds of the problem, respectively; the given population size N (population size) is divided into two subpopulations: a male subpopulation and a female subpopulation; the subpopulation sizes are respectively... and And satisfy .

[0140] The specific process of the foraging strategy described in step 4 of the embodiments of the present invention includes the following steps: Step 4-1: Calculate the adaptive spiral search operator :

[0141] (4)

[0142] (5)

[0143] Where b is a random number between [0,1] It is used to control the shape of the spiral; pi represents Π;

[0144] Step 4-2: Calculate the fitness value for each individual;

[0145] Step 4-3: When At 0.25, the snake will choose a random individual to search for food;

[0146] (6)

[0147] (7)

[0148] in, This represents the i-th male individual. z represents a random individual among males. It is a constant with a value of 0.05. This represents the i-th female individual. Represents a random individual among the females; Representative of individuals fitness value, Representative of individuals fitness value, Representative of individuals fitness value, Representative of individuals fitness value, Represents the best individual. This represents the worst individual.

[0149] The process of dynamically updating the selection mechanism in step 5 provided in this embodiment of the invention includes the following steps:

[0150] Step 5-1: Calculate using formula (8) Flight stride:

[0151] (8)

[0152] Step 5-2: In formula (8) and They all follow a Gaussian distribution:

[0153] (9)

[0154] (10)

[0155] (11)

[0156] In formula (10) It is a standard gamma function. It is usually set to 1.5, where sin represents the sine function;

[0157] Step 5-3: When When the value is greater than 0.6, the position update formula for the snake to move closer to the food location is shown in (12).

[0158] (12)

[0159] in The position representing the best individual. (t) represents the i-th individual in the t-th generation of males (or females). dimension, It is a constant with a value of 2. The individual represents the worst position in the global current position;

[0160] Furthermore, the position update process for the combat phase described in step 6 is as follows: when the temperature When Rand < 0.6 and Rand > 0.6, the snake will enter battle mode; the position update of battle mode is shown in formulas (13) and (14):

[0161] (13)

[0162] (14)

[0163] and These represent the worst male and the worst female individuals, respectively. The position representing the best individual among the females. The position representing the best individual among the males. This represents the fitness value of the best individual among the females. This represents the fitness value of the i-th individual. The fitness value represents the best individual among the males. This represents the i-th individual in the t-th generation of females. This represents the i-th individual in the t-th generation of a male;

[0164] The location update process during the mating and spawning phase described in step 7 includes the following steps:

[0165] Step 7-1: When the food quantity Q>0.25, Temp<0.6, and Rand<0.6, the snake will exhibit mating behavior; the mating pattern can be described by formulas (15) and (16):

[0166] (15)

[0167] (16)

[0168] This represents the fitness of the i-th female. This represents the fitness of the i-th male. Let represent the fitness of the i-th individual. This represents the i-th female individual. This represents the i-th male individual. This represents the i-th individual among the males of generation t. This represents the i-th individual among the females of generation t;

[0169] Step 7-2: Find the least fit individual among the males and females;

[0170] Step 7-3: After mating, the individual will enter the egg-laying stage. During this process, the individual with the worst fitness among the male and female individuals will be replaced; the update formula is shown in (17) and (18).

[0171] (17)

[0172] (18)

[0173] The quantum computing-based T-distribution mutation process described in step 8 includes the following steps:

[0174] Step 8-1: Construct the mutation factor by establishing a quantum space consisting of multiple qubits in the same dimension as the problem, as shown in formula (19).

[0175] = (19)

[0176] in the formula Let dim be the dimension of the search space. Describes a two-dimensional column vector of units, where ;

[0177] Step 8-2: Use formulas (20) and (21) to find the current optimal solution. Mapped into quantum space

[0178] (20)

[0179] (twenty one)

[0180] In formulas (20) and (21) Represents the current optimal position. dimension, This represents the magnitude of the variable, which is used to ensure diversity after mutations. The probability is usually set to 1 or -1. Describing the j-th dimension , Describing the j-th dimension ;

[0181] Step 8-3: Transform the quantum space into the corresponding solution space;

[0182] = * (twenty two)

[0183] = * (twenty three)

[0184] Step 8-4: Calculate the quantum rotation factor

[0185] (twenty four)

[0186] In formula (24) ;

[0187] Step 8-5: Add the quantum rotation factor to the corresponding solution space.

