A Method for Assessing the Synergistic Restoration Effects of Campus Building and Landscape Transition

Abstract

This invention discloses a method for assessing the synergistic restorative effects of campus building and landscape transitions. The method acquires spatiotemporally aligned campus environmental data, human physiological response data, and psychological perception data through gradient path design and synchronous multimodal data acquisition. Deep analysis of the environmental data extracts visual and auditory features, constructing a standardized environmental feature tensor. Based on the environmental feature tensor and the physiological response data, dynamic restorative effect modeling is performed, quantifying the nonlinear mapping relationship between environmental factors and restorative effects, as well as cross-sensory synergistic effects. According to the output of the dynamic model, multi-objective optimization is used to generate environmental transformation schemes. Generative artificial intelligence is used to achieve parameterized scene reconstruction and effect verification, and a dynamic restorative assessment map of campus space is constructed. This invention achieves scientific quantification and visualization assessment of the restorative potential of campus transition environments, providing decision support for refined environmental design.

Description

Technical Field

[0001] This invention relates to the intersection of environmental assessment and artificial intelligence, specifically to a method and system for assessing the synergistic restorative effects of the transition between campus buildings and landscapes (hereinafter referred to as "building-landscape"). Background Technology

[0002] Currently, the assessment of the restorative effects of the campus environment mainly relies on subjective questionnaires or isolated static analyses of single environmental parameters such as green view rate and noise level. These methods fail to reveal the real-time synergistic mechanisms of multi-sensory information such as vision and hearing, and also neglect the dynamic cumulative characteristics of physiological responses.

[0003] Existing research often neglects the nonlinear relationship between environmental factors and human autonomic nervous activity. In particular, when transitioning between different types of environments (such as buildings and landscapes), there is a lack of in-depth analysis of key issues such as the interaction between factors and the temporal patterns of environmental exposure. A set of assessment methods that can be applied in engineering has not yet been formed.

[0004] In summary, existing evaluation methods have shortcomings in terms of data objectivity, multi-sensory collaboration, dynamic nonlinear modeling, and engineering application transformation. Therefore, there is an urgent need for a systematic evaluation method and system that can integrate multimodal objective data, analyze multi-factor collaborative mechanisms, characterize dynamic response processes, and support design decisions and scheme optimization. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned shortcomings of existing technologies and provide a method for assessing the synergistic restorative effects of campus building and landscape transitions. This method first acquires spatiotemporally aligned building-landscape environment, physiological responses, and psychological perception data through gradient path design and synchronous multimodal data acquisition. Then, it analyzes the semantic features of building and landscape elements and integrates sound elements to construct an environmental feature tensor. Based on this, it innovatively establishes a dynamic mapping architecture for physiological delay correction and multimodal synergistic effects, quantifying the complex nonlinear relationship between environmental stimuli and physiological recovery. Subsequently, it generates optimal solutions, utilizes generative artificial intelligence to achieve parametric reconstruction and effect verification, and constructs a dynamic evaluation map to intuitively present the distribution patterns of the restorative effects of the campus environment, forming a closed-loop technical framework of "data acquisition - modeling and analysis - effect assessment - scene optimization".

[0006] To address at least one of the problems existing in the prior art, this invention provides a method for assessing the synergistic restorative effect of the transition between campus buildings and landscape, comprising the following steps:

[0007] S1: By designing gradient paths and collecting multimodal data synchronously, acquire spatiotemporally aligned campus building-landscape environment data, human physiological response data, psychological perception data, and spatiotemporal trajectory data;

[0008] S2: Based on the environmental data obtained in step S1, extract visual and auditory features, and construct a standardized environmental feature tensor;

[0009] S3: Based on the environmental feature tensor obtained in step S2 and the physiological response data obtained in step S1, dynamic modeling of the recovery effect is performed to obtain a dynamic model; the dynamic model is validated and adaptively fed back using psychological perception data.

[0010] S4: Based on the output of the dynamic model described in step S3, perform multi-objective optimization to generate an environmental transformation plan, utilize generative artificial intelligence to achieve parameterized scene reconstruction and effect verification, and construct a dynamic restorative assessment map of the campus space.

[0011] Furthermore, step S1 includes:

[0012] S11: Based on the campus spatial topology map, plan multiple gradient transition pedestrian paths from buildings to landscapes;

[0013] S12: Simultaneously collect environmental data, physiological response data, psychological perception data, and spatiotemporal trajectory data along the walking path;

[0014] S13: Using a time synchronization protocol and a spatial interpolation matching algorithm, the environmental data, physiological response data, psychological perception data, and spatiotemporal trajectory data collected in step S12 are processed to achieve alignment of multimodal data in a unified spatiotemporal coordinate system.

[0015] Furthermore, step S2 includes:

[0016] S21: Perform image semantic segmentation on the visual scene in the environmental data to obtain building area mask and landscape area mask, and calculate visual features based on the mask. The visual features include at least building index, landscape index and spatial openness.

