The knowledge system covered under the heading “minimum two-fold support to vector deposition prediction based on the improvement of the disassembly particle population” is a front-line multidisciplinary research direction that integrates machine learning, intelligent optimization algorithms, numerical calculations and civil engineering safety monitoring. At the core of the solution to the problem of the sinking of structures such as the physical engineering base, slope, bridge towers, and upper building infrastructure during construction and operating periods — which are characterized by non-linear, small samples, strong noise, dynamic evolution and mechanical ambiguity — traditional empirical formulas (e. G., the sum of layers) or physical models (e. G., the biot firmure theory) tend to make it difficult to balance precision and generalization, while data-driven smart modelling approaches show significant advantages. The key improvement of the minimum 2x2 support vector (least squares support vehicle machine, lsvm) as an important variant of the standard support vector (svm) is the easing of the variable restraints in the svm original optimisation issue into equations and the conversion of the target function from a secondary plan (qp) to a linear equation group to solve: i. E., replacing the ε-insensitive loss of vapnik by introducing the square error loss function, degrading the problem to a linear system (kα = y) to solve the lagrangi multiplier. This transformation has significantly reduced the complexity of calculations and increased the speed of training, especially for small and medium-sized regression missions; however, the performance of lssvm is highly dependent on three key super-parameters: regularization of parameters gamma (the trade-off between control of empirical and structural risks), consideration of nuclear function parameters (e. G. The bandwidth of the rbf nuclear, the determination of the local assembly capability of the model) and possible deviation b. If the parameters are not properly set, it is highly likely to lead to a combination (over-gaming, over-consideration) or under-synthesis (over-gaming, over-consideration), thus seriously reducing its robustness and extrapolation in the prediction of the sedimentary time series. The chaotic particle swarm optimization, cpso is precisely the enhanced smart algorithm that has been developed to overcome the weaknesses of traditional psos, which are prone to local excellence, low levels of condensation and insufficient diversity of populations. It embeds a mixed map (e. G. Logistic mapx{n+=mix n (1-x n), μ=4 in total disarray) into the pso framework: on the one hand, using a mixed sequence at the initial stage to generate an evenly distributed and dynamic initial particle position to enhance the potential for global exploration; on the other hand, mixing disturbances in the speed or location of selected particles during the iterative process to effectively break the stagnation and increase the ability to jump out of local extreme values; and, on the other hand, further balancing exploration with exploration and development (exploitation) capabilities in the context of inertial weights, compression factors or elite retention strategies. In an optimised scenario of lssvm parameters, the cpso is an adaptive function using indicators such as predicted average error (mse), average absolute percentage error (mape) or decision factor r2, which will form (gamma, think) a two-dimensional (or multi-dimensional) search space that will automatically find the best combination of broad-based parameters through a million-scale adaptive assessment and particle evolution, much more effective than traditional media such as grid search, random search or cross-checking. Matlab, as the mainstream platform for the development of a prototype of engineering computing and algorithms, has assumed the role of a full process realization carrier in this project: pre-processing (missing value plug-in, removal of anomalies, integration of slide windows), feature engineering (formation lag steps, time-series features such as cumulative rainfall, temperature gradients, load increments, etc.), lssvm modelling (relaying ls-svmlab toolbox or autonomous preparation of nuclear matrix construction and linear solvency module), to the cpso algorithm code (combination initialization, adaptive assessment, particle updating, boundary processing), multi-wheel cross-certification design, to single output (single point future t-fall) and multiple output (synthetic prediction of sedimentation at multiple spatial sites) lssvm, all of which are based on tot-matlab. Of particular note is the fact that multi-output extensions do not simply run multiple single-out models in parallel, but rather significantly increase the synergetic prediction accuracy of space-related deposition models by sharing covert features, introducing export-relevance restraints (e. G., coordinated differential correction) or using quantitative nuclear techniques. At the application level of civil engineering monitoring, the methodology has been successfully applied to typical scenarios such as the deposition warning of surrounding buildings caused by the excavation of soft land-based deep-pits, the assessment of long-term post-deposition work on high-filled road embankments, and the evolution of water reservoir-induced slope migration trends. Its engineering value is reflected not only in a projected error of 30 per cent - 50 per cent lower than that of the bp neural network, the arima model, but also in its interpretability (contribution to the inversion of the influence factor through the nuclear function), its ability to update online (support incremental learning to adapt to new monitoring data) and its quantitative potential for uncertainty (in combination with the projected confidence interval of bootstream or bates lssvm). In summary, the technology system represents an important paradigm shift in the health diagnosis of smart algorithms and is one of the key enabling technologies to drive civil engineering from “experience-driven” to “data+ mechanisms-driven”。