[0188] = (25)

[0189] Step 8-6: Map the newly created quantum space to the corresponding solution space; finally, update the position of the global optimum as shown in formula (26):

[0190] (26)

[0191] In formula (26) and This represents a point mapped to the solution space after quantum rotation. To show obedience - The variance factor of the distribution, with degrees of freedom of (Current iteration number), individual position is updated through elite selection.

[0192] like Figure 2 As shown in the figure, an embodiment of the present invention provides a loess landslide displacement prediction system based on the MSSO-ST-Transformer model, comprising:

[0193] The preprocessing module is used to preprocess the data, generate training samples, and construct the ST-Transformer network;

[0194] The calculation module is used to establish the objective function and calculate the amount of food Q and the temperature Temp in SO;

[0195] The initialization module is used to initialize the population using the best point set strategy;

[0196] The update module is used to execute a foraging strategy if the food quantity Q < 0.25; if the food quantity Q > 0.25 and Temp > 0.6, it enters the feeding phase and updates the individual's position using a dynamic update selection mechanism; if the food quantity Q > 0.25, Temp < 0.6, and Rand > 0.6, the individual enters the combat phase and the population is updated; if the food quantity Q > 0.25, Temp < 0.6, and Rand < 0.6, the individual enters the mating and egg-laying phase and iteratively updates the population; finally, it uses a T-distribution mutation based on quantum computing to perturb the current optimal position and update the optimal solution.

[0197] The judgment module is used to determine whether the iteration termination condition is met. If it is not met, return to step 4; if it is met, proceed to the next step.

[0198] The training module is used to interpret the optimal solution obtained from the search as the hyperparameter configuration of the ST-Transformer, train the ST-Transformer model with training samples, and obtain the final prediction model.

[0199] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, the computer program being executed by the processor causing the processor to perform the steps of the loess landslide displacement prediction method based on the MSSO-ST-Transformer model.

[0200] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the loess landslide displacement prediction method based on the MSSO-ST-Transformer model.

[0201] Another objective of this invention is to provide an information data processing terminal for implementing the loess landslide displacement prediction system based on the MSSO-ST-Transformer model.

[0202] Specific implementation of the present invention:

[0203] The experimental platform used in this embodiment is a 64-bit Windows 10 system with MATLAB 2022a installed, an Intel Core i7-6700 processor, and 8GB of memory.

[0204] like Figure 3 The landslide displacement prediction method based on the MSSO-ST-Transformer model of the present invention specifically includes the following steps:

[0205] Step 1: Collect characteristic data that cause landslide displacement and clean the data.

[0206] The collected features include, but are not limited to, the following: landslide displacement distance, rainfall, temperature, humidity, and soil moisture.

[0207] Data cleaning methods include, but are not limited to, the following: standardizing the format / units, removing invalid data, removing unreasonable data, removing outliers and noise, and filling in data.

[0208] Step 2: Use all features from Step 1 as input to the model, and the displacement feature (landslide displacement distance) as the model output. For any sample, the true output value at time T is the input value of the displacement data at time T+stepN, where stepN is the step size. Synthesize data from period T to period T+stepN and input it as a sample into the ST-Transformer. In this embodiment, stepN=8 (predicting landslide displacement 8 days later), and the period T is in days, collecting landslide displacement data for more than 800 days.

[0209] The total sample was divided into training samples and test samples in a 6:4 ratio.

[0210] Step 3: Establish the initial ST-Transformer model. In this embodiment, the initial model is as follows: Figure 5 As shown, the first layer (input layer) receives data with a dimension of "number of samples × 8 (time steps) × 5 (number of features)"; the second layer (Patch embedding layer) divides the input into patches according to the spatial dimension (number of features) and converts them into vectors (e.g., when the patch size = 2, 5 features are divided into 3 patches, and the vector dimension = 2 × 8 = 16); the third layer (position encoding layer) generates position encodings with the same dimension as the patch vectors and adds them to the patch vectors; the fourth layer (Transformer encoder layer) contains a multi-head attention module (number of heads = 2), layer normalization (ε = 1e-6), and a Feed-Forward network (hidden layer dimension = 64).