[0017] S22: Perform frequency band analysis and loudness calculation on the sound pressure level data in the environmental data to obtain auditory characteristics, wherein the auditory characteristics include at least the sound pressure level and instantaneous loudness of multiple frequency bands;

[0018] S23: Combine the visual features obtained in step S21 with the auditory features obtained in step S22 to construct an environmental feature tensor, and standardize the environmental feature tensor to form the standardized environmental feature tensor.

[0019] Furthermore, the dynamic modeling of the restorative effect in step S3 includes:

[0020] S31: Preprocess the physiological response data obtained in step S1 to obtain corrected physiological data suitable for modeling;

[0021] S32: Obtain individual baseline characteristics;

[0022] S33: Using the environmental feature tensor obtained in step S2 as the environmental input, the corrected physiological data obtained in step S31 as the physiological target, and fusing the individual baseline features, a dynamic model for predicting restorative physiological indicators is constructed; the dynamic model is realized by fusing and modeling multimodal environmental features and physiological data, and can also be called a multimodal fusion model.

[0023] S34: Use the psychological perception data obtained in step S1 to verify and adaptively feedback the dynamic model;

[0024] S35: Based on the dynamic model, quantify the nonlinear mapping relationship between environmental factors and restorative effects, and calculate the cross-sensory synergistic gain index;

[0025] S36: Based on the aforementioned dynamic model, a sequence modeling method is used to analyze the cumulative dynamic characteristics of restorative effects during the transition from architecture to landscape, in order to analyze the synergistic effects of cross-sensory environmental factors.

[0026] Furthermore, the preprocessing in step S31 includes: applying a Gaussian weighted integral model to perform physiological delay correction on the heart rate variability index, and applying a moving average filter to smooth and denoise the skin conductance activity signal.

[0027] Furthermore, the dynamic model constructed in step S33 is an improved Kolmogorov-Arnold network model (referred to as the improved KAN model), whose input is formed by fusing the environmental feature tensor, the individual baseline features, and the cross features of the two.

[0028] Furthermore, in step S34, the psychological perception data is used as a supervision signal to verify the prediction results of the dynamic model; when the correlation between the prediction results and the psychological perception data is lower than a preset threshold, the dynamic model is retrained.

[0029] Furthermore, the sequence modeling method in step S36 employs a bidirectional spatiotemporal long short-term memory network.

[0030] Furthermore, step S4 includes:

[0031] S41: Based on the cross-sensory synergistic gain index, determine the key environmental factor combination for synergistic effect, with maximizing the predicted output value of the dynamic model in step S3 as the first optimization objective and maximizing the spatial openness obtained from step S2 as the second optimization objective, and construct a multi-objective optimization function; under the conditions of satisfying the landscape index constraint, auditory feature constraint and transformation feasibility constraint, use a multi-objective optimization algorithm to generate a Pareto optimal environmental parameter solution set; wherein, the landscape index constraint and the auditory feature constraint are set based on the environmental features obtained from step S2.

[0032] S42: Input each set of parameters in the Pareto optimal environment parameter solution set generated in step S41 into the conditional generative artificial intelligence model to generate the corresponding three-dimensional visualization scene.

[0033] S43: For each of the three-dimensional visualization scenes generated in step S42, extract its environmental features and input them into the dynamic model constructed in step S3 to predict its restorative effect value, and compare and verify it with the predicted value of the benchmark scene.

[0034] S44: Based on the spatiotemporal trajectory data obtained in step S1, the campus space is discretized into a grid system; for each grid, based on its environmental characteristics, the dynamic model constructed in step S3 is used to predict the restorative effect, and the spatiotemporal cumulative effect is calculated to generate the dynamic restorative assessment map in the form of a dynamic spatiotemporal heat map.

[0035] This invention also provides a system for assessing the synergistic restorative effects of campus building and landscape transitions, used to implement the aforementioned method, comprising:

[0036] The data acquisition and spatiotemporal alignment module is used to perform step S1 to acquire a spatiotemporally aligned multimodal dataset;

[0037] The environmental feature deep analysis module is used to execute step S2, which parses the environmental data in the multimodal dataset and generates a standardized environmental feature tensor.

[0038] The dynamic modeling module for recovery effects is used to perform step S3, which involves dynamic modeling based on the environmental feature tensor and physiological response data to quantify the relationship between environmental factors and recovery effects.

[0039] The scene optimization and evaluation map construction module is used to perform step S4, optimize and verify the scene based on the output of the dynamic modeling, and construct a dynamic restorative evaluation map.

[0040] Compared with the prior art, the present invention can achieve at least the following beneficial effects:

[0041] This invention breaks through the limitations of traditional design relying on static indicators by quantitatively analyzing the multi-sensory collaborative mechanism of the transition between the built environment and the landscape environment; based on the environmental parameter-scene mapping capability of generative artificial intelligence, it improves the accuracy of the visualization transformation and effect verification of the scheme; the constructed dynamic evaluation map transforms the abstract feature response into an intuitive heat map expression, enabling architects to accurately identify areas with high recovery potential and optimize the configuration of elements, thereby systematically improving the recovery effect of the campus environment and the quality of the living environment. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the overall process of evaluating the synergistic restorative effect of the transition between campus buildings and landscape, as described in this embodiment.