[0211] The fifth layer (temporal aggregation layer): performs mean pooling on the "number of samples × number of patches × vector dimension" data output by the encoder to obtain features of "number of samples × vector dimension"; the sixth layer (fully connected layer): output dimension = 1, activation function is ReLU. In this embodiment, there are 4 hyperparameters to be optimized for ST-Transformer (search dimension dim = 4), as follows: Patch size: defined in the range [2, 5] (controls the granularity of spatial feature extraction); number of Transformer encoder layers: defined in the range [1, 3] (controls the temporal dependency capture capability); number of multi-head attention heads: defined in the range [2, 8] (controls the number of parallel attention mechanisms); dimension of hidden layers in Feed-Forward network: defined in the range [64, 256] (controls the feature nonlinear transformation capability).

[0212] Because the MSSO algorithm of this invention is a search algorithm in a continuous space, the lower bound Lb and upper bound Ub of the search space are respectively: Lb=[2, 1, 2, 64] (Patch size = 2, number of encoder layers = 1, number of attention heads = 2, Feed-Forward dimension = 64).

[0213] Ub=[5, 3, 8, 256] (Patch size = 5, encoder layers = 3, attention heads = 8, feed-forward dimension = 256). Note: In this embodiment, including the comparative experiments, the loss function of the ST-Transformer model is the mean squared error (MSE) function, and the backpropagation algorithm is the Adam algorithm.

[0214] Step 4: Set parameters for this method. Parameters that need to be set include: population size N, i.e., the number of male individuals. and number of female individuals Maximum number of iterations T; upper bound of search space Ub, lower bound Lb, search dimension dim.

[0215] In this embodiment, N=40 ( =20, =20), T=40, Ub, Lb, dim are obtained from step 3.

[0216] Let the number of iterations be t=1.

[0217] Step 5: Calculate two important parameters in the algorithm: food quantity Q and temperature Temp. The calculation formulas are shown below:

[0218] (1)

[0219] In equation (1), t represents the current iteration number, and T represents the maximum iteration number.

[0220] in The calculation formula is shown in (2).

[0221] (2)

[0222] Equation (2) z This represents a constant with a value of 0.5.

[0223] Step 6: Initialize the population

[0224] Step 6-1: Within the search space of step 4, generate One excellent point .in, = ,in To meet The smallest prime number. Then the position of each individual in the population is updated as shown in the following formula:

[0225] (3)

[0226] in, for The decimal part, Representing the Individual. These represent the upper and lower bounds of the search space, respectively. `dim` is obtained from step 4. This represents the position information of individual i in the j-th dimension.

[0227] Step 6-2: Calculate the fitness of each individual in the initialized population from Step 6-1, and select the position of the individual with the lowest fitness. Specifically, the population is traversed, and the value of each dimension at the position of individual i is converted into the corresponding hyperparameter in the ST-Transformer model; the model is trained using the training samples obtained in step 3; then, the test samples are substituted, and the total error is multiplied by 100 to obtain the fitness of individual i. Multiplying by 100 here is to improve the discriminative power.

[0228] Step 7: If the amount of food Q < 0.25, then execute the foraging strategy.

[0229] Step 7-1: Calculate the adaptive spiral search operator , The calculation formula is as follows:

[0230] (4)

[0231] (5)

[0232] Where b is a random number between [0,1] It is used to control the shape of the spiral, where t is the current iteration number and T is the maximum iteration number.

[0233] Step 7-2: Update the position of the individual using formulas (6) and (7).

[0234] (6)

[0235] (7)

[0236] In formulas (6) and (7), Let represent the i-th male individual, which is a 1*dim vector. z represents a random individual among males. It is a constant with a value of 0.05. This represents the i-th female individual. This refers to a random individual among the females. Representative of individuals fitness value, Representative of individuals fitness value, Representative of individuals fitness value, Representative of individuals The fitness value.

[0237] Step 8: When the food quantity Q>0.25 and Temp>0.6, the feeding stage begins.