[0043] Figure 2 This is a flowchart of the data acquisition and spatiotemporal alignment process in the embodiment.

[0044] Figure 3 This is a flowchart illustrating the dynamic modeling of the restorative effect in the embodiments.

[0045] Figure 4 This is a block diagram of the collaborative recovery effect assessment system in the embodiment. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] This embodiment proposes a collaborative restorative effect assessment method for the transition between campus buildings and landscapes. Through simultaneous multimodal data acquisition, environmental modeling and analysis, restorative effect assessment, and scene optimization, it achieves quantitative analysis and optimal design of environmental restorative effects. First, based on gradient path design, it integrates building-landscape environmental and physiological-psychological data to construct a spatiotemporally aligned multidimensional dataset. Second, through semantic decoupling of building and landscape and multisensory feature fusion, it generates a standardized environmental feature tensor. Then, it constructs a physiological delay correction model and a cross-modal dynamic mapping architecture to analyze the nonlinear correlation between environmental factors and restorative effects. Finally, it employs multi-objective optimization to generate Pareto solutions, combined with a conditional generation model to achieve 3D scene reconstruction under physical constraints, ultimately constructing a dynamic assessment map.

[0048] Specifically, such as Figure 1 As one example, a method for assessing the synergistic restorative effects of campus building and landscape transition includes the following steps:

[0049] Step S1: Data Acquisition and Spatiotemporal Alignment. Through gradient path design and synchronous multimodal data acquisition, spatiotemporally aligned campus building-landscape environment data, human physiological response data, psychological perception data, and spatiotemporal trajectory data are obtained; such as... Figure 2 The data acquisition and spatiotemporal alignment process is as follows.

[0050] To achieve a multi-dimensional analysis of environmental restorative effects, this step involved an experiment that integrated data collection from the “building-landscape” environment, physiological responses, and psychological perceptions. The distance confounding effect was eliminated by using a gradient experimental path, and high-precision synchronization technology was employed to acquire spatiotemporally matched multimodal data, providing a complete data foundation for subsequent modeling.

[0051] First, experimental path design was conducted. Based on the campus spatial topology map, multiple gradient transition pedestrian paths from buildings to landscapes were planned, and the set of path lengths was defined as follows: ,in For path numbering, For path The length of the walking distance is used to eliminate the confounding effects of the restorative effect through length gradient control.

[0052] Secondly, multimodal data was acquired synchronously. For environmental data (including visual image data and auditory sound pressure level data), a panoramic camera was used with a sampling rate of [missing information]. Acquire visual scene and simultaneously bind sound pressure level sensor (sampling rate) Record sound pressure level For physiological data acquisition, a smart bracelet integrating PPG and EDA sensors is used. Sampling rate continuously records heart rate variability (HRV) indicators , and skin electric ,in Quantified as conductivity value For psychological data collection, data is collected at fixed time intervals during the path walking process. Triggering scale completion, using the PRS scale to assess restorative perception, generating discrete time series. ,in ( ), The path start time. and These represent the serial number of the number of entries on the psychological scale and the total number of entries, respectively. For trajectory data collection, this is determined by the positioning error. High-precision differential GPS, with Recording spacetime coordinates .

[0053] All data streams are spatiotemporally aligned using the Precision Time Protocol (PTP). First, the timestamp alignment error is defined. ,in For each sensor's timestamp, This refers to the timestamp when the GPS device records data points. In the spatial interpolation matching algorithm, spatial matching is achieved through a cubic spline interpolation function. Fitted GPS trajectory points, where For the order index of the polynomial terms, These are the coefficients of the cubic spline function. For arc length parameters, Arc length parameter of This operation ensures that the data points of any sensor in the spatiotemporal coordinate system are at the power of [the power]. Spatial error with interpolation trajectory ,in This indicates the timestamp of the sensor data points. Arc length parameters mapped onto the fitted GPS trajectory Finally, a multimodal dataset with a unified spatiotemporal index was constructed. ,in For environmental feature vectors, For physiological feature vectors, This represents the spatial matching error.

[0054] Step S2: In-depth analysis of environmental features. Based on the environmental data obtained in Step S1, visual and auditory features are extracted, and a standardized environmental feature tensor is constructed. A specific example of the process is as follows.

[0055] Obtaining a spatiotemporally aligned multimodal dataset Next, environmental feature analysis first involves building-landscape semantic separation. For visual feature processing, images captured by a panoramic camera are used. Pixel-level semantic segmentation is performed using the DeepLabv3+ model, and a binary mask for the building region is defined. and landscape mask Based on this, the building index and landscape index are calculated:

[0056] ;

[0057] ;

[0058] in They represent The proportion of architecture and landscape at any given time and These represent the image width and height, respectively, used to normalize the pixel ratio of building / landscape areas. Simultaneously, a depth map is generated using the DepthAnything v2 model. Calculate the openness of the space:

[0059] ;

[0060] in, For pixels The depth value. Visual feature supplementation extracts the average saturation channel value of the HSV color space. ,in, For image At pixel position The saturation value at that location.