[0238] Step 8-1: Calculate using formula (8) Flight stride:

[0239] (8)

[0240] Step 8-2: In formula (8) and They all follow a Gaussian distribution:

[0241] (9)

[0242] (10)

[0243] (11)

[0244] In formula (10) It is a standard gamma function. It is usually set to 1.5.

[0245] Step 8-3: The formula for updating the position of the snake as it moves closer to the food location is shown in (12).

[0246] (12)

[0247] in The position representing the best individual is a 1*dim dimensional vector. Represents the i-th individual of the male (or female) dimension, It is a constant with a value of 2. The individual represents the worst position in the global current position.

[0248] Step 9: When food quantity Q > 0.25, Temp < 0.6, and Rand > 0.6, the individual enters the combat phase. The position update formula for the combat phase is as follows:

[0249] (13)

[0250] (14)

[0251] Rand is a random number between [0, 1]. and These represent the worst male and the worst female individuals, respectively. The position representing the best individual among the males. This represents the fitness value of the best individual among the females. This represents the fitness value of the i-th individual. This represents the fitness value of the best individual among the males.

[0252] Step 10: The location update process during the mating and spawning phase includes the following steps:

[0253] Step 10-1: When the food quantity Q > 0.25, Temp < 0.6, and Rand < 0.6, the individual enters the mating stage. The mating pattern can be described by formulas (15) and (16):

[0254] (15)

[0255] (16)

[0256] Step 10-2: Find the least fit individual among the males and females.

[0257] Step 10-3: After mating, the population enters the spawning stage, during which the least fit male and female individuals are replaced. The update formulas are shown in (17) and (18).

[0258] (17)

[0259] (18)

[0260] Step 11: Perform a quantum computing-based T-distribution mutation process

[0261] Step 11-1: Construct the mutation factor by establishing a quantum space consisting of multiple qubits on the same dimension dim as the problem, as shown in formula (19).

[0262] = (19)

[0263] in the formula dim represents the dimension of the search space. Describes a two-dimensional column vector of units, where .

[0264] Step 11-2: Using formulas (20) and (21), the current optimal solution is obtained. Mapped into quantum space

[0265] (20)

[0266] (twenty one)

[0267] In formulas (20) and (21) Represents the current optimal position. dimension, This represents the magnitude of the variable, which is used to ensure diversity after mutations. The probability is usually set to 1 or -1. express In the A two-dimensional column vector composed of components in the dimension.

[0268] Step 11-3: Transform the quantum space into the corresponding solution space.

[0269] = * (twenty two)

[0270] = * (twenty three)

[0271] and These represent the 1*dim dimensional row vectors after decomposition.

[0272] Step 11-4: Calculate the quantum rotation factor

[0273] (twenty four)

[0274] In formula (24) .

[0275] Step 11-5: Add the quantum rotation factor to the corresponding solution space.

[0276] = (25)

[0277] This represents the solution space after quantum rotation.

[0278] Step 11-6: Map the rotated quantum space to the corresponding solution space to generate... and .

[0279] The final update of the global optimum position is shown in formula (26):

[0280] (26)

[0281] In formula (26) and This represents a point mapped to the solution space after quantum rotation. To show obedience - The variance factor of the distribution, with degrees of freedom of (Current iteration number), individual position is updated through elite selection.

[0282] Step 11-7: The latest generated in Step 11-6 Replace the previous one .

[0283] Step 12: Determine if the iteration number t is greater than T. If not, set t = t + 1 and return to step 5; if so, proceed to step 13.

[0284] like Figure 4 Step 13: Obtain the final prediction model, train it, and make predictions.

[0285] Step 13-1: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The values ​​in each dimension are converted into the corresponding hyperparameters in the ST-Transformer model.

[0286] Step 13-2: Train the model using the training samples obtained in Step 2;

[0287] Step 13-3: Substitute the test samples obtained in Step 2 into the trained ST-Transformer model, compare the output results with the true values, and perform experimental results. Figure 7 .

[0288] The ST-Transformer method was simulated on the same experimental platform, with a predefined model such as... Figure 5 The corresponding hyperparameters are [2, 1, 2, 64, 0.0001], representing Patch size = 2, encoder layers = 1, attention heads = 2, Feed-Forward dimension = 64, and learning rate = 0.0001. The data preprocessing steps are as described in steps 1 and 2 of this method. The obtained training samples are fed into the ST-Transformer model for training, and the test samples are fed into the trained test model. The output results are compared with the true values ​​to obtain experimental results. Figure 6 .