[0061] For auditory feature processing, the time-domain signal acquired by the sound pressure level sensor is first processed. Perform 1 / 3 octave band analysis and set the center frequency. There are a total of 8 frequency bands. Calculate the sound pressure level of each frequency band:

[0062] ;

[0063] in, For time window, Center frequency Instantaneous sound pressure level corresponding to the frequency band, The reference sound pressure level is used.

[0064] Furthermore, the instantaneous loudness N(t) is calculated:

[0065] ;

[0066] in, The human hearing threshold curve function, These are the frequency band weighting coefficients. The final environmental feature tensor is constructed as follows:

[0067] ;

[0068] All features are arranged by time step. Integrate into tensors And perform Z-score standardization. .in, These are the mean and standard deviation for each dimension. This is an element-wise division method.

[0069] Step S3: Dynamic modeling of the recovery effect. Based on the environmental feature tensor obtained in step S2 and the physiological response data obtained in step S1, dynamic modeling of the recovery effect is performed to obtain a dynamic model; simultaneously, psychological perception data is used to validate and adaptively feedback the dynamic model; such as... Figure 3 The specific process of dynamic modeling of the restorative effect in this embodiment is as follows.

[0070] This embodiment addresses the complex nonlinear relationship between environmental stimuli and physiological responses by performing multi-dimensional dynamic modeling of restorative effects. The implementation consists of four core steps: physiological data preprocessing, multimodal feature fusion modeling, time-series effect analysis, and key factor analysis.

[0071] (1) Physiological data preprocessing

[0072] A complete dataset based on spatiotemporal alignment First, the raw physiological signals undergo deep correction processing. Because the autonomic nervous system exhibits a delayed response to environmental stimuli, traditional instantaneous correlation analysis can introduce systematic bias. Therefore, this embodiment designs a Gaussian weighted integral correction system:

[0073] ;

[0074] ;

[0075] ;

[0076] in, For correction SDNN physiological data at time points, The original SDNN time-domain signal was acquired. For time variables, The kernel function is Gaussian, and the weights are assigned based on the time difference. Decision. Decision made using the Gaussian kernel function. Precisely quantify the physiological inertia of the HRV index, among which Characterizing the average delay time of the population, Reflects the dispersion of individual differences. The normalization coefficients ensure signal energy conservation. For EDA signals, moving average filtering is used to eliminate motion artifacts.

[0077] ;

[0078] in, After filtering Electrodermal activity signals at any given time for The original electrodermal signal at time t, and the filter window width. The settings are optimized based on the physiological characteristics of skin conductance response. A corrected physiological tensor is then generated. Its temporal resolution is strictly aligned with environmental characteristics.

[0079] (2) Modeling of collaborative influence mechanism

[0080] To achieve accurate mapping between environmental stimuli and physiological recovery effects, this invention constructs a multimodal fusion architecture based on an improved Kolmogorov-Arnold network (KAN) model. This architecture improves upon the traditional KAN model by enhancing cross-modal interaction capabilities through structured input. First, a cross-modal fusion layer is added after the basis function mapping layer to enhance cross-sensory feature interaction capabilities. Second, a dual-gating mechanism of feature selection gates and feature transformation paths is designed to improve the fitting accuracy of high-frequency nonlinear responses. Furthermore, the Kronecker product is introduced to explicitly encode environment-individual interaction effects, addressing the shortcomings of traditional fully connected networks in modeling multimodal collaborative mechanisms. This improved Kolmogorov-Arnold network model, through a multi-level feature interaction mechanism, fully captures the complex nonlinear relationship between visual and auditory factors and human physiological responses in the architectural landscape transition environment.

[0081] In the specific implementation, the standardized environmental feature vector is first... Compared with individual baseline characteristics (A 3D vector composed of age, gender, and BMI) is deeply fused, and an enhanced input is generated through a feature cross-layer: Using Kronecker product operation Explicitly constructing context-individual interaction features effectively encodes key interaction effects, ultimately forming enhanced input. .

[0082] The network core employs learnable basis functions for feature space transformation, and designs eight sets of cubic B-spline basis functions. Each group contains 16 optimizable control points. Mapping the high-dimensional input to a smooth function space:

[0083] ;

[0084] This process adaptively adjusts the shape of the basis functions to accurately fit the local nonlinear relationship between environmental features and physiological responses. A further radial basis function layer is introduced to achieve cross-modal feature fusion.

[0085] ;

[0086] in, The higher-order feature vectors after cross-modal fusion, cluster centers The bandwidth parameter is determined by k-means pre-training. Dynamic optimization during training enables the model to automatically focus on key regions in the feature space. The final output layer employs a dual-path gating mechanism to enhance nonlinear expressive power. First, feature selection gates are generated:

[0087] ;

[0088] Among them, feature selection gate parameters and These represent the weight matrix and bias vector for the dual-gating mechanism. The flow ratio of each feature dimension is determined using the Sigmoid function. Simultaneously, the feature transformation path is generated using the GeLU activation function.

[0089] ;

[0090] Among them, feature transformation path parameters and These are the weight matrix and bias vector of the feature transformation, respectively. After Hadamard product operation, the output is normalized to restore the level. This architecture fits high-frequency nonlinear responses using basis functions and dynamically adjusts the contribution of environmental factors through a gating mechanism, demonstrating significant advantages over traditional fully connected networks in modeling complex environment-physiology mappings. Furthermore, Used as a supervisory signal to verify the recovery level predicted by the KAN model. ,when When there are significant changes, the corresponding psychological perception data PRS score should change synchronously (Pearson correlation coefficient threshold). Otherwise, the model will be retrained.

[0091] (3) Analysis of key influencing factors

[0092] To reveal the restoration mechanism of the transition from architecture to landscape environment, this invention constructs a multi-granularity interpretability analysis framework. At the feature level, the Shapley value is used to decompose the independent contributions of each environmental factor. Conditional expectation is calculated... Quantification of a fixed single factor The offset of the predicted value relative to the global average, where, The improved KAN model predicts the recovery effect. Further normalization yields the relative contribution. This operation overcomes the limitation of traditional feature importance ranking that ignores interaction effects.

[0093] At the interaction level, this invention proposes a cross-sensory synergistic gain index. :

[0094] ;

[0095] in, Environmental factors The Shapley value, i.e. The conditional expectation. For factor pairs. Calculate joint contributions With independent contributions and The relative difference. When The results indicate the existence of a positive synergistic effect. To verify significance, 1000 Monte Carlo permutation tests were performed, and a null distribution was constructed by randomly shuffling the order of factor combinations. Finally, samples meeting the criteria were selected. and Significantly co-located sets .

[0096] At the spatial response level, construct three-dimensional response surfaces for key factors:

[0097] ;

[0098] The domain is calculated through dense sampling. The expected recovery level within the region was determined using the Marching Cubes algorithm to generate isosurfaces, and gradient change areas in two key regions were labeled. ) and high-efficiency collaborative zone ( This study aims to achieve a multi-scale mechanistic analysis of the restorative effects of the campus transition environment. The gradient change region reflects the sensitive transition of the restorative effect caused by small changes in environmental parameters, while the high-efficiency synergy region indicates the optimal parameter range for multi-factor synergistic enhancement.

[0099] (4) Dynamic analysis of sequence effects

[0100] To address the cumulative restorative effects during the transition from architecture to landscape, this invention proposes a sequence dynamics modeling method based on bidirectional spatiotemporal LSTM. The model input is a feature sequence strictly ordered by timestamps. It captures the impact of historical and future environments simultaneously through a bidirectional loop architecture. Forward LSTM unit. Propagation along the forward timeline, memorizing the environmental exposure history from the path's origin to the current location; inverse LSTM unit. Backpropagation encodes expected environmental characteristics for future time periods. The hidden states of both are then fused through a nonlinear fusion layer. The generation of spatiotemporal joint representations effectively addresses the shortcomings of traditional unidirectional models in modeling "expected environmental effects." Among these, The weight matrix of the nonlinear fusion layer. Indicates the forward LSTM unit and inverse LSTM unit The splicing operation.

[0101] Based on this feature representation, the cumulative amount of the restorative effect is quantified using the path integral principle. First, the instantaneous rate of recovery is calculated:

[0102] ;

[0103] in, The gradient of environmental characteristic changes is characterized, and the inner product operation reflects the sensitivity of physiological responses to environmental changes. For at any time The output is the normalized predicted recovery level. Integrating along the time axis yields the theoretical cumulative value. ,in, for The instantaneous recovery rate at any given moment. To correct the memory decay characteristics of the autonomic nervous system, an exponential decay factor is introduced. Forming a modified model:

[0104] ;

[0105] in, The attenuation coefficient accurately reflects the time-varying attenuation pattern of the recovery effect in different individuals; This represents the end time of the current analysis path. This represents the adjusted cumulative recovery effect. This model enables dynamic quantification of the recovery effect in transitional scenarios and can predict the evolution of recovery effects at different walking rhythms.

[0106] This step integrates dynamic latency correction, an improved KAN architecture, bidirectional spatiotemporal LSTM, and interpretable machine learning techniques, through... The model accurately quantifies the physiological delay effect, and the radial basis function enhances the feature representation ability of KAN. To reveal the cumulative dynamics of environmental exposure, The indicators enable a quantitative assessment of cross-sensory synergistic effects.

[0107] Step S4: Scene Optimization and Effect Verification. Based on the output of the dynamic model in Step S3, multi-objective optimization is performed to generate an environmental transformation plan. Generative artificial intelligence is used to realize parameterized scene reconstruction and effect verification, and a dynamic restorative assessment map of the campus space is constructed for scene optimization and effect verification.

[0108] In this embodiment, based on the dynamic modeling of the restorative effect, this step performs scene optimization and effect verification. An environmental modification plan is generated using a multi-objective optimization algorithm, and a conditional generative artificial intelligence model is combined to achieve a visual transformation of parameters into the scene, which is then verified. A specific implementation process is illustrated below.