[0289] Experimental results of the ST-Transformer method are as follows: Figure 6 The experimental results of this method are as follows: Figure 7 The dashed line represents the predicted result, and the solid line represents the actual result. The closer the two curves are, the better the result.

[0290] By comparison Figure 6 and Figure 7 It can be seen that, Figure 7 The predicted results (dashed line) are closer to the actual values ​​(solid line), while Figure 6 The predicted results differ significantly from the actual data. This demonstrates that our method can better adapt to the various characteristics of landslide data and is superior to the existing ST-Transformer prediction model.

[0291] Example 1: Implementation of a Global Method for Landslide Displacement Prediction Based on MSSO-ST-Transformer

[0292] In an automated landslide monitoring project in a loess hilly area, the cumulative displacement sequence of monitoring points and rainfall-related monitoring data were selected as raw data. These data were standardized using a preprocessing module, and a training sample set was constructed using a sliding window approach. A spatiotemporal network structure was constructed according to the method in claim 1. The Patch embedding layer segments and encodes the spatial structure of the monitoring point distribution, the location encoding layer provides a temporal reference for the sequence, the Transformer encoder layer completes multi-feature interaction, and the temporal aggregation layer extracts high-dimensional dynamic features. Subsequently, the objective function drives the MSSO optimization framework to calculate temperature and food quantity. A uniform initial population is obtained through initialization with a set of optimal points, providing the model with a more sufficient global search foundation before training.

[0293] During model operation, a foraging strategy is implemented when food availability is low, guiding the population to conduct large-scale exploration through a spiral operator. When food availability increases and temperature rises, an ingestion strategy is implemented, utilizing the Levy step size to enhance leapfrog optimization capabilities. As temperature decreases, a random factor automatically distinguishes between combat and mating / spawning phases, thereby enhancing search accuracy or maintaining population diversity in local areas. At the end of the iteration, a mutation factor constructed in quantum space perturbs the current optimal position, causing the solution to escape local optima within a narrow convergence region. This results in an optimized hyperparameter configuration used to train the ST-Transformer model, achieving high-precision displacement trend prediction in actual applications.

[0294] Example 2: Explanation of the Realizability of the ST-Transformer Model Structure

[0295] According to the structure of claim 2, the ST-Transformer model receives the deformation sequence matrix through the input layer in engineering implementation. The Patch embedding layer spatially slices the distribution of monitoring points at each time step, making the spatial neighborhood features explicit. Subsequently, the position encoding layer introduces a temporal index to ensure the model's sensitivity to changes in sequence order. The main body of the model consists of a multi-layer Transformer encoder, each layer containing a multi-head attention module, which can pay attention to the displacement coupling relationship between different monitoring points in parallel. Layer normalization is used to stabilize gradient propagation, and a feedforward network is used for nonlinear feature extraction. The temporal aggregation layer compresses the multi-time step output into a single vector, which is then used by the output layer to predict subsequent displacements.

[0296] This structure can be implemented using mainstream deep learning frameworks and can be trained end-to-end using engineering monitoring data. Landslide monitoring data exhibits strong spatial heterogeneity and significant temporal fluctuations. The multi-head attention mechanism can capture the differences between local accelerated deformation and long-term stable deformation, while the feedforward network further enhances feature representation capabilities. Experimental results show that, while maintaining controllable parameter scale, this structure can simultaneously handle spatial structural constraints and temporal evolution patterns, demonstrating good engineering adaptability and providing a stable and trainable foundation network for subsequent hyperparameter optimization.

[0297] Example 3: Implementation of Initializing the Population with the Best Point Set

[0298] According to claim 4, a set of optimal points is generated in the dim-dimensional space to form the initial population. The fractional part of each candidate point in each dimension is taken to generate the base positions, which are then mapped to the hyperparameter search space through a linear scaling operation. In engineering applications, dim depends on the number of hyperparameters of the ST-Transformer model, such as the number of attention heads, encoder layers, and hidden dimensions. This method ensures high uniformity of the initial population across the entire search space, avoiding the problem of concentrated distribution in a local area in traditional random initialization, thereby improving global search capability.

[0299] The initialization process further divides the resulting population into male and female subpopulations, each executing different update equations in subsequent behavioral strategies, resulting in higher diversity in the early stages of the algorithm. In engineering verification, compared with random initialization, optimal point set initialization significantly accelerates the early convergence of the MSSO algorithm and reduces the probability of getting trapped in local optima in high-dimensional space, thus enabling the model to ultimately obtain a more stable hyperparameter combination.

[0300] Example 4: Engineering Implementation of Foraging, Feeding, Combat, and Mating / Laying Strategies

[0301] According to claims 5 and 6, when the temperature is low and food is scarce, a foraging strategy is implemented, guiding individuals to conduct a large-scale search in areas far from the current convergence center through a spiral search formula, effectively improving early global exploration capabilities. When food availability increases and the temperature is high, the feeding strategy adopts the Levy step size update method, exploring unvisited areas with large jumps, improving the algorithm's adaptability in irregular search spaces.

[0302] Once the temperature drops to the intermediate range, the system automatically switches to either a combat strategy or a mating-spawning strategy based on random factors. The combat strategy enhances aggregation towards the optimal individual through attraction and repulsion terms, thereby improving convergence. The mating-spawning strategy utilizes the interaction between males and females to generate new individuals and replace the two least fit members, maintaining a certain level of diversity in the population during the mid-to-late stages. Engineering experiments show that the combined use of the four-stage strategy effectively avoids premature convergence and keeps the hyperparameter search results stable in repeated experiments.

[0303] Example 5: Implementation of the quantum T distribution mutation mechanism

[0304] According to claim 6, a quantum space composed of multiple qubits is constructed as a mutation factor, where each qubit consists of a two-dimensional vector and satisfies the modulus constraint. By mapping the current optimal solution to the quantum space and combining it with a t-distribution perturbation with heavy-tailed distribution characteristics, jumps with different amplitudes but statistical regularities can be generated, enabling the solution to migrate from a local region to a better region in a high-dimensional space.

[0305] Quantum rotation operations are used to enhance the randomness after mutation, allowing points mapped to the solution space to cover a wider region. In engineering applications, this mechanism effectively improves the problem of premature convergence in later iterations of the MSSO algorithm, giving deeper regions in the search space a chance to be visited. Comparative experiments show that after adding the quantum mutation mechanism, the optimal hyperparameter combination remains consistent under multiple random seeds, improving the accuracy of the landslide displacement prediction model obtained in the final training.

[0306] Example 6: Deployment of a landslide prediction system

[0307] According to claims 7 to 10, each step is assigned as an independent hardware or software module to constitute the prediction system. The preprocessing module is responsible for reading and normalizing the displacement data uploaded by the automated monitoring equipment; the calculation module calculates the food quantity and temperature based on the current model state; the initialization module generates an initial population based on the optimal point set; the update module updates the population position according to foraging, feeding, fighting, or mating and spawning strategies; the judgment module determines whether the convergence condition is met based on the error curve; and the training module finally trains the spatiotemporal network using the optimal hyperparameters.

[0308] The system can be deployed on the monitoring station server or the data processing platform of the geological disaster early warning center. Modules are coupled through a data bus, and all processes can be executed automatically without manual intervention. At the deployment site, the system works in conjunction with online monitoring equipment, automatically completing the prediction process after data upload and outputting displacement trends for multiple future time steps, providing real-time auxiliary decision-making support for landslide disaster early warning.

[0309] This invention, through the fusion design of the "MSSO optimization algorithm + ST-Transformer spatiotemporal model," achieves significantly better technical results than traditional models in the scenario of loess landslide displacement prediction. The following, combining experimental platform, data design, and simulation results comparison, systematically illustrates the core evidence for the technical effectiveness:

[0310] I. Basic Experimental Conditions

[0311] 1. Experimental Platform and Tools

[0312] Hardware environment: Intel Core i7-6700 processor, 8GB memory, 64-bit Windows 10 system, to eliminate the interference of hardware performance on model training and prediction;

[0313] Software tools: MATLAB 2022a (a mainstream numerical computation and deep learning tool that ensures the universality and reproducibility of model implementation);

[0314] Core dependencies: The deep learning framework is based on MATLAB's built-in Neural Network Toolbox, and the Hyperparameter Optimization Module (MSSO) is a custom-programmed implementation that interacts with the ST-Transformer model through an interface.