[0109] (1) Generation of multi-objective optimization strategies

[0110] Contribution of key factors derived from dynamic model analysis and cross-sensory synergistic gain index A multi-objective optimization model with multiple engineering constraints is established. The objective function comprehensively considers the physiological recovery effect and spatial perception quality, maximizing the predicted value of the recovery effect. With environmental openness The weighted sum. The optimization process imposes rigid constraints based on the actual scenario: landscape index. Limiting it to a certain range to ensure ecological rationality; 1kHz sound pressure level Controlled The scope meets auditory comfort requirements; feasibility constraints of the modification. To ensure the feasibility of the plan, This represents the feature vector of the baseline scene.

[0111] The optimization algorithm employs an improved NSGA-III framework, achieving Pareto solution output through a four-stage co-evolutionary mechanism. The first step is population initialization within the feasible region. Generate 200 candidate solutions internally. Each solution corresponds to a specific combination of environmental parameters. The adaptive evolution phase performs differential mutation operations:

[0112] ;

[0113] in, The coefficient of variation is the new candidate solution generated through the difference mutation operation. Follows the annealed normal distribution ,variance The constraint decays exponentially with the number of iterations. Furthermore, the constraint violation degree is calculated to prioritize individuals with low violation degrees:

[0114] ;

[0115] in, In the optimization model, the first The actual calculated value of each constraint condition For the corresponding number Boundary values ​​of each constraint condition. This is the constraint index. The reference point-oriented selection stage divides the target space into a hyperplane mesh and uses a non-dominated sorting method to retain the solution set closest to the Pareto front. The final output includes... Pareto solution set of group optimization scheme This provides parameterized input for subsequent scene reconstruction, among which, Pareto solution set The first in One optimization scheme, This represents the total number of solution schemes.

[0116] (2) Conditional scene generation

[0117] To address the Pareto solution set of multi-objective optimization outputs, a hierarchical conditional generative AI model was developed to achieve accurate mapping of environmental parameters to 3D scenes. This model is based on a customized modification of the Stable Diffusion XL (SDXL) architecture. First, a hierarchical conditional encoder is designed. A fully connected neural network decouples the environmental feature vectors from the multi-objective optimization output into visual and auditory control signals, which are then concatenated into a conditional embedding as guidance for subsequent generation. Model training utilizes a dataset of RGB images of a campus scene and their corresponding environmental parameter labels. End-to-end optimization is performed by combining pixel-level MSE loss and physical parameter reconstruction loss. The SDXL backbone network is frozen, with only the decoder adaptation layer fine-tuned to ensure a balance between generation quality and parameter controllability. The specific implementation is as follows:

[0118] First, the environmental feature vector is processed by a parameter-learnable neural network decoder. Decoupling into two sets of physical control signals and visual control signals. Regulating vegetation density, spatial permeability, and color saturation; auditory control signals Generate a 1 / 3 octave band sound pressure spectrum, and finally the encoder outputs a spliced ​​vector. As a conditional embedding:

[0119] ;

[0120] A controllable diffusion module is constructed based on the SDXL framework, and its forward process employs a Gaussian noise injection mechanism.

[0121] ;

[0122] in, Let be the noise scheduling parameters of the diffusion model at time t. This is the identity matrix. During the reverse denoising process, conditional convolutional layers are embedded in each layer of the U-Net decoder in SDXL. Cross-modal fusion of physical parameters and generated content is achieved through feature concatenation operations, and the conditional embedding vectors are... Injecting a denoising network:

[0123] ;

[0124] in, The mean function predicted by the denoising neural network. For diffusion model in The fixed noise variance at any given time. Conditional embedding is generated by a fully connected layer performing a nonlinear transformation on the multi-physics control signals. After generating an RGB image, it is converted into a point cloud 3D model through a neural radiation field decoder, simultaneously coupling acoustic simulation and computational fluid dynamics simulation: calling the Odeon SDK to calculate acoustic parameters such as reverberation time, ultimately forming an interactive scene that collaboratively expresses visual and auditory elements.

[0125] (3) Prediction and verification of recovery effect

[0126] A three-level quantization verification process is performed on the generated optimized scene, including feature inversion verification, restoration effect prediction, and response surface construction. The first step is feature inversion verification, which extracts the actual environment feature vectors from the rendered image and calculates their Manhattan distance normalized deviation from the target parameters.

[0127] ;

[0128] in, This is the target environment feature vector output by the multi-objective optimization algorithm. This represents the actual environmental feature vector. The recovery effect prediction part inputs the actual environmental features into a pre-trained Kolmogorov-Arnold network model. :

[0129] ;

[0130] Furthermore, the recovery effect improvement rate is quantified based on baseline scenario predictions. ,in The predicted values ​​for the restorative effects of the original campus environment are given. Finally, a response surface is constructed. For the three-dimensional core parameter space, Latin hypercube sampling is used to generate 2000 uniformly distributed design points. A four-dimensional response hypersurface is then constructed using radial basis function interpolation.

[0131] ;

[0132] in, For the first The weighting coefficients of each sample point in the radial basis function interpolation. For parameter coordinates, The bandwidth parameter of the radial basis function. For the first The coordinates of each sample point in the three-dimensional core parameter space. Finally, significant cooperative regions are identified based on Hessian matrix eigenvalue analysis. ,in, In response to hypersurface At point The Hessian matrix at that location, It is the smallest eigenvalue of the corresponding matrix.