[0315] 2. Experimental Data Design

[0316] Data source: Over 800 days of continuous monitoring data from a landslide monitoring point in the Loess Hilly Area, covering 5 core characteristics (landslide displacement distance, rainfall, temperature, humidity, and soil moisture), with data volume and dimensions consistent with actual engineering scenarios;

[0317] Data preprocessing: Standardize units (displacement: mm, rainfall: mm / day, temperature: ℃), remove outlier noise points, and impute missing values ​​to avoid data quality affecting model performance;

[0318] Sample partitioning: The training set (480 days) and the test set (320 days) are divided in a 6:4 ratio, and the prediction step size is set to 8 days (that is, the displacement at time T+8 is predicted based on the data at time T, which meets the actual requirement of "early warning more than 1 week in advance" in engineering).

[0319] Comparison Model: The "MSSO-ST-Transformer model" of this invention is compared with the "unoptimized ST-Transformer model" as the benchmark (with the hyperparameters fixed as: Patch size = 2, number of encoder layers = 1, number of attention heads = 2, and Feed-Forward dimension = 64) under the same conditions.

[0320] II. Evidence of the effectiveness of core technologies

[0321] 1. Improved Prediction Accuracy: Quantitative Comparison of Error Indicators

[0322] In the experiment, mean squared error (MSE) and mean absolute error (MAE) were used as the core evaluation indicators (the smaller the error, the higher the prediction accuracy). The performance differences between the two types of models were calculated using the test set data, as shown in the table below:

[0323] Model Test set MSE (mm²) Test set MAE (mm) Accuracy improvement rate (MSE comparison) Traditional ST-Transformer 128.6 9.2 - MSSO-ST-Transformer 45.3 5.1 65.0%

[0324] After MSSO optimization, the model's MSE decreased from 128.6 to 45.3, improving accuracy by 65%; MAE decreased from 9.2mm to 5.1mm, meaning that the average deviation between the predicted value and the actual displacement was reduced by 4.1mm. This improvement is crucial in landslide early warning, as it can avoid false alarms (such as misjudging small displacements as risk) or missed alarms (such as misjudging large displacements as safe) caused by "excessive prediction deviation".

[0325] Reasons for the reduction in error: MSSO finds the optimal combination of hyperparameters for ST-Transformer through global optimization (the final optimized hyperparameters in the experiment are: Patch size = 3, number of encoder layers = 2, number of attention heads = 4, and Feed-Forward dimension = 128), which solves the defects of traditional models such as "high sensitivity to hyperparameters and easy overfitting", and enables the model to maintain high generalization ability on unseen test data.

[0326] Figure 8 shows the convergence curves of the proposed MSSO (Multi-Strategy Sparrow Optimization) and traditional SO (Sparrow Optimization) in the hyperparameter search process. It is evident from the figure that the loss of traditional SO decreases more slowly, exhibiting a slow and monotonous convergence pattern. In contrast, MSSO rapidly approaches the optimal region in the early stages of iteration, and in the later stages, it achieves continuous deep searching through the synergistic effect of multiple strategies. This allows the objective function to decrease to a level far lower than that of SO after approximately 200 iterations, ultimately improving the convergence accuracy by about 2–3 orders of magnitude. This demonstrates that MSSO effectively overcomes the shortcomings of traditional SO, such as its tendency to get trapped in local optima and insufficient search efficiency, exhibiting significant global search capabilities and stability advantages in the hyperparameter optimization process.

[0327] Figure 9 shows a bar chart comparing the performance of different models in the target prediction task, where dark gray represents MSE (mm²) and light gray represents MAE (mm). The results show that the errors of LSTM and CNN are both at a high level; although the traditional ST-Transformer improves feature coupling ability, its MSE and MAE are still relatively large. The MSSO-optimized ST-Transformer of this invention achieves significant performance improvement, with its MSE and MAE being significantly lower than traditional methods. According to experimental statistics, the performance improvement of the model of this invention compared to the traditional ST-Transformer reaches: MSE reduced by 65.0% and MAE reduced by 44.6%** (marked above the bar chart), fully verifying the effectiveness and feasibility of MSSO in improving parameter search quality and reducing model error.