[0133] (4) Dynamic evaluation map construction

[0134] Building upon scenario optimization and effect verification, this step dynamically quantifies the restorative effect using spatiotemporal heat maps, constructing a dynamic restorative assessment map of the campus space. Based on the aforementioned high-precision GPS trajectory data and restorative effect prediction model, a visual representation of the campus space's restoration effect is achieved.

[0135] To achieve a digital representation of the campus space, spatial grid discretization was performed, and a spatial indexing system was constructed using a regular grid discretization method. The target area was divided into a 2m × 2m grid unit system. Grid center point coordinates The interpolation model is accurately generated using a cubic spline interpolation algorithm. Effectively suppress boundary distortion and ensure strict alignment between the grid system and actual geographic coordinates.

[0136] Based on this, a four-dimensional spacetime tensor is constructed. Among them, spatial dimension In the corresponding grid system, T represents the time dimension, and the third dimension with three channels stores the predicted values ​​of restorative effects, environmental feature gradients, and data confidence scores, respectively. This data structure supports rapid retrieval and parallel processing of spatiotemporal information, providing underlying support for dynamic assessment.

[0137] Furthermore, a restorative effect space mapping algorithm is constructed. This is applied to each grid cell. During a specific period To assess the requirements, a multi-level computational process was designed, starting with coupling environmental feature vectors. Compared with standard individual baseline Input the pre-trained improved Kolmogorov-Arnold network and output the normalized predicted restorative effect:

[0138] ;

[0139] This model inherits the weight parameters from previous model training, ensuring the biological consistency of the prediction results with physiologically measured data. For sparse GPS trajectory regions, an inverse distance weighted interpolation algorithm is used for spatial optimization, with the target point... Centered on, search radius All observation points within the range Through the weighting function Calculate the weighted average This algorithm can effectively suppress the boundary abrupt effect in data gap regions while ensuring spatial continuity.

[0140] Finally, based on the gridded restorative effect data, a spatiotemporal heatmap is generated to quantify the restorative effect. First, the cumulative restorative effect function is defined:

[0141] ;

[0142] This function quantizes from the reference time. up to the current moment The accumulated spatial recovery potential. Discretization is used in actual calculations. ,in From the start time At the time Between The specific time point corresponding to each discrete time step. To improve computational efficiency, a GPU parallel optimization architecture was designed. By using a CUDA thread-grid mapping spatial grid system, each thread block is responsible for the computation task of 16×16 grid cells. Combined with shared memory data reuse technology, the time taken to generate a single frame heatmap is compressed to within 50ms.

[0143] The visualization rendering adopts a three-dimensional dynamic representation scheme, with vertical axis mapping accumulating to restore intensity. The planar coordinates correspond to geographical locations. Color coding uses the HSL gradient model, achieving a gradient from red (…). , From dark blue ( , A continuous gradient. Dynamic transparency is set synchronously. Visual enhancement is achieved in regions with complex rate changes.

[0144] like Figure 4 The synergistic recovery effect assessment system provided in this embodiment for implementing the method includes:

[0145] The data acquisition and spatiotemporal alignment module is used to perform step S1 to acquire a spatiotemporally aligned multimodal dataset;

[0146] The environmental feature deep analysis module is used to execute step S2, which parses the environmental data in the multimodal dataset and generates a standardized environmental feature tensor.

[0147] The dynamic modeling module for recovery effects is used to perform step S3, which involves dynamic modeling based on the environmental feature tensor and physiological response data to quantify the relationship between environmental factors and recovery effects.

[0148] The scene optimization and evaluation map construction module is used to perform step S4, optimize and verify the scene based on the output of the dynamic modeling, and construct a dynamic restorative evaluation map.

[0149] Furthermore, the data acquisition and spatiotemporal alignment module includes:

[0150] Path planning unit, used to plan gradient walking paths;

[0151] A multi-sensor synchronous acquisition unit is used to synchronously acquire environmental data, physiological response data, psychological perception data, and high-precision spatiotemporal trajectory data along the path.

[0152] The spatiotemporal alignment processing unit is used to align the data collected by the multi-sensor synchronous acquisition unit in a unified spatiotemporal coordinate system through a time synchronization protocol and a spatial interpolation algorithm.

[0153] The environmental feature deep analysis module includes:

[0154] The visual feature processing unit is used to perform semantic segmentation of visual scenes and calculate building index, landscape index and spatial openness.

[0155] The auditory feature processing unit is used to perform frequency band analysis and instantaneous loudness calculation on sound pressure level data;

[0156] The feature fusion and normalization unit is used to fuse the outputs of the visual feature processing unit and the auditory feature processing unit, and perform normalization processing to generate the environmental feature tensor.