[0328] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0329] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting loess landslide displacement based on the MSSO-ST-Transformer model, characterized in that, Includes the following necessary steps: Landslide monitoring data were preprocessed and used to create training samples; A spatiotemporal feature extraction network is constructed based on the training samples; Calculate the food quantity Q and temperature Temp based on the objective function; The initial search population is formed using a prime point set strategy; Based on food availability and temperature indicators, the population employs phased search strategies involving foraging, feeding, fighting, or mating and spawning to update its location. During the iteration process, a mutation factor constructed in quantum space is used to perturb the current optimal position and obtain a new optimal solution; When the convergence condition is met, the final optimal solution is used as the hyperparameter configuration of the spatiotemporal network to complete the model training and obtain the displacement prediction model.

2. The method as described in claim 1, characterized in that, The spatiotemporal feature extraction network includes an input layer, a patch embedding layer, a position encoding layer, a Transformer encoder layer, a temporal aggregation layer, and an output layer connected in sequence. Each Transformer encoder layer includes a multi-head attention module, a layer normalization module, and a feedforward network module.

3. The method as described in claim 1, characterized in that, The amount of food Q and the temperature Temp are calculated as follows: The temperature Temp is determined by an exponential decay formula, and its value gradually decreases as the number of iterations increases; The amount of food Q is determined by the exponential growth formula, and its value gradually increases as the number of iterations approaches the maximum number of iterations. In the calculation process, 50% of the constant z1 is used as the initial factor.

4. The method as described in claim 1, characterized in that, The initial search population is generated from a set of superior points, which constructs multiple superior points in a dim-dimensional space. The initial position of an individual is calculated based on the fractional part of each superior point, and the population is divided into male subpopulations and female subpopulations, so that the two subpopulations execute different behavioral update rules in subsequent update phases.

5. The method as described in claim 1, characterized in that, When the food quantity is less than 0.25, the foraging phase begins, and the position of random individuals is updated using an adaptive spiral search operator. When the food quantity is greater than 0.25 and the temperature is higher than 0.6, the feeding phase begins, and the feeding position is updated using an update formula consisting of the product of the Levy distribution step size generated by the random step size and the difference between the optimal individual position.

6. The method as described in claim 1, characterized in that, When the food quantity is greater than 0.25 and the temperature is lower than 0.6, the battle phase begins when the random factor is greater than 0.

6. The battle update determines the offset factor based on the ratio of the individual's fitness to that of the best individual. When the random factor is less than 0.6, the mating and egg-laying phase begins. The position is updated using the mutual attraction term between males and females, and the two individuals with the worst fitness are replaced by randomly generated new individuals. In each iteration cycle, a two-dimensional vector group constructed in quantum space is used as a mutation factor. It is mapped to the solution space through quantum rotation operations, and the perturbation term generated by the t-distribution is combined to perform mutation update on the optimal position.

7. A loess landslide displacement prediction system based on the MSSO-ST-Transformer model, characterized in that, It includes a preprocessing module, a calculation module, an initialization module, an update module, a judgment module, and a training module. The preprocessing module is used to obtain training samples and construct a spatiotemporal network. The calculation module is used to calculate the amount of food Q and the temperature Temp; the initialization module is used to form a search population based on the set of best points. The update module is used to perform position updates at different search stages; the judgment module is used to determine whether the iteration termination condition is met. The training module is used to train the spatiotemporal network and generate a prediction model by using the optimal solution as hyperparameter configuration.

8. The system as described in claim 7, characterized in that, The initialization module is configured to generate a set of optimal points in a dim-dimensional space and divide the population into a male subset and a female subset.

9. The system as described in claim 7, characterized in that, The update module is configured to perform position updates for the foraging, feeding, combat, and mating / spawning stages, and to perturb the current optimal solution using a quantum mutation factor in each iteration.

10. The system as described in claim 7, characterized in that, The training module trains the spatiotemporal network based on the optimal hyperparameters obtained from the final search and outputs model results for predicting loess landslide displacement.

Images