[0157] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in this invention may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for assessing the synergistic restorative effect of the transition between campus buildings and landscape, characterized in that, Includes the following steps: S1: By designing gradient paths and collecting multimodal data synchronously, acquire spatiotemporally aligned campus building-landscape environment data, human physiological response data, psychological perception data, and spatiotemporal trajectory data; S2: Based on the environmental data obtained in step S1, extract visual and auditory features, and construct a standardized environmental feature tensor; S3: Based on the environmental feature tensor obtained in step S2 and the physiological response data obtained in step S1, a dynamic model is obtained by performing dynamic modeling of the recovery effect; the dynamic model is then validated and adaptively fed back using psychological perception data; specifically including: S31: Preprocess the physiological response data obtained in step S1 to obtain corrected physiological data suitable for modeling; S32: Obtain individual baseline characteristics; S33: Using the environmental feature tensor obtained in step S2 as the environmental input, the corrected physiological data obtained in step S31 as the physiological target, and fusing the individual baseline features, a dynamic model for predicting restorative physiological indicators is constructed. S34: Use the psychological perception data obtained in step S1 to verify and adaptively feedback the dynamic model; S35: Based on the dynamic model, quantify the nonlinear mapping relationship between environmental factors and restorative effects, and calculate the cross-sensory synergistic gain index to analyze the synergistic effect of cross-sensory environmental factors. S36: Based on the aforementioned dynamic model, the cumulative dynamic characteristics of restorative effects during the transition from architecture to landscape are analyzed using a sequence modeling method; S4: Based on the output of the dynamic model in step S3, perform multi-objective optimization to generate an environmental transformation plan. Utilize generative artificial intelligence to reconstruct parameterized scenes and verify effects, and construct a dynamic restorative assessment map of the campus space. S41: Based on the cross-sensory synergistic gain index, determine the key environmental factor combination for synergistic enhancement. Maximize the restorative physiological indicators predicted by the dynamic model in step S3 as the first optimization objective, and maximize the spatial openness obtained from step S2 as the second optimization objective. Construct a multi-objective optimization function. Under the conditions of satisfying landscape index constraints, auditory feature constraints, and transformation feasibility constraints, use a multi-objective optimization algorithm to generate a Pareto-optimal environmental parameter solution set. The landscape index constraints and auditory feature constraints are set based on the environmental features obtained from step S2. S42: Input each set of parameters in the Pareto optimal environment parameter solution set generated in step S41 into the conditional generative artificial intelligence model to generate the corresponding three-dimensional visualization scene. S43: For each of the three-dimensional visualization scenes generated in step S42, extract its environmental features and input them into the dynamic model constructed in step S3 to predict its restorative effect value, and compare and verify it with the predicted value of the benchmark scene. S44: Based on the spatiotemporal trajectory data obtained in step S1, the campus space is discretized into a grid system; for each grid, the restorative effect is predicted using the dynamic model constructed in step S3 based on its environmental characteristics, and the spatiotemporal cumulative effect is calculated to generate the dynamic restorative assessment map in the form of a dynamic spatiotemporal heat map.

2. The method according to claim 1, characterized in that, Step S1 includes: S11: Based on the campus spatial topology map, plan multiple gradient transition pedestrian paths from buildings to landscapes; S12: Simultaneously collect environmental data, physiological response data, psychological perception data, and spatiotemporal trajectory data along the walking path; S13: Using a time synchronization protocol and a spatial interpolation matching algorithm, the environmental data, physiological response data, psychological perception data, and spatiotemporal trajectory data collected in step S12 are processed to achieve alignment of multimodal data in a unified spatiotemporal coordinate system.

3. The method according to claim 1, characterized in that, Step S2 includes: S21: Perform image semantic segmentation on the visual scene in the environmental data to obtain building area mask and landscape area mask, and calculate visual features based on the mask. The visual features include at least building index, landscape index and spatial openness. S22: Perform frequency band analysis and loudness calculation on the sound pressure level data in the environmental data to obtain auditory characteristics, wherein the auditory characteristics include at least the sound pressure level and instantaneous loudness of multiple frequency bands; S23: The visual features obtained in step S21 and the auditory features obtained in step S22 are fused to construct an environmental feature tensor, and the environmental feature tensor is standardized to form the standardized environmental feature tensor.

4. The method according to claim 1, characterized in that, The preprocessing in step S31 includes: using a Gaussian weighted integral model to perform physiological delay correction on the heart rate variability index, and using a moving average filter to smooth and denoise the skin conductance activity signal.

5. The method according to claim 1, characterized in that, The dynamic model constructed in step S33 is an improved Kolmogorov-Arnold network model, whose input is a fusion of the environmental feature tensor, the individual baseline features, and the cross features of the two.

6. The method according to claim 4, characterized in that, The sequence modeling method in step S36 employs a bidirectional spatiotemporal long short-term memory network.

7. A system for assessing the synergistic restorative effects of campus building and landscape transitions, used to implement the method described in any one of claims 1-6, characterized in that... include: The data acquisition and spatiotemporal alignment module is used to perform step S1 to acquire a spatiotemporally aligned multimodal dataset; The environmental feature deep analysis module is used to execute step S2, which parses the environmental data in the multimodal dataset and generates a standardized environmental feature tensor. The dynamic modeling module for recovery effects is used to perform step S3, which involves dynamic modeling based on the environmental feature tensor and physiological response data to quantify the relationship between environmental factors and recovery effects. The scene optimization and evaluation map construction module is used to perform step S4, optimize and verify the scene based on the output of the dynamic modeling, and construct a dynamic restorative evaluation map.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1-6.

Images